1 .. SPDX-License-Identifier: BSD-3-Clause
2 Copyright (c) 2017, Cisco Systems, Inc.
8 ENIC PMD is the DPDK poll-mode driver for the Cisco System Inc. VIC Ethernet
9 NICs. These adapters are also referred to as vNICs below. If you are running
10 or would like to run DPDK software applications on Cisco UCS servers using
11 Cisco VIC adapters the following documentation is relevant.
13 How to obtain ENIC PMD integrated DPDK
14 --------------------------------------
16 ENIC PMD support is integrated into the DPDK suite. dpdk-<version>.tar.gz
17 should be downloaded from http://core.dpdk.org/download/
20 Configuration information
21 -------------------------
23 - **DPDK Configuration Parameters**
25 The following configuration options are available for the ENIC PMD:
27 - **CONFIG_RTE_LIBRTE_ENIC_PMD** (default y): Enables or disables inclusion
28 of the ENIC PMD driver in the DPDK compilation.
30 - **vNIC Configuration Parameters**
32 - **Number of Queues**
34 The maximum number of receive queues (RQs), work queues (WQs) and
35 completion queues (CQs) are configurable on a per vNIC basis
36 through the Cisco UCS Manager (CIMC or UCSM).
38 These values should be configured as follows:
40 - The number of WQs should be greater or equal to the value of the
41 expected nb_tx_q parameter in the call to
42 rte_eth_dev_configure()
44 - The number of RQs configured in the vNIC should be greater or
45 equal to *twice* the value of the expected nb_rx_q parameter in
46 the call to rte_eth_dev_configure(). With the addition of Rx
47 scatter, a pair of RQs on the vnic is needed for each receive
48 queue used by DPDK, even if Rx scatter is not being used.
49 Having a vNIC with only 1 RQ is not a valid configuration, and
50 will fail with an error message.
52 - The number of CQs should set so that there is one CQ for each
53 WQ, and one CQ for each pair of RQs.
55 For example: If the application requires 3 Rx queues, and 3 Tx
56 queues, the vNIC should be configured to have at least 3 WQs, 6
57 RQs (3 pairs), and 6 CQs (3 for use by WQs + 3 for use by the 3
62 Likewise, the number of receive and transmit descriptors are configurable on
63 a per-vNIC basis via the UCS Manager and should be greater than or equal to
64 the nb_rx_desc and nb_tx_desc parameters expected to be used in the calls
65 to rte_eth_rx_queue_setup() and rte_eth_tx_queue_setup() respectively.
66 An application requesting more than the set size will be limited to that
69 Unless there is a lack of resources due to creating many vNICs, it
70 is recommended that the WQ and RQ sizes be set to the maximum. This
71 gives the application the greatest amount of flexibility in its
74 - *Note*: Since the introduction of Rx scatter, for performance
75 reasons, this PMD uses two RQs on the vNIC per receive queue in
76 DPDK. One RQ holds descriptors for the start of a packet, and the
77 second RQ holds the descriptors for the rest of the fragments of
78 a packet. This means that the nb_rx_desc parameter to
79 rte_eth_rx_queue_setup() can be a greater than 4096. The exact
80 amount will depend on the size of the mbufs being used for
81 receives, and the MTU size.
83 For example: If the mbuf size is 2048, and the MTU is 9000, then
84 receiving a full size packet will take 5 descriptors, 1 from the
85 start-of-packet queue, and 4 from the second queue. Assuming
86 that the RQ size was set to the maximum of 4096, then the
87 application can specify up to 1024 + 4096 as the nb_rx_desc
88 parameter to rte_eth_rx_queue_setup().
92 At least one interrupt per vNIC interface should be configured in the UCS
93 manager regardless of the number receive/transmit queues. The ENIC PMD
94 uses this interrupt to get information about link status and errors
97 In addition to the interrupt for link status and errors, when using Rx queue
98 interrupts, increase the number of configured interrupts so that there is at
99 least one interrupt for each Rx queue. For example, if the app uses 3 Rx
100 queues and wants to use per-queue interrupts, configure 4 (3 + 1) interrupts.
102 - **Receive Side Scaling**
104 In order to fully utilize RSS in DPDK, enable all RSS related settings in
105 CIMC or UCSM. These include the following items listed under
106 Receive Side Scaling:
107 TCP, IPv4, TCP-IPv4, IPv6, TCP-IPv6, IPv6 Extension, TCP-IPv6 Extension.
110 .. _enic-flow-director:
112 Flow director support
113 ---------------------
115 Advanced filtering support was added to 1300 series VIC firmware starting
116 with version 2.0.13 for C-series UCS servers and version 3.1.2 for UCSM
117 managed blade servers. In order to enable advanced filtering the 'Advanced
118 filter' radio button should be enabled via CIMC or UCSM followed by a reboot
121 With advanced filters, perfect matching of all fields of IPv4, IPv6 headers
122 as well as TCP, UDP and SCTP L4 headers is available through flow director.
123 Masking of these fields for partial match is also supported.
125 Without advanced filter support, the flow director is limited to IPv4
126 perfect filtering of the 5-tuple with no masking of fields supported.
128 SR-IOV mode utilization
129 -----------------------
131 UCS blade servers configured with dynamic vNIC connection policies in UCSM
132 are capable of supporting SR-IOV. SR-IOV virtual functions (VFs) are
133 specialized vNICs, distinct from regular Ethernet vNICs. These VFs can be
134 directly assigned to virtual machines (VMs) as 'passthrough' devices.
136 In UCS, SR-IOV VFs require the use of the Cisco Virtual Machine Fabric Extender
137 (VM-FEX), which gives the VM a dedicated
138 interface on the Fabric Interconnect (FI). Layer 2 switching is done at
139 the FI. This may eliminate the requirement for software switching on the
140 host to route intra-host VM traffic.
142 Please refer to `Creating a Dynamic vNIC Connection Policy
143 <http://www.cisco.com/c/en/us/td/docs/unified_computing/ucs/sw/vm_fex/vmware/gui/config_guide/b_GUI_VMware_VM-FEX_UCSM_Configuration_Guide/b_GUI_VMware_VM-FEX_UCSM_Configuration_Guide_chapter_010.html#task_433E01651F69464783A68E66DA8A47A5>`_
144 for information on configuring SR-IOV adapter policies and port profiles
147 Once the policies are in place and the host OS is rebooted, VFs should be
148 visible on the host, E.g.:
150 .. code-block:: console
152 # lspci | grep Cisco | grep Ethernet
153 0d:00.0 Ethernet controller: Cisco Systems Inc VIC Ethernet NIC (rev a2)
154 0d:00.1 Ethernet controller: Cisco Systems Inc VIC SR-IOV VF (rev a2)
155 0d:00.2 Ethernet controller: Cisco Systems Inc VIC SR-IOV VF (rev a2)
156 0d:00.3 Ethernet controller: Cisco Systems Inc VIC SR-IOV VF (rev a2)
157 0d:00.4 Ethernet controller: Cisco Systems Inc VIC SR-IOV VF (rev a2)
158 0d:00.5 Ethernet controller: Cisco Systems Inc VIC SR-IOV VF (rev a2)
159 0d:00.6 Ethernet controller: Cisco Systems Inc VIC SR-IOV VF (rev a2)
160 0d:00.7 Ethernet controller: Cisco Systems Inc VIC SR-IOV VF (rev a2)
162 Enable Intel IOMMU on the host and install KVM and libvirt, and reboot again as
163 required. Then, using libvirt, create a VM instance with an assigned device.
164 Below is an example ``interface`` block (part of the domain configuration XML)
165 that adds the host VF 0d:00:01 to the VM. ``profileid='pp-vlan-25'`` indicates
166 the port profile that has been configured in UCSM.
168 .. code-block:: console
170 <interface type='hostdev' managed='yes'>
171 <mac address='52:54:00:ac:ff:b6'/>
172 <driver name='vfio'/>
174 <address type='pci' domain='0x0000' bus='0x0d' slot='0x00' function='0x1'/>
176 <virtualport type='802.1Qbh'>
177 <parameters profileid='pp-vlan-25'/>
182 Alternatively, the configuration can be done in a separate file using the
183 ``network`` keyword. These methods are described in the libvirt documentation for
184 `Network XML format <https://libvirt.org/formatnetwork.html>`_.
186 When the VM instance is started, libvirt will bind the host VF to
187 vfio, complete provisioning on the FI and bring up the link.
191 It is not possible to use a VF directly from the host because it is not
192 fully provisioned until libvirt brings up the VM that it is assigned
195 In the VM instance, the VF will now be visible. E.g., here the VF 00:04.0 is
196 seen on the VM instance and should be available for binding to a DPDK.
198 .. code-block:: console
201 00:04.0 Ethernet controller: Cisco Systems Inc VIC SR-IOV VF (rev a2)
203 Follow the normal DPDK install procedure, binding the VF to either ``igb_uio``
204 or ``vfio`` in non-IOMMU mode.
206 In the VM, the kernel enic driver may be automatically bound to the VF during
207 boot. Unbinding it currently hangs due to a known issue with the driver. To
208 work around the issue, blacklist the enic module as follows.
209 Please see :ref:`Limitations <enic_limitations>` for limitations in
212 .. code-block:: console
214 # cat /etc/modprobe.d/enic.conf
221 Passthrough does not require SR-IOV. If VM-FEX is not desired, the user
222 may create as many regular vNICs as necessary and assign them to VMs as
223 passthrough devices. Since these vNICs are not SR-IOV VFs, using them as
224 passthrough devices do not require libvirt, port profiles, and VM-FEX.
227 .. _enic-generic-flow-api:
229 Generic Flow API support
230 ------------------------
232 Generic Flow API is supported. The baseline support is:
234 - **1200 series VICs**
236 5-tuple exact flow support for 1200 series adapters. This allows:
238 - Attributes: ingress
239 - Items: ipv4, ipv6, udp, tcp (must exactly match src/dst IP
240 addresses and ports and all must be specified)
241 - Actions: queue and void
244 - **1300 and later series VICS with advanced filters disabled**
246 With advanced filters disabled, an IPv4 or IPv6 item must be specified
249 - Attributes: ingress
250 - Items: eth, vlan, ipv4, ipv6, udp, tcp, vxlan, inner eth, vlan, ipv4, ipv6, udp, tcp
251 - Actions: queue and void
252 - Selectors: 'is', 'spec' and 'mask'. 'last' is not supported
253 - In total, up to 64 bytes of mask is allowed across all headers
255 - **1300 and later series VICS with advanced filters enabled**
257 - Attributes: ingress
258 - Items: eth, vlan, ipv4, ipv6, udp, tcp, vxlan, raw, inner eth, vlan, ipv4, ipv6, udp, tcp
259 - Actions: queue, mark, drop, flag, rss, passthru, and void
260 - Selectors: 'is', 'spec' and 'mask'. 'last' is not supported
261 - In total, up to 64 bytes of mask is allowed across all headers
263 The VIC performs packet matching after applying VLAN strip. If VLAN
264 stripping is enabled, EtherType in the ETH item corresponds to the
265 stripped VLAN header's EtherType. Stripping does not affect the VLAN
266 item. TCI and EtherType in the VLAN item are matched against those in
267 the (stripped) VLAN header whether stripping is enabled or disabled.
269 More features may be added in future firmware and new versions of the VIC.
270 Please refer to the release notes.
277 Recent hardware models support overlay offload. When enabled, the NIC performs
278 the following operations for VXLAN, NVGRE, and GENEVE packets. In all cases,
279 inner and outer packets can be IPv4 or IPv6.
281 - TSO for VXLAN and GENEVE packets.
283 Hardware supports NVGRE TSO, but DPDK currently has no NVGRE offload flags.
285 - Tx checksum offloads.
287 The NIC fills in IPv4/UDP/TCP checksums for both inner and outer packets.
289 - Rx checksum offloads.
291 The NIC validates IPv4/UDP/TCP checksums of both inner and outer packets.
292 Good checksum flags (e.g. ``PKT_RX_L4_CKSUM_GOOD``) indicate that the inner
293 packet has the correct checksum, and if applicable, the outer packet also
294 has the correct checksum. Bad checksum flags (e.g. ``PKT_RX_L4_CKSUM_BAD``)
295 indicate that the inner and/or outer packets have invalid checksum values.
297 - Inner Rx packet type classification
299 PMD sets inner L3/L4 packet types (e.g. ``RTE_PTYPE_INNER_L4_TCP``), and
300 ``RTE_PTYPE_TUNNEL_GRENAT`` to indicate that the packet is tunneled.
301 PMD does not set L3/L4 packet types for outer packets.
305 RSS hash calculation, therefore queue selection, is done on inner packets.
307 In order to enable overlay offload, the 'Enable VXLAN' box should be checked
308 via CIMC or UCSM followed by a reboot of the server. When PMD successfully
309 enables overlay offload, it prints the following message on the console.
311 .. code-block:: console
313 Overlay offload is enabled
315 By default, PMD enables overlay offload if hardware supports it. To disable
316 it, set ``devargs`` parameter ``disable-overlay=1``. For example::
318 -w 12:00.0,disable-overlay=1
320 By default, the NIC uses 4789 as the VXLAN port. The user may change
321 it through ``rte_eth_dev_udp_tunnel_port_{add,delete}``. However, as
322 the current NIC has a single VXLAN port number, the user cannot
323 configure multiple port numbers.
328 VIC adapters can tag, untag, or modify the VLAN headers of ingress
329 packets. The ingress VLAN rewrite mode controls this behavior. By
330 default, it is set to pass-through, where the NIC does not modify the
331 VLAN header in any way so that the application can see the original
332 header. This mode is sufficient for many applications, but may not be
333 suitable for others. Such applications may change the mode by setting
334 ``devargs`` parameter ``ig-vlan-rewrite`` to one of the following.
336 - ``pass``: Pass-through mode. The NIC does not modify the VLAN
337 header. This is the default mode.
339 - ``priority``: Priority-tag default VLAN mode. If the ingress packet
340 is tagged with the default VLAN, the NIC replaces its VLAN header
341 with the priority tag (VLAN ID 0).
343 - ``trunk``: Default trunk mode. The NIC tags untagged ingress packets
344 with the default VLAN. Tagged ingress packets are not modified. To
345 the application, every packet appears as tagged.
347 - ``untag``: Untag default VLAN mode. If the ingress packet is tagged
348 with the default VLAN, the NIC removes or untags its VLAN header so
349 that the application sees an untagged packet. As a result, the
350 default VLAN becomes `untagged`. This mode can be useful for
351 applications such as OVS-DPDK performance benchmarks that utilize
352 only the default VLAN and want to see only untagged packets.
355 Vectorized Rx Handler
356 ---------------------
358 ENIC PMD includes a version of the receive handler that is vectorized using
359 AVX2 SIMD instructions. It is meant for bulk, throughput oriented workloads
360 where reducing cycles/packet in PMD is a priority. In order to use the
361 vectorized handler, take the following steps.
363 - Use a recent version of gcc, icc, or clang and build 64-bit DPDK. If
364 the compiler is known to support AVX2, DPDK build system
365 automatically compiles the vectorized handler. Otherwise, the
366 handler is not available.
368 - Set ``devargs`` parameter ``enable-avx2-rx=1`` to explicitly request that
369 PMD consider the vectorized handler when selecting the receive handler.
372 -w 12:00.0,enable-avx2-rx=1
374 As the current implementation is intended for field trials, by default, the
375 vectorized handler is not considered (``enable-avx2-rx=0``).
377 - Run on a UCS M4 or later server with CPUs that support AVX2.
379 PMD selects the vectorized handler when the handler is compiled into
380 the driver, the user requests its use via ``enable-avx2-rx=1``, CPU
381 supports AVX2, and scatter Rx is not used. To verify that the
382 vectorized handler is selected, enable debug logging
383 (``--log-level=pmd,debug``) and check the following message.
385 .. code-block:: console
387 enic_use_vector_rx_handler use the non-scatter avx2 Rx handler
389 .. _enic_limitations:
394 - **VLAN 0 Priority Tagging**
396 If a vNIC is configured in TRUNK mode by the UCS manager, the adapter will
397 priority tag egress packets according to 802.1Q if they were not already
398 VLAN tagged by software. If the adapter is connected to a properly configured
399 switch, there will be no unexpected behavior.
401 In test setups where an Ethernet port of a Cisco adapter in TRUNK mode is
402 connected point-to-point to another adapter port or connected though a router
403 instead of a switch, all ingress packets will be VLAN tagged. Programs such
404 as l3fwd may not account for VLAN tags in packets and may misbehave. One
405 solution is to enable VLAN stripping on ingress so the VLAN tag is removed
406 from the packet and put into the mbuf->vlan_tci field. Here is an example
407 of how to accomplish this:
409 .. code-block:: console
411 vlan_offload = rte_eth_dev_get_vlan_offload(port);
412 vlan_offload |= ETH_VLAN_STRIP_OFFLOAD;
413 rte_eth_dev_set_vlan_offload(port, vlan_offload);
415 Another alternative is modify the adapter's ingress VLAN rewrite mode so that
416 packets with the default VLAN tag are stripped by the adapter and presented to
417 DPDK as untagged packets. In this case mbuf->vlan_tci and the PKT_RX_VLAN and
418 PKT_RX_VLAN_STRIPPED mbuf flags would not be set. This mode is enabled with the
419 ``devargs`` parameter ``ig-vlan-rewrite=untag``. For example::
421 -w 12:00.0,ig-vlan-rewrite=untag
423 - Limited flow director support on 1200 series and 1300 series Cisco VIC
424 adapters with old firmware. Please see :ref:`enic-flow-director`.
426 - Flow director features are not supported on generation 1 Cisco VIC adapters
431 - KVM hypervisor support only. VMware has not been tested.
432 - Requires VM-FEX, and so is only available on UCS managed servers connected
433 to Fabric Interconnects. It is not on standalone C-Series servers.
434 - VF devices are not usable directly from the host. They can only be used
435 as assigned devices on VM instances.
436 - Currently, unbind of the ENIC kernel mode driver 'enic.ko' on the VM
437 instance may hang. As a workaround, enic.ko should be blacklisted or removed
438 from the boot process.
439 - pci_generic cannot be used as the uio module in the VM. igb_uio or
440 vfio in non-IOMMU mode can be used.
441 - The number of RQs in UCSM dynamic vNIC configurations must be at least 2.
442 - The number of SR-IOV devices is limited to 256. Components on target system
443 might limit this number to fewer than 256.
447 - The number of filters that can be specified with the Generic Flow API is
448 dependent on how many header fields are being masked. Use 'flow create' in
449 a loop to determine how many filters your VIC will support (not more than
450 1000 for 1300 series VICs). Filters are checked for matching in the order they
451 were added. Since there currently is no grouping or priority support,
452 'catch-all' filters should be added last.
453 - The supported range of IDs for the 'MARK' action is 0 - 0xFFFD.
454 - RSS and PASSTHRU actions only support "receive normally". They are limited
455 to supporting MARK + RSS and PASSTHRU + MARK to allow the application to mark
456 packets and then receive them normally. These require 1400 series VIC adapters
458 - RAW items are limited to matching UDP tunnel headers like VXLAN.
462 - ``rx_good_bytes`` (ibytes) always includes VLAN header (4B) and CRC bytes (4B).
463 This behavior applies to 1300 and older series VIC adapters.
464 1400 series VICs do not count CRC bytes, and count VLAN header only when VLAN
465 stripping is disabled.
466 - When the NIC drops a packet because the Rx queue has no free buffers,
467 ``rx_good_bytes`` still increments by 4B if the packet is not VLAN tagged or
468 VLAN stripping is disabled, or by 8B if the packet is VLAN tagged and stripping
470 This behavior applies to 1300 and older series VIC adapters. 1400 series VICs
471 do not increment this byte counter when packets are dropped.
475 - Hardware enables and disables UDP and TCP RSS hashing together. The driver
476 cannot control UDP and TCP hashing individually.
478 How to build the suite
479 ----------------------
481 The build instructions for the DPDK suite should be followed. By default
482 the ENIC PMD library will be built into the DPDK library.
484 Refer to the document :ref:`compiling and testing a PMD for a NIC
485 <pmd_build_and_test>` for details.
487 For configuring and using UIO and VFIO frameworks, please refer to the
488 documentation that comes with DPDK suite.
490 Supported Cisco VIC adapters
491 ----------------------------
493 ENIC PMD supports all recent generations of Cisco VIC adapters including:
499 Supported Operating Systems
500 ---------------------------
502 Any Linux distribution fulfilling the conditions described in Dependencies
503 section of DPDK documentation.
508 - Unicast, multicast and broadcast transmission and reception
509 - Receive queue polling
510 - Port Hardware Statistics
511 - Hardware VLAN acceleration
512 - IP checksum offload
513 - Receive side VLAN stripping
514 - Multiple receive and transmit queues
515 - Flow Director ADD, UPDATE, DELETE, STATS operation support IPv4 and IPv6
517 - Setting RX VLAN (supported via UCSM/CIMC only)
518 - VLAN filtering (supported via UCSM/CIMC only)
519 - Execution of application by unprivileged system users
520 - IPV4, IPV6 and TCP RSS hashing
521 - UDP RSS hashing (1400 series and later adapters)
524 - SR-IOV on UCS managed servers connected to Fabric Interconnects
528 - Rx/Tx checksum offloads for VXLAN, NVGRE, GENEVE
529 - TSO for VXLAN and GENEVE packets
532 Known bugs and unsupported features in this release
533 ---------------------------------------------------
535 - Signature or flex byte based flow direction
536 - Drop feature of flow direction
537 - VLAN based flow direction
538 - Non-IPV4 flow direction
539 - Setting of extended VLAN
540 - MTU update only works if Scattered Rx mode is disabled
541 - Maximum receive packet length is ignored if Scattered Rx mode is used
546 - Prepare the system as recommended by DPDK suite. This includes environment
547 variables, hugepages configuration, tool-chains and configuration.
548 - Insert vfio-pci kernel module using the command 'modprobe vfio-pci' if the
549 user wants to use VFIO framework.
550 - Insert uio kernel module using the command 'modprobe uio' if the user wants
551 to use UIO framework.
552 - DPDK suite should be configured based on the user's decision to use VFIO or
554 - If the vNIC device(s) to be used is bound to the kernel mode Ethernet driver
555 use 'ip' to bring the interface down. The dpdk-devbind.py tool can
556 then be used to unbind the device's bus id from the ENIC kernel mode driver.
557 - Bind the intended vNIC to vfio-pci in case the user wants ENIC PMD to use
558 VFIO framework using dpdk-devbind.py.
559 - Bind the intended vNIC to igb_uio in case the user wants ENIC PMD to use
560 UIO framework using dpdk-devbind.py.
562 At this point the system should be ready to run DPDK applications. Once the
563 application runs to completion, the vNIC can be detached from vfio-pci or
564 igb_uio if necessary.
566 Root privilege is required to bind and unbind vNICs to/from VFIO/UIO.
567 VFIO framework helps an unprivileged user to run the applications.
568 For an unprivileged user to run the applications on DPDK and ENIC PMD,
569 it may be necessary to increase the maximum locked memory of the user.
570 The following command could be used to do this.
572 .. code-block:: console
574 sudo sh -c "ulimit -l <value in Kilo Bytes>"
576 The value depends on the memory configuration of the application, DPDK and
577 PMD. Typically, the limit has to be raised to higher than 2GB.
580 The compilation of any unused drivers can be disabled using the
581 configuration file in config/ directory (e.g., config/common_linux).
582 This would help in bringing down the time taken for building the
583 libraries and the initialization time of the application.
588 - https://www.cisco.com/c/en/us/products/servers-unified-computing/index.html
589 - https://www.cisco.com/c/en/us/products/interfaces-modules/unified-computing-system-adapters/index.html
594 Any questions or bugs should be reported to DPDK community and to the ENIC PMD
597 - John Daley <johndale@cisco.com>
598 - Hyong Youb Kim <hyonkim@cisco.com>