Re: [PATCH] replace incorrect strscpy use in FORTIFY_SOURCE

From: Daniel Micay
Date: Fri Jul 14 2017 - 20:24:11 EST


On Fri, 2017-07-14 at 16:51 -0700, Kees Cook wrote:
> On Fri, Jul 14, 2017 at 2:28 PM, Daniel Micay <danielmicay@xxxxxxxxx>
> wrote:
> > Using strscpy was wrong because FORTIFY_SOURCE is passing the
> > maximum
> > possible size of the outermost object, but strscpy defines the count
> > parameter as the exact buffer size, so this could copy past the end
> > of
> > the source. This would still be wrong with the planned usage of
> > __builtin_object_size(p, 1) for intra-object overflow checks since
> > it's
> > the maximum possible size of the specified object with no guarantee
> > of
> > it being that large.
> >
> > Reuse of the fortified functions like this currently makes the
> > runtime
> > error reporting less precise but that can be improved later on.
> >
> > Signed-off-by: Daniel Micay <danielmicay@xxxxxxxxx>
>
> Thanks for fixing this! Linus, do you want to take this directly or
> have it go via -mm where fortify landed originally?
>
> Acked-by: Kees Cook <keescook@xxxxxxxxxxxx>
>
> As far as testing goes, was the NFS tree not in -next, or was a test
> not running against -next? I'm curious why it took until the NFS tree
> landed in Linus's tree for this to get noticed. Fortify was in -next
> for a while...
>
> -Kees

Not sure but one issue is that v1 wasn't broken and that's what I most
heavily tested myself. I then switched to testing with intra-object size
checks on top of it, and there would have needed to be a case like this
to break with those stricter checks:

char a[2];
char b[3];
char *p = cond ? a : b;
strcpy(a, c);

__builtin_object_size(p, 0) / __builtin_object_size(p, 1) will both
return 3 there, which is wrong with that incorrect strscpy usage.

I wouldn't have found it via NFS since I've been testing with the string
(but not memory) functions switched to the stricter type 1.

I started on some test cases for FORTIFY_SOURCE but I was testing to
make sure it catches all the bugs it's supposed to catch, it's probably
a good idea to write some more generic string edge case tests too.

One of the somewhat subtle cases (which is handled properly already):

struct foo {
char a[2];
char b;
}

struct foo x;
x.a[0] = '1';
x.a[1] = '2';
x.b = '\0';

strlen(x.a); // BUG with stricter intra-object overflow checks on
strnlen(x.a, 3); // BUG with stricter intra-object overflow checks on
strnlen(x.a, 2); // no overflow

Anyway, it'd be really good if other people looked closely at these. I
wasn't sure what to do with test cases that I've made though. It seems
ones that are *supposed* to BUG should go in lkdtm, and should the other
tests just be together with those? Some should also pass w/o the intra
object overflow checks, but BUG with them enabled.