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_net_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>`__.
27 Besides its dependency on libibverbs (that implies libmlx5 and associated
28 kernel support), librte_net_mlx5 relies heavily on system calls for control
29 operations such as querying/updating the MTU and flow control parameters.
31 For security reasons and robustness, this driver only deals with virtual
32 memory addresses. The way resources allocations are handled by the kernel,
33 combined with hardware specifications that allow to handle virtual memory
34 addresses directly, ensure that DPDK applications cannot access random
35 physical memory (or memory that does not belong to the current process).
37 This capability allows the PMD to coexist with kernel network interfaces
38 which remain functional, although they stop receiving unicast packets as
39 long as they share the same MAC address.
40 This means legacy linux control tools (for example: ethtool, ifconfig and
41 more) can operate on the same network interfaces that owned by the DPDK
44 The PMD can use libibverbs and libmlx5 to access the device firmware
45 or directly the hardware components.
46 There are different levels of objects and bypassing abilities
47 to get the best performances:
49 - Verbs is a complete high-level generic API
50 - Direct Verbs is a device-specific API
51 - DevX allows to access firmware objects
52 - Direct Rules manages flow steering at low-level hardware layer
54 Enabling librte_net_mlx5 causes DPDK applications to be linked against
60 - Multi arch support: x86_64, POWER8, ARMv8, i686.
61 - Multiple TX and RX queues.
62 - Support for scattered TX and RX frames.
63 - IPv4, IPv6, TCPv4, TCPv6, UDPv4 and UDPv6 RSS on any number of queues.
64 - RSS using different combinations of fields: L3 only, L4 only or both,
65 and source only, destination only or both.
66 - Several RSS hash keys, one for each flow type.
67 - Default RSS operation with no hash key specification.
68 - Configurable RETA table.
69 - Link flow control (pause frame).
70 - Support for multiple MAC addresses.
74 - RX CRC stripping configuration.
75 - Promiscuous mode on PF and VF.
76 - Multicast promiscuous mode on PF and VF.
77 - Hardware checksum offloads.
78 - Flow director (RTE_FDIR_MODE_PERFECT, RTE_FDIR_MODE_PERFECT_MAC_VLAN and
80 - Flow API, including :ref:`flow_isolated_mode`.
82 - KVM and VMware ESX SR-IOV modes are supported.
83 - RSS hash result is supported.
84 - Hardware TSO for generic IP or UDP tunnel, including VXLAN and GRE.
85 - Hardware checksum Tx offload for generic IP or UDP tunnel, including VXLAN and GRE.
87 - Statistics query including Basic, Extended and per queue.
89 - Tunnel types: VXLAN, L3 VXLAN, VXLAN-GPE, GRE, MPLSoGRE, MPLSoUDP, IP-in-IP, Geneve, GTP.
90 - Tunnel HW offloads: packet type, inner/outer RSS, IP and UDP checksum verification.
91 - NIC HW offloads: encapsulation (vxlan, gre, mplsoudp, mplsogre), NAT, routing, TTL
92 increment/decrement, count, drop, mark. For details please see :ref:`mlx5_offloads_support`.
93 - Flow insertion rate of more then million flows per second, when using Direct Rules.
94 - Support for multiple rte_flow groups.
95 - Per packet no-inline hint flag to disable packet data copying into Tx descriptors.
102 - For secondary process:
104 - Forked secondary process not supported.
105 - External memory unregistered in EAL memseg list cannot be used for DMA
106 unless such memory has been registered by ``mlx5_mr_update_ext_mp()`` in
107 primary process and remapped to the same virtual address in secondary
108 process. If the external memory is registered by primary process but has
109 different virtual address in secondary process, unexpected error may happen.
111 - When using Verbs flow engine (``dv_flow_en`` = 0), flow pattern without any
112 specific VLAN will match for VLAN packets as well:
114 When VLAN spec is not specified in the pattern, the matching rule will be created with VLAN as a wild card.
115 Meaning, the flow rule::
117 flow create 0 ingress pattern eth / vlan vid is 3 / ipv4 / end ...
119 Will only match vlan packets with vid=3. and the flow rule::
121 flow create 0 ingress pattern eth / ipv4 / end ...
123 Will match any ipv4 packet (VLAN included).
125 - When using DV flow engine (``dv_flow_en`` = 1), flow pattern without VLAN item
126 will match untagged packets only.
129 flow create 0 ingress pattern eth / ipv4 / end ...
131 Will match untagged packets only.
134 flow create 0 ingress pattern eth / vlan / ipv4 / end ...
136 Will match tagged packets only, with any VLAN ID value.
139 flow create 0 ingress pattern eth / vlan vid is 3 / ipv4 / end ...
141 Will only match tagged packets with VLAN ID 3.
143 - VLAN pop offload command:
145 - Flow rules having a VLAN pop offload command as one of their actions and
146 are lacking a match on VLAN as one of their items are not supported.
147 - The command is not supported on egress traffic.
149 - VLAN push offload is not supported on ingress traffic.
151 - VLAN set PCP offload is not supported on existing headers.
153 - A multi segment packet must have not more segments than reported by dev_infos_get()
154 in tx_desc_lim.nb_seg_max field. This value depends on maximal supported Tx descriptor
155 size and ``txq_inline_min`` settings and may be from 2 (worst case forced by maximal
156 inline settings) to 58.
158 - Flows with a VXLAN Network Identifier equal (or ends to be equal)
159 to 0 are not supported.
161 - L3 VXLAN and VXLAN-GPE tunnels cannot be supported together with MPLSoGRE and MPLSoUDP.
163 - Match on Geneve header supports the following fields only:
169 Currently, the only supported options length value is 0.
171 - VF: flow rules created on VF devices can only match traffic targeted at the
172 configured MAC addresses (see ``rte_eth_dev_mac_addr_add()``).
174 - Match on GTP tunnel header item supports the following fields only:
176 - v_pt_rsv_flags: E flag, S flag, PN flag
180 - No Tx metadata go to the E-Switch steering domain for the Flow group 0.
181 The flows within group 0 and set metadata action are rejected by hardware.
185 MAC addresses not already present in the bridge table of the associated
186 kernel network device will be added and cleaned up by the PMD when closing
187 the device. In case of ungraceful program termination, some entries may
188 remain present and should be removed manually by other means.
190 - When Multi-Packet Rx queue is configured (``mprq_en``), a Rx packet can be
191 externally attached to a user-provided mbuf with having EXT_ATTACHED_MBUF in
192 ol_flags. As the mempool for the external buffer is managed by PMD, all the
193 Rx mbufs must be freed before the device is closed. Otherwise, the mempool of
194 the external buffers will be freed by PMD and the application which still
195 holds the external buffers may be corrupted.
197 - If Multi-Packet Rx queue is configured (``mprq_en``) and Rx CQE compression is
198 enabled (``rxq_cqe_comp_en``) at the same time, RSS hash result is not fully
199 supported. Some Rx packets may not have PKT_RX_RSS_HASH.
201 - IPv6 Multicast messages are not supported on VM, while promiscuous mode
202 and allmulticast mode are both set to off.
203 To receive IPv6 Multicast messages on VM, explicitly set the relevant
204 MAC address using rte_eth_dev_mac_addr_add() API.
206 - To support a mixed traffic pattern (some buffers from local host memory, some
207 buffers from other devices) with high bandwidth, a mbuf flag is used.
209 An application hints the PMD whether or not it should try to inline the
210 given mbuf data buffer. PMD should do the best effort to act upon this request.
212 The hint flag ``RTE_PMD_MLX5_FINE_GRANULARITY_INLINE`` is dynamic,
213 registered by application with rte_mbuf_dynflag_register(). This flag is
214 purely driver-specific and declared in PMD specific header ``rte_pmd_mlx5.h``,
215 which is intended to be used by the application.
217 To query the supported specific flags in runtime,
218 the function ``rte_pmd_mlx5_get_dyn_flag_names`` returns the array of
219 currently (over present hardware and configuration) supported specific flags.
220 The "not inline hint" feature operating flow is the following one:
223 - probe the devices, ports are created
224 - query the port capabilities
225 - if port supporting the feature is found
226 - register dynamic flag ``RTE_PMD_MLX5_FINE_GRANULARITY_INLINE``
227 - application starts the ports
228 - on ``dev_start()`` PMD checks whether the feature flag is registered and
229 enables the feature support in datapath
230 - application might set the registered flag bit in ``ol_flags`` field
231 of mbuf being sent and PMD will handle ones appropriately.
233 - The amount of descriptors in Tx queue may be limited by data inline settings.
234 Inline data require the more descriptor building blocks and overall block
235 amount may exceed the hardware supported limits. The application should
236 reduce the requested Tx size or adjust data inline settings with
237 ``txq_inline_max`` and ``txq_inline_mpw`` devargs keys.
239 - To provide the packet send scheduling on mbuf timestamps the ``tx_pp``
240 parameter should be specified.
241 When PMD sees the RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME set on the packet
242 being sent it tries to synchronize the time of packet appearing on
243 the wire with the specified packet timestamp. It the specified one
244 is in the past it should be ignored, if one is in the distant future
245 it should be capped with some reasonable value (in range of seconds).
246 These specific cases ("too late" and "distant future") can be optionally
247 reported via device xstats to assist applications to detect the
248 time-related problems.
250 The timestamp upper "too-distant-future" limit
251 at the moment of invoking the Tx burst routine
252 can be estimated as ``tx_pp`` option (in nanoseconds) multiplied by 2^23.
253 Please note, for the testpmd txonly mode,
254 the limit is deduced from the expression::
256 (n_tx_descriptors / burst_size + 1) * inter_burst_gap
258 There is no any packet reordering according timestamps is supposed,
259 neither within packet burst, nor between packets, it is an entirely
260 application responsibility to generate packets and its timestamps
261 in desired order. The timestamps can be put only in the first packet
262 in the burst providing the entire burst scheduling.
264 - E-Switch decapsulation Flow:
266 - can be applied to PF port only.
267 - must specify VF port action (packet redirection from PF to VF).
268 - optionally may specify tunnel inner source and destination MAC addresses.
270 - E-Switch encapsulation Flow:
272 - can be applied to VF ports only.
273 - must specify PF port action (packet redirection from VF to PF).
277 - The input buffer, used as outer header, is not validated.
281 - The decapsulation is always done up to the outermost tunnel detected by the HW.
282 - The input buffer, providing the removal size, is not validated.
283 - The buffer size must match the length of the headers to be removed.
285 - ICMP(code/type/identifier/sequence number) / ICMP6(code/type) matching, IP-in-IP and MPLS flow matching are all
286 mutually exclusive features which cannot be supported together
287 (see :ref:`mlx5_firmware_config`).
291 - Requires DevX and DV flow to be enabled.
292 - KEEP_CRC offload cannot be supported with LRO.
293 - The first mbuf length, without head-room, must be big enough to include the
295 - Rx queue with LRO offload enabled, receiving a non-LRO packet, can forward
296 it with size limited to max LRO size, not to max RX packet length.
297 - LRO can be used with outer header of TCP packets of the standard format:
298 eth (with or without vlan) / ipv4 or ipv6 / tcp / payload
300 Other TCP packets (e.g. with MPLS label) received on Rx queue with LRO enabled, will be received with bad checksum.
301 - LRO packet aggregation is performed by HW only for packet size larger than
302 ``lro_min_mss_size``. This value is reported on device start, when debug
307 - ``DEV_RX_OFFLOAD_KEEP_CRC`` cannot be supported with decapsulation
308 for some NICs (such as ConnectX-6 Dx and BlueField 2).
309 The capability bit ``scatter_fcs_w_decap_disable`` shows NIC support.
313 - Supports ``RTE_FLOW_ACTION_TYPE_SAMPLE`` action only within NIC Rx and E-Switch steering domain.
314 - The E-Switch Sample flow must have the eswitch_manager VPORT destination (PF or ECPF) and no additional actions.
315 - For ConnectX-5, the ``RTE_FLOW_ACTION_TYPE_SAMPLE`` is typically used as first action in the E-Switch egress flow if with header modify or encapsulation actions.
317 - IPv6 header item 'proto' field, indicating the next header protocol, should
318 not be set as extension header.
319 In case the next header is an extension header, it should not be specified in
320 IPv6 header item 'proto' field.
321 The last extension header item 'next header' field can specify the following
322 header protocol type.
327 MLX5 supports various methods to report statistics:
329 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.
331 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.
333 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.
341 The ibverbs libraries can be linked with this PMD in a number of ways,
342 configured by the ``ibverbs_link`` build option:
344 - ``shared`` (default): the PMD depends on some .so files.
346 - ``dlopen``: Split the dependencies glue in a separate library
347 loaded when needed by dlopen.
348 It make dependencies on libibverbs and libmlx4 optional,
349 and has no performance impact.
351 - ``static``: Embed static flavor of the dependencies libibverbs and libmlx4
352 in the PMD shared library or the executable static binary.
354 Environment variables
355 ~~~~~~~~~~~~~~~~~~~~~
359 A list of directories in which to search for the rdma-core "glue" plug-in,
360 separated by colons or semi-colons.
362 - ``MLX5_SHUT_UP_BF``
364 Configures HW Tx doorbell register as IO-mapped.
366 By default, the HW Tx doorbell is configured as a write-combining register.
367 The register would be flushed to HW usually when the write-combining buffer
368 becomes full, but it depends on CPU design.
370 Except for vectorized Tx burst routines, a write memory barrier is enforced
371 after updating the register so that the update can be immediately visible to
374 When vectorized Tx burst is called, the barrier is set only if the burst size
375 is not aligned to MLX5_VPMD_TX_MAX_BURST. However, setting this environmental
376 variable will bring better latency even though the maximum throughput can
379 Run-time configuration
380 ~~~~~~~~~~~~~~~~~~~~~~
382 - librte_net_mlx5 brings kernel network interfaces up during initialization
383 because it is affected by their state. Forcing them down prevents packets
386 - **ethtool** operations on related kernel interfaces also affect the PMD.
391 In order to run as a non-root user,
392 some capabilities must be granted to the application::
394 setcap cap_sys_admin,cap_net_admin,cap_net_raw,cap_ipc_lock+ep <dpdk-app>
396 Below are the reasons of the need for each capability:
399 When using physical addresses (PA mode), with Linux >= 4.0,
400 for access to ``/proc/self/pagemap``.
403 For device configuration.
406 For raw ethernet queue allocation through kernel driver.
409 For DMA memory pinning.
414 - ``rxq_cqe_comp_en`` parameter [int]
416 A nonzero value enables the compression of CQE on RX side. This feature
417 allows to save PCI bandwidth and improve performance. Enabled by default.
421 - x86_64 with ConnectX-4, ConnectX-4 Lx, ConnectX-5, ConnectX-6, ConnectX-6 Dx
423 - POWER9 and ARMv8 with ConnectX-4 Lx, ConnectX-5, ConnectX-6, ConnectX-6 Dx
426 - ``rxq_cqe_pad_en`` parameter [int]
428 A nonzero value enables 128B padding of CQE on RX side. The size of CQE
429 is aligned with the size of a cacheline of the core. If cacheline size is
430 128B, the CQE size is configured to be 128B even though the device writes
431 only 64B data on the cacheline. This is to avoid unnecessary cache
432 invalidation by device's two consecutive writes on to one cacheline.
433 However in some architecture, it is more beneficial to update entire
434 cacheline with padding the rest 64B rather than striding because
435 read-modify-write could drop performance a lot. On the other hand,
436 writing extra data will consume more PCIe bandwidth and could also drop
437 the maximum throughput. It is recommended to empirically set this
438 parameter. Disabled by default.
442 - CPU having 128B cacheline with ConnectX-5 and BlueField.
444 - ``rxq_pkt_pad_en`` parameter [int]
446 A nonzero value enables padding Rx packet to the size of cacheline on PCI
447 transaction. This feature would waste PCI bandwidth but could improve
448 performance by avoiding partial cacheline write which may cause costly
449 read-modify-copy in memory transaction on some architectures. Disabled by
454 - x86_64 with ConnectX-4, ConnectX-4 Lx, ConnectX-5, ConnectX-6, ConnectX-6 Dx
456 - POWER8 and ARMv8 with ConnectX-4 Lx, ConnectX-5, ConnectX-6, ConnectX-6 Dx
459 - ``mprq_en`` parameter [int]
461 A nonzero value enables configuring Multi-Packet Rx queues. Rx queue is
462 configured as Multi-Packet RQ if the total number of Rx queues is
463 ``rxqs_min_mprq`` or more. Disabled by default.
465 Multi-Packet Rx Queue (MPRQ a.k.a Striding RQ) can further save PCIe bandwidth
466 by posting a single large buffer for multiple packets. Instead of posting a
467 buffers per a packet, one large buffer is posted in order to receive multiple
468 packets on the buffer. A MPRQ buffer consists of multiple fixed-size strides
469 and each stride receives one packet. MPRQ can improve throughput for
470 small-packet traffic.
472 When MPRQ is enabled, max_rx_pkt_len can be larger than the size of
473 user-provided mbuf even if DEV_RX_OFFLOAD_SCATTER isn't enabled. PMD will
474 configure large stride size enough to accommodate max_rx_pkt_len as long as
475 device allows. Note that this can waste system memory compared to enabling Rx
476 scatter and multi-segment packet.
478 - ``mprq_log_stride_num`` parameter [int]
480 Log 2 of the number of strides for Multi-Packet Rx queue. Configuring more
481 strides can reduce PCIe traffic further. If configured value is not in the
482 range of device capability, the default value will be set with a warning
483 message. The default value is 4 which is 16 strides per a buffer, valid only
484 if ``mprq_en`` is set.
486 The size of Rx queue should be bigger than the number of strides.
488 - ``mprq_log_stride_size`` parameter [int]
490 Log 2 of the size of a stride for Multi-Packet Rx queue. Configuring a smaller
491 stride size can save some memory and reduce probability of a depletion of all
492 available strides due to unreleased packets by an application. If configured
493 value is not in the range of device capability, the default value will be set
494 with a warning message. The default value is 11 which is 2048 bytes per a
495 stride, valid only if ``mprq_en`` is set. With ``mprq_log_stride_size`` set
496 it is possible for a packet to span across multiple strides. This mode allows
497 support of jumbo frames (9K) with MPRQ. The memcopy of some packets (or part
498 of a packet if Rx scatter is configured) may be required in case there is no
499 space left for a head room at the end of a stride which incurs some
502 - ``mprq_max_memcpy_len`` parameter [int]
504 The maximum length of packet to memcpy in case of Multi-Packet Rx queue. Rx
505 packet is mem-copied to a user-provided mbuf if the size of Rx packet is less
506 than or equal to this parameter. Otherwise, PMD will attach the Rx packet to
507 the mbuf by external buffer attachment - ``rte_pktmbuf_attach_extbuf()``.
508 A mempool for external buffers will be allocated and managed by PMD. If Rx
509 packet is externally attached, ol_flags field of the mbuf will have
510 EXT_ATTACHED_MBUF and this flag must be preserved. ``RTE_MBUF_HAS_EXTBUF()``
511 checks the flag. The default value is 128, valid only if ``mprq_en`` is set.
513 - ``rxqs_min_mprq`` parameter [int]
515 Configure Rx queues as Multi-Packet RQ if the total number of Rx queues is
516 greater or equal to this value. The default value is 12, valid only if
519 - ``txq_inline`` parameter [int]
521 Amount of data to be inlined during TX operations. This parameter is
522 deprecated and converted to the new parameter ``txq_inline_max`` providing
523 partial compatibility.
525 - ``txqs_min_inline`` parameter [int]
527 Enable inline data send only when the number of TX queues is greater or equal
530 This option should be used in combination with ``txq_inline_max`` and
531 ``txq_inline_mpw`` below and does not affect ``txq_inline_min`` settings above.
533 If this option is not specified the default value 16 is used for BlueField
534 and 8 for other platforms
536 The data inlining consumes the CPU cycles, so this option is intended to
537 auto enable inline data if we have enough Tx queues, which means we have
538 enough CPU cores and PCI bandwidth is getting more critical and CPU
539 is not supposed to be bottleneck anymore.
541 The copying data into WQE improves latency and can improve PPS performance
542 when PCI back pressure is detected and may be useful for scenarios involving
543 heavy traffic on many queues.
545 Because additional software logic is necessary to handle this mode, this
546 option should be used with care, as it may lower performance when back
547 pressure is not expected.
549 If inline data are enabled it may affect the maximal size of Tx queue in
550 descriptors because the inline data increase the descriptor size and
551 queue size limits supported by hardware may be exceeded.
553 - ``txq_inline_min`` parameter [int]
555 Minimal amount of data to be inlined into WQE during Tx operations. NICs
556 may require this minimal data amount to operate correctly. The exact value
557 may depend on NIC operation mode, requested offloads, etc. It is strongly
558 recommended to omit this parameter and use the default values. Anyway,
559 applications using this parameter should take into consideration that
560 specifying an inconsistent value may prevent the NIC from sending packets.
562 If ``txq_inline_min`` key is present the specified value (may be aligned
563 by the driver in order not to exceed the limits and provide better descriptor
564 space utilization) will be used by the driver and it is guaranteed that
565 requested amount of data bytes are inlined into the WQE beside other inline
566 settings. This key also may update ``txq_inline_max`` value (default
567 or specified explicitly in devargs) to reserve the space for inline data.
569 If ``txq_inline_min`` key is not present, the value may be queried by the
570 driver from the NIC via DevX if this feature is available. If there is no DevX
571 enabled/supported the value 18 (supposing L2 header including VLAN) is set
572 for ConnectX-4 and ConnectX-4 Lx, and 0 is set by default for ConnectX-5
573 and newer NICs. If packet is shorter the ``txq_inline_min`` value, the entire
576 For ConnectX-4 NIC, driver does not allow specifying value below 18
577 (minimal L2 header, including VLAN), error will be raised.
579 For ConnectX-4 Lx NIC, it is allowed to specify values below 18, but
580 it is not recommended and may prevent NIC from sending packets over
583 Please, note, this minimal data inlining disengages eMPW feature (Enhanced
584 Multi-Packet Write), because last one does not support partial packet inlining.
585 This is not very critical due to minimal data inlining is mostly required
586 by ConnectX-4 and ConnectX-4 Lx, these NICs do not support eMPW feature.
588 - ``txq_inline_max`` parameter [int]
590 Specifies the maximal packet length to be completely inlined into WQE
591 Ethernet Segment for ordinary SEND method. If packet is larger than specified
592 value, the packet data won't be copied by the driver at all, data buffer
593 is addressed with a pointer. If packet length is less or equal all packet
594 data will be copied into WQE. This may improve PCI bandwidth utilization for
595 short packets significantly but requires the extra CPU cycles.
597 The data inline feature is controlled by number of Tx queues, if number of Tx
598 queues is larger than ``txqs_min_inline`` key parameter, the inline feature
599 is engaged, if there are not enough Tx queues (which means not enough CPU cores
600 and CPU resources are scarce), data inline is not performed by the driver.
601 Assigning ``txqs_min_inline`` with zero always enables the data inline.
603 The default ``txq_inline_max`` value is 290. The specified value may be adjusted
604 by the driver in order not to exceed the limit (930 bytes) and to provide better
605 WQE space filling without gaps, the adjustment is reflected in the debug log.
606 Also, the default value (290) may be decreased in run-time if the large transmit
607 queue size is requested and hardware does not support enough descriptor
608 amount, in this case warning is emitted. If ``txq_inline_max`` key is
609 specified and requested inline settings can not be satisfied then error
612 - ``txq_inline_mpw`` parameter [int]
614 Specifies the maximal packet length to be completely inlined into WQE for
615 Enhanced MPW method. If packet is large the specified value, the packet data
616 won't be copied, and data buffer is addressed with pointer. If packet length
617 is less or equal, all packet data will be copied into WQE. This may improve PCI
618 bandwidth utilization for short packets significantly but requires the extra
621 The data inline feature is controlled by number of TX queues, if number of Tx
622 queues is larger than ``txqs_min_inline`` key parameter, the inline feature
623 is engaged, if there are not enough Tx queues (which means not enough CPU cores
624 and CPU resources are scarce), data inline is not performed by the driver.
625 Assigning ``txqs_min_inline`` with zero always enables the data inline.
627 The default ``txq_inline_mpw`` value is 268. The specified value may be adjusted
628 by the driver in order not to exceed the limit (930 bytes) and to provide better
629 WQE space filling without gaps, the adjustment is reflected in the debug log.
630 Due to multiple packets may be included to the same WQE with Enhanced Multi
631 Packet Write Method and overall WQE size is limited it is not recommended to
632 specify large values for the ``txq_inline_mpw``. Also, the default value (268)
633 may be decreased in run-time if the large transmit queue size is requested
634 and hardware does not support enough descriptor amount, in this case warning
635 is emitted. If ``txq_inline_mpw`` key is specified and requested inline
636 settings can not be satisfied then error will be raised.
638 - ``txqs_max_vec`` parameter [int]
640 Enable vectorized Tx only when the number of TX queues is less than or
641 equal to this value. This parameter is deprecated and ignored, kept
642 for compatibility issue to not prevent driver from probing.
644 - ``txq_mpw_hdr_dseg_en`` parameter [int]
646 A nonzero value enables including two pointers in the first block of TX
647 descriptor. The parameter is deprecated and ignored, kept for compatibility
650 - ``txq_max_inline_len`` parameter [int]
652 Maximum size of packet to be inlined. This limits the size of packet to
653 be inlined. If the size of a packet is larger than configured value, the
654 packet isn't inlined even though there's enough space remained in the
655 descriptor. Instead, the packet is included with pointer. This parameter
656 is deprecated and converted directly to ``txq_inline_mpw`` providing full
657 compatibility. Valid only if eMPW feature is engaged.
659 - ``txq_mpw_en`` parameter [int]
661 A nonzero value enables Enhanced Multi-Packet Write (eMPW) for ConnectX-5,
662 ConnectX-6, ConnectX-6 Dx and BlueField. eMPW allows the TX burst function to pack
663 up multiple packets in a single descriptor session in order to save PCI bandwidth
664 and improve performance at the cost of a slightly higher CPU usage. When
665 ``txq_inline_mpw`` is set along with ``txq_mpw_en``, TX burst function copies
666 entire packet data on to TX descriptor instead of including pointer of packet.
668 The Enhanced Multi-Packet Write feature is enabled by default if NIC supports
669 it, can be disabled by explicit specifying 0 value for ``txq_mpw_en`` option.
670 Also, if minimal data inlining is requested by non-zero ``txq_inline_min``
671 option or reported by the NIC, the eMPW feature is disengaged.
673 - ``tx_db_nc`` parameter [int]
675 The rdma core library can map doorbell register in two ways, depending on the
676 environment variable "MLX5_SHUT_UP_BF":
678 - As regular cached memory (usually with write combining attribute), if the
679 variable is either missing or set to zero.
680 - As non-cached memory, if the variable is present and set to not "0" value.
682 The type of mapping may slightly affect the Tx performance, the optimal choice
683 is strongly relied on the host architecture and should be deduced practically.
685 If ``tx_db_nc`` is set to zero, the doorbell is forced to be mapped to regular
686 memory (with write combining), the PMD will perform the extra write memory barrier
687 after writing to doorbell, it might increase the needed CPU clocks per packet
688 to send, but latency might be improved.
690 If ``tx_db_nc`` is set to one, the doorbell is forced to be mapped to non
691 cached memory, the PMD will not perform the extra write memory barrier
692 after writing to doorbell, on some architectures it might improve the
695 If ``tx_db_nc`` is set to two, the doorbell is forced to be mapped to regular
696 memory, the PMD will use heuristics to decide whether write memory barrier
697 should be performed. For bursts with size multiple of recommended one (64 pkts)
698 it is supposed the next burst is coming and no need to issue the extra memory
699 barrier (it is supposed to be issued in the next coming burst, at least after
700 descriptor writing). It might increase latency (on some hosts till next
701 packets transmit) and should be used with care.
703 If ``tx_db_nc`` is omitted or set to zero, the preset (if any) environment
704 variable "MLX5_SHUT_UP_BF" value is used. If there is no "MLX5_SHUT_UP_BF",
705 the default ``tx_db_nc`` value is zero for ARM64 hosts and one for others.
707 - ``tx_pp`` parameter [int]
709 If a nonzero value is specified the driver creates all necessary internal
710 objects to provide accurate packet send scheduling on mbuf timestamps.
711 The positive value specifies the scheduling granularity in nanoseconds,
712 the packet send will be accurate up to specified digits. The allowed range is
713 from 500 to 1 million of nanoseconds. The negative value specifies the module
714 of granularity and engages the special test mode the check the schedule rate.
715 By default (if the ``tx_pp`` is not specified) send scheduling on timestamps
718 - ``tx_skew`` parameter [int]
720 The parameter adjusts the send packet scheduling on timestamps and represents
721 the average delay between beginning of the transmitting descriptor processing
722 by the hardware and appearance of actual packet data on the wire. The value
723 should be provided in nanoseconds and is valid only if ``tx_pp`` parameter is
724 specified. The default value is zero.
726 - ``tx_vec_en`` parameter [int]
728 A nonzero value enables Tx vector on ConnectX-5, ConnectX-6, ConnectX-6 Dx
729 and BlueField NICs if the number of global Tx queues on the port is less than
730 ``txqs_max_vec``. The parameter is deprecated and ignored.
732 - ``rx_vec_en`` parameter [int]
734 A nonzero value enables Rx vector if the port is not configured in
735 multi-segment otherwise this parameter is ignored.
739 - ``vf_nl_en`` parameter [int]
741 A nonzero value enables Netlink requests from the VF to add/remove MAC
742 addresses or/and enable/disable promiscuous/all multicast on the Netdevice.
743 Otherwise the relevant configuration must be run with Linux iproute2 tools.
744 This is a prerequisite to receive this kind of traffic.
746 Enabled by default, valid only on VF devices ignored otherwise.
748 - ``l3_vxlan_en`` parameter [int]
750 A nonzero value allows L3 VXLAN and VXLAN-GPE flow creation. To enable
751 L3 VXLAN or VXLAN-GPE, users has to configure firmware and enable this
752 parameter. This is a prerequisite to receive this kind of traffic.
756 - ``dv_xmeta_en`` parameter [int]
758 A nonzero value enables extensive flow metadata support if device is
759 capable and driver supports it. This can enable extensive support of
760 ``MARK`` and ``META`` item of ``rte_flow``. The newly introduced
761 ``SET_TAG`` and ``SET_META`` actions do not depend on ``dv_xmeta_en``.
763 There are some possible configurations, depending on parameter value:
765 - 0, this is default value, defines the legacy mode, the ``MARK`` and
766 ``META`` related actions and items operate only within NIC Tx and
767 NIC Rx steering domains, no ``MARK`` and ``META`` information crosses
768 the domain boundaries. The ``MARK`` item is 24 bits wide, the ``META``
769 item is 32 bits wide and match supported on egress only.
771 - 1, this engages extensive metadata mode, the ``MARK`` and ``META``
772 related actions and items operate within all supported steering domains,
773 including FDB, ``MARK`` and ``META`` information may cross the domain
774 boundaries. The ``MARK`` item is 24 bits wide, the ``META`` item width
775 depends on kernel and firmware configurations and might be 0, 16 or
776 32 bits. Within NIC Tx domain ``META`` data width is 32 bits for
777 compatibility, the actual width of data transferred to the FDB domain
778 depends on kernel configuration and may be vary. The actual supported
779 width can be retrieved in runtime by series of rte_flow_validate()
782 - 2, this engages extensive metadata mode, the ``MARK`` and ``META``
783 related actions and items operate within all supported steering domains,
784 including FDB, ``MARK`` and ``META`` information may cross the domain
785 boundaries. The ``META`` item is 32 bits wide, the ``MARK`` item width
786 depends on kernel and firmware configurations and might be 0, 16 or
787 24 bits. The actual supported width can be retrieved in runtime by
788 series of rte_flow_validate() trials.
790 +------+-----------+-----------+-------------+-------------+
791 | Mode | ``MARK`` | ``META`` | ``META`` Tx | FDB/Through |
792 +======+===========+===========+=============+=============+
793 | 0 | 24 bits | 32 bits | 32 bits | no |
794 +------+-----------+-----------+-------------+-------------+
795 | 1 | 24 bits | vary 0-32 | 32 bits | yes |
796 +------+-----------+-----------+-------------+-------------+
797 | 2 | vary 0-32 | 32 bits | 32 bits | yes |
798 +------+-----------+-----------+-------------+-------------+
800 If there is no E-Switch configuration the ``dv_xmeta_en`` parameter is
801 ignored and the device is configured to operate in legacy mode (0).
803 Disabled by default (set to 0).
805 The Direct Verbs/Rules (engaged with ``dv_flow_en`` = 1) supports all
806 of the extensive metadata features. The legacy Verbs supports FLAG and
807 MARK metadata actions over NIC Rx steering domain only.
809 - ``dv_flow_en`` parameter [int]
811 A nonzero value enables the DV flow steering assuming it is supported
812 by the driver (RDMA Core library version is rdma-core-24.0 or higher).
814 Enabled by default if supported.
816 - ``dv_esw_en`` parameter [int]
818 A nonzero value enables E-Switch using Direct Rules.
820 Enabled by default if supported.
822 - ``lacp_by_user`` parameter [int]
824 A nonzero value enables the control of LACP traffic by the user application.
825 When a bond exists in the driver, by default it should be managed by the
826 kernel and therefore LACP traffic should be steered to the kernel.
827 If this devarg is set to 1 it will allow the user to manage the bond by
828 itself and not steer LACP traffic to the kernel.
830 Disabled by default (set to 0).
832 - ``mr_ext_memseg_en`` parameter [int]
834 A nonzero value enables extending memseg when registering DMA memory. If
835 enabled, the number of entries in MR (Memory Region) lookup table on datapath
836 is minimized and it benefits performance. On the other hand, it worsens memory
837 utilization because registered memory is pinned by kernel driver. Even if a
838 page in the extended chunk is freed, that doesn't become reusable until the
839 entire memory is freed.
843 - ``representor`` parameter [list]
845 This parameter can be used to instantiate DPDK Ethernet devices from
846 existing port (or VF) representors configured on the device.
848 It is a standard parameter whose format is described in
849 :ref:`ethernet_device_standard_device_arguments`.
851 For instance, to probe port representors 0 through 2::
855 - ``max_dump_files_num`` parameter [int]
857 The maximum number of files per PMD entity that may be created for debug information.
858 The files will be created in /var/log directory or in current directory.
860 set to 128 by default.
862 - ``lro_timeout_usec`` parameter [int]
864 The maximum allowed duration of an LRO session, in micro-seconds.
865 PMD will set the nearest value supported by HW, which is not bigger than
866 the input ``lro_timeout_usec`` value.
867 If this parameter is not specified, by default PMD will set
868 the smallest value supported by HW.
870 - ``hp_buf_log_sz`` parameter [int]
872 The total data buffer size of a hairpin queue (logarithmic form), in bytes.
873 PMD will set the data buffer size to 2 ** ``hp_buf_log_sz``, both for RX & TX.
874 The capacity of the value is specified by the firmware and the initialization
875 will get a failure if it is out of scope.
876 The range of the value is from 11 to 19 right now, and the supported frame
877 size of a single packet for hairpin is from 512B to 128KB. It might change if
878 different firmware release is being used. By using a small value, it could
879 reduce memory consumption but not work with a large frame. If the value is
880 too large, the memory consumption will be high and some potential performance
881 degradation will be introduced.
882 By default, the PMD will set this value to 16, which means that 9KB jumbo
883 frames will be supported.
885 - ``reclaim_mem_mode`` parameter [int]
887 Cache some resources in flow destroy will help flow recreation more efficient.
888 While some systems may require the all the resources can be reclaimed after
890 The parameter ``reclaim_mem_mode`` provides the option for user to configure
891 if the resource cache is needed or not.
893 There are three options to choose:
895 - 0. It means the flow resources will be cached as usual. The resources will
896 be cached, helpful with flow insertion rate.
898 - 1. It will only enable the DPDK PMD level resources reclaim.
900 - 2. Both DPDK PMD level and rdma-core low level will be configured as
903 By default, the PMD will set this value to 0.
905 - ``sys_mem_en`` parameter [int]
907 A non-zero value enables the PMD memory management allocating memory
908 from system by default, without explicit rte memory flag.
910 By default, the PMD will set this value to 0.
912 - ``decap_en`` parameter [int]
914 Some devices do not support FCS (frame checksum) scattering for
915 tunnel-decapsulated packets.
916 If set to 0, this option forces the FCS feature and rejects tunnel
917 decapsulation in the flow engine for such devices.
919 By default, the PMD will set this value to 1.
921 .. _mlx5_firmware_config:
923 Firmware configuration
924 ~~~~~~~~~~~~~~~~~~~~~~
926 Firmware features can be configured as key/value pairs.
928 The command to set a value is::
930 mlxconfig -d <device> set <key>=<value>
932 The command to query a value is::
934 mlxconfig -d <device> query | grep <key>
936 The device name for the command ``mlxconfig`` can be either the PCI address,
937 or the mst device name found with::
941 Below are some firmware configurations listed.
947 value: 1=Infiniband 2=Ethernet 3=VPI(auto-sense)
953 - maximum number of SR-IOV virtual functions::
957 - enable DevX (required by Direct Rules and other features)::
961 - aggressive CQE zipping::
965 - L3 VXLAN and VXLAN-GPE destination UDP port::
968 IP_OVER_VXLAN_PORT=<udp dport>
970 - enable VXLAN-GPE tunnel flow matching::
972 FLEX_PARSER_PROFILE_ENABLE=0
974 FLEX_PARSER_PROFILE_ENABLE=2
976 - enable IP-in-IP tunnel flow matching::
978 FLEX_PARSER_PROFILE_ENABLE=0
980 - enable MPLS flow matching::
982 FLEX_PARSER_PROFILE_ENABLE=1
984 - enable ICMP(code/type/identifier/sequence number) / ICMP6(code/type) fields matching::
986 FLEX_PARSER_PROFILE_ENABLE=2
988 - enable Geneve flow matching::
990 FLEX_PARSER_PROFILE_ENABLE=0
992 FLEX_PARSER_PROFILE_ENABLE=1
994 - enable GTP flow matching::
996 FLEX_PARSER_PROFILE_ENABLE=3
998 - enable eCPRI flow matching::
1000 FLEX_PARSER_PROFILE_ENABLE=4
1006 This driver relies on external libraries and kernel drivers for resources
1007 allocations and initialization. The following dependencies are not part of
1008 DPDK and must be installed separately:
1012 User space Verbs framework used by librte_net_mlx5. This library provides
1013 a generic interface between the kernel and low-level user space drivers
1016 It allows slow and privileged operations (context initialization, hardware
1017 resources allocations) to be managed by the kernel and fast operations to
1018 never leave user space.
1022 Low-level user space driver library for Mellanox
1023 ConnectX-4/ConnectX-5/ConnectX-6/BlueField devices, it is automatically loaded
1026 This library basically implements send/receive calls to the hardware
1029 - **Kernel modules**
1031 They provide the kernel-side Verbs API and low level device drivers that
1032 manage actual hardware initialization and resources sharing with user
1035 Unlike most other PMDs, these modules must remain loaded and bound to
1038 - mlx5_core: hardware driver managing Mellanox
1039 ConnectX-4/ConnectX-5/ConnectX-6/BlueField devices and related Ethernet kernel
1041 - mlx5_ib: InifiniBand device driver.
1042 - ib_uverbs: user space driver for Verbs (entry point for libibverbs).
1044 - **Firmware update**
1046 Mellanox OFED/EN releases include firmware updates for
1047 ConnectX-4/ConnectX-5/ConnectX-6/BlueField adapters.
1049 Because each release provides new features, these updates must be applied to
1050 match the kernel modules and libraries they come with.
1054 Both libraries are BSD and GPL licensed. Linux kernel modules are GPL
1060 Either RDMA Core library with a recent enough Linux kernel release
1061 (recommended) or Mellanox OFED/EN, which provides compatibility with older
1064 RDMA Core with Linux Kernel
1065 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
1067 - Minimal kernel version : v4.14 or the most recent 4.14-rc (see `Linux installation documentation`_)
1068 - Minimal rdma-core version: v15+ commit 0c5f5765213a ("Merge pull request #227 from yishaih/tm")
1069 (see `RDMA Core installation documentation`_)
1070 - When building for i686 use:
1072 - rdma-core version 18.0 or above built with 32bit support.
1073 - Kernel version 4.14.41 or above.
1075 - Starting with rdma-core v21, static libraries can be built::
1078 CFLAGS=-fPIC cmake -DIN_PLACE=1 -DENABLE_STATIC=1 -GNinja ..
1081 .. _`Linux installation documentation`: https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git/plain/Documentation/admin-guide/README.rst
1082 .. _`RDMA Core installation documentation`: https://raw.githubusercontent.com/linux-rdma/rdma-core/master/README.md
1088 - Mellanox OFED version: **4.5** and above /
1089 Mellanox EN version: **4.5** and above
1092 - ConnectX-4: **12.21.1000** and above.
1093 - ConnectX-4 Lx: **14.21.1000** and above.
1094 - ConnectX-5: **16.21.1000** and above.
1095 - ConnectX-5 Ex: **16.21.1000** and above.
1096 - ConnectX-6: **20.27.0090** and above.
1097 - ConnectX-6 Dx: **22.27.0090** and above.
1098 - BlueField: **18.25.1010** and above.
1100 While these libraries and kernel modules are available on OpenFabrics
1101 Alliance's `website <https://www.openfabrics.org/>`__ and provided by package
1102 managers on most distributions, this PMD requires Ethernet extensions that
1103 may not be supported at the moment (this is a work in progress).
1106 <http://www.mellanox.com/page/products_dyn?product_family=26&mtag=linux>`__ and
1108 <http://www.mellanox.com/page/products_dyn?product_family=27&mtag=linux>`__
1109 include the necessary support and should be used in the meantime. For DPDK,
1110 only libibverbs, libmlx5, mlnx-ofed-kernel packages and firmware updates are
1111 required from that distribution.
1115 Several versions of Mellanox OFED/EN are available. Installing the version
1116 this DPDK release was developed and tested against is strongly
1117 recommended. Please check the `prerequisites`_.
1122 The following Mellanox device families are supported by the same mlx5 driver:
1132 Below are detailed device names:
1134 * Mellanox\ |reg| ConnectX\ |reg|-4 10G MCX4111A-XCAT (1x10G)
1135 * Mellanox\ |reg| ConnectX\ |reg|-4 10G MCX412A-XCAT (2x10G)
1136 * Mellanox\ |reg| ConnectX\ |reg|-4 25G MCX4111A-ACAT (1x25G)
1137 * Mellanox\ |reg| ConnectX\ |reg|-4 25G MCX412A-ACAT (2x25G)
1138 * Mellanox\ |reg| ConnectX\ |reg|-4 40G MCX413A-BCAT (1x40G)
1139 * Mellanox\ |reg| ConnectX\ |reg|-4 40G MCX4131A-BCAT (1x40G)
1140 * Mellanox\ |reg| ConnectX\ |reg|-4 40G MCX415A-BCAT (1x40G)
1141 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX413A-GCAT (1x50G)
1142 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX4131A-GCAT (1x50G)
1143 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX414A-BCAT (2x50G)
1144 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX415A-GCAT (1x50G)
1145 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX416A-BCAT (2x50G)
1146 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX416A-GCAT (2x50G)
1147 * Mellanox\ |reg| ConnectX\ |reg|-4 50G MCX415A-CCAT (1x100G)
1148 * Mellanox\ |reg| ConnectX\ |reg|-4 100G MCX416A-CCAT (2x100G)
1149 * Mellanox\ |reg| ConnectX\ |reg|-4 Lx 10G MCX4111A-XCAT (1x10G)
1150 * Mellanox\ |reg| ConnectX\ |reg|-4 Lx 10G MCX4121A-XCAT (2x10G)
1151 * Mellanox\ |reg| ConnectX\ |reg|-4 Lx 25G MCX4111A-ACAT (1x25G)
1152 * Mellanox\ |reg| ConnectX\ |reg|-4 Lx 25G MCX4121A-ACAT (2x25G)
1153 * Mellanox\ |reg| ConnectX\ |reg|-4 Lx 40G MCX4131A-BCAT (1x40G)
1154 * Mellanox\ |reg| ConnectX\ |reg|-5 100G MCX556A-ECAT (2x100G)
1155 * Mellanox\ |reg| ConnectX\ |reg|-5 Ex EN 100G MCX516A-CDAT (2x100G)
1156 * Mellanox\ |reg| ConnectX\ |reg|-6 200G MCX654106A-HCAT (2x200G)
1157 * Mellanox\ |reg| ConnectX\ |reg|-6 Dx EN 100G MCX623106AN-CDAT (2x100G)
1158 * Mellanox\ |reg| ConnectX\ |reg|-6 Dx EN 200G MCX623105AN-VDAT (1x200G)
1160 Quick Start Guide on OFED/EN
1161 ----------------------------
1163 1. Download latest Mellanox OFED/EN. For more info check the `prerequisites`_.
1166 2. Install the required libraries and kernel modules either by installing
1167 only the required set, or by installing the entire Mellanox OFED/EN::
1169 ./mlnxofedinstall --upstream-libs --dpdk
1171 3. Verify the firmware is the correct one::
1175 4. Verify all ports links are set to Ethernet::
1177 mlxconfig -d <mst device> query | grep LINK_TYPE
1181 Link types may have to be configured to Ethernet::
1183 mlxconfig -d <mst device> set LINK_TYPE_P1/2=1/2/3
1185 * LINK_TYPE_P1=<1|2|3> , 1=Infiniband 2=Ethernet 3=VPI(auto-sense)
1187 For hypervisors, verify SR-IOV is enabled on the NIC::
1189 mlxconfig -d <mst device> query | grep SRIOV_EN
1192 If needed, configure SR-IOV::
1194 mlxconfig -d <mst device> set SRIOV_EN=1 NUM_OF_VFS=16
1195 mlxfwreset -d <mst device> reset
1197 5. Restart the driver::
1199 /etc/init.d/openibd restart
1203 service openibd restart
1205 If link type was changed, firmware must be reset as well::
1207 mlxfwreset -d <mst device> reset
1209 For hypervisors, after reset write the sysfs number of virtual functions
1212 To dynamically instantiate a given number of virtual functions (VFs)::
1214 echo [num_vfs] > /sys/class/infiniband/mlx5_0/device/sriov_numvfs
1216 6. Install DPDK and you are ready to go.
1217 See :doc:`compilation instructions <../linux_gsg/build_dpdk>`.
1219 Enable switchdev mode
1220 ---------------------
1222 Switchdev mode is a mode in E-Switch, that binds between representor and VF.
1223 Representor is a port in DPDK that is connected to a VF in such a way
1224 that assuming there are no offload flows, each packet that is sent from the VF
1225 will be received by the corresponding representor. While each packet that is
1226 sent to a representor will be received by the VF.
1227 This is very useful in case of SRIOV mode, where the first packet that is sent
1228 by the VF will be received by the DPDK application which will decide if this
1229 flow should be offloaded to the E-Switch. After offloading the flow packet
1230 that the VF that are matching the flow will not be received any more by
1231 the DPDK application.
1233 1. Enable SRIOV mode::
1235 mlxconfig -d <mst device> set SRIOV_EN=true
1237 2. Configure the max number of VFs::
1239 mlxconfig -d <mst device> set NUM_OF_VFS=<num of vfs>
1243 mlxfwreset -d <mst device> reset
1245 3. Configure the actual number of VFs::
1247 echo <num of vfs > /sys/class/net/<net device>/device/sriov_numvfs
1249 4. Unbind the device (can be rebind after the switchdev mode)::
1251 echo -n "<device pci address" > /sys/bus/pci/drivers/mlx5_core/unbind
1253 5. Enbale switchdev mode::
1255 echo switchdev > /sys/class/net/<net device>/compat/devlink/mode
1260 1. Configure aggressive CQE Zipping for maximum performance::
1262 mlxconfig -d <mst device> s CQE_COMPRESSION=1
1264 To set it back to the default CQE Zipping mode use::
1266 mlxconfig -d <mst device> s CQE_COMPRESSION=0
1268 2. In case of virtualization:
1270 - Make sure that hypervisor kernel is 3.16 or newer.
1271 - Configure boot with ``iommu=pt``.
1272 - Use 1G huge pages.
1273 - Make sure to allocate a VM on huge pages.
1274 - Make sure to set CPU pinning.
1276 3. Use the CPU near local NUMA node to which the PCIe adapter is connected,
1277 for better performance. For VMs, verify that the right CPU
1278 and NUMA node are pinned according to the above. Run::
1282 to identify the NUMA node to which the PCIe adapter is connected.
1284 4. If more than one adapter is used, and root complex capabilities allow
1285 to put both adapters on the same NUMA node without PCI bandwidth degradation,
1286 it is recommended to locate both adapters on the same NUMA node.
1287 This in order to forward packets from one to the other without
1288 NUMA performance penalty.
1290 5. Disable pause frames::
1292 ethtool -A <netdev> rx off tx off
1294 6. Verify IO non-posted prefetch is disabled by default. This can be checked
1295 via the BIOS configuration. Please contact you server provider for more
1296 information about the settings.
1300 On some machines, depends on the machine integrator, it is beneficial
1301 to set the PCI max read request parameter to 1K. This can be
1302 done in the following way:
1304 To query the read request size use::
1306 setpci -s <NIC PCI address> 68.w
1308 If the output is different than 3XXX, set it by::
1310 setpci -s <NIC PCI address> 68.w=3XXX
1312 The XXX can be different on different systems. Make sure to configure
1313 according to the setpci output.
1315 7. To minimize overhead of searching Memory Regions:
1317 - '--socket-mem' is recommended to pin memory by predictable amount.
1318 - Configure per-lcore cache when creating Mempools for packet buffer.
1319 - Refrain from dynamically allocating/freeing memory in run-time.
1321 .. _mlx5_offloads_support:
1323 Supported hardware offloads
1324 ---------------------------
1326 .. table:: Minimal SW/HW versions for queue offloads
1328 ============== ===== ===== ========= ===== ========== ==========
1329 Offload DPDK Linux rdma-core OFED firmware hardware
1330 ============== ===== ===== ========= ===== ========== ==========
1331 common base 17.11 4.14 16 4.2-1 12.21.1000 ConnectX-4
1332 checksums 17.11 4.14 16 4.2-1 12.21.1000 ConnectX-4
1333 Rx timestamp 17.11 4.14 16 4.2-1 12.21.1000 ConnectX-4
1334 TSO 17.11 4.14 16 4.2-1 12.21.1000 ConnectX-4
1335 LRO 19.08 N/A N/A 4.6-4 16.25.6406 ConnectX-5
1336 ============== ===== ===== ========= ===== ========== ==========
1338 .. table:: Minimal SW/HW versions for rte_flow offloads
1340 +-----------------------+-----------------+-----------------+
1341 | Offload | with E-Switch | with NIC |
1342 +=======================+=================+=================+
1343 | Count | | DPDK 19.05 | | DPDK 19.02 |
1344 | | | OFED 4.6 | | OFED 4.6 |
1345 | | | rdma-core 24 | | rdma-core 23 |
1346 | | | ConnectX-5 | | ConnectX-5 |
1347 +-----------------------+-----------------+-----------------+
1348 | Drop | | DPDK 19.05 | | DPDK 18.11 |
1349 | | | OFED 4.6 | | OFED 4.5 |
1350 | | | rdma-core 24 | | rdma-core 23 |
1351 | | | ConnectX-5 | | ConnectX-4 |
1352 +-----------------------+-----------------+-----------------+
1353 | Queue / RSS | | | | DPDK 18.11 |
1354 | | | N/A | | OFED 4.5 |
1355 | | | | | rdma-core 23 |
1356 | | | | | ConnectX-4 |
1357 +-----------------------+-----------------+-----------------+
1358 | Encapsulation | | DPDK 19.05 | | DPDK 19.02 |
1359 | (VXLAN / NVGRE / RAW) | | OFED 4.7-1 | | OFED 4.6 |
1360 | | | rdma-core 24 | | rdma-core 23 |
1361 | | | ConnectX-5 | | ConnectX-5 |
1362 +-----------------------+-----------------+-----------------+
1363 | Encapsulation | | DPDK 19.11 | | DPDK 19.11 |
1364 | GENEVE | | OFED 4.7-3 | | OFED 4.7-3 |
1365 | | | rdma-core 27 | | rdma-core 27 |
1366 | | | ConnectX-5 | | ConnectX-5 |
1367 +-----------------------+-----------------+-----------------+
1368 | | Header rewrite | | DPDK 19.05 | | DPDK 19.02 |
1369 | | (set_ipv4_src / | | OFED 4.7-1 | | OFED 4.7-1 |
1370 | | set_ipv4_dst / | | rdma-core 24 | | rdma-core 24 |
1371 | | set_ipv6_src / | | ConnectX-5 | | ConnectX-5 |
1372 | | set_ipv6_dst / | | | | |
1373 | | set_tp_src / | | | | |
1374 | | set_tp_dst / | | | | |
1375 | | dec_ttl / | | | | |
1376 | | set_ttl / | | | | |
1377 | | set_mac_src / | | | | |
1378 | | set_mac_dst) | | | | |
1379 +-----------------------+-----------------+-----------------+
1380 | | Header rewrite | | DPDK 20.02 | | DPDK 20.02 |
1381 | | (set_dscp) | | OFED 5.0 | | OFED 5.0 |
1382 | | | | rdma-core 24 | | rdma-core 24 |
1383 | | | | ConnectX-5 | | ConnectX-5 |
1384 +-----------------------+-----------------+-----------------+
1385 | Jump | | DPDK 19.05 | | DPDK 19.02 |
1386 | | | OFED 4.7-1 | | OFED 4.7-1 |
1387 | | | rdma-core 24 | | N/A |
1388 | | | ConnectX-5 | | ConnectX-5 |
1389 +-----------------------+-----------------+-----------------+
1390 | Mark / Flag | | DPDK 19.05 | | DPDK 18.11 |
1391 | | | OFED 4.6 | | OFED 4.5 |
1392 | | | rdma-core 24 | | rdma-core 23 |
1393 | | | ConnectX-5 | | ConnectX-4 |
1394 +-----------------------+-----------------+-----------------+
1395 | Port ID | | DPDK 19.05 | | N/A |
1396 | | | OFED 4.7-1 | | N/A |
1397 | | | rdma-core 24 | | N/A |
1398 | | | ConnectX-5 | | N/A |
1399 +-----------------------+-----------------+-----------------+
1400 | | VLAN | | DPDK 19.11 | | DPDK 19.11 |
1401 | | (of_pop_vlan / | | OFED 4.7-1 | | OFED 4.7-1 |
1402 | | of_push_vlan / | | ConnectX-5 | | ConnectX-5 |
1403 | | of_set_vlan_pcp / | | | | |
1404 | | of_set_vlan_vid) | | | | |
1405 +-----------------------+-----------------+-----------------+
1406 | Hairpin | | | | DPDK 19.11 |
1407 | | | N/A | | OFED 4.7-3 |
1408 | | | | | rdma-core 26 |
1409 | | | | | ConnectX-5 |
1410 +-----------------------+-----------------+-----------------+
1411 | Meta data | | DPDK 19.11 | | DPDK 19.11 |
1412 | | | OFED 4.7-3 | | OFED 4.7-3 |
1413 | | | rdma-core 26 | | rdma-core 26 |
1414 | | | ConnectX-5 | | ConnectX-5 |
1415 +-----------------------+-----------------+-----------------+
1416 | Metering | | DPDK 19.11 | | DPDK 19.11 |
1417 | | | OFED 4.7-3 | | OFED 4.7-3 |
1418 | | | rdma-core 26 | | rdma-core 26 |
1419 | | | ConnectX-5 | | ConnectX-5 |
1420 +-----------------------+-----------------+-----------------+
1421 | Sampling | | DPDK 20.11 | | DPDK 20.11 |
1422 | | | OFED 5.2 | | OFED 5.2 |
1423 | | | rdma-core 32 | | rdma-core 32 |
1424 | | | ConnectX-5 | | ConnectX-5 |
1425 +-----------------------+-----------------+-----------------+
1430 MARK and META items are interrelated with datapath - they might move from/to
1431 the applications in mbuf fields. Hence, zero value for these items has the
1432 special meaning - it means "no metadata are provided", not zero values are
1433 treated by applications and PMD as valid ones.
1435 Moreover in the flow engine domain the value zero is acceptable to match and
1436 set, and we should allow to specify zero values as rte_flow parameters for the
1437 META and MARK items and actions. In the same time zero mask has no meaning and
1438 should be rejected on validation stage.
1443 Flows are not cached in the driver.
1444 When stopping a device port, all the flows created on this port from the
1445 application will be flushed automatically in the background.
1446 After stopping the device port, all flows on this port become invalid and
1447 not represented in the system.
1448 All references to these flows held by the application should be discarded
1449 directly but neither destroyed nor flushed.
1451 The application should re-create the flows as required after the port restart.
1456 Compared to librte_net_mlx4 that implements a single RSS configuration per
1457 port, librte_net_mlx5 supports per-protocol RSS configuration.
1459 Since ``testpmd`` defaults to IP RSS mode and there is currently no
1460 command-line parameter to enable additional protocols (UDP and TCP as well
1461 as IP), the following commands must be entered from its CLI to get the same
1462 behavior as librte_net_mlx4::
1465 > port config all rss all
1471 This section demonstrates how to launch **testpmd** with Mellanox
1472 ConnectX-4/ConnectX-5/ConnectX-6/BlueField devices managed by librte_net_mlx5.
1474 #. Load the kernel modules::
1476 modprobe -a ib_uverbs mlx5_core mlx5_ib
1478 Alternatively if MLNX_OFED/MLNX_EN is fully installed, the following script
1481 /etc/init.d/openibd restart
1485 User space I/O kernel modules (uio and igb_uio) are not used and do
1486 not have to be loaded.
1488 #. Make sure Ethernet interfaces are in working order and linked to kernel
1489 verbs. Related sysfs entries should be present::
1491 ls -d /sys/class/net/*/device/infiniband_verbs/uverbs* | cut -d / -f 5
1500 #. Optionally, retrieve their PCI bus addresses for whitelisting::
1503 for intf in eth2 eth3 eth4 eth5;
1505 (cd "/sys/class/net/${intf}/device/" && pwd -P);
1508 sed -n 's,.*/\(.*\),-w \1,p'
1517 #. Request huge pages::
1519 echo 1024 > /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages/nr_hugepages
1521 #. Start testpmd with basic parameters::
1523 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
1528 EAL: PCI device 0000:05:00.0 on NUMA socket 0
1529 EAL: probe driver: 15b3:1013 librte_net_mlx5
1530 PMD: librte_net_mlx5: PCI information matches, using device "mlx5_0" (VF: false)
1531 PMD: librte_net_mlx5: 1 port(s) detected
1532 PMD: librte_net_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:fe
1533 EAL: PCI device 0000:05:00.1 on NUMA socket 0
1534 EAL: probe driver: 15b3:1013 librte_net_mlx5
1535 PMD: librte_net_mlx5: PCI information matches, using device "mlx5_1" (VF: false)
1536 PMD: librte_net_mlx5: 1 port(s) detected
1537 PMD: librte_net_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:ff
1538 EAL: PCI device 0000:06:00.0 on NUMA socket 0
1539 EAL: probe driver: 15b3:1013 librte_net_mlx5
1540 PMD: librte_net_mlx5: PCI information matches, using device "mlx5_2" (VF: false)
1541 PMD: librte_net_mlx5: 1 port(s) detected
1542 PMD: librte_net_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:fa
1543 EAL: PCI device 0000:06:00.1 on NUMA socket 0
1544 EAL: probe driver: 15b3:1013 librte_net_mlx5
1545 PMD: librte_net_mlx5: PCI information matches, using device "mlx5_3" (VF: false)
1546 PMD: librte_net_mlx5: 1 port(s) detected
1547 PMD: librte_net_mlx5: port 1 MAC address is e4:1d:2d:e7:0c:fb
1548 Interactive-mode selected
1549 Configuring Port 0 (socket 0)
1550 PMD: librte_net_mlx5: 0x8cba80: TX queues number update: 0 -> 2
1551 PMD: librte_net_mlx5: 0x8cba80: RX queues number update: 0 -> 2
1552 Port 0: E4:1D:2D:E7:0C:FE
1553 Configuring Port 1 (socket 0)
1554 PMD: librte_net_mlx5: 0x8ccac8: TX queues number update: 0 -> 2
1555 PMD: librte_net_mlx5: 0x8ccac8: RX queues number update: 0 -> 2
1556 Port 1: E4:1D:2D:E7:0C:FF
1557 Configuring Port 2 (socket 0)
1558 PMD: librte_net_mlx5: 0x8cdb10: TX queues number update: 0 -> 2
1559 PMD: librte_net_mlx5: 0x8cdb10: RX queues number update: 0 -> 2
1560 Port 2: E4:1D:2D:E7:0C:FA
1561 Configuring Port 3 (socket 0)
1562 PMD: librte_net_mlx5: 0x8ceb58: TX queues number update: 0 -> 2
1563 PMD: librte_net_mlx5: 0x8ceb58: RX queues number update: 0 -> 2
1564 Port 3: E4:1D:2D:E7:0C:FB
1565 Checking link statuses...
1566 Port 0 Link Up - speed 40000 Mbps - full-duplex
1567 Port 1 Link Up - speed 40000 Mbps - full-duplex
1568 Port 2 Link Up - speed 10000 Mbps - full-duplex
1569 Port 3 Link Up - speed 10000 Mbps - full-duplex
1576 This section demonstrates how to dump flows. Currently, it's possible to dump
1577 all flows with assistance of external tools.
1579 #. 2 ways to get flow raw file:
1581 - Using testpmd CLI:
1583 .. code-block:: console
1585 testpmd> flow dump <port> <output_file>
1587 - call rte_flow_dev_dump api:
1589 .. code-block:: console
1591 rte_flow_dev_dump(port, file, NULL);
1593 #. Dump human-readable flows from raw file:
1595 Get flow parsing tool from: https://github.com/Mellanox/mlx_steering_dump
1597 .. code-block:: console
1599 mlx_steering_dump.py -f <output_file>