[PATCH] structleak: disable BYREF_ALL in combination with KASAN_STACK

[Ard Biesheuvel]

On Fri, 21 Jun 2019 at 15:44, Arnd Bergmann <arnd at arndb.de> wrote:
>
> On Fri, Jun 21, 2019 at 3:32 PM Ard Biesheuvel
> <ard.biesheuvel at linaro.org> wrote:
> > 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.
>
> One pattern I have seen here is temporary variables from macros or
> inline functions whose lifetime now extends over the entire function
> rather than just the basic block in which they are defined, see e.g.
> lpfc_debug_dump_qe() being inlined multiple times into
> lpfc_debug_dump_all_queues(). Each instance of the local
> "char line_buf[LPFC_LBUF_SZ];" seems to add on to the previous
> one now, where the behavior without the structleak plugin is that
> they don't.
>

Right, that seems to be due to the fact that this code

/* split the first bb where we can put the forced initializers */
gcc_assert(single_succ_p(ENTRY_BLOCK_PTR_FOR_FN(cfun)));
bb = single_succ(ENTRY_BLOCK_PTR_FOR_FN(cfun));
if (!single_pred_p(bb)) {
    split_edge(single_succ_edge(ENTRY_BLOCK_PTR_FOR_FN(cfun)));
    gcc_assert(single_succ_p(ENTRY_BLOCK_PTR_FOR_FN(cfun)));
}

puts all the initializers at the beginning of the function rather than
inside the scope of the definition.