Re: [PATCH 0/5] x86/ftrace: Cure boot time W+X mapping

[Linus Torvalds]

On Tue, Oct 25, 2022 at 1:11 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> These few patches re-work and re-order boot things enough to avoid ftrace
> creating boot time W+X maps.

Thanks, looks fine.

> The patches compile and boot for the one config I tested things on (with
> ftrace=function enabled; *slooooow*).

So this might be just tracing overhead, but it might also be that you
slowed down text_poke() at bootup a _lot_.

The only part that the NX^W checking cared about was that

-       if (likely(system_state != SYSTEM_BOOTING))
-               set_memory_ro((unsigned long)trampoline, npages);
+       set_memory_ro((unsigned long)trampoline, npages);
        set_memory_x((unsigned long)trampoline, npages);

for the create_trampoline(), because without the 'set_memory_ro()',
the 'set_memory_x()' will complain.

It does strike me that it's stupid to make those be two calls that do
exactly the same thing, and we should have a combined "set it
read-only and executable" function, but that's a separate issue.

The slowness is probably not the trampilines, but just the regular
"text_poke of kernel text" that we probably want to keep special just
because otherwise it's _so_ slow to do for every alternative etc.

                Linus