ENA driver is supported on all FreeBSD releases starting from 12.0
The driver was verified on the following distributions:
Releases:
- FreeBSD 12.3
- FreeBSD 13.1
Development: +-----------+-------------+ | Branch | SHA | +-----------+-------------+ | stable/12 | dad7c072f48 | | stable/13 | 185159f77c9 | | HEAD | 8020c05683f | +-----------+-------------+
Bug Fixes
- Fix KASAN and KMSAN kernel sanitizers build errors, by changing the
first_interruptvariable type frombooltouint8_t. - Fix LLQ descriptor reconfiguration after the device reset. This could cause driver misbehavior on the 6-generation AWS instances.
Minor Changes
- Align
req_idandqidlogging order. - Fix typo in a comment.
New Features
- Rework the Makefile to automatically detect the kernel options. This also changes the supported compilation options.
Bug Fixes
- Fix invalid KASSERT in the netmap code.
- Fix ENI metrics probing, by moving it to a separate task. That fixes the issue with the sleep inside the callout.
- Prevent LLQ initialization when membar isn't exposed.
Minor Changes
- Remove write-only variables.
- Code style improvements.
- Align the driver with the latest upstream kernel driver changes.
- Use the generic function for setting the device description in probe.
- Documentation update regarding the TSO.
- Use atomic accessors for the first_interrupt variable.
- Store last Tx cleanup ticks for easier device reset debugging.
- Logging improvements.
New Features
- Add NUMA awareness to the driver which runs on the kernel with enabled option RSS.
- Add support for the IPv6 L4 Tx checksum offloads.
Bug Fixes
- Increase lifetime of the timer service to work from the attach to detach in order to be able to handle more device failures.
- Do not trigger the reset if the ENA device is not responsive.
Minor Changes
- Optimize logic of the Tx path 'req_id' value validation.
- Rework Makefile in order to read the kernel build options from the SYSDIR. This change makes the requirement for manual Makefile modifications redundant.
New Features
- Add full kernel RSS option support.
- Add RSS hash key reconfiguration feature.
- Add RSS indirection table reconfiguration feature.
Bug Fixes
- Fix building with gcc9.
- Trigger device reset whenever ena_com_prepare_tx fails.
- Prevent potential race condition between reset routine and ena_destroy_device.
- Add global lock for all driver instances. This fixes a race between sysctl procedures and driver reconfiguration routines.
- Disable meta descriptor caching when LLQ is being used. This fixes netmap support on ENAv3 hardware.
Minor Changes
- Avoid unnecessary mbuf collapse for fragmented mbufs when driver is working in the LLQ mode.
- Add extra log messages.
- Add locking assertions for critical sections.
- Restructure code, by moving RSS related logic to separate source files.
- Add extra statistics counters - csum_good for properly calculated checksum and unmask_interrupt_num for the number of interrupt unmasking events.
New Features
- Large LLQ headers support. In order to use LLQ with packet headers greater than 96B, the large LLQ headers should be enabled by modifying the sysctl node: hw.ena.force_large_llq_header.
- Rework logging system, to be able to compile out logs on the IO path and simplify the logging levels.
Bug Fixes
- Hide sysctl nodes for the unused queues.
Minor Changes
- Style adjustments.
- Convert README document into reStructuredText format and update the installation section regarding git.
- Remove surplus NULL checks when freeing resources.
Bug Fixes
- Fix resource allocation if the MSIx vector table is not sharing the BAR with the registers. FreeBSD requires all resources to be allocated by the driver, so having unavailable BAR with MSIx vector table on it, will prevent PCI code from allocating MSIx interrupts.
New Features
- Add ENI (Elastic Network Interface) metrics support. They can be enabled and accessed using the sysctl.
- Add support for the Rx offsets (HW feature)
Bug Fixes
- Fix driver when running on kernel with RSS option being enabled (please note, that the standalone driver should also be built with 'CFLAGS += -DRSS' option in the Makefile).
- Fix alignment of the ena_com_io_cq descriptors.
- Fix validation of the Rx requested ID.
Minor Changes
- Add SPDX license identifiers.
- Add description of all driver tunables to the README file.
- Fix names/description of the device IDs supported by the driver.
- Update HAL to the newer version.
- Remove deprecated ENA_WARN macro.
New Features
- Add 'Tx drops' statistic which can be acquired by probing sysctl: dev.ena.#.hw_stats.tx_drops Where # stands for the interface number (0, 1, 2, ...).
- Mark the driver as epoch ready, to properly support latest FreeBSD 13 stack changes.
- Allow changing the number of IO queues during runtime by using sysctl: dev.ena.#.io_queues_nb Where # stands for the interface number (0, 1, 2, ...).
- Generate default RSS key using RNG in order to improve security.
Bug Fixes
- Fix ENA IO memory read/write macros - add appropriate barriers where needed and remove them where they are no longer needed.
- Tx DMA mapping in case of the LLQ is being used was totally reworked to improve stability and simplicity.
- Rx refill optimization for low memory conditions - by default page size clusters are now being used. It can be reverted, as it may affect performance a bit, by setting sysctl "hw.ena.enable_9k_mbufs" to "1" in "/boot/loader.conf" file.
- Prevent driver from limiting IO queue number in case number of MSIx vectors changed between device resets.
- Update 'make clean' to remove 'opt_global.h' which started to appear on FreeBSD 13.
- Fix build for arm64 FreeBSD 12 as pmap_change_attr() isn't public there, yet
Minor Changes
- Update HAL to the newer version.
- Driver version can now be acquired probing sysctl: hw.ena.driver_version
- Improve locking on the configuration path - the driver was using two different locks and now it is unified.
- Rework Rx queue size configuration to do not use hardcoded values and to avoid device reset upon doing so.
- Tx buffer ring size reconfiguration was also reworked for the same purpose.
- Create queues with optional backoff in case the OS cannot provide enough resources while setting up the queues.
- Allow to disable meta caching for Tx path - it may be requested by the device.
- Set all sysctl nodes and procedures as MP safe.
New Features
- Initialize cleanup task as NET_TASK - it is required by the latest stack updates as NET_TASK is entering epoch mode upon execution.
Bug Fixes
- Fix race on multiple queues when LLQ is being enabled with Tx checksum offload. It could lead to data corruption.
- Fix compilation for gcc by defining validate_rx_req_id() as 'static inline'.
New Features
- Add netmap support.
- Add LLQ support for the aarch64 by enabling WC.
Bug Fixes
- Fix infinite missing keep alive reset loop in case the reset routine takes too long because of the host being overloaded.
Minor Changes
- Split controller and datapath code into separate files.
New Features
- LLQ (Low Latency Queue) feature was added. New placement policy enables storing Tx descriptor rings and packet headers in the device memory. LLQs improve average and tail latencies. As part of LLQs enablement new admin capability was defined, device memory was mapped as write-combined and the transmit processing flow was updated.
- Rx checksum offloads can now be used on some of the HW.
- Optimization of the locks on both Tx and Rx paths. Rx locks were removed and Tx locks were optimized by adding queue management (stop and start).
- Additional Tx doorbells checks were added - supported only by some of the hardware.
Bug Fixes
- Suppress excessive error printouts in Tx hot path.
- Fix validation of the Rx OOO completion.
- Prevent double activation of admin interrupt (issue on A1).
- The reset is being triggered if too many Rx descriptors were received.
- The vaddr and paddr must be zeroed if allocation of the coherent memory fails using bus_dma_* API.
- The error is being set if allocation of IO interrupts fails.
- Driver now correctly supports partial MSI-x allocation - previously it was assuming, that the OS will always give enough MSI-x if the allocation won't return error.
- The ifp must be released detached before the last ena_down is called, to make sure ena_up will not be called once the ifp is released.
- The ifp should be allocated at the very end of the attach() function, as ifp cleanup is not working properly from the attach context.
- The validation of the Tx req_id was fixed.
- Error handling upon ENA reset is now fixed - it was causing memory leak and device failure.
- Tx offloads for fragmented pkt headers are fixed.
- Add check for msix_entries in case reset fails to prevent NULL pointer reference in ena_up().
- The driver is now checking for missing MSI-x.
- DMA synchronization calls were added to the Tx and Rx paths.
Minor Changes
- Add module PNP information.
- Update HAL to the newest version for ENAv2 driver.
- Logging of the ENA admin errors.
- Kernel RSS support was removed, as it was not compatible with ENA device.
- The host info structure is now receiving information about number of CPUs on the host as well as bdf (hints for the device).
- Rx refill threshold was limited, due to maximum number of Rx ring increased to 8k.
- ENA reset is now split into restore and destroy stages.
- New line characters in printouts were fixed.
- Make the reset triggering safer, by checking if the reset was already triggered.
- Checking for missing Tx completion was reworked for better readability and compatibility with above feature.
- AENQ notification handler was added.
- Device state is now being tracked using bitfields.
Bug Fixes
- Fix a mutex not owned ASSERT panic in ENA control path
- Skip setting MTU for ENA if it is not changing
- Do not pass header length to the NIC, as extracting it from the mbuf can be complicated in some cases
- Change BIT macro to work on unsigned value
New Features
- Update ena_com HAL to newest version
- Add reset reasons when triggering reset of the device
- Cover code with branch predictioning statements
- Rework counting of HW statistics
- Add RX OOO completion
- Check for RX ring state to prevent from stall - run rx cleanup in another thread if ring was stalled
- Allow logging system to be dynamically adjusted from the userspace from sysctl (hw.ena.log_level)
- Read max MTU value from the device
- Allow usage of more RX descriptors than 1
Bug Fixes
- Reset device after attach fails
- In the reset task. free management irq after calling ena_down. Admin queue can still be used before ena_down is called, or when it is being handled
- Do not reset device if ena_reset_task fails
- Move call of the ena_com_dev_reset() to the ena_down() routine - it should be called only if interface was up
- Fix error handling in attach
- Destroy admin queue resources after freeing interrupts (resources are used by the interrupt handler)
- Fix error handling in many places of the driver
- Lock drbr_free() call - it must be called with lock acquired
- Allow partial MSI-x allocation - update number of queues respectively
- Fix checking for the value of the DF flag
- Fix warnings that appeared when compiling driver with gcc
- Fix comparing L3 type with L4 enum on setting RX hash
- Fix calculating IO queues number
- Fix setting AENQ group to indicate only implemented handlers
Minor Changes
- Use different function for checking empty space on the sq ring (ena-com API change)
- Fix typo on ENA_TX_CLEANUP_THRESHOLD
- Change checking for EPERM with EOPNOTSUPP - change in the ena-com API
- Remove deprecated debugging methods
- Style fixes to align driver with FreeBSD style guide
- Remove deprecated and unsused counters
- Remove unsused macros and fields from ENA header file
- Update README file
- Remove NULL checks for malloc with M_WAITOK
- Do not check conditions explicitly for bool variables
New Features
- Add mbuf collapse feature in case the mbuf has more segments than the device can handle
Bug Fixes
- Remove unsused statement from the platform file to enable gcc compilation with standard warning flags
- Fix error check for Rx mbuf allocation
- Fix creation of the DMA tags and TSO settings
- Call drbr_advance() before leaving Tx routine to prevent panic
- Unmask all IO irqs after driver state is set as running, otherwise the interrupts can stay masked if they were in that state during interface initialization routine
- Acquire locks before calling drbr_flush() as it is required by the API
- Add missing lock upon initialization of the interface
- Add additional locks when releasing Tx resources and bufs
- Move hw stats updating routine to separate task to prevent from sleeping while unsleepable lock is held
- Add error handling if init of the reset task fails
- Add locks before each ena_up and ena_down call
Minor Changes
- Remove Rx mutex as it is no longer needed
New Features
- Acquire dma transaction width from the NIC
- Allow to control level of printouts from the Makefile
- Add watchdog service which is running periodically checks for:
- missing keep alive
- admin queue state
- missing tx completions
- if one of above fails, the reset is triggered
- additional sysctl statistics were added
- Add reset task
- Add fast path for the TX - if drbr is empty, send packet in current thread instead of creating task
- Add additional TX cleanup calls on TX path if queue is almost full
- Use single ithread routine for cleanup of both TX and RX queues - it allows synchronization of the TX and RX queues
- Add host info feature - driver is sending information about OS to the device
Bug Fixes
- Initialize part of the io queues resources during attach to prevent from race condition during ena_down() call
- Fix freeing RX counters twice
- Unmask all irqs when creating queues - this fixes issue with missing MSI-x when reloading interface during heavy network traffic
- Add locks in ioctl call to prevent from resource concurrency when handling simultaneous up/down of the interface
- If sending packet to the NIC failed, leave transmitting loop instead of trying until success - this could cause infinite loop
- Fixed RSS handling mechanism by properly assigning flag to the mbuf depending on type of the packet
- Defer creation of the RSS to allow load ENA driver during boot
- Fix error handling paths:
- ena_enable_msix: possible memory leak if pci_alloc_msix fails
- ena_create_io_queues: memory already allocated for part of tx/rx queues is not released on error
- ena_setup_rx_resources: lack of error check after initialization of enqueue task
- ena_request_mgmnt_irq: ignored fail of irq resource activation
- ena_sysctl_(dump|update)_stats: ignored errors returned by functions
- ena_reset_task: kernel page fault happens after failed reset
- Add locks in qflush to prevent race with ena_start_xmit()
- Add additional checks for driver status to prevent driver from hanging in the infinite loop on TX path
- Add error checks after malloc with M_NOWAIT
- Add lock for the reset task to prevent conflict with ioctl
- Move BPF_MTAP to track packet only if it was successfully sent to the NIC
Minor Changes
- Update README instructions for installing driver
- Update description of the RX and TX path to reflect current state of the driver
- Update license to the 2-clause BSD
- Move macros describing version of the driver to the ena.h file
- Change printouts levels meaning and place from which some of them are called
- Use one RX buffer size for all MTU settings
- Remove link speed change code - feature is not supported by the device
- Align version system with Linux driver
- Upgrade ena_com to 1.1.4.1
- Reduce amount of DMA tags in use - one per type of the queue instead one per queue
- Remove transmit modes
- Remove vlan support
- Check for number of fragments before doing DMA mapping
- Synchronize device initialization routine with Linux driver
- Simplify TX cleanup routine loop
- Reword descriptions of the sysctls
- Remove references to NETMAP
- Added budget to TX cleanup
Initial commit of the FreeBSD driver.