Re: usb/storage/uas: slab-out-of-bounds in uas_probe

From: Alan Stern
Date: Fri Sep 22 2017 - 10:58:42 EST


On Fri, 22 Sep 2017, Greg Kroah-Hartman wrote:

> On Fri, Sep 22, 2017 at 09:58:15AM +0200, Greg Kroah-Hartman wrote:
> > On Thu, Sep 21, 2017 at 03:04:05PM -0400, Alan Stern wrote:
> > > On Thu, 21 Sep 2017, Andrey Konovalov wrote:
> > >
> > > > On Thu, Sep 21, 2017 at 6:10 PM, Greg Kroah-Hartman
> > > > <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > > > > On Thu, Sep 21, 2017 at 05:39:05PM +0200, Andrey Konovalov wrote:
> > > > >> Hi!
> > > > >>
> > > > >> I've got the following report while fuzzing the kernel with syzkaller.
> > > > >>
> > > > >> On commit ebb2c2437d8008d46796902ff390653822af6cc4 (Sep 18).
> > > > >>
> > > > >> The issue occurs when we iterate over interface altsettings, but I
> > > > >> don't see the driver doing anything wrong. I might be missing
> > > > >> something, or this might be an issue in USB core altsettings parsing.
> > > > >
> > > > >
> > > > > Any chance you happen to have the descriptors that you were feeding into
> > > > > the kernel at this crash? That might help in figuring out what "went
> > > > > wrong".
> > > >
> > > > Here's the data that I feed into dummy_udc:
> > > >
> > > > 00 00 00 00 09 02 12 00 01 34 05 80 07 09 04 6e
> > > > 09 00 08 06 62 00 12 01 05 00 cb f7 71 83 04 00
> > > > 05 00 ab f6 07 81 08 01
> > > >
> > > > Also attaching a C reproducer (requires dummy_hcd and gadgetfs) and my .config.
> > >
> > > Why do your reproducers use an mmap'ed array for their data instead of
> > > a plain old statically allocated array?
> > >
> > > Anyway, this turns out to be a genuine (and subtle!) bug in the uas
> > > driver. The uas_find_uas_alt_setting() routine in uas-detect.h returns
> > > a bAlternateSetting value, but then the uas_use_uas_driver() routine
> > > uses this value as an index to the altsetting array -- which it isn't.
> > >
> > > Normally this doesn't cause any problems because the the entries in the
> > > array have bAlternateSetting values 0, 1, etc., so the value is equal
> > > to the index. But in your fuzzed case, that wasn't true.
> > >
> > > The patch below fixes this bug, by returning a pointer to the
> > > alt-setting entry instead of either the value or the index. Pointers
> > > are less subject to misinterpretation.
> >
> > Ugh, messy, nice catch and fix, I'll go queue it up now, thanks for
> > resolving this.
>
> Oops, no, you didn't send this as a "real" patch yet, I'll wait :)

I was waiting for Andrey to verify that the problem was really gone.
I'll submit it officially later today.

On Fri, 22 Sep 2017, Felipe Balbi wrote:

> > Felipe, any objection for me taking this, and the other gadget driver
> > fixes that Alan just sent out, directly in my tree?
>
> none whatsoever, for all of them:

There's still more to come. I've got four patches for dummy-hcd just
about ready to send. :-)

Alan Stern