rxrpc: Fix congestion management

rxrpc has a problem in its congestion management in that it saves the congestion window size (cwnd) from one call to another, but if this is 0 at the time is saved, then the next call may not actually manage to ever transmit anything. To this end: (1) Don't save cwnd between calls, but rather reset back down to the initial cwnd and re-enter slow-start if data transmission is idle for more than an RTT. (2) Preserve ssthresh instead, as that is a handy estimate of pipe capacity. Knowing roughly when to stop slow start and enter congestion avoidance can reduce the tendency to overshoot and drop larger amounts of packets when probing. In future, cwind growth also needs to be constrained when the window isn't being filled due to being application limited.

David Howells committed 4 years ago

Commit 1fc4fa21 parent 6869ddb11 files · +60 −20Browse files

Unified · Split

11 files

include/trace/events/rxrpc.h +1 −0
net/rxrpc/ar-internal.h +5 −4
net/rxrpc/call_accept.c +2 −1
net/rxrpc/call_object.c +6 −1
net/rxrpc/conn_client.c +2 −1
net/rxrpc/conn_object.c +1 −1
net/rxrpc/input.c +20 −1
net/rxrpc/output.c +1 −0
net/rxrpc/peer_object.c +1 −6
net/rxrpc/proc.c +2 −2
net/rxrpc/sendmsg.c +19 −3

include/trace/events/rxrpc.h+1−0

@@ -193,6 +193,7 @@ · 192 unmodified lines
193	193	EM(rxrpc_cong_new_low_nack, " NewLowN") \
194	194	EM(rxrpc_cong_no_change, " -") \
195	195	EM(rxrpc_cong_progress, " Progres") \
	196	+ EM(rxrpc_cong_idle_reset, " IdleRes") \
196	197	EM(rxrpc_cong_retransmit_again, " ReTxAgn") \
197	198	EM(rxrpc_cong_rtt_window_end, " RttWinE") \
198	199	E_(rxrpc_cong_saw_nack, " SawNack")

net/rxrpc/ar-internal.h+5−4

@@ -332,7 +332,7 @@ struct rxrpc_peer { · 331 unmodified lines
332	332	u32 rto_j; /* Retransmission timeout in jiffies */
333	333	u8 backoff; /* Backoff timeout */
334	334
335		− u8 cong_cwnd; /* Congestion window size */
	335	+ u8 cong_ssthresh; /* Congestion slow-start threshold */
336	336	};
337	337
338	338	/*
@@ -626,6 +626,7 @@ struct rxrpc_call { · 287 unmodified lines
626	626	u16 tx_backoff; /* Delay to insert due to Tx failure */
627	627	u8 tx_winsize; /* Maximum size of Tx window */
628	628	#define RXRPC_TX_MAX_WINDOW 128
	629	+ ktime_t tx_last_sent; /* Last time a transmission occurred */
629	630
630	631	/* Received data tracking */
631	632	struct sk_buff_head recvmsg_queue; /* Queue of packets ready for recvmsg() */
@@ -687,10 +688,10 @@ struct rxrpc_call { · 55 unmodified lines
687	688	* Summary of a new ACK and the changes it made to the Tx buffer packet states.
688	689	*/
689	690	struct rxrpc_ack_summary {
	691	+ u16 nr_acks; /* Number of ACKs in packet */
	692	+ u16 nr_new_acks; /* Number of new ACKs in packet */
	693	+ u16 nr_rot_new_acks; /* Number of rotated new ACKs */
690	694	u8 ack_reason;
691		− u8 nr_acks; /* Number of ACKs in packet */
692		− u8 nr_new_acks; /* Number of new ACKs in packet */
693		− u8 nr_rot_new_acks; /* Number of rotated new ACKs */
694	695	bool saw_nacks; /* Saw NACKs in packet */
695	696	bool new_low_nack; /* T if new low NACK found */
696	697	bool retrans_timeo; /* T if reTx due to timeout happened */

net/rxrpc/call_accept.c+2−1

@@ -324,7 +324,8 @@ static struct rxrpc_call rxrpc_alloc_incoming_call(struct rxrpc_sock rx, · 323 unmodified lines
324	324	call->security = conn->security;
325	325	call->security_ix = conn->security_ix;
326	326	call->peer = rxrpc_get_peer(conn->params.peer);
327		− call->cong_cwnd = call->peer->cong_cwnd;
	327	+ call->cong_ssthresh = call->peer->cong_ssthresh;
	328	+ call->tx_last_sent = ktime_get_real();
328	329	return call;
329	330	}
330	331

net/rxrpc/call_object.c+6−1

@@ -166,7 +166,12 @@ struct rxrpc_call rxrpc_alloc_call(struct rxrpc_sock rx, gfp_t gfp, · 165 unmodified lines
166	166	call->rx_winsize = rxrpc_rx_window_size;
167	167	call->tx_winsize = 16;
168	168
169		− call->cong_cwnd = 2;
	169	+ if (RXRPC_TX_SMSS > 2190)
	170	+ call->cong_cwnd = 2;
	171	+ else if (RXRPC_TX_SMSS > 1095)
	172	+ call->cong_cwnd = 3;
	173	+ else
	174	+ call->cong_cwnd = 4;
170	175	call->cong_ssthresh = RXRPC_TX_MAX_WINDOW;
171	176
172	177	call->rxnet = rxnet;

net/rxrpc/conn_client.c+2−1

@@ -363,7 +363,8 @@ static struct rxrpc_bundle rxrpc_prep_call(struct rxrpc_sock rx, · 362 unmodified lines
363	363	if (!cp->peer)
364	364	goto error;
365	365
366		− call->cong_cwnd = cp->peer->cong_cwnd;
	366	+ call->tx_last_sent = ktime_get_real();
	367	+ call->cong_ssthresh = cp->peer->cong_ssthresh;
367	368	if (call->cong_cwnd >= call->cong_ssthresh)
368	369	call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
369	370	else

net/rxrpc/conn_object.c+1−1

@@ -207,7 +207,7 @@ void rxrpc_disconnect_call(struct rxrpc_call *call) · 206 unmodified lines
207	207	{
208	208	struct rxrpc_connection *conn = call->conn;
209	209
210		− call->peer->cong_cwnd = call->cong_cwnd;
	210	+ call->peer->cong_ssthresh = call->cong_ssthresh;
211	211
212	212	if (!hlist_unhashed(&call->error_link)) {
213	213	spin_lock_bh(&call->peer->lock);

net/rxrpc/input.c+20−1

@@ -58,6 +58,25 @@ static void rxrpc_congestion_management(struct rxrpc_call *call, · 57 unmodified lines
58	58	summary->cumulative_acks = cumulative_acks;
59	59	summary->dup_acks = call->cong_dup_acks;
60	60
	61	+ /* If we haven't transmitted anything for >1RTT, we should reset the
	62	+ * congestion management state.
	63	+ */
	64	+ if ((call->cong_mode == RXRPC_CALL_SLOW_START \|\|
	65	+ call->cong_mode == RXRPC_CALL_CONGEST_AVOIDANCE) &&
	66	+ ktime_before(ktime_add_us(call->tx_last_sent,
	67	+ call->peer->srtt_us >> 3),
	68	+ ktime_get_real())
	69	+ ) {
	70	+ change = rxrpc_cong_idle_reset;
	71	+ summary->mode = RXRPC_CALL_SLOW_START;
	72	+ if (RXRPC_TX_SMSS > 2190)
	73	+ summary->cwnd = 2;
	74	+ else if (RXRPC_TX_SMSS > 1095)
	75	+ summary->cwnd = 3;
	76	+ else
	77	+ summary->cwnd = 4;
	78	+ }
	79	+
61	80	switch (call->cong_mode) {
62	81	case RXRPC_CALL_SLOW_START:
63	82	if (summary->saw_nacks)
@@ -205,7 +224,7 @@ static bool rxrpc_rotate_tx_window(struct rxrpc_call *call, rxrpc_seq_t to, · 141 unmodified lines
205	224
206	225	if (call->acks_lowest_nak == call->acks_hard_ack) {
207	226	call->acks_lowest_nak = to;
208		− } else if (before_eq(call->acks_lowest_nak, to)) {
	227	+ } else if (after(to, call->acks_lowest_nak)) {
209	228	summary->new_low_nack = true;
210	229	call->acks_lowest_nak = to;
211	230	}

net/rxrpc/output.c+1−0

@@ -501,6 +501,7 @@ dont_set_request_ack: · 500 unmodified lines
501	501
502	502	done:
503	503	if (ret >= 0) {
	504	+ call->tx_last_sent = txb->last_sent;
504	505	if (txb->wire.flags & RXRPC_REQUEST_ACK) {
505	506	call->peer->rtt_last_req = txb->last_sent;
506	507	if (call->peer->rtt_count > 1) {

net/rxrpc/peer_object.c+1−6

@@ -227,12 +227,7 @@ struct rxrpc_peer rxrpc_alloc_peer(struct rxrpc_local local, gfp_t gfp) · 226 unmodified lines
227	227
228	228	rxrpc_peer_init_rtt(peer);
229	229
230		− if (RXRPC_TX_SMSS > 2190)
231		− peer->cong_cwnd = 2;
232		− else if (RXRPC_TX_SMSS > 1095)
233		− peer->cong_cwnd = 3;
234		− else
235		− peer->cong_cwnd = 4;
	230	+ peer->cong_ssthresh = RXRPC_TX_MAX_WINDOW;
236	231	trace_rxrpc_peer(peer->debug_id, rxrpc_peer_new, 1, here);
237	232	}
238	233

net/rxrpc/proc.c+2−2

@@ -217,7 +217,7 @@ static int rxrpc_peer_seq_show(struct seq_file seq, void v) · 216 unmodified lines
217	217	seq_puts(seq,
218	218	"Proto Local "
219	219	" Remote "
220		− " Use CW MTU LastUse RTT RTO\n"
	220	+ " Use SST MTU LastUse RTT RTO\n"
221	221	);
222	222	return 0;
223	223	}
@@ -235,7 +235,7 @@ static int rxrpc_peer_seq_show(struct seq_file seq, void v) · 11 unmodified lines
235	235	lbuff,
236	236	rbuff,
237	237	refcount_read(&peer->ref),
238		− peer->cong_cwnd,
	238	+ peer->cong_ssthresh,
239	239	peer->mtu,
240	240	now - peer->last_tx_at,
241	241	peer->srtt_us >> 3,

net/rxrpc/sendmsg.c+19−3

@@ -22,11 +22,27 @@ · 21 unmodified lines
22	22	*/
23	23	static bool rxrpc_check_tx_space(struct rxrpc_call call, rxrpc_seq_t _tx_win)
24	24	{
25		− unsigned int win_size =
26		− min_t(unsigned int, call->tx_winsize,
27		− call->cong_cwnd + call->cong_extra);
	25	+ unsigned int win_size;
28	26	rxrpc_seq_t tx_win = smp_load_acquire(&call->acks_hard_ack);
29	27
	28	+ /* If we haven't transmitted anything for >1RTT, we should reset the
	29	+ * congestion management state.
	30	+ */
	31	+ if (ktime_before(ktime_add_us(call->tx_last_sent,
	32	+ call->peer->srtt_us >> 3),
	33	+ ktime_get_real())) {
	34	+ if (RXRPC_TX_SMSS > 2190)
	35	+ win_size = 2;
	36	+ else if (RXRPC_TX_SMSS > 1095)
	37	+ win_size = 3;
	38	+ else
	39	+ win_size = 4;
	40	+ win_size += call->cong_extra;
	41	+ } else {
	42	+ win_size = min_t(unsigned int, call->tx_winsize,
	43	+ call->cong_cwnd + call->cong_extra);
	44	+ }
	45	+
30	46	if (_tx_win)
31	47	*_tx_win = tx_win;
32	48	return call->tx_top - tx_win < win_size;