1 .. SPDX-License-Identifier: BSD-3-Clause
2 Copyright 2015 6WIND S.A.
3 Copyright 2015 Mellanox Technologies, Ltd
5 .. include:: <isonum.txt>
10 The MLX5 poll mode driver library (**librte_pmd_mlx5**) provides support
11 for **Mellanox ConnectX-4**, **Mellanox ConnectX-4 Lx** , **Mellanox
12 ConnectX-5**, **Mellanox ConnectX-6**, **Mellanox ConnectX-6 Dx** and
13 **Mellanox BlueField** families of 10/25/40/50/100/200 Gb/s adapters
14 as well as their virtual functions (VF) in SR-IOV context.
16 Information and documentation about these adapters can be found on the
17 `Mellanox website <http://www.mellanox.com>`__. Help is also provided by the
18 `Mellanox community <http://community.mellanox.com/welcome>`__.
20 There is also a `section dedicated to this poll mode driver
21 <http://www.mellanox.com/page/products_dyn?product_family=209&mtag=pmd_for_dpdk>`__.
25 Due to external dependencies, this driver is disabled in default configuration
26 of the "make" build. It can be enabled with ``CONFIG_RTE_LIBRTE_MLX5_PMD=y``
27 or by using "meson" build system which will detect dependencies.
32 Besides its dependency on libibverbs (that implies libmlx5 and associated
33 kernel support), librte_pmd_mlx5 relies heavily on system calls for control
34 operations such as querying/updating the MTU and flow control parameters.
36 For security reasons and robustness, this driver only deals with virtual
37 memory addresses. The way resources allocations are handled by the kernel,
38 combined with hardware specifications that allow to handle virtual memory
39 addresses directly, ensure that DPDK applications cannot access random
40 physical memory (or memory that does not belong to the current process).
42 This capability allows the PMD to coexist with kernel network interfaces
43 which remain functional, although they stop receiving unicast packets as
44 long as they share the same MAC address.
45 This means legacy linux control tools (for example: ethtool, ifconfig and
46 more) can operate on the same network interfaces that owned by the DPDK
49 The PMD can use libibverbs and libmlx5 to access the device firmware
50 or directly the hardware components.
51 There are different levels of objects and bypassing abilities
52 to get the best performances:
54 - Verbs is a complete high-level generic API
55 - Direct Verbs is a device-specific API
56 - DevX allows to access firmware objects
57 - Direct Rules manages flow steering at low-level hardware layer
59 Enabling librte_pmd_mlx5 causes DPDK applications to be linked against
65 - Multi arch support: x86_64, POWER8, ARMv8, i686.
66 - Multiple TX and RX queues.
67 - Support for scattered TX and RX frames.
68 - IPv4, IPv6, TCPv4, TCPv6, UDPv4 and UDPv6 RSS on any number of queues.
69 - RSS using different combinations of fields: L3 only, L4 only or both,
70 and source only, destination only or both.
71 - Several RSS hash keys, one for each flow type.
72 - Default RSS operation with no hash key specification.
73 - Configurable RETA table.
74 - Link flow control (pause frame).
75 - Support for multiple MAC addresses.
79 - RX CRC stripping configuration.
80 - Promiscuous mode on PF and VF.
81 - Multicast promiscuous mode on PF and VF.
82 - Hardware checksum offloads.
83 - Flow director (RTE_FDIR_MODE_PERFECT, RTE_FDIR_MODE_PERFECT_MAC_VLAN and
85 - Flow API, including :ref:`flow_isolated_mode`.
87 - KVM and VMware ESX SR-IOV modes are supported.
88 - RSS hash result is supported.
89 - Hardware TSO for generic IP or UDP tunnel, including VXLAN and GRE.
90 - Hardware checksum Tx offload for generic IP or UDP tunnel, including VXLAN and GRE.
92 - Statistics query including Basic, Extended and per queue.
94 - Tunnel types: VXLAN, L3 VXLAN, VXLAN-GPE, GRE, MPLSoGRE, MPLSoUDP, IP-in-IP, Geneve, GTP.
95 - Tunnel HW offloads: packet type, inner/outer RSS, IP and UDP checksum verification.
96 - NIC HW offloads: encapsulation (vxlan, gre, mplsoudp, mplsogre), NAT, routing, TTL
97 increment/decrement, count, drop, mark. For details please see :ref:`mlx5_offloads_support`.
98 - Flow insertion rate of more then million flows per second, when using Direct Rules.
99 - Support for multiple rte_flow groups.
100 - Per packet no-inline hint flag to disable packet data copying into Tx descriptors.
107 - For secondary process:
109 - Forked secondary process not supported.
110 - External memory unregistered in EAL memseg list cannot be used for DMA
111 unless such memory has been registered by ``mlx5_mr_update_ext_mp()`` in
112 primary process and remapped to the same virtual address in secondary
113 process. If the external memory is registered by primary process but has
114 different virtual address in secondary process, unexpected error may happen.
116 - When using Verbs flow engine (``dv_flow_en`` = 0), flow pattern without any
117 specific VLAN will match for VLAN packets as well:
119 When VLAN spec is not specified in the pattern, the matching rule will be created with VLAN as a wild card.
120 Meaning, the flow rule::
122 flow create 0 ingress pattern eth / vlan vid is 3 / ipv4 / end ...
124 Will only match vlan packets with vid=3. and the flow rule::
126 flow create 0 ingress pattern eth / ipv4 / end ...
128 Will match any ipv4 packet (VLAN included).
130 - When using DV flow engine (``dv_flow_en`` = 1), flow pattern without VLAN item
131 will match untagged packets only.
134 flow create 0 ingress pattern eth / ipv4 / end ...
136 Will match untagged packets only.
139 flow create 0 ingress pattern eth / vlan / ipv4 / end ...
141 Will match tagged packets only, with any VLAN ID value.
144 flow create 0 ingress pattern eth / vlan vid is 3 / ipv4 / end ...
146 Will only match tagged packets with VLAN ID 3.
148 - VLAN pop offload command:
150 - Flow rules having a VLAN pop offload command as one of their actions and
151 are lacking a match on VLAN as one of their items are not supported.
152 - The command is not supported on egress traffic.
154 - VLAN push offload is not supported on ingress traffic.
156 - VLAN set PCP offload is not supported on existing headers.
158 - A multi segment packet must have not more segments than reported by dev_infos_get()
159 in tx_desc_lim.nb_seg_max field. This value depends on maximal supported Tx descriptor
160 size and ``txq_inline_min`` settings and may be from 2 (worst case forced by maximal
161 inline settings) to 58.
163 - Flows with a VXLAN Network Identifier equal (or ends to be equal)
164 to 0 are not supported.
166 - L3 VXLAN and VXLAN-GPE tunnels cannot be supported together with MPLSoGRE and MPLSoUDP.
168 - Match on Geneve header supports the following fields only:
174 Currently, the only supported options length value is 0.
176 - VF: flow rules created on VF devices can only match traffic targeted at the
177 configured MAC addresses (see ``rte_eth_dev_mac_addr_add()``).
179 - Match on GTP tunnel header item supports the following fields only:
181 - v_pt_rsv_flags: E flag, S flag, PN flag
185 - No Tx metadata go to the E-Switch steering domain for the Flow group 0.
186 The flows within group 0 and set metadata action are rejected by hardware.
190 MAC addresses not already present in the bridge table of the associated
191 kernel network device will be added and cleaned up by the PMD when closing
192 the device. In case of ungraceful program termination, some entries may
193 remain present and should be removed manually by other means.
195 - When Multi-Packet Rx queue is configured (``mprq_en``), a Rx packet can be
196 externally attached to a user-provided mbuf with having EXT_ATTACHED_MBUF in
197 ol_flags. As the mempool for the external buffer is managed by PMD, all the
198 Rx mbufs must be freed before the device is closed. Otherwise, the mempool of
199 the external buffers will be freed by PMD and the application which still
200 holds the external buffers may be corrupted.
202 - If Multi-Packet Rx queue is configured (``mprq_en``) and Rx CQE compression is
203 enabled (``rxq_cqe_comp_en``) at the same time, RSS hash result is not fully
204 supported. Some Rx packets may not have PKT_RX_RSS_HASH.
206 - IPv6 Multicast messages are not supported on VM, while promiscuous mode
207 and allmulticast mode are both set to off.
208 To receive IPv6 Multicast messages on VM, explicitly set the relevant
209 MAC address using rte_eth_dev_mac_addr_add() API.
211 - To support a mixed traffic pattern (some buffers from local host memory, some
212 buffers from other devices) with high bandwidth, a mbuf flag is used.
214 An application hints the PMD whether or not it should try to inline the
215 given mbuf data buffer. PMD should do the best effort to act upon this request.
217 The hint flag ``RTE_PMD_MLX5_FINE_GRANULARITY_INLINE`` is dynamic,
218 registered by application with rte_mbuf_dynflag_register(). This flag is
219 purely driver-specific and declared in PMD specific header ``rte_pmd_mlx5.h``,
220 which is intended to be used by the application.
222 To query the supported specific flags in runtime,
223 the function ``rte_pmd_mlx5_get_dyn_flag_names`` returns the array of
224 currently (over present hardware and configuration) supported specific flags.
225 The "not inline hint" feature operating flow is the following one:
228 - probe the devices, ports are created
229 - query the port capabilities
230 - if port supporting the feature is found
231 - register dynamic flag ``RTE_PMD_MLX5_FINE_GRANULARITY_INLINE``
232 - application starts the ports
233 - on ``dev_start()`` PMD checks whether the feature flag is registered and
234 enables the feature support in datapath
235 - application might set the registered flag bit in ``ol_flags`` field
236 of mbuf being sent and PMD will handle ones appropriately.
238 - The amount of descriptors in Tx queue may be limited by data inline settings.
239 Inline data require the more descriptor building blocks and overall block
240 amount may exceed the hardware supported limits. The application should
241 reduce the requested Tx size or adjust data inline settings with
242 ``txq_inline_max`` and ``txq_inline_mpw`` devargs keys.
244 - E-Switch decapsulation Flow:
246 - can be applied to PF port only.
247 - must specify VF port action (packet redirection from PF to VF).
248 - optionally may specify tunnel inner source and destination MAC addresses.
250 - E-Switch encapsulation Flow:
252 - can be applied to VF ports only.
253 - must specify PF port action (packet redirection from VF to PF).
257 - The input buffer, used as outer header, is not validated.
261 - The decapsulation is always done up to the outermost tunnel detected by the HW.
262 - The input buffer, providing the removal size, is not validated.
263 - The buffer size must match the length of the headers to be removed.
265 - ICMP/ICMP6 code/type matching, IP-in-IP and MPLS flow matching are all
266 mutually exclusive features which cannot be supported together
267 (see :ref:`mlx5_firmware_config`).
271 - Requires DevX and DV flow to be enabled.
272 - KEEP_CRC offload cannot be supported with LRO.
273 - The first mbuf length, without head-room, must be big enough to include the
275 - Rx queue with LRO offload enabled, receiving a non-LRO packet, can forward
276 it with size limited to max LRO size, not to max RX packet length.
277 - LRO can be used with outer header of TCP packets of the standard format:
278 eth (with or without vlan) / ipv4 or ipv6 / tcp / payload
280 Other TCP packets (e.g. with MPLS label) received on Rx queue with LRO enabled, will be received with bad checksum.
285 MLX5 supports various methods to report statistics:
287 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.
289 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 :ref:`Bifurcated Linux Driver <linux_gsg_linux_drivers>` 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.
291 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.
299 These options can be modified in the ``.config`` file.
301 - ``CONFIG_RTE_LIBRTE_MLX5_PMD`` (default **n**)
303 Toggle compilation of librte_pmd_mlx5 itself.
305 - ``CONFIG_RTE_IBVERBS_LINK_DLOPEN`` (default **n**)
307 Build PMD with additional code to make it loadable without hard
308 dependencies on **libibverbs** nor **libmlx5**, which may not be installed
309 on the target system.
311 In this mode, their presence is still required for it to run properly,
312 however their absence won't prevent a DPDK application from starting (with
313 ``CONFIG_RTE_BUILD_SHARED_LIB`` disabled) and they won't show up as
314 missing with ``ldd(1)``.
316 It works by moving these dependencies to a purpose-built rdma-core "glue"
317 plug-in which must either be installed in a directory whose name is based
318 on ``CONFIG_RTE_EAL_PMD_PATH`` suffixed with ``-glue`` if set, or in a
319 standard location for the dynamic linker (e.g. ``/lib``) if left to the
320 default empty string (``""``).
322 This option has no performance impact.
324 - ``CONFIG_RTE_IBVERBS_LINK_STATIC`` (default **n**)
326 Embed static flavor of the dependencies **libibverbs** and **libmlx5**
327 in the PMD shared library or the executable static binary.
329 - ``CONFIG_RTE_LIBRTE_MLX5_DEBUG`` (default **n**)
331 Toggle debugging code and stricter compilation flags. Enabling this option
332 adds additional run-time checks and debugging messages at the cost of
337 For BlueField, target should be set to ``arm64-bluefield-linux-gcc``. This
338 will enable ``CONFIG_RTE_LIBRTE_MLX5_PMD`` and set ``RTE_CACHE_LINE_SIZE`` to
339 64. Default armv8a configuration of make build and meson build set it to 128
340 then brings performance degradation.
342 This option is available in meson:
344 - ``ibverbs_link`` can be ``static``, ``shared``, or ``dlopen``.
346 Environment variables
347 ~~~~~~~~~~~~~~~~~~~~~
351 A list of directories in which to search for the rdma-core "glue" plug-in,
352 separated by colons or semi-colons.
354 Only matters when compiled with ``CONFIG_RTE_IBVERBS_LINK_DLOPEN``
355 enabled and most useful when ``CONFIG_RTE_EAL_PMD_PATH`` is also set,
356 since ``LD_LIBRARY_PATH`` has no effect in this case.
358 - ``MLX5_SHUT_UP_BF``
360 Configures HW Tx doorbell register as IO-mapped.
362 By default, the HW Tx doorbell is configured as a write-combining register.
363 The register would be flushed to HW usually when the write-combining buffer
364 becomes full, but it depends on CPU design.
366 Except for vectorized Tx burst routines, a write memory barrier is enforced
367 after updating the register so that the update can be immediately visible to
370 When vectorized Tx burst is called, the barrier is set only if the burst size
371 is not aligned to MLX5_VPMD_TX_MAX_BURST. However, setting this environmental
372 variable will bring better latency even though the maximum throughput can
375 Run-time configuration
376 ~~~~~~~~~~~~~~~~~~~~~~
378 - librte_pmd_mlx5 brings kernel network interfaces up during initialization
379 because it is affected by their state. Forcing them down prevents packets
382 - **ethtool** operations on related kernel interfaces also affect the PMD.
387 In order to run as a non-root user,
388 some capabilities must be granted to the application::
390 setcap cap_sys_admin,cap_net_admin,cap_net_raw,cap_ipc_lock+ep <dpdk-app>
392 Below are the reasons of the need for each capability:
395 When using physical addresses (PA mode), with Linux >= 4.0,
396 for access to ``/proc/self/pagemap``.
399 For device configuration.
402 For raw ethernet queue allocation through kernel driver.
405 For DMA memory pinning.
410 - ``rxq_cqe_comp_en`` parameter [int]
412 A nonzero value enables the compression of CQE on RX side. This feature
413 allows to save PCI bandwidth and improve performance. Enabled by default.
417 - x86_64 with ConnectX-4, ConnectX-4 Lx, ConnectX-5, ConnectX-6, ConnectX-6 Dx
419 - POWER9 and ARMv8 with ConnectX-4 Lx, ConnectX-5, ConnectX-6, ConnectX-6 Dx
422 - ``rxq_cqe_pad_en`` parameter [int]
424 A nonzero value enables 128B padding of CQE on RX side. The size of CQE
425 is aligned with the size of a cacheline of the core. If cacheline size is
426 128B, the CQE size is configured to be 128B even though the device writes
427 only 64B data on the cacheline. This is to avoid unnecessary cache
428 invalidation by device's two consecutive writes on to one cacheline.
429 However in some architecture, it is more beneficial to update entire
430 cacheline with padding the rest 64B rather than striding because
431 read-modify-write could drop performance a lot. On the other hand,
432 writing extra data will consume more PCIe bandwidth and could also drop
433 the maximum throughput. It is recommended to empirically set this
434 parameter. Disabled by default.
438 - CPU having 128B cacheline with ConnectX-5 and BlueField.
440 - ``rxq_pkt_pad_en`` parameter [int]
442 A nonzero value enables padding Rx packet to the size of cacheline on PCI
443 transaction. This feature would waste PCI bandwidth but could improve
444 performance by avoiding partial cacheline write which may cause costly
445 read-modify-copy in memory transaction on some architectures. Disabled by
450 - x86_64 with ConnectX-4, ConnectX-4 Lx, ConnectX-5, ConnectX-6, ConnectX-6 Dx
452 - POWER8 and ARMv8 with ConnectX-4 Lx, ConnectX-5, ConnectX-6, ConnectX-6 Dx
455 - ``mprq_en`` parameter [int]
457 A nonzero value enables configuring Multi-Packet Rx queues. Rx queue is
458 configured as Multi-Packet RQ if the total number of Rx queues is
459 ``rxqs_min_mprq`` or more. Disabled by default.
461 Multi-Packet Rx Queue (MPRQ a.k.a Striding RQ) can further save PCIe bandwidth
462 by posting a single large buffer for multiple packets. Instead of posting a
463 buffers per a packet, one large buffer is posted in order to receive multiple
464 packets on the buffer. A MPRQ buffer consists of multiple fixed-size strides
465 and each stride receives one packet. MPRQ can improve throughput for
466 small-packet traffic.
468 When MPRQ is enabled, max_rx_pkt_len can be larger than the size of
469 user-provided mbuf even if DEV_RX_OFFLOAD_SCATTER isn't enabled. PMD will
470 configure large stride size enough to accommodate max_rx_pkt_len as long as
471 device allows. Note that this can waste system memory compared to enabling Rx
472 scatter and multi-segment packet.
474 - ``mprq_log_stride_num`` parameter [int]
476 Log 2 of the number of strides for Multi-Packet Rx queue. Configuring more
477 strides can reduce PCIe traffic further. If configured value is not in the
478 range of device capability, the default value will be set with a warning
479 message. The default value is 4 which is 16 strides per a buffer, valid only
480 if ``mprq_en`` is set.
482 The size of Rx queue should be bigger than the number of strides.
484 - ``mprq_log_stride_size`` parameter [int]
486 Log 2 of the size of a stride for Multi-Packet Rx queue. Configuring a smaller
487 stride size can save some memory and reduce probability of a depletion of all
488 available strides due to unreleased packets by an application. If configured
489 value is not in the range of device capability, the default value will be set
490 with a warning message. The default value is 11 which is 2048 bytes per a
491 stride, valid only if ``mprq_en`` is set. With ``mprq_log_stride_size`` set
492 it is possible for a pcaket to span across multiple strides. This mode allows
493 support of jumbo frames (9K) with MPRQ. The memcopy of some packets (or part
494 of a packet if Rx scatter is configured) may be required in case there is no
495 space left for a head room at the end of a stride which incurs some
498 - ``mprq_max_memcpy_len`` parameter [int]
500 The maximum length of packet to memcpy in case of Multi-Packet Rx queue. Rx
501 packet is mem-copied to a user-provided mbuf if the size of Rx packet is less
502 than or equal to this parameter. Otherwise, PMD will attach the Rx packet to
503 the mbuf by external buffer attachment - ``rte_pktmbuf_attach_extbuf()``.
504 A mempool for external buffers will be allocated and managed by PMD. If Rx
505 packet is externally attached, ol_flags field of the mbuf will have
506 EXT_ATTACHED_MBUF and this flag must be preserved. ``RTE_MBUF_HAS_EXTBUF()``
507 checks the flag. The default value is 128, valid only if ``mprq_en`` is set.
509 - ``rxqs_min_mprq`` parameter [int]
511 Configure Rx queues as Multi-Packet RQ if the total number of Rx queues is
512 greater or equal to this value. The default value is 12, valid only if
515 - ``txq_inline`` parameter [int]
517 Amount of data to be inlined during TX operations. This parameter is
518 deprecated and converted to the new parameter ``txq_inline_max`` providing
519 partial compatibility.
521 - ``txqs_min_inline`` parameter [int]
523 Enable inline data send only when the number of TX queues is greater or equal
526 This option should be used in combination with ``txq_inline_max`` and
527 ``txq_inline_mpw`` below and does not affect ``txq_inline_min`` settings above.
529 If this option is not specified the default value 16 is used for BlueField
530 and 8 for other platforms
532 The data inlining consumes the CPU cycles, so this option is intended to
533 auto enable inline data if we have enough Tx queues, which means we have
534 enough CPU cores and PCI bandwidth is getting more critical and CPU
535 is not supposed to be bottleneck anymore.
537 The copying data into WQE improves latency and can improve PPS performance
538 when PCI back pressure is detected and may be useful for scenarios involving
539 heavy traffic on many queues.
541 Because additional software logic is necessary to handle this mode, this
542 option should be used with care, as it may lower performance when back
543 pressure is not expected.
545 If inline data are enabled it may affect the maximal size of Tx queue in
546 descriptors because the inline data increase the descriptor size and
547 queue size limits supported by hardware may be exceeded.
549 - ``txq_inline_min`` parameter [int]
551 Minimal amount of data to be inlined into WQE during Tx operations. NICs
552 may require this minimal data amount to operate correctly. The exact value
553 may depend on NIC operation mode, requested offloads, etc. It is strongly
554 recommended to omit this parameter and use the default values. Anyway,
555 applications using this parameter should take into consideration that
556 specifying an inconsistent value may prevent the NIC from sending packets.
558 If ``txq_inline_min`` key is present the specified value (may be aligned
559 by the driver in order not to exceed the limits and provide better descriptor
560 space utilization) will be used by the driver and it is guaranteed that
561 requested amount of data bytes are inlined into the WQE beside other inline
562 settings. This key also may update ``txq_inline_max`` value (default
563 or specified explicitly in devargs) to reserve the space for inline data.
565 If ``txq_inline_min`` key is not present, the value may be queried by the
566 driver from the NIC via DevX if this feature is available. If there is no DevX
567 enabled/supported the value 18 (supposing L2 header including VLAN) is set
568 for ConnectX-4 and ConnectX-4 Lx, and 0 is set by default for ConnectX-5
569 and newer NICs. If packet is shorter the ``txq_inline_min`` value, the entire
572 For ConnectX-4 NIC, driver does not allow specifying value below 18
573 (minimal L2 header, including VLAN), error will be raised.
575 For ConnectX-4 Lx NIC, it is allowed to specify values below 18, but
576 it is not recommended and may prevent NIC from sending packets over
579 Please, note, this minimal data inlining disengages eMPW feature (Enhanced
580 Multi-Packet Write), because last one does not support partial packet inlining.
581 This is not very critical due to minimal data inlining is mostly required
582 by ConnectX-4 and ConnectX-4 Lx, these NICs do not support eMPW feature.
584 - ``txq_inline_max`` parameter [int]
586 Specifies the maximal packet length to be completely inlined into WQE
587 Ethernet Segment for ordinary SEND method. If packet is larger than specified
588 value, the packet data won't be copied by the driver at all, data buffer
589 is addressed with a pointer. If packet length is less or equal all packet
590 data will be copied into WQE. This may improve PCI bandwidth utilization for
591 short packets significantly but requires the extra CPU cycles.
593 The data inline feature is controlled by number of Tx queues, if number of Tx
594 queues is larger than ``txqs_min_inline`` key parameter, the inline feature
595 is engaged, if there are not enough Tx queues (which means not enough CPU cores
596 and CPU resources are scarce), data inline is not performed by the driver.
597 Assigning ``txqs_min_inline`` with zero always enables the data inline.
599 The default ``txq_inline_max`` value is 290. The specified value may be adjusted
600 by the driver in order not to exceed the limit (930 bytes) and to provide better
601 WQE space filling without gaps, the adjustment is reflected in the debug log.
602 Also, the default value (290) may be decreased in run-time if the large transmit
603 queue size is requested and hardware does not support enough descriptor
604 amount, in this case warning is emitted. If ``txq_inline_max`` key is
605 specified and requested inline settings can not be satisfied then error
608 - ``txq_inline_mpw`` parameter [int]
610 Specifies the maximal packet length to be completely inlined into WQE for
611 Enhanced MPW method. If packet is large the specified value, the packet data
612 won't be copied, and data buffer is addressed with pointer. If packet length
613 is less or equal, all packet data will be copied into WQE. This may improve PCI
614 bandwidth utilization for short packets significantly but requires the extra
617 The data inline feature is controlled by number of TX queues, if number of Tx
618 queues is larger than ``txqs_min_inline`` key parameter, the inline feature
619 is engaged, if there are not enough Tx queues (which means not enough CPU cores
620 and CPU resources are scarce), data inline is not performed by the driver.
621 Assigning ``txqs_min_inline`` with zero always enables the data inline.
623 The default ``txq_inline_mpw`` value is 268. The specified value may be adjusted
624 by the driver in order not to exceed the limit (930 bytes) and to provide better
625 WQE space filling without gaps, the adjustment is reflected in the debug log.
626 Due to multiple packets may be included to the same WQE with Enhanced Multi
627 Packet Write Method and overall WQE size is limited it is not recommended to
628 specify large values for the ``txq_inline_mpw``. Also, the default value (268)
629 may be decreased in run-time if the large transmit queue size is requested
630 and hardware does not support enough descriptor amount, in this case warning
631 is emitted. If ``txq_inline_mpw`` key is specified and requested inline
632 settings can not be satisfied then error will be raised.
634 - ``txqs_max_vec`` parameter [int]
636 Enable vectorized Tx only when the number of TX queues is less than or
637 equal to this value. This parameter is deprecated and ignored, kept
638 for compatibility issue to not prevent driver from probing.
640 - ``txq_mpw_hdr_dseg_en`` parameter [int]
642 A nonzero value enables including two pointers in the first block of TX
643 descriptor. The parameter is deprecated and ignored, kept for compatibility
646 - ``txq_max_inline_len`` parameter [int]
648 Maximum size of packet to be inlined. This limits the size of packet to
649 be inlined. If the size of a packet is larger than configured value, the
650 packet isn't inlined even though there's enough space remained in the
651 descriptor. Instead, the packet is included with pointer. This parameter
652 is deprecated and converted directly to ``txq_inline_mpw`` providing full
653 compatibility. Valid only if eMPW feature is engaged.
655 - ``txq_mpw_en`` parameter [int]
657 A nonzero value enables Enhanced Multi-Packet Write (eMPW) for ConnectX-5,
658 ConnectX-6, ConnectX-6 Dx and BlueField. eMPW allows the TX burst function to pack
659 up multiple packets in a single descriptor session in order to save PCI bandwidth
660 and improve performance at the cost of a slightly higher CPU usage. When
661 ``txq_inline_mpw`` is set along with ``txq_mpw_en``, TX burst function copies
662 entire packet data on to TX descriptor instead of including pointer of packet.
664 The Enhanced Multi-Packet Write feature is enabled by default if NIC supports
665 it, can be disabled by explicit specifying 0 value for ``txq_mpw_en`` option.
666 Also, if minimal data inlining is requested by non-zero ``txq_inline_min``
667 option or reported by the NIC, the eMPW feature is disengaged.
669 - ``tx_db_nc`` parameter [int]
671 The rdma core library can map doorbell register in two ways, depending on the
672 environment variable "MLX5_SHUT_UP_BF":
674 - As regular cached memory (usually with write combining attribute), if the
675 variable is either missing or set to zero.
676 - As non-cached memory, if the variable is present and set to not "0" value.
678 The type of mapping may slightly affect the Tx performance, the optimal choice
679 is strongly relied on the host architecture and should be deduced practically.
681 If ``tx_db_nc`` is set to zero, the doorbell is forced to be mapped to regular
682 memory (with write combining), the PMD will perform the extra write memory barrier
683 after writing to doorbell, it might increase the needed CPU clocks per packet
684 to send, but latency might be improved.
686 If ``tx_db_nc`` is set to one, the doorbell is forced to be mapped to non
687 cached memory, the PMD will not perform the extra write memory barrier
688 after writing to doorbell, on some architectures it might improve the
691 If ``tx_db_nc`` is set to two, the doorbell is forced to be mapped to regular
692 memory, the PMD will use heuristics to decide whether write memory barrier
693 should be performed. For bursts with size multiple of recommended one (64 pkts)
694 it is supposed the next burst is coming and no need to issue the extra memory
695 barrier (it is supposed to be issued in the next coming burst, at least after
696 descriptor writing). It might increase latency (on some hosts till next
697 packets transmit) and should be used with care.
699 If ``tx_db_nc`` is omitted or set to zero, the preset (if any) environment
700 variable "MLX5_SHUT_UP_BF" value is used. If there is no "MLX5_SHUT_UP_BF",
701 the default ``tx_db_nc`` value is zero for ARM64 hosts and one for others.
703 - ``tx_vec_en`` parameter [int]
705 A nonzero value enables Tx vector on ConnectX-5, ConnectX-6, ConnectX-6 Dx
706 and BlueField NICs if the number of global Tx queues on the port is less than
707 ``txqs_max_vec``. The parameter is deprecated and ignored.
709 - ``rx_vec_en`` parameter [int]
711 A nonzero value enables Rx vector if the port is not configured in
712 multi-segment otherwise this parameter is ignored.
716 - ``vf_nl_en`` parameter [int]
718 A nonzero value enables Netlink requests from the VF to add/remove MAC
719 addresses or/and enable/disable promiscuous/all multicast on the Netdevice.
720 Otherwise the relevant configuration must be run with Linux iproute2 tools.
721 This is a prerequisite to receive this kind of traffic.
723 Enabled by default, valid only on VF devices ignored otherwise.
725 - ``l3_vxlan_en`` parameter [int]
727 A nonzero value allows L3 VXLAN and VXLAN-GPE flow creation. To enable
728 L3 VXLAN or VXLAN-GPE, users has to configure firmware and enable this
729 parameter. This is a prerequisite to receive this kind of traffic.
733 - ``dv_xmeta_en`` parameter [int]
735 A nonzero value enables extensive flow metadata support if device is
736 capable and driver supports it. This can enable extensive support of
737 ``MARK`` and ``META`` item of ``rte_flow``. The newly introduced
738 ``SET_TAG`` and ``SET_META`` actions do not depend on ``dv_xmeta_en``.
740 There are some possible configurations, depending on parameter value:
742 - 0, this is default value, defines the legacy mode, the ``MARK`` and
743 ``META`` related actions and items operate only within NIC Tx and
744 NIC Rx steering domains, no ``MARK`` and ``META`` information crosses
745 the domain boundaries. The ``MARK`` item is 24 bits wide, the ``META``
746 item is 32 bits wide and match supported on egress only.
748 - 1, this engages extensive metadata mode, the ``MARK`` and ``META``
749 related actions and items operate within all supported steering domains,
750 including FDB, ``MARK`` and ``META`` information may cross the domain
751 boundaries. The ``MARK`` item is 24 bits wide, the ``META`` item width
752 depends on kernel and firmware configurations and might be 0, 16 or
753 32 bits. Within NIC Tx domain ``META`` data width is 32 bits for
754 compatibility, the actual width of data transferred to the FDB domain
755 depends on kernel configuration and may be vary. The actual supported
756 width can be retrieved in runtime by series of rte_flow_validate()
759 - 2, this engages extensive metadata mode, the ``MARK`` and ``META``
760 related actions and items operate within all supported steering domains,
761 including FDB, ``MARK`` and ``META`` information may cross the domain
762 boundaries. The ``META`` item is 32 bits wide, the ``MARK`` item width
763 depends on kernel and firmware configurations and might be 0, 16 or
764 24 bits. The actual supported width can be retrieved in runtime by
765 series of rte_flow_validate() trials.
767 +------+-----------+-----------+-------------+-------------+
768 | Mode | ``MARK`` | ``META`` | ``META`` Tx | FDB/Through |
769 +======+===========+===========+=============+=============+
770 | 0 | 24 bits | 32 bits | 32 bits | no |
771 +------+-----------+-----------+-------------+-------------+
772 | 1 | 24 bits | vary 0-32 | 32 bits | yes |
773 +------+-----------+-----------+-------------+-------------+
774 | 2 | vary 0-32 | 32 bits | 32 bits | yes |
775 +------+-----------+-----------+-------------+-------------+
777 If there is no E-Switch configuration the ``dv_xmeta_en`` parameter is
778 ignored and the device is configured to operate in legacy mode (0).
780 Disabled by default (set to 0).
782 The Direct Verbs/Rules (engaged with ``dv_flow_en`` = 1) supports all
783 of the extensive metadata features. The legacy Verbs supports FLAG and
784 MARK metadata actions over NIC Rx steering domain only.
786 - ``dv_flow_en`` parameter [int]
788 A nonzero value enables the DV flow steering assuming it is supported
789 by the driver (RDMA Core library version is rdma-core-24.0 or higher).
791 Enabled by default if supported.
793 - ``dv_esw_en`` parameter [int]
795 A nonzero value enables E-Switch using Direct Rules.
797 Enabled by default if supported.
799 - ``lacp_by_user`` parameter [int]
801 A nonzero value enables the control of LACP traffic by the user application.
802 When a bond exists in the driver, by default it should be managed by the
803 kernel and therefore LACP traffic should be steered to the kernel.
804 If this devarg is set to 1 it will allow the user to manage the bond by
805 itself and not steer LACP traffic to the kernel.
807 Disabled by default (set to 0).
809 - ``mr_ext_memseg_en`` parameter [int]
811 A nonzero value enables extending memseg when registering DMA memory. If
812 enabled, the number of entries in MR (Memory Region) lookup table on datapath
813 is minimized and it benefits performance. On the other hand, it worsens memory
814 utilization because registered memory is pinned by kernel driver. Even if a
815 page in the extended chunk is freed, that doesn't become reusable until the
816 entire memory is freed.
820 - ``representor`` parameter [list]
822 This parameter can be used to instantiate DPDK Ethernet devices from
823 existing port (or VF) representors configured on the device.
825 It is a standard parameter whose format is described in
826 :ref:`ethernet_device_standard_device_arguments`.
828 For instance, to probe port representors 0 through 2::
832 - ``max_dump_files_num`` parameter [int]
834 The maximum number of files per PMD entity that may be created for debug information.
835 The files will be created in /var/log directory or in current directory.
837 set to 128 by default.
839 - ``lro_timeout_usec`` parameter [int]
841 The maximum allowed duration of an LRO session, in micro-seconds.
842 PMD will set the nearest value supported by HW, which is not bigger than
843 the input ``lro_timeout_usec`` value.
844 If this parameter is not specified, by default PMD will set
845 the smallest value supported by HW.
847 - ``hp_buf_log_sz`` parameter [int]
849 The total data buffer size of a hairpin queue (logarithmic form), in bytes.
850 PMD will set the data buffer size to 2 ** ``hp_buf_log_sz``, both for RX & TX.
851 The capacity of the value is specified by the firmware and the initialization
852 will get a failure if it is out of scope.
853 The range of the value is from 11 to 19 right now, and the supported frame
854 size of a single packet for hairpin is from 512B to 128KB. It might change if
855 different firmware release is being used. By using a small value, it could
856 reduce memory consumption but not work with a large frame. If the value is
857 too large, the memory consumption will be high and some potential performance
858 degradation will be introduced.
859 By default, the PMD will set this value to 16, which means that 9KB jumbo
860 frames will be supported.
862 - ``reclaim_mem_mode`` parameter [int]
864 Cache some resources in flow destroy will help flow recreation more efficient.
865 While some systems may require the all the resources can be reclaimed after
867 The parameter ``reclaim_mem_mode`` provides the option for user to configure
868 if the resource cache is needed or not.
870 There are three options to choose:
872 - 0. It means the flow resources will be cached as usual. The resources will
873 be cached, helpful with flow insertion rate.
875 - 1. It will only enable the DPDK PMD level resources reclaim.
877 - 2. Both DPDK PMD level and rdma-core low level will be configured as
880 By default, the PMD will set this value to 0.
882 .. _mlx5_firmware_config:
884 Firmware configuration
885 ~~~~~~~~~~~~~~~~~~~~~~
887 Firmware features can be configured as key/value pairs.
889 The command to set a value is::
891 mlxconfig -d <device> set <key>=<value>
893 The command to query a value is::
895 mlxconfig -d <device> query | grep <key>
897 The device name for the command ``mlxconfig`` can be either the PCI address,
898 or the mst device name found with::
902 Below are some firmware configurations listed.
908 value: 1=Infiniband 2=Ethernet 3=VPI(auto-sense)
914 - maximum number of SR-IOV virtual functions::
918 - enable DevX (required by Direct Rules and other features)::
922 - aggressive CQE zipping::
926 - L3 VXLAN and VXLAN-GPE destination UDP port::
929 IP_OVER_VXLAN_PORT=<udp dport>
931 - enable VXLAN-GPE tunnel flow matching::
933 FLEX_PARSER_PROFILE_ENABLE=0
935 FLEX_PARSER_PROFILE_ENABLE=2
937 - enable IP-in-IP tunnel flow matching::
939 FLEX_PARSER_PROFILE_ENABLE=0
941 - enable MPLS flow matching::
943 FLEX_PARSER_PROFILE_ENABLE=1
945 - enable ICMP/ICMP6 code/type fields matching::
947 FLEX_PARSER_PROFILE_ENABLE=2
949 - enable Geneve flow matching::
951 FLEX_PARSER_PROFILE_ENABLE=0
953 FLEX_PARSER_PROFILE_ENABLE=1
955 - enable GTP flow matching::
957 FLEX_PARSER_PROFILE_ENABLE=3
962 This driver relies on external libraries and kernel drivers for resources
963 allocations and initialization. The following dependencies are not part of
964 DPDK and must be installed separately:
968 User space Verbs framework used by librte_pmd_mlx5. This library provides
969 a generic interface between the kernel and low-level user space drivers
972 It allows slow and privileged operations (context initialization, hardware
973 resources allocations) to be managed by the kernel and fast operations to
974 never leave user space.
978 Low-level user space driver library for Mellanox
979 ConnectX-4/ConnectX-5/ConnectX-6/BlueField devices, it is automatically loaded
982 This library basically implements send/receive calls to the hardware
987 They provide the kernel-side Verbs API and low level device drivers that
988 manage actual hardware initialization and resources sharing with user
991 Unlike most other PMDs, these modules must remain loaded and bound to
994 - mlx5_core: hardware driver managing Mellanox
995 ConnectX-4/ConnectX-5/ConnectX-6/BlueField devices and related Ethernet kernel
997 - mlx5_ib: InifiniBand device driver.
998 - ib_uverbs: user space driver for Verbs (entry point for libibverbs).
1000 - **Firmware update**
1002 Mellanox OFED/EN releases include firmware updates for
1003 ConnectX-4/ConnectX-5/ConnectX-6/BlueField adapters.
1005 Because each release provides new features, these updates must be applied to
1006 match the kernel modules and libraries they come with.
1010 Both libraries are BSD and GPL licensed. Linux kernel modules are GPL
1016 Either RDMA Core library with a recent enough Linux kernel release
1017 (recommended) or Mellanox OFED/EN, which provides compatibility with older
1020 RDMA Core with Linux Kernel
1021 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
1023 - Minimal kernel version : v4.14 or the most recent 4.14-rc (see `Linux installation documentation`_)
1024 - Minimal rdma-core version: v15+ commit 0c5f5765213a ("Merge pull request #227 from yishaih/tm")
1025 (see `RDMA Core installation documentation`_)
1026 - When building for i686 use:
1028 - rdma-core version 18.0 or above built with 32bit support.
1029 - Kernel version 4.14.41 or above.
1031 - Starting with rdma-core v21, static libraries can be built::
1034 CFLAGS=-fPIC cmake -DIN_PLACE=1 -DENABLE_STATIC=1 -GNinja ..
1037 .. _`Linux installation documentation`: https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git/plain/Documentation/admin-guide/README.rst
1038 .. _`RDMA Core installation documentation`: https://raw.githubusercontent.com/linux-rdma/rdma-core/master/README.md
1040 If rdma-core libraries are built but not installed, DPDK makefile can link them,
1041 thanks to these environment variables:
1043 - ``EXTRA_CFLAGS=-I/path/to/rdma-core/build/include``
1044 - ``EXTRA_LDFLAGS=-L/path/to/rdma-core/build/lib``
1045 - ``PKG_CONFIG_PATH=/path/to/rdma-core/build/lib/pkgconfig``
1050 - Mellanox OFED version: **4.5** and above /
1051 Mellanox EN version: **4.5** and above
1054 - ConnectX-4: **12.21.1000** and above.
1055 - ConnectX-4 Lx: **14.21.1000** and above.
1056 - ConnectX-5: **16.21.1000** and above.
1057 - ConnectX-5 Ex: **16.21.1000** and above.
1058 - ConnectX-6: **20.27.0090** and above.
1059 - ConnectX-6 Dx: **22.27.0090** and above.
1060 - BlueField: **18.25.1010** and above.
1062 While these libraries and kernel modules are available on OpenFabrics
1063 Alliance's `website <https://www.openfabrics.org/>`__ and provided by package
1064 managers on most distributions, this PMD requires Ethernet extensions that
1065 may not be supported at the moment (this is a work in progress).
1068 <http://www.mellanox.com/page/products_dyn?product_family=26&mtag=linux>`__ and
1070 <http://www.mellanox.com/page/products_dyn?product_family=27&mtag=linux>`__
1071 include the necessary support and should be used in the meantime. For DPDK,
1072 only libibverbs, libmlx5, mlnx-ofed-kernel packages and firmware updates are
1073 required from that distribution.
1077 Several versions of Mellanox OFED/EN are available. Installing the version
1078 this DPDK release was developed and tested against is strongly
1079 recommended. Please check the `prerequisites`_.
1084 The following Mellanox device families are supported by the same mlx5 driver:
1094 Below are detailed device names:
1096 * Mellanox\ |reg| ConnectX\ |reg|-4 10G MCX4111A-XCAT (1x10G)
1097 * Mellanox\ |reg| ConnectX\ |reg|-4 10G MCX412A-XCAT (2x10G)
1098 * Mellanox\ |reg| ConnectX\ |reg|-4 25G MCX4111A-ACAT (1x25G)
1099 * Mellanox\ |reg| ConnectX\ |reg|-4 25G MCX412A-ACAT (2x25G)
1100 * Mellanox\ |reg| ConnectX\ |reg|-4 40G MCX413A-BCAT (1x40G)
1101 * Mellanox\ |reg| ConnectX\ |reg|-4 40G MCX4131A-BCAT (1x40G)
1102 * Mellanox\ |reg| ConnectX\ |reg|-4 40G MCX415A-BCAT (1x40G)
1103 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX413A-GCAT (1x50G)
1104 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX4131A-GCAT (1x50G)
1105 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX414A-BCAT (2x50G)
1106 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX415A-GCAT (1x50G)
1107 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX416A-BCAT (2x50G)
1108 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX416A-GCAT (2x50G)
1109 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX415A-CCAT (1x100G)
1110 * Mellanox\ |reg| ConnectX\ |reg|-4 100G MCX416A-CCAT (2x100G)
1111 * Mellanox\ |reg| ConnectX\ |reg|-4 Lx 10G MCX4111A-XCAT (1x10G)
1112 * Mellanox\ |reg| ConnectX\ |reg|-4 Lx 10G MCX4121A-XCAT (2x10G)
1113 * Mellanox\ |reg| ConnectX\ |reg|-4 Lx 25G MCX4111A-ACAT (1x25G)
1114 * Mellanox\ |reg| ConnectX\ |reg|-4 Lx 25G MCX4121A-ACAT (2x25G)
1115 * Mellanox\ |reg| ConnectX\ |reg|-4 Lx 40G MCX4131A-BCAT (1x40G)
1116 * Mellanox\ |reg| ConnectX\ |reg|-5 100G MCX556A-ECAT (2x100G)
1117 * Mellanox\ |reg| ConnectX\ |reg|-5 Ex EN 100G MCX516A-CDAT (2x100G)
1118 * Mellanox\ |reg| ConnectX\ |reg|-6 200G MCX654106A-HCAT (2x200G)
1119 * Mellanox\ |reg| ConnectX\ |reg|-6 Dx EN 100G MCX623106AN-CDAT (2x100G)
1120 * Mellanox\ |reg| ConnectX\ |reg|-6 Dx EN 200G MCX623105AN-VDAT (1x200G)
1122 Quick Start Guide on OFED/EN
1123 ----------------------------
1125 1. Download latest Mellanox OFED/EN. For more info check the `prerequisites`_.
1128 2. Install the required libraries and kernel modules either by installing
1129 only the required set, or by installing the entire Mellanox OFED/EN::
1131 ./mlnxofedinstall --upstream-libs --dpdk
1133 3. Verify the firmware is the correct one::
1137 4. Verify all ports links are set to Ethernet::
1139 mlxconfig -d <mst device> query | grep LINK_TYPE
1143 Link types may have to be configured to Ethernet::
1145 mlxconfig -d <mst device> set LINK_TYPE_P1/2=1/2/3
1147 * LINK_TYPE_P1=<1|2|3> , 1=Infiniband 2=Ethernet 3=VPI(auto-sense)
1149 For hypervisors, verify SR-IOV is enabled on the NIC::
1151 mlxconfig -d <mst device> query | grep SRIOV_EN
1154 If needed, configure SR-IOV::
1156 mlxconfig -d <mst device> set SRIOV_EN=1 NUM_OF_VFS=16
1157 mlxfwreset -d <mst device> reset
1159 5. Restart the driver::
1161 /etc/init.d/openibd restart
1165 service openibd restart
1167 If link type was changed, firmware must be reset as well::
1169 mlxfwreset -d <mst device> reset
1171 For hypervisors, after reset write the sysfs number of virtual functions
1174 To dynamically instantiate a given number of virtual functions (VFs)::
1176 echo [num_vfs] > /sys/class/infiniband/mlx5_0/device/sriov_numvfs
1178 6. Compile DPDK and you are ready to go. See instructions on
1179 :ref:`Development Kit Build System <Development_Kit_Build_System>`
1181 Enable switchdev mode
1182 ---------------------
1184 Switchdev mode is a mode in E-Switch, that binds between representor and VF.
1185 Representor is a port in DPDK that is connected to a VF in such a way
1186 that assuming there are no offload flows, each packet that is sent from the VF
1187 will be received by the corresponding representor. While each packet that is
1188 sent to a representor will be received by the VF.
1189 This is very useful in case of SRIOV mode, where the first packet that is sent
1190 by the VF will be received by the DPDK application which will decide if this
1191 flow should be offloaded to the E-Switch. After offloading the flow packet
1192 that the VF that are matching the flow will not be received any more by
1193 the DPDK application.
1195 1. Enable SRIOV mode::
1197 mlxconfig -d <mst device> set SRIOV_EN=true
1199 2. Configure the max number of VFs::
1201 mlxconfig -d <mst device> set NUM_OF_VFS=<num of vfs>
1205 mlxfwreset -d <mst device> reset
1207 3. Configure the actual number of VFs::
1209 echo <num of vfs > /sys/class/net/<net device>/device/sriov_numvfs
1211 4. Unbind the device (can be rebind after the switchdev mode)::
1213 echo -n "<device pci address" > /sys/bus/pci/drivers/mlx5_core/unbind
1215 5. Enbale switchdev mode::
1217 echo switchdev > /sys/class/net/<net device>/compat/devlink/mode
1222 1. Configure aggressive CQE Zipping for maximum performance::
1224 mlxconfig -d <mst device> s CQE_COMPRESSION=1
1226 To set it back to the default CQE Zipping mode use::
1228 mlxconfig -d <mst device> s CQE_COMPRESSION=0
1230 2. In case of virtualization:
1232 - Make sure that hypervisor kernel is 3.16 or newer.
1233 - Configure boot with ``iommu=pt``.
1234 - Use 1G huge pages.
1235 - Make sure to allocate a VM on huge pages.
1236 - Make sure to set CPU pinning.
1238 3. Use the CPU near local NUMA node to which the PCIe adapter is connected,
1239 for better performance. For VMs, verify that the right CPU
1240 and NUMA node are pinned according to the above. Run::
1244 to identify the NUMA node to which the PCIe adapter is connected.
1246 4. If more than one adapter is used, and root complex capabilities allow
1247 to put both adapters on the same NUMA node without PCI bandwidth degradation,
1248 it is recommended to locate both adapters on the same NUMA node.
1249 This in order to forward packets from one to the other without
1250 NUMA performance penalty.
1252 5. Disable pause frames::
1254 ethtool -A <netdev> rx off tx off
1256 6. Verify IO non-posted prefetch is disabled by default. This can be checked
1257 via the BIOS configuration. Please contact you server provider for more
1258 information about the settings.
1262 On some machines, depends on the machine integrator, it is beneficial
1263 to set the PCI max read request parameter to 1K. This can be
1264 done in the following way:
1266 To query the read request size use::
1268 setpci -s <NIC PCI address> 68.w
1270 If the output is different than 3XXX, set it by::
1272 setpci -s <NIC PCI address> 68.w=3XXX
1274 The XXX can be different on different systems. Make sure to configure
1275 according to the setpci output.
1277 7. To minimize overhead of searching Memory Regions:
1279 - '--socket-mem' is recommended to pin memory by predictable amount.
1280 - Configure per-lcore cache when creating Mempools for packet buffer.
1281 - Refrain from dynamically allocating/freeing memory in run-time.
1283 .. _mlx5_offloads_support:
1285 Supported hardware offloads
1286 ---------------------------
1288 .. table:: Minimal SW/HW versions for queue offloads
1290 ============== ===== ===== ========= ===== ========== ==========
1291 Offload DPDK Linux rdma-core OFED firmware hardware
1292 ============== ===== ===== ========= ===== ========== ==========
1293 common base 17.11 4.14 16 4.2-1 12.21.1000 ConnectX-4
1294 checksums 17.11 4.14 16 4.2-1 12.21.1000 ConnectX-4
1295 Rx timestamp 17.11 4.14 16 4.2-1 12.21.1000 ConnectX-4
1296 TSO 17.11 4.14 16 4.2-1 12.21.1000 ConnectX-4
1297 LRO 19.08 N/A N/A 4.6-4 16.25.6406 ConnectX-5
1298 ============== ===== ===== ========= ===== ========== ==========
1300 .. table:: Minimal SW/HW versions for rte_flow offloads
1302 +-----------------------+-----------------+-----------------+
1303 | Offload | with E-Switch | with NIC |
1304 +=======================+=================+=================+
1305 | Count | | DPDK 19.05 | | DPDK 19.02 |
1306 | | | OFED 4.6 | | OFED 4.6 |
1307 | | | rdma-core 24 | | rdma-core 23 |
1308 | | | ConnectX-5 | | ConnectX-5 |
1309 +-----------------------+-----------------+-----------------+
1310 | Drop | | DPDK 19.05 | | DPDK 18.11 |
1311 | | | OFED 4.6 | | OFED 4.5 |
1312 | | | rdma-core 24 | | rdma-core 23 |
1313 | | | ConnectX-5 | | ConnectX-4 |
1314 +-----------------------+-----------------+-----------------+
1315 | Queue / RSS | | | | DPDK 18.11 |
1316 | | | N/A | | OFED 4.5 |
1317 | | | | | rdma-core 23 |
1318 | | | | | ConnectX-4 |
1319 +-----------------------+-----------------+-----------------+
1320 | Encapsulation | | DPDK 19.05 | | DPDK 19.02 |
1321 | (VXLAN / NVGRE / RAW) | | OFED 4.7-1 | | OFED 4.6 |
1322 | | | rdma-core 24 | | rdma-core 23 |
1323 | | | ConnectX-5 | | ConnectX-5 |
1324 +-----------------------+-----------------+-----------------+
1325 | Encapsulation | | DPDK 19.11 | | DPDK 19.11 |
1326 | GENEVE | | OFED 4.7-3 | | OFED 4.7-3 |
1327 | | | rdma-core 27 | | rdma-core 27 |
1328 | | | ConnectX-5 | | ConnectX-5 |
1329 +-----------------------+-----------------+-----------------+
1330 | | Header rewrite | | DPDK 19.05 | | DPDK 19.02 |
1331 | | (set_ipv4_src / | | OFED 4.7-1 | | OFED 4.7-1 |
1332 | | set_ipv4_dst / | | rdma-core 24 | | rdma-core 24 |
1333 | | set_ipv6_src / | | ConnectX-5 | | ConnectX-5 |
1334 | | set_ipv6_dst / | | | | |
1335 | | set_tp_src / | | | | |
1336 | | set_tp_dst / | | | | |
1337 | | dec_ttl / | | | | |
1338 | | set_ttl / | | | | |
1339 | | set_mac_src / | | | | |
1340 | | set_mac_dst) | | | | |
1341 +-----------------------+-----------------+-----------------+
1342 | | Header rewrite | | DPDK 20.02 | | DPDK 20.02 |
1343 | | (set_dscp) | | OFED 5.0 | | OFED 5.0 |
1344 | | | | rdma-core 24 | | rdma-core 24 |
1345 | | | | ConnectX-5 | | ConnectX-5 |
1346 +-----------------------+-----------------+-----------------+
1347 | Jump | | DPDK 19.05 | | DPDK 19.02 |
1348 | | | OFED 4.7-1 | | OFED 4.7-1 |
1349 | | | rdma-core 24 | | N/A |
1350 | | | ConnectX-5 | | ConnectX-5 |
1351 +-----------------------+-----------------+-----------------+
1352 | Mark / Flag | | DPDK 19.05 | | DPDK 18.11 |
1353 | | | OFED 4.6 | | OFED 4.5 |
1354 | | | rdma-core 24 | | rdma-core 23 |
1355 | | | ConnectX-5 | | ConnectX-4 |
1356 +-----------------------+-----------------+-----------------+
1357 | Port ID | | DPDK 19.05 | | N/A |
1358 | | | OFED 4.7-1 | | N/A |
1359 | | | rdma-core 24 | | N/A |
1360 | | | ConnectX-5 | | N/A |
1361 +-----------------------+-----------------+-----------------+
1362 | | VLAN | | DPDK 19.11 | | DPDK 19.11 |
1363 | | (of_pop_vlan / | | OFED 4.7-1 | | OFED 4.7-1 |
1364 | | of_push_vlan / | | ConnectX-5 | | ConnectX-5 |
1365 | | of_set_vlan_pcp / | | | | |
1366 | | of_set_vlan_vid) | | | | |
1367 +-----------------------+-----------------+-----------------+
1368 | Hairpin | | | | DPDK 19.11 |
1369 | | | N/A | | OFED 4.7-3 |
1370 | | | | | rdma-core 26 |
1371 | | | | | ConnectX-5 |
1372 +-----------------------+-----------------+-----------------+
1373 | Meta data | | DPDK 19.11 | | DPDK 19.11 |
1374 | | | OFED 4.7-3 | | OFED 4.7-3 |
1375 | | | rdma-core 26 | | rdma-core 26 |
1376 | | | ConnectX-5 | | ConnectX-5 |
1377 +-----------------------+-----------------+-----------------+
1378 | Metering | | DPDK 19.11 | | DPDK 19.11 |
1379 | | | OFED 4.7-3 | | OFED 4.7-3 |
1380 | | | rdma-core 26 | | rdma-core 26 |
1381 | | | ConnectX-5 | | ConnectX-5 |
1382 +-----------------------+-----------------+-----------------+
1387 MARK and META items are interrelated with datapath - they might move from/to
1388 the applications in mbuf fields. Hence, zero value for these items has the
1389 special meaning - it means "no metadata are provided", not zero values are
1390 treated by applications and PMD as valid ones.
1392 Moreover in the flow engine domain the value zero is acceptable to match and
1393 set, and we should allow to specify zero values as rte_flow parameters for the
1394 META and MARK items and actions. In the same time zero mask has no meaning and
1395 should be rejected on validation stage.
1400 Flows are not cached in the driver.
1401 When stopping a device port, all the flows created on this port from the
1402 application will be flushed automatically in the background.
1403 After stopping the device port, all flows on this port become invalid and
1404 not represented in the system.
1405 All references to these flows held by the application should be discarded
1406 directly but neither destroyed nor flushed.
1408 The application should re-create the flows as required after the port restart.
1413 Compared to librte_pmd_mlx4 that implements a single RSS configuration per
1414 port, librte_pmd_mlx5 supports per-protocol RSS configuration.
1416 Since ``testpmd`` defaults to IP RSS mode and there is currently no
1417 command-line parameter to enable additional protocols (UDP and TCP as well
1418 as IP), the following commands must be entered from its CLI to get the same
1419 behavior as librte_pmd_mlx4::
1422 > port config all rss all
1428 This section demonstrates how to launch **testpmd** with Mellanox
1429 ConnectX-4/ConnectX-5/ConnectX-6/BlueField devices managed by librte_pmd_mlx5.
1431 #. Load the kernel modules::
1433 modprobe -a ib_uverbs mlx5_core mlx5_ib
1435 Alternatively if MLNX_OFED/MLNX_EN is fully installed, the following script
1438 /etc/init.d/openibd restart
1442 User space I/O kernel modules (uio and igb_uio) are not used and do
1443 not have to be loaded.
1445 #. Make sure Ethernet interfaces are in working order and linked to kernel
1446 verbs. Related sysfs entries should be present::
1448 ls -d /sys/class/net/*/device/infiniband_verbs/uverbs* | cut -d / -f 5
1457 #. Optionally, retrieve their PCI bus addresses for whitelisting::
1460 for intf in eth2 eth3 eth4 eth5;
1462 (cd "/sys/class/net/${intf}/device/" && pwd -P);
1465 sed -n 's,.*/\(.*\),-w \1,p'
1474 #. Request huge pages::
1476 echo 1024 > /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages/nr_hugepages
1478 #. Start testpmd with basic parameters::
1480 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
1485 EAL: PCI device 0000:05:00.0 on NUMA socket 0
1486 EAL: probe driver: 15b3:1013 librte_pmd_mlx5
1487 PMD: librte_pmd_mlx5: PCI information matches, using device "mlx5_0" (VF: false)
1488 PMD: librte_pmd_mlx5: 1 port(s) detected
1489 PMD: librte_pmd_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:fe
1490 EAL: PCI device 0000:05:00.1 on NUMA socket 0
1491 EAL: probe driver: 15b3:1013 librte_pmd_mlx5
1492 PMD: librte_pmd_mlx5: PCI information matches, using device "mlx5_1" (VF: false)
1493 PMD: librte_pmd_mlx5: 1 port(s) detected
1494 PMD: librte_pmd_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:ff
1495 EAL: PCI device 0000:06:00.0 on NUMA socket 0
1496 EAL: probe driver: 15b3:1013 librte_pmd_mlx5
1497 PMD: librte_pmd_mlx5: PCI information matches, using device "mlx5_2" (VF: false)
1498 PMD: librte_pmd_mlx5: 1 port(s) detected
1499 PMD: librte_pmd_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:fa
1500 EAL: PCI device 0000:06:00.1 on NUMA socket 0
1501 EAL: probe driver: 15b3:1013 librte_pmd_mlx5
1502 PMD: librte_pmd_mlx5: PCI information matches, using device "mlx5_3" (VF: false)
1503 PMD: librte_pmd_mlx5: 1 port(s) detected
1504 PMD: librte_pmd_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:fb
1505 Interactive-mode selected
1506 Configuring Port 0 (socket 0)
1507 PMD: librte_pmd_mlx5: 0x8cba80: TX queues number update: 0 -> 2
1508 PMD: librte_pmd_mlx5: 0x8cba80: RX queues number update: 0 -> 2
1509 Port 0: E4:1D:2D:E7:0C:FE
1510 Configuring Port 1 (socket 0)
1511 PMD: librte_pmd_mlx5: 0x8ccac8: TX queues number update: 0 -> 2
1512 PMD: librte_pmd_mlx5: 0x8ccac8: RX queues number update: 0 -> 2
1513 Port 1: E4:1D:2D:E7:0C:FF
1514 Configuring Port 2 (socket 0)
1515 PMD: librte_pmd_mlx5: 0x8cdb10: TX queues number update: 0 -> 2
1516 PMD: librte_pmd_mlx5: 0x8cdb10: RX queues number update: 0 -> 2
1517 Port 2: E4:1D:2D:E7:0C:FA
1518 Configuring Port 3 (socket 0)
1519 PMD: librte_pmd_mlx5: 0x8ceb58: TX queues number update: 0 -> 2
1520 PMD: librte_pmd_mlx5: 0x8ceb58: RX queues number update: 0 -> 2
1521 Port 3: E4:1D:2D:E7:0C:FB
1522 Checking link statuses...
1523 Port 0 Link Up - speed 40000 Mbps - full-duplex
1524 Port 1 Link Up - speed 40000 Mbps - full-duplex
1525 Port 2 Link Up - speed 10000 Mbps - full-duplex
1526 Port 3 Link Up - speed 10000 Mbps - full-duplex
1533 This section demonstrates how to dump flows. Currently, it's possible to dump
1534 all flows with assistance of external tools.
1536 #. 2 ways to get flow raw file:
1538 - Using testpmd CLI:
1540 .. code-block:: console
1542 testpmd> flow dump <port> <output_file>
1544 - call rte_flow_dev_dump api:
1546 .. code-block:: console
1548 rte_flow_dev_dump(port, file, NULL);
1550 #. Dump human-readable flows from raw file:
1552 Get flow parsing tool from: https://github.com/Mellanox/mlx_steering_dump
1554 .. code-block:: console
1556 mlx_steering_dump.py -f <output_file>