Re: [PATCH 6/6] powerpc/livepatch: Add live patching support on ppc64le

[Michael Ellerman <mpe@ellerman.id.au>]

On Thu, 2016-03-24 at 14:42 +0100, Torsten Duwe wrote:

> On Thu, Mar 24, 2016 at 10:04:05PM +1100, Michael Ellerman wrote:

> > +livepatch_handler:
> > +	CURRENT_THREAD_INFO(r12, r1)
> [...]

> > +	/* Put ctr in r12 for global entry and branch there */
> > +	mfctr	r12
> > +	bctrl
>             ^
> I like this piece. No need to fiddle out the return helper address.

Good.

> > +	/*
> > +	 * Now we are returning from the patched function to the original
> > +	 * caller A. We are free to use r0 and r12, and we can use r2 until we
> > +	 * restore it.
> > +	 */
> > +
> > +	CURRENT_THREAD_INFO(r12, r1)
> > +
> > +	/* Save stack pointer into r0 */
> > +	mr	r0, r1
> > +
> > +	ld	r1, TI_livepatch_sp(r12)
> > +
> > +	/* Check stack marker hasn't been trashed */
> > +	lis     r2,  STACK_END_MAGIC@h
> > +	ori     r2,  r2, STACK_END_MAGIC@l
> > +	ld	r12, -8(r1)
> > +1:	tdne	r12, r2
> > +	EMIT_BUG_ENTRY 1b, __FILE__, __LINE__ - 1, 0
> 
> This however worries me a bit. Sure, in the end, a stack overflow is
> a stack overflow, and if all the information does not fit there,
> there's little you can do.

Yeah stack overflow in the kernel is very very fatal.

> But wouldn't it be better to kmalloc that area and realloc in
> klp_arch_set_pc when it's full? Maybe along with a warning message?

You can't realloc in klp_arch_set_pc(), you might be patching sl*b and holding
one of its locks. You might also recurse.

We could allocate a larger buffer as a "klp stack" for each task when the first
live patch is installed, and for every task created afterward. But that
potentially significantly increases memory usage on live patched kernels :)

> That way a live patched kernel will not run into stack size problems
> any earlier than an unpatched kernel would.

Yeah that's true. I'm not sure what the best trade off is.

cheers