On 2022/04/21 20:15, Tetsuo Handa wrote: >> I feel that amount of output above is too short for "char program[2053]". >> How can TCP/IPv6 socket be created from this quite limited operations? > > Since bpf_skb_load_helper_8() nothing but reads a byte, I don't think that > bpf(BPF_PROG_LOAD) / setsockopt(SOL_SOCKET, SO_ATTACH_BPF) can affect this > use-after-free bug, unless "char program[2053]" is doing something other > than reading a byte. It turned out that only first 37 bytes of "char program[2053]" was meaningful. const union bpf_attr attr = { .prog_type = BPF_PROG_TYPE_SOCKET_FILTER, .insn_cnt = 5, .insns = (unsigned long long) "\xbf\x16\x00\x00\x00\x00\x00\x00\xb7\x07\x00\x00\x01\x00\xf0\xff" "\x50\x70\x00\x00\x00\x00\x00\x00\x30\x00\x00\x00\x00\x00\xc0\x00" "\x95\x00\x00\x00\x00", .license = (unsigned long long) "GPL", }; Thus, I think that the output of "tools/bpf/bpftool/bpftool prog dump xlat id $NUM" was correct. Moreover, it turned out that even no-op program which does only r0 = 0 exit can trigger this use-after-free bug. const union bpf_attr attr = { .prog_type = BPF_PROG_TYPE_SOCKET_FILTER, .insn_cnt = 2, .insns = (unsigned long long) "\xb7\x00\x00\x00\x00\x00\x00\x00\x95\x00\x00\x00\x00\x00\x00\x00", .license = (unsigned long long) "GPL", }; That is, it seems that loading a BPF program using bpf(BPF_PROG_LOAD) and attaching that program using setsockopt(SOL_SOCKET, SO_ATTACH_BPF) itself somehow affects this use-after-free bug. Now, I suspect that some refcount is taken by setsockopt(SOL_SOCKET, SO_ATTACH_BPF) and is causing the socket to stay alive long enough to fire tcp_retransmit_timer(). Any ideas?