Network adapters with multiple physical ports and/or SR-IOV capabilities usually support the offload of traffic steering rules between their virtual functions (VFs), physical functions (PFs) and ports.
Like for standard Ethernet switches, this involves a combination of automatic MAC learning and manual configuration. For most purposes it is managed by the host system and fully transparent to users and applications.
On the other hand, applications typically found on hypervisors that process layer 2 (L2) traffic (such as OVS) need to steer traffic themselves according on their own criteria.
Without a standard software interface to manage traffic steering rules between VFs, PFs and the various physical ports of a given device, applications cannot take advantage of these offloads; software processing is mandatory even for traffic which ends up re-injected into the device it originates from.
This document describes how such steering rules can be configured through the DPDK flow API (rte_flow), with emphasis on the SR-IOV use case (PF/VF steering) using a single physical port for clarity, however the same logic applies to any number of ports without necessarily involving SR-IOV.
In many cases, traffic steering rules cannot be determined in advance; applications usually have to process a bit of traffic in software before thinking about offloading specific flows to hardware.
Applications therefore need the ability to receive and inject traffic to various device endpoints (other VFs, PFs or physical ports) before connecting them together. Device drivers must provide means to hook the “other end” of these endpoints and to refer them when configuring flow rules.
This role is left to so-called “port representors” (also known as “VF representors” in the specific context of VFs), which are to DPDK what the Ethernet switch device driver model (switchdev) [1] is to Linux, and which can be thought as a software “patch panel” front-end for applications.
-w pci:dbdf,representor=0
-w pci:dbdf,representor=[0-3]
-w pci:dbdf,representor=[0,5-11]
struct rte_eth_dev_info {
...
uint32_t dev_flags; /**< Device flags */
...
};
/**
* Ethernet device associated switch information
*/
struct rte_eth_switch_info {
const char *name; /**< switch name */
uint16_t domain_id; /**< switch domain id */
uint16_t port_id; /**< switch port id */
};
[1] | Ethernet switch device driver model (switchdev) |
“Basic” in the sense that it is not managed by applications, which nonetheless expect traffic to flow between the various endpoints and the outside as if everything was linked by an Ethernet hub.
The following diagram pictures a setup involving a device with one PF, two VFs and one shared physical port
.-------------. .-------------. .-------------.
| hypervisor | | VM 1 | | VM 2 |
| application | | application | | application |
`--+----------' `----------+--' `--+----------'
| | |
.-----+-----. | |
| port_id 3 | | |
`-----+-----' | |
| | |
.-+--. .---+--. .--+---.
| PF | | VF 1 | | VF 2 |
`-+--' `---+--' `--+---'
| | |
`---------. .-----------------------' |
| | .-------------------------'
| | |
.--+-----+-----+--.
| interconnection |
`--------+--------'
|
.----+-----.
| physical |
| port 0 |
`----------'
Subsequent sections in this document describe means for DPDK applications running on the hypervisor to freely assign specific flows between PF, VFs and physical ports based on traffic properties, by managing this interconnection.
When a DPDK application gets assigned a PF device and is deliberately not started in basic SR-IOV mode, any traffic coming from physical ports is received by PF according to default rules, while VFs remain isolated.
.-------------. .-------------. .-------------.
| hypervisor | | VM 1 | | VM 2 |
| application | | application | | application |
`--+----------' `----------+--' `--+----------'
| | |
.-----+-----. | |
| port_id 3 | | |
`-----+-----' | |
| | |
.-+--. .---+--. .--+---.
| PF | | VF 1 | | VF 2 |
`-+--' `------' `------'
|
`-----.
|
.--+----------------------.
| managed interconnection |
`------------+------------'
|
.----+-----.
| physical |
| port 0 |
`----------'
In this mode, interconnection must be configured by the application to enable VF communication, for instance by explicitly directing traffic with a given destination MAC address to VF 1 and allowing that with the same source MAC address to come out of it.
For this to work, hypervisor applications need a way to refer to either VF 1 or VF 2 in addition to the PF. This is addressed by VF representors.
VF representors are virtual but standard DPDK network devices (albeit with limited capabilities) created by PMDs when managing a PF device.
Since they represent VF instances used by other applications, configuring them (e.g. assigning a MAC address or setting up promiscuous mode) affects interconnection accordingly. If supported, they may also be used as two-way communication ports with VFs (assuming switchdev topology)
.-------------. .-------------. .-------------.
| hypervisor | | VM 1 | | VM 2 |
| application | | application | | application |
`--+---+---+--' `----------+--' `--+----------'
| | | | |
| | `-------------------. | |
| `---------. | | |
| | | | |
.-----+-----. .-----+-----. .-----+-----. | |
| port_id 3 | | port_id 4 | | port_id 5 | | |
`-----+-----' `-----+-----' `-----+-----' | |
| | | | |
.-+--. .-----+-----. .-----+-----. .---+--. .--+---.
| PF | | VF 1 rep. | | VF 2 rep. | | VF 1 | | VF 2 |
`-+--' `-----+-----' `-----+-----' `---+--' `--+---'
| | | | |
| | .---------' | |
`-----. | | .-----------------' |
| | | | .---------------------'
| | | | |
.--+-------+---+---+---+--.
| managed interconnection |
`------------+------------'
|
.----+-----.
| physical |
| port 0 |
`----------'
As previously described, the “interconnection” block represents a logical concept. Interconnection occurs when hardware configuration enables traffic flows from one place to another (e.g. physical port 0 to VF 1) according to some criteria.
This is discussed in more detail in traffic steering.
In the following diagram, each meaningful traffic origin or endpoint as seen by the hypervisor application is tagged with a unique letter from A to F.
.-------------. .-------------. .-------------.
| hypervisor | | VM 1 | | VM 2 |
| application | | application | | application |
`--+---+---+--' `----------+--' `--+----------'
| | | | |
| | `-------------------. | |
| `---------. | | |
| | | | |
.----(A)----. .----(B)----. .----(C)----. | |
| port_id 3 | | port_id 4 | | port_id 5 | | |
`-----+-----' `-----+-----' `-----+-----' | |
| | | | |
.-+--. .-----+-----. .-----+-----. .---+--. .--+---.
| PF | | VF 1 rep. | | VF 2 rep. | | VF 1 | | VF 2 |
`-+--' `-----+-----' `-----+-----' `--(D)-' `-(E)--'
| | | | |
| | .---------' | |
`-----. | | .-----------------' |
| | | | .---------------------'
| | | | |
.--+-------+---+---+---+--.
| managed interconnection |
`------------+------------'
|
.---(F)----.
| physical |
| port 0 |
`----------'
Although uncommon, some devices do not enforce a one to one mapping between PF and physical ports. For instance, by default all ports of mlx4 adapters are available to all their PF/VF instances, in which case additional ports appear next to F in the above diagram.
Assuming no interconnection is provided by default in this mode, setting up a basic SR-IOV configuration involving physical port 0 could be broken down as:
PF:
VF 1:
VF 2:
Devices may additionally support advanced matching criteria such as IPv4/IPv6 addresses or TCP/UDP ports.
The combination of matching criteria with target endpoints fits well with rte_flow [6], which expresses flow rules as combinations of patterns and actions.
Enhancing rte_flow with the ability to make flow rules match and target these endpoints provides a standard interface to manage their interconnection without introducing new concepts and whole new API to implement them. This is described in flow API (rte_flow).
[6] | Generic flow API (rte_flow) |
Compared to creating a brand new dedicated interface, rte_flow was deemed flexible enough to manage representor traffic only with minor extensions:
For advanced uses:
From an application standpoint, “ingress” and “egress” flow rule attributes apply to the DPDK port ID they are associated with. They select a traffic direction for matching patterns, but have no impact on actions.
When matching traffic coming from or going to a different place than the immediate port ID a flow rule is associated with, these attributes keep their meaning while applying to the chosen origin, as highlighted by the following diagram
.-------------. .-------------. .-------------.
| hypervisor | | VM 1 | | VM 2 |
| application | | application | | application |
`--+---+---+--' `----------+--' `--+----------'
| | | | |
| | `-------------------. | |
| `---------. | | |
| ^ | ^ | ^ | |
| | ingress | | ingress | | ingress | |
| | egress | | egress | | egress | |
| v | v | v | |
.----(A)----. .----(B)----. .----(C)----. | |
| port_id 3 | | port_id 4 | | port_id 5 | | |
`-----+-----' `-----+-----' `-----+-----' | |
| | | | |
.-+--. .-----+-----. .-----+-----. .---+--. .--+---.
| PF | | VF 1 rep. | | VF 2 rep. | | VF 1 | | VF 2 |
`-+--' `-----+-----' `-----+-----' `--(D)-' `-(E)--'
| | | ^ | | ^
| | | egress | | | | egress
| | | ingress | | | | ingress
| | .---------' v | | v
`-----. | | .-----------------' |
| | | | .---------------------'
| | | | |
.--+-------+---+---+---+--.
| managed interconnection |
`------------+------------'
^ |
ingress | |
egress | |
v |
.---(F)----.
| physical |
| port 0 |
`----------'
Ingress and egress are defined as relative to the application creating the flow rule.
For instance, matching traffic sent by VM 2 would be done through an ingress flow rule on VF 2 (E). Likewise for incoming traffic on physical port (F). This also applies to C and A respectively.
Traffic direction describes how an application could match traffic coming from or going to a specific place reachable from a DPDK port ID. This makes sense when the traffic in question is normally seen (i.e. sent or received) by the application creating the flow rule (e.g. as in “redirect all traffic coming from VF 1 to local queue 6”).
However this does not force such traffic to take a specific route. Creating a flow rule on A matching traffic coming from D is only meaningful if it can be received by A in the first place, otherwise doing so simply has no effect.
A new flow rule attribute named “transfer” is necessary for that. Combining it with “ingress” or “egress” and a specific origin requests a flow rule to be applied at the lowest level
ingress only : ingress + transfer
:
.-------------. .-------------. : .-------------. .-------------.
| hypervisor | | VM 1 | : | hypervisor | | VM 1 |
| application | | application | : | application | | application |
`------+------' `--+----------' : `------+------' `--+----------'
| | | traffic : | | | traffic
.----(A)----. | v : .----(A)----. | v
| port_id 3 | | : | port_id 3 | |
`-----+-----' | : `-----+-----' |
| | : | ^ |
| | : | | traffic |
.-+--. .---+--. : .-+--. .---+--.
| PF | | VF 1 | : | PF | | VF 1 |
`-+--' `--(D)-' : `-+--' `--(D)-'
| | | traffic : | ^ | | traffic
| | v : | | traffic | v
.--+-----------+--. : .--+-----------+--.
| interconnection | : | interconnection |
`--------+--------' : `--------+--------'
| | traffic : |
| v : |
.---(F)----. : .---(F)----.
| physical | : | physical |
| port 0 | : | port 0 |
`----------' : `----------'
With “ingress” only, traffic is matched on A thus still goes to physical port F by default
testpmd> flow create 3 ingress pattern vf id is 1 / end
actions queue index 6 / end
With “ingress + transfer”, traffic is matched on D and is therefore successfully assigned to queue 6 on A
testpmd> flow create 3 ingress transfer pattern vf id is 1 / end
actions queue index 6 / end
When port representors exist, implicit flow rules with the “transfer” attribute (described in without port representors) are be assumed to exist between them and their represented resources. These may be immutable.
In this case, traffic is received by default through the representor and neither the “transfer” attribute nor traffic origin in flow rule patterns are necessary. They simply have to be created on the representor port directly and may target a different representor as described in PORT_ID action.
Implicit traffic flow with port representor
.-------------. .-------------.
| hypervisor | | VM 1 |
| application | | application |
`--+-------+--' `----------+--'
| | ^ | | traffic
| | | traffic | v
| `-----. |
| | |
.----(A)----. .----(B)----. |
| port_id 3 | | port_id 4 | |
`-----+-----' `-----+-----' |
| | |
.-+--. .-----+-----. .---+--.
| PF | | VF 1 rep. | | VF 1 |
`-+--' `-----+-----' `--(D)-'
| | |
.--|-------------|-----------|--.
| | | | |
| | `-----------' |
| | <-- traffic |
`--|----------------------------'
|
.---(F)----.
| physical |
| port 0 |
`----------'
Matches traffic originating from (ingress) or going to (egress) a physical port of the underlying device.
Using this pattern item without specifying a port index matches the physical port associated with the current DPDK port ID by default. As described in traffic steering, specifying it should be rarely needed.
Directs matching traffic to a given physical port index.
Matches traffic originating from (ingress) or going to (egress) a given DPDK port ID.
Normally only supported if the port ID in question is known by the underlying PMD and related to the device the flow rule is created against.
This must not be confused with the PORT pattern item which refers to the physical port of a device. PORT_ID refers to a struct rte_eth_dev object on the application side (also known as “port representor” depending on the kind of underlying device).
Directs matching traffic to a given DPDK port ID.
Same restrictions as PORT_ID pattern item.
Matches traffic originating from (ingress) or going to (egress) the physical function of the current device.
If supported, should work even if the physical function is not managed by the application and thus not associated with a DPDK port ID. Its behavior is otherwise similar to PORT_ID pattern item using PF port ID.
Directs matching traffic to the physical function of the current device.
Same restrictions as PF pattern item.
Matches traffic originating from (ingress) or going to (egress) a given virtual function of the current device.
If supported, should work even if the virtual function is not managed by the application and thus not associated with a DPDK port ID. Its behavior is otherwise similar to PORT_ID pattern item using VF port ID.
Note this pattern item does not match VF representors traffic which, as separate entities, should be addressed through their own port IDs.
Directs matching traffic to a given virtual function of the current device.
Same restrictions as VF pattern item.
These actions are named according to the protocol they encapsulate traffic with (e.g. VXLAN_ENCAP) and using specific parameters (e.g. VNI for VXLAN).
While they modify traffic and can be used multiple times (order matters), unlike PORT_ID action and friends, they have no impact on steering.
As described in actions order and repetition this means they are useless if used alone in an action list, the resulting traffic gets dropped unless combined with either PASSTHRU or other endpoint-targeting actions.
They perform the reverse of *_ENCAP actions by popping protocol headers from traffic instead of pushing them. They can be used multiple times as well.
Note that using these actions on non-matching traffic results in undefined behavior. It is recommended to match the protocol headers to decapsulate on the pattern side of a flow rule in order to use these actions or otherwise make sure only matching traffic goes through.
Flow rules are currently restricted to at most a single action of each supported type, performed in an unpredictable order (or all at once). To repeat actions in a predictable fashion, applications have to make rules pass-through and use priority levels.
It’s now clear that PMD support for chaining multiple non-terminating flow rules of varying priority levels is prohibitively difficult to implement compared to simply allowing multiple identical actions performed in a defined order by a single flow rule.
This section provides practical examples based on the established testpmd flow command syntax [2], in the context described in traffic steering
.-------------. .-------------. .-------------.
| hypervisor | | VM 1 | | VM 2 |
| application | | application | | application |
`--+---+---+--' `----------+--' `--+----------'
| | | | |
| | `-------------------. | |
| `---------. | | |
| | | | |
.----(A)----. .----(B)----. .----(C)----. | |
| port_id 3 | | port_id 4 | | port_id 5 | | |
`-----+-----' `-----+-----' `-----+-----' | |
| | | | |
.-+--. .-----+-----. .-----+-----. .---+--. .--+---.
| PF | | VF 1 rep. | | VF 2 rep. | | VF 1 | | VF 2 |
`-+--' `-----+-----' `-----+-----' `--(D)-' `-(E)--'
| | | | |
| | .---------' | |
`-----. | | .-----------------' |
| | | | .---------------------'
| | | | |
.--|-------|---|---|---|--.
| | | `---|---' |
| | `-------' |
| `---------. |
`------------|------------'
|
.---(F)----.
| physical |
| port 0 |
`----------'
By default, PF (A) can communicate with the physical port it is associated with (F), while VF 1 (D) and VF 2 (E) are isolated and restricted to communicate with the hypervisor application through their respective representors (B and C) if supported.
Examples in subsequent sections apply to hypervisor applications only and are based on port representors A, B and C.
[2] | Flow syntax |
Assign all port traffic (F) to VF 1 (D) indiscriminately through their representors
flow create 3 ingress pattern / end actions port_id id 4 / end
flow create 4 ingress pattern / end actions port_id id 3 / end
More practical example with MAC address restrictions
flow create 3 ingress
pattern eth dst is {VF 1 MAC} / end
actions port_id id 4 / end
flow create 4 ingress
pattern eth src is {VF 1 MAC} / end
actions port_id id 3 / end
From outside to PF and VFs
flow create 3 ingress
pattern eth dst is ff:ff:ff:ff:ff:ff / end
actions port_id id 3 / port_id id 4 / port_id id 5 / end
Note port_id id 3 is necessary otherwise only VFs would receive matching traffic.
From PF to outside and VFs
flow create 3 egress
pattern eth dst is ff:ff:ff:ff:ff:ff / end
actions port / port_id id 4 / port_id id 5 / end
From VFs to outside and PF
flow create 4 ingress
pattern eth dst is ff:ff:ff:ff:ff:ff src is {VF 1 MAC} / end
actions port_id id 3 / port_id id 5 / end
flow create 5 ingress
pattern eth dst is ff:ff:ff:ff:ff:ff src is {VF 2 MAC} / end
actions port_id id 4 / port_id id 4 / end
Similar 33:33:* rules based on known MAC addresses should be added for IPv6 traffic.
Assuming pass-through flow rules are supported
flow create 5 ingress
pattern eth / end
actions vxlan_encap vni 42 / passthru / end
flow create 5 egress
pattern vxlan vni is 42 / end
actions vxlan_decap / passthru / end
Here passthru is needed since as described in actions order and repetition, flow rules are otherwise terminating; if supported, a rule without a target endpoint will drop traffic.
Without pass-through support, ingress encapsulation on the destination endpoint might not be supported and action list must provide one
flow create 5 ingress
pattern eth src is {VF 2 MAC} / end
actions vxlan_encap vni 42 / port_id id 3 / end
flow create 3 ingress
pattern vxlan vni is 42 / end
actions vxlan_decap / port_id id 5 / end