On Mon, Jun 10, 2019 at 12:19:29PM -0500, Josh Poimboeuf wrote:
> On Fri, Jun 07, 2019 at 10:37:56AM +0200, Peter Zijlstra wrote:
> > > > +}
> > > > +
> > > > +static int static_call_module_notify(struct notifier_block *nb,
> > > > + unsigned long val, void *data)
> > > > +{
> > > > + struct module *mod = data;
> > > > + int ret = 0;
> > > > +
> > > > + cpus_read_lock();
> > > > + static_call_lock();
> > > > +
> > > > + switch (val) {
> > > > + case MODULE_STATE_COMING:
> > > > + module_disable_ro(mod);
> > > > + ret = static_call_add_module(mod);
> > > > + module_enable_ro(mod, false);
> > >
> > > Doesn’t it cause some pages to be W+X ?
>
> How so?
This is after complete_formation() which does RO,X. If we then disable
RO we end up with W+X pages, which is bad.
That said, alternatives, ftrace, dynamic_debug all run before
complete_formation() specifically such that they can directly poke text.
Possibly we should add a notifier callback for MODULE_STATE_UNFORMED,
but that is for another day.
> >> Can it be avoided?
> >
> > I don't know why it does this, jump_labels doesn't seem to need this,
> > and I'm not seeing what static_call needs differently.
>
> I forgot why I did this, but it's probably for the case where there's a
> static call site in module init code. It deserves a comment.
>
> Theoretically, jump labels need this to.
>
> BTW, there's a change coming that will require the text_mutex before
> calling module_{disable,enable}_ro().
I can't find why it would need this (and I'm going to remove it).
Specifically complete_formation() does enable_ro(.after_init=false),
which leaves .ro_after_init writable so
{jump_label,static_call}_sort_entries() will work.
But both jump_label and static_call then use the full text_poke(), not
text_poke_early(), for modules.
