On 8/26/24 6:37 PM, zijianzhang@xxxxxxxxxxxxx wrote:
From: Amery Hung <amery.hung@xxxxxxxxxxxxx>
This series prevents sockops users from accidentally causing packet
drops. This can happen when a BPF_SOCK_OPS_HDR_OPT_LEN_CB program
reserves different option lengths in tcp_sendmsg().
Initially, sockops BPF_SOCK_OPS_HDR_OPT_LEN_CB program will be called to
reserve a space in tcp_send_mss(), which will return the MSS for TSO.
Then, BPF_SOCK_OPS_HDR_OPT_LEN_CB will be called in __tcp_transmit_skb()
again to calculate the actual tcp_option_size and skb_push() the total
header size.
skb->gso_size is restored from TCP_SKB_CB(skb)->tcp_gso_size, which is
derived from tcp_send_mss() where we first call HDR_OPT_LEN. If the
reserved opt size is smaller than the actual header size, the len of the
skb can exceed the MTU. As a result, ip(6)_fragment will drop the
packet if skb->ignore_df is not set.
To prevent this accidental packet drop, we need to make sure the
second call to the BPF_SOCK_OPS_HDR_OPT_LEN_CB program reserves space
not more than the first time.
iiuc, it is a bug in the bpf prog itself that did not reserve the same header
length and caused a drop. It is not the only drop case though for an incorrect
bpf prog. There are other cases where a bpf prog can accidentally drop a packet.
Do you have an actual use case that the bpf prog cannot reserve the correct
header length for the same sk ?
