Re: [linux-usb-devel] Bug creating USB endpoints in 2.6.20.x (kernel bug 8198)

["Dmitry Torokhov" <dmitry.torokhov@gmail.com>]

Hi Tejun,

On 5/9/07, Tejun Heo <htejun@gmail.com> wrote:
> Chris Rankin wrote:
> > --- Tejun Heo <htejun@gmail.com> wrote:
> >> Okay, here's a half-assed fix.  With this patch applied, if you try to
> >> unload a module while you're opening it's dev attribute, kernel will
> >> oops later when the file is accessed or closed later but it should fix
> >> the bug winecfg triggers.  I really dunno how to fix this the right way
> >> in the stable kernel.  Better ideas?  Anyone?
> >
> > How about a WARN() and a small(?) memory leak? Better than an oops, surely?
>
> Device node creation/deletion can be quite often depending on
> configuration, so I don't think we can afford memory leak here.  It can
> develop into a big problem for long running hosts.  IMHO, just
> introducing module unload/deletion race is much better.  It's the lesser
> evil, difficult to trigger and already broken in other places anyway.
>
> I think we need to hear what other people are thinking about it.  Cc'ing
> Maneesh, Dmitry and Cornelia.  The whole thread can be read at...
>
>  http://thread.gmane.org/gmane.linux.usb.devel/53559
>  http://thread.gmane.org/gmane.linux.usb.devel/53846
>
> The thread is rather long but just reading the message from the second
> URL should be enough.  The problem is that dev->devt_attr (class dev has
> the same problem) is deallocated when the device is deleted.  If the dev
> sysfs attribute has users at that point, the dev sysfs node is left with
> garbled struct attribute causing oops later.
>
> IMHO, the proper fix for this is immediate-disconnect which is no in -mm
> as the problem is caused by expecting immediate-disconnect behavior when
> it isn't implemented.
>
> As written above, I think it's better to risk module unload / sysfs race
> than keeping the current sysfs deletion / open race.  What do you guys
> think?
>

How about embedding struct attribute fro devt into struct
[class_]device for now? It is not too big and device is still going to
be pinned into memory while there are sysfs users... I don't like
fattening of device structures but leaks and/or oopses are worse in my
book.

-- 
Dmitry