Re: [PATCH] Add a text_poke syscall v2

[Oleg Nesterov <oleg@redhat.com>]

Sorry for completely offtopic question, but while we are here...

On 11/30, Jiri Kosina wrote:
>
> We have moved from using stop_machine() to int3-based patching exactly
> because it's much more lightweight.

I don't really understans the barriers in poke_int3_handler() and
text_poke_bp(). To the point, I do not really understand why do we
actually need bp_patching_in_progress, but lets ignore this.

Lets look at the end of text_poke_bp(), it does

	on_each_cpu(do_sync_core, NULL, 1);

	bp_patching_in_progress = false;
	smp_wmb();

First of all, this smp_wmb() is not clear. But what I actually
can't understand is why it is safe to clear bp_patching_in_progress.

OK, on_each_cpu() should serialize us with do_int3(), but only if
poke_int3_handler() is called with irqs disabled.

However, do_int3() does preempt_conditional_sti() and this looks
as if it can be called with irqs enabled? If this is actually
possible then text_poke_bp() needs synchronize_sched() to avoid
the races with poke_int3_handler(), afaics.

OTOH, int3 is GATE_INTERRUPT, doesn't this mean that that do_int3()
can enable irqs unconditionally and on_each_cpu() also acts as a
synchronization barrier?

Oleg.