haproxy/src/quic_loss.c
Frederic Lecaille 167e38e0e0 MINOR: quic: Add a counter for reordered packets
A packet is considered as reordered when it is detected as lost because its packet
number is above the largest acknowledeged packet number by at least the
packet reordering threshold value.

Add ->nb_reordered_pkt new quic_loss struct member at the same location that
the number of lost packets to count such packets.

Should be backported to 2.6.
2024-02-14 11:32:29 +01:00

313 lines
9.6 KiB
C

#include <import/eb64tree.h>
#include <haproxy/quic_conn-t.h>
#include <haproxy/quic_loss.h>
#include <haproxy/quic_tls.h>
#include <haproxy/quic_trace.h>
#include <haproxy/atomic.h>
#include <haproxy/list.h>
#include <haproxy/ticks.h>
#include <haproxy/trace.h>
/* Update <ql> QUIC loss information with new <rtt> measurement and <ack_delay>
* on ACK frame receipt which MUST be min(ack->ack_delay, max_ack_delay)
* before the handshake is confirmed.
*/
void quic_loss_srtt_update(struct quic_loss *ql,
unsigned int rtt, unsigned int ack_delay,
struct quic_conn *qc)
{
TRACE_ENTER(QUIC_EV_CONN_RTTUPDT, qc);
TRACE_PROTO("TX loss srtt update", QUIC_EV_CONN_RTTUPDT, qc, &rtt, &ack_delay, ql);
ql->latest_rtt = rtt;
if (!ql->rtt_min) {
/* No previous measurement. */
ql->srtt = rtt;
ql->rtt_var = rtt / 2;
ql->rtt_min = rtt;
}
else {
int diff;
ql->rtt_min = QUIC_MIN(rtt, ql->rtt_min);
/* Specific to QUIC (RTT adjustment). */
if (ack_delay && rtt >= ql->rtt_min + ack_delay)
rtt -= ack_delay;
diff = ql->srtt - rtt;
if (diff < 0)
diff = -diff;
ql->rtt_var = (3 * ql->rtt_var + diff) / 4;
ql->srtt = (7 * ql->srtt + rtt) / 8;
}
TRACE_PROTO("TX loss srtt update", QUIC_EV_CONN_RTTUPDT, qc,,, ql);
TRACE_LEAVE(QUIC_EV_CONN_RTTUPDT, qc);
}
/* Returns for <qc> QUIC connection the first packet number space which
* experienced packet loss, if any or a packet number space with
* TICK_ETERNITY as packet loss time if not.
*/
struct quic_pktns *quic_loss_pktns(struct quic_conn *qc)
{
struct quic_pktns *pktns, *p;
TRACE_ENTER(QUIC_EV_CONN_SPTO, qc);
BUG_ON(LIST_ISEMPTY(&qc->pktns_list));
pktns = p = LIST_NEXT(&qc->pktns_list, struct quic_pktns *, list);
do {
TRACE_PROTO("TX loss pktns", QUIC_EV_CONN_SPTO, qc, p);
if (!tick_isset(pktns->tx.loss_time) ||
tick_is_lt(p->tx.loss_time, pktns->tx.loss_time)) {
pktns = p;
}
p = LIST_NEXT(&p->list, struct quic_pktns *, list);
} while (&p->list != &qc->pktns_list);
TRACE_LEAVE(QUIC_EV_CONN_SPTO, qc);
return pktns;
}
/* Returns for <qc> QUIC connection the first packet number space to
* arm the PTO for if any or a packet number space with TICK_ETERNITY
* as PTO value if not.
*/
struct quic_pktns *quic_pto_pktns(struct quic_conn *qc,
int handshake_confirmed,
unsigned int *pto)
{
unsigned int duration, lpto;
struct quic_loss *ql = &qc->path->loss;
struct quic_pktns *pktns, *p;
TRACE_ENTER(QUIC_EV_CONN_SPTO, qc);
BUG_ON(LIST_ISEMPTY(&qc->pktns_list));
duration =
ql->srtt +
(QUIC_MAX(4 * ql->rtt_var, QUIC_TIMER_GRANULARITY) << ql->pto_count);
/* RFC 9002 6.2.2.1. Before Address Validation
*
* the client MUST set the PTO timer if the client has not received an
* acknowledgment for any of its Handshake packets and the handshake is
* not confirmed (see Section 4.1.2 of [QUIC-TLS]), even if there are no
* packets in flight.
*
* TODO implement the above paragraph for QUIC on backend side. Note
* that if now_ms is used this function is not reentrant anymore and can
* not be used anytime without side-effect (for example after QUIC
* connection migration).
*/
lpto = TICK_ETERNITY;
pktns = p = LIST_NEXT(&qc->pktns_list, struct quic_pktns *, list);
do {
unsigned int tmp_pto;
if (p->tx.in_flight) {
if (p == qc->apktns) {
if (!handshake_confirmed) {
TRACE_STATE("TX PTO handshake not already confirmed", QUIC_EV_CONN_SPTO, qc);
goto out;
}
duration += qc->max_ack_delay << ql->pto_count;
}
tmp_pto = tick_add(p->tx.time_of_last_eliciting, duration);
if (!tick_isset(lpto) || tick_is_lt(tmp_pto, lpto)) {
lpto = tmp_pto;
pktns = p;
}
TRACE_PROTO("TX PTO", QUIC_EV_CONN_SPTO, qc, p);
}
p = LIST_NEXT(&p->list, struct quic_pktns *, list);
} while (&p->list != &qc->pktns_list);
out:
if (pto)
*pto = lpto;
TRACE_PROTO("TX PTO", QUIC_EV_CONN_SPTO, qc, pktns, &duration);
TRACE_LEAVE(QUIC_EV_CONN_SPTO, qc);
return pktns;
}
/* Look for packet loss from sent packets for <qel> encryption level of a
* connection with <ctx> as I/O handler context. If remove is true, remove them from
* their tree if deemed as lost or set the <loss_time> value the packet number
* space if any not deemed lost.
* Should be called after having received an ACK frame with newly acknowledged
* packets or when the the loss detection timer has expired.
* Always succeeds.
*/
void qc_packet_loss_lookup(struct quic_pktns *pktns, struct quic_conn *qc,
struct list *lost_pkts)
{
struct eb_root *pkts;
struct eb64_node *node;
struct quic_loss *ql;
unsigned int loss_delay;
uint64_t pktthresh;
TRACE_ENTER(QUIC_EV_CONN_PKTLOSS, qc);
TRACE_PROTO("TX loss", QUIC_EV_CONN_PKTLOSS, qc, pktns);
pkts = &pktns->tx.pkts;
pktns->tx.loss_time = TICK_ETERNITY;
if (eb_is_empty(pkts))
goto out;
ql = &qc->path->loss;
loss_delay = QUIC_MAX(ql->latest_rtt, ql->srtt);
loss_delay = QUIC_MAX(loss_delay, MS_TO_TICKS(QUIC_TIMER_GRANULARITY)) *
QUIC_LOSS_TIME_THRESHOLD_MULTIPLICAND / QUIC_LOSS_TIME_THRESHOLD_DIVISOR;
node = eb64_first(pkts);
/* RFC 9002 6.1.1. Packet Threshold
* The RECOMMENDED initial value for the packet reordering threshold
* (kPacketThreshold) is 3, based on best practices for TCP loss detection
* [RFC5681] [RFC6675]. In order to remain similar to TCP, implementations
* SHOULD NOT use a packet threshold less than 3; see [RFC5681].
* Some networks may exhibit higher degrees of packet reordering, causing a
* sender to detect spurious losses. Additionally, packet reordering could be
* more common with QUIC than TCP because network elements that could observe
* and reorder TCP packets cannot do that for QUIC and also because QUIC
* packet numbers are encrypted.
*/
/* Dynamic packet reordering threshold calculation depending on the distance
* (in packets) between the last transmitted packet and the oldest still in
* flight before loss detection.
*/
pktthresh = pktns->tx.next_pn - 1 - eb64_entry(node, struct quic_tx_packet, pn_node)->pn_node.key;
/* Apply a ratio to this threshold and add it to QUIC_LOSS_PACKET_THRESHOLD. */
pktthresh = pktthresh * global.tune.quic_reorder_ratio / 100 + QUIC_LOSS_PACKET_THRESHOLD;
while (node) {
struct quic_tx_packet *pkt;
int64_t largest_acked_pn;
unsigned int loss_time_limit, time_sent;
int reordered;
pkt = eb64_entry(&node->node, struct quic_tx_packet, pn_node);
largest_acked_pn = pktns->rx.largest_acked_pn;
node = eb64_next(node);
if ((int64_t)pkt->pn_node.key > largest_acked_pn)
break;
time_sent = pkt->time_sent;
loss_time_limit = tick_add(time_sent, loss_delay);
reordered = (int64_t)largest_acked_pn >= pkt->pn_node.key + pktthresh;
if (reordered)
ql->nb_reordered_pkt++;
if (tick_is_le(loss_time_limit, now_ms) || reordered) {
eb64_delete(&pkt->pn_node);
LIST_APPEND(lost_pkts, &pkt->list);
ql->nb_lost_pkt++;
}
else {
if (tick_isset(pktns->tx.loss_time))
pktns->tx.loss_time = tick_first(pktns->tx.loss_time, loss_time_limit);
else
pktns->tx.loss_time = loss_time_limit;
break;
}
}
out:
TRACE_PROTO("TX loss", QUIC_EV_CONN_PKTLOSS, qc, pktns, lost_pkts);
TRACE_LEAVE(QUIC_EV_CONN_PKTLOSS, qc);
}
/* Handle <pkts> list of lost packets detected at <now_us> handling their TX
* frames. Send a packet loss event to the congestion controller if in flight
* packet have been lost. Also frees the packet in <pkts> list.
*
* Returns 1 on success else 0 if loss limit has been exceeded. A
* CONNECTION_CLOSE was prepared to close the connection ASAP.
*/
int qc_release_lost_pkts(struct quic_conn *qc, struct quic_pktns *pktns,
struct list *pkts, uint64_t now_us)
{
struct quic_tx_packet *pkt, *tmp, *oldest_lost, *newest_lost;
int close = 0;
TRACE_ENTER(QUIC_EV_CONN_PRSAFRM, qc);
if (LIST_ISEMPTY(pkts))
goto leave;
oldest_lost = newest_lost = NULL;
list_for_each_entry_safe(pkt, tmp, pkts, list) {
struct list tmp = LIST_HEAD_INIT(tmp);
pkt->pktns->tx.in_flight -= pkt->in_flight_len;
qc->path->prep_in_flight -= pkt->in_flight_len;
qc->path->in_flight -= pkt->in_flight_len;
if (pkt->flags & QUIC_FL_TX_PACKET_ACK_ELICITING)
qc->path->ifae_pkts--;
/* Treat the frames of this lost packet. */
if (!qc_handle_frms_of_lost_pkt(qc, pkt, &pktns->tx.frms))
close = 1;
LIST_DELETE(&pkt->list);
if (!oldest_lost) {
oldest_lost = newest_lost = pkt;
}
else {
if (newest_lost != oldest_lost)
quic_tx_packet_refdec(newest_lost);
newest_lost = pkt;
}
}
if (!close) {
if (newest_lost) {
/* Sent a congestion event to the controller */
struct quic_cc_event ev = { };
ev.type = QUIC_CC_EVT_LOSS;
ev.loss.time_sent = newest_lost->time_sent;
quic_cc_event(&qc->path->cc, &ev);
}
/* If an RTT have been already sampled, <rtt_min> has been set.
* We must check if we are experiencing a persistent congestion.
* If this is the case, the congestion controller must re-enter
* slow start state.
*/
if (qc->path->loss.rtt_min && newest_lost != oldest_lost) {
unsigned int period = newest_lost->time_sent - oldest_lost->time_sent;
if (quic_loss_persistent_congestion(&qc->path->loss, period,
now_ms, qc->max_ack_delay))
qc->path->cc.algo->slow_start(&qc->path->cc);
}
}
/* <oldest_lost> cannot be NULL at this stage because we have ensured
* that <pkts> list is not empty. Without this, GCC 12.2.0 reports a
* possible overflow on a 0 byte region with O2 optimization.
*/
ALREADY_CHECKED(oldest_lost);
quic_tx_packet_refdec(oldest_lost);
if (newest_lost != oldest_lost)
quic_tx_packet_refdec(newest_lost);
leave:
TRACE_LEAVE(QUIC_EV_CONN_PRSAFRM, qc);
return !close;
}