On Sun, Mar 28, 2021 at 04:42:03PM +0200, Thomas Gleixner wrote:
> On Fri, Mar 19 2021 at 14:28, Kees Cook wrote:
> > +/*
> > + * Do not use this anywhere else in the kernel. This is used here because
> > + * it provides an arch-agnostic way to grow the stack with correct
> > + * alignment. Also, since this use is being explicitly masked to a max of
> > + * 10 bits, stack-clash style attacks are unlikely. For more details see
> > + * "VLAs" in Documentation/process/deprecated.rst
>
> VLAs are bad, VLAs to the rescue! :)
I'm aware of the irony, but luto's idea really makes things easy. As
documented there, though, this has a hard-coded (low) upper bound, so
it's not like "regular" VLA use.
>
> > + * The asm statement is designed to convince the compiler to keep the
> > + * allocation around even after "ptr" goes out of scope.
> > + */
> > +void *__builtin_alloca(size_t size);
> > +
> > +#define add_random_kstack_offset() do { \
> > + if (static_branch_maybe(CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT, \
> > + &randomize_kstack_offset)) { \
> > + u32 offset = this_cpu_read(kstack_offset); \
>
> Not that it matters on x86, but as this has to be called in the
> interrupt disabled region of the syscall entry, shouldn't this be a
> raw_cpu_read(). The asm-generic version has a preempt_disable/enable
> pair around the raw read for native wordsize reads, otherwise a
> irqsave/restore pair.
>
> __this_cpu_read() is fine as well, but that has an sanity check before
> the raw read when CONFIG_DEBUG_PREEMPT is on, which is harmless but
> also pointless in this case.
>
> Probably the same for the counterpart this_cpu_write().
Oh! Excellent point. I think this will make a big difference on arm64. I
will adjust and test.
--
Kees Cook
