This adds transmit buffering to DCCP.
I have tested with CCID2/3 and with loss and rate limiting.
The only slight downside I have observed is that there can be memory pressure
on the receiver as they receive packets faster. However I can reproduce
without my patch also if I send fast enough. I believe that this is due to
lack of buffer limiting and/or slight flaws in congestion algorithm - neither
of which is due to this patch! I checked for memory consumption on the
transmitter and did not observe any problems here.
I would like this to be considered for 2.6.18 and I believe it could help
Andrea's work with CCID2.
Signed-off-by: Ian McDonald <ian.mcdonald@xxxxxxxxxxx>
---
diff --git a/include/linux/dccp.h b/include/linux/dccp.h
index 676333b..2d7671c 100644
--- a/include/linux/dccp.h
+++ b/include/linux/dccp.h
@@ -438,6 +438,7 @@ struct dccp_ackvec;
* @dccps_role - Role of this sock, one of %dccp_role
* @dccps_ndp_count - number of Non Data Packets since last data packet
* @dccps_hc_rx_ackvec - rx half connection ack vector
+ * @dccps_xmit_timer - timer for when CCID is not ready to send
*/
struct dccp_sock {
/* inet_connection_sock has to be the first member of dccp_sock */
@@ -470,6 +471,7 @@ struct dccp_sock {
enum dccp_role dccps_role:2;
__u8 dccps_hc_rx_insert_options:1;
__u8 dccps_hc_tx_insert_options:1;
+ struct timer_list dccps_xmit_timer;
};
static inline struct dccp_sock *dccp_sk(const struct sock *sk)
diff --git a/net/dccp/dccp.h b/net/dccp/dccp.h
index 1fe5091..1ba5ac5 100644
--- a/net/dccp/dccp.h
+++ b/net/dccp/dccp.h
@@ -5,7 +5,7 @@ #define _DCCP_H
*
* An implementation of the DCCP protocol
* Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@xxxxxxxxxxxxxxxx>
- * Copyright (c) 2005 Ian McDonald <iam4@xxxxxxxxxxxxxxxx>
+ * Copyright (c) 2005-6 Ian McDonald <ian.mcdonald@xxxxxxxxxxx>
*
* This program is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 as
@@ -123,7 +123,7 @@ extern void dccp_send_delayed_ack(struct
extern void dccp_send_sync(struct sock *sk, const u64 seq,
const enum dccp_pkt_type pkt_type);
-extern int dccp_write_xmit(struct sock *sk, struct sk_buff *skb, long *timeo);
+extern void dccp_write_xmit(struct sock *sk, int block);
extern void dccp_write_space(struct sock *sk);
extern void dccp_init_xmit_timers(struct sock *sk);
diff --git a/net/dccp/output.c b/net/dccp/output.c
index 7409e4a..73da2b6 100644
--- a/net/dccp/output.c
+++ b/net/dccp/output.c
@@ -199,7 +199,7 @@ static int dccp_wait_for_ccid(struct soc
while (1) {
prepare_to_wait(sk->sk_sleep, &wait, TASK_INTERRUPTIBLE);
- if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
+ if (sk->sk_err)
goto do_error;
if (!*timeo)
goto do_nonblock;
@@ -235,37 +235,70 @@ do_interrupted:
goto out;
}
-int dccp_write_xmit(struct sock *sk, struct sk_buff *skb, long *timeo)
+static void dccp_write_xmit_timer(unsigned long data) {
+ struct sock *sk = (struct sock *)data;
+ struct dccp_sock *dp = dccp_sk(sk);
+
+ bh_lock_sock(sk);
+ if (sock_owned_by_user(sk))
+ sk_reset_timer(sk, &dp->dccps_xmit_timer, jiffies+1);
+ else
+ dccp_write_xmit(sk, 0);
+ bh_unlock_sock(sk);
+ sock_put(sk);
+}
+
+void dccp_write_xmit(struct sock *sk, int block)
{
- const struct dccp_sock *dp = dccp_sk(sk);
- int err = ccid_hc_tx_send_packet(dp->dccps_hc_tx_ccid, sk, skb,
- skb->len);
+ struct dccp_sock *dp = dccp_sk(sk);
+ struct sk_buff *skb;
+ long timeo = 2000; /* If a packet is taking longer than 2 secs
+ we have other issues */
- if (err > 0)
- err = dccp_wait_for_ccid(sk, skb, timeo);
+ while ((skb = skb_peek(&sk->sk_write_queue))) {
+ int err = ccid_hc_tx_send_packet(dp->dccps_hc_tx_ccid, sk, skb,
+ skb->len);
+
+ if (err > 0) {
+ if (!block) {
+ sk_reset_timer(sk, &dp->dccps_xmit_timer,
+ msecs_to_jiffies(err)+jiffies);
+ break;
+ } else
+ err = dccp_wait_for_ccid(sk, skb, &timeo);
+ }
- if (err == 0) {
- struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
- const int len = skb->len;
+ skb_dequeue(&sk->sk_write_queue);
+ if (err == 0) {
+ struct dccp_skb_cb *dcb = DCCP_SKB_CB(skb);
+ const int len = skb->len;
- if (sk->sk_state == DCCP_PARTOPEN) {
- /* See 8.1.5. Handshake Completion */
- inet_csk_schedule_ack(sk);
- inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
+ if (sk->sk_state == DCCP_PARTOPEN) {
+ /* See 8.1.5. Handshake Completion */
+ inet_csk_schedule_ack(sk);
+ inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
inet_csk(sk)->icsk_rto,
DCCP_RTO_MAX);
- dcb->dccpd_type = DCCP_PKT_DATAACK;
- } else if (dccp_ack_pending(sk))
- dcb->dccpd_type = DCCP_PKT_DATAACK;
- else
- dcb->dccpd_type = DCCP_PKT_DATA;
-
- err = dccp_transmit_skb(sk, skb);
- ccid_hc_tx_packet_sent(dp->dccps_hc_tx_ccid, sk, 0, len);
- } else
- kfree_skb(skb);
+ dcb->dccpd_type = DCCP_PKT_DATAACK;
+ } else if (dccp_ack_pending(sk))
+ dcb->dccpd_type = DCCP_PKT_DATAACK;
+ else
+ dcb->dccpd_type = DCCP_PKT_DATA;
+
+ err = dccp_transmit_skb(sk, skb);
+ ccid_hc_tx_packet_sent(dp->dccps_hc_tx_ccid, sk, 0, len);
+ if (err != 0) {
+ BUG();
+ break;
+ }
+ } else
+ kfree(skb);
+ if (err != 0) {
+ BUG();
+ break;
+ }
- return err;
+ }
}
int dccp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
@@ -427,6 +460,9 @@ static inline void dccp_connect_init(str
dccp_set_seqno(&dp->dccps_awl, max48(dp->dccps_awl, dp->dccps_iss));
icsk->icsk_retransmits = 0;
+ init_timer(&dp->dccps_xmit_timer);
+ dp->dccps_xmit_timer.data = (unsigned long)sk;
+ dp->dccps_xmit_timer.function = dccp_write_xmit_timer;
}
int dccp_connect(struct sock *sk)
@@ -561,8 +597,10 @@ void dccp_send_close(struct sock *sk, co
DCCP_PKT_CLOSE : DCCP_PKT_CLOSEREQ;
if (active) {
+ dccp_write_xmit(sk, 1);
dccp_skb_entail(sk, skb);
dccp_transmit_skb(sk, skb_clone(skb, prio));
+ /* FIXME do we need a retransmit timer here? */
} else
dccp_transmit_skb(sk, skb);
}
diff --git a/net/dccp/proto.c b/net/dccp/proto.c
index 5317fd3..6432a60 100644
--- a/net/dccp/proto.c
+++ b/net/dccp/proto.c
@@ -663,17 +663,8 @@ int dccp_sendmsg(struct kiocb *iocb, str
if (rc != 0)
goto out_discard;
- rc = dccp_write_xmit(sk, skb, &timeo);
- /*
- * XXX we don't use sk_write_queue, so just discard the packet.
- * Current plan however is to _use_ sk_write_queue with
- * an algorith similar to tcp_sendmsg, where the main difference
- * is that in DCCP we have to respect packet boundaries, so
- * no coalescing of skbs.
- *
- * This bug was _quickly_ found & fixed by just looking at an OSTRA
- * generated callgraph 8) -acme
- */
+ skb_queue_tail(&sk->sk_write_queue, skb);
+ dccp_write_xmit(sk,0);
out_release:
release_sock(sk);
return rc ? : len;
@@ -847,6 +838,7 @@ static int dccp_close_state(struct sock
void dccp_close(struct sock *sk, long timeout)
{
+ struct dccp_sock *dp = dccp_sk(sk);
struct sk_buff *skb;
int state;
@@ -863,6 +855,8 @@ void dccp_close(struct sock *sk, long ti
goto adjudge_to_death;
}
+ sk_stop_timer(sk, &dp->dccps_xmit_timer);
+
/*
* We need to flush the recv. buffs. We do this only on the
* descriptor close, not protocol-sourced closes, because the
-
: send the line "unsubscribe dccp" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
