Solarflare 5.X Linux network driver =================================== Version: v5.3.18.1012 Overview -------- This package is the Solarflare Linux NET driver DKMS (Dynamic Kernel Module support) package. Dynamic Kernel Module Support (DKMS) is a framework where device driver source can reside outside the kernel source tree. It supports an easy method to rebuild modules as you upgrade kernels. Driver updates -------------- The latest version of this driver can always be downloaded from . Documentation ------------- For detailed instructions of how to install and configure the driver please refer to the "Solarflare Server Adapter User's Guide" (part number SF-103837-CD) which is available as a PDF from . For more details on obtaining public keys for verifying signed RPMs please refer to the "Solarflare RPM Signing Keys Release note" (part number SF-117279-LS) which is available in ZIP format from the Linux section of . Supported hardware platforms ---------------------------- The drivers contained in this package support the following adapters: - Xilinx Alveo Adapters - Xilinx Alveo U25 SmartNIC Accelerator Card - Solarflare Server Adapters: - XtremeScale X2522 (10G) PCIe Dual Port SFP28 - XtremeScale X2522-25G PCIe Dual Port SFP28 - XtremeScale X2541 PCIe Single Port QSFP28 - XtremeScale X2542 PCIe Dual Port QSFP28 - XtremeScale X2552 OCP 2.0 Dual Port SFP28 - XtremeScale X2562 OCP 3.0 Dual Port SFP28 - Solarflare SFN8522M Dual Port SFP+ Server Adapter - Solarflare SFN8042 Dual Port QSFP+ Server Adapter - Solarflare SFN8522 Dual Port SFP+ Server Adapter - Solarflare SFN8542 Dual Port QSFP+ Server Adapter - Solarflare SFN8722 Dual Port SFP+ OCP Server Adapter - The following OEM adapters are also supported, but note each OEM may recommend a different driver version to match the version used for testing/qualification - Xilinx SFN8522M-PLUS Ethernet 10Gb 2-port SFP+ Adapter for HPE - Xilinx SFN8042 Ethernet 10/40Gb 2-port QSFP+ Adapter for HPE - Xilinx 2522-25G-PLUS Ethernet 10/25Gb 2-port SFP28 Adapter for HPE - Xilinx 2522-25G Ethernet 10/25Gb 2-port SFP28 Adapter for HPE Building and installation ------------------------- The DKMS system must be installed before the Xilinx DKMS RPM Execute the command "dkms --version" to determine whether DKMS is installed. If not install from your distribution or failing that: http://linux.dell.com/dkms/ Make sure DKMS is version 3.0.3 or newer, or 2.8.7 or older. N.B. From DKMS version 2.8.8 the REMAKE_INITRD feature of DKMS is no longer available and it will be necessary to manually trigger the rebuild of initramfs whenever the driver or kernel version changes. See KI 33605 for details. To install the Xilinx driver DKMS package execute the following command: rpm -i sfc-dkms-.noarch.rpm Linux distribution support for Solarflare Adapters -------------------------------------------------- This driver package is supported on: - RedHat Enterprise Linux 7.9 - RedHat Enterprise Linux 8.8, 8.9, 8.10 - RedHat Enterprise Linux 9.2, 9.3, 9.4 - SuSE Linux Enterprise Server 15 sp4, sp5 - Canonical Ubuntu Server LTS 20.04, 22.04, 24.04 - Debian 11 "Bullseye" 11.0 - Kernel.org linux kernels 4.15 to 6.6 AMD aims to support the current and previous major release of each supported distribution at the time this driver was released (plus, possibly, the latest long term support release if this is neither the current nor previous major release). Support includes all minor updates/releases/service packs of the above major releases, for which the distributor has not yet declared end of life/support. Support ------- Please contact your local AMD support representative or email . Changelog --------- ### v5.3.18.1012 [General changes] ~ Improve kernel compatibility for efx_ptp_get_host_time SWNETLINUX-5201 ~ Add support for RHEL 8.10 SWNETLINUX-5200 ~ Add support for RHEL 9.4 SWNETLINUX-5126 ~ Fix skb leak when attempting to drop PTP packets and check if PTP channel exists before transmitting PTP packets SWNETLINUX-5215 ~ Fix RSS indirection table reporting SWNETLINUX-5310 ~ Report PORT_FIBRE to ethtool for optical module connectors SWNETLINUX-5165 ~ Fix null pointer issue when using separate_tx_channels=1 SWNETLINUX-5206 ~ Don't log unnecessary error messages when trying to enable hardware PPS when it has already been enabled SWNETLINUX-5002 ~ Implement setting and querying hardware timestamp settings SWNETLINUX-5185 ~ Enable physical hardware clock support when PTP_1588_CLOCK is included in kernel as a module SWNETLINUX-5204 ~ Ensure that PPS workqueues are always correctly cleaned up. SWNETLINUX-5136 ~ Add support to enable and disable hardware PPS input via userspace EXTTS events SWNETLINUX-4550 ~ Improve kernel compatibility when using struct netlink_ext_ack and add generic support for symmetric-xor RSS hash. SWNETLINUX-4193 ~ Correctly reflect interface names in /proc/mtd even when these interface are renamed. ~ Add PCI slot info to MTD partition names and extend the MTD partition name buffer to avoid truncation ### v5.3.16.1004 [General changes] ~ Fixes to support newer upstream kernels (v6.6) SWNETLINUX-5141 ~ When HW PPS is enabled ensure it is re-enabled if the NIC is reset SWNETLINUX-5081 ~ Resolve issue if NIC doesn't have IP address configured and a second firmware update is applied by sfupdate without a driver reload. SWNETLINUX-5127 ~ Don't try to allocate XDP TX resources when XDP is disabled SWNETLINUX-5069 ~ Avoid kernel panic in 'dev_gro_receive' by adding a check for zero length in EF10 RX prefix. SWNETLINUX-4975 ~ Remove support for SFA7942Q SWNETLINUX-5135 ~ Provide hardware PPS reading via PHC (/dev/ptpX) to support third-party applications ### v5.3.14.1019 [General changes] ~ fixes to support newer upstream kernels (v6.1) ~ removal of kcompat for old kernels until RHEL 6.10 SWNETLINUX-4904 ~ Remove 128-bit register writes to fix concurrency bug (TX DMA error) SWNETLINUX-5058 ~ Support LC optical connector type SWNETLINUX-4106 ~ fix MTU update "Operation not permitted" problem on VFs SWNETLINUX-4969 ~ monitor periodic statistics DMA updates to avoid stats stop incrementing SWNETLINUX-4930 ~ Use phc.adjfine value scaled ppm value SWNETLINUX-2610 ~ Add PTP cross-timestamping (PTP_SYS_OFFSET_PRECISE ioctl) for higher quality timestamps and compatibility with future kernels. SWNETLINUX-4952 ~ fix use after free error when disabling sriov SWNETLINUX-4674 ~ Add PTP gettimex64 callback for PTP_SYS_OFFSET_EXTENDED ioctl SWNETLINUX-4941 ~ fix build errors with CONFIG_NET_DEVLINK=n SWNETLINUX-4871 ~ replace ptp_clock_info .adjfreq method with .adjfine SWNETLINUX-4885 ~ Fix module eeprom reporting for QSFP modules to avoid "netlink error: Invalid argument" when running ethtool -m SWNETLINUX-4909 ~ Avoid crash when running out of VI resources, to allow the driver to keep trying to assign resources when using num_vis=1 and many VFs ### v5.3.13.1006 [General changes] ~ fixes to support newer upstream kernels ~ removal of kcompat for very old kernels ~ simplify DMA mask setting ~ remove MDIO support Issue 4818 ~ Use short interface names, without the "np0" suffix. These days the suffix is normally set to the port name when the NIC has more than one network port on the same PCI function. The EF10 adapters only support 1 network port per PCI function, so the suffix never adds any useful information. The suffix numbering was unrelated to the physical cage number provided on the backplate of the adapter. The original issue (Issue 64065 below) was resolved by adding the phys_port_id and phys_port_name files in sysfs. Back then the udev rules provided by distributions would not have the side effect of adding the suffix to the interface name. Over time those rules have evolved, and removing the suffix again is best to avoid confusion and misinterpretation. This means the phys_port_name file in sysfs is not supported any more for EF10 adapters, but phys_port_id is still supported. Issue 4732 ~ Add support for SLES15SP4 checking if vdpa.ko is installed Issue 4734 ~ Add support for RHEL 9.0 including PTP Issue 4824 ~ For multi-PF provide ifconfig stats by using vadaptor stats Issue 4708 ~ Provide VF ifconfig stats by using vadaptor stats Issue 4837 ~ Fix param separate_tx_channels=1 by not increasing driverlink IRQs when "reducing" them Issue 4563 ~ Rename external PPS device from 'sfc' to scheme 'ptp#.ext' Issue 4841 ~ Change VF mac via PF as first preference if available Issue 3401 ~ Migrate sfc.ko from virtbus to auxbus Issue 4834 ~ reduce mcdi busy waiting poll Issue 4840 ~ Fix unsupported link speeds warning Issue 4867 ~ Fix datapath reset on nic resume ### v5.3.12.1021 [General changes] ~ Export fibre-specific supported link modes ~ Implement ethtool::get_fec_stats ~ Failure to allocate the RX page_ring is no longer an error (but may reduce performance) ~ Depending on the maximum NIC speed, adjust the size of the Rx recycle ring and provide an RX improvement of up to 18% Issue 4708 ~ Address issue where VF ifconfig stats remain at zero Issue 4704 ~ Disable softirqs for ptp TX Issue 4205 ~ Keep track of ntuple filters while interface is down Issue 4584 ~ Support SIOCDEVPRIVATE from kernel v5.15 Issue 2390 ~ Increase EF_MAX_CHANNELS to 255 ### v5.3.4.1004 Issue 4084 ~ Use spinlock rather than TC mutex in counter update ### v5.3.4.1003 Issue 4061: ~ More fixes. Issue 4052: ~ initialise TC datastructures before channels Issue 4069: ~ correct integer types in tc.c for 32-bit Issue 4077: ~ fix page leak on representor Rx ### v5.3.4.1002 Issue 3470: ~ For remote representers pretend the physical link is up Issue 3944: ~ reset RSS indirection table on ifup Issue 4061: ~ check rep_dev is valid before using it Issue 3905: ~ Ignore IP TOS on egress ### v5.3.4.1001 Issue 3946: ~ Create PF representors Issue 3993: ~ Disconnect encap->neigh when releasing even if not freeing Issue 4042: ~ Correct filter_table_down callback on EF100 VF Issue 4001: ~ Reinsert filters after MTU changes. ### v5.3.4.1000 Issue 3945: ~ Create representors only if we have control of the MAE. Issue 4039: ~ fix VDPA for RHEL 8.3 ### v5.3.3.1006 Issue 3906: ~ Ignore IP TOS on egress ### v5.3.3.1005 Issue 4019: ~ Build fixes for various kernels. Issue 3801: ~ Reattach after not resetting when setting tunnel ports. Issue 4031: ~ Report correct vi_min to driverlink clients. ### v5.3.3.1004 [General changes] Issue 3986: ~ disable TC on RHEL 8.3 Issue 3941: ~ Partial VDPA support. Issue 3312: ~ support encap TSO on EF100 Issue 3927: ~ MAE fixes Issue 3961: ~ Fix missing sensor readings. Issue 3988: ~ Set default BAR to 0 Issue 3937: ~ MAE v2 counter packet handling. Issue 3172: ~ Add support to change ring size in EF100 Issue 3869: ~ Handle VI allocation < requested VIS Issue 3875: ~ Abandon probe earlier when no port exists on a function ### v5.1.0.1001 [General changes] ~ Add support for the SKB receive hash; which can be used by GRO. Issue 80717: ~ Improve race handling for MCDI completions. ### v5.0.1.1003 [General changes] ~ Initial support for Riverhead VCU1525. This is provided by the sfc_ef100.ko kernel module. ### v4_15_17_1000 Issue 5201 ~ Add support for RHEL 8.10 Issue 4193 ~ Correctly reflect interface names in /proc/mtd even when these interface are renamed. ~ Add PCI slot info to MTD partition names and extend the MTD partition name buffer to avoid truncation ### v4_15_16_1000 Issue 4975: ~ Remove support for SFA7942Q Issue 4991 and 5131: ~ Monitor periodic statistics DMA updates except when resetting and when stats-block-usecs 0 ### v4_15_15_1003 Issue 4934 ~ Add kcompat for netif_napi_add changes ### v4_15_15_1002 Issue 4890 ~ Add support for RHEL 8.7 ### v4_15_15_1001 Issue 3157 ~ Add ethtool private flag to toggle link on resets Issue 4824 ~ Use vadaptor stats for non-link control PFs ### v4_15_15_1000 Issue 4731 ~ Add support for RHEL 8.6 Issue 4563 ~ Rename external PPS device from 'sfc' to use the 'ptp#.ext' scheme ### v4_15_14_1003 Issue 4598 ~ Support for CentOS removed Issue 4596 ~ Added compatibility for DKMS 3.0 Issue 4708 ~ Fix VF ifconfig stats ### v4.15.14.1001 Issue 4466: ~ Restore ioctl that enables/disables hardware PPS Issue 4411 ~ Fix UDP tunnel offloads on older kernels Issue 4412 ~ Fix UDP tunnel offload failure at driver unload ### v4.15.13.1000 Issue 2706: ~ Enhance driver filter table size for EF10 hardware Issue 4180: ~ Fix support for small ring sizes ### v4.15.12.1008 Issue 4166: ~ Only use vAdapter stats when VFs are being used Issue 4144: ~ Fix memory leak when unloading driver Issue 4102: ~ Fix memory leak on Rx ring resize Issue 4085: ~ Support new UDP tunnel offload API Issue 4030: ~ Fix build on kernel 5.10 and 5.11 Issue 3854: ~ PTP TX support for RT kernels Issue 3626: ~ PTP, PPS and PHC improvements ### v4.15.12.1001 Issue 3901: ~ Identify QSFP28 properly Issue 774: ~ Expose PPS devices ### v4.15.10.1002 Issue 3809: ~ Fix build on kernel 5.8 Issue 3262: ~ Add support PTP_EXTTS_REQUEST Issue 3747: ~ Show sensor value even when min1/max1 are not defined Issue 3601: ~ Fix build without PTP support ### v4.15.9.1003 Issue 3723: ~ Fix build error on RHEL 8.3 Issue 3523: ~ Correct sensor displays on RHEL 6 and 7 ### v4.15.9.1002 Issue 3722: ~ Set link state when suspending. Issue 3710: ~ Support different XDP APIs on different kernels. ### v4.15.9.1001 Issue 3676: ~ Added supported coalesce params ethtool_ops of latest kernel. ### v4.15.8.1001 Issue 3647: ~ Fix compiliation issue when kernel doesn't support MTD. Issue 3491: ~ Set link state on unbind. ### v4.15.8.1000 Issue 3605 ~ Reduce EFX_RX_USR_BUF_SIZE by XDP_PACKET_HEADROOM to increase XDP efficiency. Issue 3564: ~ Change reset on failed to flush to RECOVER_OR_ALL. This will improve stability on heavily loaded systems. Issue 3612: ~ Only run online selftests on firmware variants that support them. ### v4.15.7.1000 Issue 3491: ~ Add LINKDOWN bit to SET_LINK ### v4.15.6.1006 Issue 3547: ~ U25 support for sensors. ### v4.15.6.1005 Issue 3523: ~ Display sensor labels on RHEL 6 and RHEL 7. ### v4.15.6.1004 SF-123551-KI / Issue 3281: ~ Changing ring sizes when using XDP could cause a kernel panic. Issue 3060: ~ Warn when a non-primary function associates with the PHC. Issue 3547: ~ Correct timestamp reconstruction at 16 bit rollover points on SFC8000. Issue 3527, 3274: ~ Decouple MTD cleanup more from net device. ### v4.15.6.1000 Issues 3479, 3405: ~ Adjust compatibility layer to improve range of kernel support. ### v4.15.5.1007 Issue 3405: ~ Enhance the ARFS expiry mechanism. Issue 3376: ~ Link down was sometimes misinterpreted as link-up. Issue 3332: ~ Avoid too many MC reboots for UDP tunnels. ### v4.15.5.1003 Issue 3336: ~ Fix a mistaken attempt to use PTP on NICs for which PTP is not licensed. Issue 876: ~ Do not print an error when the old private ioctl EFX_TS_INIT is issued on new kernels. Issue 3350: ~ An optimisation that reduces latency in some situations. Issue 88589: ~ Fix KBUILD_SRC detection to be compatible with linux-5.3. Issue 89184: ~ Correct the channel count for ethtool -x. Issue 89155: ~ If we notice the that NIC has gone into an infinite reset loop, disable it. ### v4.15.4.1003 - 19.2 release Issue 89009: ~ Fix PTP support on Debian 8. Issue 76254: ~ Allow changing the MAC address of a VM on RHEL 7.4 and later. Issue 88295: ~ Remove stack trace when removing the PIO license. ### v4.15.4.1002 SF-122939-KI / Issue 88364,88912: ~ Fix kernel soft lockup when retrieving PF stats. SF-122787-KI / Issue 88259: ~ Fix issue with delivering PTP SYNC packets to sfptpd on VLANed interfaces. ~ This regression is present in driver versions 4.15.2.1001 to 4.15.3.1011 inclusive. Issue 88012: ~ Fix build error when CONFIG_SFC_SRIOV is disabled. Issue 87939,88582: ~ Fix for warnings during driver load of the form sysfs: cannot create duplicate filename '/devices/virtual/workqueue/sfc_pps_0000:02:00' Issue 87608,87609: ~ Address potential WARNING efx_mtd_creation_work() seen on driver load. ### v4.15.3.1011 [General changes] ~ Initial kernel compat fixes for linux kernel 5.2. ~ Reset stats when a network interface is created. Note this will not reset stats on other interfaces that share the physical network port i.e. when sfboot pf-count>=2 Alternatively reset stats with "ethtool --reset mac" or use "sfctool" if your copy of "ethtool" does not support "--reset". ~ LNXNET-62: Report the version of the driver to adapter firmware at driver load. This can then be reported by management protocols (e.g. NC-SI). ~ The 8000 series and newer adapters all get hardware timestamps from the MAC and hence can supply a hardware timestamp for any packet when requested. This enables the use of PTP layer 2 timestamps. ~ add "rss_cpus=rss_numa_local_cores" and "rss_cpus=numa_local_hyperthreads" module parameters to only allocate channels on a given NUMA node. This is useful for very large systems and can significantly reduce the number of interrupts consumed. Issue 87503: ~ Add PCIe domain to PTP/PPS workqueue name to avoid naming collisions in /sys. Issue 86926: ~ Ensure ethtool can return 256bytes of DAC module info (instead of none) when the module is SFF-8472 compliant but where the DDM compliance bit isn't set. Issue 86925: ~ Link speed shows as "0 Mbps full duplex" for bonds on linux kernels < 4.6 Issue 86413: ~ For kernels >=4.19 avoid duplicate names with MTD/NVMEM symlinks caused by interface renaming races in udev. Issue 85750: ~ Fix registration of PHC clocks where there are more than two ports or functions on an adapter. Issue 81246: ~ Fix timestamping for PTP packets over VLANed interfaces on <=SFN6XXX hardware. Issue 80717: ~ Improve race handling in MCDI completions. This was seen to cause LIST_POISON1 assertions in the kernel. ### v4.15.2.1001 -- maintenance release for 18.3 Issue 85216: ~ Some MTD partitions were not correctly populated, with the missing partitions depending on uninitialised data. This could potentially cause issues when updating firmware. Issue 84673: ~ In some situations the driver could fail to recover from a failure to allocate a receive buffer. This may happen when the system is under very serious memory pressure. The driver should now recover once memory becomes available. ### v4.15.1.1008 -- 18.3 release [General changes] ~ Performance improvements for 100G NICs. On a 100G NIC in a single port mode, setting the performance_profile module parameter to "throughput" will put the NIC in a mode that makes more efficient use of the NIC. Please note that this mode is incompatible with EFVI and Onload. ~ Kernel compat fixes to support Linux 4.20. [Deprecation warnings] ~ The hybrid SRIOV support for 5000- and 6000-series NICs will be removed in an upcoming release. ### v4.15.0.1012 -- update release for OpenOnload-201811 [General changes] ~ Kernel compat fixes to support linux 4.19. ~ Use irq_set_affinity_hint when backport is available in RHEL kernels. ~ Calculate descriptors per SKB based off the NIC type. This can allow smaller TX rings. Issue 82232: ~ Previously, invalid tx ring lengths would be corrected and applied when using the ethtool -G|--set-ring interface. Invalid lengths will now be rejected. This also fixed an issue where the uncorrected length would be reported when later using ethtool -g|--show-ring. Issue 82073: ~ Avoid issue if PF0 bound after PF1 when adapter has a PTP license. This has been seen when using netplan on Ubuntu and could result in a panic from queue_work_on() where efx_ptp_event() is also present in the stack. This issue was introduced in v4.14.0.1014. Issue 77401: ~ The PCI VPD is no longer used during driver load. Instead of printing the part number we now show the PCI device and subsystem numbers, which are registered in the public PCI ID database. ### v4.14.0.1018 -- 18.2 release [General changes] ~ Kernel compat for v4.19-rc kernels and RHEL 7.6 beta. [25/50/100G] ~ Added a warning on driver load if using a 25/50/100G capable NIC with a kernel that doesn't support 25/50/100. The warning text is: "This NIC has link speeds that are not supported by your kernel" ### v4.14.0.1014 -- update release for OpenOnload-201805 [General changes] ~ Kernel compat fixes to support linux 4.18 and RHEL 6.10 ~ Removal of obsolete kernel compat support [ESXi in-guest PTP] ~ Allow VFs to register PHC clocks ~ Register a single PHC clock per adapter (grouping on adapter base MAC address) ~ Ensure all adapter interfaces (including VFs) point to the single PHC clock. ~ VFs will only be able to read/sync to adapter clock unless extra privileges are granted. [Removed features] ~ Removal of previously deprecated features: SARFS, SXPS, driverlink RX packet callback, proxy MCDI authorization. Issue 81032: ~ Channels used for XDP transmit are now included in the output from ethtool --get-channels, listed as "Other". These channels are only created if the driver is loaded with the module parameter xdp_alloc_tx_resources set. Issue 81011: ~ Ethtool -s (sset/slinksettings) commands are now interpreted more simply, using the autoneg setting to choose between using the advertise bitmap or the speed/duplex settings. This works around a problem in some RHEL7 kernels where 'ethtool -s ... autoneg off' had no effect. Issue 80588: ~ Fix for a memory leak when replacing an XDP programme. The driver was retaining an additional reference to the old programme when a new one was installed, resulting in it not being freed by the kernel. Issue 80393: ~ Fix initialisation of filter semaphore for 5000/6000 series NICs. A semaphore was not correctly initialised, causing errors from debug kernels. Issue 80006: ~ Fix the calculation of required resources when setting up XDP. One XDP transmit queue is required per CPU, but we share an event queue between multiplt transmit queues. This was not accounted for correctly, meaning the driver could refuse to enable XDP despite having resources. ### v4.13.1.1034 -- X2 GA release, OpenOnload-201805 [General changes] ~ Kernel compat fixes to support linux 4.15 [timespec_add_ns(), setup_timer(), mtd_erase_callback(), TICK_USEC ] ~ Support RSS spreading of ethtool ntuple filters and add this to driverlink APIs. To configure this requires upstreamed ethtool changes or a copy of sfctool2 from the sfutils package. ~ Remove "ctpio_dmabuf_start" that was erroneously included in ethtool stats. ~ Expose flash partitions with default settings via MTD. These are now used by sfboot. ~ Replace asynchronous filter operations (e.g. ARFS) to ensure all filter table functions are called in a sleepable context. [Deprecation warnings] ~ Warnings will be printed if the following features of the driver are used: SARFS, SXPS, driverlink RX packet callback, proxy MCDI authorization. ~ Please see the section on deprecated features elsewhere in this release note. Issue 79102: ~ Filter locking improvements. ~ This had been seen to cause a command such as "ip link set ethX vf 0 vlan 10" to cause hung task timeouts with "efx_ef10_filter_insert+0x94/0x900" in the stack. Issue 78429: ~ Avoid using RSS filters with "capture-packed-stream" firmware variant. The error seen would be "efx_ef10_filter_insert failed rc=-95". SF-119883-KI / Issue 78299: ~ Fix for kernel panic or NMI. The backtrace will include "efx_mcdi_rpc_completer" and a "__wake_up" function. Issue 77962: ~ Handle delayed release of MTD memory by reference counting partition info. ~ This could cause a kernel WARNING and backtrace at: drivers/mtd/mtdcore.c:508 add_mtd_device+0x246/0x350 [mtd] Issue 77898: ~ Clean up channel->tx_queues when retrying after VI shortage. ~ This was only seen with 16PFs (sfboot pf-count=16) and resulted in the error "failed to initialise TXQ 8" Issue 77378: ~ Avoid "Unable to set UDP tunnel ports" when using L3xUDP firmware variant. ### v4.13.1.1021 -- X2 pre-production release [General changes] ~ Initial support for Solarflare X2 adapter series. This has not completed testing and qualification at Solarflare and hence should not be used in a production environment. ~ Kernel compatibility for upstream 4.16-rc1. ~ Kernel compatibility for RHEL 7.5. ~ Improve error handling when removing an MTD partition fails. This can lead to repeated messages of the form "Removing MTD device #0 with use count 1". ~ Fix build error with gcc version 7.2.1. ~ Expose the PHC clock for 1 VF per VM. ~ Increase the number of multicast filters to 512. ~ Disabled SKB cache on recent kernels due to netdev_alloc_skb improvements making it redundant. ~ Check if MCDI event pending, when MCDI times out. ~ Expose FEC statistics on X2 adapters. ~ Add support for 25/50/100G supported/advertised speeds. ~ expose CTPIO stats on NICs on X2 adapters. ~ Add control of forward_fcs through ethtool feature flags. ~ support X2 VI strides other than 8k. SF-119883-KI / Issue 78131: ~ Address a potential kernel panic where the backtrace will include "efx_mcdi_rpc_completer". ~ This affected systems running sfc v4.12 prior to v4.12.2.1013. Issue 78042: ~ Don't restore filters marked BUSY in efx_ef10_filter_table_restore(). ~ This could lead to the following errors, but only after the rare event that the MC resets: MC command 0x8a inlen 108 failed rc=-22 (raw=0) arg=0 Issue 76297: ~ Add 'immediate send' feature to async MCDI and use it to ensure that PTP time sync commands are sent in the same thread as the process_times logic is done in. [Stratus ftServer support] - Note these issues only affect PCI hot-plug cases Issue 76858: ~ Fix kernel crash in show_max_adjfreq+0x20/0x40. ### v4.12.2.1014 -- update release for OpenOnload-201710 [General changes] ~ Kernel compatibility for upstream 4.15, 4.16-rc5 and RHEL 7.5. ~ Increase the number of multicast filters used before reverting to the 'multicast all' filter. In response to customer feedback the maximum number of multicast filters inserted has been increased from 256 to 512. Above that the individual filters will still be replaced with a 'multicast all' filter. SF-119883-KI / Issue 78131: ~ A race condition in the MCDI (driver<=>adapter communication) state machine could result in stack data being accessed from a different thread after it had gone out of scope. This would result in invalid addresses being read or a hard lockup, both with efx_mcdi_rpc_completer and __wake_up in the call stack. A similar race (not observed in use) was identified and fixed in the PTP/MCDI interaction. ### v4.12.2.1006 -- update release for OpenOnload-201710 [General changes] ~ GCC 7.2 introduces additional warnings related to potential string truncation. The warnings generated by the drivers are not causes for concern, since the actual possible numerical range is small; however, the format string checking isn't aware of the actual numerical range. Format strings have been updated and types changed to prevent warnings. ~ Kernel compatibility for upstream 4.14, 4.15-rc6. Issue 74885: ~ A race condition in the MCDI (driver<=>adapter communication) state machine could result in cleanup code being run twice, resulting in various errors with efx_mcdi_process_cleanup_list in the call stack. Issue 75614: ~ During a link flap caused by a switch the driver would log link up/down messages at an excessive rate. These messages are now rate limited. SF-119179-KI / Issue 75087: ~ Interrupts were wrongly named "-xdp" instead of "-tx" when using separate_tx_channels mode. Issue 73384: ~ Adapter license checks must be completed before the driver attempts to read any capability flags. ### v4.12.1.1021 -- GA release (SF17.2) SF-119052-KI / Issue 74628: ~ Fix "rx event arrived on queue xx labeled as queue yy" that could disable the interface. ~ This could only occur with: * >32 RXQs in-use (i.e. >32 logical CPUs or rss_cpus module parameter set). * sfboot msix-limit of 64 or above (a non-default configuration). * SFC net driver v4.12.1.1016 (i.e. OpenOnload-201710). Issue 74078: ~ Remove warning when expiring an ARFS filter that has been replaced with a higher priority filter (such as an ethtool ntuple filter). ### v4.12.1.1016 -- release with OpenOnload-201710 [General changes] ~ Rework of MCDI (driver<=>adapter communication) state machine (to better support long running background operations on the adapter). ~ Increase maximum number of channels (as determined by the rss_cpus module parameter) from 32 to 64. This allows more RX and TX queues which is useful on systems with large numbers of cores. Note that most NICs will have a default MSI-X interrupt count limit of 32. This can be increased using "sfboot msix-limit=". ~ Support configuring a 2-tuple RSS hash for UDP. This ensures that a UDP flow always hashes to the same CPU regardless of how many IP fragments the UDP packet spans, To configure a 2-uple of IP source and destination address: # ethtool -N rx-flow-hash udp4 sd To confirm which fields are used for the RSS hash: # ethtool -n rx-flow-hash udp4 To use the The default is a 4-way hash: # ethtool -N rx-flow-hash udp4 sdfn [PTP changes] ~ Increase the frequency adjustment range supported by the net driver to +-10% and create mechanism for sfptpd to detect this. [XDP preview] ~ This release includes experimental driver support for XDP (Express Data Path). XDP is a kernel feature that allows eBPF programs to be run on incoming packets before they are passed to the rest of the network stack. The program can modify the packet and instruct the driver to: - accept the packet - drop the packet - transmit the packet on the same interface - transmit the packet on a different interface ~ Support for the various features depends on different upstream kernel versions. Please check XDP support is available in your kernel: - Support for drop and TX appeared in linux v4.8. - Support for redirect is appearing in linux v4.14. ~ For XDP transmit, additional transmit queues are used by the driver. To avoid increasing the resource utilization in general use, this is NOT enabled by default. To enable allocation of the necessary resources, the module parameter "xdp_alloc_tx_resources=1" should be specified when loading the driver. The intention is for a future release of this driver to allow such resources to be allocated dynamically, avoiding the need for this load time configuration; as such, it is likely that this module parameter will be removed in a future release. Issue 73931: ~ Check structure pointer is valid before use in debugfs. ~ This was seen to cause a NULL pointer dereference in debugfs access only if certain parts of driver initialization had failed. There is no issue if the driver successfully creates a network interface. Issue 73700: ~ Candidate fix for "kernel BUG at lib/dynamic_queue_limits.c:26!" by adding memory barriers to address race between netdev_tx_sent and netdev_tx_completed. Issue 73384: ~ [SFN8000 series] No licenses available if the sfc driver loaded immediately after firmware restart. ~ This could only be triggered with "ethtool -t enp1s0f0; onload_tool reload" Issue 73143: ~ Fixed reading of ef10 data on non-ef10 NIC (i.e. SFN5000 or SFN6000 series) ~ This was detected using KASAN by Stefano Brivio at Redhat: BUG: KASAN: slab-out-of-bounds in efx_mcdi_mac_stats Issue 73104: ~ Only report link down if it was previously up. ~ This can occur as the adapter firmware may report changes to PCS states while the overall link state is still down. Issue 73085: ~ [PTP] Erroneous PTP offset sometimes reported on multi CPU-package machines, believed to be caused by QPI routing. ~ Improve sample selection by using the one which is closest to the average of good samples. Issue 72543: ~ [onload] Increase event queue 0 allowance for driverlink events. This had been seen to lead to the following errors on TX queue 0: MCDI request was completed without an event TX queue 0 spurious TX completion id 1480 Issue 72328: ~ Fix possible overflow of unicast address list. e.g. by configuration of a 32nd macvlan interface all on the same network interface. The error was of the form: WARNING: at driver/linux_net/ef10.c:7032 efx_ef10_filter_insert_addr_list+0x4d2/0x510 [sfc]() Issue 72220: ~ [32bit PHC support]. Fix bug in efx_phc_gettime32() that returns junk time. ~ This only affects systems without 64bit PHC support e.g. some SLES11 (kernel 3.0.101) Issue 71931: ~ Correct endianness of filters debugfs output. Issue 70902: ~ [SFN5000 and SFN6000 series] Fix panic in efx_ethtool_get_regs() from "ethtool -d". Issue 68849: ~ Fix "hwmon_device_register() is deprecated" warning seen in Ubuntu 17.04 Issue 68040: ~ [PTP + SFN8000 series] Report correct TX timestamp corrections for (a) normal data path used for PTP transmit. (b) general TX timestamps from the MAC. Issue 67150: ~ Reduce race condition waiting for ethtool -t offline self-test to complete. Issue 66860: ~ Correct VF handling when "ethtool -t offline" self-test is run on any privileged PF. Issue 63057: ~ [SFN8000 series] fallback to default speeds if none advertised. ~ There can be an issue getting a link on SFN8000 series adapters after a RHEL 6.8/7.3 in-box driver has been loaded first. Issue 62723: ~ Remove PCIe FLR after "Failed to acquire MCDI state" as this does not seem to improve driver robustness. Issue 58792: ~ Fix WARNINGs on driver unload for (debug) kernels with the option DEBUG_KOBJECT_RELEASE enabled. Issue 56902: ~ Ensure filter is removed during an MC reboot (e.g. sfupdate) when onload is being used. This has been seen to cause the following errors: [onload] tcp_helper_endpoint_clear_filters: [0:28] sfc 0000:01:00.0 enp1s0f0: MC command 0x8a inlen 108 failed rc=-22 (raw=0) and [onload] __oof_manager_addr_del: 0:79 CLEAR TCP 172.16.132.67:35769 sfc 0000:01:00.0 enp1s0f0: MC command 0x8a inlen 108 failed rc=-22 (raw=0) Issue 48050: ~ Handle adapter VPD data residing at offsets greater than 512bytes in PCI config space. ~ This resulted in the message "Unable to read VPD" [Stratus ftServer support] - Note these issues only affect PCI hot-plug cases Issue 68343: ~ Fix deadlock between efx_channel_reassign_irq and efx_irq_notify. ~ Could also be hit when running "ethtool -C" and "ethtool -G" in a loop. ### v4.10.7.1004 -- Update release for EOL5 Issue 75470: ~ Error handling in the transmit path in efx_dequeue_buffers could cause a null pointer dereference. This error handling has not been hit in any known production environment. Issue 73143: ~ Fixed reading of ef10 data on non-ef10 NIC (i.e. SFN5000 or SFN6000 series) ~ This was detected using KASAN by Stefano Brivio at Redhat: BUG: KASAN: slab-out-of-bounds in efx_mcdi_mac_stats Issue 72328: ~ Fix possible overflow of unicast address list. e.g. by configuration of a 32nd macvlan interface all on the same network interface. The error was of the form: WARNING: at driver/linux_net/ef10.c:7032 efx_ef10_filter_insert_addr_list+0x4d2/0x510 [sfc]() ### v4.10.7.1001 -- Update release for EOL5 Issue 51996: ~ Kernel compatibility fix for early RHEL 6.x releases. WQ_MEM_RECLAIM was not defined. ### v4.10.7.1000 -- Update release to match EO5 release [General changes] ~ Kernel compatibility for RHEL 7.4 ~ Change to newer API to avoid the following deprecation warning: hwmon_device_register() is deprecated. Please convert the driver to use hwmon_device_register_with_info(). ~ Expose the sfc_pps work queue in sysfs so that CPU affinity can be controlled. ### v4.10.6.1002 -- Update release (SF17.1) [General changes]: ~ Kernel compatibility extended to Linux 4.11. ~ Adjust to core networking busy polling support available in Linux 4.11. [XDP preview] ~ Preview of RX side XDP support. ~ This allows for RX filtering to use the DROP target and for RX packet modification (e.g. decapsulation). ~ Additionally header expansion is supported (e.g. encapsulation). Issue 69569: ~ Fix timestamp being 1s behind if the minor nanosecond timestamp happened to be exactly on 0. ### v4.10.4.1005 -- GA release to match openonload-201606-u1.2 [General changes] ~ Kernel compatability extended to Linux 4.10. Issue 68158: ~ Fix driver build without PTP support. Issue 67994: ~ Fix for pps_get_ts() only returning microsecond resolution with RHEL 7.2. ~ (Note this did not affect other RHEL 7.X releases). ~ This would affect the PTP convergence accuracy that is achieved. ### v4.10.4.1000 -- GA release for ESXi adding support for SFN8542 adapters. [General VMware ESXi 5.5, 6.0 and 6.5 support] ~ GA support for ESXi and SFN8542 Solarflare adapters. ~ Support VXLAN acceleration for ESXi. ### v4.10.0.1025 -- GA release for overlays support on SFN8XXX adapters. [General changes] ~ Improve error message when a TX queue times out to give more information. Issue 66936: ~ Ensure the packet prefix is available to timestamping. ~ Without this patch invalid memory accesses to addresses ending 0xfb will be seen in __efx_rx_skb_attach_timestamp() if GRO or LRO is being used. Issue 66058: ~ Clear napi_hash related state when copying channels. ~ This is only seen on driver unload when >256 NAPI channels are being used (across all adapters on the server; partly determined by the sfc module parameter "rss_cpus"). ~ This has been seen to cause: - general protection fault in efx_stop_all+0x38a/0x5a0. - kernel NULL pointer dereference in napi_hash_add+0x59/0xc0. [Stratus ftServer support] - Note these issues only affect PCI hot-plug cases: Issue 67007: ~ Fix race between efx_pci_remove_main() and efx_irq_notify(). Issue 61922: ~ Hold RTNL while we modify the CPU channel map. ### v4.10.0.1011 -- Intermediate release as part of openonload-201606-u1 [General changes] ~ GA release of acceleration for overlay networks, VxLAN / NV-GRE / GENEVE (previously beta) ~ Check MTU against minimum threshold. ~ Better detection of BIOS/kernel failing to provide addresses for PCI BARs. Issue 66869 (ESX only): ~ Fix a race in the admin side of MCDI proxy authorization. ~ The error message reported would be of the form: efx_proxy_auth_handle_request: invalid req state 5 Issue 65427: ~ Correctly report UDP RSS hashing mode as 2-tuple (address only) or 4-tuple (address and port) with "ethtool -n rx-flow-hash udp4" Issue 65511: ~ Avoid kernel warning from efx_ef10_ev_process() when the sfc.ko module parameter forward_fcs=1, as CRCs should be ignored in this case. Issue 65013: ~ An interaction with the kernel stack can lead to no traffic being transmitted if a very small (less than 68) MTU is set. The driver will now reject such configurations as invalid. Issue 64919: ~ Fix crash seen in efx_pci_probe_main() when sf-boot pf-count was altered and the driver was reloaded. Issue 64574/65176 (SFN8XXX only): ~ Packets with transmit timestamps with the xmit_more flag (set by the kernel queuing discipline) could be delayed before transmission, causing missing packet reports from PTP daemons. In unusual cases this could lead to TX watchdog timeouts. ~ The sfptpd error message reported would be of the form: sfptpd: retrieved transmit timestamp but no packet kernel: sfc 0000:82:00.1 sfc1: TX queue 4 spurious TX completion id 1a ### v4.9.0.1016 -- GA release adding support for SFN8542 adapters. Please note that network interface names may change across this driver update. See issue 64065 below. Support for stateless checksum offloads for NVGRE,VXLAN,GENEVE encapsulation on SFN8XXX adapters is still considered a BETA at this time. [General changes] ~ Fail gracefully when PCIe BARs are unassigned (better detection of BIOS issues which have been seen after a PCI bus rescan). ~ Add support for software TX timestamping. [SFN8000 series support] ~ Add support for "ethtool -C stats-block-usecs". Default is 1 sec, but this can be reduced in order to accurately sample every second from a script. [overlays support] ~ Add encap TNI to debugfs filter dump. ~ Set UDP tunnel port list (for GENEVE and VxLAN) on driver load. This may cause the microcontroller on the adapter to reboot in order to reprogram some elements in the adapter datapath. Issue 64478: ~ In firmware-variant=low-latency if there are more than 256 multicast filter subscriptions then use multicast promiscuous. Issue 64065: ~ Stop setting dev_port and instead implement get_phys_port_id for PFs. This makes network interface names assigned by systemd more consistent with other network drivers e.g. enp130s0f1d1 becomes enp130s0f1. Issue 64013: ~ Correct a kernel WARNING in skb_checksum_help() when using TSO over a VLAN over a bond. The error message seen would be: WARNING: at net/core/dev.c:1907 skb_warn_bad_offload+0xc2/0xf0() This issue was introduced between v4.7.0.1035 and .1036. Issue 63865: ~ Fix checking of PTP license when interface is taken down and then up. This would affect MCFW tx 4096 Issue 63389: ~ Prevent allocating a PTP interrupt in some cases where it is not required. Issue 62350: ~ Only claim checksum offload for encapsulation types that SFN8XXX hardware can handle (GENEVE, VXLAN, NVGRE). Issue 61338: ~ Fix for VFs being broken after sfupdate was run while onload stacks were constructed. The error message seen would be: resetting (DATAPATH) failed to restore vswitching rc=-71, VFs may not function unable to restore all filters ### v4.8.4.1000 -- GA release for ESXi adding support for SFN8522 adapters. [SFN8522 support] ~ GA support for ESXi and SFN8522 Solarflare adapters. ~ Please note that VXLAN acceleration for ESXi, including VXLAN IOVP certification, will follow in a future driver update. ### v4.8.2.1004 -- GA release adding support for SFN8522 adapters. [SFN8522 support] ~ GA support for SFN8522 Solarflare adapters, to match all features of the SFN7XXX series including SR-IOV and NIC partitioning. ~ Support for more efficient TSO on SFN8XXX adapters. ~ Use a symmetric RSS hash by default. This is most useful when sniffing both sides of a TCP conversation. This ensures that all packets are directed to the same RX queue and all are handled in-order. ~ Add "tso_version" and "csum_offload" to TX queue information in debugfs. ~ Attach hardware RX timestamps to SKBs in GRO and LRO code-paths. ~ Support IPv6 flow-steering in ethtool rxnfc API (still needs support in ethtool) ~ Support encap-over-IPv6 in ethtool n-tuple API (still needs support in ethtool) ~ Support for SKB xmit_more flags in compatible kernels. ~ Change to use the 4-tuple UDP RSS hashing, as the resulting hash value will help improve GRO performance. ~ Read adadpter PCIe configuration to warn on PCIe lane width or speed not being optimal for the adapter in use. ~ Use 4K PIO buffers on SFN8XXX adapters. ~ Adds a module parameter "performance_profile" to allow selection of broad tuning options, allowing selection of low latency ('latency') or high throughput ('throughput') profiles. The default is 'auto', in which the profile is selected based off licensed options on the adapter. At present this only changes the event queue initialization options on 8000 series adapters. [BETA quality support for stateless checksum offloads for NVGRE,VXLAN,GENEVE encapsulation on SFN8XXX adapters] ~ Support for checksum offloads in the presense of encapsulation/ overlay networks (tunnels) on SFN8XXX hardware. Outer frames of (NV)GRE / VxLAN / GENEVE are supported. Linux native tunnels and OVS will utilise this checksum offload. ~ Add support for encapsulated TSO for 8000 series adapters. ~ Add message when adapter resets due to VXLAN port list change. ~ Add debugfs entry "vxlan_ports" to print the list of VXLAN ports. [onload] Issue 62104: ~ More conservatively account for the number of wakeup events that the driver might have to handle, one for each possible onload stack. This prevents event queue overflows in rare cases. [EF-VI support] ~ Do not treat a broadcast filter as exclusive in filter logic. This prevented ef_vi from installing >1 UDP filter with the local IP set to broadcast. [Stratus ftServer support] Issue 57944: ~ Take a read lock on the filter table when performing driverlink (onload/EFVI) operations on filters; as there was a possible NULL pointer dereference if there was a race with PCIe hot-plug. [General fixes] Issue 62086: ~ On MC reset (e.g. running sfupdate or ethtool -t offline tests), clear PIO buffer linkage in TXQs. Otherwise, if we fail to allocate new PIO buffers, TXQs will try to use the old deallocated PIO buffers and this will cause an error. Issue 58215: ~ Remove suprious WARNING from efx_hard_start_xmit() seen after "ethtool -t" off-line self-tests. Use of "ethtool -t" can be identified from "MC entered BIST mode" in the kernel log. Issue 57217: ~ Fix taking the network interface into promiscuous (or allmulti) mode when running with low-latency firmware variant. i.e "sfboot firmware-variant=ultra-low-latency". The error message seen would be: Unicast mismatch filter insert failed rc=-93 Multicast mismatch filter insert failed rc=-93 Issue 41741: ~ Implement shutdown() driver entry point. This helps quiesce the Solarflare adapter when using kexec to transition to a new kernel. This issue was seen on IBM POWER machines that raised EEH errors during the kexec. Issue 33781: ~ "ethtool -S" statistic "port_rx_bad_bytes" was not being correctly reset in all cases. ### v4.7.1.1000 -- Update release [General changes] ~ Build fixes for linux 4.5 kernels. ~ Provide an option to place external network port PHY into power down when the network interface is administratively down. This means the link partner also sees link down. Currently this is controlled via an ethtool private flag: - To enable/disable: ethtool --set-priv-flags DEVNAME phy-power-follows-link (on|off) - To show current state: ethtool --show-priv-flags DEVNAME ~ support IPv6 flow steering (needs ethtool support that is not yet released). Issue 53130 ~ Create MTD partitions (to allow sfupdate) even if NVRAM metadata cannot be successfully read. [Stratus ftServer support] Issue 60796: ~ Fix efx_mcdi_acquire_sync() returning -ETIMEDOUT on impolite unplug. ### v4.7.0.1049 -- Update release for EOL 4 Issue 65013: ~ Upstream kernel 4.10 centralised checking of minimum and maximum MTU values. In kernels without this checking the kernel doesn't handle small MTUs correctly, so we introduce a check for overly small MTU - the new minimum is 68 bytes. ### v4.7.0.1048 -- Update release for EOL 4 Issue 72220: ~ [32bit PHC support]. Fix bug in efx_phc_gettime32() that returns junk time. ~ This only affects systems without 64bit PHC support e.g. some SLES11 (kernel 3.0.101) Issue 73143: ~ Fixed reading of ef10 data on non-ef10 NIC (i.e. SFN5000 or SFN6000 series) ~ This was detected using KASAN by Stefano Brivio at Redhat: BUG: KASAN: slab-out-of-bounds in efx_mcdi_mac_stats Issue 72328: ~ Fix possible overflow of unicast address list. e.g. by configuration of a 32nd macvlan interface all on the same network interface. The error was of the form: WARNING: at driver/linux_net/ef10.c:7032 efx_ef10_filter_insert_addr_list+0x4d2/0x510 [sfc]() ### v4.7.0.1046 -- Update release for EOL 3 Please note that network interface names may change across this driver update. See issue 64065 below. [General changes] ~ Build fixes for linux 4.10 kernels. ~ Build fixes for RHEL 7.4 kernels. ~ Provide an option to place external network port PHY into power down when the network interface is administratively down. This means the link partner also sees link down. Currently this is controlled via an ethtool private flag: - To enable/disable: ethtool --set-priv-flags DEVNAME phy-power-follows-link (on|off) - To show current state: ethtool --show-priv-flags DEVNAME [SFN8522 support] ~ Support IPv6 flow-steering in ethtool rxnfc API (still needs support in ethtool) ~ Support encap-over-IPv6 in ethtool n-tuple API (still needs support in ethtool) Issue 69569: ~ Fix timestamp being 1s behind if the minor nanosecond timestamp happened to be exactly on 0. Issue 67994: ~ Fix for pps_get_ts() only returning microsecond resolution with RHEL 7.2. ~ (Note this did not affect other RHEL 7.X releases). ~ This would affect the PTP convergence accuracy that is achieved. Issue 64065: ~ Stop setting dev_port and instead implement get_phys_port_id for PFs. This makes network interface names assigned by systemd more consistent with other network drivers e.g. enp130s0f1d1 becomes enp130s0f1. ### v4.7.0.1035 -- Update release [General changes] ~ expose SFA7942Q flash partitions via MTD (to speed up sfupdate) ~ Include Stratus ftServer improvements from v4.5.1.1037 ~ Use timespec64 if available for y2038 compatibility. Issue 59734: ~ Fix detection of PHC so that /dev/ptp* is registered on SLES11/12 Issue 59701 & 58885: ~ Take the mac_lock in efx_ef10_sriov_set_vf_mac() This fixes a kernel panic in efx_ef10_filter_vlan_sync_rx_mode() when using SR-IOV direct pass-thru mode. ### v4.7.0.1031 -- Feature release [General changes] ~ Support adapter firmware state dump via ethtool. ~ Set IRQ interrupt hints for 3.x kernels. The irqbalance service using policy=strict will follow these hints. ~ NAPI budget now only applies to receive processing (previously TX completion and other internal events were incorrectly counted towards the budget). ~ Use symmetric RSS hash key by default. This allows applications such as IDS to ensure that both directions of a TCP stream will be spread via RSS to the same channel. ~ Implement VLAN filtering. (only on full-featured adapter firmware variant) [SR-IOV] ~ Implement sriov_set_vf_spoofchk() for SFN7XXX adapters. [Performance] ~ Automatically configure XPS at driver load (on kernels where this is supported). ~ Implement IRQ notifiers so that XPS can match any changes to interrupt affinity. ~ Report TX completions to BQL after all TX events from an interrupt/NAPI poll. This gives a significant improvement to UDP transaction rates using RHEL7 or SLES12 kernels. ~ read EVQ in cache line batches. ~ allocate rx pages on the same node as the interrupt. ~ Improve performance when the driver is asked to add/strip VLAN tags for linux kernels >=3.0. [General VMware ESXi support] Issue 43845: ~ Avoid installation of duplicate default unicast and multicast Rx filters. Issue 57409: ~ Disable RSS by default ('num_rss_netqs' module parameter with value greater than zero may be used to enable RSS on specified number of NetQueues). [Support SR-IOV on VMware ESXi 5.5 and 6.0] First release supporting SR-IOV on VMware ESXi (SFN7xxx adapters only). This support requires sfutils >=v4.7.0.1004 which contains adapter firmware v4.7.0.1019 and firmware-variant=full-feature (setup with sfboot). See chapter "SR-IOV Virtualization Using ESXi" of the "Solarflare Server Adapter User's Guide". ~ Support 'max_vfs' module parameter to request specified number of VFs for each PF. ~ Enforce PF promiscuous mode if VFs are requested (required on ESXi 5.5 only since ESXi 6.0 enforces this itself). ~ Support VF MAC address configuration; security settings of the port group that the VM network interface belongs to should be used to control it. ~ Support VF default VLAN tagging done by the NIC. ~ Support VF anti-spoofing control configured as "allow/deny forged transmits" in security settings of the port group the VM network interface belongs to. ~ Support reading VF statistics from the hypervisor (for example, esxcli network sriovnic vf stats -n vmnic2 -v 0). ~ Support MTU change requested on VF, if allowed by hypervisor. This is configurable in VM network interface settings. ~ Support VF promiscuous mode, if allowed by hypervisor (ESXi 6.0 only). ~ Support VF VLAN membership control by hypervisor; only untagged traffic is allowed by default. Issue 51699: ~ Fix lockdep warning on driver unload. Issue 56887: ~ Fix memory leak for EFX_MCDI_REQUEST private ioctl calls e.g. from sfupdate. Issue 56934: ~ avoid rx copybreak on GRO tail packet to avoid potential reordering. ### v4.5.1.1038 -- Update release [Stratus ftServer support] Issue 59975: ~ Add workaround for loss of RX doorbell in a "breaker" test. ### v4.5.1.1037 -- Update release [General changes] ~ Support for linux v4.3 ~ Simplify locking when CONFIG_NET_RX_BUSY_POLL is enabled (~20ns latency win) ~ Make RSS hashing symmetric by default. The key chosen has been shown to provide good spreading as well as being symmetric. Symmetry is important for applications that sniff both directions of TCP conversations (eg. IDS and other security apps). ~ Support for Solarflare SFA7942Q QSFP+ AOE Adapter (but please use specific packages that are targeted at this adapter). Issue 55559: ~ Cleanup IRQ affinity hints correctly in probe failure path which caused this warning: "WARNING: at kernel/irq/manage.c:1221 __free_irq+0x19a/0x1e0()" Issue 56887: ~ Fix small memory leak on every EFX_MCDI_REQUEST private ioctl call. Issue 56934: ~ Avoid rx copybreak on GRO tail packet. ~ If a fragmented TCP packet is being handled by GRO, rx copybreak should not be used on any small tail fragment as this can result in packets being delivered out of order. These small fragments could benefit from reassembly in GRO if part of a stream of larger packets. [OpenOnload support] Issue 57548: ~ IPv4 broadcast address should not be treated as exclusive. ~ this allows ef_vi to install >1 UDP filters with a local broadcast IP. [Stratus ftServer OpenOnload support] ~ Further work to strengthen driver robustness on surprise device removal; especially when OpenOnload is being used. Issue 53701: ~ Support changing MAC address with filters installed. This is required for OpenOnload to add new slaves to a bond. ### v4.5.1.1026 -- Update release [General changes] ~ Support for linux v4.2. ~ Make locking between NAPI and busy-polling less expensive. [Stratus ftServer support] ~ Further work to remove warnings if device isolation and fail-over occurs. ~ Fix infinite loop in efx_pause_napi() on module unload if onload is being used and a hardware fault causes the driver to enter a disabled state. Issue 54445: ~ Address "NOHZ: local_softirq_pending 08" seen on sfc.ko module unload. Issue 54748: ~ When sfc.ko module parameters rss_numa_local=1 and rss_cpus is greater than the number of logical CPUs on a node, remove the false limitation that onload could only use interrupts on the local NUMA node. Issue 54840: ~ Fix efx_ef10_filter_is_exclusive() which did not identify the ethertype and IP-protocol filters as exclusive. Issue 55443: ~ Fix rare case where after several failed attempts to allocate memory (in order to refill the RX descriptor ring) the driver did not continue to retry. This causes the ethtool stat "rx_nodesc_drops" to increase. Issue 56015: ~ Fix potential NULL pointer deference in efx_mcdi_mon_remove(). ### v4.5.1.1020 -- Update release Issue 50755: ~ Fix serious page reference counting issue relating to the RX recycle ring. ~ For full details please see the following known issue on the Solarflare support site where you can also sign up for e-mail notifications Issue 53778: ~ Resolve deadlock when setting VF VLAN via "ip link" command. Issue 54347: ~ Add legacy method for changing a PF's MAC address. ~ This restores compatability with Solarflare adapter firmware <= 4.1.1.1023 ### v4.5.1.1010 -- Feature release [General changes] ~ Kernel compat support for v4.0.0. ~ Support ethtool filters on IP protocol. ~ Allow setting and querying the RSS hash key through ethtool interface. ~ Implement ethtool -l|--show-channels. ~ Add known good default RSS (Topelitz) key (vs. using a random key) and add a module parameter 'rss_use_fixed_key' (default:1) ~ Rework allocation of TX queues to allow driver to take advantage of upcoming firmware performance improvements on SFN7142Q adapters in 2*40G mode. [Stratus ftServer support] ~ Introduce support for Stratus ftServers (including polite and impolite removal cases). Driver module parameter 'monitor_interval_ms' (default 200ms) is used to test see if a PCIe isolation event has occurred. ~ Please use the version of the driver that is supplied by Stratus that has been fully qualified for your ftServer. [General fixes] Issue 52542: ~ Fix rare 'kernel BUG at include/linux/skbuff.h:1485 in efx_rx_mk_skb+0x418/0x440' seen with RHEL7 by filling the skb_cache before initialising the Rx queue. Issue 52296: ~ Add timeout in efx_mcdi_acquire_sync() and print a warning on failure. Issue 44816: ~ Fix possible ARFS duplicate flows with the same ID causing the error 'MC command 0x8a inlen 12 failed rc=-22 (raw=22) arg=2' by tracking RPS flow IDs per channel instead of per function [PTP fixes] Issue 47477: ~ If adapter has no PTP license then display "no PTP support" instead of the obscure message "MC command 0xb failed". Issue 45960: ~ fix sfptpd warnings seen when SolarCapture started with ptp=1 and then exited by correcting RX filter insertion actions required in this use-case. [Onload fixes] Issue 52368: ~ [onload] disable and re-enable NAPI more safely during driverlink unregister. Issue 52324: ~ [onload] Fix NULL pointer dereference in Onload efrm_pd_alloc() after passing a device into a VM guest. Issue 51102: ~ [onload/EFVI] don't try to insert mismatch or broadcast filters when in kernel-block mode. This would cause warnings "X filter insert failed" Issue 50721: ~ [onload] Reduce size of initial EF10 memory mapping in preparation of future Onload support on Stratus servers Issue 50320: ~ Fix possible 'rcu_preempt self-detected stall' warnings seen where the stack trace includes 'efx_dl_handle_event' by obeying NAPI budget accounting for work done in driverlink events. [Build and kernel compat fixes] Issue 53241: ~ Build fix for 64bit PHC operations 'ptp.c:2187: error: unknown field 'gettime' specified in initializer' Issue 53103: ~ Build fix for 3.181.11 error: 'struct pci_dev' has no member named 'physfn' Issue 52632: ~ Fix build issue when CONFIG_SFC_SRIOV is not set. 'error: 'struct efx_ef10_nic_data' has no member named 'vf'' Issue 52404: ~ Build fix when CONFIG_PCI_IOV is not set to avoid the error 'error: 'efx_sriov_get_phys_port_id' undeclared here (not in a function)' Issue 52379: ~ Improve kcompat for 3.18.10 that caused 'error: implicit declaration of function 'set_cpus_allowed'' Issue 52371: ~ add RHEL5 kernel compat for IS_ERR_OR_NULL used in efx_init_debugfs_netdev() Issue 51721: ~ Avoid the following warning on RHEL7.1 'WARNING: at net/core/dev.c:5450 register_netdevice+0x1cc/0x410()' Issue 51627: ~ Build fix for 'sfc_minimal=1' which gave the error 'struct efx_tx_queue' has no member named 'sarfs_update' Issue 51443: ~ Build fix in mcdi_proxy.c for kernels without CONFIG_NET_NS set. 'error: 'struct net_device' has no member named 'nd_net'' Issue 50816: ~ kcompat changes to resolve 'ethtool -K eth2 rx off' failures on RHEL6.6 Issue 42338: ~ Build fix to use pci_enable_msix_range() instead of pci_enable_msix() for latest kernels. ### v4.4.1.1017 -- Feature release New features: ~ Support for an access control daemon to control multicast subscriptions (please contact Solarflare support for more details of how to use this). ~ Performance improvements. - enable rss_numa_local by default - set the default rx_recycle_ring_size to 512 - set rx_copybreak default to 192 - Reduce lumpiness of the sarfs hash function - Avoid GRO path for small packets so that data is copied. - Change to fixed RX recycle ring - Support RX descriptor reposting. - Replace atomic_set() with atomic_add() when incrementing rx_buf page - Add compiler hints to reduce icache miss rate for efx_enqueue_skb() - Move re-enable of TX path to end of napi poll loop ~ Experimental support for NIC partitioning + SRIOV used simultaneously. ~ Support setting the MAC and VLAN of a VF via the PF (via the "ip" command) Issue 48262: ~ Avoid inserting unicast filters when kernel is blocked (Solarcapture/EF-VI) Issue 48771: ~ Support SFF8472 diagnostics for SFP+ modules connected via QSFP. Issue 48310: ~ Allow tx_copybreak=0 to disable this feature. Issue 49178/49179: ~ enable cascaded multicast filters in adapter firmware (if present). Issue 48910: ~ [SFN5/6XXX] Fix regression "num_vis=-1" was exporting no VIs to onload. Issue 45957: ~ kernel 3.18 and 3.19 compatibility. Issue 49573: ~ Implement destination port/ip filtering option for SARFS. ~ Add debugfs counter for the number of updates to filters that have been requested from tx_queue. Issue 48771: ~ Pass QSFP module information through the driver to ethtool. Issue 50830: ~ Fill in rxhash correctly in LRO and non-receive offload paths. Issue 49577: ~ Fix rare crash when inserting hardware filters using ethtool. Issue 50992: ~ Fix issue where "ip link show" gives wrong mac address for vfs in direct passthrough mode, by updating the PF's record of the VF's MAC address if possible. Issue 49598: ~ Reduce the frequency and verbosity of messages generated from (expected) denied firmware requests by unprivileged PFs/VFs. ~ Make failed filter removal less noisy after an MC reset. ### v4.2.2.1016 -- Feature release First release supporting NIC partitioning, SR-IOV and PF-IOV modes on SFN7XXX adapters. This support requires sfutils >=v4.3.0.1010 which contains adapter firmware v4.2.1.1014 and firmware-variant=full-feature (setup with sfboot) General changes: ~ the old module parameter "enable_multi_pf" has been removed as NIC partitioning (with multiple PFs) is now fully supported by this driver ~ add sysfs entry and module parameter to control MCDI tracing between the driver and the adapter firmware. This is an extra debug feature that is for the use of Solarflare support. Issue 44671: ~ Work around TRIGGER_INTERRUPT command not being supported on SFN7142Q. Issue 44372: ~ Disable RSS on packed stream firmware variant (used for SolarCapture) as it is not supported Issue 44737: ~ Increase BIST mode timeout from 10s to 30s to accommodate longer SFN7142Q BIST operations. Issue 46845: ~ remove bogus __GFP_COLD flag in page allocation to improve performance. Issue 47102: ~ replace get_page() with atomic_set() to improve performance. Issue 44509: ~ PTP. Fix logical error in efx_ptp_restart() Issue 44510: ~ PTP. To increase robustness, efx_ptp_change_mode() should always disable PTP in the firmware if disable is requested even if the driver believes that PTP is already stopped Issue 47551: ~ Only fallback to lower interrupt mode if the lower mode is supported. Issue 47169: ~ Kernel compat fixes for RHEL 6.6 Issue 47339: ~ Kernel compat for linux 3.16. Issue 47268: ~ Allow for a smaller RX ring to be configured (minimum of 128 reduced to 16) Issue 47603: ~ Move selection of SXPS from compile time to run time. ### v4.1.2.1003 -- Feature release New features to support the SolarCapture 1.3 release: ~ New forward_fcs attribute in sysfs which when enabled, causes the driver to not strip CRCs on ingress path. When enabled frames with bad CRCs are not discarded. Changing forward_fcs requires reconfiguring the adapter's MAC ~ Disable ARFS if the required adapter filter flags are not supported ~ Disable RSS on Packed Stream firmware variant Other new features: ~ Add sysfs entry to control MCDI tracing ~ Add module parameter to enable MCDI logging on new functions Issue 45309: ~ Take mac_lock when setting or unsetting the kernel block (from EFVI or onload) in order to avoid races in sync_rx_mode. Issue 44095: ~ change EFX_BUG_ON_PARANOID on mcdi failure in ptp_synchronize to a proper error path. Issue 43849: ~ recheck mcdi->state after timeout in efx_mcdi_await_completion Issue 42806: ~ handle nonlinear SKBs in efx_filter_rfs() Issue 45444: ~ fix calling of free_irq with already free vector Issue 40785: ~ add module option "underreport_skb_truesize" to falsify skb->truesize for LRO packets Issue 43391: ~ use __GFP_NOWARN when allocating RX pages from atomic context. Issue 44829: ~ don't BUG_ON efx->max_channels == 0 in probe Issue 46587: ~ Add RW semaphore to protect against MCDI timeouts in efx_ef10_filter_table_remove() which could cause a kernel panic in efx_ef10_filter_table_restore(). There should not be MCDI timeouts seen in production systems so this is a defensive fix rather than a bug fix. ### v4.1.0.6734B -- Update release for VMware ESXi Issue 51531: ~ Some ethtool operations could cause an RTNL deadlock. This could be experienced when running ethtool from the command line or if openmanage is installed. This can cause a hang when the adapters page is opened in vSphere. ### v4.1.0.6734A -- Update release for VMware ESXi Issue 44988: ~ Use receive ring with 512 entries by default to consume less packet buffers. ~ Change default number of NetQueues from 4 to 8. Issue 46436: ~ If available, put RSS hash information to skb. ### v4.1.0.6734 -- Feature release New Features: ~ Add support for busy polling available in Linux 3.11. See section 1.8 of http://kernelnewbies.org/Linux_3.11 Note that only connect()'ed sockets can make use of this latency reducing feature. ~ Implement SARFS; Solarflare Accelerated Receive Flow Steering (disabled by default) This inserts hardware RX filters in order to direct flows back to the same CPU where the user-land process is transmitted. This should help cache locality. This feature can be used when Accelerated RFS is not available in the currently running kernel, and if ARFS is in use will be disabled. To enable SARFS, set the "sarfs_sample_rate module" parameter to non-zero. 20 is recommended. In addition, some form of XPS is needed for SARFS to function correctly. The easiest way to do this is to enable Solarflare XPS by setting the module parameter "sxps_enabled=1" Alternatively if your kernel supports it you can enable XPS by following the instructions at: https://www.kernel.org/doc/Documentation/networking/scaling.txt or in the Solarflare Adapter User Guide (SF-103837-CD) section 3.20 Performance Tuning on Linux ~ 7xxx: Add support for SFC9140 40G QSFP network controller silicon. (network adapter part number SFN7142Q). ~ Preferentially affinitise interrupts to CPUs on the NUMA node local to the adapter and restrict RSS to only local CPUs (disabled by default). To enable this feature set the module parameter "rss_numa_local=1" ~ add RX hook for driverlink to allow Solarsecure Filter Engine to inspect packets. PTP issues/features: Issue 35011: ~ PTP: Use net_ratelimit() for PTP event queue overflow messages. Issue 41277: ~ PTP: Advertise translated hardware timestamps. Issue 41479: ~ PTP: Remove warning when synchronize operation gets no usable samples. Issue 38893: ~ PTP: Avoid assumptions about skb header area length in efx_ptp_rx() ~ PTP: Remove two redundant efx_ptp_synchronize() calls ~ PTP: Fix preprocessor conditions using EFX_HAVE_PHC_SUPPORT ~ PTP: Always request 'enhanced' PTPv2 filtering on Siena ~ PTP: Switch back to the original PTP clock adjustment operation Issue 41335: ~ PTP: Accept fake PPS events whether or not packet timestamping of PTP packets is enabled in the NIC. This allows phc2sys to be used without requiring linuxptp to be running Issue 42466: ~ PTP: check for NULL efx->ptp_data in efx_ptp_event Issue 42956: ~ PTP: Test for ethtool get_ts_info support directly Issue 41055: ~ PTP: Check for non-PTP or non-primary function in private clock ioctls ~ In some cases this could cause a kernel NULL pointer dereference in efx_ptp_ts_settime() when calling ioctl(SIOCSHWTSTAMP) Issue 41022: ~ PTP: Fix kernel NULL pointer dereference in efx_hard_start_xmit() that could be caused if hardware timestamping on raw sockets were enabled. Issue 40999: ~ PTP: Expose ptp_caps on primary function, not all functions on port 0 Issue 42321: ~ 7xxx: PTP. Fix issue that can cause firmware watchdog to expire while host compares system and NIC clocks. Driver support for OpenOnload / EFVI / SolarSecure Filtering Engine: Issue 40538: ~ onload: multicast loopback: net_driver filter support ~ onload: net_driver filter support for stack_id Issue 41295: ~ Updates driver link filter_block_kernel API (used by onload and EFVI apps) to have a unicast/multicast split, to allow for a closer match to previous semantics. Issue 41992: ~ Sanity check filter specs when inserting from driverlink clients. General Issues / Features / Improvements: Issue 40635: ~ Remove use of procfs in net driver when CONFIG_DEBUG_FS not defined. Issue 41727: ~ Pass correct pointer to free_irq() on failure path Issue 41197: ~ added ioctl support for MCDI licensed app state call used by "sfkey" utility Issue 41248: ~ netif_device_detach in efx_schedule_reset so the tx watchdog won't fire while we're waiting for reset. Issue 42437: ~ use ether_addr_copy and similar helpers. Issue 37976: ~ On MCDI timeout, issue an FLR (and mark MCDI to fail-fast) Issue 42757: ~ Fix TX DMA errors seen when using module parameter "separate_tx_channels=1" Issue 42199: ~ allow PIO more often as PIO conditionals were incorrectly sensitive to some cleanup actions performed during hardware setup. Issue 40373: ~ rewrite pcie_enable_error_reporting error to be less alarming (normally when AER cannot be enabled as the platform firmware is handling this) Issue 39820: ~ Don't count discarded packets as checksum errors as well. Issue 39528: ~ Improve 'Fatal error' logs which occur when MC reboot is detected Issue 43431: ~ Make ethtool_set_flags understand rxvlan and txvlan offload options ~ This is needed allow ethtool -K to operate on RHEL 6.5 Issue 41279: ~ Fix reporting of MC BIST results through ethtool Now 1=>pass, 0=>not run, <0 fail Issue 41791: ~ 5xxx/6xxx: Fix over-accounting of buffer table space for virtual functions Issue 37075: ~ 7xxx: MAC stats (incl loopback) reported per vadapter Issue 41673: ~ 7xxx: Update link fault reporting VMware ESXi Support: Issue 42738: ~ Do not use LRO for IPv6 on ESXi 5.1 and earlier. Issue 43023: ~ Fix jumbo MTU issue when NetQueue is used. Issue 43245: ~ Fix problem with NetQueue filters removal. ### v4.0.2.6640 -- VMware ESXi feature release Support VMware ESXi 5.5: ~ Support SFN5xxx/SFN6xxx and SFN7xxx. ~ Support VLAN tagging on transmit and stripping on receive in driver. ~ Support paired NetQueue feature which says that transmit and receive NetQueues with the same number share event queue/interrupt. ~ Support dynamic NetQueue feature which allows dynamically allocalte NetQueue with specified features. ~ Support RSS NetQueue feature. ~ Support preemptible NetQueue feature which allows to reallocation of a NetQueue with different features. ~ Support dynamic RSS NetQueue feature which allows ESXi kernel to configure RSS hash key and indirection table. ~ Remove 'rss_vlans' module parameter since RSS NetQueue feature is supported. ~ Support 'netq_force_rss' module parameter to enforce RSS on all applicable NetQueues. ~ Support 'num_rss_netqs' module parameter to control number of NetQueues with RSS support (one by default on ESXi 5.1 and newer version, zero on earlier versions). ~ Support LRO and corresponding NetQueue feature which allows to control LRO enabled/disabled when NetQueue is allocated. ~ Support 'netq_force_lro' module parameter to force LRO on all NetQueues including default one. ### v4.0.2.6628.1 -- Hot-fix release ~ use 64-bit writes for PIO as 32bit PIO writes to write-combined mappings can, in very rare circumstances cause data corruption (before TCP checksum offload occurs). Restrict PIO to x86_64 systems. ### v4.0.2.6628 -- Feature release Add support for RX time-stamping and PTP on SFN7XXX family adapters: ~ This release provides the driver support necessary for both the Solarflare Enhanced PTP daemon ("sfptpd") and linuxptpd which is provided in some distributions. Issue 37568: ~ Properly deliver VLAN-tagged PTP packets when disabling PTP. Issue 39894: ~ Fix DMA unmapping of incorrect addresses for TSO TX buffers on SFN7XXX family adapters. This bug would be observed as unmap page errors, and would affect users with TSO (enabled by default) and an IOMMU enabled. Issue 38352: ~ Remove build time warning message when MTD is disabled, as sfutils can now work on linux using an IOCTL to the driver. ### v4.0.0.6585 -- Feature release Support for SFN7XXX family adapters (using SFC91XX family silicon): ~ This release contains full GA quality support for the new SFN7XXX series of network adapters. ~ Currently running PTP on SFN7XXX adapters is not supported. This support will be in a subsequent release in 2013Q4. ~ There are new counters exposed via "ethtool -S" to expose any new internal packet counters. ~ driver support for the "sfkey" utility to modify feature licenses on the adapter. TX improvements: ~ Support for SFC91XX TX PIO (Programmed IO) feature for lower TX latency. This is controlled by the new module parameter "piobuf_size" (default on x86 is 256 bytes). For frames that are smaller than this, and if there are adapter resources available, a packet will be transfered directly to the adapter via PIO rather than incur the latency of a bus mastered DMA. RX improvements: ~ The driver now uses a single optimized RX buffer allocation strategy (the "rx_alloc_method" module parameter has been deprecated and has no effect other than to print a warning). ~ Use memory more efficiently by supporting RX scatter for large MTU values ~ Make initial fill of RX descriptors synchronous. This is to avoid a non-zero "rx_nodesc_drops" count which could previously increment during driver load when there was incoming traffic. Support for POWER systems, including: ~ AER/EEH support in the net driver ~ Improve performance by reusing DMA-mapped pages ~ Improve performance by allocating RX buffers more efficiently in pages of memory Ethtool offline self-test behaviour for SFN7XXX: ~ For SFN7XXX series adapters the ethtool offline selftest (which can be used to verify hardware operation) can cause operations on other network interfaces on the same adapter to fail (as the whole adapter goes into a test mode). Only run the ethtool offline self-test on one port at a time, and expect to see loss of link for the duration of the test on all ports. Note the ethtool online test does not have this limitation. Issue 39031: ~ Remove ethtool RESET operation that could be used by an unpriviledged local user to cause a local denial of service. Issue 31933: ~ Cancel interrupt self-test on interface-down on Linux <2.6.23 Issue 36071: ~ Avoid kernel warning in skb_try_coalesce Issue 34973: ~ Fix RX buffer handling on non-x86 and non-POWER systems Issue 35013: ~ Fix potential kernel panic when using certain (uncommon) MTU values Issue 35019: ~ Fix build failure on RHEL5 kernel Issue 34716: ~ Fix rare memory leak on RX data path Issue 33921 / 33922: ~ Fix RX filter insertion conflicts Issue 34253: ~ Disable AOE on RHEL4 to fix build failure Issue 34299: ~ Fix potential flood of syslog if PTP packet storm occurs Issue 16820: ~ Detach network device during reconfiguration to avoid TX watchdog event Issue 34335: ~ Fix driver load failure due to symbols exported on some kernel versions ### v3.3.3.6329 -- Update release Issue 38482: ~ Fix issue in dealing with multicast mismatch present in v3.3.0.6262, which was caused by the fix for Issue 31533. This could cause old packet data to be received (and hence could look like packet loss, or out-of-order packets). The bug hits when the following events coincide: 1) The arrival of a frame that is discarded by the driver. 2) The receiving channel is in "page-mode" for RX buffer allocation. 3) The previous packet received has been consumed and freed. By default the RX allocation mode is selected automatically by the driver according to the traffic profile, and will prefer page-mode when packets that can be handled by kernel GRO dominate. A workaround for affected drivers is to set the driver parameter "rx_alloc_method=1" (to force SKB based allocations). Issue 37963: ~ Allow PTP to work with vlans post 2.6.37 Issue 37405: ~ Add kernel compat for RHEL 6.5 backport of ARFS ### v3.3.0.6309 -- Update release Issue 34948: ~ Fix CPU soft-lockup on interrupted self-test Issue 34511: ~ Fix kernel panic when VLAN group rapidly created/destroyed Issue 30759: ~ Changes to max buffer table and VI advertised via driverlink ~ Respect num_vis setting even if it exceeds the heuristic limit Issue 35637: ~ Fix kernel panic on RX completion with unexpected index Issue 35644: ~ Use hardware checksum offload under circumstances when we previously didn't (specifically, packets not aggregated by LRO and not passed into GRO) Issue 35774: ~ Remove sfc_tune from RPM builds and disable CONFIG_SFC_TUNE Issue 36530: ~ Kernel compat fix for getnstime_raw_and_real() on 2.6.26.8 Issue 37257: ~ Build fixes for Linux 3.10: VLAN API, I2C API Issue 37256: ~ Remove write permission from phy_type attribute (Linux 3.10 compatibility) Issue 37280: ~ Update netdev notifier function for API change in Linux 3.11 Issue 37110: ~ Always store port number in private data, not net_device::dev_id Issue 35400: ~ PTP: Fix incorrect filter address for PTP Peer Delay Multicast Issue ~ PTP: Fix crash if PTP hardware clock cannot be registered Issue 37248: ~ PTP: mask out PCP and DEI fields in VLAN matching ### v3.3.0.6262 -- Update release Issue 31533: ~ Fix sync of RX DMA buffer when it is not the last in the page (non issue on x86 servers) Issue 31533: ~ Fix efx_rx_buf_offset() in the presence of swiotlb (used on 32bit systems with >3Gb RAM) Issue 33851: ~ Fix very rare data corruption - use TX push only with a single descriptor Issue ~ AOE: Expose FC partition for use by sfupdate Issue 33711: ~ PTP: Fix for MC statistics. Issue 33719: ~ PTP: Check that PTP is enabled in the kernel before compiling PHC support Issue 34355: ~ PTP: Guard against using PPS ioctl API on non-PTP adapters Issue 34356: ~ PTP: Improve accuracy of PPS timestamps Issue 34482: ~ PTP: Fix to SO_TIMESTAMPING system timestamps (in conversion of hardware timestamps) ### v3.3.0.6246 -- Update release Support for Enhanced PTP daemon, including: ~ Support for VLAN tagged PTP packets ~ Configurable receive filtering for VLANs, PTP Domain number and UUID ~ Timestamping of multicast PTP packets on PTP Primary and Peer Delay Multicast addresses ~ Timestamping of unicast PTP packets sent directly to host ~ Separate interfaces to apply a clock offset and adjust the clock frequency ~ IOCTL interface to facilitate synchronization to an external PPS signal ~ Support for recent kernels that include the PHC API Early support for AOE SFA6902F boards General improvements / fixes: ~ Various kernel compatibility fixes up to linux 3.7 Issue 21056: ~ Dropped packet stat ("rx_nodesc_drop_cnt" in ethtool -S) no longer counts drops until interface is fully up Issue 32292: ~ Fix potential kernel warning from efx_hard_start_xmit when changing MTUs (lock TX queues before calling netif_device_detach()) ### v3.3.0.6222B -- Feature release General improvements / fixes: ~ Support for byte queue limits (requires Linux 3.3) ~ Do not fail probe when NVRAM MAC address is invalid ~ Fix races between net device operations, online reset, power management and hotplug ~ Keep NICs quiescent after they have been disabled ~ Improve log messages in case we abort probe due to a pending reset ~ Correct the reported priorities of different filter types ~ Fix spurious array bounds error in efx_filter_redirect_id() ~ Fix PTP synchronisation on big-endian architectures Issue 21781 / 24969: ~ Reduce TX latency by using copy buffers for short fragmented packets This adds the "tx_copybreak" module parameter which sets the size of a TX copy buffer. The size must be a power of 2 between 128 and 4096 (default 128) Issue 24887: ~ Implement VLAN tag insertion for SFC9000 family Issue 31308: ~ Enable VLAN SG and TSO offloads for SFC9000 family on RHEL 5.5+ Issue 20784: ~ Add hardware monitor for sensors managed by firmware Issue 23165: ~ Implement the VF spoofchk (TX filtering) attribute (requires Linux 3.2) Issue 30065: ~ Fix compilation issues with 3.4 kernels Issue 31530: ~ Fix reporting of IPv4 full filters through ethtool/sfctool Issues 31576 / 22910: ~ Really disable flow control while flushing Issue 31671: ~ Work-around flush timeout when flushes have completed ### v3.2.1.6110 -- Update release Issue 30436: ~ Remove duplicate calls to suspend and resume driverlink during self test reset Issue 31308: ~ Fix reduced performance under some older RHEL versions with VLANs Issue 31045: ~ Add RHEL4.8 support Issue 31092: ~ Fix incorrect usage line in sfctool --help output ### v3.2.1.6099 -- Update release Issue 30312: ~ Latency reduction: Use 128bit write for TX descriptor push Issue 28547: ~ Enhancement: Expose SFP+ module vendor information via ethtool Issue 30286 / CVE-2012-3412: ~ DoS protection - prevent extreme TSO parameters from stalling TX queues Issue 30845: ~ Self-test: disable VFs and never set RX_DESCQ_EN for safety Issue 26646 / 22607: ~ Fix possibility of some statistics counters being non-contiguous Issue 30038: ~ Build fix for kernel 3.4-rc5 Issue 30762: ~ Fix kernel-compat test for PCI VPD in sysfs ### v3.2.1.6090 -- Update release Issue 28324: ~ Fix recovery path from memory parity hardware errors ~ Add module option "efx_disable_port_on_mpe" to top recovery and disable interface (to allow bonding to fail over on such errors) Issue 29760: ~ Report serial number of adapter on memory parity hardware error Issue 25050: ~ Limit number of retries on RX queue flush hardware failures Issue: ~ Fix missing cleanup in failure path of efx_pci_probe() Issue 29758: ~ Reporting of MC scheduler message downgraded Issue 26972: ~ Fix start of day NAPI race < 2.6.24 Issue 23581: ~ Display warning of fail to create MTD partition Issue 28008: ~ Timer quantum is incorrect for Falcon user-mode queues Issue 20567: ~ Fix kernel panic with efx_mcdi_phy_remove on failure path to probe port Issue 27481 ~ Fix possible race when stopping TX queues Issue 28139: ~ Self test can fail on machines that disable interrupts for > 10ms Issue 28767: ~ Build fix for kernel 3.3 Issue 28406: ~ SR-IOV: Fix race condition involving rtnetlink and pci_enable_sriov Issue 27202: ~ SR-IOV: Fix initialisation and teardown race Issue 28110: ~ SR-IOV: Correctly set TX MAC filter search limits Issue 27924: ~ SR-IOV: Pass timer quantum over VFDI for SFN6122F support Issue 28404: ~ SR-IOV: RSS only works on VF if also enabled on PF Issue 28528: ~ SR-IOV: Correct validation of allocation size in peer_page_count Issue 27838: ~ Implement EFX_FILTER_FLAG_RX_OVERRIDE_IP for 'default' filters Issue 27571: ~ RX n-tuple interface now deprecated, replaced by RX NFC (users that are configuring RX n-tuple filters with sfctool must also update sfutils to version 3.2.1.6079 or later) Issue 27514: ~ Allow RX hash insertion to be disabled Issue 26944: ~ Self test on driver load removed from Onload build of driver Issue 22798: ~ Set IRQ affinity_hint where supported Issue 28546: ~ Display adapter part number (from VPD) in syslog Issue 28086: ~ Display turbo mode in sysfs Issue 29546: ~ Refill RX rings as soon as space for a batch (reducing latency jitter) ~ Fill RX rings completely full, rather than to 95% full ### v3.2.0.6044 -- Update release Issue T2716: ~ Build fix for sfc_minimal=1 Issue 27569: ~ Build fix for gcc 4.6 Issue 27486: ~ Build fix for RHEL5.7 2.6.18-274.12.1.el5 Issue 27466 ~ Build fix for Linux 3.3 Issue 27470: ~ Fix ethtool -x support Issue ~ SR-IOV: Explicitly keep track of which channel is used for VFDI ### v3.2.0.6040 Support for the SFN5322F SFP+ Precision Time Synchronization Server Adapter Support for the Solarflare SFN6122F Dual Port SFP+ Server Adapter Issue 19619: ~ Remove TX completions from adaptive IRQ scoring Issue 27202: ~ SR-IOV Prevent VFDI and netlink VF ops racing with init/fini of VF state Issue 27001: ~ Don't run on-load self-tests on RT kernels Issue 26985: ~ Work around compiler bug causing incorrect checksums ### v3.1.0.4091 -- Update release Issue 26616: ~ Fix possible self-test failure on driver load - remove dependence on NAPI polling Issue 26697: ~ Fix logging in debug driver on Linux 2.6.34+ Issue 26408 ~ Build fixes for Linux 3.1 Issue 26406 ~ Redhat 6.2 build fix - EFX_HAVE_NTUPLE_SUPPORT compat test Issue 24109 ~ Make all MAC statistics consistently 64 bits wide Issue ~ SR-IOV: correct warning on error from efx_mcdi_rpc(MC_CMD_FLUSH_RX_QUEUES) Issue 26841 ~ SR-IOV: Use full memory barriers at entry and exit of efx_sriov_memcpy() Issue 26436: ~ SR-IOV: Correct validation of VI indices within VF Issue 26437: ~ SR-IOV: Report unsupported VFDI operations specifically and quietly ### v3.1.0.4083 -- Update release Issue ~ Fix assertion failure when using SFC4000 with internal SRAM Issue 25985 ~ Export rss_spread through driverlink Issue 26000 ~ Build fixes for Linux 3.1 Issue 6531 ~ Use correct fields of struct ethtool_coalesce Issue 24380 ~ Correct reporting and validation of TX interrupt moderation Issue 24380 ~ Correct error code for unsupported interrupt moderation parameters Issue 25815 ~ Change rss_cpus default to "cores" Issue ~ Fix kernel compat condition for EFX_HAVE_ETHTOOL_RESET Issue 25542 ~ Fix build failure on Linux 2.6.25 Issue 25542 ~ Avoid build errors with some gcc versions Issue 25420 ~ Export efx_filter_set_{uc,mc}_def() Issue 25414 ~ Fix statistics lock handling Issue 23143 ~ Revert use of write-combining - unsafe on some platforms Issue 23128 ~ Update driverlink API version to 5 Issue 23177 ~ Check for irregular filter tables in efx_restore_filters() Issue 23128 ~ Add support for RX unicast/multicast default filters Issue 22885 ~ Flush software RX pipeline when stopping datapath Issue 24612 ~ Install sfc_tune.ko under 'extra' subdirectory Issue ~ Make LRO module parameter read-only Issue 19703 ~ Reduce number of workqueues used Issue ~ Make reset_nic try ETHTOOL_RESET before private ioctl Issue ~ Cleanup Company/adapter names Issue ~ Use napi_gro_(receive_frags)_gr() on RHEL6.1 Issue ~ Fix MAC statistics on big endian platforms Issue ~ Add EFX_USE_GRO build time parameter Issue 25756 ~ SR-IOV: Don't reset the PF if a VM causes a descriptor fetch error Issue 25486 ~ SR-IOV: Add SCATTER_EN support to VFDI_INIT_RXQ Issue 24004 ~ SR-IOV: Change the vf_limit that rtnetlink exports to 32 Issue 25091 ~ SR-IOV: Only broadcast the peer list when it changes Issue 21004 ~ SR-IOV: Move VFDI processing onto its own workqueue for onload Issue 23954 ~ SR-IOV: export remaining VIs via driverlink when target_num_vis=0 Issue 24582 ~ SR-IOV: Fix "sleeping whilst atomic" during FLR Issue 24373 ~ SR-IOV: Remove the VF MAC address from the peer list when there is no RX filter. Issue 24373 ~ SR-IOV: Fix crash when init/fini races with sysfs/netlink access Issue 24373 ~ SR-IOV: Broadcast the peer list from a work item Issue 23696: ~ SR-IOV: Reduce the maximum VF EVQ size to 8192 entries Issue 24293: ~ SR-IOV: Add module parameter to control tx filtering Issue 21004 ~ SR-IOV: Insert a VF security filter when MAC address non-zero Issue 23954: ~ SR-IOV: Advertise configuration through driverlink Issue 23905 ~ SR-IOV: Reprogram the rx filter when the VF address changes Issue ~ SR-IOV: Fix calculation of driverlink buffer table limits Issue 23843 ~ SR-IOV: Support legacy interrupts on the PF Issue 23676: ~ SR-IOV: Fix mac_addr/tci export ### v3.1.0.4047 Performance: - TX push: reduces TX latency - Write-combining: reduces TX latency (dependent on kernel version) - RX hash insertion: allows faster flow lookup in software - Driver structures reorganised for improved cache locality Features: - SR-IOV support: host-side support for accelerated guest networking in KVM and Xen - RX n-tuple support: use 'ethtool -U' or 'sfctool -U' as an alternative to the sfc_affinity module - Soft-TSO statistics: included in statistics reported by 'ethtool -S' - Driver logging control: for diagnostic purposes, use 'ethtool -s ethX msglvl ...' to enable and disable messages - Register dump support: for diagnostic purposes, use 'ethtool -d' to dump register state ### v3.0.11.2248 -- Update release Issue 22752 ~ Fix CPU topology package selection Issue 22539 ~ Fix potential crash on load - drop packets received during on-load selftest setup Issue 22157 ~ kernel compat fix for 2.6.37 - print_hex_dump() Issue 20990 ~ LRO possible data corruption - remove padding in first skb in super-packet Issue 21637 ~ Report link-down through ethtool API whenever interface is down Issue 21679 ~ Fix build failure on recent non-SMP kernels Issue 21675 ~ Fix build failure with Debian 2.6.36 ### v3.0.8.2221 -- Update release Issue 21669 ~ Fix issue with multiple ports/adapters, where each port has different number of MSI-X interrupts allocated by the kernel Issue 21641 ~ Changing RX/TX ring size can crash when using legacy interrupt Issue 21204 ~ Log message at start and end of ethtool self-test Issue 21639 ~ Build fix for MRG 1.2 (change TX multiqueue compat tests), use proper functions to tell networking core how many queues we have Issue 21499 ~ Include Onload events in NAPI budget Issue 21530 ~ CPU topology fix for some AMD CPUs Issue 21477 ~ ethtool get permanent MAC address fix for 2.6.23+ Issue 21478 ~ WoL - use currently configured MAC address Issue 21465 ~ Standardise basic RPM build dependencies Issue 21437 ~ fix bogus gcc 4.5 build warning/error Issue 21351 ~ Fix for dual speed optics fix using ethtool -s speed Issue 21267 ~ Event based MCDI completion and MC REBOOT/CMDDONE ordering issue Issue 21197 ~ Store the real advertising mask for PHYs - allow us to push back changes correct after an MC reset Issue 19930 ~ Additional validation of userland MCDI request length Issue 20165 ~ Do not preallocate MSI-X vectors (to workaround bugs in older kernels) - can cause interrupt breakage on newer kernels Issue 21069 ~ Remove MDIO dependencies in minimal build Issue 21111 ~ Remove broken automatic fallback for invalid SFC4000n chip/board config Issue 21099 ~ Always read MC firmware version when requested through ethtool Issue 20990 ~ LRO possible data corruption - remove padding from skb fragments Issue 21025 ~ Return error code for invalid ring parameters - prevents kernel panic on invalid settings Issue 20803 ~ SFN4112F: fix issue where self-tests could stop further RX traffic Issue 20807 ~ Handle configs with > 32 TX queues Issue 20798 ~ Fix rss_cpus=cores on mrg-1.3 Issue 19792 ~ SFN4122F: Clear LM87 alarm/interrupt status at start of day Issue 20662 ~ Some MCDI errors are expected - downgrade error message Issue 20649 ~ Fix error message when setting invalid ring parameters ### v3.0.6.2199 -- Update release Issue 20541 ~ Support large rx ring >2048 entries on 64bit kernels Issue 20441 ~ Fix rss_cpus=cores for kernel header only builds Issue 20602 ~ Fix build issue on RHEL 5.6 Issue 18658 ~ Enable TSO for VLAN devices on RHEL 5.5+ Issue 20567 ~ Fix failure paths in efx_probe_port (could cause kernel panic on unload) Issue 19129 ~ Don't display packet drops as errors in ifconfig Issue 18658 ~ Enable TX checksum offload and TSO for VLAN devices on RHEL 5.5+ Issue 20359 ~ Avoid aliasing warning from gcc 4.4.4 Issue 20262 ~ Fix SIOCEFX for 32bit processes (used by sfupdate to reboot MC firmware) Issue 20065 ~ Allocate DMA and event rings using GFP_KERNEL Issue 20055 ~ Log ethtool and MCDI params when set link fails Issue 19906 ~ Change rx_bad to be FCS error count Issue 19996 ~ Midrange NIC support Issue 20006 ~ Support builds without HWMON Issue 19983 ~ Define unlocked_ioctl instead of ioctl for 2.6.36 Issue 19985 ~ Adapt to 64-bit net_device statistics for 2.6.36 Issue 19951 ~ Fix 2.6.27.19-5-xen module load failure Deprecated features ------------------- Support for SuSE Linux Enterprise Server is deprecated. Support for RHEL 7.9 is deprecated. Support for EF100 is deprecated. Removed features ---------------- The following features were previously marked for deprecation and have now been removed. Please see the following descriptions for any impact and alternatives: MSI: ~ MSI facilities are no longer supported. ~ MSI-X facilities are supported. Hybrid SRIOV for older NICs: ~ Hybrid SRIOV support for 5000- and 6000-series NICs with Linux KVM. This feature has not been actively used or maintained for some time. SARFS: ~ This is a feature similar to the kernel feature ARFS (Accelerated Receive Flow Steering) but implemented purely within the sfc.ko driver. All kernels that this driver supports are new enough to support ARFS. Information on using in-kernel ARFS is at https://www.kernel.org/doc/Documentation/networking/scaling.txt or please consult the Solarflare Server Adapter User's Guide [SF-103837-CD] available at https://www.xilinx.com/support/download/nic-software-and-drivers.html#linux SXPS: ~ This is a feature similar to the kernel feature XPS (Transmit Packet Steering) but implemented purely within the sfc.ko driver. All kernels that this driver supports are new enough to support XPS. Information on using in-kernel XPS is at https://www.kernel.org/doc/Documentation/networking/scaling.txt or please consult the Solarflare Server Adapter User's Guide [SF-103837-CD] available at https://www.xilinx.com/support/download/nic-software-and-drivers.html#linux Proxy MCDI authorization: ~ This feature had been used by a proof-of-concept userland utility to vet and control requests to subscribe to a multicast group from an unprivileged function (e.g. a VF in a VM). Driverlink RX packet callback: ~ This was used for inspecting/permitting all RX packets via a driverlink callback. Legacy interrupts: ~ All servers in use are expected to support MSI or MSI-X interrupts. Support for SFN5000, SFN6000 and SFN7000 family adapters has been removed. NIC partitioning, SR-IOV and PF-IOV modes on SFN7XXX/8XXX adapters ------------------------------------------------------------------ This release of the Solarflare Linux Net driver supports NIC partitioning (multiple PFs) and SR-IOV (VFs). See "Solarflare Server Adapter User's Guide" SF-103837-CD Issue >=12 available from https://www.xilinx.com/support/download/nic-software-and-drivers.html#linux for more detailed information on setting up NIC partitioning, SR-IOV or PF-IOV modes. Some additional details as well as known issues are documented below: General: ~ NIC partitioning/SRIOV/PF-IOV features require sfutils >=v4.3.0.1010 which contains adapter firmware v4.2.1.1014 and firmware-variant=full-feature (setup with sfboot) ~ Please note that Hybrid SR-IOV support for SFN5XXX|6XXX drivers is now deprecated. Please use an older driver to continue using this feature. ~ Link control (e.g. setting pause frames with ethtool and controlling MTU) is only available on the first PF on each port; (this is referred to as "PF0" or the "primary PF") ~ Only link control functions have the full MAC statistics available to them (via ethtool -S). All statistics that come from the MAC are now prefixed "port_" as they refer to a single physical network port. All other functions have statistics which show the number of packets and bytes both received and transmitted by that function, classified into unicast, multicast and broadcast traffic. ~ MTD partitions are only exposed from PF0 on each port, hence the utilities sfupdate, sfboot and sfkey need to be run where they have access to PF0. ~ It is now possible to identify which physical port an interface is bound to. See /sys/class/net//device/physical_port (note this is zero based whereas the identifiers printed on the adapter brackets are one based) ~ The driver supports the use of shared RSS indirection tables, after the 16 unique hardware RSS indirection tables entries have been consumed. ~ In order for active-backup bonding to work, the slave interface normally takes the MAC address of the master interface. Only PF0 is allowed to change its MAC address. Therefore for this to work with other functions it is necessary to set "sfboot mac-spoofing=enabled". An alternative is to set the bonding module option "fail_over_mac" and allow the gratuitous ARP to inform the other link partners of the change in MAC address. ~ At this time PTP will only work with the PF0 on each port and will not work with a VF. NIC partitioning support on SFN7XXX/8XXX adapters: ~ The sfc.ko driver will bind to multiple PFs on the same physical adapter port and create a network interface for each PF. ~ Configure with "sfboot switch-mode=partitioning pf-count=X" and cold reboot the server if warned to do so. ~ If using NIC partitioning in the same layer2 broadcast domain you may need the kernel parameter "arp_ignore=2" as linux will reply to ARP on any interface. This might cause traffic to go over unexpected interfaces. ~ If using NIC partitioning with VLANs all PFs must be on a unique VLAN (which can include VLAN 0). Traffic that is sent/received from this PF will have VLAN tags transparently inserted/stripped by the adapter. A single PF on each port can be on VLAN0 where it will both receive and transmit untagged traffic. ~ Configure NIC partitioning with VLANs using "sfboot pf-vlans=,..." ~ Do not use LACP bonding modes when using NIC partitioning as the LACP partner (switch) will not be aware of the partitioning. ~ The driver parameter "num_vis" is now more important. This controls how many extra VIs (RX/TX/event rings) are allocated for the use of onload on each PF. If not using onload you may wish to set "num_vis=0" so that there are no false resource limitations when using multiple PFs. ~ When PF0 is put into promiscuous mode it will receive traffic for its own MAC address as well as any other incoming traffic that does not match another PF's MAC address. For other PFs there is no change in received traffic if the interface is put into promiscuous mode. SR-IOV support on SFN7XXX/8XXX adapters: ~ The sfc.ko driver will bind to all VFs and create network interfaces. ~ Configure with "sfboot switch-mode=sriov pf-count=1 vf-count=X" and cold reboot the server if warned to do so. ~ RHEL6.5 and RHEL7 are supported for the host/hypervisor OS. ~ It is possible to consume the VFs using these KVM/libirt modes: - network hostdev (VF passed into the VM; migration is only possible if the VF is hot-unplugged) - direct-passthrough (macvtap interface used over the VF and virtio-net driver used in the VM guest) ~ Note that VFs will initially be assigned random MAC addresses, but libvirt (and ip link) can override the MAC address. ~ A VF in a VM will be unable to set its own MAC address. See the item on active-backup bonding. ~ To enable VFs (for RHEL6.5 and RHEL 7) echo > /sys/class/ethX/device/sriov_numvfs ~ To enable VFs on older operating systems set the sfc.ko module parameter "max_vfs". If set to a single integer this is the number of VFs to enable and will apply to all PFs. If you wish to configure this to be a different value for each PF set this to a comma separated list. ~ Please ensure that you enable all BIOS settings that refer to virtualization and SR-IOV (note there might be several e.g. SRIOV, IOMMU, VT, VT-x, VT-d, IO virtualization) ~ Solarflare recommends setting "pci=realloc" as a kernel parameter (e.g. in /boot/grub/grub.conf). This allows the kernel to reprogram the initial BIOS PCI apertures in case PCI aperture space for the VFs is not assigned by the BIOS. ~ To view which VFs are associated with which PFs and the VF MAC addresses please use "ip link show" ~ A PF put into promiscuous mode will not see the traffic to/from a VF. ~ A VF put into promiscuous mode will not see any additional traffic. ~ Setting VF MAC/VLANs/link state is possible via the ip command (or libvirt configuration). From "ip link help" "ip link set { dev DEVICE} [ vf NUM [ mac LLADDR ] [ vlan VLANID ] [ state { auto | enable | disable} ] ]" ~ Setting VF spoofcheck/rate/qos (e.g. via the ip command) is not supported at this time. SR-IOV with NIC partitioning mode support on SFN7XXX/8XXX adapters: ~ This mode offers partitioning of the adapter. Each partition is on a unique VLAN and within this layer2 broadcast domain there is switching on the adapter between the PF and VFs. ~ This support is currently considered EXPERIMENTAL i.e. all features are present, but there has only been limited testing. Solarflare will support the use of this feature in a test environment, but does not recommend use in a production environment until this feature becomes GA. Note that both SR-IOV and NIC partitioning as individual features are both considered GA quality. ~ Ensure you are using adapter firmware >= v4.4.2.1011 ~ It is necessary to configure NIC partitioning with unique VLANs in this mode ~ Configure with: "sfboot switch-mode=partitioning-with-sriov \ pf-count=X pf-vlans=[,]... \ vf-count=Y" and cold reboot the server if warned to do so. PF-IOV mode support on SFN7XXX/8XXX adapters: ~ This mode offers switching between PFs. Use this mode where a server does not support SR-IOV (otherwise please use SR-IOV mode as this provides up-to 240VFs versus only 16PFs) ~ Configure with "sfboot switch-mode=pfiov pf-count=X" and cold reboot the server if warned to do so. ~ In RHEL7 passing through a PCI function uses the new VFIO mechanism. Due to a missing capability flag on SFN7XXX/8XXX hardware, VFIO will only allow all PFs to be passed into a guest together. This will be addressed with a kernel quirk. Known issues with the platform (KVM/libvirt/kernel) --------------------------------------------------- ~ libvirt direct-bridged mode on-top of a VF is not supported. Please use libvirt direct-bridged mode over a PF or direct-passthrough mode with VFs. Issue 4732: ~ On SLES15 SP4 vDPA functionality is only supported if the kernel-default-extra package is installed when the driver is build. If the kernel-default-extra package is removed after building the driver it will subsequently fail to load due to unknown vdpa symbols. A driver rebuild is then needed to remove vDPA support from the driver. Note that loading the VDPA module will taint the kernel as "unsupported". Issue 4596 ~ From DKMS version 2.8.8 the REMAKE_INITRD feature has been removed. For details see KI 33605. Issue 4351 ~ On SLES15 direct-passthrough mode is not supported. Issue 46876: ~ Multicast does not work in libvirt direct-passthrough mode. This is because the multicast address list is not propagated to the network interface below the macvtap driver. Without this the sfc.ko/sfc_ef100.ko driver is unable to program the adapter switch to efficiently replicate multicast traffic to only the subscribed functions. Solarflare will work with upstream to find a solution to this issue. Issue 47742: ~ IPv6 traffic does not pass in direct-bridged and direct-passthrough modes. In direct-passthrough mode multicast neighbour solicitation fails as per the previous point. In direct-bridged mode this is because an interface calculates an IPv6 link address based on its MAC address. There will be a neighbour solicitation multicast filter in the host, but when a global address is assigned in the VM guest, a multicast filter corresponding to the new global address is not inserted in the host because macvtap doesn't propagate this information. Issue 46948: ~ The macvtap driver does not propagate the link state from the real network interface. This means that it is not possible to use bonded interfaces that use link state detection (miimon). Instead please use ARP probes. This affects libvirt direct-bridged and direct-passthrough modes. Issue 47490: ~ Bad interactions can occur with NetworkManager scripts when it sees new VFs with random MAC addresses. Alternatively use the older networking service/scripts. Known Issues with Solarflare driver/firmware -------------------------------------------- SWNETLINUX-5350 / SWNETLINUX-5126: ~ When enabling the "separate_tx_channels" module parameter in combination with sfptpd, user may experience reduced timestamping accuracy due to the system falling back to software timestamps instead of using hardware timestamps. SWNETLINUX-5070: ~ Due to a race between the adjustment of ring sizes and NIC reset it is recommended to avoid using ethtool during boot. For more details see: SWNETLINUX-4855: ~ For some overlay network encapsulation offloads the NIC must maintain a list of UDP ports used for overlays. Due to EF10 hardware limitations, some changes to this list requires a full datapath reset, triggered by a firmware initiated MC reboot. ~ Applications like OvS can update UDP ports (that are used for overlay networks) in quick succession. This in turn can lead to multiple MC reboots being triggered from Firmware. ~ MC reboots require full datapath resets and drivers would need to recreate their queues and filters. EF10 detects these in flight resets and marks tunnels as dirty. It could happen that tunnels are marked dirty and never correctly configured. ~ This situation is exacerbated by newer kernels which backfill slots in the UDP port list array and when operating on the boundary of our offload capabilities (max 16 UDP ports for overlays). Issue 4991: ~ Message "Periodic DMA statistics are not updating. Restarting periodic collection" might appear in systems logs. This can be ignored SF-119015-KI / Issue 72914: ~ [SFN7xxx series firmware only]. There is an issue when using MCFW where a transmit queue using TSO may lockup and cause a NETDEV watchdog. ~ In the default configuration of "pf-count=1" (where there is no NIC partitioning) the NETDEV watchdog will occur on TX queue >=31, this will only be used if there are more than 32 logical CPUs in the machine, or if the "rss_cpus" module parameter is explicitly set. ~ As a workaround please set the module parameter to at most "rss_cpus=31". ~ As a fix please update adapter firmware using the "sfutils" package from https://www.xilinx.com/support/download/nic-software-and-drivers.html#linux to MCFW >= v6.2.7.1000 Issue 50868: ~ If using a SFN7142Q adapter with adapter firmware v4.1.0.6742 and running sfupdate, after the update the two network interfaces from the second QSFP+ port will be in a disabled state. Please reload the driver to correct this. This issue has been resolved in adapter firmware >=v4.1.1.1022. Copyright --------- Copyright (C) 2019-2022, Xilinx, Inc. Copyright (C) 2022-2024, Advanced Micro Devices, Inc. All rights reserved. Use is subject to license terms.