bpf: sockmap, sock_map_delete needs to use xchg
authorJohn Fastabend <john.fastabend@gmail.com>
Fri, 19 Jul 2019 17:29:19 +0000 (10:29 -0700)
committerDaniel Borkmann <daniel@iogearbox.net>
Mon, 22 Jul 2019 14:04:17 +0000 (16:04 +0200)
commit45a4521dcbd92e71c9e53031b40e34211d3b4feb
tree08c6b20c0f3f00de939f03be1d6f57e941e78ae3
parent32857cf57f920cdc03b5095f08febec94cf9c36b
bpf: sockmap, sock_map_delete needs to use xchg

__sock_map_delete() may be called from a tcp event such as unhash or
close from the following trace,

  tcp_bpf_close()
    tcp_bpf_remove()
      sk_psock_unlink()
        sock_map_delete_from_link()
          __sock_map_delete()

In this case the sock lock is held but this only protects against
duplicate removals on the TCP side. If the map is free'd then we have
this trace,

  sock_map_free
    xchg()                  <- replaces map entry
    sock_map_unref()
      sk_psock_put()
        sock_map_del_link()

The __sock_map_delete() call however uses a read, test, null over the
map entry which can result in both paths trying to free the map
entry.

To fix use xchg in TCP paths as well so we avoid having two references
to the same map entry.

Fixes: 604326b41a6fb ("bpf, sockmap: convert to generic sk_msg interface")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
net/core/sock_map.c