Re: [RFC PATCH v16 0/6] mm: security: ro protection for dynamic data

[Dave Chinner]

On Tue, Feb 20, 2018 at 03:56:00PM -0800, Matthew Wilcox wrote:
> On Wed, Feb 21, 2018 at 08:36:04AM +1100, Dave Chinner wrote:
> > FWIW, I'm not wanting to use it to replace static variables. All the
> > structures are dynamically allocated right now, and get assigned to
> > other dynamically allocated pointers. I'd likely split the current
> > structures into a "ro after init" structure and rw structure, so
> > how does the "__ro_after_init" attribute work in that case? Is it
> > something like this?
> > 
> > struct xfs_mount {
> > 	struct xfs_mount_ro{
> > 		.......
> > 	} *ro __ro_after_init;
       ^^^^^^^^

pointer, not embedded structure....

> > 	......
> 
> No, you'd do:
> 
> struct xfs_mount_ro {
> 	[...]
> };
> 
> struct xfs_mount {
> 	const struct xfs_mount_ro *ro;
> 	[...]
> };

.... so that's pretty much the same thing :P

> > Also, what compile time checks are in place to catch writes to
> > ro structure members? Is sparse going to be able to check this sort
> > of thing, like is does with endian-specific variables?
> 
> Just labelling the pointer const should be enough for the compiler to
> catch unintended writes.

Ok.

> > > I'd be interested to have your review of the pmalloc API, if you think
> > > something is missing, once I send out the next revision.
> > 
> > I'll look at it in more depth when it comes past again. :P
> 
> I think the key question is whether you want a slab-style interface
> or whether you want a kmalloc-style interface.  I'd been assuming
> the former, but Igor has implemented the latter already.

Slabs are rally only useful when you have lots of a specific type of
object. I'm concerned mostly about one-off per-mount point
structures, of which there are relatively few. A heap-like pool per
mount is fine for this.

Cheers,

Dave.
-- 
Dave Chinner
dchinner@xxxxxxxxxx