Re: [patch 4/4] prctl: PR_SET_MM -- Introduce PR_SET_MM_MAP operation, v3

[Cyrill Gorcunov <gorcunov@gmail.com>]

On Fri, Aug 22, 2014 at 01:46:28PM -0700, Andrew Morton wrote:
> On Sat, 23 Aug 2014 00:38:09 +0400 Cyrill Gorcunov <gorcunov@gmail.com> wrote:
> 
> > > 
> > > Or will we?  What happens if we later decide that some additional field
> > > needs to be added?  Do we version the interface?  Add a new prctl()
> > > mode?  Let's cook up a plan for that and at least add to changelog?
> > 
> > I don't expect to change it anytime soon but we still have an option --
> > if we decide to extend or shrink it we always can use sizeof/offsetof
> > helpers to check which exactly version userspace asks us to use.
> 
> How does that work?  We just have a blob of bytes coming in from
> userspace.

Not just blob. We have it as a structure where all fields have a
constant size. Say we have

struct prctl_mm_map {
	__u64 start_code;
	__u64 start_code;
	__u64 some-new-field;
};

in the kernel, so its size will be 24 bytes but userspace
uses old definition without @some-new-field member (16 bytes).
So when we get a reguest with 16 bytes from userspace we can
find the userspace have passed old definition. It's not as
explicit as if we would have some @version field in struct
prctl_mm_mmap, but looks fine for me. Still I can add @version
into the structure if you prefer.

> > As far as I understand the mm_struct is not the structure which
> > changes that frequently, right?
> 
> We might find existing things which criu wants to access.  And criu
> lives forever, yes?  The mm_struct is likely to change over that time
> period ;)

Hopefully criu will live long enough so I would have a chance to update
prctl_mm_map accordingly :) Still the good thing is that once mm_struct
get changed the kernel fails to build in sys.c and the change will
be noticed immediately so we update sys.c as well.