2008-11-27 14:22:42

by Andreas Petlund

[permalink] [raw]
Subject: RFC: Latency reducing TCP modifications for thin-stream interactive applications

A wide range of Internet-based services that use reliable transport
protocols display what we call thin-stream properties. This means
that the application sends data with such a low rate that the
retransmission mechanisms of the transport protocol are not fully
effective. In time-dependent scenarios (like online games, control
systems or some sensor networks) where the user experience depends
on the data delivery latency, packet loss can be devastating for
the service quality. Extreme latencies are caused by TCP's
dependency on the arrival of new data from the application to trigger
retransmissions effectively through fast retransmit instead of
waiting for long timeouts. After analyzing a large number of
time-dependent interactive applications, we have seen that they
often produce thin streams (as described above) and also stay with
this traffic pattern throughout its entire lifespan. The
combination of time-dependency and the fact that the streams
provoke high latencies when using TCP is unfortunate.

In order to reduce application-layer latency when packets are lost,
we have implemented modifications to the TCP retransmission
mechanisms in the Linux kernel. We have also implemented a
bundling mechanisms that introduces redundancy in order to
preempt the experience of packet loss. In short, if the kernel
detects a thin stream, we trade a small amount of bandwidth for
latency reduction and apply:

Removal of exponential backoff: To prevent an exponential increase
in retransmission delay for a repeatedly lost packet, we remove
the exponential factor.

FASTER Fast Retransmit: Instead of waiting for 3 duplicate
acknowledgments before sending a fast retransmission, we retransmit
after receiving only one.

Redundant Data Bundling: We copy (bundle) data from the
unacknowledged packets in the send buffer into the next packet if
space is available.

These enhancements are applied only if the stream is detected as
thin.
This is accomplished by defining thresholds for packet size and
packets in flight. Also, we consider the redundancy introduced
by our mechanisms acceptable because the streams are so thin
that normal congestion mechanisms do not come into effect.

We have implemented these changes in the Linux kernel (2.6.23.8),
and have tested the modifications on a wide range of different
thin-stream applications (Skype, BZFlag, SSH, ...) under varying
network conditions. Our results show that applications which use
TCP for interactive time-dependent traffic will experience a
reduction in both maximum and average latency, giving the users
quicker feedback to their interactions.

Availability of this kind of mechanisms will help provide
customizability for interactive network services. The quickly
growing market for Linux gaming may benefit from lowered latency.
As an example, most of the large MMORPG's today use TCP (like World
of Warcraft and Age of Conan) and several multimedia applications
(like Skype) use TCP fallback if UDP is blocked.

The modifications are all TCP standard compliant and transparent
to the receiver. As such, a game server could implement the
modifications and get a one-way latency benefit without touching
any of the clients.

In the following papers, we discuss the benefits and tradeoffs of
the decribed mechanisms:
"The Fun of using TCP for an MMORPG":
http://simula.no/research/networks/publications/Griwodz.2006.1
"TCP Enhancements For Interactive Thin-Stream Applications":
http://simula.no/research/networks/publications/Simula.ND.83
"Improving application layer latency for reliable thin-stream game
traffic": http://simula.no/research/networks/publications/Simula.ND.185
"TCP mechanisms for improving the user experience for time-dependent
thin-stream applications":
http://simula.no/research/networks/publications/Simula.ND.159
Our presentation from the 2008 Linux-Kongress can be found here:
http://data.guug.de/slides/lk2008/lk-2008-Andreas-Petlund.pdf

We have included a patch for the 2.6.23.8 kernel which implements the
modifications. The patch is not properly segmented and formatted, but
attached as a reference. We are currently working on an updated patch
set which we hopefully will be able to post in a couple of weeks. This
will also give us time to integrate any ideas that may arise from the
discussions here.

We are happy for all feedback regarding this:
Is something like this viable to introduce into the kernel? Is the
scheme for thin-stream detection mechanism acceptable. Any viewpoints
on the architecture and design?


diff -Nur linux-2.6.23.8.vanilla/include/linux/sysctl.h linux-2.6.23.8-tcp-thin/include/linux/sysctl.h
--- linux-2.6.23.8.vanilla/include/linux/sysctl.h 2007-11-16 19:14:27.000000000 +0100
+++ linux-2.6.23.8-tcp-thin/include/linux/sysctl.h 2008-07-03 11:47:21.000000000 +0200
@@ -355,6 +355,11 @@
NET_IPV4_ROUTE=18,
NET_IPV4_FIB_HASH=19,
NET_IPV4_NETFILTER=20,
+
+ NET_IPV4_TCP_FORCE_THIN_RDB=29, /* Added @ Simula */
+ NET_IPV4_TCP_FORCE_THIN_RM_EXPB=30, /* Added @ Simula */
+ NET_IPV4_TCP_FORCE_THIN_DUPACK=31, /* Added @ Simula */
+ NET_IPV4_TCP_RDB_MAX_BUNDLE_BYTES=32, /* Added @ Simula */

NET_IPV4_TCP_TIMESTAMPS=33,
NET_IPV4_TCP_WINDOW_SCALING=34,
diff -Nur linux-2.6.23.8.vanilla/include/linux/tcp.h linux-2.6.23.8-tcp-thin/include/linux/tcp.h
--- linux-2.6.23.8.vanilla/include/linux/tcp.h 2007-11-16 19:14:27.000000000 +0100
+++ linux-2.6.23.8-tcp-thin/include/linux/tcp.h 2008-07-02 15:17:38.000000000 +0200
@@ -97,6 +97,10 @@
#define TCP_CONGESTION 13 /* Congestion control algorithm */
#define TCP_MD5SIG 14 /* TCP MD5 Signature (RFC2385) */

+#define TCP_THIN_RDB 15 /* Added @ Simula - Enable redundant data bundling */
+#define TCP_THIN_RM_EXPB 16 /* Added @ Simula - Remove exponential backoff */
+#define TCP_THIN_DUPACK 17 /* Added @ Simula - Reduce number of dupAcks needed */
+
#define TCPI_OPT_TIMESTAMPS 1
#define TCPI_OPT_SACK 2
#define TCPI_OPT_WSCALE 4
@@ -296,6 +300,10 @@
u8 nonagle; /* Disable Nagle algorithm? */
u8 keepalive_probes; /* num of allowed keep alive probes */

+ u8 thin_rdb; /* Enable RDB */
+ u8 thin_rm_expb; /* Remove exp. backoff */
+ u8 thin_dupack; /* Remove dupack */
+
/* RTT measurement */
u32 srtt; /* smoothed round trip time << 3 */
u32 mdev; /* medium deviation */
diff -Nur linux-2.6.23.8.vanilla/include/net/sock.h linux-2.6.23.8-tcp-thin/include/net/sock.h
--- linux-2.6.23.8.vanilla/include/net/sock.h 2007-11-16 19:14:27.000000000 +0100
+++ linux-2.6.23.8-tcp-thin/include/net/sock.h 2008-07-02 17:07:10.000000000 +0200
@@ -462,7 +462,10 @@

static inline void sk_stream_free_skb(struct sock *sk, struct sk_buff *skb)
{
- skb_truesize_check(skb);
+ /* Modified @ Simula
+ skb_truesize_check creates unnecessary
+ noise when combined with RDB */
+ //skb_truesize_check(skb);
sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
sk->sk_wmem_queued -= skb->truesize;
sk->sk_forward_alloc += skb->truesize;
diff -Nur linux-2.6.23.8.vanilla/include/net/tcp.h linux-2.6.23.8-tcp-thin/include/net/tcp.h
--- linux-2.6.23.8.vanilla/include/net/tcp.h 2007-11-16 19:14:27.000000000 +0100
+++ linux-2.6.23.8-tcp-thin/include/net/tcp.h 2008-07-03 11:48:54.000000000 +0200
@@ -188,9 +188,19 @@
#define TCP_NAGLE_CORK 2 /* Socket is corked */
#define TCP_NAGLE_PUSH 4 /* Cork is overridden for already queued data */

+/* Added @ Simula - Thin stream support */
+#define TCP_FORCE_THIN_RDB 0 /* Thin streams: exp. backoff default off */
+#define TCP_FORCE_THIN_RM_EXPB 0 /* Thin streams: dynamic dupack default off */
+#define TCP_FORCE_THIN_DUPACK 0 /* Thin streams: smaller minRTO default off */
+#define TCP_RDB_MAX_BUNDLE_BYTES 0 /* Thin streams: Limit maximum bundled bytes */
+
extern struct inet_timewait_death_row tcp_death_row;

/* sysctl variables for tcp */
+extern int sysctl_tcp_force_thin_rdb; /* Added @ Simula */
+extern int sysctl_tcp_force_thin_rm_expb; /* Added @ Simula */
+extern int sysctl_tcp_force_thin_dupack; /* Added @ Simula */
+extern int sysctl_tcp_rdb_max_bundle_bytes; /* Added @ Simula */
extern int sysctl_tcp_timestamps;
extern int sysctl_tcp_window_scaling;
extern int sysctl_tcp_sack;
@@ -723,6 +733,16 @@
return (tp->packets_out - tp->left_out + tp->retrans_out);
}

+/* Added @ Simula
+ *
+ * To determine whether a stream is thin or not
+ * return 1 if thin, 0 othervice
+ */
+static inline unsigned int tcp_stream_is_thin(const struct tcp_sock *tp)
+{
+ return (tp->packets_out < 4 ? 1 : 0);
+}
+
/* If cwnd > ssthresh, we may raise ssthresh to be half-way to cwnd.
* The exception is rate halving phase, when cwnd is decreasing towards
* ssthresh.
diff -Nur linux-2.6.23.8.vanilla/net/ipv4/sysctl_net_ipv4.c linux-2.6.23.8-tcp-thin/net/ipv4/sysctl_net_ipv4.c
--- linux-2.6.23.8.vanilla/net/ipv4/sysctl_net_ipv4.c 2007-11-16 19:14:27.000000000 +0100
+++ linux-2.6.23.8-tcp-thin/net/ipv4/sysctl_net_ipv4.c 2008-07-03 11:49:59.000000000 +0200
@@ -187,6 +187,38 @@
}

ctl_table ipv4_table[] = {
+ { /* Added @ Simula for thin streams */
+ .ctl_name = NET_IPV4_TCP_FORCE_THIN_RDB,
+ .procname = "tcp_force_thin_rdb",
+ .data = &sysctl_tcp_force_thin_rdb,
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = &proc_dointvec
+ },
+ { /* Added @ Simula for thin streams */
+ .ctl_name = NET_IPV4_TCP_FORCE_THIN_RM_EXPB,
+ .procname = "tcp_force_thin_rm_expb",
+ .data = &sysctl_tcp_force_thin_rm_expb,
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = &proc_dointvec
+ },
+ { /* Added @ Simula for thin streams */
+ .ctl_name = NET_IPV4_TCP_FORCE_THIN_DUPACK,
+ .procname = "tcp_force_thin_dupack",
+ .data = &sysctl_tcp_force_thin_dupack,
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = &proc_dointvec
+ },
+ { /* Added @ Simula for thin streams */
+ .ctl_name = NET_IPV4_TCP_RDB_MAX_BUNDLE_BYTES,
+ .procname = "tcp_rdb_max_bundle_bytes",
+ .data = &sysctl_tcp_rdb_max_bundle_bytes,
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = &proc_dointvec
+ },
{
.ctl_name = NET_IPV4_TCP_TIMESTAMPS,
.procname = "tcp_timestamps",
diff -Nur linux-2.6.23.8.vanilla/net/ipv4/tcp.c linux-2.6.23.8-tcp-thin/net/ipv4/tcp.c
--- linux-2.6.23.8.vanilla/net/ipv4/tcp.c 2007-11-16 19:14:27.000000000 +0100
+++ linux-2.6.23.8-tcp-thin/net/ipv4/tcp.c 2008-07-03 11:51:55.000000000 +0200
@@ -270,6 +270,10 @@

int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;

+/* Added @ Simula */
+int sysctl_tcp_force_thin_rdb __read_mostly = TCP_FORCE_THIN_RDB;
+int sysctl_tcp_rdb_max_bundle_bytes __read_mostly = TCP_RDB_MAX_BUNDLE_BYTES;
+
DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;

atomic_t tcp_orphan_count = ATOMIC_INIT(0);
@@ -658,6 +662,167 @@
return tmp;
}

+/* Added at Simula to support RDB */
+static int tcp_trans_merge_prev(struct sock *sk, struct sk_buff *skb, int mss_now)
+{
+ struct tcp_sock *tp = tcp_sk(sk);
+
+ /* Make sure that this isn't referenced by somebody else */
+
+ if(!skb_cloned(skb)){
+ struct sk_buff *prev_skb = skb->prev;
+ int skb_size = skb->len;
+ int old_headlen = 0;
+ int ua_data = 0;
+ int uad_head = 0;
+ int uad_frags = 0;
+ int ua_nr_frags = 0;
+ int ua_frags_diff = 0;
+
+ /* Since this technique currently does not support SACK, I
+ * return -1 if the previous has been SACK'd. */
+ if(TCP_SKB_CB(prev_skb)->sacked & TCPCB_SACKED_ACKED){
+ return -1;
+ }
+
+ /* Current skb is out of window. */
+ if (after(TCP_SKB_CB(skb)->end_seq, tp->snd_una+tp->snd_wnd)){
+ return -1;
+ }
+
+ /*TODO: Optimize this part with regards to how the
+ variables are initialized */
+
+ /*Calculates the ammount of unacked data that is available*/
+ ua_data = (TCP_SKB_CB(prev_skb)->end_seq - tp->snd_una >
+ prev_skb->len ? prev_skb->len :
+ TCP_SKB_CB(prev_skb)->end_seq - tp->snd_una);
+ ua_frags_diff = ua_data - prev_skb->data_len;
+ uad_frags = (ua_frags_diff > 0 ? prev_skb->data_len : ua_data);
+ uad_head = (ua_frags_diff > 0 ? ua_data - uad_frags : 0);
+
+ if(ua_data <= 0)
+ return -1;
+
+ if(uad_frags > 0){
+ int i = 0;
+ int bytes_frags = 0;
+
+ if(uad_frags == prev_skb->data_len){
+ ua_nr_frags = skb_shinfo(prev_skb)->nr_frags;
+ } else{
+ for(i=skb_shinfo(prev_skb)->nr_frags - 1; i>=0; i--){
+ if(skb_shinfo(prev_skb)->frags[i].size
+ + bytes_frags == uad_frags){
+ ua_nr_frags += 1;
+ break;
+ }
+ ua_nr_frags += 1;
+ bytes_frags += skb_shinfo(prev_skb)->frags[i].size;
+ }
+ }
+ }
+
+ /*
+ * Do the diffrenet checks on size and content, and return if
+ * something will not work.
+ *
+ * TODO: Support copying some bytes
+ *
+ * 1. Larger than MSS.
+ * 2. Enough room for the stuff stored in the linear area
+ * 3. Enoug room for the pages
+ * 4. If both skbs have some data stored in the linear area, and prev_skb
+ * also has some stored in the paged area, they cannot be merged easily.
+ * 5. If prev_skb is linear, then this one has to be it as well.
+ */
+ if ((sysctl_tcp_rdb_max_bundle_bytes == 0 && ((skb_size + ua_data) > mss_now))
+ || (sysctl_tcp_rdb_max_bundle_bytes > 0 && ((skb_size + ua_data) >
+ sysctl_tcp_rdb_max_bundle_bytes))){
+ return -1;
+ }
+
+ /* We need to know tailroom, even if it is nonlinear */
+ if(uad_head > (skb->end - skb->tail)){
+ return -1;
+ }
+
+ if(skb_is_nonlinear(skb) && (uad_frags > 0)){
+ if((ua_nr_frags +
+ skb_shinfo(skb)->nr_frags) > MAX_SKB_FRAGS){
+ return -1;
+ }
+
+ if(skb_headlen(skb) > 0){
+ return -1;
+ }
+ }
+
+ if((uad_frags > 0) && skb_headlen(skb) > 0){
+ return -1;
+ }
+
+ /* To avoid duplicate copies (and copies
+ where parts have been acked) */
+ if(TCP_SKB_CB(skb)->seq <= (TCP_SKB_CB(prev_skb)->end_seq - ua_data)){
+ return -1;
+ }
+
+ /*SYN's are holy*/
+ if(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN || TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN){
+ return -1;
+ }
+
+ /* Copy linear data */
+ if(uad_head > 0){
+
+ /* Add required space to the header. Can't use put due to linearity */
+ old_headlen = skb_headlen(skb);
+ skb->tail += uad_head;
+ skb->len += uad_head;
+
+ if(skb_headlen(skb) > 0){
+ memmove(skb->data + uad_head, skb->data, old_headlen);
+ }
+
+ skb_copy_to_linear_data(skb, prev_skb->data + (skb_headlen(prev_skb) - uad_head), uad_head);
+ }
+
+ /*Copy paged data*/
+ if(uad_frags > 0){
+ int i = 0;
+ /*Must move data backwards in the array.*/
+ if(skb_is_nonlinear(skb)){
+ memmove(skb_shinfo(skb)->frags + ua_nr_frags,
+ skb_shinfo(skb)->frags,
+ skb_shinfo(skb)->nr_frags*sizeof(skb_frag_t));
+ }
+
+ /*Copy info and update pages*/
+ memcpy(skb_shinfo(skb)->frags,
+ skb_shinfo(prev_skb)->frags + (skb_shinfo(prev_skb)->nr_frags - ua_nr_frags),
+ ua_nr_frags*sizeof(skb_frag_t));
+
+ for(i=0; i<ua_nr_frags;i++){
+ get_page(skb_shinfo(skb)->frags[i].page);
+ }
+
+ skb_shinfo(skb)->nr_frags += ua_nr_frags;
+ skb->data_len += uad_frags;
+ skb->len += uad_frags;
+ }
+
+ TCP_SKB_CB(skb)->seq = TCP_SKB_CB(prev_skb)->end_seq - ua_data;
+
+ if(skb->ip_summed == CHECKSUM_PARTIAL)
+ skb->csum = CHECKSUM_PARTIAL;
+ else
+ skb->csum = skb_checksum(skb, 0, skb->len, 0);
+ }
+
+ return 1;
+}
+
int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
size_t size)
{
@@ -825,6 +990,16 @@

from += copy;
copied += copy;
+
+ /* Added at Simula to support RDB */
+ if(((tp->thin_rdb || sysctl_tcp_force_thin_rdb)) && skb->len < mss_now){
+ if(skb->prev != (struct sk_buff*) &(sk)->sk_write_queue
+ && !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)
+ && !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)){
+ tcp_trans_merge_prev(sk, skb, mss_now);
+ }
+ } /* End - Simula */
+
if ((seglen -= copy) == 0 && iovlen == 0)
goto out;

@@ -1870,7 +2045,25 @@
tcp_push_pending_frames(sk);
}
break;
-
+
+ /* Added @ Simula. Support for thin streams */
+ case TCP_THIN_RDB:
+ if(val)
+ tp->thin_rdb = 1;
+ break;
+
+ /* Added @ Simula. Support for thin streams */
+ case TCP_THIN_RM_EXPB:
+ if(val)
+ tp->thin_rm_expb = 1;
+ break;
+
+ /* Added @ Simula. Support for thin streams */
+ case TCP_THIN_DUPACK:
+ if(val)
+ tp->thin_dupack = 1;
+ break;
+
case TCP_KEEPIDLE:
if (val < 1 || val > MAX_TCP_KEEPIDLE)
err = -EINVAL;
diff -Nur linux-2.6.23.8.vanilla/net/ipv4/tcp_input.c linux-2.6.23.8-tcp-thin/net/ipv4/tcp_input.c
--- linux-2.6.23.8.vanilla/net/ipv4/tcp_input.c 2007-11-16 19:14:27.000000000 +0100
+++ linux-2.6.23.8-tcp-thin/net/ipv4/tcp_input.c 2008-07-03 11:57:08.000000000 +0200
@@ -89,6 +89,9 @@
int sysctl_tcp_frto_response __read_mostly;
int sysctl_tcp_nometrics_save __read_mostly;

+/* Added @ Simula */
+int sysctl_tcp_force_thin_dupack __read_mostly = TCP_FORCE_THIN_DUPACK;
+
int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
int sysctl_tcp_abc __read_mostly;

@@ -1709,6 +1712,12 @@
*/
return 1;
}
+
+ /*Added at Simula to modify fast retransmit */
+ if ((tp->thin_dupack || sysctl_tcp_force_thin_dupack) &&
+ tcp_fackets_out(tp) > 1 && tcp_stream_is_thin(tp)){
+ return 1;
+ }

return 0;
}
@@ -2442,30 +2451,127 @@
{
struct tcp_sock *tp = tcp_sk(sk);
const struct inet_connection_sock *icsk = inet_csk(sk);
- struct sk_buff *skb;
+ struct sk_buff *skb = tcp_write_queue_head(sk);
+ struct sk_buff *next_skb;
+
__u32 now = tcp_time_stamp;
int acked = 0;
int prior_packets = tp->packets_out;
+
+ /*Added at Simula for RDB support*/
+ __u8 done = 0;
+ int remove = 0;
+ int remove_head = 0;
+ int remove_frags = 0;
+ int no_frags;
+ int data_frags;
+ int i;
+
__s32 seq_rtt = -1;
ktime_t last_ackt = net_invalid_timestamp();
-
- while ((skb = tcp_write_queue_head(sk)) &&
- skb != tcp_send_head(sk)) {
+
+ while (skb != NULL
+ && ((!(tp->thin_rdb || sysctl_tcp_force_thin_rdb)
+ && skb != tcp_send_head(sk)
+ && skb != (struct sk_buff *)&sk->sk_write_queue)
+ || ((tp->thin_rdb || sysctl_tcp_force_thin_rdb)
+ && skb != (struct sk_buff *)&sk->sk_write_queue))){
struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
__u8 sacked = scb->sacked;
-
+
+ if(skb == NULL){
+ break;
+ }
+
+ if(skb == tcp_send_head(sk)){
+ break;
+ }
+
+ if(skb == (struct sk_buff *)&sk->sk_write_queue){
+ break;
+ }
+
/* If our packet is before the ack sequence we can
* discard it as it's confirmed to have arrived at
* the other end.
*/
if (after(scb->end_seq, tp->snd_una)) {
- if (tcp_skb_pcount(skb) > 1 &&
- after(tp->snd_una, scb->seq))
- acked |= tcp_tso_acked(sk, skb,
- now, &seq_rtt);
- break;
+ if (tcp_skb_pcount(skb) > 1 && after(tp->snd_una, scb->seq))
+ acked |= tcp_tso_acked(sk, skb, now, &seq_rtt);
+
+ done = 1;
+
+ /* Added at Simula for RDB support*/
+ if ((tp->thin_rdb || sysctl_tcp_force_thin_rdb) && after(tp->snd_una, scb->seq)) {
+ if (!skb_cloned(skb) && !(scb->flags & TCPCB_FLAG_SYN)){
+ remove = tp->snd_una - scb->seq;
+ remove_head = (remove > skb_headlen(skb) ?
+ skb_headlen(skb) : remove);
+ remove_frags = (remove > skb_headlen(skb) ?
+ remove - remove_head : 0);
+
+ /* Has linear data */
+ if(skb_headlen(skb) > 0 && remove_head > 0){
+ memmove(skb->data,
+ skb->data + remove_head,
+ skb_headlen(skb) - remove_head);
+
+ skb->tail -= remove_head;
+ }
+
+ if(skb_is_nonlinear(skb) && remove_frags > 0){
+ no_frags = 0;
+ data_frags = 0;
+
+ /*Remove unecessary pages*/
+ for(i=0; i<skb_shinfo(skb)->nr_frags; i++){
+ if(data_frags + skb_shinfo(skb)->frags[i].size
+ == remove_frags){
+ put_page(skb_shinfo(skb)->frags[i].page);
+ no_frags += 1;
+ break;
+ }
+ put_page(skb_shinfo(skb)->frags[i].page);
+ no_frags += 1;
+ data_frags += skb_shinfo(skb)->frags[i].size;
+ }
+
+ if(skb_shinfo(skb)->nr_frags > no_frags)
+ memmove(skb_shinfo(skb)->frags,
+ skb_shinfo(skb)->frags + no_frags,
+ (skb_shinfo(skb)->nr_frags
+ - no_frags)*sizeof(skb_frag_t));
+
+ skb->data_len -= remove_frags;
+ skb_shinfo(skb)->nr_frags -= no_frags;
+
+ }
+
+ scb->seq += remove;
+ skb->len -= remove;
+
+ if(skb->ip_summed == CHECKSUM_PARTIAL)
+ skb->csum = CHECKSUM_PARTIAL;
+ else
+ skb->csum = skb_checksum(skb, 0, skb->len, 0);
+
+ }
+
+ /*Only move forward if data could be removed from this packet*/
+ done = 2;
+
+ }
+
+ if(done == 1 || tcp_skb_is_last(sk,skb)){
+ break;
+ } else if(done == 2){
+ skb = skb->next;
+ done = 1;
+ continue;
+ }
+
}
-
+
/* Initial outgoing SYN's get put onto the write_queue
* just like anything else we transmit. It is not
* true data, and if we misinform our callers that
@@ -2479,14 +2585,14 @@
acked |= FLAG_SYN_ACKED;
tp->retrans_stamp = 0;
}
-
+
/* MTU probing checks */
if (icsk->icsk_mtup.probe_size) {
if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
tcp_mtup_probe_success(sk, skb);
}
}
-
+
if (sacked) {
if (sacked & TCPCB_RETRANS) {
if (sacked & TCPCB_SACKED_RETRANS)
@@ -2510,24 +2616,32 @@
seq_rtt = now - scb->when;
last_ackt = skb->tstamp;
}
+
+ if ((tp->thin_rdb || sysctl_tcp_force_thin_rdb) && skb == tcp_send_head(sk)) {
+ tcp_advance_send_head(sk, skb);
+ }
+
tcp_dec_pcount_approx(&tp->fackets_out, skb);
tcp_packets_out_dec(tp, skb);
+ next_skb = skb->next;
tcp_unlink_write_queue(skb, sk);
sk_stream_free_skb(sk, skb);
clear_all_retrans_hints(tp);
+ /* Added at Simula to support RDB */
+ skb = next_skb;
}
-
+
if (acked&FLAG_ACKED) {
u32 pkts_acked = prior_packets - tp->packets_out;
const struct tcp_congestion_ops *ca_ops
= inet_csk(sk)->icsk_ca_ops;
-
+
tcp_ack_update_rtt(sk, acked, seq_rtt);
tcp_ack_packets_out(sk);
-
+
if (ca_ops->pkts_acked) {
s32 rtt_us = -1;
-
+
/* Is the ACK triggering packet unambiguous? */
if (!(acked & FLAG_RETRANS_DATA_ACKED)) {
/* High resolution needed and available? */
diff -Nur linux-2.6.23.8.vanilla/net/ipv4/tcp_output.c linux-2.6.23.8-tcp-thin/net/ipv4/tcp_output.c
--- linux-2.6.23.8.vanilla/net/ipv4/tcp_output.c 2007-11-16 19:14:27.000000000 +0100
+++ linux-2.6.23.8-tcp-thin/net/ipv4/tcp_output.c 2008-07-03 11:55:45.000000000 +0200
@@ -1653,7 +1653,7 @@

BUG_ON(tcp_skb_pcount(skb) != 1 ||
tcp_skb_pcount(next_skb) != 1);
-
+
/* changing transmit queue under us so clear hints */
clear_all_retrans_hints(tp);

@@ -1702,6 +1702,166 @@
}
}

+/* Added at Simula. Variation of the regular collapse,
+ adapted to support RDB */
+static void tcp_retrans_merge_redundant(struct sock *sk,
+ struct sk_buff *skb,
+ int mss_now)
+{
+ struct tcp_sock *tp = tcp_sk(sk);
+ struct sk_buff *next_skb = skb->next;
+ int skb_size = skb->len;
+ int new_data = 0;
+ int new_data_head = 0;
+ int new_data_frags = 0;
+ int new_frags = 0;
+ int old_headlen = 0;
+
+ int i;
+ int data_frags = 0;
+
+ /* Loop through as many packets as possible
+ * (will create a lot of redundant data, but WHATEVER).
+ * The only packet this MIGHT be critical for is
+ * if this packet is the last in the retrans-queue.
+ *
+ * Make sure that the first skb isnt already in
+ * use by somebody else. */
+
+ if (!skb_cloned(skb)) {
+ /* Iterate through the retransmit queue */
+ for (; (next_skb != (sk)->sk_send_head) &&
+ (next_skb != (struct sk_buff *) &(sk)->sk_write_queue);
+ next_skb = next_skb->next) {
+
+ /* Reset variables */
+ new_frags = 0;
+ data_frags = 0;
+ new_data = TCP_SKB_CB(next_skb)->end_seq - TCP_SKB_CB(skb)->end_seq;
+
+ /* New data will be stored at skb->start_add + some_offset,
+ in other words the last N bytes */
+ new_data_frags = (new_data > next_skb->data_len ?
+ next_skb->data_len : new_data);
+ new_data_head = (new_data > next_skb->data_len ?
+ new_data - skb->data_len : 0);
+
+ /*
+ * 1. Contains the same data
+ * 2. Size
+ * 3. Sack
+ * 4. Window
+ * 5. Cannot merge with a later packet that has linear data
+ * 6. The new number of frags will exceed the limit
+ * 7. Enough tailroom
+ */
+
+ if(new_data <= 0){
+ return;
+ }
+
+ if ((sysctl_tcp_rdb_max_bundle_bytes == 0 && ((skb_size + new_data) > mss_now))
+ || (sysctl_tcp_rdb_max_bundle_bytes > 0 && ((skb_size + new_data) >
+ sysctl_tcp_rdb_max_bundle_bytes))){
+ return;
+ }
+
+ if(TCP_SKB_CB(next_skb)->flags & TCPCB_FLAG_FIN){
+ return;
+ }
+
+ if((TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
+ (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)){
+ return;
+ }
+
+ if(after(TCP_SKB_CB(skb)->end_seq + new_data, tp->snd_una + tp->snd_wnd)){
+ return;
+ }
+
+ if(skb_shinfo(skb)->frag_list || skb_shinfo(skb)->frag_list){
+ return;
+ }
+
+ /* Calculate number of new fragments. Any new data will be
+ stored in the back. */
+ if(skb_is_nonlinear(next_skb)){
+ i = (skb_shinfo(next_skb)->nr_frags == 0 ?
+ 0 : skb_shinfo(next_skb)->nr_frags - 1);
+ for( ; i>=0;i--){
+ if(data_frags + skb_shinfo(next_skb)->frags[i].size ==
+ new_data_frags){
+ new_frags += 1;
+ break;
+ }
+
+ data_frags += skb_shinfo(next_skb)->frags[i].size;
+ new_frags += 1;
+ }
+ }
+
+ /* If dealing with a fragmented skb, only merge
+ with an skb that ONLY contain frags */
+ if(skb_is_nonlinear(skb)){
+
+ /*Due to the way packets are processed, no later data*/
+ if(skb_headlen(next_skb) && new_data_head > 0){
+ return;
+ }
+
+ if(skb_is_nonlinear(next_skb) && (new_data_frags > 0) &&
+ ((skb_shinfo(skb)->nr_frags + new_frags) > MAX_SKB_FRAGS)){
+ return;
+ }
+
+ } else {
+ if(skb_headlen(next_skb) && (new_data_head > (skb->end - skb->tail))){
+ return;
+ }
+ }
+
+ /*Copy linear data. This will only occur if both are linear,
+ or only A is linear*/
+ if(skb_headlen(next_skb) && (new_data_head > 0)){
+ old_headlen = skb_headlen(skb);
+ skb->tail += new_data_head;
+ skb->len += new_data_head;
+
+ /* The new data starts in the linear area,
+ and the correct offset will then be given by
+ removing new_data ammount of bytes from length. */
+ skb_copy_to_linear_data_offset(skb, old_headlen, next_skb->tail -
+ new_data_head, new_data_head);
+ }
+
+ if(skb_is_nonlinear(next_skb) && (new_data_frags > 0)){
+ memcpy(skb_shinfo(skb)->frags + skb_shinfo(skb)->nr_frags,
+ skb_shinfo(next_skb)->frags +
+ (skb_shinfo(next_skb)->nr_frags - new_frags),
+ new_frags*sizeof(skb_frag_t));
+
+ for(i=skb_shinfo(skb)->nr_frags;
+ i < skb_shinfo(skb)->nr_frags + new_frags; i++)
+ get_page(skb_shinfo(skb)->frags[i].page);
+
+ skb_shinfo(skb)->nr_frags += new_frags;
+ skb->data_len += new_data_frags;
+ skb->len += new_data_frags;
+ }
+
+ TCP_SKB_CB(skb)->end_seq += new_data;
+
+ if(skb->ip_summed == CHECKSUM_PARTIAL)
+ skb->csum = CHECKSUM_PARTIAL;
+ else
+ skb->csum = skb_checksum(skb, 0, skb->len, 0);
+
+ skb_size = skb->len;
+ }
+
+ }
+}
+
/* Do a simple retransmit without using the backoff mechanisms in
* tcp_timer. This is used for path mtu discovery.
* The socket is already locked here.
@@ -1756,6 +1916,8 @@
/* This retransmits one SKB. Policy decisions and retransmit queue
* state updates are done by the caller. Returns non-zero if an
* error occurred which prevented the send.
+ * Modified at Simula to support thin stream optimizations
+ * TODO: Update to use new helpers (like tcp_write_queue_next())
*/
int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
{
@@ -1802,10 +1964,21 @@
(skb->len < (cur_mss >> 1)) &&
(tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
(!tcp_skb_is_last(sk, skb)) &&
- (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
- (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
- (sysctl_tcp_retrans_collapse != 0))
+ (skb_shinfo(skb)->nr_frags == 0
+ && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0)
+ && (tcp_skb_pcount(skb) == 1
+ && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1)
+ && (sysctl_tcp_retrans_collapse != 0)
+ && !((tp->thin_rdb || sysctl_tcp_force_thin_rdb))) {
tcp_retrans_try_collapse(sk, skb, cur_mss);
+ } else if ((tp->thin_rdb || sysctl_tcp_force_thin_rdb)) {
+ if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
+ !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
+ (skb->next != tcp_send_head(sk)) &&
+ (skb->next != (struct sk_buff *) &sk->sk_write_queue)) {
+ tcp_retrans_merge_redundant(sk, skb, cur_mss);
+ }
+ }

if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
return -EHOSTUNREACH; /* Routing failure or similar. */
diff -Nur linux-2.6.23.8.vanilla/net/ipv4/tcp_timer.c linux-2.6.23.8-tcp-thin/net/ipv4/tcp_timer.c
--- linux-2.6.23.8.vanilla/net/ipv4/tcp_timer.c 2007-11-16 19:14:27.000000000 +0100
+++ linux-2.6.23.8-tcp-thin/net/ipv4/tcp_timer.c 2008-07-02 15:17:38.000000000 +0200
@@ -32,6 +32,9 @@
int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
int sysctl_tcp_orphan_retries __read_mostly;

+/* Added @ Simula */
+int sysctl_tcp_force_thin_rm_expb __read_mostly = TCP_FORCE_THIN_RM_EXPB;
+
static void tcp_write_timer(unsigned long);
static void tcp_delack_timer(unsigned long);
static void tcp_keepalive_timer (unsigned long data);
@@ -368,13 +371,26 @@
*/
icsk->icsk_backoff++;
icsk->icsk_retransmits++;
-
+
out_reset_timer:
- icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
+ /* Added @ Simula removal of exponential backoff for thin streams */
+ if ((tp->thin_rm_expb || sysctl_tcp_force_thin_rm_expb) && tcp_stream_is_thin(tp)) {
+ /* Since 'icsk_backoff' is used to reset timer, set to 0
+ * Recalculate 'icsk_rto' as this might be increased if stream oscillates
+ * between thin and thick, thus the old value might already be too high
+ * compared to the value set by 'tcp_set_rto' in tcp_input.c which resets
+ * the rto without backoff. */
+ icsk->icsk_backoff = 0;
+ icsk->icsk_rto = min(((tp->srtt >> 3) + tp->rttvar), TCP_RTO_MAX);
+ } else {
+ /* Use normal backoff */
+ icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
+ }
+ /* End Simula*/
inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
if (icsk->icsk_retransmits > sysctl_tcp_retries1)
__sk_dst_reset(sk);
-
+
out:;
}


2008-11-28 12:25:50

by Ilpo Järvinen

[permalink] [raw]
Subject: Re: RFC: Latency reducing TCP modifications for thin-stream interactive applications

linux-net is users list, use netdev instead on matters relating
development...

On Thu, 27 Nov 2008, Andreas Petlund wrote:

> A wide range of Internet-based services that use reliable transport
> protocols display what we call thin-stream properties. This means
> that the application sends data with such a low rate that the
> retransmission mechanisms of the transport protocol are not fully
> effective. In time-dependent scenarios (like online games, control
> systems or some sensor networks) where the user experience depends
> on the data delivery latency, packet loss can be devastating for
> the service quality. Extreme latencies are caused by TCP's
> dependency on the arrival of new data from the application to trigger
> retransmissions effectively through fast retransmit instead of
> waiting for long timeouts. After analyzing a large number of
> time-dependent interactive applications, we have seen that they
> often produce thin streams (as described above) and also stay with
> this traffic pattern throughout its entire lifespan. The
> combination of time-dependency and the fact that the streams
> provoke high latencies when using TCP is unfortunate.
>
> In order to reduce application-layer latency when packets are lost,
> we have implemented modifications to the TCP retransmission
> mechanisms in the Linux kernel. We have also implemented a
> bundling mechanisms that introduces redundancy in order to
> preempt the experience of packet loss. In short, if the kernel
> detects a thin stream, we trade a small amount of bandwidth for
> latency reduction and apply:
>
> Removal of exponential backoff: To prevent an exponential increase
> in retransmission delay for a repeatedly lost packet, we remove
> the exponential factor.
>
> FASTER Fast Retransmit: Instead of waiting for 3 duplicate
> acknowledgments before sending a fast retransmission, we retransmit
> after receiving only one.
>
> Redundant Data Bundling: We copy (bundle) data from the
> unacknowledged packets in the send buffer into the next packet if
> space is available.
>
> These enhancements are applied only if the stream is detected as
> thin.
> This is accomplished by defining thresholds for packet size and
> packets in flight. Also, we consider the redundancy introduced
> by our mechanisms acceptable because the streams are so thin
> that normal congestion mechanisms do not come into effect.
>
> We have implemented these changes in the Linux kernel (2.6.23.8),
> and have tested the modifications on a wide range of different
> thin-stream applications (Skype, BZFlag, SSH, ...) under varying
> network conditions. Our results show that applications which use
> TCP for interactive time-dependent traffic will experience a
> reduction in both maximum and average latency, giving the users
> quicker feedback to their interactions.
>
> Availability of this kind of mechanisms will help provide
> customizability for interactive network services. The quickly
> growing market for Linux gaming may benefit from lowered latency.
> As an example, most of the large MMORPG's today use TCP (like World
> of Warcraft and Age of Conan) and several multimedia applications
> (like Skype) use TCP fallback if UDP is blocked.
>
> The modifications are all TCP standard compliant and transparent
> to the receiver. As such, a game server could implement the
> modifications and get a one-way latency benefit without touching
> any of the clients.
>
> In the following papers, we discuss the benefits and tradeoffs of
> the decribed mechanisms:
> "The Fun of using TCP for an MMORPG":
> http://simula.no/research/networks/publications/Griwodz.2006.1
> "TCP Enhancements For Interactive Thin-Stream Applications":
> http://simula.no/research/networks/publications/Simula.ND.83
> "Improving application layer latency for reliable thin-stream game
> traffic": http://simula.no/research/networks/publications/Simula.ND.185
> "TCP mechanisms for improving the user experience for time-dependent
> thin-stream applications":
> http://simula.no/research/networks/publications/Simula.ND.159
> Our presentation from the 2008 Linux-Kongress can be found here:
> http://data.guug.de/slides/lk2008/lk-2008-Andreas-Petlund.pdf
>
> We have included a patch for the 2.6.23.8 kernel which implements the
> modifications. The patch is not properly segmented and formatted, but
> attached as a reference. We are currently working on an updated patch
> set which we hopefully will be able to post in a couple of weeks.

Please make sure that it's based on net-next tree next time since there
are lots of changes already in the areas you're touching.

> This will also give us time to integrate any ideas that may arise from
> the discussions here.
>
> We are happy for all feedback regarding this:
> Is something like this viable to introduce into the kernel?

No idea in general. But it has to be (nearly) minimal and clean if such
thing is ever to be considered. The one below is definately not even close
minimal in many areas...

> Is the scheme for thin-stream detection mechanism acceptable.

Does reduncancy happen in the initial slow start as well (depends on
write pattern)? Why it is so in case the stream is to become thick
right after initial rtts?

> Any viewpoints on the architecture and design?

I'll go through the patch below (some things are mostly "educational"
since I also propose dropping the current way of use here and there).
Some overall comments below everything.

> diff -Nur linux-2.6.23.8.vanilla/include/linux/sysctl.h linux-2.6.23.8-tcp-thin/include/linux/sysctl.h
> --- linux-2.6.23.8.vanilla/include/linux/sysctl.h 2007-11-16 19:14:27.000000000 +0100
> +++ linux-2.6.23.8-tcp-thin/include/linux/sysctl.h 2008-07-03 11:47:21.000000000 +0200
> @@ -355,6 +355,11 @@
> NET_IPV4_ROUTE=18,
> NET_IPV4_FIB_HASH=19,
> NET_IPV4_NETFILTER=20,
> +
> + NET_IPV4_TCP_FORCE_THIN_RDB=29, /* Added @ Simula */
> + NET_IPV4_TCP_FORCE_THIN_RM_EXPB=30, /* Added @ Simula */
> + NET_IPV4_TCP_FORCE_THIN_DUPACK=31, /* Added @ Simula */
> + NET_IPV4_TCP_RDB_MAX_BUNDLE_BYTES=32, /* Added @ Simula */

I don't think you could reuse empty ranges. In any case all new sysctls
to be added must use CTL_UNNUMBERED. These number are deprecated and
should not be added anymore. And you add too many sysctl anyway
considering the change (e.g., max & rdb are redundant if proper mapping
of values is used).

> NET_IPV4_TCP_TIMESTAMPS=33,
> NET_IPV4_TCP_WINDOW_SCALING=34,
> diff -Nur linux-2.6.23.8.vanilla/include/linux/tcp.h linux-2.6.23.8-tcp-thin/include/linux/tcp.h
> --- linux-2.6.23.8.vanilla/include/linux/tcp.h 2007-11-16 19:14:27.000000000 +0100
> +++ linux-2.6.23.8-tcp-thin/include/linux/tcp.h 2008-07-02 15:17:38.000000000 +0200
> @@ -97,6 +97,10 @@
> #define TCP_CONGESTION 13 /* Congestion control algorithm */
> #define TCP_MD5SIG 14 /* TCP MD5 Signature (RFC2385) */
>
> +#define TCP_THIN_RDB 15 /* Added @ Simula - Enable redundant data bundling */
> +#define TCP_THIN_RM_EXPB 16 /* Added @ Simula - Remove exponential backoff */
> +#define TCP_THIN_DUPACK 17 /* Added @ Simula - Reduce number of dupAcks needed */
> +
> #define TCPI_OPT_TIMESTAMPS 1
> #define TCPI_OPT_SACK 2
> #define TCPI_OPT_WSCALE 4
> @@ -296,6 +300,10 @@
> u8 nonagle; /* Disable Nagle algorithm? */
> u8 keepalive_probes; /* num of allowed keep alive probes */
>
> + u8 thin_rdb; /* Enable RDB */
> + u8 thin_rm_expb; /* Remove exp. backoff */
> + u8 thin_dupack; /* Remove dupack */
> +

Only one bit used per field?

> /* RTT measurement */
> u32 srtt; /* smoothed round trip time << 3 */
> u32 mdev; /* medium deviation */
> diff -Nur linux-2.6.23.8.vanilla/include/net/sock.h linux-2.6.23.8-tcp-thin/include/net/sock.h
> --- linux-2.6.23.8.vanilla/include/net/sock.h 2007-11-16 19:14:27.000000000 +0100
> +++ linux-2.6.23.8-tcp-thin/include/net/sock.h 2008-07-02 17:07:10.000000000 +0200
> @@ -462,7 +462,10 @@
>
> static inline void sk_stream_free_skb(struct sock *sk, struct sk_buff *skb)
> {
> - skb_truesize_check(skb);
> + /* Modified @ Simula
> + skb_truesize_check creates unnecessary
> + noise when combined with RDB */
> + //skb_truesize_check(skb);

Please fix the bugs you added instead of hacking around them. You'll have
to do proper accounting.

> sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
> sk->sk_wmem_queued -= skb->truesize;
> sk->sk_forward_alloc += skb->truesize;
> diff -Nur linux-2.6.23.8.vanilla/include/net/tcp.h linux-2.6.23.8-tcp-thin/include/net/tcp.h
> --- linux-2.6.23.8.vanilla/include/net/tcp.h 2007-11-16 19:14:27.000000000 +0100
> +++ linux-2.6.23.8-tcp-thin/include/net/tcp.h 2008-07-03 11:48:54.000000000 +0200
> @@ -188,9 +188,19 @@
> #define TCP_NAGLE_CORK 2 /* Socket is corked */
> #define TCP_NAGLE_PUSH 4 /* Cork is overridden for already queued data */
>
> +/* Added @ Simula - Thin stream support */
> +#define TCP_FORCE_THIN_RDB 0 /* Thin streams: exp. backoff default off */
> +#define TCP_FORCE_THIN_RM_EXPB 0 /* Thin streams: dynamic dupack default off */
> +#define TCP_FORCE_THIN_DUPACK 0 /* Thin streams: smaller minRTO default off */
> +#define TCP_RDB_MAX_BUNDLE_BYTES 0 /* Thin streams: Limit maximum bundled bytes */

Please drop these and just assign directly.

> +
> extern struct inet_timewait_death_row tcp_death_row;
>
> /* sysctl variables for tcp */
> +extern int sysctl_tcp_force_thin_rdb; /* Added @ Simula */
> +extern int sysctl_tcp_force_thin_rm_expb; /* Added @ Simula */
> +extern int sysctl_tcp_force_thin_dupack; /* Added @ Simula */
> +extern int sysctl_tcp_rdb_max_bundle_bytes; /* Added @ Simula */
> extern int sysctl_tcp_timestamps;
> extern int sysctl_tcp_window_scaling;
> extern int sysctl_tcp_sack;
> @@ -723,6 +733,16 @@
> return (tp->packets_out - tp->left_out + tp->retrans_out);
> }
>
> +/* Added @ Simula

Please remove the from the patch.

> + *
> + * To determine whether a stream is thin or not
> + * return 1 if thin, 0 othervice

I don't find this comment too useful. The function name is obvious enough.

> + */
> +static inline unsigned int tcp_stream_is_thin(const struct tcp_sock *tp)
> +{
> + return (tp->packets_out < 4 ? 1 : 0);

return tp->packets < 4; is enough??

> +}
> +
> /* If cwnd > ssthresh, we may raise ssthresh to be half-way to cwnd.
> * The exception is rate halving phase, when cwnd is decreasing towards
> * ssthresh.
> diff -Nur linux-2.6.23.8.vanilla/net/ipv4/sysctl_net_ipv4.c linux-2.6.23.8-tcp-thin/net/ipv4/sysctl_net_ipv4.c
> --- linux-2.6.23.8.vanilla/net/ipv4/sysctl_net_ipv4.c 2007-11-16 19:14:27.000000000 +0100
> +++ linux-2.6.23.8-tcp-thin/net/ipv4/sysctl_net_ipv4.c 2008-07-03 11:49:59.000000000 +0200
> @@ -187,6 +187,38 @@
> }
>
> ctl_table ipv4_table[] = {
> + { /* Added @ Simula for thin streams */
> + .ctl_name = NET_IPV4_TCP_FORCE_THIN_RDB,

CTL_UNNUMBERED as mentioned above.

> + .procname = "tcp_force_thin_rdb",
> + .data = &sysctl_tcp_force_thin_rdb,
> + .maxlen = sizeof(int),
> + .mode = 0644,
> + .proc_handler = &proc_dointvec
> + },
> + { /* Added @ Simula for thin streams */
> + .ctl_name = NET_IPV4_TCP_FORCE_THIN_RM_EXPB,
> + .procname = "tcp_force_thin_rm_expb",
> + .data = &sysctl_tcp_force_thin_rm_expb,
> + .maxlen = sizeof(int),
> + .mode = 0644,
> + .proc_handler = &proc_dointvec
> + },
> + { /* Added @ Simula for thin streams */
> + .ctl_name = NET_IPV4_TCP_FORCE_THIN_DUPACK,
> + .procname = "tcp_force_thin_dupack",
> + .data = &sysctl_tcp_force_thin_dupack,
> + .maxlen = sizeof(int),
> + .mode = 0644,
> + .proc_handler = &proc_dointvec
> + },
> + { /* Added @ Simula for thin streams */
> + .ctl_name = NET_IPV4_TCP_RDB_MAX_BUNDLE_BYTES,
> + .procname = "tcp_rdb_max_bundle_bytes",
> + .data = &sysctl_tcp_rdb_max_bundle_bytes,
> + .maxlen = sizeof(int),
> + .mode = 0644,
> + .proc_handler = &proc_dointvec
> + },
> {
> .ctl_name = NET_IPV4_TCP_TIMESTAMPS,
> .procname = "tcp_timestamps",
> diff -Nur linux-2.6.23.8.vanilla/net/ipv4/tcp.c linux-2.6.23.8-tcp-thin/net/ipv4/tcp.c
> --- linux-2.6.23.8.vanilla/net/ipv4/tcp.c 2007-11-16 19:14:27.000000000 +0100
> +++ linux-2.6.23.8-tcp-thin/net/ipv4/tcp.c 2008-07-03 11:51:55.000000000 +0200
> @@ -270,6 +270,10 @@
>
> int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
>
> +/* Added @ Simula */
> +int sysctl_tcp_force_thin_rdb __read_mostly = TCP_FORCE_THIN_RDB;
> +int sysctl_tcp_rdb_max_bundle_bytes __read_mostly = TCP_RDB_MAX_BUNDLE_BYTES;

No need for these defines, just assign directly.

> +
> DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
>
> atomic_t tcp_orphan_count = ATOMIC_INIT(0);
> @@ -658,6 +662,167 @@
> return tmp;
> }
>
> +/* Added at Simula to support RDB */
> +static int tcp_trans_merge_prev(struct sock *sk, struct sk_buff *skb, int mss_now)

why int?

...In general this function should be split, having a separate skb
(non-TCP) function(s) and tcp-side would be ideal.

It's misnamed as well, it won't merge but duplicates?

Also, this approach is extremely intrusive and adding non-linear seqno
things into write queue will require _you_ to do _full audit_ over every
single place to verify that seqno leaps backwards won't break anything
(and you'll still probably miss some cases). I wonder if you realize
how easily this kind of change manifests itself as a silent data
corruption on stream level and have taken appropriate actions to validate
that not a single one of scenario leads to data coming as different out as
was sent in (every single byte, it's not enough to declare that
application worked which could well happen with corrupted data too). TCP
is very coreish and such bugs will definately hit people hard.

> +{
> + struct tcp_sock *tp = tcp_sk(sk);
> +
> + /* Make sure that this isn't referenced by somebody else */
> +
> + if(!skb_cloned(skb)){

I'd inverse the condition and return to reduce indentation level of the
rest.

> + struct sk_buff *prev_skb = skb->prev;
> + int skb_size = skb->len;
> + int old_headlen = 0;
> + int ua_data = 0;
> + int uad_head = 0;
> + int uad_frags = 0;
> + int ua_nr_frags = 0;
> + int ua_frags_diff = 0;
> +
> + /* Since this technique currently does not support SACK, I
> + * return -1 if the previous has been SACK'd. */
> + if(TCP_SKB_CB(prev_skb)->sacked & TCPCB_SACKED_ACKED){
> + return -1;
> + }

Drop block braces if the block's content in all fits on a single line
(except in constructs where else would become braces unbalanced).

> +
> + /* Current skb is out of window. */
> + if (after(TCP_SKB_CB(skb)->end_seq, tp->snd_una+tp->snd_wnd)){

...tcp_wnd_end(tp).

> + return -1;
> + }
> +
> + /*TODO: Optimize this part with regards to how the
> + variables are initialized */
> +
> + /*Calculates the ammount of unacked data that is available*/
> + ua_data = (TCP_SKB_CB(prev_skb)->end_seq - tp->snd_una >
> + prev_skb->len ? prev_skb->len :
> + TCP_SKB_CB(prev_skb)->end_seq - tp->snd_una);

min(...) ?. But see the comment below...

> + ua_frags_diff = ua_data - prev_skb->data_len;
> + uad_frags = (ua_frags_diff > 0 ? prev_skb->data_len : ua_data);
> + uad_head = (ua_frags_diff > 0 ? ua_data - uad_frags : 0);
> +
> + if(ua_data <= 0)
> + return -1;

Can this ever happen except with syn/fin?

> +
> + if(uad_frags > 0){
> + int i = 0;
> + int bytes_frags = 0;
> +
> + if(uad_frags == prev_skb->data_len){
> + ua_nr_frags = skb_shinfo(prev_skb)->nr_frags;
> + } else{
> + for(i=skb_shinfo(prev_skb)->nr_frags - 1; i>=0; i--){
> + if(skb_shinfo(prev_skb)->frags[i].size
> + + bytes_frags == uad_frags){
> + ua_nr_frags += 1;
> + break;
> + }
> + ua_nr_frags += 1;
> + bytes_frags += skb_shinfo(prev_skb)->frags[i].size;
> + }
> + }
> + }
> +
> + /*
> + * Do the diffrenet checks on size and content, and return if
> + * something will not work.
> + *
> + * TODO: Support copying some bytes
> + *
> + * 1. Larger than MSS.
> + * 2. Enough room for the stuff stored in the linear area
> + * 3. Enoug room for the pages
> + * 4. If both skbs have some data stored in the linear area, and prev_skb
> + * also has some stored in the paged area, they cannot be merged easily.
> + * 5. If prev_skb is linear, then this one has to be it as well.
> + */
> + if ((sysctl_tcp_rdb_max_bundle_bytes == 0 && ((skb_size + ua_data) > mss_now))
> + || (sysctl_tcp_rdb_max_bundle_bytes > 0 && ((skb_size + ua_data) >
> + sysctl_tcp_rdb_max_bundle_bytes))){

max_size = sysctl_... ? : mss_now;
if (skb_size + ua_data > max_size)

?

> + return -1;
> + }
> +
> + /* We need to know tailroom, even if it is nonlinear */
> + if(uad_head > (skb->end - skb->tail)){

skb_tailroom()

> + return -1;
> + }
> +
> + if(skb_is_nonlinear(skb) && (uad_frags > 0)){
> + if((ua_nr_frags +
> + skb_shinfo(skb)->nr_frags) > MAX_SKB_FRAGS){
> + return -1;
> + }
> +
> + if(skb_headlen(skb) > 0){
> + return -1;
> + }
> + }
> +
> + if((uad_frags > 0) && skb_headlen(skb) > 0){
> + return -1;
> + }
> +
> + /* To avoid duplicate copies (and copies
> + where parts have been acked) */
> + if(TCP_SKB_CB(skb)->seq <= (TCP_SKB_CB(prev_skb)->end_seq - ua_data)){

So ua_data must be prev_skb->len or this fails? Why ua_data needs such
complex setup above?

> + return -1;
> + }
> +
> + /*SYN's are holy*/
> + if(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN || TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN){

flags & (...SYN|...FIN) ??

> + return -1;
> + }
> +
> + /* Copy linear data */
> + if(uad_head > 0){
> +
> + /* Add required space to the header. Can't use put due to linearity */
> + old_headlen = skb_headlen(skb);
> + skb->tail += uad_head;
> + skb->len += uad_head;
> +
> + if(skb_headlen(skb) > 0){
> + memmove(skb->data + uad_head, skb->data, old_headlen);
> + }
> +
> + skb_copy_to_linear_data(skb, prev_skb->data + (skb_headlen(prev_skb) - uad_head), uad_head);
> + }
> +
> + /*Copy paged data*/
> + if(uad_frags > 0){
> + int i = 0;
> + /*Must move data backwards in the array.*/
> + if(skb_is_nonlinear(skb)){
> + memmove(skb_shinfo(skb)->frags + ua_nr_frags,
> + skb_shinfo(skb)->frags,
> + skb_shinfo(skb)->nr_frags*sizeof(skb_frag_t));
> + }
> +
> + /*Copy info and update pages*/
> + memcpy(skb_shinfo(skb)->frags,
> + skb_shinfo(prev_skb)->frags + (skb_shinfo(prev_skb)->nr_frags - ua_nr_frags),
> + ua_nr_frags*sizeof(skb_frag_t));
> +
> + for(i=0; i<ua_nr_frags;i++){
> + get_page(skb_shinfo(skb)->frags[i].page);
> + }
> +
> + skb_shinfo(skb)->nr_frags += ua_nr_frags;
> + skb->data_len += uad_frags;
> + skb->len += uad_frags;
> + }
> +
> + TCP_SKB_CB(skb)->seq = TCP_SKB_CB(prev_skb)->end_seq - ua_data;
> +
> + if(skb->ip_summed == CHECKSUM_PARTIAL)
> + skb->csum = CHECKSUM_PARTIAL;
> + else
> + skb->csum = skb_checksum(skb, 0, skb->len, 0);
> + }
> +
> + return 1;
> +}

This function was far too complex to understand and very hard to validate
as is, and I'm not that sure it will be needed anyway (see notes below).

> +
> int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
> size_t size)
> {
> @@ -825,6 +990,16 @@
>
> from += copy;
> copied += copy;
> +
> + /* Added at Simula to support RDB */
> + if(((tp->thin_rdb || sysctl_tcp_force_thin_rdb)) && skb->len < mss_now){
> + if(skb->prev != (struct sk_buff*) &(sk)->sk_write_queue
> + && !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)
> + && !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)){

Why SYN/FIN are checked by both the caller and in tcp_trans_merge_prev

> + tcp_trans_merge_prev(sk, skb, mss_now);
> + }
> + } /* End - Simula */
> +
> if ((seglen -= copy) == 0 && iovlen == 0)
> goto out;
>
> @@ -1870,7 +2045,25 @@
> tcp_push_pending_frames(sk);
> }
> break;
> -
> +
> + /* Added @ Simula. Support for thin streams */
> + case TCP_THIN_RDB:
> + if(val)
> + tp->thin_rdb = 1;
> + break;
> +
> + /* Added @ Simula. Support for thin streams */
> + case TCP_THIN_RM_EXPB:
> + if(val)
> + tp->thin_rm_expb = 1;
> + break;
> +
> + /* Added @ Simula. Support for thin streams */
> + case TCP_THIN_DUPACK:
> + if(val)
> + tp->thin_dupack = 1;
> + break;
> +

These are probably for some experimentations? Should probably be dropped
altogether.

> case TCP_KEEPIDLE:
> if (val < 1 || val > MAX_TCP_KEEPIDLE)
> err = -EINVAL;
> diff -Nur linux-2.6.23.8.vanilla/net/ipv4/tcp_input.c linux-2.6.23.8-tcp-thin/net/ipv4/tcp_input.c
> --- linux-2.6.23.8.vanilla/net/ipv4/tcp_input.c 2007-11-16 19:14:27.000000000 +0100
> +++ linux-2.6.23.8-tcp-thin/net/ipv4/tcp_input.c 2008-07-03 11:57:08.000000000 +0200
> @@ -89,6 +89,9 @@
> int sysctl_tcp_frto_response __read_mostly;
> int sysctl_tcp_nometrics_save __read_mostly;
>
> +/* Added @ Simula */
> +int sysctl_tcp_force_thin_dupack __read_mostly = TCP_FORCE_THIN_DUPACK;
> +
> int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
> int sysctl_tcp_abc __read_mostly;
>
> @@ -1709,6 +1712,12 @@
> */
> return 1;
> }
> +
> + /*Added at Simula to modify fast retransmit */
> + if ((tp->thin_dupack || sysctl_tcp_force_thin_dupack) &&
> + tcp_fackets_out(tp) > 1 && tcp_stream_is_thin(tp)){

Nowadays tcp_dupack_heurestics(tp)

> + return 1;
> + }
>
> return 0;
> }
> @@ -2442,30 +2451,127 @@
> {
> struct tcp_sock *tp = tcp_sk(sk);
> const struct inet_connection_sock *icsk = inet_csk(sk);
> - struct sk_buff *skb;
> + struct sk_buff *skb = tcp_write_queue_head(sk);
> + struct sk_buff *next_skb;

If you need next_skb, you might want to iterate with _safe variant below
unless there's a good reason to not use it (e.g., in the latest sack code
it wasn't safe enough).

> +
> __u32 now = tcp_time_stamp;
> int acked = 0;
> int prior_packets = tp->packets_out;
> +
> + /*Added at Simula for RDB support*/
> + __u8 done = 0;
> + int remove = 0;
> + int remove_head = 0;
> + int remove_frags = 0;
> + int no_frags;
> + int data_frags;
> + int i;
> +
> __s32 seq_rtt = -1;
> ktime_t last_ackt = net_invalid_timestamp();
> -
> - while ((skb = tcp_write_queue_head(sk)) &&
> - skb != tcp_send_head(sk)) {
> +
> + while (skb != NULL
> + && ((!(tp->thin_rdb || sysctl_tcp_force_thin_rdb)
> + && skb != tcp_send_head(sk)
> + && skb != (struct sk_buff *)&sk->sk_write_queue)
> + || ((tp->thin_rdb || sysctl_tcp_force_thin_rdb)
> + && skb != (struct sk_buff *)&sk->sk_write_queue))){

Too complex while condition for most of the mortals to understand, rather
use while(1) (or preferrably the _safe walk variant) and if (...) break;s

In addition, there has been large rewrite since 2.6.23 days in
tcp_clean_rtx_queue...

> struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
> __u8 sacked = scb->sacked;
> -
> +
> + if(skb == NULL){
> + break;

??? You already checked for that, no? Unless of course that while
condition defeated your too and you needed this to avoid imminent
crashing... :-) The real fix is to simplify while condition.

> + }
> +
> + if(skb == tcp_send_head(sk)){

ditto?

> + break;
> + }
> +
> + if(skb == (struct sk_buff *)&sk->sk_write_queue){

????

> + break;
> + }
> +
> /* If our packet is before the ack sequence we can
> * discard it as it's confirmed to have arrived at
> * the other end.
> */
> if (after(scb->end_seq, tp->snd_una)) {
> - if (tcp_skb_pcount(skb) > 1 &&
> - after(tp->snd_una, scb->seq))
> - acked |= tcp_tso_acked(sk, skb,
> - now, &seq_rtt);
> - break;
> + if (tcp_skb_pcount(skb) > 1 && after(tp->snd_una, scb->seq))
> + acked |= tcp_tso_acked(sk, skb, now, &seq_rtt);

Not changing unrelated formats in the same patch would be nice (you
probably knew that, just mentioning), it makes review a lot harder.

> +
> + done = 1;
> +
> + /* Added at Simula for RDB support*/
> + if ((tp->thin_rdb || sysctl_tcp_force_thin_rdb) && after(tp->snd_una, scb->seq)) {
> + if (!skb_cloned(skb) && !(scb->flags & TCPCB_FLAG_SYN)){

Does this thing of youes depend on skb being not cloned?

> + remove = tp->snd_una - scb->seq;
> + remove_head = (remove > skb_headlen(skb) ?
> + skb_headlen(skb) : remove);
> + remove_frags = (remove > skb_headlen(skb) ?
> + remove - remove_head : 0);
> +
> + /* Has linear data */
> + if(skb_headlen(skb) > 0 && remove_head > 0){
> + memmove(skb->data,
> + skb->data + remove_head,
> + skb_headlen(skb) - remove_head);
> +
> + skb->tail -= remove_head;
> + }
> +
> + if(skb_is_nonlinear(skb) && remove_frags > 0){
> + no_frags = 0;
> + data_frags = 0;
> +
> + /*Remove unecessary pages*/
> + for(i=0; i<skb_shinfo(skb)->nr_frags; i++){
> + if(data_frags + skb_shinfo(skb)->frags[i].size
> + == remove_frags){
> + put_page(skb_shinfo(skb)->frags[i].page);
> + no_frags += 1;
> + break;
> + }
> + put_page(skb_shinfo(skb)->frags[i].page);
> + no_frags += 1;
> + data_frags += skb_shinfo(skb)->frags[i].size;
> + }
> +
> + if(skb_shinfo(skb)->nr_frags > no_frags)
> + memmove(skb_shinfo(skb)->frags,
> + skb_shinfo(skb)->frags + no_frags,
> + (skb_shinfo(skb)->nr_frags
> + - no_frags)*sizeof(skb_frag_t));
> +
> + skb->data_len -= remove_frags;
> + skb_shinfo(skb)->nr_frags -= no_frags;
> +
> + }
> +
> + scb->seq += remove;
> + skb->len -= remove;
> +
> + if(skb->ip_summed == CHECKSUM_PARTIAL)
> + skb->csum = CHECKSUM_PARTIAL;
> + else
> + skb->csum = skb_checksum(skb, 0, skb->len, 0);
> +
> + }

Open coded tcp_trim_head?

> +
> + /*Only move forward if data could be removed from this packet*/
> + done = 2;
> +
> + }
> +
> + if(done == 1 || tcp_skb_is_last(sk,skb)){
> + break;
> + } else if(done == 2){
> + skb = skb->next;
> + done = 1;
> + continue;
> + }
> +
> }
> -
> +

Please make sure that no this kind of space changes will be present in the
patch you're going do send.

> /* Initial outgoing SYN's get put onto the write_queue
> * just like anything else we transmit. It is not
> * true data, and if we misinform our callers that
> @@ -2479,14 +2585,14 @@
> acked |= FLAG_SYN_ACKED;
> tp->retrans_stamp = 0;
> }
> -
> +
> /* MTU probing checks */
> if (icsk->icsk_mtup.probe_size) {
> if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
> tcp_mtup_probe_success(sk, skb);
> }
> }
> -
> +
> if (sacked) {
> if (sacked & TCPCB_RETRANS) {
> if (sacked & TCPCB_SACKED_RETRANS)
> @@ -2510,24 +2616,32 @@
> seq_rtt = now - scb->when;
> last_ackt = skb->tstamp;
> }
> +
> + if ((tp->thin_rdb || sysctl_tcp_force_thin_rdb) && skb == tcp_send_head(sk)) {
> + tcp_advance_send_head(sk, skb);
> + }
> +

I kind of missed the point of this change, can you please justify it if
it's still needed? I think it must either be bug in your code causing this
to happen or unnecessary.

> tcp_dec_pcount_approx(&tp->fackets_out, skb);
> tcp_packets_out_dec(tp, skb);
> + next_skb = skb->next;
> tcp_unlink_write_queue(skb, sk);
> sk_stream_free_skb(sk, skb);
> clear_all_retrans_hints(tp);
> + /* Added at Simula to support RDB */
> + skb = next_skb;

Use _safe wq walking variant if you need this.

> }
> -
> +
> if (acked&FLAG_ACKED) {
> u32 pkts_acked = prior_packets - tp->packets_out;
> const struct tcp_congestion_ops *ca_ops
> = inet_csk(sk)->icsk_ca_ops;
> -
> +
> tcp_ack_update_rtt(sk, acked, seq_rtt);
> tcp_ack_packets_out(sk);
> -
> +
> if (ca_ops->pkts_acked) {
> s32 rtt_us = -1;
> -
> +
> /* Is the ACK triggering packet unambiguous? */
> if (!(acked & FLAG_RETRANS_DATA_ACKED)) {
> /* High resolution needed and available? */
> diff -Nur linux-2.6.23.8.vanilla/net/ipv4/tcp_output.c linux-2.6.23.8-tcp-thin/net/ipv4/tcp_output.c
> --- linux-2.6.23.8.vanilla/net/ipv4/tcp_output.c 2007-11-16 19:14:27.000000000 +0100
> +++ linux-2.6.23.8-tcp-thin/net/ipv4/tcp_output.c 2008-07-03 11:55:45.000000000 +0200
> @@ -1653,7 +1653,7 @@
>
> BUG_ON(tcp_skb_pcount(skb) != 1 ||
> tcp_skb_pcount(next_skb) != 1);
> -
> +
> /* changing transmit queue under us so clear hints */
> clear_all_retrans_hints(tp);
>
> @@ -1702,6 +1702,166 @@
> }
> }
>
> +/* Added at Simula. Variation of the regular collapse,
> + adapted to support RDB */
> +static void tcp_retrans_merge_redundant(struct sock *sk,
> + struct sk_buff *skb,
> + int mss_now)
> +{
> + struct tcp_sock *tp = tcp_sk(sk);
> + struct sk_buff *next_skb = skb->next;
> + int skb_size = skb->len;
> + int new_data = 0;
> + int new_data_head = 0;
> + int new_data_frags = 0;
> + int new_frags = 0;> + int old_headlen = 0;
> +
> + int i;
> + int data_frags = 0;
> +
> + /* Loop through as many packets as possible
> + * (will create a lot of redundant data, but WHATEVER).
> + * The only packet this MIGHT be critical for is
> + * if this packet is the last in the retrans-queue.
> + *
> + * Make sure that the first skb isnt already in
> + * use by somebody else. */
> +
> + if (!skb_cloned(skb)) {

Relying on skb's not being cloned will make your change to work in a
minority of the cases on many hardwares that have tx reclaims happening
late enough. I recently got some numbers about clones after rtt and can
claim this to happen for sure!

> + /* Iterate through the retransmit queue */
> + for (; (next_skb != (sk)->sk_send_head) &&
> + (next_skb != (struct sk_buff *) &(sk)->sk_write_queue);
> + next_skb = next_skb->next) {

_safe(...) walk.

> +
> + /* Reset variables */
> + new_frags = 0;
> + data_frags = 0;
> + new_data = TCP_SKB_CB(next_skb)->end_seq - TCP_SKB_CB(skb)->end_seq;
> +
> + /* New data will be stored at skb->start_add + some_offset,
> + in other words the last N bytes */
> + new_data_frags = (new_data > next_skb->data_len ?
> + next_skb->data_len : new_data);

min()...?

> + new_data_head = (new_data > next_skb->data_len ?
> + new_data - skb->data_len : 0);

max(new_data - skb->data_len, 0);

> +
> + /*
> + * 1. Contains the same data
> + * 2. Size
> + * 3. Sack
> + * 4. Window
> + * 5. Cannot merge with a later packet that has linear data
> + * 6. The new number of frags will exceed the limit
> + * 7. Enough tailroom
> + */
> +
> + if(new_data <= 0){
> + return;
> + }
> +
> + if ((sysctl_tcp_rdb_max_bundle_bytes == 0 && ((skb_size + new_data) > mss_now))
> + || (sysctl_tcp_rdb_max_bundle_bytes > 0 && ((skb_size + new_data) >
> + sysctl_tcp_rdb_max_bundle_bytes))){
> + return;
> + }
> +
> + if(TCP_SKB_CB(next_skb)->flags & TCPCB_FLAG_FIN){
> + return;
> + }
> +
> + if((TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
> + (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)){
> + return;
> + }
> +
> + if(after(TCP_SKB_CB(skb)->end_seq + new_data, tp->snd_una + tp->snd_wnd)){

tcp_wnd_end(tp)

> + return;
> + }
> +
> + if(skb_shinfo(skb)->frag_list || skb_shinfo(skb)->frag_list){

If it wasn't cloned, why you do this check?

> + return;
> + }
> +
> + /* Calculate number of new fragments. Any new data will be
> + stored in the back. */
> + if(skb_is_nonlinear(next_skb)){
> + i = (skb_shinfo(next_skb)->nr_frags == 0 ?
> + 0 : skb_shinfo(next_skb)->nr_frags - 1);
> + for( ; i>=0;i--){
> + if(data_frags + skb_shinfo(next_skb)->frags[i].size ==
> + new_data_frags){
> + new_frags += 1;
> + break;
> + }
> +
> + data_frags += skb_shinfo(next_skb)->frags[i].size;
> + new_frags += 1;
> + }
> + }
> +
> + /* If dealing with a fragmented skb, only merge
> + with an skb that ONLY contain frags */
> + if(skb_is_nonlinear(skb)){
> +
> + /*Due to the way packets are processed, no later data*/
> + if(skb_headlen(next_skb) && new_data_head > 0){
> + return;
> + }
> +
> + if(skb_is_nonlinear(next_skb) && (new_data_frags > 0) &&
> + ((skb_shinfo(skb)->nr_frags + new_frags) > MAX_SKB_FRAGS)){
> + return;
> + }
> +
> + } else {
> + if(skb_headlen(next_skb) && (new_data_head > (skb->end - skb->tail))){
> + return;
> + }
> + }
> +
> + /*Copy linear data. This will only occur if both are linear,
> + or only A is linear*/
> + if(skb_headlen(next_skb) && (new_data_head > 0)){
> + old_headlen = skb_headlen(skb);
> + skb->tail += new_data_head;
> + skb->len += new_data_head;
> +
> + /* The new data starts in the linear area,
> + and the correct offset will then be given by
> + removing new_data ammount of bytes from length. */
> + skb_copy_to_linear_data_offset(skb, old_headlen, next_skb->tail -
> + new_data_head, new_data_head);
> + }
> +
> + if(skb_is_nonlinear(next_skb) && (new_data_frags > 0)){
> + memcpy(skb_shinfo(skb)->frags + skb_shinfo(skb)->nr_frags,
> + skb_shinfo(next_skb)->frags +
> + (skb_shinfo(next_skb)->nr_frags - new_frags),
> + new_frags*sizeof(skb_frag_t));
> +
> + for(i=skb_shinfo(skb)->nr_frags;
> + i < skb_shinfo(skb)->nr_frags + new_frags; i++)
> + get_page(skb_shinfo(skb)->frags[i].page);
> +
> + skb_shinfo(skb)->nr_frags += new_frags;
> + skb->data_len += new_data_frags;
> + skb->len += new_data_frags;
> + }
> +
> + TCP_SKB_CB(skb)->end_seq += new_data;
> +
> + if(skb->ip_summed == CHECKSUM_PARTIAL)
> + skb->csum = CHECKSUM_PARTIAL;
> + else
> + skb->csum = skb_checksum(skb, 0, skb->len, 0);
> +
> + skb_size = skb->len;
> + }
> +
> + }
> +}

I suppose this function was effectively:

{
tcp_write_queue_walk_safe(...) {
...checks...
if (tcp_trim_head(...))
break;
tcp_collapse_retrans(...);
}
}

It's sort of counter-intuitive that first you use redundancy but now in
here when the path _has shown problems_ you now remove the redundancy if
I understand you correctly? The logic is...?

> +
> /* Do a simple retransmit without using the backoff mechanisms in
> * tcp_timer. This is used for path mtu discovery.
> * The socket is already locked here.
> @@ -1756,6 +1916,8 @@
> /* This retransmits one SKB. Policy decisions and retransmit queue
> * state updates are done by the caller. Returns non-zero if an
> * error occurred which prevented the send.
> + * Modified at Simula to support thin stream optimizations
> + * TODO: Update to use new helpers (like tcp_write_queue_next())
> */
> int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
> {
> @@ -1802,10 +1964,21 @@
> (skb->len < (cur_mss >> 1)) &&
> (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
> (!tcp_skb_is_last(sk, skb)) &&
> - (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
> - (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
> - (sysctl_tcp_retrans_collapse != 0))
> + (skb_shinfo(skb)->nr_frags == 0
> + && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0)
> + && (tcp_skb_pcount(skb) == 1
> + && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1)
> + && (sysctl_tcp_retrans_collapse != 0)
> + && !((tp->thin_rdb || sysctl_tcp_force_thin_rdb))) {
> tcp_retrans_try_collapse(sk, skb, cur_mss);
> + } else if ((tp->thin_rdb || sysctl_tcp_force_thin_rdb)) {
> + if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
> + !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
> + (skb->next != tcp_send_head(sk)) &&
> + (skb->next != (struct sk_buff *) &sk->sk_write_queue)) {
> + tcp_retrans_merge_redundant(sk, skb, cur_mss);
> + }
> + }

This function is recoded in net-next. And some rather simple changes are
still expected to that area to handle paged data once I get to that.

>
> if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
> return -EHOSTUNREACH; /* Routing failure or similar. */
> diff -Nur linux-2.6.23.8.vanilla/net/ipv4/tcp_timer.c linux-2.6.23.8-tcp-thin/net/ipv4/tcp_timer.c
> --- linux-2.6.23.8.vanilla/net/ipv4/tcp_timer.c 2007-11-16 19:14:27.000000000 +0100
> +++ linux-2.6.23.8-tcp-thin/net/ipv4/tcp_timer.c 2008-07-02 15:17:38.000000000 +0200
> @@ -32,6 +32,9 @@
> int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
> int sysctl_tcp_orphan_retries __read_mostly;
>
> +/* Added @ Simula */
> +int sysctl_tcp_force_thin_rm_expb __read_mostly = TCP_FORCE_THIN_RM_EXPB;
> +
> static void tcp_write_timer(unsigned long);
> static void tcp_delack_timer(unsigned long);
> static void tcp_keepalive_timer (unsigned long data);
> @@ -368,13 +371,26 @@
> */
> icsk->icsk_backoff++;
> icsk->icsk_retransmits++;
> -
> +
> out_reset_timer:
> - icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
> + /* Added @ Simula removal of exponential backoff for thin streams */
> + if ((tp->thin_rm_expb || sysctl_tcp_force_thin_rm_expb) && tcp_stream_is_thin(tp)) {
> + /* Since 'icsk_backoff' is used to reset timer, set to 0
> + * Recalculate 'icsk_rto' as this might be increased if stream oscillates
> + * between thin and thick, thus the old value might already be too high
> + * compared to the value set by 'tcp_set_rto' in tcp_input.c which resets
> + * the rto without backoff. */
> + icsk->icsk_backoff = 0;
> + icsk->icsk_rto = min(((tp->srtt >> 3) + tp->rttvar), TCP_RTO_MAX);

tcp_set_rto(sk);
tcp_bound_rto(sk);

Though combine + nuking tcp_bound_rto would definately make sense
considering the callsites do both.

> + } else {
> + /* Use normal backoff */
> + icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
> + }
> + /* End Simula*/
> inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
> if (icsk->icsk_retransmits > sysctl_tcp_retries1)
> __sk_dst_reset(sk);
> -
> +
> out:;
> }

It seems very intrusive solution in general. I doubt you succeed in
pulling it off as is without breaking something. To me it seems rather
fragile approach to make write queue seqno backleaps you're proposing. It
also leads to troubles in the truesize as you have noticed. Why not just
building those redundancy containing segments at the write time in case
the stream is thin, then all other parts would not have to bother about
dealing these things? Number of sysctls should be minimized, if they're
to be added at all. Skb work functions should be separated from tcp layer
things.

If you depend on non-changing sysctl value to select right branch, you're
asking for trouble as the userspave is allowed to change it during the
flow as well and even during the ack processing.


--
i.

2008-12-04 12:33:21

by Sven-Thorsten Dietrich

[permalink] [raw]
Subject: Re: RFC: Latency reducing TCP modifications for thin-stream interactive applications

On Thu, 2008-11-27 at 14:39 +0100, Andreas Petlund wrote:

Hi -

Just a minor nit (trying your patch on a 2.6.22 Kernel):

These little white space tweaks are to be avoided, adding to the patch
size unnecessarily.

Do you have a version against 2.6.28, or Linux-tip?

Regards,

Sven

> }
> -
> +
> /* Initial outgoing SYN's get put onto the write_queue
> * just like anything else we transmit. It is not
> * true data, and if we misinform our callers that
> @@ -2479,14 +2585,14 @@
> acked |= FLAG_SYN_ACKED;
> tp->retrans_stamp = 0;
> }
> -
> +
> /* MTU probing checks */
> if (icsk->icsk_mtup.probe_size) {
> if (!after(tp->mtu_probe.probe_seq_end,
> TCP_SKB_CB(skb)->end_seq)) {
> tcp_mtup_probe_success(sk, skb);
> }
> }
> -
> +
> if (sacked) {
> if (sacked & TCPCB_RETRANS) {
> if (sacked & TCPCB_SACKED_RETRANS)
> @@ -2510,24 +2616,32 @@

2009-01-12 15:19:00

by Andreas Petlund

[permalink] [raw]
Subject: Re: RFC: Latency reducing TCP modifications for thin-stream interactive applications

Thank you for comprehensive feedback. We're sorry it took so long to reply.
We will begin the work on a patch set against the net-next tree now and
try to follow your formatting and functional advice.

Ilpo J?rvinen wrote:
> linux-net is users list, use netdev instead on matters relating
> development...

Bad case of typo :)

>
> Please make sure that it's based on net-next tree next time since there
> are lots of changes already in the areas you're touching.

We will do that for the updated patch set.

>> This will also give us time to integrate any ideas that may arise from
>> the discussions here.
>>
>> We are happy for all feedback regarding this:
>> Is something like this viable to introduce into the kernel?
>>
> No idea in general. But it has to be (nearly) minimal and clean if such
> thing is ever to be considered. The one below is definately not even close
> minimal in many areas...

We will clean up the noise, redo the formatting and divide it into logical
segments for the new patch set.

>> Is the scheme for thin-stream detection mechanism acceptable.
>
> Does reduncancy happen in the initial slow start as well (depends on
> write pattern)? Why it is so in case the stream is to become thick
> right after initial rtts?
>

If the window is halved, (but still not at less than 4 segment), the mechanisms
will be turned off (due to the < 4 packets in flight limit). If the stream backs off
to a minimal window, the mechanisms will be turned on.

The bundling mechanism (if activated) will stay active since it relies on
the packet sizes and interarrival time, not on number of packets in flight.

> ...In general this function should be split, having a separate skb
> (non-TCP) function(s) and tcp-side would be ideal.
>
> It's misnamed as well, it won't merge but duplicates?
>
> Also, this approach is extremely intrusive and adding non-linear seqno
> things into write queue will require _you_ to do _full audit_ over every
> single place to verify that seqno leaps backwards won't break anything
> (and you'll still probably miss some cases). I wonder if you realize
> how easily this kind of change manifests itself as a silent data
> corruption on stream level and have taken appropriate actions to validate
> that not a single one of scenario leads to data coming as different out as
> was sent in (every single byte, it's not enough to declare that
> application worked which could well happen with corrupted data too). TCP
> is very coreish and such bugs will definately hit people hard.
>

We have to admit we don't quite see the problem. Since a page can never be removed
before all owners have decleared that they no longer needs it, all data will be correctly
present. Also, since the packets are placed linearly on the queue and we don't support
when SACK is enabled, no gaps will occur. But maybe we have misunderstood your
comment, please let us know if that is the case.


>> + if(TCP_SKB_CB(skb)->seq <= (TCP_SKB_CB(prev_skb)->end_seq - ua_data)){
>
> So ua_data must be prev_skb->len or this fails? Why ua_data needs such
> complex setup above?
>

>
> Does this thing of youes depend on skb being not cloned?
>

We will look into making the calculation ua_data easier and also the cloning requirement. The check
is there to avoid duplicate data in the case when a previous packet has been bundled with later ones.
However, when we think about it, this might not be neccessary since bundling on retransmission is not
allowed to bundle with packets that are yet to be sent. Thank you for pointing this out.

>> +
>> + if ((tp->thin_rdb || sysctl_tcp_force_thin_rdb) && skb == tcp_send_head(sk)) {
>> + tcp_advance_send_head(sk, skb);
>> + }
>> +
>
> I kind of missed the point of this change, can you please justify it if
> it's still needed? I think it must either be bug in your code causing this
> to happen or unnecessary.
>

Thank you very much, this was placed there for a part of the development that we found out
was unecessary.

>> + * Make sure that the first skb isnt already in
>> + * use by somebody else. */
>> +
>> + if (!skb_cloned(skb)) {
>
> Relying on skb's not being cloned will make your change to work in a
> minority of the cases on many hardwares that have tx reclaims happening
> late enough. I recently got some numbers about clones after rtt and can
> claim this to happen for sure!
>

We were not aware of this and will look into the whole cloned-thin in this patch.

>> +
>> + if(skb_shinfo(skb)->frag_list || skb_shinfo(skb)->frag_list){
>
> If it wasn't cloned, why you do this check?
>

Good question, oversight by us, thank you for pointing it out.

> I suppose this function was effectively:
>
> {
> tcp_write_queue_walk_safe(...) {
> ...checks...
> if (tcp_trim_head(...))
> break;
> tcp_collapse_retrans(...);
> }
> }

Correct, except we dont collapse but bundle, will be modified and simplified in the next version of the patch.

> It's sort of counter-intuitive that first you use redundancy but now in
> here when the path _has shown problems_ you now remove the redundancy if
> I understand you correctly? The logic is...?
>

Even though the path has problems (as shown by the need for retransmission), we still bundle as many
packets as possible. If you are still not sure, please let us know what part of the code that is confusing.

> It seems very intrusive solution in general. I doubt you succeed in
> pulling it off as is without breaking something. To me it seems rather
> fragile approach to make write queue seqno backleaps you're proposing. It
> also leads to troubles in the truesize as you have noticed. Why not just
> building those redundancy containing segments at the write time in case
> the stream is thin, then all other parts would not have to bother about
> dealing these things? Number of sysctls should be minimized, if they're
> to be added at all. Skb work functions should be separated from tcp layer
> things.
>
> If you depend on non-changing sysctl value to select right branch, you're
> asking for trouble as the userspave is allowed to change it during the
> flow as well and even during the ack processing.
>

As far as we understand this comment, you want us to do it on the application layer instead? Do you mean as a
middleware, application-specific solution or something similar? Also, we believe doing it on the application layer
will lead to the same delays that we try to prevent, since sent data will be delayed on the transport layer in case
of loss.

-Andreas Petlund & Kristian Evensen

2009-01-14 15:32:56

by Ilpo Järvinen

[permalink] [raw]
Subject: Re: RFC: Latency reducing TCP modifications for thin-stream interactive applications

On Mon, 12 Jan 2009, Andreas Petlund wrote:

> Thank you for comprehensive feedback. We're sorry it took so long to reply.
> We will begin the work on a patch set against the net-next tree now and
> try to follow your formatting and functional advice.

Sadly it will be quite a lot of work (I know, many changes have happened).

> Ilpo J?rvinen wrote:
> > linux-net is users list, use netdev instead on matters relating
> > development...
>
> Bad case of typo :)

Also you could have had a bit less exhaustive cc list, IMHO :-).

> > Please make sure that it's based on net-next tree next time since there
> > are lots of changes already in the areas you're touching.
>
> We will do that for the updated patch set.

Thanks, that will greatly help review when I don't have to apply into some
very oldish kernel version (-p style patch would have helped to show the
context better).

> >> This will also give us time to integrate any ideas that may arise from
> >> the discussions here.
> >>
> >> We are happy for all feedback regarding this:
> >> Is something like this viable to introduce into the kernel?
> >>
> > No idea in general. But it has to be (nearly) minimal and clean if such
> > thing is ever to be considered. The one below is definately not even close
> > minimal in many areas...
>
> We will clean up the noise, redo the formatting and divide it into logical
> segments for the new patch set.

I hope you get more feedback on viability then, my comments were mainly
on technical side.

> >> Is the scheme for thin-stream detection mechanism acceptable.
> >
> > Does reduncancy happen in the initial slow start as well (depends on
> > write pattern)? Why it is so in case the stream is to become thick
> > right after initial rtts?

This is probably partially my misunderstanding as well, I just assumed
tcp_stream_is_thin() is used everywhere without reading too carefully
all those (a && b || (c || (!e && d))) mess.

But in general the problem is still there, ie., if even if some magic
would claim that stream is "thin" we cannot apply that to the streams
which, after the initial slow-start, will no longer match to that magic.
Otherwise stating that we only apply this and this to thin streams won't
be valid.

> If the window is halved, (but still not at less than 4 segment), the
> mechanisms will be turned off (due to the < 4 packets in flight limit).

I'm sorry, this makes not sense to me...

> If the stream backs off to a minimal window, the mechanisms will be
> turned on.
>
> The bundling mechanism (if activated) will stay active since it relies on
> the packet sizes and interarrival time, not on number of packets in flight.

Likewise, maybe you're talking of some other version of the patch since
I cannot find such parts that do what you describe (or the patch is just
too messy :-)).

> > ...In general this function should be split, having a separate skb
> > (non-TCP) function(s) and tcp-side would be ideal.
> >
> > It's misnamed as well, it won't merge but duplicates?
> >
> > Also, this approach is extremely intrusive and adding non-linear seqno
> > things into write queue will require _you_ to do _full audit_ over every
> > single place to verify that seqno leaps backwards won't break anything
> > (and you'll still probably miss some cases). I wonder if you realize
> > how easily this kind of change manifests itself as a silent data
> > corruption on stream level and have taken appropriate actions to validate
> > that not a single one of scenario leads to data coming as different out as
> > was sent in (every single byte, it's not enough to declare that
> > application worked which could well happen with corrupted data too). TCP
> > is very coreish and such bugs will definately hit people hard.
>
> We have to admit we don't quite see the problem. Since a page can never
> be removed before all owners have decleared that they no longer needs
> it, all data will be correctly present. Also, since the packets are
> placed linearly on the queue and we don't support when SACK is enabled,
> no gaps will occur. But maybe we have misunderstood your comment, please
> let us know if that is the case.

Yes, the problems _may_ arise because you create afaict:
skb->end_seq > next_skb->seq situations which is something that certainly
needs an audit over every single seqno operation to see that they don't
implicitly assume otherwise (ie., omit redundant compares)! There are
certainly places I know of already which will break... I'm afraid
that using the approach you've select you'll end up having very many
places needing some extra tweaking, that's why I suggested the alternative
approach to just construct the redundant things on-the-fly while keeping
the write queue as is.

> >> + if(TCP_SKB_CB(skb)->seq <= (TCP_SKB_CB(prev_skb)->end_seq - ua_data)){
> >
> > So ua_data must be prev_skb->len or this fails? Why ua_data needs such
> > complex setup above?
> >
>
> >
> > Does this thing of youes depend on skb being not cloned?
> >
>
> We will look into making the calculation ua_data easier and also the
> cloning requirement.

> The check is there to avoid duplicate data in the case when a previous
> packet has been bundled with later ones.

Ah, this is the already bundled check... Then it would be simpler to do
just: if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(prev_skb)->end_seq)),
wouldn't it?

> However, when we think about it, this might not be neccessary since
> bundling on retransmission is not allowed to bundle with packets that
> are yet to be sent. Thank you for pointing this out.

I guess you must still have some way to protect against duplication on
in tcp_trans_merge_prev() on successive send calls.

> >> +
> >> + if(skb_shinfo(skb)->frag_list || skb_shinfo(skb)->frag_list){
> >
> > If it wasn't cloned, why you do this check?
>
> Good question, oversight by us, thank you for pointing it out.

Funny, it seems to be even buggy for the original purpose (if that ever
was valid) as it is not (x || y) but (x || x) :-D.

> > I suppose this function was effectively:
> >
> > {
> > tcp_write_queue_walk_safe(...) {
> > ...checks...
> > if (tcp_trim_head(...))
> > break;
> > tcp_collapse_retrans(...);
> > }
> > }
>
> Correct, except we dont collapse but bundle, will be modified and
> simplified in the next version of the patch.

Ah, I misunderstood this part, so my next comment is not valid:

> > It's sort of counter-intuitive that first you use redundancy but now in
> > here when the path _has shown problems_ you now remove the redundancy if
> > I understand you correctly? The logic is...?
> >
>
> Even though the path has problems (as shown by the need for
> retransmission), we still bundle as many packets as possible. If you are
> still not sure, please let us know what part of the code that is
> confusing.

...see above.

Though there is not another question, why send them as small at all here
but instead just combining two together and sending twice??? I somewhat
can understand why including from the previous skb is necessary at the
first time sending since the previous is already out but in here we don't
have such obstacle.

> > It seems very intrusive solution in general. I doubt you succeed in
> > pulling it off as is without breaking something. To me it seems rather
> > fragile approach to make write queue seqno backleaps you're proposing. It
> > also leads to troubles in the truesize as you have noticed. Why not just
> > building those redundancy containing segments at the write time in case
> > the stream is thin, then all other parts would not have to bother about
> > dealing these things? Number of sysctls should be minimized, if they're
> > to be added at all. Skb work functions should be separated from tcp layer
> > things.
> >
> > If you depend on non-changing sysctl value to select right branch, you're
> > asking for trouble as the userspave is allowed to change it during the
> > flow as well and even during the ack processing.
> >
>
> As far as we understand this comment, you want us to do it on the
> application layer instead? Do you mean as a middleware,
> application-specific solution or something similar? Also, we believe
> doing it on the application layer will lead to the same delays that we
> try to prevent, since sent data will be delayed on the transport layer
> in case of loss.

No but at the time we're supposed to actually send an skb to the lower
layer and keeping the rest intact.


--
i.

2009-01-16 10:37:59

by kristrev

[permalink] [raw]
Subject: Re: RFC: Latency reducing TCP modifications for thin-stream interactive applications

Hi Ilpo,

>> We have to admit we don't quite see the problem. Since a page can never
>> be removed before all owners have decleared that they no longer needs
>> it, all data will be correctly present. Also, since the packets are
>> placed linearly on the queue and we don't support when SACK is enabled,
>> no gaps will occur. But maybe we have misunderstood your comment, please
>> let us know if that is the case.
>
> Yes, the problems _may_ arise because you create afaict:
> skb->end_seq > next_skb->seq situations which is something that certainly
> needs an audit over every single seqno operation to see that they don't
> implicitly assume otherwise (ie., omit redundant compares)! There are
> certainly places I know of already which will break... I'm afraid
> that using the approach you've select you'll end up having very many
> places needing some extra tweaking, that's why I suggested the alternative
> approach to just construct the redundant things on-the-fly while keeping
> the write queue as is.
>

>> As far as we understand this comment, you want us to do it on the
>> application layer instead? Do you mean as a middleware,
>> application-specific solution or something similar? Also, we believe
>> doing it on the application layer will lead to the same delays that we
>> try to prevent, since sent data will be delayed on the transport layer
>> in case of loss.
>
> No but at the time we're supposed to actually send an skb to the lower
> layer and keeping the rest intact.

I have a small question regarding these two comments. Are you allowed to
modify the packet data stored in a clone? Based on my understanding of the
code and what a clone is, I thought the data was shared between the
original and the clone. Thus, any data appended/inserted into the clone
will also be appended to the original.

If I have understood the code correctly, what will then be the difference
between our current solution and the one you suggest (except we can remove
one of the bundling methods and when a packet is retransmitted)? If I have
not understood the code correctly, feel free to yell :) (if it is a
misunderstanding, it also explains all the checks for skb->cloned).

-Kristian