network delay simulation

network delay simulation

[Stephen Hemminger]

Still bumming around for how to easily simulate long latencies.
There is NISTnet but that is fugly old 2.2 code; and Dummynet, and hitbox 
which are FreeBSD based. The existing traffic shaper might do, but it seems
to be limited to lower speed lines and needs some work (using atomic_set
do own locking for instance).

Probably "the cowboy way" would be to right a new net/scheduler to just
do FIFO delay.

Any ideas/comments? Somebody have some code sitting in a drawer?


-- 
Stephen Hemminger 		mailto:shemminger@osdl.org
Open Source Development Lab	http://developer.osdl.org/shemminger

Re: network delay simulation

[Andi Kleen]

On Tue, 16 Mar 2004 15:10:58 -0800
Stephen Hemminger <shemminger@osdl.org> wrote:

> Still bumming around for how to easily simulate long latencies.
> There is NISTnet but that is fugly old 2.2 code; and Dummynet, and hitbox 
> which are FreeBSD based. The existing traffic shaper might do, but it seems
> to be limited to lower speed lines and needs some work (using atomic_set
> do own locking for instance).
> 
> Probably "the cowboy way" would be to right a new net/scheduler to just
> do FIFO delay.
> 
> Any ideas/comments? Somebody have some code sitting in a drawer?

iirc there was a 2.4 port of NistNet. Porting it to 2.6 is probably
not that much work.

-Andi

Re: network delay simulation

[David S. Miller]

On Wed, 17 Mar 2004 00:00:48 +0100
Andi Kleen <ak@suse.de> wrote:

> iirc there was a 2.4 port of NistNet. Porting it to 2.6 is probably
> not that much work.

nistnet works, but even akpm remembers it as pretty grotty, who knows
maybe the most recent 2.4.x variant was much better, but I doubt it.

Re: network delay simulation

[Andi Kleen]

On Tue, 16 Mar 2004 17:50:52 -0800
"David S. Miller" <davem@redhat.com> wrote:

> On Wed, 17 Mar 2004 00:00:48 +0100
> Andi Kleen <ak@suse.de> wrote:
> 
> > iirc there was a 2.4 port of NistNet. Porting it to 2.6 is probably
> > not that much work.
> 
> nistnet works, but even akpm remembers it as pretty grotty, who knows
> maybe the most recent 2.4.x variant was much better, but I doubt it.

For testing it's probably good enough. No need to merge it ;-)

-Andi

Re: network delay simulation

[Andrew Morton]

"David S. Miller" <davem@redhat.com> wrote:
>
> On Wed, 17 Mar 2004 00:00:48 +0100
> Andi Kleen <ak@suse.de> wrote:
> 
> > iirc there was a 2.4 port of NistNet. Porting it to 2.6 is probably
> > not that much work.
> 
> nistnet works, but even akpm remembers it as pretty grotty, who knows
> maybe the most recent 2.4.x variant was much better, but I doubt it.

whaddya mean "even"?

I wrote a userspace thingy many moons ago which provides variable delays
and bandwidth restriction.  It's for simulating long, thin pipes.  It uses
the tap/tun interface to route packets up to userspace where the
delay/throttling is implemented.  It requires weird policy routing rules,
but there are scripts there which set everything up.

http://www.zipworld.com.au/~akpm/linux/patches/stuff/packet-delay.tar.gz

Re: network delay simulation

[David S. Miller]

On Tue, 16 Mar 2004 15:10:58 -0800
Stephen Hemminger <shemminger@osdl.org> wrote:

> Probably "the cowboy way" would be to right a new net/scheduler to just
> do FIFO delay.
> 
> Any ideas/comments? Somebody have some code sitting in a drawer?

Probably a packet scheduler would be my first choice, you could then even
tag packets using classification and therefore delay differently for different
flows.

Another thing I've always wanted to do was add a "struct netdev *" argument
to dev_alloc_skb(), then we could do things like do SKB allocation failure
tests on drivers to stress them in low memory situations and weird patterns
of alloc failures.  Probably do this in a transitional way via a new routine
with a new name that drivers gradually move over to using.

The idea is that we provide an interface by which to upload some bit string
of alloc success/failure binary states, and at each alloc the state machine
just runs through the bits to decide whether to just quickly return NULL or
not.

Re: network delay simulation

[John Heffner]

On Tue, 16 Mar 2004, David S. Miller wrote:

> On Tue, 16 Mar 2004 15:10:58 -0800
> Stephen Hemminger <shemminger@osdl.org> wrote:
>
> > Probably "the cowboy way" would be to right a new net/scheduler to just
> > do FIFO delay.
> >
> > Any ideas/comments? Somebody have some code sitting in a drawer?
>
> Probably a packet scheduler would be my first choice, you could then even
> tag packets using classification and therefore delay differently for different
> flows.

One insight I had a while ago is that packet pacing and artificial delay
require a very similar mechanism.  I hacked something a while ago that did
this in a general way, though it had some problems and I never really
followed up on it.  Given the recent discussion of pacing on linux-net,
this might be something to think about.

  -John

Re: network delay simulation

[jamal]

On Tue, 2004-03-16 at 20:41, David S. Miller wrote:
 
> > Any ideas/comments? Somebody have some code sitting in a drawer?

I had something trying to replace nistnet a while back - probably
have to go 10 harddrives back to find it.

> Probably a packet scheduler would be my first choice, you could then even
> tag packets using classification and therefore delay differently for different
> flows.

This is exactly what i had. You need to be able to choose what flow gets
affected and which ones are left alone.
A better approach is to use the action code patches i have because then
you can cascade actions that you want and keep adding new error
injection actions
example delay for random time between 30 seconds to 1 minute, then fuck
the checksum, then simulate congestion ... with each of those actions
selectable on how it gets added to the datapath.

> Another thing I've always wanted to do was add a "struct netdev *" argument
> to dev_alloc_skb(), then we could do things like do SKB allocation failure
> tests on drivers to stress them in low memory situations and weird patterns
> of alloc failures.  Probably do this in a transitional way via a new routine
> with a new name that drivers gradually move over to using.

this is valuable. I can bet you over 70% of the drivers will puke like
theres no tommorow.
Not just that, the net stack code will too.

> The idea is that we provide an interface by which to upload some bit string
> of alloc success/failure binary states, and at each alloc the state machine
> just runs through the bits to decide whether to just quickly return NULL or
> not.

A simple algorithm is random().

cheers,
jamal

Re: network delay simulation

[David S. Miller]

On 17 Mar 2004 11:09:26 -0500
jamal <hadi@cyberus.ca> wrote:

> A simple algorithm is random().

Yes, but you want to weight it just a bit.  Userland could use random()
to generate the bit-stream in fact :)

Re: network delay simulation

[Stephen Hemminger]

On Wed, 17 Mar 2004 09:51:07 -0800
"David S. Miller" <davem@redhat.com> wrote:

> On 17 Mar 2004 11:09:26 -0500
> jamal <hadi@cyberus.ca> wrote:
> 
> > A simple algorithm is random().
> 
> Yes, but you want to weight it just a bit.  Userland could use random()
> to generate the bit-stream in fact :)


In an earlier project we had a test suite that failed the Nth allocation,
then repeated startup failing from 1 till all allocations in a typical
interaction round trip had been tested.  Sure led to some nasty failures
in the test bed, which was way better than the customer finding it.

-- 
Stephen Hemminger 		mailto:shemminger@osdl.org
Open Source Development Lab	http://developer.osdl.org/shemminger

Re: network delay simulation

[David S. Miller]

On Wed, 17 Mar 2004 10:00:01 -0800
Stephen Hemminger <shemminger@osdl.org> wrote:

> In an earlier project we had a test suite that failed the Nth allocation,
> then repeated startup failing from 1 till all allocations in a typical
> interaction round trip had been tested.  Sure led to some nasty failures
> in the test bed, which was way better than the customer finding it.

Yes, but we don't really want alloc_skb() doing this, this makes the
failures go to other places (like TCP, netlink, whatever) instead of
what you're trying to actually test (RX out of buffer handling in
a specific driver).

It is an interesting thing to test too, no doubt.

[LARTC] [PATCH] packet delay scheduler

[Stephen Hemminger]

Okay, here is a very simple QOS scheduler that delays packets for 2.6.
It is good for testing, and might be useful for people who want to put
some class of traffic into a "penalty box".

Alexey's tc command is really easy to extend to new disciplines.  There is
a version which knows about this scheduler at:
	http://developer.osdl.org/shemminger/tcp/iproute2-delay.tar.bz2
The only changes were to tc/Makefile and new file tc/q_delay.c

A trivial example is:
	 tc qdisc add dev eth0 root delay latency 25ms rate 100mbit

Try it, if others find it useful please consider adding it.

diff -Nru a/include/linux/pkt_sched.h b/include/linux/pkt_sched.h
--- a/include/linux/pkt_sched.h	Thu Mar 18 12:04:02 2004
+++ b/include/linux/pkt_sched.h	Thu Mar 18 12:04:02 2004
@@ -432,4 +432,10 @@
 
 #define TCA_ATM_MAX	TCA_ATM_STATE
 
+/* Delay section */
+struct tc_dly_qopt
+{
+	__u32	latency;
+	__u32   limit;
+};	
 #endif
diff -Nru a/net/sched/Kconfig b/net/sched/Kconfig
--- a/net/sched/Kconfig	Thu Mar 18 12:04:02 2004
+++ b/net/sched/Kconfig	Thu Mar 18 12:04:02 2004
@@ -164,6 +164,17 @@
 	  To compile this code as a module, choose M here: the
 	  module will be called sch_dsmark.
 
+config NET_SCH_DELAY
+	tristate "Delay simulator"
+	depends on NET_SCHED
+	help
+	  Say Y if you want to delay packets by a fixed amount of
+	  time. This is often useful to simulate network delay when
+	  testing applications or protocols.
+
+	  To compile this driver as a module, choose M here: the module
+	  will be called sch_delay.
+
 config NET_SCH_INGRESS
 	tristate "Ingress Qdisc"
 	depends on NET_SCHED && NETFILTER
diff -Nru a/net/sched/Makefile b/net/sched/Makefile
--- a/net/sched/Makefile	Thu Mar 18 12:04:02 2004
+++ b/net/sched/Makefile	Thu Mar 18 12:04:02 2004
@@ -22,6 +22,7 @@
 obj-$(CONFIG_NET_SCH_TEQL)	+= sch_teql.o
 obj-$(CONFIG_NET_SCH_PRIO)	+= sch_prio.o
 obj-$(CONFIG_NET_SCH_ATM)	+= sch_atm.o
+obj-$(CONFIG_NET_SCH_DELAY)	+= sch_delay.o
 obj-$(CONFIG_NET_CLS_U32)	+= cls_u32.o
 obj-$(CONFIG_NET_CLS_ROUTE4)	+= cls_route.o
 obj-$(CONFIG_NET_CLS_FW)	+= cls_fw.o
diff -Nru a/net/sched/sch_delay.c b/net/sched/sch_delay.c
--- /dev/null	Wed Dec 31 16:00:00 1969
+++ b/net/sched/sch_delay.c	Thu Mar 18 12:04:02 2004
@@ -0,0 +1,269 @@
+/*
+ * net/sched/sch_delay.c	Simple constant delay
+ *
+ * 		This program is free software; you can redistribute it and/or
+ * 		modify it under the terms of the GNU General Public License
+ * 		as published by the Free Software Foundation; either version
+ * 		2 of the License, or (at your option) any later version.
+ *
+ * Authors:	Stephen Hemminger <shemminger@osdl.org>
+ */
+
+#include <linux/config.h>
+#include <linux/module.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+
+#include <linux/string.h>
+#include <linux/mm.h>
+#include <linux/socket.h>
+#include <linux/sockios.h>
+#include <linux/in.h>
+#include <linux/errno.h>
+#include <linux/interrupt.h>
+#include <linux/if_ether.h>
+#include <linux/inet.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/notifier.h>
+#include <net/ip.h>
+#include <net/route.h>
+#include <linux/skbuff.h>
+#include <net/sock.h>
+#include <net/pkt_sched.h>
+
+/*	Network delay simulator
+	This scheduler adds a fixed delay to all packets.
+	Similar to NISTnet and BSD Dummynet.
+
+	It uses byte fifo underneath similar to TBF */
+struct dly_sched_data {
+	u32	latency;
+	u32	limit;
+	struct timer_list timer;
+	struct Qdisc *qdisc;
+};
+
+/* Time stamp put into socket buffer control block */
+struct dly_skb_cb {
+	psched_time_t	queuetime;
+};
+
+/* Enqueue packets with underlying discipline (fifo)
+ * but mark them with current time first.
+ */
+static int dly_enqueue(struct sk_buff *skb, struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct dly_skb_cb *cb = (struct dly_skb_cb *)skb->cb;
+	int ret;
+
+	PSCHED_GET_TIME(cb->queuetime);
+
+	/* Queue to underlying scheduler */
+	ret = q->qdisc->enqueue(skb, q->qdisc);
+	if (ret)
+		sch->stats.drops++;
+	else {
+		sch->q.qlen++;
+		sch->stats.bytes += skb->len;
+		sch->stats.packets++;
+	}
+	return 0;
+}
+
+/* Requeue packets but don't change time stamp */
+static int dly_requeue(struct sk_buff *skb, struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	int ret;
+
+	ret = q->qdisc->ops->requeue(skb, q->qdisc);
+	if (ret = 0)
+		sch->q.qlen++;
+	return ret;
+}
+
+static unsigned int dly_drop(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	unsigned int len;
+
+	len = q->qdisc->ops->drop(q->qdisc);
+	if (len) {
+		sch->q.qlen--;
+		sch->stats.drops++;
+	}
+	return len;
+}
+
+/* Dequeue packet.
+ * If packet needs to be held up, then stop the
+ * queue and set timer to wakeup later.
+ */
+static struct sk_buff *dly_dequeue(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct sk_buff *skb = q->qdisc->dequeue(q->qdisc);
+
+	if (skb) {
+		struct dly_skb_cb *cb = (struct dly_skb_cb *)skb->cb;
+		psched_time_t now;
+		long diff;
+
+		PSCHED_GET_TIME(now);
+		diff = q->latency - PSCHED_TDIFF(now, cb->queuetime);
+
+		if (diff <= 0) {
+			sch->q.qlen--;
+			sch->flags &= ~TCQ_F_THROTTLED;
+			return skb;
+		}
+
+		if (!netif_queue_stopped(sch->dev)) {
+			long delay = PSCHED_US2JIFFIE(diff);
+			if (delay <= 0)
+				delay = 1;
+			mod_timer(&q->timer, jiffies+delay);
+		}
+
+		if (q->qdisc->ops->requeue(skb, q->qdisc) != NET_XMIT_SUCCESS) {
+			sch->q.qlen--;
+			sch->stats.drops++;
+		}
+		sch->flags |= TCQ_F_THROTTLED;
+	}
+	return NULL;
+}
+
+static void dly_reset(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+
+	qdisc_reset(q->qdisc);
+	sch->q.qlen = 0;
+	sch->flags &= ~TCQ_F_THROTTLED;
+	del_timer(&q->timer);
+}
+
+static void dly_timer(unsigned long arg)
+{
+	struct Qdisc *sch = (struct Qdisc *)arg;
+
+	sch->flags &= ~TCQ_F_THROTTLED;
+	netif_schedule(sch->dev);
+}
+
+/* Tell Fifo the new limit. */
+static int change_limit(struct Qdisc *q, u32 limit)
+{
+	struct rtattr *rta;
+	int ret;
+
+	rta = kmalloc(RTA_LENGTH(sizeof(struct tc_fifo_qopt)), GFP_KERNEL);
+	if (!rta)
+		return -ENOMEM;
+
+	rta->rta_type = RTM_NEWQDISC;
+	((struct tc_fifo_qopt *)RTA_DATA(rta))->limit = limit;
+	ret = q->ops->change(q, rta);
+	kfree(rta);
+
+	return ret;
+}
+
+/* Setup underlying FIFO discipline */
+static int dly_change(struct Qdisc *sch, struct rtattr *opt)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct tc_dly_qopt *qopt = RTA_DATA(opt);
+	int err;
+
+	if (q->qdisc = &noop_qdisc) {
+		struct Qdisc *child
+			= qdisc_create_dflt(sch->dev, &bfifo_qdisc_ops);
+		if (!child)
+			return -EINVAL;
+		q->qdisc = child;
+	}
+
+	err = change_limit(q->qdisc, qopt->limit);
+	if (err) {
+		qdisc_destroy(q->qdisc);
+		q->qdisc = &noop_qdisc;
+	} else {
+		q->latency = qopt->latency;
+		q->limit = qopt->limit;
+	}
+	return err;
+}
+
+static int dly_init(struct Qdisc *sch, struct rtattr *opt)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+
+	if (!opt)
+		return -EINVAL;
+
+	init_timer(&q->timer);
+	q->timer.function = dly_timer;
+	q->timer.data = (unsigned long) sch;
+	q->qdisc = &noop_qdisc;
+
+	return dly_change(sch, opt);
+}
+
+static void dly_destroy(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+
+	del_timer(&q->timer);
+	qdisc_destroy(q->qdisc);
+	q->qdisc = &noop_qdisc;
+}
+
+static int dly_dump(struct Qdisc *sch, struct sk_buff *skb)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	unsigned char	 *b = skb->tail;
+	struct tc_dly_qopt qopt;
+
+	qopt.latency = q->latency;
+	qopt.limit = q->limit;
+
+	RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
+
+	return skb->len;
+
+rtattr_failure:
+	skb_trim(skb, b - skb->data);
+	return -1;
+}
+
+static struct Qdisc_ops dly_qdisc_ops = {
+	.id		=	"delay",
+	.priv_size	=	sizeof(struct dly_sched_data),
+	.enqueue	=	dly_enqueue,
+	.dequeue	=	dly_dequeue,
+	.requeue	=	dly_requeue,
+	.drop		=	dly_drop,
+	.init		=	dly_init,
+	.reset		=	dly_reset,
+	.destroy	=	dly_destroy,
+	.change		=	dly_change,
+	.dump		=	dly_dump,
+	.owner		=	THIS_MODULE,
+};
+
+
+static int __init dly_module_init(void)
+{
+	return register_qdisc(&dly_qdisc_ops);
+}
+static void __exit dly_module_exit(void)
+{
+	unregister_qdisc(&dly_qdisc_ops);
+}
+module_init(dly_module_init)
+module_exit(dly_module_exit)
+MODULE_LICENSE("GPL");
_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

[PATCH] packet delay scheduler

[Stephen Hemminger]

Okay, here is a very simple QOS scheduler that delays packets for 2.6.
It is good for testing, and might be useful for people who want to put
some class of traffic into a "penalty box".

Alexey's tc command is really easy to extend to new disciplines.  There is
a version which knows about this scheduler at:
	http://developer.osdl.org/shemminger/tcp/iproute2-delay.tar.bz2
The only changes were to tc/Makefile and new file tc/q_delay.c

A trivial example is:
	 tc qdisc add dev eth0 root delay latency 25ms rate 100mbit

Try it, if others find it useful please consider adding it.

diff -Nru a/include/linux/pkt_sched.h b/include/linux/pkt_sched.h
--- a/include/linux/pkt_sched.h	Thu Mar 18 12:04:02 2004
+++ b/include/linux/pkt_sched.h	Thu Mar 18 12:04:02 2004
@@ -432,4 +432,10 @@
 
 #define TCA_ATM_MAX	TCA_ATM_STATE
 
+/* Delay section */
+struct tc_dly_qopt
+{
+	__u32	latency;
+	__u32   limit;
+};	
 #endif
diff -Nru a/net/sched/Kconfig b/net/sched/Kconfig
--- a/net/sched/Kconfig	Thu Mar 18 12:04:02 2004
+++ b/net/sched/Kconfig	Thu Mar 18 12:04:02 2004
@@ -164,6 +164,17 @@
 	  To compile this code as a module, choose M here: the
 	  module will be called sch_dsmark.
 
+config NET_SCH_DELAY
+	tristate "Delay simulator"
+	depends on NET_SCHED
+	help
+	  Say Y if you want to delay packets by a fixed amount of
+	  time. This is often useful to simulate network delay when
+	  testing applications or protocols.
+
+	  To compile this driver as a module, choose M here: the module
+	  will be called sch_delay.
+
 config NET_SCH_INGRESS
 	tristate "Ingress Qdisc"
 	depends on NET_SCHED && NETFILTER
diff -Nru a/net/sched/Makefile b/net/sched/Makefile
--- a/net/sched/Makefile	Thu Mar 18 12:04:02 2004
+++ b/net/sched/Makefile	Thu Mar 18 12:04:02 2004
@@ -22,6 +22,7 @@
 obj-$(CONFIG_NET_SCH_TEQL)	+= sch_teql.o
 obj-$(CONFIG_NET_SCH_PRIO)	+= sch_prio.o
 obj-$(CONFIG_NET_SCH_ATM)	+= sch_atm.o
+obj-$(CONFIG_NET_SCH_DELAY)	+= sch_delay.o
 obj-$(CONFIG_NET_CLS_U32)	+= cls_u32.o
 obj-$(CONFIG_NET_CLS_ROUTE4)	+= cls_route.o
 obj-$(CONFIG_NET_CLS_FW)	+= cls_fw.o
diff -Nru a/net/sched/sch_delay.c b/net/sched/sch_delay.c
--- /dev/null	Wed Dec 31 16:00:00 1969
+++ b/net/sched/sch_delay.c	Thu Mar 18 12:04:02 2004
@@ -0,0 +1,269 @@
+/*
+ * net/sched/sch_delay.c	Simple constant delay
+ *
+ * 		This program is free software; you can redistribute it and/or
+ * 		modify it under the terms of the GNU General Public License
+ * 		as published by the Free Software Foundation; either version
+ * 		2 of the License, or (at your option) any later version.
+ *
+ * Authors:	Stephen Hemminger <shemminger@osdl.org>
+ */
+
+#include <linux/config.h>
+#include <linux/module.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+
+#include <linux/string.h>
+#include <linux/mm.h>
+#include <linux/socket.h>
+#include <linux/sockios.h>
+#include <linux/in.h>
+#include <linux/errno.h>
+#include <linux/interrupt.h>
+#include <linux/if_ether.h>
+#include <linux/inet.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/notifier.h>
+#include <net/ip.h>
+#include <net/route.h>
+#include <linux/skbuff.h>
+#include <net/sock.h>
+#include <net/pkt_sched.h>
+
+/*	Network delay simulator
+	This scheduler adds a fixed delay to all packets.
+	Similar to NISTnet and BSD Dummynet.
+
+	It uses byte fifo underneath similar to TBF */
+struct dly_sched_data {
+	u32	latency;
+	u32	limit;
+	struct timer_list timer;
+	struct Qdisc *qdisc;
+};
+
+/* Time stamp put into socket buffer control block */
+struct dly_skb_cb {
+	psched_time_t	queuetime;
+};
+
+/* Enqueue packets with underlying discipline (fifo)
+ * but mark them with current time first.
+ */
+static int dly_enqueue(struct sk_buff *skb, struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct dly_skb_cb *cb = (struct dly_skb_cb *)skb->cb;
+	int ret;
+
+	PSCHED_GET_TIME(cb->queuetime);
+
+	/* Queue to underlying scheduler */
+	ret = q->qdisc->enqueue(skb, q->qdisc);
+	if (ret)
+		sch->stats.drops++;
+	else {
+		sch->q.qlen++;
+		sch->stats.bytes += skb->len;
+		sch->stats.packets++;
+	}
+	return 0;
+}
+
+/* Requeue packets but don't change time stamp */
+static int dly_requeue(struct sk_buff *skb, struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	int ret;
+
+	ret = q->qdisc->ops->requeue(skb, q->qdisc);
+	if (ret == 0)
+		sch->q.qlen++;
+	return ret;
+}
+
+static unsigned int dly_drop(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	unsigned int len;
+
+	len = q->qdisc->ops->drop(q->qdisc);
+	if (len) {
+		sch->q.qlen--;
+		sch->stats.drops++;
+	}
+	return len;
+}
+
+/* Dequeue packet.
+ * If packet needs to be held up, then stop the
+ * queue and set timer to wakeup later.
+ */
+static struct sk_buff *dly_dequeue(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct sk_buff *skb = q->qdisc->dequeue(q->qdisc);
+
+	if (skb) {
+		struct dly_skb_cb *cb = (struct dly_skb_cb *)skb->cb;
+		psched_time_t now;
+		long diff;
+
+		PSCHED_GET_TIME(now);
+		diff = q->latency - PSCHED_TDIFF(now, cb->queuetime);
+
+		if (diff <= 0) {
+			sch->q.qlen--;
+			sch->flags &= ~TCQ_F_THROTTLED;
+			return skb;
+		}
+
+		if (!netif_queue_stopped(sch->dev)) {
+			long delay = PSCHED_US2JIFFIE(diff);
+			if (delay <= 0)
+				delay = 1;
+			mod_timer(&q->timer, jiffies+delay);
+		}
+
+		if (q->qdisc->ops->requeue(skb, q->qdisc) != NET_XMIT_SUCCESS) {
+			sch->q.qlen--;
+			sch->stats.drops++;
+		}
+		sch->flags |= TCQ_F_THROTTLED;
+	}
+	return NULL;
+}
+
+static void dly_reset(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+
+	qdisc_reset(q->qdisc);
+	sch->q.qlen = 0;
+	sch->flags &= ~TCQ_F_THROTTLED;
+	del_timer(&q->timer);
+}
+
+static void dly_timer(unsigned long arg)
+{
+	struct Qdisc *sch = (struct Qdisc *)arg;
+
+	sch->flags &= ~TCQ_F_THROTTLED;
+	netif_schedule(sch->dev);
+}
+
+/* Tell Fifo the new limit. */
+static int change_limit(struct Qdisc *q, u32 limit)
+{
+	struct rtattr *rta;
+	int ret;
+
+	rta = kmalloc(RTA_LENGTH(sizeof(struct tc_fifo_qopt)), GFP_KERNEL);
+	if (!rta)
+		return -ENOMEM;
+
+	rta->rta_type = RTM_NEWQDISC;
+	((struct tc_fifo_qopt *)RTA_DATA(rta))->limit = limit;
+	ret = q->ops->change(q, rta);
+	kfree(rta);
+
+	return ret;
+}
+
+/* Setup underlying FIFO discipline */
+static int dly_change(struct Qdisc *sch, struct rtattr *opt)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct tc_dly_qopt *qopt = RTA_DATA(opt);
+	int err;
+
+	if (q->qdisc == &noop_qdisc) {
+		struct Qdisc *child
+			= qdisc_create_dflt(sch->dev, &bfifo_qdisc_ops);
+		if (!child)
+			return -EINVAL;
+		q->qdisc = child;
+	}
+
+	err = change_limit(q->qdisc, qopt->limit);
+	if (err) {
+		qdisc_destroy(q->qdisc);
+		q->qdisc = &noop_qdisc;
+	} else {
+		q->latency = qopt->latency;
+		q->limit = qopt->limit;
+	}
+	return err;
+}
+
+static int dly_init(struct Qdisc *sch, struct rtattr *opt)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+
+	if (!opt)
+		return -EINVAL;
+
+	init_timer(&q->timer);
+	q->timer.function = dly_timer;
+	q->timer.data = (unsigned long) sch;
+	q->qdisc = &noop_qdisc;
+
+	return dly_change(sch, opt);
+}
+
+static void dly_destroy(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+
+	del_timer(&q->timer);
+	qdisc_destroy(q->qdisc);
+	q->qdisc = &noop_qdisc;
+}
+
+static int dly_dump(struct Qdisc *sch, struct sk_buff *skb)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	unsigned char	 *b = skb->tail;
+	struct tc_dly_qopt qopt;
+
+	qopt.latency = q->latency;
+	qopt.limit = q->limit;
+
+	RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
+
+	return skb->len;
+
+rtattr_failure:
+	skb_trim(skb, b - skb->data);
+	return -1;
+}
+
+static struct Qdisc_ops dly_qdisc_ops = {
+	.id		=	"delay",
+	.priv_size	=	sizeof(struct dly_sched_data),
+	.enqueue	=	dly_enqueue,
+	.dequeue	=	dly_dequeue,
+	.requeue	=	dly_requeue,
+	.drop		=	dly_drop,
+	.init		=	dly_init,
+	.reset		=	dly_reset,
+	.destroy	=	dly_destroy,
+	.change		=	dly_change,
+	.dump		=	dly_dump,
+	.owner		=	THIS_MODULE,
+};
+
+
+static int __init dly_module_init(void)
+{
+	return register_qdisc(&dly_qdisc_ops);
+}
+static void __exit dly_module_exit(void)
+{
+	unregister_qdisc(&dly_qdisc_ops);
+}
+module_init(dly_module_init)
+module_exit(dly_module_exit)
+MODULE_LICENSE("GPL");

Re: [PATCH] packet delay scheduler

[David S. Miller]

On Thu, 18 Mar 2004 12:04:51 -0800
Stephen Hemminger <shemminger@osdl.org> wrote:

> Okay, here is a very simple QOS scheduler that delays packets for 2.6.
> It is good for testing, and might be useful for people who want to put
> some class of traffic into a "penalty box".

I'm happy with this and added it to my 2.6.x tree.

Could you cook up a 2.4.x variant for me?  Thanks.

[LARTC] Re: [PATCH] packet delay scheduler

[Stephen Hemminger]

Okay, here is a 2.4 version.


diff -Nru a/Documentation/Configure.help b/Documentation/Configure.help
--- a/Documentation/Configure.help	Fri Mar 19 10:46:19 2004
+++ b/Documentation/Configure.help	Fri Mar 19 10:46:19 2004
@@ -10829,6 +10829,14 @@
   whenever you want). If you want to compile it as a module, say M
   here and read <file:Documentation/modules.txt>.
 
+Network delay simualtor  
+CONFIG_NET_SCH_DELAY
+  Say Y if you want to delay packets by a fixed amount of
+  time. This is often useful to simulate network delay when
+  testing applications or protocols.
+  
+  This code is also available as a module called sch_delay.o
+
 Ingress Qdisc
 CONFIG_NET_SCH_INGRESS
   If you say Y here, you will be able to police incoming bandwidth
diff -Nru a/include/linux/pkt_sched.h b/include/linux/pkt_sched.h
--- a/include/linux/pkt_sched.h	Fri Mar 19 10:46:19 2004
+++ b/include/linux/pkt_sched.h	Fri Mar 19 10:46:19 2004
@@ -432,4 +432,11 @@
 
 #define TCA_ATM_MAX	TCA_ATM_STATE
 
+/* Delay section */
+struct tc_dly_qopt
+{
+	__u32	latency;
+	__u32   limit;
+};	
+
 #endif
diff -Nru a/net/sched/Config.in b/net/sched/Config.in
--- a/net/sched/Config.in	Fri Mar 19 10:46:19 2004
+++ b/net/sched/Config.in	Fri Mar 19 10:46:19 2004
@@ -15,6 +15,7 @@
 tristate '  TEQL queue' CONFIG_NET_SCH_TEQL
 tristate '  TBF queue' CONFIG_NET_SCH_TBF
 tristate '  GRED queue' CONFIG_NET_SCH_GRED
+tristate '  Network delay simulator' CONFIG_NET_SCH_DELAY
 tristate '  Diffserv field marker' CONFIG_NET_SCH_DSMARK
 if [ "$CONFIG_NETFILTER" = "y" ]; then
    tristate '  Ingress Qdisc' CONFIG_NET_SCH_INGRESS
diff -Nru a/net/sched/Makefile b/net/sched/Makefile
--- a/net/sched/Makefile	Fri Mar 19 10:46:19 2004
+++ b/net/sched/Makefile	Fri Mar 19 10:46:19 2004
@@ -14,6 +14,7 @@
 obj-$(CONFIG_NET_SCH_INGRESS)	+= sch_ingress.o 
 obj-$(CONFIG_NET_SCH_CBQ)	+= sch_cbq.o
 obj-$(CONFIG_NET_SCH_CSZ)	+= sch_csz.o
+obj-$(CONFIG_NET_SCH_DELAY)	+= sch_delay.o
 obj-$(CONFIG_NET_SCH_HPFQ)	+= sch_hpfq.o
 obj-$(CONFIG_NET_SCH_HFSC)	+= sch_hfsc.o
 obj-$(CONFIG_NET_SCH_HTB)	+= sch_htb.o
diff -Nru a/net/sched/sch_delay.c b/net/sched/sch_delay.c
--- /dev/null	Wed Dec 31 16:00:00 1969
+++ b/net/sched/sch_delay.c	Fri Mar 19 10:46:19 2004
@@ -0,0 +1,277 @@
+/*
+ * net/sched/sch_delay.c	Simple constant delay
+ *
+ * 		This program is free software; you can redistribute it and/or
+ * 		modify it under the terms of the GNU General Public License
+ * 		as published by the Free Software Foundation; either version
+ * 		2 of the License, or (at your option) any later version.
+ *
+ * Authors:	Stephen Hemminger <shemminger@osdl.org>
+ */
+
+#include <linux/config.h>
+#include <linux/module.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+
+#include <linux/string.h>
+#include <linux/mm.h>
+#include <linux/socket.h>
+#include <linux/sockios.h>
+#include <linux/in.h>
+#include <linux/errno.h>
+#include <linux/interrupt.h>
+#include <linux/if_ether.h>
+#include <linux/inet.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/notifier.h>
+#include <net/ip.h>
+#include <net/route.h>
+#include <linux/skbuff.h>
+#include <net/sock.h>
+#include <net/pkt_sched.h>
+
+/*	Network delay simulator
+	This scheduler adds a fixed delay to all packets.
+	Similar to NISTnet and BSD Dummynet.
+
+	It uses byte fifo underneath similar to TBF */
+struct dly_sched_data {
+	u32	latency;
+	u32	limit;
+	struct timer_list timer;
+	struct Qdisc *qdisc;
+};
+
+/* Time stamp put into socket buffer control block */
+struct dly_skb_cb {
+	psched_time_t	queuetime;
+};
+
+/* Enqueue packets with underlying discipline (fifo)
+ * but mark them with current time first.
+ */
+static int dly_enqueue(struct sk_buff *skb, struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct dly_skb_cb *cb = (struct dly_skb_cb *)skb->cb;
+	int ret;
+
+	PSCHED_GET_TIME(cb->queuetime);
+
+	/* Queue to underlying scheduler */
+	ret = q->qdisc->enqueue(skb, q->qdisc);
+	if (ret)
+		sch->stats.drops++;
+	else {
+		sch->q.qlen++;
+		sch->stats.bytes += skb->len;
+		sch->stats.packets++;
+	}
+	return 0;
+}
+
+/* Requeue packets but don't change time stamp */
+static int dly_requeue(struct sk_buff *skb, struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	int ret;
+
+	ret = q->qdisc->ops->requeue(skb, q->qdisc);
+	if (ret = 0)
+		sch->q.qlen++;
+	return ret;
+}
+
+static unsigned int dly_drop(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	unsigned int len;
+
+	len = q->qdisc->ops->drop(q->qdisc);
+	if (len) {
+		sch->q.qlen--;
+		sch->stats.drops++;
+	}
+	return len;
+}
+
+/* Dequeue packet.
+ * If packet needs to be held up, then stop the
+ * queue and set timer to wakeup later.
+ */
+static struct sk_buff *dly_dequeue(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct sk_buff *skb = q->qdisc->dequeue(q->qdisc);
+
+	if (skb) {
+		struct dly_skb_cb *cb = (struct dly_skb_cb *)skb->cb;
+		psched_time_t now;
+		long diff;
+
+		PSCHED_GET_TIME(now);
+		diff = q->latency - PSCHED_TDIFF(now, cb->queuetime);
+
+		if (diff <= 0) {
+			sch->q.qlen--;
+			sch->flags &= ~TCQ_F_THROTTLED;
+			return skb;
+		}
+
+		if (!netif_queue_stopped(sch->dev)) {
+			long delay = PSCHED_US2JIFFIE(diff);
+			if (delay <= 0)
+				delay = 1;
+			mod_timer(&q->timer, jiffies+delay);
+		}
+
+		if (q->qdisc->ops->requeue(skb, q->qdisc) != NET_XMIT_SUCCESS) {
+			sch->q.qlen--;
+			sch->stats.drops++;
+		}
+		sch->flags |= TCQ_F_THROTTLED;
+	}
+	return NULL;
+}
+
+static void dly_reset(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+
+	qdisc_reset(q->qdisc);
+	sch->q.qlen = 0;
+	sch->flags &= ~TCQ_F_THROTTLED;
+	del_timer(&q->timer);
+}
+
+static void dly_timer(unsigned long arg)
+{
+	struct Qdisc *sch = (struct Qdisc *)arg;
+
+	sch->flags &= ~TCQ_F_THROTTLED;
+	netif_schedule(sch->dev);
+}
+
+/* Tell Fifo the new limit. */
+static int change_limit(struct Qdisc *q, u32 limit)
+{
+	struct rtattr *rta;
+	int ret;
+
+	rta = kmalloc(RTA_LENGTH(sizeof(struct tc_fifo_qopt)), GFP_KERNEL);
+	if (!rta)
+		return -ENOMEM;
+
+	rta->rta_type = RTM_NEWQDISC;
+	((struct tc_fifo_qopt *)RTA_DATA(rta))->limit = limit;
+	ret = q->ops->change(q, rta);
+	kfree(rta);
+
+	return ret;
+}
+
+/* Setup underlying FIFO discipline */
+static int dly_change(struct Qdisc *sch, struct rtattr *opt)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct tc_dly_qopt *qopt = RTA_DATA(opt);
+	int err;
+
+	if (q->qdisc = &noop_qdisc) {
+		struct Qdisc *child
+			= qdisc_create_dflt(sch->dev, &bfifo_qdisc_ops);
+		if (!child)
+			return -EINVAL;
+		q->qdisc = child;
+	}
+
+	err = change_limit(q->qdisc, qopt->limit);
+	if (err) {
+		qdisc_destroy(q->qdisc);
+		q->qdisc = &noop_qdisc;
+	} else {
+		q->latency = qopt->latency;
+		q->limit = qopt->limit;
+	}
+	return err;
+}
+
+static int dly_init(struct Qdisc *sch, struct rtattr *opt)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	int err;
+
+	if (!opt)
+		return -EINVAL;
+
+	MOD_INC_USE_COUNT;
+
+	init_timer(&q->timer);
+	q->timer.function = dly_timer;
+	q->timer.data = (unsigned long) sch;
+	q->qdisc = &noop_qdisc;
+
+	err = dly_change(sch, opt);
+	if (err)
+		MOD_DEC_USE_COUNT;
+
+	return err;
+}
+
+static void dly_destroy(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+
+	del_timer(&q->timer);
+	qdisc_destroy(q->qdisc);
+	q->qdisc = &noop_qdisc;
+
+	MOD_DEC_USE_COUNT;
+}
+
+static int dly_dump(struct Qdisc *sch, struct sk_buff *skb)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	unsigned char	 *b = skb->tail;
+	struct tc_dly_qopt qopt;
+
+	qopt.latency = q->latency;
+	qopt.limit = q->limit;
+
+	RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
+
+	return skb->len;
+
+rtattr_failure:
+	skb_trim(skb, b - skb->data);
+	return -1;
+}
+
+struct Qdisc_ops dly_qdisc_ops = {
+	.id		=	"delay",
+	.priv_size	=	sizeof(struct dly_sched_data),
+	.enqueue	=	dly_enqueue,
+	.dequeue	=	dly_dequeue,
+	.requeue	=	dly_requeue,
+	.drop		=	dly_drop,
+	.init		=	dly_init,
+	.reset		=	dly_reset,
+	.destroy	=	dly_destroy,
+	.change		=	dly_change,
+	.dump		=	dly_dump,
+};
+
+#ifdef MODULE
+int init_module(void)
+{
+	return register_qdisc(&dly_qdisc_ops);
+}
+
+void cleanup_module(void)
+{
+	unregister_qdisc(&dly_qdisc_ops);
+}
+#endif
+MODULE_LICENSE("GPL");
_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

Re: [PATCH] packet delay scheduler

[Stephen Hemminger]

Okay, here is a 2.4 version.


diff -Nru a/Documentation/Configure.help b/Documentation/Configure.help
--- a/Documentation/Configure.help	Fri Mar 19 10:46:19 2004
+++ b/Documentation/Configure.help	Fri Mar 19 10:46:19 2004
@@ -10829,6 +10829,14 @@
   whenever you want). If you want to compile it as a module, say M
   here and read <file:Documentation/modules.txt>.
 
+Network delay simualtor  
+CONFIG_NET_SCH_DELAY
+  Say Y if you want to delay packets by a fixed amount of
+  time. This is often useful to simulate network delay when
+  testing applications or protocols.
+  
+  This code is also available as a module called sch_delay.o
+
 Ingress Qdisc
 CONFIG_NET_SCH_INGRESS
   If you say Y here, you will be able to police incoming bandwidth
diff -Nru a/include/linux/pkt_sched.h b/include/linux/pkt_sched.h
--- a/include/linux/pkt_sched.h	Fri Mar 19 10:46:19 2004
+++ b/include/linux/pkt_sched.h	Fri Mar 19 10:46:19 2004
@@ -432,4 +432,11 @@
 
 #define TCA_ATM_MAX	TCA_ATM_STATE
 
+/* Delay section */
+struct tc_dly_qopt
+{
+	__u32	latency;
+	__u32   limit;
+};	
+
 #endif
diff -Nru a/net/sched/Config.in b/net/sched/Config.in
--- a/net/sched/Config.in	Fri Mar 19 10:46:19 2004
+++ b/net/sched/Config.in	Fri Mar 19 10:46:19 2004
@@ -15,6 +15,7 @@
 tristate '  TEQL queue' CONFIG_NET_SCH_TEQL
 tristate '  TBF queue' CONFIG_NET_SCH_TBF
 tristate '  GRED queue' CONFIG_NET_SCH_GRED
+tristate '  Network delay simulator' CONFIG_NET_SCH_DELAY
 tristate '  Diffserv field marker' CONFIG_NET_SCH_DSMARK
 if [ "$CONFIG_NETFILTER" = "y" ]; then
    tristate '  Ingress Qdisc' CONFIG_NET_SCH_INGRESS
diff -Nru a/net/sched/Makefile b/net/sched/Makefile
--- a/net/sched/Makefile	Fri Mar 19 10:46:19 2004
+++ b/net/sched/Makefile	Fri Mar 19 10:46:19 2004
@@ -14,6 +14,7 @@
 obj-$(CONFIG_NET_SCH_INGRESS)	+= sch_ingress.o 
 obj-$(CONFIG_NET_SCH_CBQ)	+= sch_cbq.o
 obj-$(CONFIG_NET_SCH_CSZ)	+= sch_csz.o
+obj-$(CONFIG_NET_SCH_DELAY)	+= sch_delay.o
 obj-$(CONFIG_NET_SCH_HPFQ)	+= sch_hpfq.o
 obj-$(CONFIG_NET_SCH_HFSC)	+= sch_hfsc.o
 obj-$(CONFIG_NET_SCH_HTB)	+= sch_htb.o
diff -Nru a/net/sched/sch_delay.c b/net/sched/sch_delay.c
--- /dev/null	Wed Dec 31 16:00:00 1969
+++ b/net/sched/sch_delay.c	Fri Mar 19 10:46:19 2004
@@ -0,0 +1,277 @@
+/*
+ * net/sched/sch_delay.c	Simple constant delay
+ *
+ * 		This program is free software; you can redistribute it and/or
+ * 		modify it under the terms of the GNU General Public License
+ * 		as published by the Free Software Foundation; either version
+ * 		2 of the License, or (at your option) any later version.
+ *
+ * Authors:	Stephen Hemminger <shemminger@osdl.org>
+ */
+
+#include <linux/config.h>
+#include <linux/module.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+
+#include <linux/string.h>
+#include <linux/mm.h>
+#include <linux/socket.h>
+#include <linux/sockios.h>
+#include <linux/in.h>
+#include <linux/errno.h>
+#include <linux/interrupt.h>
+#include <linux/if_ether.h>
+#include <linux/inet.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/notifier.h>
+#include <net/ip.h>
+#include <net/route.h>
+#include <linux/skbuff.h>
+#include <net/sock.h>
+#include <net/pkt_sched.h>
+
+/*	Network delay simulator
+	This scheduler adds a fixed delay to all packets.
+	Similar to NISTnet and BSD Dummynet.
+
+	It uses byte fifo underneath similar to TBF */
+struct dly_sched_data {
+	u32	latency;
+	u32	limit;
+	struct timer_list timer;
+	struct Qdisc *qdisc;
+};
+
+/* Time stamp put into socket buffer control block */
+struct dly_skb_cb {
+	psched_time_t	queuetime;
+};
+
+/* Enqueue packets with underlying discipline (fifo)
+ * but mark them with current time first.
+ */
+static int dly_enqueue(struct sk_buff *skb, struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct dly_skb_cb *cb = (struct dly_skb_cb *)skb->cb;
+	int ret;
+
+	PSCHED_GET_TIME(cb->queuetime);
+
+	/* Queue to underlying scheduler */
+	ret = q->qdisc->enqueue(skb, q->qdisc);
+	if (ret)
+		sch->stats.drops++;
+	else {
+		sch->q.qlen++;
+		sch->stats.bytes += skb->len;
+		sch->stats.packets++;
+	}
+	return 0;
+}
+
+/* Requeue packets but don't change time stamp */
+static int dly_requeue(struct sk_buff *skb, struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	int ret;
+
+	ret = q->qdisc->ops->requeue(skb, q->qdisc);
+	if (ret == 0)
+		sch->q.qlen++;
+	return ret;
+}
+
+static unsigned int dly_drop(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	unsigned int len;
+
+	len = q->qdisc->ops->drop(q->qdisc);
+	if (len) {
+		sch->q.qlen--;
+		sch->stats.drops++;
+	}
+	return len;
+}
+
+/* Dequeue packet.
+ * If packet needs to be held up, then stop the
+ * queue and set timer to wakeup later.
+ */
+static struct sk_buff *dly_dequeue(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct sk_buff *skb = q->qdisc->dequeue(q->qdisc);
+
+	if (skb) {
+		struct dly_skb_cb *cb = (struct dly_skb_cb *)skb->cb;
+		psched_time_t now;
+		long diff;
+
+		PSCHED_GET_TIME(now);
+		diff = q->latency - PSCHED_TDIFF(now, cb->queuetime);
+
+		if (diff <= 0) {
+			sch->q.qlen--;
+			sch->flags &= ~TCQ_F_THROTTLED;
+			return skb;
+		}
+
+		if (!netif_queue_stopped(sch->dev)) {
+			long delay = PSCHED_US2JIFFIE(diff);
+			if (delay <= 0)
+				delay = 1;
+			mod_timer(&q->timer, jiffies+delay);
+		}
+
+		if (q->qdisc->ops->requeue(skb, q->qdisc) != NET_XMIT_SUCCESS) {
+			sch->q.qlen--;
+			sch->stats.drops++;
+		}
+		sch->flags |= TCQ_F_THROTTLED;
+	}
+	return NULL;
+}
+
+static void dly_reset(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+
+	qdisc_reset(q->qdisc);
+	sch->q.qlen = 0;
+	sch->flags &= ~TCQ_F_THROTTLED;
+	del_timer(&q->timer);
+}
+
+static void dly_timer(unsigned long arg)
+{
+	struct Qdisc *sch = (struct Qdisc *)arg;
+
+	sch->flags &= ~TCQ_F_THROTTLED;
+	netif_schedule(sch->dev);
+}
+
+/* Tell Fifo the new limit. */
+static int change_limit(struct Qdisc *q, u32 limit)
+{
+	struct rtattr *rta;
+	int ret;
+
+	rta = kmalloc(RTA_LENGTH(sizeof(struct tc_fifo_qopt)), GFP_KERNEL);
+	if (!rta)
+		return -ENOMEM;
+
+	rta->rta_type = RTM_NEWQDISC;
+	((struct tc_fifo_qopt *)RTA_DATA(rta))->limit = limit;
+	ret = q->ops->change(q, rta);
+	kfree(rta);
+
+	return ret;
+}
+
+/* Setup underlying FIFO discipline */
+static int dly_change(struct Qdisc *sch, struct rtattr *opt)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	struct tc_dly_qopt *qopt = RTA_DATA(opt);
+	int err;
+
+	if (q->qdisc == &noop_qdisc) {
+		struct Qdisc *child
+			= qdisc_create_dflt(sch->dev, &bfifo_qdisc_ops);
+		if (!child)
+			return -EINVAL;
+		q->qdisc = child;
+	}
+
+	err = change_limit(q->qdisc, qopt->limit);
+	if (err) {
+		qdisc_destroy(q->qdisc);
+		q->qdisc = &noop_qdisc;
+	} else {
+		q->latency = qopt->latency;
+		q->limit = qopt->limit;
+	}
+	return err;
+}
+
+static int dly_init(struct Qdisc *sch, struct rtattr *opt)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	int err;
+
+	if (!opt)
+		return -EINVAL;
+
+	MOD_INC_USE_COUNT;
+
+	init_timer(&q->timer);
+	q->timer.function = dly_timer;
+	q->timer.data = (unsigned long) sch;
+	q->qdisc = &noop_qdisc;
+
+	err = dly_change(sch, opt);
+	if (err)
+		MOD_DEC_USE_COUNT;
+
+	return err;
+}
+
+static void dly_destroy(struct Qdisc *sch)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+
+	del_timer(&q->timer);
+	qdisc_destroy(q->qdisc);
+	q->qdisc = &noop_qdisc;
+
+	MOD_DEC_USE_COUNT;
+}
+
+static int dly_dump(struct Qdisc *sch, struct sk_buff *skb)
+{
+	struct dly_sched_data *q = (struct dly_sched_data *)sch->data;
+	unsigned char	 *b = skb->tail;
+	struct tc_dly_qopt qopt;
+
+	qopt.latency = q->latency;
+	qopt.limit = q->limit;
+
+	RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
+
+	return skb->len;
+
+rtattr_failure:
+	skb_trim(skb, b - skb->data);
+	return -1;
+}
+
+struct Qdisc_ops dly_qdisc_ops = {
+	.id		=	"delay",
+	.priv_size	=	sizeof(struct dly_sched_data),
+	.enqueue	=	dly_enqueue,
+	.dequeue	=	dly_dequeue,
+	.requeue	=	dly_requeue,
+	.drop		=	dly_drop,
+	.init		=	dly_init,
+	.reset		=	dly_reset,
+	.destroy	=	dly_destroy,
+	.change		=	dly_change,
+	.dump		=	dly_dump,
+};
+
+#ifdef MODULE
+int init_module(void)
+{
+	return register_qdisc(&dly_qdisc_ops);
+}
+
+void cleanup_module(void)
+{
+	unregister_qdisc(&dly_qdisc_ops);
+}
+#endif
+MODULE_LICENSE("GPL");

Re: [PATCH] packet delay scheduler

[David S. Miller]

On Fri, 19 Mar 2004 10:47:05 -0800
Stephen Hemminger <shemminger@osdl.org> wrote:

> Okay, here is a 2.4 version.

Applied, thanks Stephen.

Re: [PATCH] packet delay scheduler

[jamal]

Hi Stephen,

Good effort but we could do more to capture the essence of nistnet. I am
not familiar with dummynet so cant comment on it. 
For example just with the delaying action nistnet does a lot more
(cutnpaste):
Fixed delays, variable delays with loadable delay probability
distribution (Default distribution is pareto-normal, empirically
parametrized to match observed packet delays), settable mean/variance.

It can also introduce Packet reordering, packet loss, packet
duplication, as well as bandwidth limitation to induce congestion.

You get the point: essentially you have barely scratched the surface.

I think putting all the above in one huge monolithic module is wrong
this being one of the problems i had with nistnet.

The better alternative is to use the tc extension patches i have
because then you could write little modules which do different things
and cascade them. 

Maybe as an example i can rewrite the fixed delayer you wrote as an
action module. The only catch is the patches i have are still in 2.4.x
at the moment. Let me know if you are interested and i could pass you
the patches.

cheers,
jamal

Re: [PATCH] packet delay scheduler

[David S. Miller]

On 19 Mar 2004 09:52:07 -0500
jamal <hadi@cyberus.ca> wrote:

> The better alternative is to use the tc extension patches i have
> because then you could write little modules which do different things
> and cascade them. 

I agree that this chaining idea of tc actions is the end-all-be-all
way to do this kind of stuff.

But for now we can put Stephen's delay scheduler in the tree, and I
don't see any real problem with that until your tc action changes
are ready for integration.

Re: [PATCH] packet delay scheduler

[jamal]

On Fri, 2004-03-19 at 17:21, David S. Miller wrote:

> But for now we can put Stephen's delay scheduler in the tree, and I
> don't see any real problem with that until your tc action changes
> are ready for integration.

Absolutely - sorry intent was just to motivate not reject.

I would say the extension patches are ready for inclusion but we can
take that offline. I have a few people testing them under more rigorous
conditions at the moment.

cheers,
jamal

Re: network delay simulation

[Ben Greear]

Stephen Hemminger wrote:
> Still bumming around for how to easily simulate long latencies.
> There is NISTnet but that is fugly old 2.2 code; and Dummynet, and hitbox 
> which are FreeBSD based. The existing traffic shaper might do, but it seems
> to be limited to lower speed lines and needs some work (using atomic_set
> do own locking for instance).
> 
> Probably "the cowboy way" would be to right a new net/scheduler to just
> do FIFO delay.
> 
> Any ideas/comments? Somebody have some code sitting in a drawer?

I'll give you licenses to my (closed source) LANforge-ICE WAN emulator which runs
as a Linux module and has been benchmarked at 1Gbps full-duplex (well, 999Mbps).

It can do up to about 1 second delay at 1Gbps, but with more delay
at this high speed it runs out of memory (seems skb_alloc, or the
way I am calling it, will not use high-memory).  At lower speeds you
can have several seconds of delay.

Please send me email off the list if you are interested.

Thanks,
Ben

-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com