netfilter: nf_tables: nft_compat: fix refcount leak on xt module

netfilter: nf_tables: nft_compat: fix refcount leak on xt module

Taehee Yoo reported following bug:
    iptables-compat -I OUTPUT -m cpu --cpu 0
    iptables-compat -F
    lsmod |grep xt_cpu
    xt_cpu                 16384  1

Quote:
"When above command is given, a netlink message has two expressions that
are the cpu compat and the nft_counter.
The nft_expr_type_get() in the nf_tables_expr_parse() successes
first expression then, calls select_ops callback.
(allocates memory and holds module)
But, second nft_expr_type_get() in the nf_tables_expr_parse()
returns -EAGAIN because of request_module().
In that point, by the 'goto err1',
the 'module_put(info[i].ops->type->owner)' is called.
There is no release routine."

The core problem is that unlike all other expression,
nft_compat select_ops has side effects.

1. it allocates dynamic memory which holds an nft ops struct.
   In all other expressions, ops has static storage duration.
2. It grabs references to the xt module that it is supposed to
   invoke.

Depending on where things go wrong, error unwinding doesn't
always do the right thing.

In the above scenario, a new nft_compat_expr is created and
xt_cpu module gets loaded with a refcount of 1.

Due to to -EAGAIN, the netlink messages get re-parsed.
When that happens, nft_compat finds that xt_cpu is already present
and increments module refcount again.

This fixes the problem by making select_ops to have no visible
side effects and removes all extra module_get/put.

When select_ops creates a new nft_compat expression, the new
expression has a refcount of 0, and the xt module gets its refcount
incremented.

When error happens, the next call finds existing entry, but will no
longer increase the reference count -- the presence of existing
nft_xt means we already hold a module reference.

Because nft_xt_put is only called from nft_compat destroy hook,
it will never see the initial zero reference count.
->destroy can only be called after ->init(), and that will increase the
refcount.

Lastly, we now free nft_xt struct with kfree_rcu.
Else, we get use-after free in nf_tables_rule_destroy:

  while (expr != nft_expr_last(rule) && expr->ops) {
    nf_tables_expr_destroy(ctx, expr);
    expr = nft_expr_next(expr); // here

nft_expr_next() dereferences expr->ops. This is safe
for all users, as ops have static storage duration.
In nft_compat case however, its ->destroy callback can
free the memory that hold the ops structure.

Tested-by: Taehee Yoo <ap420073@gmail.com>
Reported-by: Taehee Yoo <ap420073@gmail.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>