Re: [PATCH bpf] bpf: Add sk_is_inet check in tls_sw_has_ctx_tx/rx

[Stanislav Fomichev <stfomichev@gmail.com>]

On 10/29, Zijian Zhang wrote:
> 
> 
> On 10/29/24 4:07 PM, Stanislav Fomichev wrote:
> > On 10/29, zijianzhang@bytedance.com wrote:
> > > From: Zijian Zhang <zijianzhang@bytedance.com>
> > > 
> > > As the introduction of the support for vsock and unix sockets in sockmap,
> > > tls_sw_has_ctx_tx/rx cannot presume the socket passed in must be inet.
> > > Otherwise, tls_get_ctx may return an invalid pointer and result in page
> > > fault in function tls_sw_ctx_rx.
> > > 
> > > BUG: unable to handle page fault for address: 0000000000040030
> > > Workqueue: vsock-loopback vsock_loopback_work
> > > RIP: 0010:sk_psock_strp_data_ready+0x23/0x60
> > > Call Trace:
> > >   ? __die+0x81/0xc3
> > >   ? no_context+0x194/0x350
> > >   ? do_page_fault+0x30/0x110
> > >   ? async_page_fault+0x3e/0x50
> > >   ? sk_psock_strp_data_ready+0x23/0x60
> > >   virtio_transport_recv_pkt+0x750/0x800
> > >   ? update_load_avg+0x7e/0x620
> > >   vsock_loopback_work+0xd0/0x100
> > >   process_one_work+0x1a7/0x360
> > >   worker_thread+0x30/0x390
> > >   ? create_worker+0x1a0/0x1a0
> > >   kthread+0x112/0x130
> > >   ? __kthread_cancel_work+0x40/0x40
> > >   ret_from_fork+0x1f/0x40
> > > 
> > > Fixes: 0608c69c9a80 ("bpf: sk_msg, sock{map|hash} redirect through ULP")
> > > Fixes: e91de6afa81c ("bpf: Fix running sk_skb program types with ktls")
> > > 
> > > Signed-off-by: Zijian Zhang <zijianzhang@bytedance.com>
> > > ---
> > >   include/net/tls.h | 12 ++++++++++--
> > >   1 file changed, 10 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/include/net/tls.h b/include/net/tls.h
> > > index 3a33924db2bc..a65939c7ad61 100644
> > > --- a/include/net/tls.h
> > > +++ b/include/net/tls.h
> > > @@ -390,8 +390,12 @@ tls_offload_ctx_tx(const struct tls_context *tls_ctx)
> > >   static inline bool tls_sw_has_ctx_tx(const struct sock *sk)
> > >   {
> > > -	struct tls_context *ctx = tls_get_ctx(sk);
> > > +	struct tls_context *ctx;
> > > +
> > > +	if (!sk_is_inet(sk))
> > > +		return false;
> > > +	ctx = tls_get_ctx(sk);
> > >   	if (!ctx)
> > >   		return false;
> > >   	return !!tls_sw_ctx_tx(ctx);
> > > @@ -399,8 +403,12 @@ static inline bool tls_sw_has_ctx_tx(const struct sock *sk)
> > >   static inline bool tls_sw_has_ctx_rx(const struct sock *sk)
> > >   {
> > > -	struct tls_context *ctx = tls_get_ctx(sk);
> > > +	struct tls_context *ctx;
> > > +
> > > +	if (!sk_is_inet(sk))
> > > +		return false;
> > > +	ctx = tls_get_ctx(sk);
> > >   	if (!ctx)
> > >   		return false;
> > >   	return !!tls_sw_ctx_rx(ctx);
> > 
> > This seems like a strange place to fix it. Why does tls_get_ctx return
> > invalid pointer for non-tls/ulp sockets? Shouldn't it be NULL?
> > Is sockmap even supposed to work with vsock?
> 
> Here is my understanding, please correct me if I am wrong :)
> ```
> static inline struct tls_context *tls_get_ctx(const struct sock *sk)
> {
> 	const struct inet_connection_sock *icsk = inet_csk(sk);
> 	return (__force void *)icsk->icsk_ulp_data;
> }
> ```
> tls_get_ctx assumes the socket passed is icsk_socket. However, unix
> and vsock do not have inet_connection_sock, they have unix_sock and
> vsock_sock. The offset of icsk_ulp_data are meaningless for them, and
> they might point to some other values which might not be NULL.
> 
> Afaik, sockmap started to support vsock in 634f1a7110b4 ("vsock: support
> sockmap"), and support unix in 94531cfcbe79 ("af_unix: Add
> unix_stream_proto for sockmap").
> 
> If the above is correct, I find that using inet_test_bit(IS_ICSK, sk)
> instead of sk_is_inet will be more accurate.

Thanks for the context, makes sense. And consolidating this sk_is_inet
check inside tls_get_ctx is worse because it gets called outside of
sockmap?