From: Zac Ecob <zacecob@xxxxxxxxxxxxxx> Date: Sat, 22 Jun 2024 12:20:05 +0000 > Problem is title. > > To trigger, I attached an EBPF prof that just returned -1, and send ~1000 > packets through it. If you want to drop the packet, the prog must return 0. You can see sk_filter_trim_cap() where the returned value from bpf prog is cast to unsigned int. Then, pskb_trim() does nothing because skb->len is smaller than (unsigned int)-1, and 0 is set to err. unsigned int pkt_len; pkt_len = bpf_prog_run_save_cb(filter->prog, skb); err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM; After calling sk_filter() from unix_dgram_sendmsg(), the skb is just queued to the peer. > > After doing some investigation, the `sk_wmem_alloc` member of `struct sk` > seems to only be increasing, presumably missing some refcnt_dec somewhere. So, no refcnt is leaked. What is missing is recv() on the peer side. > > At a certain point, in `sock_alloc_send_pskb`, we fail the check: > > ` > if (sk_wmem_alloc_get(sk) < READ_ONCE(sk->sk_sndbuf)) > ` > > Upon which we enter `sock_wait_for_wmem` and schedule a massive timeout > (at least that's what happened in my tests). > > Not sure where the missing refcnt subs are, must admit unfamiliarity with > the network code. The paired sub is sock_wfree() in unix_destruct_scm(), which is set to skb->destructor() in unix_scm_to_skb() and called from kfree_skb(). > > Please let me know if I need to add anything. > > Thanks
