Re: WARN_ON in commit 0ffdaf5 "net/sock: add WARN_ON(parent->sk) in sock_graft()"

On (08/04/17 19:13), Guoqing Jiang wrote:

HI,

With commit 0ffdaf5b41cf4435ece14d1d3e977ce69012a20d "net/sock: add
WARN_ON(parent->sk) in sock_graft()",
I can see the WARN_ON is triggered easily by dlm subsystem.

tcp_accept_from_sock() in  fs/dlm/lowcomms.c has the following snippet:

1.    result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
                                      SOCK_STREAM, IPPROTO_TCP, &newsock);
        sk is set by the path "sock_create_kern -> __sock_creat ->
pf->create => inet_create -> sock_init_data"
    ...
2.    result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK, true);
        Then sock_graft is called by "con->sock->ops->accept => inet_accept
-> sock_graft "

Does it mean call accept() after the just created socket is not a acceptable
behavior? Any comment will be appreciate, thanks.

It means that you have a sk leak, just like rds-tcp did. See
	https://patchwork.ozlabs.org/patch/780356/
for a discussion.

You should be calling sock_create_lite, instead of sock_create_kern.

The ->sk that triggers the WARN_ON is the one that is leaking (you will
see that the ->sk befrore the ->accept (i.e., the one set up by  sock_create_kern)
is different than the one after ->accept)

--Sowmini