Re: [PATCH] ntfs: change check order in ntfs_attr_find

[Hawkins Jiawei]

On Sat, 27 Aug 2022 at 18:59, Dan Carpenter <dan.carpenter@xxxxxxxxxx> wrote:
>
> On Sat, Aug 27, 2022 at 05:02:31PM +0800, Hawkins Jiawei wrote:
> > On Sat, 27 Aug 2022 at 14:42, Dan Carpenter <dan.carpenter@xxxxxxxxxx> wrote:
> > >
> > > On Fri, Aug 26, 2022 at 11:42:32PM +0800, Hawkins Jiawei wrote:
> > > > On Fri, 26 Aug 2022 at 23:15, Dan Carpenter <dan.carpenter@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Fri, Aug 26, 2022 at 08:32:57PM +0800, Hawkins Jiawei wrote:
> > > > > > > syz test https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git     master
> > > > > > >
> > > > > > > Looks like it is improper check order that causes this bug.
> > > > > >
> > > > > > Sorry for wrong command.
> > > > > > #syz test https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git     master
> > > > > >
> > > > > > diff --git a/fs/ntfs/attrib.c b/fs/ntfs/attrib.c
> > > > > > index 52615e6090e1..6480cd2d371d 100644
> > > > > > --- a/fs/ntfs/attrib.c
> > > > > > +++ b/fs/ntfs/attrib.c
> > > > > > @@ -594,10 +594,11 @@ static int ntfs_attr_find(const ATTR_TYPE type, const ntfschar *name,
> > > > > >       for (;; a = (ATTR_RECORD*)((u8*)a + le32_to_cpu(a->length))) {
> > > > > >               u8 *mrec_end = (u8 *)ctx->mrec +
> > > > > >                              le32_to_cpu(ctx->mrec->bytes_allocated);
> > > > > > +             if ((u8*)a < (u8*)ctx->mrec || (u8*)a > mrec_end)
> > > > > > +                     break;
> > > > >
> > > > > This definitely seems like a bug.  But your code won't build.  Syzbot
> > > > > must have -Werror turned off?
> > > > Hi Dan,
> > > > Did you mean we should put the variable declares at the beginning of the function?
> > > > (Correct me if I understand anything wrong)
> > >
> > > You can declare it at the beginning of the block.
> > OK, I will do like that.
> >
> > >
> > > >
> > > > >
> > > > > Btw, this was in the original code, but those casts are ugly.  Ideally
> > > > > there would be some way to get rid of them.  But otherwise at least
> > > > > put a space after the u8.  "(u8 *)a < (u8 *)ctx->mrec".
> > > > >
> > > > > >               u8 *name_end = (u8 *)a + le16_to_cpu(a->name_offset) +
> > > > > >                              a->name_length * sizeof(ntfschar);
> > > > > > -             if ((u8*)a < (u8*)ctx->mrec || (u8*)a > mrec_end ||
> > > > > > -                 name_end > mrec_end)
> > > > > > +             if (name_end > mrec_end)
> > > > > >                       break;
> > > > >
> > > > > regards,
> > > > > dan carpenter
> > > > So maybe I can try to refactor these codes. But I wonder if this can be
> > > > done in a seperate bug
> > >
> > > The kernel has a strict "one thing per patch rule".  Those rules are
> > > for reviewers and easier backporting.  So the trick is to write the
> > > commit message to persuade the reviewer that the way you've written the
> > > patch is the easiest way to review it.  So here is how I would write the
> > > commit message:
> > >
> > > [PATCH] ntfs: fix out of bounds read in ntfs_attr_find()
> > >
> > > This code deferences "a" to calculate "name_end" and then it checks to
> > > ensure that "a" is within bounds.  Move the bounds checks earlier and
> > > add some comments to make it more clear what they're doing.  Then
> > > calculate "name_end" and check that.
> > >
> > > (Btw, are the wrap around checks really sufficient?  It seems like it
> > > could wrap to something still within the ->mrec buffer but before the
> > > current entry so it would end up in a forever loop or something?)
> > I am not for sure, but it seems that it is OK before.
> > As for the forever loop, there is a break when a->length is 0 in the loop,
> > So I think it probably would not end up in a forever loop?(Correct me if
> > I am wrong)
> >
>
> Checking for zero is not sufficient because it can wrap on 32 bit
> systems.  It needs something like:
>
>                         return -ENOENT;
>                 len = le32_to_cpu(a->length);
>                 if (!len ||
>                     (void *)a + len < (void *)a ||
>                     (void *)a + len > mrec_end)
>                         break;
>                 if (a->type != type)
>                         continue;
>
> Sort of ugly, but hopefully it gives the idea of what I'm saying.
>
> regards,
> dan carpenter
Hi, Dan
Do you mean there may be an overflow on 32bit systems?
It seems that it is, so it may end up in a forever loop
as you point out before.

I will try to add an overflow checking helper to fix it.