On Fri, Jun 4, 2021 at 2:54 AM Christian Brauner
[off-list ref] wrote:

On Thu, Jun 03, 2021 at 03:52:29PM -0700, Cong Wang wrote:

quoted

On Wed, Jun 2, 2021 at 2:14 AM Christian Brauner
[off-list ref] wrote:

quoted

But the point is that ns->ops should never be accessed when that
namespace type is disabled. Or in other words, the bug is that something
in netns makes use of namespace features when they are disabled. If we
handle ->ops being NULL we might be tapering over a real bug somewhere.

It is merely a protocol between fs/nsfs.c and other namespace users,
so there is certainly no right or wrong here, the only question is which
one is better.

quoted

Jakub's proposal in the other mail makes sense and falls in line with
how the rest of the netns getters are implemented. For example
get_net_ns_fd_fd():

It does not make any sense to me. get_net_ns() merely increases
the netns refcount, which is certainly fine for init_net too, no matter
CONFIG_NET_NS is enabled or disabled. Returning EOPNOTSUPP
there is literally saying we do not support increasing init_net refcount,
which is of course false.

quoted

struct net *get_net_ns_by_fd(int fd)
{
        return ERR_PTR(-EINVAL);
}

There is a huge difference between just increasing netns refcount
and retrieving it by fd, right? I have no idea why you bring this up,
calling them getters is missing their difference.

This argument doesn't hold up. All netns helpers ultimately increase the
reference count of the net namespace they find. And if any of them
perform operations where they are called in environments wherey they
need CONFIG_NET_NS they handle this case at compile time.

Let me explain it in this more straight way: what is the protocol here
for indication of !CONFIG_XXX_NS? Clearly it must be ns->ops==NULL,
because all namespaces use the following similar pattern:

#ifdef CONFIG_NET_NS
        net->ns.ops = &netns_operations;
#endif

Now you are arguing the protocol is not this, but it is the getter of
open_related_ns() returns an error pointer.

(Pluse they are defined in a central place in net/net_namespace.{c,h}.
That includes the low-level get_net() function and all the others.
get_net_ns() is the only one that's defined out of band. So get_net_ns()
currently is arguably also misplaced.)

Of course they do, only struct ns_common is generic. What's your
point? Each ns.ops is defined by each namespace too.

The problem I have with fixing this in nsfs is that it gives the
impression that this is a bug in nsfs whereas it isn't and it
potentially helps tapering over other bugs.

Like I keep saying, this is just a protocol, there is no right or
wrong here. If the protocol is just ops==NULL, then there is nothing
wrong use it.

(BTW, we have a lot of places that use ops==NULL as a protocol,
they work really well.)

get_net_ns() is only called for codepaths that call into nsfs via
open_related_ns() and it's the only namespace that does this. But

I am pretty sure userns does the same:

197         case NS_GET_USERNS:
198                 return open_related_ns(ns, ns_get_owner);

open_related_ns() is only well defined if CONFIG_<NAMESPACE_TYPE> is
set. For example, none of the procfs namespace f_ops will be set for
!CONFIG_NET_NS. So clearly the socket specific getter here is buggy as
it doesn't account for !CONFIG_NET_NS and it should be fixed.

If the protocol is just ops==NULL, then the core part should just check
ops==NULL. Pure and simple. I have no idea why you do not admit the
fact that every namespace intentionally leaves ops as NULL when its
config is disabled.

Plus your fix leaks references to init netns without fixing get_net_ns()
too.

I thought it is 100% clear that this patch is not from me?

Plus, the PoC patch from me actually suggests to change
open_related_ns(), not __ns_get_path(). I have no idea why you
both miss it.

Thanks.