CVE-2026-99248 covers a runtime PM teardown that runs in the wrong order in the Xilinx I2C controller driver, i2c-xiic. It affects kernels from 7.1 to 7.2.12; the fix is in 7.2.13 and 7.3, and older kernels never had the code. On an affected kernel, each removal of the driver disables the controller clock twice and produces a kernel warning.
CVE-2026-99248 was published on 3 November 2026 by the Linux kernel CNA for the Xilinx I2C controller driver, drivers/i2c/busses/i2c-xiic.c. It describes a runtime PM teardown that runs in the wrong order when the driver is removed from its device. Three questions decide whether it matters to your product: is your kernel between 7.1 and 7.2.12, does your image enable CONFIG_I2C_XILINX, and does your tree carry the fix from 7.2.13 and 7.3. The bug class is one that any driver author can write: a driver was converted to managed (devm) helpers, and those helpers release resources in an order that the clock state did not allow.
- Component:
drivers/i2c/busses/i2c-xiic.c, functionsxiic_i2c_probe()andxiic_i2c_remove(); built byCONFIG_I2C_XILINX. - Introduced in: mainline 7.1, by commit 50c63491ff26 (“i2c: xiic: switch to devres managed APIs”).
- Fixed in: mainline 7.3 (commit d33373c565af) and stable 7.2.13 (commit 47e105bdc863). The 7.1.y series is not fixed.
- Not affected: 7.0 and older, including the 6.18, 6.12, 6.6, 6.1, 5.15 and 5.10 longterm series.
- Severity: no CVSS score has been published. Checked on 17 November 2026: the announcement has no severity section, the CNA container on cve.org has no metrics, and NVD shows the record as “Received”, with no metrics and no analysis of its own.
- Reached by: removal of the driver from the device, when the runtime PM teardown runs.
- Effect: the controller clock is disabled twice, and
clk_core_disable()prints a WARN on each removal. - Exploitation: not listed in the CISA Known Exploited Vulnerabilities catalogue (version 2026.11.13, checked on 17 November 2026).
What the code does
xiic_i2c_probe() sets up the Xilinx I2C controller. It obtains and enables the clock with devm_clk_get_enabled(), configures runtime PM with an autosuspend delay, requests the interrupt and registers the I2C adapter. xiic_i2c_remove() runs when the driver is detached: it deletes the adapter, resumes the device, calls xiic_deinit() and drops the runtime PM reference.
Runtime PM lets the driver stop the clock while the bus is idle: the suspend callback, xiic_i2c_runtime_suspend(), calls clk_disable(). Because the driver uses autosuspend, a put arms a timer, and the suspend happens when the delay expires.
The clock enable count therefore has two users: devm_clk_get_enabled(), which disables the clock when the managed resources are released, and runtime PM. The count stays balanced only if the device is runtime-active, with the clock on, when the managed clock release runs.
The bug: a runtime PM teardown in the wrong order
Commit 50c63491ff26, merged for 7.1, converted probe to managed helpers to simplify its error paths. The hand-written runtime PM setup became one managed call, and the runtime PM teardown lines were deleted from xiic_i2c_remove(). This is the probe code before the fix:
pm_runtime_set_autosuspend_delay(dev, XIIC_PM_TIMEOUT);
pm_runtime_use_autosuspend(dev);
ret = devm_pm_runtime_set_active_enabled(dev);
if (ret)
return ret;The managed helper registers a release action in the driver core. This is the action as it stands in drivers/base/power/runtime.c in 7.2.12:
static void pm_runtime_disable_action(void *data)
{
pm_runtime_dont_use_autosuspend(data);
pm_runtime_disable(data);
}Managed resources are released after remove() returns, in the reverse order of registration. The clock was registered before runtime PM, so its release runs after this action. The fix commit describes the sequence on removal:
xiic_i2c_remove()callspm_runtime_put_sync(). With autosuspend in use, the device remains in the active state and the autosuspend timer is armed.- The release action calls
pm_runtime_dont_use_autosuspend()while runtime PM is still enabled. With autosuspend cleared,rpm_idle()suspends the device at once, andxiic_i2c_runtime_suspend()callsclk_disable(). pm_runtime_disable()runs next, too late to prevent that suspend.- The
devm_clk_get_enabled()release callsclk_disable_unprepare()on a clock that is already disabled.
The commit message reports the result as a WARN from clk_core_disable(), with the text “clkN already disabled”, on every teardown. The record reports no memory corruption and no crash.
The commit identifies a second, narrower fault that a hand-written runtime PM teardown can also contain. Calling pm_runtime_put_sync() before pm_runtime_disable() arms the autosuspend timer; if the task is preempted before runtime PM is disabled, the timer can expire and the suspend callback disables the clock.
How it is reached
The path runs when the driver is unbound from the device, when the driver module is unloaded, or when the platform device is removed. Unbinding through sysfs and unloading a module are administrator operations. The CVE record and the fix commit state no attack vector and no remote trigger, and this article does not add one.
The configuration requirement is CONFIG_I2C_XILINX, built in or as a module. In 7.3 its only Kconfig dependency is HAS_IOMEM, but the driver binds only on boards whose design contains the controller.
The effect of each event is a warning with a stack trace in the kernel log. A kernel booted with panic_on_warn treats any WARN as a panic; that is general kernel behaviour, not a statement from the record. A product that never unbinds the driver does not run this path in normal operation.
The fix
Commit d33373c565af (“i2c: xiic: restore non-managed runtime PM to fix clk WARN flood”, dated 14 October 2026) changes one file and does three things. First, probe enables runtime PM by hand again:
/* before */
ret = devm_pm_runtime_set_active_enabled(dev);
if (ret)
return ret;
/* after */
ret = pm_runtime_set_active(dev);
if (ret) {
pm_runtime_dont_use_autosuspend(dev);
return ret;
}
pm_runtime_enable(dev);Second, remove performs the runtime PM teardown explicitly, in an order that cannot suspend the device:
/* before */
pm_runtime_put_sync(dev);
/* after */
pm_runtime_disable(dev);
pm_runtime_put_noidle(dev);
pm_runtime_set_suspended(dev);
pm_runtime_dont_use_autosuspend(dev);pm_runtime_disable() runs first, so no suspend callback can start after it. pm_runtime_put_noidle() then drops the usage count without an idle check, and the runtime PM core refuses to suspend a device whose runtime PM is disabled. The device keeps its clock on, which is the state the managed clock release expects, and that release disables the clock exactly once.
Third, the interrupt goes back to request_threaded_irq(), with an explicit free_irq(). The commit gives the reason: probe now unwinds runtime PM by hand, and a handler registered with a managed call would remain registered across that manual runtime PM teardown.
This is sufficient because runtime PM is now enabled once in probe and disabled once in remove, at points where the order relative to the clock is explicit.
The fix has not been amended: it is the newest change to the file in the 7.2.y and 7.3.y branches, and the only later change in mainline is an unrelated tree-wide conversion.
Which kernels are fixed
| Series | Status | Detail |
|---|---|---|
| Mainline | Fixed in 7.3 | Commit d33373c565af; affected from 7.1 |
| 7.3.y | Not affected | 7.3 was released with the fix |
| 7.2.y | Fixed in 7.2.13 | Commit 47e105bdc863; 7.2 to 7.2.12 affected; end of life at 7.2.15 |
| 7.1.y | Not fixed | The branch has the introducing commit and not the fix; no longer listed on kernel.org |
| 7.0 and older | Not affected | Per the CVE record; includes 6.18, 6.12, 6.6, 6.1, 5.15 and 5.10 |
The CVE was listed in the weekly round-up for 1 – 7 November 2026. Because 7.2 is end of life, the practical target for a product on 7.2.y or 7.1.y is 7.3.y, where the newest release is 7.3.4. A vendor BSP on an older kernel is affected only if the vendor backported the conversion commit.
Check your tree
With the kernel source in a directory named linux, search the git history of the file for the two commit subjects:
raghu@techveda.org:~$ git -C linux log --oneline --grep="i2c: xiic: switch to devres managed APIs" -- drivers/i2c/busses/i2c-xiic.c
raghu@techveda.org:~$ git -C linux log --oneline --grep="i2c: xiic: restore non-managed runtime PM" -- drivers/i2c/busses/i2c-xiic.cIf the first command prints a commit and the second prints nothing, the tree is affected. If both print a commit, the fix is present.
For a BSP delivered without history, search the file and the kernel configuration:
raghu@techveda.org:~$ grep -n "ret = devm_pm_runtime_set_active_enabled" linux/drivers/i2c/busses/i2c-xiic.c
raghu@techveda.org:~$ grep -n "Enable runtime PM by hand" linux/drivers/i2c/busses/i2c-xiic.c
raghu@techveda.org:~$ grep -n "CONFIG_I2C_XILINX" linux/.configA match for the first pattern means that the managed helper is still called, so the tree is affected. The second pattern is a comment that the fix adds in probe. If CONFIG_I2C_XILINX is not set, the file is not built and the CVE does not apply to your image.
The lesson: check the release order after a devm conversion
- Write down the release order. Managed resources are released in reverse order of registration, after
remove()returns. For each release action, check the state that the earlier actions leave behind. - Find enable counts with two users. A driver that calls
devm_clk_get_enabled()and also callsclk_disable()on the same clock in its runtime suspend callback is correct only if the device is runtime-active when the clock release runs. - Remember what autosuspend changes. A put arms a timer, and clearing autosuspend while runtime PM is enabled can suspend the device at once. In a hand-written runtime PM teardown, disable runtime PM first and then call
pm_runtime_put_noidle().
The warning appeared on every removal, so one bind and unbind cycle would have shown it. A conversion to managed helpers changes the runtime PM teardown, and it needs that test even when the diff only removes lines.
What was checked and what could not be verified
Read on 17 November 2026: the announcement on the linux-cve-announce list; the CVE record (last updated 3 November 2026, no metrics in the CNA container); the NVD entry (last modified 3 November 2026, status “Received”, no metrics); the fix commit, its 7.2.y backport and the introducing commit; the history of the file in the 7.1.y, 7.2.y and 7.3.y branches and in mainline; the runtime PM helper source in 7.2.12; the kernel.org release list; and the CISA catalogue.
Not verified: the fault was not reproduced on hardware, so no command output is shown. Longterm branches older than 7.0 were not opened; their status comes from the CVE record. Vendor trees were not examined.
Key takeaways
- CVE-2026-99248 affects kernels from 7.1 to 7.2.12 that build
CONFIG_I2C_XILINX; 7.2.13 and 7.3 are fixed, and 7.1.y is not. - The faulty runtime PM teardown runs only when the driver is removed, and the reported effect is a clock warning.
- No CVSS score has been published, and the CVE is not in the CISA catalogue.
- After a devm conversion, check the release order and test bind and unbind.
Frequently asked questions
Which kernel versions are affected by CVE-2026-99248?
Kernels from 7.1 to 7.2.12. The fix is in 7.2.13 and in 7.3. The 7.1.y series was not fixed. Kernels older than 7.1 do not contain the affected code.
Are the 6.12 and 6.18 longterm kernels affected?
No, according to the CVE record, which marks every version before 7.1 as unaffected. A vendor tree on an older kernel is affected only if it backported commit 50c63491ff26.
How do I check a BSP that has no git history?
Search drivers/i2c/busses/i2c-xiic.c for the text “ret = devm_pm_runtime_set_active_enabled”. A match means the tree is affected.
References
- CVE-2026-99248 announcement, linux-cve-announce list
- CVE-2026-99248 record, cve.org
- Fix commit d33373c565af (mainline 7.3)
- Stable backport 47e105bdc863 (7.2.13)
- Introducing commit 50c63491ff26 (mainline 7.1)
- drivers/base/power/runtime.c in 7.2.12
- CVE-2026-99248, National Vulnerability Database
- CISA Known Exploited Vulnerabilities catalogue
- kernel.org release list




