On Tue, Feb 21, 2017 at 10:36:05PM +0900, Ho-Eun Ryu wrote:
> > On 20 Feb 2017, at 7:30 PM, Mark Rutland <mark.rutl...@arm.com> wrote:
> > On Sun, Feb 19, 2017 at 07:04:06PM +0900, Hoeun Ryu wrote:
> >> @@ -3396,8 +3399,11 @@ static noinline int do_init_module(struct module
> >> *mod)
> >>
> >> do_mod_ctors(mod);
> >> /* Start the module */
> >> - if (mod->init != NULL)
> >> + if (mod->init != NULL) {
> >> + set_ro_mostly_after_init_rw();
> >> ret = do_one_initcall(mod->init);
> >> + set_ro_mostly_after_init_ro();
> >> + }
> >
> > This looks very much like the pax_{open,close}_kernel() approach for
> > write-rarely data.
>
> I read the discussion [1] and I agree that __ro_mostly_after_init marker
> looks very similar to __write_rarely.
>
> > I think it would be better to implement a first class write-rarely
> > mechanism rather than trying to extend __ro_after_init to cover this
> > case.
>
> I’m not extending __ro_after_init. __ro_mostly_after_init resides in
> the same section of rodata though.
Sorry; I was confused when I wrote that email. I now understand that
you're adding a separate annotation.
> > As mentioned previously, I *think* we can have a generic implementation
> > that uses an mm to temporarily map a (thread/cpu-local) RW alias of the
> > data in question in what would otherwise be the user half of the address
> > space. Regardless, we can have a generic interface [1] that can cater
> > for that style of approach and/or something like ARM's domains or x86's
> > pkeys.
> >
>
> I’m still learning cpu/kernel architectures, It would be very thankful if you
> tell me more about the detail of the implementation itself.
>
> The mm that maps temporary RW alias is like
> * special mm like idmap/init_mm which have its own page tables?
> * the page tables have the same content of page tables of
> init_mm’s swapper_pg_dir except for RW permissions for a
> specific section (let’s say __write_rarely)
This would be a special mm, like a user mm, that only mapped the
relevant VA(s).
That might map the relevant variable on-demand, or the mapping could
cover the whole write_rarely area.
> * then use switch_mm(special_rw_mm) to change the address space
> before the access happens to the section
> * then use switch_mm(current->mm) to change the address space to
> original after the access is done
Yes.
> And the interface itself. rare_write(__val, __val), is it a single
> value access interface.
> I’m intending to make data in __ro_mostly_after_init section RW during
> multiple accesses like during module_init/exit.
> and __rare_rw_map()/unmap() used in rare_write() seems to work like
> open/close api.
The __rare_rw_{map,unmap}() functions would map in the RW alias, but do
not necessarily change the RO alias to RW. This is why __rare_rw_ptr()
would be necessary, and is the major difference to the open/close API.
We could certainly allow several writes between a map/unmap. The key
requirement is that each write is instrumented so that it goes via the
RW alias.
> How could __rare_rw_ptr() be implemented and what happens when
> `__rw_var = __rare_rw_ptr(&(__var))` is done ?
__rare_rw_ptr() would take a pointer to the usual RO alias, and derive
its RW alias. What exactly this should do depends on how the RW alias is
implemented.
On a system using an RW mm, let's assume we place all __write_rarely
variables in a region bounded by __rare_write_begin/__rare_write_end,
and when the mm is installed place, we have an RW alias of this region
beginning at __rw_alias_start. In this case, it'd look something like:
#define __rare_rw_ptr(ptr) ({ \
unsigned long __ptr = (unsigned long)(ptr); \
__ptr -= __rare_write_start; \
__ptr += __rw_alias_start; \
(typeof(ptr))__ptr; \
})
... does that make sense?
For systems where you can freely/easily alter (local) permissions (e.g.
using ARM's domains), that can be done within __rare_rw_{map,unmap}(),
and __rare_rw_ptr can just return the original pointer.
> However the interface will look like, Do we still need a special data
> section that is mapped RO in general but RW in some cases ?
With the above, I think the usual mapping can always be RO.
> if then, doesn’t __ro_mostly_after_init marker itself make sense and
> we still need it ?
We may need a marker to bound the set of variables we wish to map in
this way.
Thanks,
Mark.
