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

[Zijian Zhang <zijianzhang@bytedance.com>]

On 10/30/24 8:38 AM, Stanislav Fomichev wrote:
> On 10/29, Zijian Zhang wrote:
>>
>> On 10/29/24 5:22 PM, Stanislav Fomichev wrote:
>>> On 10/29, Zijian Zhang wrote:
>>>>
>>>>
>>>> On 10/29/24 4:07 PM, Stanislav Fomichev wrote:
>>>>> On 10/29, zijianzhang@bytedance.com wrote:
>> ...
>>>>>> 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?
>>
>> Yes, tls_get_ctx is invoked in multiple locations, and I want to only
>> fix sockmap related calls.
> 
> Sounds convincing. Unless John/Jakub have better suggestions:
> 
> Acked-by: Stanislav Fomichev <sdf@fomichev.me>

Thanks for the Ack and reviewing!

In order to make it more accurate, I added inet_test_bit(IS_ICSK, sk)
check in version2. I just found that sk_is_inet only cannot assure
inet_csk is valid. For example, udp_sock does not have inet_connection_sock.