On 12/13/24 6:42 AM, Jason Xing wrote:
I just noticed a trickier one, sockops bpf prog can write to sk->sk_txhash. The
same should go for reading from sk. Also, sockops prog assumes a fullsock sk is
a tcp_sock which also won't work for the udp case. A quick thought is to do
something similar to is_fullsock. May be repurpose the is_fullsock somehow or a
new u8 is needed. Take a look at SOCK_OPS_{GET,SET}_FIELD. It avoids
writing/reading the sk when is_fullsock is false.
May be this message buried in the earlier reply or some piece was not clear, so
worth to highlight here.
Take a look at how is_fullsock is used in SOCK_OPS_{GET,SET}_FIELD. I think a
similar idea can be borrowed here.
Do you mean that if we introduce a new field, then bpf prog can
read/write the socket?
The same goes for writing the sk, e.g. writing the sk->sk_txhash. It needs the
sk_lock held. Reading may be ok-ish. The bpf prog can read it anyway by
bpf_probe_read...etc.
When adding udp timestamp callback later, it needs to stop reading the tcp_sock
through skops from the udp callback for sure. Do take a look at
SOCK_OPS_GET_TCP_SOCK_FIELD. I think we need to ensure the udp timestamp
callback won't break here before moving forward.
Agreed. Removing the "sock_ops.sk = sk;" is simple, but I still want
the bpf prog to be able to read some fields from the socket under
those new callbacks.
No need to remove "sock_ops.sk = sk;". Try to borrow the is_fullsock idea.
Overall, the new timestamp callback breaks assumptions like, sk_lock is held and
is_fullsock must be a tcp_sock. This needs to be audited. In particular, please
check sock_ops_func_proto() for all accessible bpf helpers. Also check the
sock_ops_is_valid_access() and sock_ops_convert_ctx_access() for directly
accessible fields without the helpers. In particular, the BPF_WRITE (able)
fields and the tcp_sock fields.
