Re: [PATCH v3] HID: hid-oxp: fix UAF on pending work in remove()

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 5 potential issue(s) to consider:

New issues:
- [High] Calling `cancel_delayed_work_sync()` on uninitialized work structures triggers a kernel `WARN_ON`.

Pre-existing issues:
- [Critical] Global `drvdata` leads to cross-device memory corruption and broken isolation.
- [High] Missing delayed work cancellation in probe error path leads to Use-After-Free.
- [High] Re-arming of delayed works during and after `remove()` causes Use-After-Free.
- [High] Missing report size validation in raw event handlers leads to out-of-bounds memory read.
--

--- Patch [1]: [PATCH v3] HID: hid-oxp: fix UAF on pending work in remove() ---
commit 626193303fb8331bebf2a55e5dc615127fc1f7e7
Author: Shengzhuo Wei <me@cherr.cc>

HID: hid-oxp: fix UAF on pending work in remove()

This commit addresses a use-after-free regression by changing
cancel_delayed_work() to cancel_delayed_work_sync() in oxp_hid_remove().
This ensures workers exit before the device is freed, preventing asleep
workers from waking up and dereferencing a stale hid_device.

Link: https://lore.kernel.org/r/20260804-oxp-fix-v2-1-b2d56e4c8a2c@cherr.cc
Link: https://lore.kernel.org/r/20260804-oxp-fix-v1-1-51a4fe787167@cherr.cc

> diff --git a/drivers/hid/hid-oxp.c b/drivers/hid/hid-oxp.c
> index 20a54f337220d..d8fb6a69d40d4 100644
> --- a/drivers/hid/hid-oxp.c
> +++ b/drivers/hid/hid-oxp.c
> @@ -1552,9 +1552,9 @@ static int oxp_hid_probe(struct hid_device *hdev,

[Severity: Critical]
This isn't a bug introduced by this patch, but does using a single static global
structure for per-device state cause cross-device memory corruption?

When oxp_cfg_probe() runs for a new device, it unconditionally overwrites
drvdata.hdev, drvdata.bmap_1, and drvdata.bmap_2 with its own allocations:

oxp_cfg_probe()
    ...
    hid_set_drvdata(hdev, &drvdata);
    mutex_init(&drvdata.cfg_mutex);
    drvdata.hdev = hdev;
    ...
    drvdata.bmap_1 = bmap_1;
    drvdata.bmap_2 = bmap_2;

If multiple supported interfaces or gamepads are connected, won't they
clobber each other's state, leading to a use-after-free regression if one
is unbound while the other remains active?

[Severity: High]
This isn't a bug introduced by this patch, but could the probe error path leak
a running delayed work?

In oxp_cfg_probe(), drvdata.oxp_mcu_init is armed via mod_delayed_work().
If subsequent initialization like devm_device_add_group() fails, it
returns an error.

The error handler in oxp_hid_probe() cleans up hardware state but doesn't
cancel the pending delayed work before returning.

Could this allow the work to run after devres frees the device memory,
causing a use-after-free regression when oxp_mcu_init_fn() dereferences
the freed pointers?

[Severity: High]
This isn't a bug introduced by this patch, but do the raw event handlers
validate the incoming report size?

In both oxp_hid_raw_event_gen_1() and oxp_hid_raw_event_gen_2(), the code
casts the buffer and reads fields like rgb_rep->effect or checks indices
like data[3] without checking the size of the incoming HID report buffer.

If a device sends a truncated report, could this result in an out-of-bounds
memory read regression?

>  static void oxp_hid_remove(struct hid_device *hdev)
>  {
> -	cancel_delayed_work(&drvdata.oxp_rgb_queue);
> -	cancel_delayed_work(&drvdata.oxp_btn_queue);
> -	cancel_delayed_work(&drvdata.oxp_mcu_init);
> +	cancel_delayed_work_sync(&drvdata.oxp_rgb_queue);
> +	cancel_delayed_work_sync(&drvdata.oxp_btn_queue);
> +	cancel_delayed_work_sync(&drvdata.oxp_mcu_init);

[Severity: High]
If a device connects and exposes a standard HID interface, oxp_hid_probe()
can return success without calling oxp_cfg_probe(), leaving the work
structures uninitialized.

When the device is disconnected, oxp_hid_remove() will unconditionally
call cancel_delayed_work_sync() on these zero-initialized structures.

Could this trigger a WARN_ON(!work->func) inside __flush_work()?

[Severity: High]
This is a pre-existing issue, but modifying these lines highlights a
regression where cancellation happens before hid_hw_stop().

Because the hardware is not stopped, incoming USB reports can still invoke
oxp_hid_raw_event_gen_2(), which unconditionally calls mod_delayed_work().

Furthermore, sysfs attributes managed by devres remain accessible to
userspace until devres cleanup occurs after remove().

Can an unprivileged user writing to sysfs (like map_button_store) or
incoming hardware events re-arm the works after this synchronization,
causing a use-after-free regression when devres frees the memory?

>  	hid_hw_close(hdev);
>  	hid_hw_stop(hdev);
>  }

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260824-oxp-fix-v3-1-599b12c91a77@cherr.cc?part=1