On Thu, Dec 21, 2023 at 11:06 PM D. Wythe <alibuda@xxxxxxxxxxxxxxxxx> wrote:
>
>
>
> On 12/21/23 5:11 AM, Alexei Starovoitov wrote:
> > On Wed, Dec 20, 2023 at 6:09 AM D. Wythe <alibuda@xxxxxxxxxxxxxxxxx> wrote:
> >> From: "D. Wythe" <alibuda@xxxxxxxxxxxxxxxxx>
> >>
> >> To support the prog update, we need to ensure that the prog seen
> >> within the hook is always valid. Considering that hooks are always
> >> protected by rcu_read_lock(), which provide us the ability to
> >> access the prog under rcu.
> >>
> >> Signed-off-by: D. Wythe <alibuda@xxxxxxxxxxxxxxxxx>
> >> ---
> >> net/netfilter/nf_bpf_link.c | 63 ++++++++++++++++++++++++++++++++++-----------
> >> 1 file changed, 48 insertions(+), 15 deletions(-)
> >>
> >> diff --git a/net/netfilter/nf_bpf_link.c b/net/netfilter/nf_bpf_link.c
> >> index e502ec0..9bc91d1 100644
> >> --- a/net/netfilter/nf_bpf_link.c
> >> +++ b/net/netfilter/nf_bpf_link.c
> >> @@ -8,17 +8,8 @@
> >> #include <net/netfilter/nf_bpf_link.h>
> >> #include <uapi/linux/netfilter_ipv4.h>
> >>
> >> -static unsigned int nf_hook_run_bpf(void *bpf_prog, struct sk_buff *skb,
> >> - const struct nf_hook_state *s)
> >> -{
> >> - const struct bpf_prog *prog = bpf_prog;
> >> - struct bpf_nf_ctx ctx = {
> >> - .state = s,
> >> - .skb = skb,
> >> - };
> >> -
> >> - return bpf_prog_run(prog, &ctx);
> >> -}
> >> +/* protect link update in parallel */
> >> +static DEFINE_MUTEX(bpf_nf_mutex);
> >>
> >> struct bpf_nf_link {
> >> struct bpf_link link;
> >> @@ -26,8 +17,20 @@ struct bpf_nf_link {
> >> struct net *net;
> >> u32 dead;
> >> const struct nf_defrag_hook *defrag_hook;
> >> + struct rcu_head head;
> > I have to point out the same issues as before, but
> > will ask them differently...
> >
> > Why do you think above rcu_head is necessary?
> >
> >> };
> >>
> >> +static unsigned int nf_hook_run_bpf(void *bpf_link, struct sk_buff *skb,
> >> + const struct nf_hook_state *s)
> >> +{
> >> + const struct bpf_nf_link *nf_link = bpf_link;
> >> + struct bpf_nf_ctx ctx = {
> >> + .state = s,
> >> + .skb = skb,
> >> + };
> >> + return bpf_prog_run(rcu_dereference_raw(nf_link->link.prog), &ctx);
> >> +}
> >> +
> >> #if IS_ENABLED(CONFIG_NF_DEFRAG_IPV4) || IS_ENABLED(CONFIG_NF_DEFRAG_IPV6)
> >> static const struct nf_defrag_hook *
> >> get_proto_defrag_hook(struct bpf_nf_link *link,
> >> @@ -126,8 +129,7 @@ static void bpf_nf_link_release(struct bpf_link *link)
> >> static void bpf_nf_link_dealloc(struct bpf_link *link)
> >> {
> >> struct bpf_nf_link *nf_link = container_of(link, struct bpf_nf_link, link);
> >> -
> >> - kfree(nf_link);
> >> + kfree_rcu(nf_link, head);
> > Why is this needed ?
> > Have you looked at tcx_link_lops ?
>
> Introducing rcu_head/kfree_rcu is to address the situation where the
> netfilter hooks might
> still access the link after bpf_nf_link_dealloc.
Why do you think so?
>
> nf_hook_run_bpf
> const struct
> bpf_nf_link *nf_link = bpf_link;
>
> bpf_nf_link_release
> nf_unregister_net_hook(nf_link->net, &nf_link->hook_ops);
>
> bpf_nf_link_dealloc
> free(link)
> bpf_prog_run(link->prog);
>
>
> I had checked the tcx_link_lops ,it's seems it use the synchronize_rcu()
> to solve the
Where do you see such code in tcx_link_lops ?
> same problem, which is also the way we used in the first version.
>
> https://lore.kernel.org/bpf/1702467945-38866-1-git-send-email-alibuda@xxxxxxxxxxxxxxxxx/
>
> However, we have received some opposing views, believing that this is a
> bit overkill,
> so we decided to use kfree_rcu.
>
> https://lore.kernel.org/bpf/20231213222415.GA13818@xxxxxxxxxxxxx/
>
> >> }
> >>
> >> static int bpf_nf_link_detach(struct bpf_link *link)
> >> @@ -162,7 +164,34 @@ static int bpf_nf_link_fill_link_info(const struct bpf_link *link,
> >> static int bpf_nf_link_update(struct bpf_link *link, struct bpf_prog *new_prog,
> >> struct bpf_prog *old_prog)
> >> {
> >> - return -EOPNOTSUPP;
> >> + struct bpf_nf_link *nf_link = container_of(link, struct bpf_nf_link, link);
> >> + int err = 0;
> >> +
> >> + mutex_lock(&bpf_nf_mutex);
> > Why do you need this mutex?
> > What race does it solve?
>
> To avoid user update a link with differ prog at the same time. I noticed
> that sys_bpf()
> doesn't seem to prevent being invoked by user at the same time. Have I
> missed something?
You're correct that sys_bpf() doesn't lock anything.
But what are you serializing in this bpf_nf_link_update() ?
What will happen if multiple bpf_nf_link_update()
without mutex run on different CPUs in parallel ?
