/*
* Before killing the tcp socket this needs to serialize with callbacks. The
* caller has already grabbed the sending sem so we're serialized with other
* senders.
*
* TCP calls the callbacks with the sock lock so we hold it while we reset the
* callbacks to those set by TCP. Our callbacks won't execute again once we
* hold the sock lock.
*/
void rds_tcp_conn_path_shutdown(struct rds_conn_path *cp)
{
struct rds_tcp_connection *tc = cp->cp_transport_data;
struct socket *sock = tc->t_sock;
rdsdebug("shutting down conn %p tc %p sock %p\n",
cp->cp_conn, tc, sock);
if (sock) {
if (rds_destroy_pending(cp->cp_conn))
rds_tcp_set_linger(sock);
sock->ops->shutdown(sock, RCV_SHUTDOWN | SEND_SHUTDOWN);
lock_sock(sock->sk);
rds_tcp_restore_callbacks(sock, tc); /* tc->tc_sock = NULL */
release_sock(sock->sk);
sock_release(sock);
}
if (tc->t_tinc) {
rds_inc_put(&tc->t_tinc->ti_inc);
tc->t_tinc = NULL;
}
tc->t_tinc_hdr_rem = sizeof(struct rds_header);
tc->t_tinc_data_rem = 0;
}
/*
* The transport must make sure that this is serialized against other
* rx and conn reset on this specific conn.
*
* We currently assert that only one fragmented message will be sent
* down a connection at a time. This lets us reassemble in the conn
* instead of per-flow which means that we don't have to go digging through
* flows to tear down partial reassembly progress on conn failure and
* we save flow lookup and locking for each frag arrival. It does mean
* that small messages will wait behind large ones. Fragmenting at all
* is only to reduce the memory consumption of pre-posted buffers.
*
* The caller passes in saddr and daddr instead of us getting it from the
* conn. This lets loopback, who only has one conn for both directions,
* tell us which roles the addrs in the conn are playing for this message.
*/
void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
struct rds_incoming *inc, gfp_t gfp)
{
struct sk_buff *skb;
struct rds_sock *rs;
struct sock *sk;
struct rds_nf_hdr *dst, *org;
int ret;
rdsdebug(KERN_ALERT "incoming: conn %p, inc %p, %u.%u.%u.%u : %d -> %u.%u.%u.%u : %d\n",
conn, inc, NIPQUAD(saddr), inc->i_hdr.h_sport, NIPQUAD(daddr), inc->i_hdr.h_dport);
/* initialize some globals */
rs = NULL;
sk = NULL;
/* save off the original connection against which the request arrived */
inc->i_oconn = conn;
inc->i_skb = NULL;
/* lets find a socket to which this request belongs */
rs = rds_find_bound(daddr, inc->i_hdr.h_dport);
/* pass it on locally if there is no socket bound, or if netfilter is
* disabled for this socket */
if (NULL == rs || !rs->rs_netfilter_enabled) {
/* drop the reference if we had taken one */
if (NULL != rs)
rds_sock_put(rs);
rds_recv_local(conn, saddr, daddr, inc, gfp);
return;
}
/* otherwise pull out the socket */
sk = rds_rs_to_sk(rs);
/* create an skb with some additional space to store our rds_nf_hdr info */
skb = alloc_skb(sizeof(struct rds_nf_hdr) * 2, gfp);
if (NULL == skb) {
/* if we have allocation problems, then we just need to depart */
rdsdebug("failure to allocate space for inc %p, %u.%u.%u.%u -> %u.%d.%u.%u\n",
inc, NIPQUAD(saddr), NIPQUAD(daddr));
rds_recv_local(conn, saddr, daddr, inc, gfp);
return;
}
/* once we've allocated an skb, also store it in our structures */
inc->i_skb = skb;
/* now pull out the rds headers */
dst = rds_nf_hdr_dst(skb);
org = rds_nf_hdr_org(skb);
/* now update our rds_nf_hdr for tracking locations of the request */
dst->saddr = saddr;
dst->daddr = daddr;
dst->sport = inc->i_hdr.h_sport;
dst->dport = inc->i_hdr.h_dport;
dst->flags = 0;
/* assign the appropriate protocol if any */
if (NULL != sk) {
dst->protocol = sk->sk_protocol;
dst->sk = sk;
} else {
dst->protocol = 0;
dst->sk = NULL;
}
/* cleanup any references taken */
if (NULL != rs)
rds_sock_put(rs);
/* the original info is just a copy */
memcpy(org, dst, sizeof(struct rds_nf_hdr));
/* convert our local data structures in the message to a generalized skb form */
if (conn->c_trans->inc_to_skb(inc, skb)) {
rdsdebug("handing off to PRE_ROUTING hook\n");
/* call down through the hook layers */
ret = NF_HOOK(PF_RDS_HOOK, NF_RDS_PRE_ROUTING, skb, NULL, NULL, rds_recv_ok);
}
/* if we had a failure to convert, then just assuming to continue as local */
else {
rdsdebug("failed to create skb form, conn %p, inc %p, %u.%u.%u.%u -> %u.%u.%u.%u\n",
conn, inc, NIPQUAD(saddr), NIPQUAD(daddr));
ret = 1;
}
/* pull back out the rds headers */
dst = rds_nf_hdr_dst(skb);
org = rds_nf_hdr_org(skb);
/* now depending upon we got back we can perform appropriate activities */
if (dst->flags & RDS_NF_HDR_FLAG_DONE) {
rds_recv_drop(conn, saddr, daddr, inc, gfp);
}
/* this is the normal good processed state */
else if (ret >= 0) {
/* check the original header and if changed do the needful */
if (dst->saddr == org->saddr && dst->daddr == org->daddr &&
conn->c_trans->skb_local(skb)) {
rds_recv_local(conn, saddr, daddr, inc, gfp);
}
/* the send both case does both a local recv and a reroute */
else if (dst->flags & RDS_NF_HDR_FLAG_BOTH) {
/* we must be sure to take an extra reference on the inc
* to be sure it doesn't accidentally get freed in between */
rds_inc_addref(inc);
/* send it up the stream locally */
rds_recv_local(conn, saddr, daddr, inc, gfp);
/* and also reroute the request */
rds_recv_route(conn, inc, gfp);
/* since we are done with processing we can drop this additional reference */
rds_inc_put(inc);
}
/* anything else is a change in possible destination so pass to route */
else
rds_recv_route(conn, inc, gfp);
}
/* we don't really expect an error state from this call that isn't the done above */
else {
/* we don't really know how to handle this yet - just ignore for now */
printk(KERN_ERR "unacceptible state for skb ret %d, conn %p, inc %p, "
"%u.%u.%u.%u -> %u.%u.%u.%u\n",
ret, conn, inc, NIPQUAD(saddr), NIPQUAD(daddr));
}
}