/*
* be very careful here. This is being called as the condition in
* wait_event_*() needs to cope with being called many times.
*/
static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc)
{
unsigned long flags;
if (!*inc) {
read_lock_irqsave(&rs->rs_recv_lock, flags);
if (!list_empty(&rs->rs_recv_queue)) {
*inc = list_entry(rs->rs_recv_queue.next,
struct rds_incoming,
i_item);
rds_inc_addref(*inc);
}
/*
* 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 rds_sock *rs = NULL;
struct sock *sk;
unsigned long flags;
struct rds_conn_path *cp;
inc->i_conn = conn;
inc->i_rx_jiffies = jiffies;
if (conn->c_trans->t_mp_capable)
cp = inc->i_conn_path;
else
cp = &conn->c_path[0];
rdsdebug("conn %p next %llu inc %p seq %llu len %u sport %u dport %u "
"flags 0x%x rx_jiffies %lu\n", conn,
(unsigned long long)cp->cp_next_rx_seq,
inc,
(unsigned long long)be64_to_cpu(inc->i_hdr.h_sequence),
be32_to_cpu(inc->i_hdr.h_len),
be16_to_cpu(inc->i_hdr.h_sport),
be16_to_cpu(inc->i_hdr.h_dport),
inc->i_hdr.h_flags,
inc->i_rx_jiffies);
/*
* Sequence numbers should only increase. Messages get their
* sequence number as they're queued in a sending conn. They
* can be dropped, though, if the sending socket is closed before
* they hit the wire. So sequence numbers can skip forward
* under normal operation. They can also drop back in the conn
* failover case as previously sent messages are resent down the
* new instance of a conn. We drop those, otherwise we have
* to assume that the next valid seq does not come after a
* hole in the fragment stream.
*
* The headers don't give us a way to realize if fragments of
* a message have been dropped. We assume that frags that arrive
* to a flow are part of the current message on the flow that is
* being reassembled. This means that senders can't drop messages
* from the sending conn until all their frags are sent.
*
* XXX we could spend more on the wire to get more robust failure
* detection, arguably worth it to avoid data corruption.
*/
if (be64_to_cpu(inc->i_hdr.h_sequence) < cp->cp_next_rx_seq &&
(inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) {
rds_stats_inc(s_recv_drop_old_seq);
goto out;
}
cp->cp_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1;
if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) {
if (inc->i_hdr.h_sport == 0) {
rdsdebug("ignore ping with 0 sport from 0x%x\n", saddr);
goto out;
}
rds_stats_inc(s_recv_ping);
rds_send_pong(cp, inc->i_hdr.h_sport);
/* if this is a handshake ping, start multipath if necessary */
if (RDS_HS_PROBE(inc->i_hdr.h_sport, inc->i_hdr.h_dport)) {
rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn);
rds_start_mprds(cp->cp_conn);
}
goto out;
}
if (inc->i_hdr.h_dport == RDS_FLAG_PROBE_PORT &&
inc->i_hdr.h_sport == 0) {
rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn);
/* if this is a handshake pong, start multipath if necessary */
rds_start_mprds(cp->cp_conn);
wake_up(&cp->cp_conn->c_hs_waitq);
goto out;
}
rs = rds_find_bound(daddr, inc->i_hdr.h_dport);
if (!rs) {
rds_stats_inc(s_recv_drop_no_sock);
goto out;
}
/* Process extension headers */
rds_recv_incoming_exthdrs(inc, rs);
/* We can be racing with rds_release() which marks the socket dead. */
sk = rds_rs_to_sk(rs);
/* serialize with rds_release -> sock_orphan */
write_lock_irqsave(&rs->rs_recv_lock, flags);
if (!sock_flag(sk, SOCK_DEAD)) {
rdsdebug("adding inc %p to rs %p's recv queue\n", inc, rs);
rds_stats_inc(s_recv_queued);
rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
be32_to_cpu(inc->i_hdr.h_len),
inc->i_hdr.h_dport);
if (sock_flag(sk, SOCK_RCVTSTAMP))
do_gettimeofday(&inc->i_rx_tstamp);
rds_inc_addref(inc);
list_add_tail(&inc->i_item, &rs->rs_recv_queue);
__rds_wake_sk_sleep(sk);
} else {
rds_stats_inc(s_recv_drop_dead_sock);
}
write_unlock_irqrestore(&rs->rs_recv_lock, flags);
out:
if (rs)
rds_sock_put(rs);
}
/*
* 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));
}
}
