On Wed, Dec 29, 2021 at 6:07 PM Menglong Dong <menglong8.dong@xxxxxxxxx> wrote: > > On Thu, Dec 30, 2021 at 12:46 AM Alexei Starovoitov > <alexei.starovoitov@xxxxxxxxx> wrote: > > > > On Wed, Dec 29, 2021 at 3:33 AM <menglong8.dong@xxxxxxxxx> wrote: > > > > > > From: Menglong Dong <imagedong@xxxxxxxxxxx> > > > > > > The cgroup eBPF attach type 'CGROUP_SOCK_OPS' is able to monitor the > > > state change of a tcp connect with 'BPF_SOCK_OPS_STATE_CB' ops. > > > > > > However, it can't trace the whole state change of a tcp connect. While > > > a connect becomes 'TCP_TIME_WAIT' state, this sock will be release and > > > a tw sock will be created. While tcp sock release, 'TCP_CLOSE' state > > > change will be passed to eBPF program. Howeven, the real state of this > > > connect is 'TCP_TIME_WAIT'. > > > > > > To make eBPF get the real state change of a tcp connect, add > > > 'CGROUP_TWSK_CLOSE' cgroup attach type, which will be called when > > > tw sock release and tcp connect become CLOSE. > > > > The use case is not explained. > > Sorry for the absence of use cases and selftests. In my case, it is for NAT of > a docker container. > > Simply speaking, I'll add an element to a hash map during sys_connect() with > 'BPF_SOCK_OPS_TCP_CONNECT_CB' ops of 'BPF_CGROUP_SOCK_OPS' > cgroup attach type. Therefore, the received packet of the host can do DNAT > according to the hash map. > > I need to release the element in the hashmap when the connection closes. > With the help of 'BPF_SOCK_OPS_STATE_CB', I can monitor the TCP_CLOSE > of the connection. However, as I mentioned above, it doesn't work well when > it comes to tw sock. When the connect become 'FIN_WAIT2' or 'TIME_WAIT', > the state of the tcp sock becomes 'TCP_CLOSE', which doesn't match the connect > state. Therefore, the 'fin' packet that the host received can't be DNAT, as the > element is already removed. > > In this patch, BPF_SOCK_OPS_TW_CLOSE_FLAG is introduced, which is used > make 'BPF_SOCK_OPS_STATE_CB' not called when this sock becomes > TCP_CLOSE if it is being replaced with a tw sock. > > > Why bpf tracing cannot be used to achieve the same? > > En...do you mean kprobe based eBPF trace? It can work, but I don't think it's > high-performance, especially for network NAT. Strictly speaking, attach types, > such as 'CGROUP_INET_SOCK_RELEASE', can be replaced by bpf tracing, but > they exist out of performance. kprobe at the entry is pretty fast. fentry is even faster. It's actually faster than cgroup based hook. Give it a shot.
