Re: [PATCH 1/2] x86/kprobes: Fix kprobes instruction boudary check with CONFIG_RETHUNK

[Peter Zijlstra]

On Wed, Sep 07, 2022 at 11:14:50PM +0900, Masami Hiramatsu wrote:
> On Wed, 7 Sep 2022 15:05:13 +0200
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> > On Wed, Sep 07, 2022 at 10:02:41AM +0200, Peter Zijlstra wrote:
> > 
> > > 	struct queue q;
> > > 
> > > 	start = paddr - offset;
> > > 	end = start + size;
> > > 	push(&q, paddr - offset);
> > > 
> > > 	while (start = pop(&q)) {
> > > 		for_each_insn(&insn, start, end, buf) {
> > > 			if (insn.kaddr == paddr)
> > > 				return 1;
> > > 
> > > 			target = insn_get_branch_addr(&insn);
> > > 			if (target)
> > > 				push(&q, target);
> > > 
> > > 			if (dead_end_insn(&insn))
> > > 				break;
> > > 		}
> > > 	}
> > 
> > There is the very rare case of intra-function-calls; but I *think*
> > they're all in noinstr/nokprobe code anyway.
> > 
> > For instance we have RSB stuffing code like:
> > 
> > 	.rept 16
> > 	call 1f;
> > 	int3
> > 	1:
> > 	.endr
> > 	add $(BITS_PER_LONG/8) * 16, %_ASM_SP
> > 
> > And the proposed will be horribly confused by that. But like said; it
> > should also never try and untangle it.
> 
> Yeah, but I guess if we break the decoding (internal) loop when we
> hit an INT3, it maybe possible to be handled?

If you make insn_get_branch_addr() return the target of CALL
instructions when this target is between function start and end, it
should work I think.

But like said; this construct is rare and all instances I can remember
should not be kprobes to begin with. These are all 'fun' things like
retpoline stubs and the the above RSB stuff loop.