[PATCH] structleak: disable BYREF_ALL in combination with KASAN_STACK

[Ard Biesheuvel]

On Fri, 21 Jun 2019 at 11:44, Arnd Bergmann <arnd at arndb.de> wrote:
>
> On Thu, Jun 20, 2019 at 7:36 PM Kees Cook <keescook at chromium.org> wrote:
> >
> > On Tue, Jun 18, 2019 at 11:47:13AM +0200, Arnd Bergmann wrote:
> > > The combination of KASAN_STACK and GCC_PLUGIN_STRUCTLEAK_BYREF_ALL
> > > leads to much larger kernel stack usage, as seen from the warnings
> > > about functions that now exceed the 2048 byte limit:
> >
> > Is the preference that this go into v5.2 (there's not much time left),
> > or should this be v5.3? (You didn't mark it as Cc: stable?)
>
> Having it in 5.2 would be great. I had not done much build testing in the last
> months, so I didn't actually realize that your patch was merged a while ago
> rather than only in linux-next.
>
> BTW, I have now run into a small number of files that are still affected
> by a stack overflow warning from STRUCTLEAK_BYREF_ALL. I'm trying
> to come up with patches for those as well, we can probably do it in a way
> that also improves the affected drivers. I'll put you on Cc when I
> find another one.
>

There is something fundamentally wrong here, though. BYREF_ALL only
initializes variables that have their address taken, which does not
explain why the size of the stack frame should increase (since in
order to have an address in the first place, the variable must already
have a stack slot assigned)

So I suspect that BYREF_ALL is defeating some optimizations where.
e.g., the call involving the address of the variable is optimized
away, but the the initialization remains, thus forcing the variable to
be allocated in the stack frame even though the initializer is the
only thing that references it.