8. Known Issues and Limitations in Legacy Releases

This section describes known issues with the DPDK software that aren’t covered in the version specific release notes sections.

8.2. Pause Frame Forwarding does not work properly on igb

Description:
For igb devices rte_eth_flow_ctrl_set does not work as expected. Pause frames are always forwarded on igb, regardless of the RFCE, MPMCF and DPF registers.
Implication:
Pause frames will never be rejected by the host on 1G NICs and they will always be forwarded.
Resolution/Workaround:
There is no workaround available.
Affected Environment/Platform:
All.
Driver/Module:
Poll Mode Driver (PMD).

8.3. In packets provided by the PMD, some flags are missing

Description:
In packets provided by the PMD, some flags are missing. The application does not have access to information provided by the hardware (packet is broadcast, packet is multicast, packet is IPv4 and so on).
Implication:
The ol_flags field in the rte_mbuf structure is not correct and should not be used.
Resolution/Workaround:
The application has to parse the Ethernet header itself to get the information, which is slower.
Affected Environment/Platform:
All.
Driver/Module:
Poll Mode Driver (PMD).

8.4. The rte_malloc library is not fully implemented

Description:
The rte_malloc library is not fully implemented.
Implication:
All debugging features of rte_malloc library described in architecture documentation are not yet implemented.
Resolution/Workaround:
No workaround available.
Affected Environment/Platform:
All.
Driver/Module:
rte_malloc.

8.5. HPET reading is slow

Description:
Reading the HPET chip is slow.
Implication:
An application that calls rte_get_hpet_cycles() or rte_timer_manage() runs slower.
Resolution/Workaround:
The application should not call these functions too often in the main loop. An alternative is to use the TSC register through rte_rdtsc() which is faster, but specific to an lcore and is a cycle reference, not a time reference.
Affected Environment/Platform:
All.
Driver/Module:
Environment Abstraction Layer (EAL).

8.6. HPET timers do not work on the Osage customer reference platform

Description:
HPET timers do not work on the Osage customer reference platform which includes an Intel® Xeon® processor 5500 series processor) using the released BIOS from Intel.
Implication:
On Osage boards, the implementation of the rte_delay_us() function must be changed to not use the HPET timer.
Resolution/Workaround:
This can be addressed by building the system with the CONFIG_RTE_LIBEAL_USE_HPET=n configuration option or by using the --no-hpet EAL option.
Affected Environment/Platform:
The Osage customer reference platform. Other vendor platforms with Intel® Xeon® processor 5500 series processors should work correctly, provided the BIOS supports HPET.
Driver/Module:
lib/librte_eal/common/include/rte_cycles.h

8.7. Not all variants of supported NIC types have been used in testing

Description:

The supported network interface cards can come in a number of variants with different device ID’s. Not all of these variants have been tested with the DPDK.

The NIC device identifiers used during testing:

  • Intel® Ethernet Controller XL710 for 40GbE QSFP+ [8086:1584]
  • Intel® Ethernet Controller XL710 for 40GbE QSFP+ [8086:1583]
  • Intel® Ethernet Controller X710 for 10GbE SFP+ [8086:1572]
  • Intel® 82576 Gigabit Ethernet Controller [8086:10c9]
  • Intel® 82576 Quad Copper Gigabit Ethernet Controller [8086:10e8]
  • Intel® 82580 Dual Copper Gigabit Ethernet Controller [8086:150e]
  • Intel® I350 Quad Copper Gigabit Ethernet Controller [8086:1521]
  • Intel® 82599 Dual Fibre 10 Gigabit Ethernet Controller [8086:10fb]
  • Intel® Ethernet Server Adapter X520-T2 [8086: 151c]
  • Intel® Ethernet Controller X540-T2 [8086:1528]
  • Intel® 82574L Gigabit Network Connection [8086:10d3]
  • Emulated Intel® 82540EM Gigabit Ethernet Controller [8086:100e]
  • Emulated Intel® 82545EM Gigabit Ethernet Controller [8086:100f]
  • Intel® Ethernet Server Adapter X520-4 [8086:154a]
  • Intel® Ethernet Controller I210 [8086:1533]
Implication:
Risk of issues with untested variants.
Resolution/Workaround:
Use tested NIC variants. For those supported Ethernet controllers, additional device IDs may be added to the software if required.
Affected Environment/Platform:
All.
Driver/Module:
Poll-mode drivers

8.8. Multi-process sample app requires exact memory mapping

Description:
The multi-process example application assumes that it is possible to map the hugepage memory to the same virtual addresses in client and server applications. Occasionally, very rarely with 64-bit, this does not occur and a client application will fail on startup. The Linux “address-space layout randomization” security feature can sometimes cause this to occur.
Implication:
A multi-process client application fails to initialize.
Resolution/Workaround:
See the “Multi-process Limitations” section in the DPDK Programmer’s Guide for more information.
Affected Environment/Platform:
All.
Driver/Module:
Multi-process example application

8.9. Packets are not sent by the 1 GbE/10 GbE SR-IOV driver when the source MAC is not the MAC assigned to the VF NIC

Description:
The 1 GbE/10 GbE SR-IOV driver can only send packets when the Ethernet header’s source MAC address is the same as that of the VF NIC. The reason for this is that the Linux ixgbe driver module in the host OS has its anti-spoofing feature enabled.
Implication:
Packets sent using the 1 GbE/10 GbE SR-IOV driver must have the source MAC address correctly set to that of the VF NIC. Packets with other source address values are dropped by the NIC if the application attempts to transmit them.
Resolution/Workaround:
Configure the Ethernet source address in each packet to match that of the VF NIC.
Affected Environment/Platform:
All.
Driver/Module:
1 GbE/10 GbE VF Poll Mode Driver (PMD).

8.10. SR-IOV drivers do not fully implement the rte_ethdev API

Description:

The SR-IOV drivers only supports the following rte_ethdev API functions:

  • rte_eth_dev_configure()
  • rte_eth_tx_queue_setup()
  • rte_eth_rx_queue_setup()
  • rte_eth_dev_info_get()
  • rte_eth_dev_start()
  • rte_eth_tx_burst()
  • rte_eth_rx_burst()
  • rte_eth_dev_stop()
  • rte_eth_stats_get()
  • rte_eth_stats_reset()
  • rte_eth_link_get()
  • rte_eth_link_get_no_wait()
Implication:
Calling an unsupported function will result in an application error.
Resolution/Workaround:
Do not use other rte_ethdev API functions in applications that use the SR-IOV drivers.
Affected Environment/Platform:
All.
Driver/Module:
VF Poll Mode Driver (PMD).

8.11. PMD does not work with –no-huge EAL command line parameter

Description:
Currently, the DPDK does not store any information about memory allocated by malloc()` (for example, NUMA node, physical address), hence PMD drivers do not work when the ``--no-huge command line parameter is supplied to EAL.
Implication:
Sending and receiving data with PMD will not work.
Resolution/Workaround:
Use huge page memory or use VFIO to map devices.
Affected Environment/Platform:
Systems running the DPDK on Linux
Driver/Module:
Poll Mode Driver (PMD).

8.12. Some hardware off-load functions are not supported by the VF Driver

Description:

Currently, configuration of the following items is not supported by the VF driver:

  • IP/UDP/TCP checksum offload
  • Jumbo Frame Receipt
  • HW Strip CRC
Implication:
Any configuration for these items in the VF register will be ignored. The behavior is dependent on the current PF setting.
Resolution/Workaround:
For the PF (Physical Function) status on which the VF driver depends, there is an option item under PMD in the config file. For others, the VF will keep the same behavior as PF setting.
Affected Environment/Platform:
All.
Driver/Module:
VF (SR-IOV) Poll Mode Driver (PMD).

8.13. Kernel crash on IGB port unbinding

Description:
Kernel crash may occur when unbinding 1G ports from the igb_uio driver, on 2.6.3x kernels such as shipped with Fedora 14.
Implication:
Kernel crash occurs.
Resolution/Workaround:
Use newer kernels or do not unbind ports.
Affected Environment/Platform:
2.6.3x kernels such as shipped with Fedora 14
Driver/Module:
IGB Poll Mode Driver (PMD).

8.15. Discrepancies between statistics reported by different NICs

Description:
Gigabit Ethernet devices from Intel include CRC bytes when calculating packet reception statistics regardless of hardware CRC stripping state, while 10-Gigabit Ethernet devices from Intel do so only when hardware CRC stripping is disabled.
Implication:
There may be a discrepancy in how different NICs display packet reception statistics.
Resolution/Workaround:
None
Affected Environment/Platform:
All.
Driver/Module:
Poll Mode Driver (PMD).

8.16. Error reported opening files on DPDK initialization

Description:
On DPDK application startup, errors may be reported when opening files as part of the initialization process. This occurs if a large number, for example, 500 or more, or if hugepages are used, due to the per-process limit on the number of open files.
Implication:
The DPDK application may fail to run.
Resolution/Workaround:
If using 2 MB hugepages, consider switching to a fewer number of 1 GB pages. Alternatively, use the ulimit command to increase the number of files which can be opened by a process.
Affected Environment/Platform:
All.
Driver/Module:
Environment Abstraction Layer (EAL).

8.17. Intel® QuickAssist Technology sample application does not work on a 32-bit OS on Shumway

Description:
The Intel® Communications Chipset 89xx Series device does not fully support NUMA on a 32-bit OS. Consequently, the sample application cannot work properly on Shumway, since it requires NUMA on both nodes.
Implication:
The sample application cannot work in 32-bit mode with emulated NUMA, on multi-socket boards.
Resolution/Workaround:
There is no workaround available.
Affected Environment/Platform:
Shumway
Driver/Module:
All.

8.18. Differences in how different Intel NICs handle maximum packet length for jumbo frame

Description:
10 Gigabit Ethernet devices from Intel do not take VLAN tags into account when calculating packet size while Gigabit Ethernet devices do so for jumbo frames.
Implication:
When receiving packets with VLAN tags, the actual maximum size of useful payload that Intel Gigabit Ethernet devices are able to receive is 4 bytes (or 8 bytes in the case of packets with extended VLAN tags) less than that of Intel 10 Gigabit Ethernet devices.
Resolution/Workaround:
Increase the configured maximum packet size when using Intel Gigabit Ethernet devices.
Affected Environment/Platform:
All.
Driver/Module:
Poll Mode Driver (PMD).

8.19. Binding PCI devices to igb_uio fails on Linux kernel 3.9 when more than one device is used

Description:
A known bug in the uio driver included in Linux kernel version 3.9 prevents more than one PCI device to be bound to the igb_uio driver.
Implication:
The Poll Mode Driver (PMD) will crash on initialization.
Resolution/Workaround:
Use earlier or later kernel versions, or apply the following patch.
Affected Environment/Platform:
Linux systems with kernel version 3.9
Driver/Module:
igb_uio module

8.20. GCC might generate Intel® AVX instructions for processors without Intel® AVX support

Description:
When compiling DPDK (and any DPDK app), gcc may generate Intel® AVX instructions, even when the processor does not support Intel® AVX.
Implication:
Any DPDK app might crash while starting up.
Resolution/Workaround:
Either compile using icc or set EXTRA_CFLAGS='-O3' prior to compilation.
Affected Environment/Platform:
Platforms which processor does not support Intel® AVX.
Driver/Module:
Environment Abstraction Layer (EAL).

8.21. Ethertype filter could receive other packets (non-assigned) in Niantic

Description:

On Intel® Ethernet Controller 82599EB When Ethertype filter (priority enable) was set, unmatched packets also could be received on the assigned queue, such as ARP packets without 802.1q tags or with the user priority not equal to set value. Launch the testpmd by disabling RSS and with multiply queues, then add the ethertype filter like the following and then start forwarding:

add_ethertype_filter 0 ethertype 0x0806 priority enable 3 queue 2 index 1

When sending ARP packets without 802.1q tag and with user priority as non-3 by tester, all the ARP packets can be received on the assigned queue.

Implication:
The user priority comparing in Ethertype filter cannot work probably. It is a NIC’s issue due to the following: “In fact, ETQF.UP is not functional, and the information will be added in errata of 82599 and X540.”
Resolution/Workaround:
None
Affected Environment/Platform:
All.
Driver/Module:
Poll Mode Driver (PMD).

8.24. Devices bound to igb_uio with VT-d enabled do not work on Linux kernel 3.15-3.17

Description:
When VT-d is enabled (iommu=pt intel_iommu=on), devices are 1:1 mapped. In the Linux kernel unbinding devices from drivers removes that mapping which result in IOMMU errors. Introduced in Linux kernel 3.15 commit, solved in Linux kernel 3.18 commit.
Implication:

Devices will not be allowed to access memory, resulting in following kernel errors:

dmar: DRHD: handling fault status reg 2
dmar: DMAR:[DMA Read] Request device [02:00.0] fault addr a0c58000
DMAR:[fault reason 02] Present bit in context entry is clear
Resolution/Workaround:
Use earlier or later kernel versions, or avoid driver binding on boot by blacklisting the driver modules. I.e., in the case of ixgbe, we can pass the kernel command line option: modprobe.blacklist=ixgbe. This way we do not need to unbind the device to bind it to igb_uio.
Affected Environment/Platform:
Linux systems with kernel versions 3.15 to 3.17.
Driver/Module:
igb_uio module.

8.25. VM power manager may not work on systems with more than 64 cores

Description:
When using VM power manager on a system with more than 64 cores, VM(s) should not use cores 64 or higher.
Implication:
VM power manager should not be used with VM(s) that are using cores 64 or above.
Resolution/Workaround:
Do not use cores 64 or above.
Affected Environment/Platform:
Platforms with more than 64 cores.
Driver/Module:
VM power manager application.

8.26. DPDK may not build on some Intel CPUs using clang < 3.7.0

Description:
When compiling DPDK with an earlier version than 3.7.0 of clang, CPU flags are not detected on some Intel platforms such as Intel Broadwell/Skylake (and possibly future CPUs), and therefore compilation fails due to missing intrinsics.
Implication:
DPDK will not build when using a clang version < 3.7.0.
Resolution/Workaround:
Use clang 3.7.0 or higher, or gcc.
Affected Environment/Platform:
Platforms with Intel Broadwell/Skylake using an old clang version.
Driver/Module:
Environment Abstraction Layer (EAL).