/**
* Dump relayed or locally-emitted packet.
* If ``from'' is NULL, packet was emitted locally.
*/
static void
dump_packet_from_to(struct dump *dump,
const struct gnutella_node *from, const struct gnutella_node *to,
const pmsg_t *mb)
{
struct dump_header dh_to;
struct dump_header dh_from;
g_assert(to != NULL);
g_assert(mb != NULL);
g_assert(pmsg_read_base(mb) == pmsg_start(mb));
if (!dump_initialize(dump))
return;
/*
* This is only for Gnutella packets, leave DHT messages out.
*/
if (GTA_MSG_DHT == gnutella_header_get_function(pmsg_start(mb)))
return;
if (!ipset_contains_addr(&dump_tx_to_addrs, to->addr, TRUE))
return;
if (NULL == from) {
struct gnutella_node local;
local.peermode = NODE_IS_UDP(to) ? NODE_P_UDP : NODE_P_NORMAL;
local.addr = listen_addr();
local.port = GNET_PROPERTY(listen_port);
if (!ipset_contains_addr(&dump_tx_from_addrs, local.addr, TRUE))
return;
dump_header_set(&dh_from, &local);
} else {
if (!ipset_contains_addr(&dump_tx_from_addrs, from->addr, TRUE))
return;
dump_header_set(&dh_from, from);
}
dump_header_set(&dh_to, to);
dh_to.data[0] |= DH_F_TO;
if (pmsg_prio(mb) != PMSG_P_DATA)
dh_to.data[0] |= DH_F_CTRL;
dump_append(dump, dh_to.data, sizeof dh_to.data);
dump_append(dump, dh_from.data, sizeof dh_from.data);
dump_append(dump, pmsg_read_base(mb), pmsg_size(mb));
dump_flush(dump);
}
/**
* Send message (eslist iterator callback).
*
* @return TRUE if message was sent and freed up.
*/
static bool
udp_tx_desc_send(void *data, void *udata)
{
struct udp_tx_desc *txd = data;
udp_sched_t *us = udata;
unsigned prio;
udp_sched_check(us);
udp_tx_desc_check(txd);
if (us->used_all)
return FALSE;
/*
* Avoid flushing consecutive queued messages to the same destination,
* for regular (non-prioritary) messages.
*
* This serves two purposes:
*
* 1- It makes sure one single host does not capture all the available
* outgoing bandwidth.
*
* 2- It somehow delays consecutive packets to a given host thereby reducing
* flooding and hopefully avoiding saturation of its RX flow.
*/
prio = pmsg_prio(txd->mb);
if (PMSG_P_DATA == prio && hset_contains(us->seen, txd->to)) {
udp_sched_log(2, "%p: skipping mb=%p (%d bytes) to %s",
us, txd->mb, pmsg_size(txd->mb), gnet_host_to_string(txd->to));
return FALSE;
}
if (udp_sched_mb_sendto(us, txd->mb, txd->to, txd->tx, txd->cb)) {
if (PMSG_P_DATA == prio && pmsg_was_sent(txd->mb))
hset_insert(us->seen, atom_host_get(txd->to));
} else {
return FALSE; /* Unsent, leave it in the queue */
}
us->buffered = size_saturate_sub(us->buffered, pmsg_size(txd->mb));
udp_tx_desc_flag_release(txd, us);
return TRUE;
}
/**
* Remove expired messages (eslist iterator).
*
* @return TRUE if message has expired and was freed up.
*/
static bool
udp_tx_desc_expired(void *data, void *udata)
{
struct udp_tx_desc *txd = data;
udp_sched_t *us = udata;
udp_sched_check(us);
udp_tx_desc_check(txd);
if (delta_time(tm_time(), txd->expire) > 0) {
udp_sched_log(1, "%p: expiring mb=%p (%d bytes) prio=%u",
us, txd->mb, pmsg_size(txd->mb), pmsg_prio(txd->mb));
if (txd->cb->add_tx_dropped != NULL)
(*txd->cb->add_tx_dropped)(txd->tx->owner, 1); /* Dropped in TX */
return udp_tx_desc_drop(data, udata); /* Returns TRUE */
}
return FALSE;
}
/**
* Send datagram.
*
* @param us the UDP scheduler responsible for sending the datagram
* @param mb the message to send
* @param to the IP:port destination of the message
* @param tx the TX stack sending the message
* @param cb callback actions on the datagram
*
* @return 0 if message was unsent, length of message if sent, queued or
* dropped.
*/
size_t
udp_sched_send(udp_sched_t *us, pmsg_t *mb, const gnet_host_t *to,
const txdrv_t *tx, const struct tx_dgram_cb *cb)
{
int len;
struct udp_tx_desc *txd;
uint prio;
len = pmsg_size(mb);
/*
* Try to send immediately if we have bandwidth.
*/
if (!us->used_all && udp_sched_mb_sendto(us, mb, to, tx, cb))
return len; /* Message "sent" */
/*
* If we already have enough data enqueued, flow-control the upper
* layer by acting as if we do not have enough bandwidth.
*
* However, we now always accept traffic sent with the highest priority
* since it is important to send those as soon as possible, i.e. ahead
* of any other pending data we would otherwise flush locally before
* servicing upper queues.
* --RAM, 2012-10-12
*/
prio = pmsg_prio(mb);
if (
PMSG_P_HIGHEST != prio &&
us->buffered >= UDP_SCHED_FACTOR * udp_sched_bw_per_second(us)
) {
udp_sched_log(1, "%p: flow-controlled", us);
us->flow_controlled = TRUE;
return 0; /* Flow control upper layers */
}
/*
* Message is going to be enqueued.
*
* However, from the upper layers (the message queue in particular),
* the message is considered as being sent, and therefore these layers
* are going to call pmsg_free() on the message.
*
* We do not want to pmsg_clone() the message because that would render
* uses of pmsg_was_sent() useless in free routines, and upper layers
* would think the message was dropped if they installed a free routine
* on the message.
*
* Hence we use pmsg_ref().
*/
txd = palloc(us->txpool);
txd->magic = UDP_TX_DESC_MAGIC;
txd->mb = pmsg_ref(mb); /* Take ownership of message */
txd->to = atom_host_get(to);
txd->tx = tx;
txd->cb = cb;
txd->expire = time_advance(tm_time(), UDP_SCHED_EXPIRE);
udp_sched_log(4, "%p: queuing mb=%p (%d bytes) prio=%u",
us, mb, pmsg_size(mb), pmsg_prio(mb));
/*
* The queue used is a LIFO to avoid buffering delaying all the messages.
* Since UDP traffic is unordered, it's better to send the most recent
* datagrams first, to reduce the perceived average latency.
*/
g_assert(prio < N_ITEMS(us->lifo));
eslist_prepend(&us->lifo[prio], txd);
us->buffered = size_saturate_add(us->buffered, len);
return len; /* Message queued, but tell upper layers it's sent */
}
/**
* Send message block to IP:port.
*
* @param us the UDP scheduler
* @param mb the message to send
* @param to the IP:port destination of the message
* @param tx the TX stack sending the message
* @param cb callback actions on the datagram
*
* @return TRUE if message was sent or dropped, FALSE if there is no more
* bandwidth to send anything.
*/
static bool
udp_sched_mb_sendto(udp_sched_t *us, pmsg_t *mb, const gnet_host_t *to,
const txdrv_t *tx, const struct tx_dgram_cb *cb)
{
ssize_t r;
int len = pmsg_size(mb);
bio_source_t *bio = NULL;
if (0 == gnet_host_get_port(to))
return TRUE;
/*
* Check whether message still needs to be sent.
*/
if (!pmsg_hook_check(mb))
return TRUE; /* Dropped */
/*
* Select the proper I/O source depending on the network address type.
*/
switch (gnet_host_get_net(to)) {
case NET_TYPE_IPV4:
bio = us->bio[UDP_SCHED_IPv4];
break;
case NET_TYPE_IPV6:
bio = us->bio[UDP_SCHED_IPv6];
break;
case NET_TYPE_NONE:
case NET_TYPE_LOCAL:
g_assert_not_reached();
}
/*
* If there is no I/O source, then the socket to send that type of traffic
* was cleared, hence we simply need to discard the message.
*/
if (NULL == bio) {
udp_sched_log(4, "%p: discarding mb=%p (%d bytes) to %s",
us, mb, pmsg_size(mb), gnet_host_to_string(to));
return udp_tx_drop(tx, cb); /* TRUE, for "sent" */
}
/*
* OK, proceed if we have bandwidth.
*/
r = bio_sendto(bio, to, pmsg_start(mb), len);
if (r < 0) { /* Error, or no bandwidth */
if (udp_sched_write_error(us, to, mb, G_STRFUNC)) {
udp_sched_log(4, "%p: dropped mb=%p (%d bytes): %m",
us, mb, pmsg_size(mb));
return udp_tx_drop(tx, cb); /* TRUE, for "sent" */
}
udp_sched_log(3, "%p: no bandwidth for mb=%p (%d bytes)",
us, mb, pmsg_size(mb));
us->used_all = TRUE;
return FALSE;
}
if (r != len) {
g_warning("%s: partial UDP write (%zd bytes) to %s "
"for %d-byte datagram",
G_STRFUNC, r, gnet_host_to_string(to), len);
} else {
udp_sched_log(5, "%p: sent mb=%p (%d bytes) prio=%u",
us, mb, pmsg_size(mb), pmsg_prio(mb));
pmsg_mark_sent(mb);
if (cb->msg_account != NULL)
(*cb->msg_account)(tx->owner, mb);
inet_udp_record_sent(gnet_host_get_addr(to));
}
return TRUE; /* Message sent */
}
