CVE-2026-64584

In the Linux kernel, the following vulnerability has been resolved:

usb: gadget: f_midi: cancel pending IN work before freeing the midi object

The f_midi driver embeds a work item (midi->work) whose handler, f_midi_in_work(), dereferences the enclosing struct f_midi through container_of(). This work is armed from two sites: f_midi_complete(), on a normal IN-endpoint completion, and f_midi_in_trigger(), on an ALSA rawmidi output-stream start.

Neither f_midi_disable() nor f_midi_unbind() cancels midi->work. f_midi_disable() only disables the endpoints and drains the in_req_fifo; it does not synchronize the work item, and the sound card is released asynchronously to the final free of the midi object.

The midi object is reference-counted (midi->free_ref) and is freed in f_midi_free() only once both the usb_function reference and the rawmidi private_data reference have been dropped. In f_midi_unbind(), f_midi_disable() runs before the sound card is released, so while the USB endpoints are already disabled the rawmidi device is still usable by an open substream. A concurrent userspace write on such a substream can reach f_midi_in_trigger() and queue midi->work again after f_midi_disable() has returned. A work item armed this way may still be pending when the last reference drops and f_midi_free() proceeds to kfree(midi), letting f_midi_in_work() dereference the struct after it has been freed, a use-after-free.

For this reason cancelling midi->work in f_midi_disable() would not be sufficient: the ALSA trigger path can rearm the work after disable() returns. Cancelling at the refcount-zero free site is the boundary after which neither arming source can survive, because by then both references that keep the midi object alive have been dropped: the USB endpoints are already disabled and the rawmidi device has been released.

Fix this by calling cancel_work_sync(&midi->work) in the refcount-zero block of f_midi_free(), before the embedded work_struct is freed along with the rest of the structure. opts->lock is a sleeping mutex, so calling cancel_work_sync() under it is permitted, and the handler takes midi->transmit_lock rather than opts->lock, so no self-deadlock can occur while it waits for a running instance of the work to finish.

This issue was found by an in-house static analysis tool.

Affected Software

References

• https://git.kernel.org/stable/c/380b4bef46c2eb260c7a9c6bb2c5be33ce5a38f9
• https://git.kernel.org/stable/c/5650c18d93a1db7e27cb5a40b394747eb4686d5b
• https://git.kernel.org/stable/c/87bc316dd6fc90072297c635e10b9aa6075ecda1
• https://git.kernel.org/stable/c/ac9a51d910bb7465c554c45320cb6c09f3d0b49d
• https://git.kernel.org/stable/c/f45089eaad0a083d71d84ff175741d7e157d9b69