DIY allocation or embedding of URBs in larger structures

[Michal Pecio]

On Thu, 03 Sep 2026 11:44:45 +0200, Takashi Iwai wrote:
> On Thu, 03 Sep 2026 11:24:17 +0200,
> Michal Pecio wrote:
> > So what happens here is that USB core continues to use a URB after
> > completion to implement things like usb_kill_urb(), so URBs are
> > reference counted. Then core decrements the count - one more use.
> > 
> > If a driver waits for completion or even usb_kill_urb() to return
> > and then proceeds to free the URB's storage, this becomes a UAF.
> > This driver embeds 2 URBs in its priv and does just that.
> > 
> > A URB can only exist as an independent allocation, core will free
> > it if upon finding zero reference count in such case.
>
> So, IIUC, now the URB *must* be always allocated via usb_alloc_urb()
> and an embedded URB isn't allowed?  If so, we'd need to address other
> drivers, too.

Yes, that's basically the case and actually has been for a long time.

The only thing that works is for both USB and the driver to call
usb_free_urb() and whoever does last will actually free the storage.
This means URB can't share storage with anything else.

Some drivers got away with making URB the first member of a struct
which is freed together with it. AFAIK this pattern doesn't crash, but
it's deprecated too because it puts a flexible member (in the URB) in
the middle of (the outer) struct.

This pattern here in caiaq has always been one race away from UAF.
Maybe it wasn't very likely to happen, but stuff like PREEMPT_RT and
hypervisors can insert unexpected delays anywhere these days.

Greg KH puts it thusly:
https://lore.kernel.org/linux-usb/2025120716-sway-hypnotic-8cb6@gregkh/

Reragds,
Michal