From: Rasmus Villemoes
> Sent: 21 September 2020 11:22
> On 19/09/2020 02.17, Al Viro wrote:
> > On Fri, Sep 18, 2020 at 05:07:43PM -0700, Andy Lutomirski wrote:
> >> On Fri, Sep 18, 2020 at 4:55 PM Al Viro <v...@zeniv.linux.org.uk> wrote:
> >>>
> >>> On Fri, Sep 18, 2020 at 04:31:33PM -0700, Andy Lutomirski wrote:
> >>>
> >>>> check_zeroed_user() looks buggy. It does:
> >>>>
> >>>> if (!user_access_begin(from, size))
> >>>> return -EFAULT;
> >>>>
> >>>> unsafe_get_user(val, (unsigned long __user *) from, err_fault);
> >>>>
> >>>> This is wrong if size < sizeof(unsigned long) -- you read outside the
> >>>> area you verified using user_access_begin().
> >>>
> >>> Read the code immediately prior to that. from will be word-aligned,
> >>> and size will be extended accordingly. If the area acceptable for
> >>> user_access_begin() ends *NOT* on a word boundary, you have a problem
> >>> and I would strongly recommend to seek professional help.
> >>>
> >>> All reads in that thing are word-aligned and word-sized. So I very
> >>> much doubt that your analysis is correct.
> >>
> >> Maybe -ETOOTIRED, but I seriously question the math in here. Suppose
> >> from == (unsigned long *)1 and size == 1. Then align is 1, and we do:
> >>
> >> from -= align;
> >> size += align;
> >>
> >> So now from = 0 and size = 2. Now we do user_access_begin(0, 2) and
> >> then immediately read 4 or 8 bytes. No good.
> >
> > Could you explain what kind of insane hardware manages to do #PF-related
> > checks (including SMAP, whatever) with *sub*WORD* granularity?
> >
> > If it's OK with 16bit read from word-aligned address, but barfs on 64bit
> > one... I want to know what the hell had its authors been smoking.
> >
>
> So, not sure how the above got triggered, but I notice there might be an
> edge case in check_zeroed_user():
>
> from -= align;
> size += align;
>
> if (!user_read_access_begin(from, size))
> return -EFAULT;
>
> unsafe_get_user(val, (unsigned long __user *) from, err_fault);
>
>
> Suppose size is (size_t)-3 and align is 3. What's the convention for
> access_ok(whatever, 0)? Is that equivalent to access_ok(whatever, 1), or
> is it always true (or $ARCH-dependent)?
Doesn't matter, it will be doing access_ok(xxx, 8) regardless of
the user-supplied transfer length.
> But, AFAICT, no current caller of check_zeroed_user can end up passing
> in a size that can overflow to 0. E.g. for the case at hand, size cannot
> be more than SIZE_MAX-24.
Basically KASAN doesn't like you reading full words and masking
off the unused bytes.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT,
UK
Registration No: 1397386 (Wales)
