The MLX5 poll mode driver library (librte_pmd_mlx5) provides support for Mellanox ConnectX-4, Mellanox ConnectX-4 Lx , Mellanox ConnectX-5, Mellanox ConnectX-6, Mellanox ConnectX-6 Dx and Mellanox BlueField families of 10/25/40/50/100/200 Gb/s adapters as well as their virtual functions (VF) in SR-IOV context.
Information and documentation about these adapters can be found on the Mellanox website. Help is also provided by the Mellanox community.
There is also a section dedicated to this poll mode driver.
Note
Due to external dependencies, this driver is disabled in default configuration of the “make” build. It can be enabled with CONFIG_RTE_LIBRTE_MLX5_PMD=y or by using “meson” build system which will detect dependencies.
Besides its dependency on libibverbs (that implies libmlx5 and associated kernel support), librte_pmd_mlx5 relies heavily on system calls for control operations such as querying/updating the MTU and flow control parameters.
For security reasons and robustness, this driver only deals with virtual memory addresses. The way resources allocations are handled by the kernel, combined with hardware specifications that allow to handle virtual memory addresses directly, ensure that DPDK applications cannot access random physical memory (or memory that does not belong to the current process).
This capability allows the PMD to coexist with kernel network interfaces which remain functional, although they stop receiving unicast packets as long as they share the same MAC address. This means legacy linux control tools (for example: ethtool, ifconfig and more) can operate on the same network interfaces that owned by the DPDK application.
The PMD can use libibverbs and libmlx5 to access the device firmware or directly the hardware components. There are different levels of objects and bypassing abilities to get the best performances:
Enabling librte_pmd_mlx5 causes DPDK applications to be linked against libibverbs.
For secondary process:
When using Verbs flow engine (dv_flow_en = 0), flow pattern without any specific VLAN will match for VLAN packets as well:
When VLAN spec is not specified in the pattern, the matching rule will be created with VLAN as a wild card. Meaning, the flow rule:
flow create 0 ingress pattern eth / vlan vid is 3 / ipv4 / end ...
Will only match vlan packets with vid=3. and the flow rule:
flow create 0 ingress pattern eth / ipv4 / end ...
Will match any ipv4 packet (VLAN included).
When using DV flow engine (dv_flow_en = 1), flow pattern without VLAN item will match untagged packets only. The flow rule:
flow create 0 ingress pattern eth / ipv4 / end ...
Will match untagged packets only. The flow rule:
flow create 0 ingress pattern eth / vlan / ipv4 / end ...
Will match tagged packets only, with any VLAN ID value. The flow rule:
flow create 0 ingress pattern eth / vlan vid is 3 / ipv4 / end ...
Will only match tagged packets with VLAN ID 3.
VLAN pop offload command:
VLAN push offload is not supported on ingress traffic.
VLAN set PCP offload is not supported on existing headers.
A multi segment packet must have not more segments than reported by dev_infos_get() in tx_desc_lim.nb_seg_max field. This value depends on maximal supported Tx descriptor size and txq_inline_min settings and may be from 2 (worst case forced by maximal inline settings) to 58.
Flows with a VXLAN Network Identifier equal (or ends to be equal) to 0 are not supported.
L3 VXLAN and VXLAN-GPE tunnels cannot be supported together with MPLSoGRE and MPLSoUDP.
Match on Geneve header supports the following fields only:
- VNI
- OAM
- protocol type
- options length Currently, the only supported options length value is 0.
VF: flow rules created on VF devices can only match traffic targeted at the configured MAC addresses (see rte_eth_dev_mac_addr_add()).
Match on GTP tunnel header item supports the following fields only:
- v_pt_rsv_flags: E flag, S flag, PN flag
- msg_type
- teid
No Tx metadata go to the E-Switch steering domain for the Flow group 0. The flows within group 0 and set metadata action are rejected by hardware.
Note
MAC addresses not already present in the bridge table of the associated kernel network device will be added and cleaned up by the PMD when closing the device. In case of ungraceful program termination, some entries may remain present and should be removed manually by other means.
When Multi-Packet Rx queue is configured (mprq_en), a Rx packet can be externally attached to a user-provided mbuf with having EXT_ATTACHED_MBUF in ol_flags. As the mempool for the external buffer is managed by PMD, all the Rx mbufs must be freed before the device is closed. Otherwise, the mempool of the external buffers will be freed by PMD and the application which still holds the external buffers may be corrupted.
If Multi-Packet Rx queue is configured (mprq_en) and Rx CQE compression is enabled (rxq_cqe_comp_en) at the same time, RSS hash result is not fully supported. Some Rx packets may not have PKT_RX_RSS_HASH.
IPv6 Multicast messages are not supported on VM, while promiscuous mode and allmulticast mode are both set to off. To receive IPv6 Multicast messages on VM, explicitly set the relevant MAC address using rte_eth_dev_mac_addr_add() API.
To support a mixed traffic pattern (some buffers from local host memory, some buffers from other devices) with high bandwidth, a mbuf flag is used.
An application hints the PMD whether or not it should try to inline the given mbuf data buffer. PMD should do the best effort to act upon this request.
The hint flag RTE_PMD_MLX5_FINE_GRANULARITY_INLINE is dynamic, registered by application with rte_mbuf_dynflag_register(). This flag is purely driver-specific and declared in PMD specific header rte_pmd_mlx5.h, which is intended to be used by the application.
To query the supported specific flags in runtime, the function rte_pmd_mlx5_get_dyn_flag_names returns the array of currently (over present hardware and configuration) supported specific flags. The “not inline hint” feature operating flow is the following one:
- application starts
- probe the devices, ports are created
- query the port capabilities
- if port supporting the feature is found
- register dynamic flag RTE_PMD_MLX5_FINE_GRANULARITY_INLINE
- application starts the ports
- on dev_start() PMD checks whether the feature flag is registered and enables the feature support in datapath
- application might set the registered flag bit in ol_flags field of mbuf being sent and PMD will handle ones appropriately.
The amount of descriptors in Tx queue may be limited by data inline settings. Inline data require the more descriptor building blocks and overall block amount may exceed the hardware supported limits. The application should reduce the requested Tx size or adjust data inline settings with txq_inline_max and txq_inline_mpw devargs keys.
E-Switch decapsulation Flow:
E-Switch encapsulation Flow:
Raw encapsulation:
Raw decapsulation:
ICMP/ICMP6 code/type matching, IP-in-IP and MPLS flow matching are all mutually exclusive features which cannot be supported together (see Firmware configuration).
LRO:
Requires DevX and DV flow to be enabled.
KEEP_CRC offload cannot be supported with LRO.
The first mbuf length, without head-room, must be big enough to include the TCP header (122B).
Rx queue with LRO offload enabled, receiving a non-LRO packet, can forward it with size limited to max LRO size, not to max RX packet length.
eth (with or without vlan) / ipv4 or ipv6 / tcp / payload
Other TCP packets (e.g. with MPLS label) received on Rx queue with LRO enabled, will be received with bad checksum.
KEEP_CRC:
Sample flow:
MLX5 supports various methods to report statistics:
Port statistics can be queried using rte_eth_stats_get(). The received and sent statistics are through SW only and counts the number of packets received or sent successfully by the PMD. The imissed counter is the amount of packets that could not be delivered to SW because a queue was full. Packets not received due to congestion in the bus or on the NIC can be queried via the rx_discards_phy xstats counter.
Extended statistics can be queried using rte_eth_xstats_get(). The extended statistics expose a wider set of counters counted by the device. The extended port statistics counts the number of packets received or sent successfully by the port. As Mellanox NICs are using the Bifurcated Linux Driver those counters counts also packet received or sent by the Linux kernel. The counters with _phy suffix counts the total events on the physical port, therefore not valid for VF.
Finally per-flow statistics can by queried using rte_flow_query when attaching a count action for specific flow. The flow counter counts the number of packets received successfully by the port and match the specific flow.
These options can be modified in the .config file.
CONFIG_RTE_LIBRTE_MLX5_PMD (default n)
Toggle compilation of librte_pmd_mlx5 itself.
CONFIG_RTE_IBVERBS_LINK_DLOPEN (default n)
Build PMD with additional code to make it loadable without hard dependencies on libibverbs nor libmlx5, which may not be installed on the target system.
In this mode, their presence is still required for it to run properly, however their absence won’t prevent a DPDK application from starting (with CONFIG_RTE_BUILD_SHARED_LIB disabled) and they won’t show up as missing with ldd(1).
It works by moving these dependencies to a purpose-built rdma-core “glue” plug-in which must either be installed in a directory whose name is based on CONFIG_RTE_EAL_PMD_PATH suffixed with -glue if set, or in a standard location for the dynamic linker (e.g. /lib) if left to the default empty string ("").
This option has no performance impact.
CONFIG_RTE_IBVERBS_LINK_STATIC (default n)
Embed static flavor of the dependencies libibverbs and libmlx5 in the PMD shared library or the executable static binary.
CONFIG_RTE_LIBRTE_MLX5_DEBUG (default n)
Toggle debugging code and stricter compilation flags. Enabling this option adds additional run-time checks and debugging messages at the cost of lower performance.
Note
For BlueField, target should be set to arm64-bluefield-linux-gcc. This will enable CONFIG_RTE_LIBRTE_MLX5_PMD and set RTE_CACHE_LINE_SIZE to 64. Default armv8a configuration of make build and meson build set it to 128 then brings performance degradation.
This option is available in meson:
MLX5_GLUE_PATH
A list of directories in which to search for the rdma-core “glue” plug-in, separated by colons or semi-colons.
Only matters when compiled with CONFIG_RTE_IBVERBS_LINK_DLOPEN enabled and most useful when CONFIG_RTE_EAL_PMD_PATH is also set, since LD_LIBRARY_PATH has no effect in this case.
MLX5_SHUT_UP_BF
Configures HW Tx doorbell register as IO-mapped.
By default, the HW Tx doorbell is configured as a write-combining register. The register would be flushed to HW usually when the write-combining buffer becomes full, but it depends on CPU design.
Except for vectorized Tx burst routines, a write memory barrier is enforced after updating the register so that the update can be immediately visible to HW.
When vectorized Tx burst is called, the barrier is set only if the burst size is not aligned to MLX5_VPMD_TX_MAX_BURST. However, setting this environmental variable will bring better latency even though the maximum throughput can slightly decline.
In order to run as a non-root user, some capabilities must be granted to the application:
setcap cap_sys_admin,cap_net_admin,cap_net_raw,cap_ipc_lock+ep <dpdk-app>
Below are the reasons of the need for each capability:
rxq_cqe_comp_en parameter [int]
A nonzero value enables the compression of CQE on RX side. This feature allows to save PCI bandwidth and improve performance. Enabled by default.
Supported on:
rxq_cqe_pad_en parameter [int]
A nonzero value enables 128B padding of CQE on RX side. The size of CQE is aligned with the size of a cacheline of the core. If cacheline size is 128B, the CQE size is configured to be 128B even though the device writes only 64B data on the cacheline. This is to avoid unnecessary cache invalidation by device’s two consecutive writes on to one cacheline. However in some architecture, it is more beneficial to update entire cacheline with padding the rest 64B rather than striding because read-modify-write could drop performance a lot. On the other hand, writing extra data will consume more PCIe bandwidth and could also drop the maximum throughput. It is recommended to empirically set this parameter. Disabled by default.
Supported on:
rxq_pkt_pad_en parameter [int]
A nonzero value enables padding Rx packet to the size of cacheline on PCI transaction. This feature would waste PCI bandwidth but could improve performance by avoiding partial cacheline write which may cause costly read-modify-copy in memory transaction on some architectures. Disabled by default.
Supported on:
mprq_en parameter [int]
A nonzero value enables configuring Multi-Packet Rx queues. Rx queue is configured as Multi-Packet RQ if the total number of Rx queues is rxqs_min_mprq or more. Disabled by default.
Multi-Packet Rx Queue (MPRQ a.k.a Striding RQ) can further save PCIe bandwidth by posting a single large buffer for multiple packets. Instead of posting a buffers per a packet, one large buffer is posted in order to receive multiple packets on the buffer. A MPRQ buffer consists of multiple fixed-size strides and each stride receives one packet. MPRQ can improve throughput for small-packet traffic.
When MPRQ is enabled, max_rx_pkt_len can be larger than the size of user-provided mbuf even if DEV_RX_OFFLOAD_SCATTER isn’t enabled. PMD will configure large stride size enough to accommodate max_rx_pkt_len as long as device allows. Note that this can waste system memory compared to enabling Rx scatter and multi-segment packet.
mprq_log_stride_num parameter [int]
Log 2 of the number of strides for Multi-Packet Rx queue. Configuring more strides can reduce PCIe traffic further. If configured value is not in the range of device capability, the default value will be set with a warning message. The default value is 4 which is 16 strides per a buffer, valid only if mprq_en is set.
The size of Rx queue should be bigger than the number of strides.
mprq_log_stride_size parameter [int]
Log 2 of the size of a stride for Multi-Packet Rx queue. Configuring a smaller stride size can save some memory and reduce probability of a depletion of all available strides due to unreleased packets by an application. If configured value is not in the range of device capability, the default value will be set with a warning message. The default value is 11 which is 2048 bytes per a stride, valid only if mprq_en is set. With mprq_log_stride_size set it is possible for a pcaket to span across multiple strides. This mode allows support of jumbo frames (9K) with MPRQ. The memcopy of some packets (or part of a packet if Rx scatter is configured) may be required in case there is no space left for a head room at the end of a stride which incurs some performance penalty.
mprq_max_memcpy_len parameter [int]
The maximum length of packet to memcpy in case of Multi-Packet Rx queue. Rx packet is mem-copied to a user-provided mbuf if the size of Rx packet is less than or equal to this parameter. Otherwise, PMD will attach the Rx packet to the mbuf by external buffer attachment - rte_pktmbuf_attach_extbuf(). A mempool for external buffers will be allocated and managed by PMD. If Rx packet is externally attached, ol_flags field of the mbuf will have EXT_ATTACHED_MBUF and this flag must be preserved. RTE_MBUF_HAS_EXTBUF() checks the flag. The default value is 128, valid only if mprq_en is set.
rxqs_min_mprq parameter [int]
Configure Rx queues as Multi-Packet RQ if the total number of Rx queues is greater or equal to this value. The default value is 12, valid only if mprq_en is set.
txq_inline parameter [int]
Amount of data to be inlined during TX operations. This parameter is deprecated and converted to the new parameter txq_inline_max providing partial compatibility.
txqs_min_inline parameter [int]
Enable inline data send only when the number of TX queues is greater or equal to this value.
This option should be used in combination with txq_inline_max and txq_inline_mpw below and does not affect txq_inline_min settings above.
If this option is not specified the default value 16 is used for BlueField and 8 for other platforms
The data inlining consumes the CPU cycles, so this option is intended to auto enable inline data if we have enough Tx queues, which means we have enough CPU cores and PCI bandwidth is getting more critical and CPU is not supposed to be bottleneck anymore.
The copying data into WQE improves latency and can improve PPS performance when PCI back pressure is detected and may be useful for scenarios involving heavy traffic on many queues.
Because additional software logic is necessary to handle this mode, this option should be used with care, as it may lower performance when back pressure is not expected.
If inline data are enabled it may affect the maximal size of Tx queue in descriptors because the inline data increase the descriptor size and queue size limits supported by hardware may be exceeded.
txq_inline_min parameter [int]
Minimal amount of data to be inlined into WQE during Tx operations. NICs may require this minimal data amount to operate correctly. The exact value may depend on NIC operation mode, requested offloads, etc. It is strongly recommended to omit this parameter and use the default values. Anyway, applications using this parameter should take into consideration that specifying an inconsistent value may prevent the NIC from sending packets.
If txq_inline_min key is present the specified value (may be aligned by the driver in order not to exceed the limits and provide better descriptor space utilization) will be used by the driver and it is guaranteed that requested amount of data bytes are inlined into the WQE beside other inline settings. This key also may update txq_inline_max value (default or specified explicitly in devargs) to reserve the space for inline data.
If txq_inline_min key is not present, the value may be queried by the driver from the NIC via DevX if this feature is available. If there is no DevX enabled/supported the value 18 (supposing L2 header including VLAN) is set for ConnectX-4 and ConnectX-4 Lx, and 0 is set by default for ConnectX-5 and newer NICs. If packet is shorter the txq_inline_min value, the entire packet is inlined.
For ConnectX-4 NIC, driver does not allow specifying value below 18 (minimal L2 header, including VLAN), error will be raised.
For ConnectX-4 Lx NIC, it is allowed to specify values below 18, but it is not recommended and may prevent NIC from sending packets over some configurations.
Please, note, this minimal data inlining disengages eMPW feature (Enhanced Multi-Packet Write), because last one does not support partial packet inlining. This is not very critical due to minimal data inlining is mostly required by ConnectX-4 and ConnectX-4 Lx, these NICs do not support eMPW feature.
txq_inline_max parameter [int]
Specifies the maximal packet length to be completely inlined into WQE Ethernet Segment for ordinary SEND method. If packet is larger than specified value, the packet data won’t be copied by the driver at all, data buffer is addressed with a pointer. If packet length is less or equal all packet data will be copied into WQE. This may improve PCI bandwidth utilization for short packets significantly but requires the extra CPU cycles.
The data inline feature is controlled by number of Tx queues, if number of Tx queues is larger than txqs_min_inline key parameter, the inline feature is engaged, if there are not enough Tx queues (which means not enough CPU cores and CPU resources are scarce), data inline is not performed by the driver. Assigning txqs_min_inline with zero always enables the data inline.
The default txq_inline_max value is 290. The specified value may be adjusted by the driver in order not to exceed the limit (930 bytes) and to provide better WQE space filling without gaps, the adjustment is reflected in the debug log. Also, the default value (290) may be decreased in run-time if the large transmit queue size is requested and hardware does not support enough descriptor amount, in this case warning is emitted. If txq_inline_max key is specified and requested inline settings can not be satisfied then error will be raised.
txq_inline_mpw parameter [int]
Specifies the maximal packet length to be completely inlined into WQE for Enhanced MPW method. If packet is large the specified value, the packet data won’t be copied, and data buffer is addressed with pointer. If packet length is less or equal, all packet data will be copied into WQE. This may improve PCI bandwidth utilization for short packets significantly but requires the extra CPU cycles.
The data inline feature is controlled by number of TX queues, if number of Tx queues is larger than txqs_min_inline key parameter, the inline feature is engaged, if there are not enough Tx queues (which means not enough CPU cores and CPU resources are scarce), data inline is not performed by the driver. Assigning txqs_min_inline with zero always enables the data inline.
The default txq_inline_mpw value is 268. The specified value may be adjusted by the driver in order not to exceed the limit (930 bytes) and to provide better WQE space filling without gaps, the adjustment is reflected in the debug log. Due to multiple packets may be included to the same WQE with Enhanced Multi Packet Write Method and overall WQE size is limited it is not recommended to specify large values for the txq_inline_mpw. Also, the default value (268) may be decreased in run-time if the large transmit queue size is requested and hardware does not support enough descriptor amount, in this case warning is emitted. If txq_inline_mpw key is specified and requested inline settings can not be satisfied then error will be raised.
txqs_max_vec parameter [int]
Enable vectorized Tx only when the number of TX queues is less than or equal to this value. This parameter is deprecated and ignored, kept for compatibility issue to not prevent driver from probing.
txq_mpw_hdr_dseg_en parameter [int]
A nonzero value enables including two pointers in the first block of TX descriptor. The parameter is deprecated and ignored, kept for compatibility issue.
txq_max_inline_len parameter [int]
Maximum size of packet to be inlined. This limits the size of packet to be inlined. If the size of a packet is larger than configured value, the packet isn’t inlined even though there’s enough space remained in the descriptor. Instead, the packet is included with pointer. This parameter is deprecated and converted directly to txq_inline_mpw providing full compatibility. Valid only if eMPW feature is engaged.
txq_mpw_en parameter [int]
A nonzero value enables Enhanced Multi-Packet Write (eMPW) for ConnectX-5, ConnectX-6, ConnectX-6 Dx and BlueField. eMPW allows the TX burst function to pack up multiple packets in a single descriptor session in order to save PCI bandwidth and improve performance at the cost of a slightly higher CPU usage. When txq_inline_mpw is set along with txq_mpw_en, TX burst function copies entire packet data on to TX descriptor instead of including pointer of packet.
The Enhanced Multi-Packet Write feature is enabled by default if NIC supports it, can be disabled by explicit specifying 0 value for txq_mpw_en option. Also, if minimal data inlining is requested by non-zero txq_inline_min option or reported by the NIC, the eMPW feature is disengaged.
tx_db_nc parameter [int]
The rdma core library can map doorbell register in two ways, depending on the environment variable “MLX5_SHUT_UP_BF”:
The type of mapping may slightly affect the Tx performance, the optimal choice is strongly relied on the host architecture and should be deduced practically.
If tx_db_nc is set to zero, the doorbell is forced to be mapped to regular memory (with write combining), the PMD will perform the extra write memory barrier after writing to doorbell, it might increase the needed CPU clocks per packet to send, but latency might be improved.
If tx_db_nc is set to one, the doorbell is forced to be mapped to non cached memory, the PMD will not perform the extra write memory barrier after writing to doorbell, on some architectures it might improve the performance.
If tx_db_nc is set to two, the doorbell is forced to be mapped to regular memory, the PMD will use heuristics to decide whether write memory barrier should be performed. For bursts with size multiple of recommended one (64 pkts) it is supposed the next burst is coming and no need to issue the extra memory barrier (it is supposed to be issued in the next coming burst, at least after descriptor writing). It might increase latency (on some hosts till next packets transmit) and should be used with care.
If tx_db_nc is omitted or set to zero, the preset (if any) environment variable “MLX5_SHUT_UP_BF” value is used. If there is no “MLX5_SHUT_UP_BF”, the default tx_db_nc value is zero for ARM64 hosts and one for others.
tx_vec_en parameter [int]
A nonzero value enables Tx vector on ConnectX-5, ConnectX-6, ConnectX-6 Dx and BlueField NICs if the number of global Tx queues on the port is less than txqs_max_vec. The parameter is deprecated and ignored.
rx_vec_en parameter [int]
A nonzero value enables Rx vector if the port is not configured in multi-segment otherwise this parameter is ignored.
Enabled by default.
vf_nl_en parameter [int]
A nonzero value enables Netlink requests from the VF to add/remove MAC addresses or/and enable/disable promiscuous/all multicast on the Netdevice. Otherwise the relevant configuration must be run with Linux iproute2 tools. This is a prerequisite to receive this kind of traffic.
Enabled by default, valid only on VF devices ignored otherwise.
l3_vxlan_en parameter [int]
A nonzero value allows L3 VXLAN and VXLAN-GPE flow creation. To enable L3 VXLAN or VXLAN-GPE, users has to configure firmware and enable this parameter. This is a prerequisite to receive this kind of traffic.
Disabled by default.
dv_xmeta_en parameter [int]
A nonzero value enables extensive flow metadata support if device is capable and driver supports it. This can enable extensive support of MARK and META item of rte_flow. The newly introduced SET_TAG and SET_META actions do not depend on dv_xmeta_en.
There are some possible configurations, depending on parameter value:
Mode | MARK | META | META Tx | FDB/Through |
---|---|---|---|---|
0 | 24 bits | 32 bits | 32 bits | no |
1 | 24 bits | vary 0-32 | 32 bits | yes |
2 | vary 0-32 | 32 bits | 32 bits | yes |
If there is no E-Switch configuration the dv_xmeta_en parameter is ignored and the device is configured to operate in legacy mode (0).
Disabled by default (set to 0).
The Direct Verbs/Rules (engaged with dv_flow_en = 1) supports all of the extensive metadata features. The legacy Verbs supports FLAG and MARK metadata actions over NIC Rx steering domain only.
dv_flow_en parameter [int]
A nonzero value enables the DV flow steering assuming it is supported by the driver (RDMA Core library version is rdma-core-24.0 or higher).
Enabled by default if supported.
dv_esw_en parameter [int]
A nonzero value enables E-Switch using Direct Rules.
Enabled by default if supported.
lacp_by_user parameter [int]
A nonzero value enables the control of LACP traffic by the user application. When a bond exists in the driver, by default it should be managed by the kernel and therefore LACP traffic should be steered to the kernel. If this devarg is set to 1 it will allow the user to manage the bond by itself and not steer LACP traffic to the kernel.
Disabled by default (set to 0).
mr_ext_memseg_en parameter [int]
A nonzero value enables extending memseg when registering DMA memory. If enabled, the number of entries in MR (Memory Region) lookup table on datapath is minimized and it benefits performance. On the other hand, it worsens memory utilization because registered memory is pinned by kernel driver. Even if a page in the extended chunk is freed, that doesn’t become reusable until the entire memory is freed.
Enabled by default.
representor parameter [list]
This parameter can be used to instantiate DPDK Ethernet devices from existing port (or VF) representors configured on the device.
It is a standard parameter whose format is described in Ethernet Device Standard Device Arguments.
For instance, to probe port representors 0 through 2:
representor=[0-2]
max_dump_files_num parameter [int]
The maximum number of files per PMD entity that may be created for debug information. The files will be created in /var/log directory or in current directory.
set to 128 by default.
lro_timeout_usec parameter [int]
The maximum allowed duration of an LRO session, in micro-seconds. PMD will set the nearest value supported by HW, which is not bigger than the input lro_timeout_usec value. If this parameter is not specified, by default PMD will set the smallest value supported by HW.
isolated_mode parameter [int]
A nonzero value configures the port to initialize in Flow isolated mode.
Set to zero by default.
hp_buf_log_sz parameter [int]
The total data buffer size of a hairpin queue (logarithmic form), in bytes. PMD will set the data buffer size to 2 ** hp_buf_log_sz, both for RX & TX. The capacity of the value is specified by the firmware and the initialization will get a failure if it is out of scope. The range of the value is from 11 to 19 right now, and the supported frame size of a single packet for hairpin is from 512B to 128KB. It might change if different firmware release is being used. By using a small value, it could reduce memory consumption but not work with a large frame. If the value is too large, the memory consumption will be high and some potential performance degradation will be introduced. By default, the PMD will set this value to 16, which means that 9KB jumbo frames will be supported.
reclaim_mem_mode parameter [int]
Cache some resources in flow destroy will help flow recreation more efficient. While some systems may require the all the resources can be reclaimed after flow destroyed. The parameter reclaim_mem_mode provides the option for user to configure if the resource cache is needed or not.
There are three options to choose:
By default, the PMD will set this value to 0.
sys_mem_en parameter [int]
A nonzero value enables the PMD memory management function allocate memory from system by default without explicitly rte memory flag.
By default, the PMD will set this value to 0.
decap_en parameter [int]
The new devarg decap_en is introduced to provide the choice to the NICs which does not support FCS (frame checksum) scattering for the tunnel decapsulated packets. By set the devarg value 0 to force the FCS feature enable and rejecting tunnel decap actions in the rte_flow engine for these special NICs.
By default, the PMD will set this value to 1.
Firmware features can be configured as key/value pairs.
The command to set a value is:
mlxconfig -d <device> set <key>=<value>
The command to query a value is:
mlxconfig -d <device> query | grep <key>
The device name for the command mlxconfig can be either the PCI address, or the mst device name found with:
mst status
Below are some firmware configurations listed.
link type:
LINK_TYPE_P1
LINK_TYPE_P2
value: 1=Infiniband 2=Ethernet 3=VPI(auto-sense)
enable SR-IOV:
SRIOV_EN=1
maximum number of SR-IOV virtual functions:
NUM_OF_VFS=<max>
enable DevX (required by Direct Rules and other features):
UCTX_EN=1
aggressive CQE zipping:
CQE_COMPRESSION=1
L3 VXLAN and VXLAN-GPE destination UDP port:
IP_OVER_VXLAN_EN=1
IP_OVER_VXLAN_PORT=<udp dport>
enable VXLAN-GPE tunnel flow matching:
FLEX_PARSER_PROFILE_ENABLE=0
or
FLEX_PARSER_PROFILE_ENABLE=2
enable IP-in-IP tunnel flow matching:
FLEX_PARSER_PROFILE_ENABLE=0
enable MPLS flow matching:
FLEX_PARSER_PROFILE_ENABLE=1
enable ICMP/ICMP6 code/type fields matching:
FLEX_PARSER_PROFILE_ENABLE=2
enable Geneve flow matching:
FLEX_PARSER_PROFILE_ENABLE=0
or
FLEX_PARSER_PROFILE_ENABLE=1
enable GTP flow matching:
FLEX_PARSER_PROFILE_ENABLE=3
This driver relies on external libraries and kernel drivers for resources allocations and initialization. The following dependencies are not part of DPDK and must be installed separately:
libibverbs
User space Verbs framework used by librte_pmd_mlx5. This library provides a generic interface between the kernel and low-level user space drivers such as libmlx5.
It allows slow and privileged operations (context initialization, hardware resources allocations) to be managed by the kernel and fast operations to never leave user space.
libmlx5
Low-level user space driver library for Mellanox ConnectX-4/ConnectX-5/ConnectX-6/BlueField devices, it is automatically loaded by libibverbs.
This library basically implements send/receive calls to the hardware queues.
Kernel modules
They provide the kernel-side Verbs API and low level device drivers that manage actual hardware initialization and resources sharing with user space processes.
Unlike most other PMDs, these modules must remain loaded and bound to their devices:
Firmware update
Mellanox OFED/EN releases include firmware updates for ConnectX-4/ConnectX-5/ConnectX-6/BlueField adapters.
Because each release provides new features, these updates must be applied to match the kernel modules and libraries they come with.
Note
Both libraries are BSD and GPL licensed. Linux kernel modules are GPL licensed.
Either RDMA Core library with a recent enough Linux kernel release (recommended) or Mellanox OFED/EN, which provides compatibility with older releases.
Minimal kernel version : v4.14 or the most recent 4.14-rc (see Linux installation documentation)
Minimal rdma-core version: v15+ commit 0c5f5765213a (“Merge pull request #227 from yishaih/tm”) (see RDMA Core installation documentation)
When building for i686 use:
Starting with rdma-core v21, static libraries can be built:
cd build
CFLAGS=-fPIC cmake -DIN_PLACE=1 -DENABLE_STATIC=1 -GNinja ..
ninja
If rdma-core libraries are built but not installed, DPDK makefile can link them, thanks to these environment variables:
- EXTRA_CFLAGS=-I/path/to/rdma-core/build/include
- EXTRA_LDFLAGS=-L/path/to/rdma-core/build/lib
- PKG_CONFIG_PATH=/path/to/rdma-core/build/lib/pkgconfig
While these libraries and kernel modules are available on OpenFabrics Alliance’s website and provided by package managers on most distributions, this PMD requires Ethernet extensions that may not be supported at the moment (this is a work in progress).
Mellanox OFED and Mellanox EN include the necessary support and should be used in the meantime. For DPDK, only libibverbs, libmlx5, mlnx-ofed-kernel packages and firmware updates are required from that distribution.
Note
Several versions of Mellanox OFED/EN are available. Installing the version this DPDK release was developed and tested against is strongly recommended. Please check the prerequisites.
The following Mellanox device families are supported by the same mlx5 driver:
- ConnectX-4
- ConnectX-4 Lx
- ConnectX-5
- ConnectX-5 Ex
- ConnectX-6
- ConnectX-6 Dx
- BlueField
Below are detailed device names:
Download latest Mellanox OFED/EN. For more info check the prerequisites.
Install the required libraries and kernel modules either by installing only the required set, or by installing the entire Mellanox OFED/EN:
./mlnxofedinstall --upstream-libs --dpdk
Verify the firmware is the correct one:
ibv_devinfo
Verify all ports links are set to Ethernet:
mlxconfig -d <mst device> query | grep LINK_TYPE
LINK_TYPE_P1 ETH(2)
LINK_TYPE_P2 ETH(2)
Link types may have to be configured to Ethernet:
mlxconfig -d <mst device> set LINK_TYPE_P1/2=1/2/3
* LINK_TYPE_P1=<1|2|3> , 1=Infiniband 2=Ethernet 3=VPI(auto-sense)
For hypervisors, verify SR-IOV is enabled on the NIC:
mlxconfig -d <mst device> query | grep SRIOV_EN
SRIOV_EN True(1)
If needed, configure SR-IOV:
mlxconfig -d <mst device> set SRIOV_EN=1 NUM_OF_VFS=16
mlxfwreset -d <mst device> reset
Restart the driver:
/etc/init.d/openibd restart
or:
service openibd restart
If link type was changed, firmware must be reset as well:
mlxfwreset -d <mst device> reset
For hypervisors, after reset write the sysfs number of virtual functions needed for the PF.
To dynamically instantiate a given number of virtual functions (VFs):
echo [num_vfs] > /sys/class/infiniband/mlx5_0/device/sriov_numvfs
Compile DPDK and you are ready to go. See instructions on Development Kit Build System
Switchdev mode is a mode in E-Switch, that binds between representor and VF. Representor is a port in DPDK that is connected to a VF in such a way that assuming there are no offload flows, each packet that is sent from the VF will be received by the corresponding representor. While each packet that is sent to a representor will be received by the VF. This is very useful in case of SRIOV mode, where the first packet that is sent by the VF will be received by the DPDK application which will decide if this flow should be offloaded to the E-Switch. After offloading the flow packet that the VF that are matching the flow will not be received any more by the DPDK application.
Enable SRIOV mode:
mlxconfig -d <mst device> set SRIOV_EN=true
Configure the max number of VFs:
mlxconfig -d <mst device> set NUM_OF_VFS=<num of vfs>
Reset the FW:
mlxfwreset -d <mst device> reset
Configure the actual number of VFs:
echo <num of vfs > /sys/class/net/<net device>/device/sriov_numvfs
Unbind the device (can be rebind after the switchdev mode):
echo -n "<device pci address" > /sys/bus/pci/drivers/mlx5_core/unbind
Enbale switchdev mode:
echo switchdev > /sys/class/net/<net device>/compat/devlink/mode
Configure aggressive CQE Zipping for maximum performance:
mlxconfig -d <mst device> s CQE_COMPRESSION=1
To set it back to the default CQE Zipping mode use:
mlxconfig -d <mst device> s CQE_COMPRESSION=0
In case of virtualization:
Use the CPU near local NUMA node to which the PCIe adapter is connected, for better performance. For VMs, verify that the right CPU and NUMA node are pinned according to the above. Run:
lstopo-no-graphics
to identify the NUMA node to which the PCIe adapter is connected.
If more than one adapter is used, and root complex capabilities allow to put both adapters on the same NUMA node without PCI bandwidth degradation, it is recommended to locate both adapters on the same NUMA node. This in order to forward packets from one to the other without NUMA performance penalty.
Disable pause frames:
ethtool -A <netdev> rx off tx off
Verify IO non-posted prefetch is disabled by default. This can be checked via the BIOS configuration. Please contact you server provider for more information about the settings.
Note
On some machines, depends on the machine integrator, it is beneficial to set the PCI max read request parameter to 1K. This can be done in the following way:
To query the read request size use:
setpci -s <NIC PCI address> 68.w
If the output is different than 3XXX, set it by:
setpci -s <NIC PCI address> 68.w=3XXX
The XXX can be different on different systems. Make sure to configure according to the setpci output.
Offload | DPDK | Linux | rdma-core | OFED | firmware | hardware |
---|---|---|---|---|---|---|
common base | 17.11 | 4.14 | 16 | 4.2-1 | 12.21.1000 | ConnectX-4 |
checksums | 17.11 | 4.14 | 16 | 4.2-1 | 12.21.1000 | ConnectX-4 |
Rx timestamp | 17.11 | 4.14 | 16 | 4.2-1 | 12.21.1000 | ConnectX-4 |
TSO | 17.11 | 4.14 | 16 | 4.2-1 | 12.21.1000 | ConnectX-4 |
LRO | 19.08 | N/A | N/A | 4.6-4 | 16.25.6406 | ConnectX-5 |
Offload | with E-Switch | with NIC |
---|---|---|
Count | DPDK 19.05
OFED 4.6
rdma-core 24
ConnectX-5
|
DPDK 19.02
OFED 4.6
rdma-core 23
ConnectX-5
|
Drop | DPDK 19.05
OFED 4.6
rdma-core 24
ConnectX-5
|
DPDK 18.11
OFED 4.5
rdma-core 23
ConnectX-4
|
Queue / RSS | N/A
|
DPDK 18.11
OFED 4.5
rdma-core 23
ConnectX-4
|
Encapsulation (VXLAN / NVGRE / RAW) | DPDK 19.05
OFED 4.7-1
rdma-core 24
ConnectX-5
|
DPDK 19.02
OFED 4.6
rdma-core 23
ConnectX-5
|
Encapsulation GENEVE | DPDK 19.11
OFED 4.7-3
rdma-core 27
ConnectX-5
|
DPDK 19.11
OFED 4.7-3
rdma-core 27
ConnectX-5
|
Header rewrite
(set_ipv4_src /
set_ipv4_dst /
set_ipv6_src /
set_ipv6_dst /
set_tp_src /
set_tp_dst /
dec_ttl /
set_ttl /
set_mac_src /
set_mac_dst)
|
DPDK 19.05
OFED 4.7-1
rdma-core 24
ConnectX-5
|
DPDK 19.02
OFED 4.7-1
rdma-core 24
ConnectX-5
|
Header rewrite
(set_dscp)
|
DPDK 20.02
OFED 5.0
rdma-core 24
ConnectX-5
|
DPDK 20.02
OFED 5.0
rdma-core 24
ConnectX-5
|
Jump | DPDK 19.05
OFED 4.7-1
rdma-core 24
ConnectX-5
|
DPDK 19.02
OFED 4.7-1
N/A
ConnectX-5
|
Mark / Flag | DPDK 19.05
OFED 4.6
rdma-core 24
ConnectX-5
|
DPDK 18.11
OFED 4.5
rdma-core 23
ConnectX-4
|
Port ID | DPDK 19.05
OFED 4.7-1
rdma-core 24
ConnectX-5
|
N/A
N/A
N/A
N/A
|
VLAN
(of_pop_vlan /
of_push_vlan /
of_set_vlan_pcp /
of_set_vlan_vid)
|
DPDK 19.11
OFED 4.7-1
ConnectX-5
|
DPDK 19.11
OFED 4.7-1
ConnectX-5
|
Hairpin | N/A
|
DPDK 19.11
OFED 4.7-3
rdma-core 26
ConnectX-5
|
Meta data | DPDK 19.11
OFED 4.7-3
rdma-core 26
ConnectX-5
|
DPDK 19.11
OFED 4.7-3
rdma-core 26
ConnectX-5
|
Metering | DPDK 19.11
OFED 4.7-3
rdma-core 26
ConnectX-5
|
DPDK 19.11
OFED 4.7-3
rdma-core 26
ConnectX-5
|
Sampling | DPDK 20.11
OFED 5.2
rdma-core 32
ConnectX-5
|
DPDK 20.11
OFED 5.2
rdma-core 32
ConnectX-5
|
MARK and META items are interrelated with datapath - they might move from/to the applications in mbuf fields. Hence, zero value for these items has the special meaning - it means “no metadata are provided”, not zero values are treated by applications and PMD as valid ones.
Moreover in the flow engine domain the value zero is acceptable to match and set, and we should allow to specify zero values as rte_flow parameters for the META and MARK items and actions. In the same time zero mask has no meaning and should be rejected on validation stage.
Flows are not cached in the driver. When stopping a device port, all the flows created on this port from the application will be flushed automatically in the background. After stopping the device port, all flows on this port become invalid and not represented in the system. All references to these flows held by the application should be discarded directly but neither destroyed nor flushed.
The application should re-create the flows as required after the port restart.
Compared to librte_pmd_mlx4 that implements a single RSS configuration per port, librte_pmd_mlx5 supports per-protocol RSS configuration.
Since testpmd defaults to IP RSS mode and there is currently no command-line parameter to enable additional protocols (UDP and TCP as well as IP), the following commands must be entered from its CLI to get the same behavior as librte_pmd_mlx4:
> port stop all
> port config all rss all
> port start all
This section demonstrates how to launch testpmd with Mellanox ConnectX-4/ConnectX-5/ConnectX-6/BlueField devices managed by librte_pmd_mlx5.
Load the kernel modules:
modprobe -a ib_uverbs mlx5_core mlx5_ib
Alternatively if MLNX_OFED/MLNX_EN is fully installed, the following script can be run:
/etc/init.d/openibd restart
Note
User space I/O kernel modules (uio and igb_uio) are not used and do not have to be loaded.
Make sure Ethernet interfaces are in working order and linked to kernel verbs. Related sysfs entries should be present:
ls -d /sys/class/net/*/device/infiniband_verbs/uverbs* | cut -d / -f 5
Example output:
eth30
eth31
eth32
eth33
Optionally, retrieve their PCI bus addresses for whitelisting:
{
for intf in eth2 eth3 eth4 eth5;
do
(cd "/sys/class/net/${intf}/device/" && pwd -P);
done;
} |
sed -n 's,.*/\(.*\),-w \1,p'
Example output:
-w 0000:05:00.1
-w 0000:06:00.0
-w 0000:06:00.1
-w 0000:05:00.0
Request huge pages:
echo 1024 > /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages/nr_hugepages
Start testpmd with basic parameters:
testpmd -l 8-15 -n 4 -w 05:00.0 -w 05:00.1 -w 06:00.0 -w 06:00.1 -- --rxq=2 --txq=2 -i
Example output:
[...]
EAL: PCI device 0000:05:00.0 on NUMA socket 0
EAL: probe driver: 15b3:1013 librte_pmd_mlx5
PMD: librte_pmd_mlx5: PCI information matches, using device "mlx5_0" (VF: false)
PMD: librte_pmd_mlx5: 1 port(s) detected
PMD: librte_pmd_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:fe
EAL: PCI device 0000:05:00.1 on NUMA socket 0
EAL: probe driver: 15b3:1013 librte_pmd_mlx5
PMD: librte_pmd_mlx5: PCI information matches, using device "mlx5_1" (VF: false)
PMD: librte_pmd_mlx5: 1 port(s) detected
PMD: librte_pmd_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:ff
EAL: PCI device 0000:06:00.0 on NUMA socket 0
EAL: probe driver: 15b3:1013 librte_pmd_mlx5
PMD: librte_pmd_mlx5: PCI information matches, using device "mlx5_2" (VF: false)
PMD: librte_pmd_mlx5: 1 port(s) detected
PMD: librte_pmd_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:fa
EAL: PCI device 0000:06:00.1 on NUMA socket 0
EAL: probe driver: 15b3:1013 librte_pmd_mlx5
PMD: librte_pmd_mlx5: PCI information matches, using device "mlx5_3" (VF: false)
PMD: librte_pmd_mlx5: 1 port(s) detected
PMD: librte_pmd_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:fb
Interactive-mode selected
Configuring Port 0 (socket 0)
PMD: librte_pmd_mlx5: 0x8cba80: TX queues number update: 0 -> 2
PMD: librte_pmd_mlx5: 0x8cba80: RX queues number update: 0 -> 2
Port 0: E4:1D:2D:E7:0C:FE
Configuring Port 1 (socket 0)
PMD: librte_pmd_mlx5: 0x8ccac8: TX queues number update: 0 -> 2
PMD: librte_pmd_mlx5: 0x8ccac8: RX queues number update: 0 -> 2
Port 1: E4:1D:2D:E7:0C:FF
Configuring Port 2 (socket 0)
PMD: librte_pmd_mlx5: 0x8cdb10: TX queues number update: 0 -> 2
PMD: librte_pmd_mlx5: 0x8cdb10: RX queues number update: 0 -> 2
Port 2: E4:1D:2D:E7:0C:FA
Configuring Port 3 (socket 0)
PMD: librte_pmd_mlx5: 0x8ceb58: TX queues number update: 0 -> 2
PMD: librte_pmd_mlx5: 0x8ceb58: RX queues number update: 0 -> 2
Port 3: E4:1D:2D:E7:0C:FB
Checking link statuses...
Port 0 Link Up - speed 40000 Mbps - full-duplex
Port 1 Link Up - speed 40000 Mbps - full-duplex
Port 2 Link Up - speed 10000 Mbps - full-duplex
Port 3 Link Up - speed 10000 Mbps - full-duplex
Done
testpmd>
This section demonstrates how to dump flows. Currently, it’s possible to dump all flows with assistance of external tools.
2 ways to get flow raw file:
testpmd> flow dump <port> <output_file>
rte_flow_dev_dump(port, file, NULL);
Dump human-readable flows from raw file:
Get flow parsing tool from: https://github.com/Mellanox/mlx_steering_dump
mlx_steering_dump.py -f <output_file>