Re: [RFC 7/7] arm64: map seperately rodata sections for __ro_mostly_after_init section

From: Kees Cook
Date: Tue Feb 21 2017 - 15:38:20 EST

On Mon, Feb 20, 2017 at 4:45 AM, Mark Rutland <mark.rutland@xxxxxxx> wrote:
> On Sun, Feb 19, 2017 at 11:35:51AM +0000, Ard Biesheuvel wrote:
>> On 19 February 2017 at 10:04, Hoeun Ryu <hoeun.ryu@xxxxxxxxx> wrote:
>> > Map rodata sections seperately for the new __ro_mostly_after_init section.
>> > Attribute of memory for __ro_mostly_after_init section can be changed later
>> > so we need a dedicated vmalloced region for set_memory_rw/ro api.
>> While it is correct that you are splitting this into three separate
>> segments (otherwise we would not be able to change the permissions
>> later without risking splitting to occur), I think this leads to
>> unnecessary fragmentation.
>> If there is demand for this feature (but you still need to make the
>> argument for that), I wonder if it wouldn't be sufficient, and much
>> more straightforward, to redefine the __ro_after_init semantics to
>> include the kind of subsystem registration and module init context you
>> are targeting, and implement some hooks to temporarily lift the
>> __ro_after_init r/o permission restrictions in a controlled manner.
> From a look over the series, I think this is just __write_rarely in
> disguise. I personally think that we should keep __write_rarely and
> __ro_after_init separate, the later being a strictly one-shot affair.

That's my thinking too.

> I had some ideas [1] as to how we could implement __write_rarely without
> carving up the kernel mapping further (and keeping the RW permissions
> local to the thread needing it), but I have not had the time to look
> into that further.

I'm working on a series to do this for x86, but I keep getting
distracted. I hope to get an RFC posted this week.


Kees Cook
Pixel Security