Re: Possible race condition in usb-serial.c

[Oliver Neukum <oliver@neukum.org>]

Am Dienstag, 19. Dezember 2006 23:33 schrieb J:
> Thank you for the response.
> 
> > This code depends on protection from BKL.
> 
> Really? I cannot find many lock_kernel calls in 
> USB directory and those, which I can find, 
> don't appear to protect usb_serial_disconnect
> and serial_close from being called at the same time.

serial_close is safe because serial_disconnect lowers the refcount
by one. usb_serial_probe() and usb_serial_open() both increment
the refcount; the former implicitly.

> May be the protection is at a higher level? 
> Personally I don't beleive it.
> If you know how this thing is supposed to work,
> please, tell me.

The data structure to protect is serial_table. Everything else is
protected by refcounts. Therefore the interesting race is between
open and disconnect. Open is called with BKL (fs/char_dev.c::chrdev_open)

Now, regarding disconnect. It used to be called with BKL held. I haven't been
able to verify that this is still the case. If not, then there's a race.

In addition usb_serial_probe() uses get_free_serial() early in the process
before the device is ready. Without BKL, this too, races with open.

People, do we take BKL in khubd?

	Regards
		Oliver