Release notes¶
Changes by version to the ha.ko driver and to the operational tools and packages distributed with it. These notes cover operational features only; changes to the application interface are in the SDK release notes. The latest version is at the top.
2.6.0¶
Released 2026-11-11. Makes copies between the host and a card work in a VM guest, adds ha-wipe for clearing a card before it goes to the next user, and changes the verdict for a card whose reset failed and the path masking in ha-hotreset.
New¶
tw_window_sizeanddmabuf_exportsnodes.tw_window_sizeshows the size of one translation window, anddmabuf_exportsshows the dma-buf export windows reserved so that another device such as an RDMA NIC can access the card's DRAM directly, with the processes that requested the exports. They also appear in the-qand--jsonoutput ofhyperaccel-smi. See sysfs nodes.ha-wipe. A tool that resets a card withha-hotreset, fills all allocatable DRAM with zeros, and reads sample blocks back, before the card goes to the next user. The deb package andmake installinstall it as/usr/bin/ha-wipe. It does not clear SRAM. See ha-wipe.ha_hdma_cpu_latency_us(default0). Places a CPU wake-up latency request while the DMA engines copy and for 10 ms after, so that the thread waiting for a copy does not wake late from a deep C-state. The default blocks every C-state during that time. With-1, no request is placed. See Parameter reference.ha_hdma_completion(defaultauto). Selects how the driver receives DMA engine completion.autouses interrupts (MSI-X) only when bare metal is confirmed and polling otherwise;irqandpolloverride that decision. Set it topollwhen the driver mistakes a VM for bare metal and every copy fails. See Parameter reference.ha_ready_timeout_ms(default3000). Sets how long (ms) the driver waits after a reset for the firmware to report ready. Before, this was fixed at 3 seconds. If the kernel log getsdevice came back but firmware did not report readyafter a reset andfw_versionshows a version a little later, raise the value. See Parameter reference.ha_host_mem_max_mb(default1024). A cap (MiB) on the total host memory that applications allocate through the driver for cards to read and write directly. Allocations from all cards and processes count together, and an allocation past the cap fails.0means no cap. See Parameter reference.
Changes¶
- Copies between the host and a card work in a VM guest. When a card was passed through to a VM, the DMA engine's completion interrupt did not reach the driver, so every copy failed after 5 seconds with
ll: timeout. Now, unless bare metal is confirmed, the driver polls the completion mark that the engine leaves in host memory. TheHDMA completion: <mode>line logged when the driver loads shows which method is used, and with polling the per-engineHDMA … MSI-Xlines are not logged. A transfer whose completion never arrives is reported by the new linell: engine stopped but its completion <mode> never arrived. x86 bare-metal servers keep using interrupts. An arm64 server whose kernel runs at EL1 switches to polling even on bare metal. On such a server, setha_hdma_completion=irqto receive completion by interrupt as before and not spend CPU on polling. See Kernel log messages. - A card-to-card copy that crosses a translation-window boundary is now transferred correctly. Before, if a piece went beyond the range covered by one translation window of the peer card (
tw_window_size), the overflowing part reached a neighboring region. A read returned another region's data as a success, and a write could overwrite the memory of another process. Now the copy is cut at the boundary and transferred one window at a time. If the peer card is being reset or is FATAL, the copy is not started and ends with an error. See Multiple devices. - A card whose card-side BAR2 address does not arrive is not bound. The driver sets up its access windows at the card-side BAR2 address reported by the firmware. If this address does not arrive within 1 second at probe time, the driver creates no
/dev/haNfor that card and, depending on whether the firmware booted, logsfirmware booted but did not publish the BAR2 inbound base …orfirmware has not booted, so the BAR2 inbound base is unknown …. See Errors that stop probe. - Card numbers and device numbers survive a module reload. When the module is loaded, the driver numbers the installed cards in PCI address order, and the number of a card whose probe fails stays empty. All cards share one major device number, and the minor is the card number. See Card numbers.
- The
pool_barparameter has been removed. The access windows are always on BAR2. faultedalso stays after a failed SDK reset. Before,healthstayedfaultedonly when an automatic reset failed. Now it also staysfaultedwhen a reset requested through the SDK fails, and automatic reset stops for that card.- Guidance after a failed reset. When a reset does not recover the card, the kernel log gets a
device reset failed (...)line, which tells you whether to reset again, runha-hotreseton the host, or power-cycle the host. A card for which the driver refused the SBR because its parent port does not mask Surprise Down can now be reset withha-hotresetwithout any option. See If the reset fails. ha-hotresetmasks the path by default. A card whose parent port is not a root port and does not mask Surprise Down is now reset too: the tool masks the path from the parent port up to the root port during the reset and restores it when the card responds again. Before, such a card was refused, and the path was masked only with--allow-switch-link-drop. With--no-switch-link-drop, the tool refuses such a card as before.--allow-switch-link-dropis accepted for compatibility but has no effect. See Path masking.- Exit codes for refusals in
ha-hotreset. A refusal that touched nothing now has its own exit code:3when a device node is open or the wait for anotherha-hotresetrun runs out, and4when an SBR is not safe at this card's PCIe position. Before, an open node and a refused SBR both ended with1. ha-hotresetfails when the driver does not bind the card. For a card that had no/dev/haNbefore the reset, the tool now also checks that thehadriver bound it after the reset. If it did not, the tool printsthe card was reset but the ha driver did not bind itand ends with exit code1. Previously it printedHot reset OKfor such a card. See ha-hotreset.ha-hotresetruns one at a time on a host. The ports it masks are shared by the cards below them. A second run waits up to 600 seconds for the first to end.ha-fw-updatesays so first when its reset step has to wait.- Interrupt handling in
ha-hotreset. If the tool fails or is interrupted after removing the card, it has the card probed again and restores the masking only if the card responds. If the card does not respond, it leaves the masking in place and prints thesetpcicommands to restore it by hand. Before, it restored the masking without checking that the card responds. - The driver's command protocol is now v9. When the driver recognizes a card, the kernel log gets
ABI protocol v<N> (driver v9, min v6). The minimum version for connecting is still v6, so the firmware does not need an update. See Supported environments. - Operation on arm64 and RHEL 9 is confirmed. It was confirmed on Ubuntu 24.04 arm64 (4K page kernel) and on RHEL 9.8 (x86_64). Before, arm64 was under verification and the RHEL 9 family was planned for verification. See Supported environments.
- The automatic-reset kernel log messages are now in English. These are the start line (
auto-TDR: the device went FATAL; resetting it), the failure line (auto-TDR: reset failed (<err>); ...), and the removal-during-reset line (auto-TDR: this instance was removed during the reset (<err>); ...). If a script searches the kernel log for the earlier Korean wording, change it to the new wording. - The AEXP kernel log messages changed. The three lines that start with
AEXP:now saybootloader default. Update anything that matches these lines. - Kernel log cleanup. The
wait_boot: elapsed=…line, which was logged every 50 ms while waiting for the boot signal, was lowered to debug level so that it no longer appears in the default kernel log. - The
helpnode no longer gives the SRAM address formula forpeer_sram. See Multiple devices for the formula. - The deb package depends on
pciutils.ha-hotreset, andha-fw-updateandha-wipethrough it, usesetpciandlspci.make installfrom the DKMS tarball also installspciutilsalong with the other packages it needs. - The deb package announces the driver interface version.
hyperdriver-dkmsprovideshyperaccel-kmd-uapi(the version this driver implements) andhyperaccel-kmd-uapi-min(the lowest version this driver still supports). SDK packages use them to require a driver they can work with. - The hot reset script in the
.tar.gzpackages is renamed.hotreset.shis nowha-hotreset.sh, andha-fw-updateandha-wipelook for that name in their own directory. The installed command is stillha-hotreset. - The driver source is licensed under GPL-2.0-only. The
hakernel module source carries the license marks, and the packages include the full text asCOPYING. In the DKMS source, all headers are now underinclude/, andsrc/holds only.cfiles.
2.5.0¶
Released 2026-10-19. Handles copies between the host and a card with several DMA engines, adds ha-led for finding a card in the chassis, and lets the driver build on arm64 and on newer kernels.
New¶
- Copies between the host and a card are handled by several DMA engines. Several engines per direction on each card process a copy in parallel, and preparing host pages overlaps with transfers, so throughput has increased. Previously, copies smaller than 4 KiB were processed by CPU copy (PIO). Now sending every copy through the DMA engines regardless of its size is the default (
ha_hdma_min_bytes, default0); if you giveN, only copies smaller thanNbytes are processed by CPU copy. See the Parameter reference. ha-led. A tool that blinks the board LED of one card so that you can find the slot the card is in within the chassis. You specify, by PCI address, the card thathyperaccel-smishows asBus-Id. The deb package andmake installinstall it as/usr/bin/ha-led. It works with firmware ABI protocol 8 or later, and the board must have an LED. See ha-led.- arm64 build. You can build the driver with DKMS on an arm64 kernel. The driver supports only 4K-page kernels, and DKMS excludes 16K and 64K page-size kernels from the build targets.
abi_prefetch_statsnode. Shows the running totals ofha_abi_prefetchdeterminations on one line. It contains the number of times a copy was prefetched (issued), the number of times its result was used (consumed), the number of refusals by reason (refused_*), and so on, and you use it to find the cause when prefetching does not happen as much as expected.
Changes¶
- Building on newer kernels. The driver is adapted to the kernel API changes in kernel 6.3 and later and in RHEL 9.4 and later kernels. For the kernels on which the build has been verified, see Supported environments.
make installverifies the build first. Themake installof the DKMS tarball checks in a temporary DKMS tree whether the new version builds on the running kernel, and only then replaces the existing driver. If it does not build, it stops without touching the installed driver.- Installation guidance from the DKMS tarball
make install. - On the RHEL family (RHEL, CentOS Stream, AlmaLinux, Rocky Linux) and Fedora, if the required packages cannot be installed, it also prints an action for each dnf error message. ForNo match for argument: dkms, it prints the command that enables EPEL (RHEL family only), and forNo match for argument: kernel-devel-..., it advises you to update the kernel, reboot, and install again. - On EL9 and later and on Fedora, it also installskernel-devel-matched, which matches the newest installed kernel. It does not pull a new kernel from the repository. - Ifmodprobe hafails, it prints the command that enrolls the DKMS signing key in the MOK under Secure Boot. - The deb package keeps the existing driver even if the build fails.
hyperdriver-dkmsfirst tries to build the new version on the running kernel. If it does not build, the installation ends in failure and prints the currently loaded version, the installed version, and the location of the build log. Removing the package deletes all versions registered with DKMS. - The card becomes FATAL when a DMA engine does not respond or does not stop. The device becomes FATAL when the card does not respond to the status register of a DMA engine (
ll: no response), or when an engine does not stop after a timeout or abort (HDMA engine at …). Before, even when an engine could not be stopped, only a warning (ch <ch> still running after force stop) was logged and only that transfer failed. Now the host pages that the engine may still write to are not released but quarantined, and new DMA transfers are blocked until a reset stops the engine. See FATAL and recovery. - The kernel log wording for DMA and translation windows changed.
ch=<ch>in the DMA lines becamechannel=<ch> <direction>(ll: timeout (n=<n> channel=<ch> <direction>)), andTW: every pool window is pinned; … failed, logged when translation windows run short, becameTW: no free pool window (…); … failed, which shows by count what is holding the windows. If you have scripts that search the kernel log for the old wording, change them to the new wording. See Kernel log messages. - Kernel log cleanup. Messages that used to be logged for every interrupt and every command (such as
MSI-X recv) were lowered to debug level so that they no longer appear in the default kernel log. The firmware error dump (Firmware error: resq[...]) is logged at most 10 times in 5 seconds, all cards combined.devcmd_fw_errorscounts every error without omission. - Stricter
ha_abi_prefetchdetermination. If a preceding task in the same stream reads or writes the destination buffer, that copy is not prefetched. A copy that is not prefetched runs in the original order. - Value check for
--bdfinha-hotreset. If you give no PCI address after--bdf, or give an empty value as in--bdf=, it prints--bdf needs a PCI addressand the usage and ends with exit code 2. Previously, a--bdfwithout a value ended with exit code 1, and an empty--bdf=selected the card when there was only one card. - Reset-refusal message while waiting for a firmware update response.
reset refused: last fw_update unanswered …now advises you to rerun the update after the firmware has responded, and to contact the support team if that does not help. The previous message suggested reloading the module,ha-hotreset --force, and reapplying power.
When upgrading from a deb package of 2.4.0 or earlier
The behavior of keeping the existing driver when the build fails applies from the upgrade after the 2.5.0 package is installed. Packages of 2.4.0 and earlier remove the installed driver first when the upgrade starts, so if the new version does not build, no installed driver remains. Before upgrading, check that the headers of the running kernel are installed and that the kernel is listed in Supported environments.
2.4.0¶
Released 2026-09-22. Adds the firmware update tool and changes the conditions under which a reset issues an SBR (Secondary Bus Reset).
Firmware ABI protocol 6 or later is required
The minimum firmware ABI protocol version that the 2.4.0 driver requires is 6 (it was 1 from 2.3.1 through 2.3.6). If the firmware advertises a lower protocol, the driver does not register that card and logs ABI protocol mismatch: firmware vN, driver requires >= v6; refusing probe in the kernel log. Check the protocol of the firmware you are using from the N in ABI protocol vN (driver vM, min vK), which is logged in the kernel log at probe (the stage in which the driver probes the card). Before upgrading, check that N is 6 or later. Drivers from 2.3.1 onward log this line. If this line is missing, or predates the ABI protocol handshake is logged instead, see the firmware entry in Supported environments.
New¶
ha-fw-update. A tool that writes a firmware image to the card's boot flash through the running firmware, then performs a hot reset of the card so that it boots with the new firmware. The deb package andmake installinstall it as/usr/bin/ha-fw-update. It works with firmware ABI protocol 7 or later, and while an update is running, other tasks on that card and resets (including automatic resets) are rejected. For the procedure and safety rules, see Firmware update.fw_update_progressnode. Shows firmware update progress as<done> <total>in units of 4 KiB sectors. It isunknownon firmware that does not support updates.- Card-to-card SRAM path and the
peer_sramnode. At probe, the driver sets up, for each card, a path to the SRAM of the other LPUs, and sets it up again after a reset.peer_sramshows, one line per LPU, the base address that a kernel running on this card uses to access each LPU's SRAM. ha_abi_prefetch(on by default). Starts copying in advance the ABI data that is sent from the host to the device before the next kernel runs, while the preceding kernel is running. With0, nothing is prefetched. It can be changed at run time.
Changes¶
- SBR conditions. The SDK device reset and the automatic reset proceed to SBR when firmware re-initialization fails. They now issue an SBR only when the parent port is a root port or the Surprise Down Error is masked on the parent port. This is the same criterion as
ha-hotreset. Up to 2.3.6, an SBR was issued unless the parent was a PCIe switch downstream port. When an SBR is refused, the reason is logged in the kernel log, and when an SBR is issued,issuing SBR via parent <BDF>is logged. - Reopen the device after a reset. When a device is reset (SDK device reset, automatic reset, PCIe error recovery, and so on), you can no longer work with a handle that was opened before the reset. Applications must reopen the device. The handle that requested the reset through the SDK can continue to be used.
- Polling stops for a FATAL device. While a device is FATAL, telemetry polling and watchdog determinations stop until it is reset. During this time
healthisfaulted, and polled values such as temperature and power becomeunknownonce the last sample is more than 5 seconds old. - Card numbers are kept. While the module stays loaded, if a card with the same PCI address is probed again (for example, re-enumeration during a reset), it gets the same
haNnumber as before, provided no other card took that number in the meantime. For the rules, see Card numbers. - Watchdog.
- If the configuration cannot catch a stall because the threshold plus the observation interval is greater than or equal to the wait ceiling, a warning is logged in the kernel log. The values are not changed.
- A hang dump is logged only once per hang.
- When a command that waits with no ceiling takes a long time, the batch progress and verdict at that point are recorded in
last_hang. - hang dump. First logs, in a single human-readable line (
hang (...)), the name of the last firmware command sent, the verdict (when a verdict was reached because the response ceiling passed), and the process (PID and name) that opened the device handle that sent the command. The stream value in the machine-readablehang dumpline is not a user-space address but a value obscured with a key that differs on every boot. last_hang. Anrc,cmd_idx, oropcodethat has never been recorded is shown as-instead of0.hyperaccel-smialso reads and shows these values.- Kernel log cleanup. The messages that were logged on every device memory allocation and release were lowered to debug level.
2.3.6¶
Released 2026-09-09. Adds the watchdog, hang diagnostics, and PCIe error recovery.
New¶
- Watchdog. Periodically observes, from outside, the waits for device commands and DMA transfers, and when a threshold is exceeded, logs
watchdog: ... wait stalledin the kernel log. With the default values, it only reports and does not cut off the wait. -ha_wd_hang_ms(default8000): The threshold for device command waits.0turns it off. -ha_wd_hdma_hang_ms(default4000): The threshold for DMA transfer waits.0turns it off. -ha_wd_poll_ms(default500, minimum50): The observation interval. -ha_wd_preempt(default0): When on, cuts off a stalled wait. Cutting off the wait of an application task makes the device FATAL, so use it together withha_auto_reset=1. - hang dump. When it finds a device command wait for which the response ceiling has passed or which the watchdog has judged to be a stall, it logs one
hang dumpline in the kernel log. The line contains the last command sent and its body, the command queue progress, and the DMA engine state. When a command that waits with no ceiling takes a long time, the elapsed time and the process waiting for that command are also logged periodically. ha_devcmd_ceiling_ms(default12000, minimum100). The ceiling on how long the driver waits for a firmware response. If there is no response within this time, it is treated as a hang anddevcmd_timeoutsis incremented. If the driver was waiting for an application task, the device becomes FATAL.- Progress information and verdict in
last_hang. The batch progress at the time of the last hang (batch_seq,exec_index,resq_tail) and the verdict (verdict) were added. The verdict distinguishes whether the device got stuck in a command (stuck-in-command) or only missed the completion interrupt (completed-irq-lost), among other cases, and isunsupportedif the firmware does not provide progress information.hyperaccel-smi -qshows the verdict and the progress information together. peer_routenode. Shows, for each peer, the PCIe path between the cards asdirect(directly within a switch) orhost-mem(via the root complex). It is for observation only and does not block transfers.- PCIe error recovery. When the kernel runs PCIe error recovery (AER, DPC), the driver stops the tasks in progress, checks the firmware boot after recovery, and then accepts requests again.
ha_aer_enable(default0, set at load time only). Sets whether the driver turns on PCIe AER error reporting for the card. The kernel's recovery action for a fatal error reported through AER is a parent-bridge SBR, which can lead to a host restart depending on the configuration, so it is turned off by default. For differences by kernel version, see FATAL and recovery.ha_sysmon_poll_ms(default200, range50~2500). You can change the telemetry polling interval.0is not a value that turns it off but the shortest interval (50).ha_tw_pin_wait_ms(default12000, range0~60000). The ceiling on how long a device-to-device copy waits for a translation window slot of the other card. When the ceiling passes, that transfer fails and is logged in the kernel log. With0, it does not wait.
Changes¶
- Firmware readiness check. At probe and reset, the driver waits until the firmware reports that it is ready. Even if it does not, the card is registered, but telemetry read from the firmware is
unknownand application requests are rejected until it is ready. If the firmware becomes ready later, normal operation starts from then. - FATAL determination. If the error code that the firmware returns for an application task is not a code defined as failing only that command (including a code never seen before), the device is treated as FATAL. Up to 2.3.1, only no response, a firmware timeout, and a core fatal error were treated as FATAL.
- Telemetry during a reset. During a device reset, telemetry polling and sysfs reads that access the device are stopped. Polled values become
unknownonce the last sample is more than 5 seconds old. - Removing a device that has open handles. The driver handles device removal caused by a driver unbind or a hot reset even if processes that opened the device remain. Requests from the remaining processes fail, and accessing a device memory mapping can raise
SIGBUS. - Kernel log cleanup. Messages that were logged for every command (
devcmd stage,devcmd flush) were lowered to debug level. The response queue dump that was logged for every command is now logged only along with a firmware error (Firmware error: resq[...]).
2.3.1¶
Released 2026-08-20. Adds a per-card hot reset tool and protects telemetry from continuing to read a card that does not respond.
New¶
ha-hotreset. A tool that performs a PCIe hot reset of a single card. The deb package andmake installinstall it as/usr/bin/ha-hotreset. It removes only the specified card, issues an SBR on the parent bridge, and rescans, so the module stays loaded and other cards are not reset. The previous reset script unloaded the module and detached all cards. - If the parent port is not a root port and does not mask Surprise Down Error, if the card's device node is open, or if there is another device under the parent bridge, it stops without touching anything. Some of these checks can be bypassed with options. ---allchecks all cards first, and if even one card fails the check, no card is reset. - For the options, see ha-hotreset.peer_mapnode. Shows, for each peer, the mapping state of device-to-device copies from this card to other cards asmappedorunmapped. A copy in anunmappeddirection fails and is not redirected to another address. If there are no other cards, it isnone.pool_bar(default1, set at load time only). Sets the BAR that holds the windows through which the host accesses device memory. With the default1(BAR2), the feature in which another card accesses this card's SRAM directly can be used together with computation, and0(BAR0) is the previous behavior. In a configuration in which the firmware does not report the BAR2 address and the card-side PCIe address differs from the host's, it falls back to BAR0 and logsusing BAR0 poolin the kernel log.ha_sysmon_fail_limit(default15). If telemetry polling gets no response this many times in a row, polling stops and this is logged in the kernel log. This is because continuing to read a card that does not respond causes a PCIe Completion Timeout, and a firmware-first platform can reset the server because of it. While polling is stopped, polled values areunknown, and resetting the device restarts polling. With0, polling does not stop.- Firmware compatibility check. At probe, the ABI protocol version that the firmware advertises is logged in the kernel log as
ABI protocol vN (driver vM, min vK), and if it is lower than the driver's minimum version, the card is not registered.
Changes¶
fw_version. Showsunknownif the firmware version value reads as0or all-ones, or if the firmware boot has not been confirmed.hyperaccel-smi. In the table and the-qoutput, per-IP utilization shows onlycore,mpu,vpu,cdma, andcgdma. The--jsonoutput includes all items of theutilnode.
2.3.0¶
Released 2026-07-31. First release to provide sysfs telemetry, hwmon, hyperaccel-smi, automatic reset, and the deb package.
New¶
- sysfs telemetry. Added read-only nodes under
/sys/class/ha/haN/. They can be read without root privileges or a device handle, and they respond even while computation is stalled. When a sensor value cannot be trusted,unknownis shown. - Static information:fw_version,hw_version,serial,dram_size,sram_size- Memory usage:dram_used,dram_free,sram_used,sram_free- Sensors:temp,temp_avg,temp_blocks,power_mw,rails,util- Tasks and errors:inflight,stream_count,devcmd_timeouts,devcmd_fw_errors,hdma_timeouts,last_hang- State and description:health(ok,degraded,faulted,unknown),help - hwmon. If the kernel supports hwmon, sensors are registered under the name
ha. lm-sensors or node_exporter can read 15 temperature values (peak, average, and 13 blocks), the voltage and current of 4 rails, and the total power with no additional configuration. hyperaccel-smi. A command that shows sysfs telemetry as a table.make installand the deb package install it as/usr/bin/hyperaccel-smi. It supports-q(detailed),--json,-l(repeat), and-i(specify a card), and you can limit the cards it shows withHYPERACCEL_VISIBLE_DEVICES. It returns exit code2when no LPU is visible, so it can be used for health checks. Python 3 is required.ha_auto_reset(default0). When on, the driver resets a device automatically when it becomes FATAL. If the automatic reset fails, it is not retried until a manual reset succeeds, andhealthshowsfaulted.- deb package
hyperdriver-dkms. Installs the DKMS sources, the udev rules, the setting that loads the module automatically at boot, andhyperaccel-smi. It builds the driver for the running kernel and for every other kernel whose headers are installed. If the headers of the running kernel are missing, it aborts the installation, and it warns if, after an upgrade, the previous driver remains loaded because the device is in use. It does not include the CDI spec generator.
Changes¶
- Reset procedure. The SDK device reset stops and cleans up tasks in progress, then first tries firmware re-initialization, and proceeds to SBR only if that fails. If the parent is a PCIe switch downstream port, it refuses the SBR to prevent a host failure. Automatic reset follows the same procedure.
- User-space library split. The driver package no longer installs
libha_driver(the library, header, and pkg-config file). This library is installed with the SDK. - Cleanup scope of the prebuilt tarball. The
make installandmake removeof the prebuilt tarball also deleteha.koinstalled at other paths or through DKMS, as well as the DKMS registration.
Version list¶
| Version | Release date |
|---|---|
| 2.6.0 | 2026-11-11 |
| 2.5.0 | 2026-10-19 |
| 2.4.0 | 2026-09-22 |
| 2.3.6 | 2026-09-09 |
| 2.3.1 | 2026-08-20 |
| 2.3.0 | 2026-07-31 |
Check the version of the loaded driver with cat /sys/module/ha/version, and the version of the module file installed for the running kernel with modinfo -F version ha.