On Tue Oct 3, 2023 at 1:47 AM EEST, Jarkko Sakkinen wrote:
> On Wed Sep 27, 2023 at 4:56 AM EEST, Haitao Huang wrote:
> > On Tue, 26 Sep 2023 08:13:18 -0500, Jarkko Sakkinen <jarkko@xxxxxxxxxx>
> > wrote:
> >
> > ...
> > >> > >> /**
> > >> > >> @@ -410,7 +429,14 @@ misc_cg_alloc(struct cgroup_subsys_state
> > >> > >> *parent_css)
> > >> > >> */
> > >> > >> static void misc_cg_free(struct cgroup_subsys_state *css)
> > >> > >> {
> > >> > >> - kfree(css_misc(css));
> > >> > >> + struct misc_cg *cg = css_misc(css);
> > >> > >> + enum misc_res_type i;
> > >> > >> +
> > >> > >> + for (i = 0; i < MISC_CG_RES_TYPES; i++)
> > >> > >> + if (cg->res[i].free)
> > >> > >> + cg->res[i].free(cg);
> > >> > >> +
> > >> > >> + kfree(cg);
> > >> > >> }
> > >> > >>
> > >> > >> /* Cgroup controller callbacks */
> > >> > >> --
> > >> > >> 2.25.1
> > >> > >
> > >> > > Since the only existing client feature requires all callbacks,
> > >> should
> > >> > > this not have that as an invariant?
> > >> > >
> > >> > > I.e. it might be better to fail unless *all* ops are non-nil (e.g.
> > >> to
> > >> > > catch issues in the kernel code).
> > >> > >
> > >> >
> > >> > These callbacks are chained from cgroup_subsys, and they are defined
> > >> > separately so it'd be better follow the same pattern. Or change
> > >> together
> > >> > with cgroup_subsys if we want to do that. Reasonable?
> > >>
> > >> I noticed this one later:
> > >>
> > >> It would better to create a separate ops struct and declare the instance
> > >> as const at minimum.
> > >>
> > >> Then there is no need for dynamic assigment of ops and all that is in
> > >> rodata. This is improves both security and also allows static analysis
> > >> bit better.
> > >>
> > >> Now you have to dynamically trace the struct instance, e.g. in case of
> > >> a bug. If this one done, it would be already in the vmlinux.
> > >I.e. then in the driver you can have static const struct declaration
> > > with *all* pointers pre-assigned.
> > >
> > > Not sure if cgroups follows this or not but it is *objectively*
> > > better. Previous work is not always best possible work...
> > >
> >
> > IIUC, like vm_ops field in vma structs. Although function pointers in
> > vm_ops are assigned statically, but you still need dynamically assign
> > vm_ops for each instance of vma.
> >
> > So the code will look like this:
> >
> > if (parent_cg->res[i].misc_ops && parent_cg->res[i].misc_ops->alloc)
> > {
> > ...
> > }
> >
> > I don't see this is the pattern used in cgroups and no strong opinion
> > either way.
> >
> > TJ, do you have preference on this?
>
> I do have strong opinion on this. In the client side we want as much
> things declared statically as we can because it gives more tools for
> statical analysis.
>
> I don't want to see dynamic assignments in the SGX driver, when they
> are not actually needed, no matter things are done in cgroups.
I.e. I don't really even care what crazy things cgroups subsystem
might do or not do. It's not my problem.
All I care is that we *do not* have any use for assigning those
pointers at run-time. So do whatever you want with cgroups side
as long as this is not the case.
BR, Jarkko
