Skip to contentSkip to conversation

sch_htb: turn intermediate classes into leaves

PATCH In review Jarek Poplawski 3 participants 8 messages Nov 27, 2006 – Dec 8, 2006

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sch_htb: turn intermediate classes into leaves

Here is a trial to do something suggested by Patrick McHardy.

[NET_SCHED] sch_htb:

- turn intermediate classes into leaves again when their last child is deleted (qdisc of deleted class is reused; struct htb_class changed)

- sch_tree_lock added in htb_put before htb_destroy_class (for consistency with htb_delete and htb_destroy) - my own suggestion

PS: qdisc_reset added to htb_delete by P. McHardy's patch: "perform qlen adjustment immediately in ->delete" should be reconsidered.

---

net/sched/sch_htb.c +46 -1
diff -Nurp linux-2.6.19-rc6-/net/sched/sch_htb.c linux-2.6.19-rc6/net/sched/sch_htb.c
@@ -147,6 +147,10 @@ struct htb_class {
 	psched_tdiff_t mbuffer;	/* max wait time */
 	long tokens, ctokens;	/* current number of tokens */
 	psched_time_t t_c;	/* checkpoint time */
+	
+	int prio;		/* For parent to leaf return possible here */
+	int quantum;		/* we do backup. Finally full replacement  */
+				/* of un.leaf originals should be done. */
 };
 
 /* TODO: maybe compute rate when size is too large .. or drop ? */
@@ -1266,6 +1270,37 @@ static void htb_destroy_filters(struct t
 	}
 }
 
+static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl)
+{
+	struct htb_class *parent = cl->parent;
+
+	if (!parent)
+		/* the root class */
+		return;
+
+	BUG_TRAP(!cl->level && cl->un.leaf.q && !cl->prio_activity);
+
+	if (!(parent->children.next == &cl->sibling &&
+		parent->children.prev == &cl->sibling))
+		/* not the last child */
+		return;	
+
+	parent->level = 0;
+	memset(&parent->un.inner, 0, sizeof(parent->un.inner));
+	INIT_LIST_HEAD(&parent->un.leaf.drop_list);
+	parent->un.leaf.q = cl->un.leaf.q;
+	cl->un.leaf.q = &noop_qdisc;
+	parent->un.leaf.quantum = parent->quantum;
+	parent->un.leaf.prio = parent->prio;
+	parent->tokens = parent->buffer;
+	parent->ctokens = parent->cbuffer;
+	PSCHED_GET_TIME(parent->t_c);
+	parent->cmode = HTB_CAN_SEND;
+
+	if (parent->un.leaf.q->q.qlen)
+		htb_activate(q, parent);
+}
+
 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
 {
 	struct htb_sched *q = qdisc_priv(sch);
@@ -1337,6 +1372,9 @@ static int htb_delete(struct Qdisc *sch,
 	if (cl->prio_activity)
 		htb_deactivate(q, cl);
 
+	if (!cl->level)
+		htb_parent_to_leaf(q, cl);
+		
 	if (--cl->refcnt == 0)
 		htb_destroy_class(sch, cl);
 
@@ -1348,8 +1386,11 @@ static void htb_put(struct Qdisc *sch, u
 {
 	struct htb_class *cl = (struct htb_class *)arg;
 
-	if (--cl->refcnt == 0)
+	if (--cl->refcnt == 0) {
+		sch_tree_lock(sch);
 		htb_destroy_class(sch, cl);
+		sch_tree_unlock(sch);
+	}
 }
 
 static int htb_change_class(struct Qdisc *sch, u32 classid,
@@ -1468,6 +1509,10 @@ static int htb_change_class(struct Qdisc
 			cl->un.leaf.quantum = hopt->quantum;
 		if ((cl->un.leaf.prio = hopt->prio) >= TC_HTB_NUMPRIO)
 			cl->un.leaf.prio = TC_HTB_NUMPRIO - 1;
+
+		/* backup for htb_parent_to_leaf */
+		cl->quantum = cl->un.leaf.quantum;
+		cl->prio = cl->un.leaf.prio;
 	}
 
 	cl->buffer = hopt->buffer;
#1 lore raw
Quoted 9 lines · Jarek Poplawski

Here is a trial to do something suggested by Patrick McHardy.

[NET_SCHED] sch_htb:

- turn intermediate classes into leaves again when their last child is deleted (qdisc of deleted class is reused; struct htb_class changed)

- sch_tree_lock added in htb_put before htb_destroy_class (for consistency with htb_delete and htb_destroy) - my own suggestion

->put() doesn't need the lock - if a class is deleted it must be completely unlinked in ->delete().

Quoted 2 lines

PS: qdisc_reset added to htb_delete by P. McHardy's patch: "perform qlen adjustment immediately in ->delete" should be reconsidered.

No, that is a necessary fix, even though it only causes a shortly visible error.

Quoted 28 lines

diff -Nurp linux-2.6.19-rc6-/net/sched/sch_htb.c linux-2.6.19-rc6/net/sched/sch_htb.c

net/sched/sch_htb.c +14 -0
@@ -147,6 +147,10 @@ struct htb_class {
 	psched_tdiff_t mbuffer;	/* max wait time */
 	long tokens, ctokens;	/* current number of tokens */
 	psched_time_t t_c;	/* checkpoint time */
+	
+	int prio;		/* For parent to leaf return possible here */
+	int quantum;		/* we do backup. Finally full replacement  */
+				/* of un.leaf originals should be done. */
 };
 
 /* TODO: maybe compute rate when size is too large .. or drop ? */
@@ -1266,6 +1270,37 @@ static void htb_destroy_filters(struct t
 	}
 }
 
+static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl)
+{
+	struct htb_class *parent = cl->parent;
+
+	if (!parent)
+		/* the root class */
+		return;
+
+	BUG_TRAP(!cl->level && cl->un.leaf.q && !cl->prio_activity);
+

net/sched/sch_htb.c static void htb_destroy_filters(struct t · line 1292

Quoted 10 lines
net/sched/sch_htb.c +10 -0
+	if (!(parent->children.next == &cl->sibling &&
+		parent->children.prev == &cl->sibling))
+		/* not the last child */
+		return;	
+
+	parent->level = 0;
+	memset(&parent->un.inner, 0, sizeof(parent->un.inner));
+	INIT_LIST_HEAD(&parent->un.leaf.drop_list);
+	parent->un.leaf.q = cl->un.leaf.q;
+	cl->un.leaf.q = &noop_qdisc;

default pfifo would be a better choice. Might be a bit ugly though because you're holding the qdisc lock here and can't call qdisc_create_dflt. Since you're already keeping backup values from the union we could consider just turning it into two seperate structures and keep the child qdisc when turning a class into a parent.

net/sched/sch_htb.c static void htb_destroy_filters(struct t · line 1301

Quoted 9 lines
net/sched/sch_htb.c +9 -0
+	parent->un.leaf.quantum = parent->quantum;
+	parent->un.leaf.prio = parent->prio;
+	parent->tokens = parent->buffer;
+	parent->ctokens = parent->cbuffer;
+	PSCHED_GET_TIME(parent->t_c);
+	parent->cmode = HTB_CAN_SEND;
+
+	if (parent->un.leaf.q->q.qlen)
+		htb_activate(q, parent);

Not possible right now and even if we reuse the old child qdisc it should be empty.

net/sched/sch_htb.c static int htb_delete(struct Qdisc *sch, · line 1377

Quoted 12 lines
net/sched/sch_htb.c +5 -0
+}
+
 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
 {
 	struct htb_sched *q = qdisc_priv(sch);
@@ -1337,6 +1372,9 @@ static int htb_delete(struct Qdisc *sch,
 	if (cl->prio_activity)
 		htb_deactivate(q, cl);
 
+	if (!cl->level)
+		htb_parent_to_leaf(q, cl);
+		

You have to manually adjust the classes level before checking for zero, it is not done currently.

Quoted 4 lines
net/sched/sch_htb.c +0 -0
 	if (--cl->refcnt == 0)
 		htb_destroy_class(sch, cl);
 

net/sched/sch_htb.c static void htb_put(struct Qdisc *sch, u · line 1394

Quoted 10 lines
net/sched/sch_htb.c +4 -1
@@ -1348,8 +1386,11 @@ static void htb_put(struct Qdisc *sch, u
 {
 	struct htb_class *cl = (struct htb_class *)arg;
 
-	if (--cl->refcnt == 0)
+	if (--cl->refcnt == 0) {
+		sch_tree_lock(sch);
 		htb_destroy_class(sch, cl);
+		sch_tree_unlock(sch);
+	}
 }

See above.

#2 lore raw
Quoted 5 lines · Patrick McHardy
Quoted 4 lines · Jarek Poplawski

Here is a trial to do something suggested by Patrick McHardy.

[NET_SCHED] sch_htb:

Quoted 8 lines
Quoted 5 lines

- turn intermediate classes into leaves again when their last child is deleted (qdisc of deleted class is reused; struct htb_class changed)

- sch_tree_lock added in htb_put before htb_destroy_class (for consistency with htb_delete and htb_destroy) - my own suggestion

->put() doesn't need the lock - if a class is deleted it must be completely unlinked in ->delete().

Could you give me some hint why not unlock before
htb_destroy_class call in htb_delete, please?

Quoted 38 lines
Quoted 2 lines

PS: qdisc_reset added to htb_delete by P. McHardy's patch: "perform qlen adjustment immediately in ->delete" should be reconsidered.

No, that is a necessary fix, even though it only causes a shortly visible error.

Quoted 32 lines

diff -Nurp linux-2.6.19-rc6-/net/sched/sch_htb.c linux-2.6.19-rc6/net/sched/sch_htb.c

net/sched/sch_htb.c +18 -0
@@ -147,6 +147,10 @@ struct htb_class {
 	psched_tdiff_t mbuffer;	/* max wait time */
 	long tokens, ctokens;	/* current number of tokens */
 	psched_time_t t_c;	/* checkpoint time */
+	
+	int prio;		/* For parent to leaf return possible here */
+	int quantum;		/* we do backup. Finally full replacement  */
+				/* of un.leaf originals should be done. */
 };
 
 /* TODO: maybe compute rate when size is too large .. or drop ? */
@@ -1266,6 +1270,37 @@ static void htb_destroy_filters(struct t
 	}
 }
 
+static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl)
+{
+	struct htb_class *parent = cl->parent;
+
+	if (!parent)
+		/* the root class */
+		return;
+
+	BUG_TRAP(!cl->level && cl->un.leaf.q && !cl->prio_activity);
+
+	if (!(parent->children.next == &cl->sibling &&
+		parent->children.prev == &cl->sibling))
+		/* not the last child */
+		return;	
Quoted 8 lines

net/sched/sch_htb.c static void htb_destroy_filters(struct t · line 1292

Quoted 6 lines
net/sched/sch_htb.c +6 -0
+
+	parent->level = 0;
+	memset(&parent->un.inner, 0, sizeof(parent->un.inner));
+	INIT_LIST_HEAD(&parent->un.leaf.drop_list);
+	parent->un.leaf.q = cl->un.leaf.q;
+	cl->un.leaf.q = &noop_qdisc;

default pfifo would be a better choice. Might be a bit ugly

I considered this unnecessary risk. But it can be done as you like. If I understand, you are not sure yet.

Quoted 5 lines

though because you're holding the qdisc lock here and can't call qdisc_create_dflt. Since you're already keeping backup values from the union we could consider just turning it into two seperate structures and keep the child qdisc when turning a class into a parent.

I've thought about this but it seems there could still be some memory saving with the union - probably - with large number of inner classes.

Quoted 4 lines
Quoted 4 lines
net/sched/sch_htb.c +4 -0
+	parent->un.leaf.quantum = parent->quantum;
+	parent->un.leaf.prio = parent->prio;
+	parent->tokens = parent->buffer;
+	parent->ctokens = parent->cbuffer;
Quoted 8 lines

net/sched/sch_htb.c static void htb_destroy_filters(struct t · line 1301

Quoted 5 lines
net/sched/sch_htb.c +5 -0
+	PSCHED_GET_TIME(parent->t_c);
+	parent->cmode = HTB_CAN_SEND;
+
+	if (parent->un.leaf.q->q.qlen)
+		htb_activate(q, parent);

Not possible right now and even if we reuse the old child qdisc it should be empty.

I'll remove this but then qdisc_reset have to be added. Or maybe I should forget rc6 and redo this only on top of your patch?

Quoted 9 lines
Quoted 9 lines
net/sched/sch_htb.c +2 -0
+}
+
 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
 {
 	struct htb_sched *q = qdisc_priv(sch);
@@ -1337,6 +1372,9 @@ static int htb_delete(struct Qdisc *sch,
 	if (cl->prio_activity)
 		htb_deactivate(q, cl);
 
Quoted 6 lines

net/sched/sch_htb.c static int htb_delete(struct Qdisc *sch, · line 1377

Quoted 3 lines
net/sched/sch_htb.c +3 -0
+	if (!cl->level)
+		htb_parent_to_leaf(q, cl);
+		

You have to manually adjust the classes level before checking for zero, it is not done currently.

Could you explain? Probably I miss something here.

Quoted 20 lines
Quoted 5 lines
net/sched/sch_htb.c +0 -0
 	if (--cl->refcnt == 0)
 		htb_destroy_class(sch, cl);
 

net/sched/sch_htb.c static void htb_put(struct Qdisc *sch, u · line 1394

Quoted 10 lines
net/sched/sch_htb.c +4 -1
@@ -1348,8 +1386,11 @@ static void htb_put(struct Qdisc *sch, u
 {
 	struct htb_class *cl = (struct htb_class *)arg;
 
-	if (--cl->refcnt == 0)
+	if (--cl->refcnt == 0) {
+		sch_tree_lock(sch);
 		htb_destroy_class(sch, cl);
+		sch_tree_unlock(sch);
+	}
 }

See above.

#3 lore raw

PATCH 1/1 sch_htb: turn intermediate classes into leaves

In reply to Patrick McHardy

[NET_SCHED] sch_htb:

[PATCH 2.6.19-rc6 with "Fix endless loops" set of patches]

- turn intermediate classes into leaves again when their last child is deleted (struct htb_class changed)

---

net/sched/sch_htb.c +51 -0
diff -Nurp linux-2.6.19-rc6-endless-/net/sched/sch_htb.c linux-2.6.19-rc6-endless/net/sched/sch_htb.c
@@ -147,6 +147,10 @@ struct htb_class {
 	psched_tdiff_t mbuffer;	/* max wait time */
 	long tokens, ctokens;	/* current number of tokens */
 	psched_time_t t_c;	/* checkpoint time */
+	
+	int prio;		/* For parent to leaf return possible here */
+	int quantum;		/* we do backup. Finally full replacement  */
+				/* of un.leaf originals should be done. */
 };
 
 /* TODO: maybe compute rate when size is too large .. or drop ? */
@@ -1271,6 +1275,38 @@ static void htb_destroy_filters(struct t
 	}
 }
 
+static inline int htb_parent_last_child(struct htb_class *cl)
+{
+	if (!cl->parent)
+		/* the root class */
+		return 0;
+
+	if (!(cl->parent->children.next == &cl->sibling &&
+		cl->parent->children.prev == &cl->sibling))
+		/* not the last child */
+		return 0;	
+
+	return 1;
+}
+
+static void htb_parent_to_leaf(struct htb_class *cl, struct Qdisc *new_q)
+{
+	struct htb_class *parent = cl->parent;
+
+	BUG_TRAP(!cl->level && cl->un.leaf.q && !cl->prio_activity);
+
+	parent->level = 0;
+	memset(&parent->un.inner, 0, sizeof(parent->un.inner));
+	INIT_LIST_HEAD(&parent->un.leaf.drop_list);
+	parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
+	parent->un.leaf.quantum = parent->quantum;
+	parent->un.leaf.prio = parent->prio;
+	parent->tokens = parent->buffer;
+	parent->ctokens = parent->cbuffer;
+	PSCHED_GET_TIME(parent->t_c);
+	parent->cmode = HTB_CAN_SEND;
+}
+
 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
 {
 	struct htb_sched *q = qdisc_priv(sch);
@@ -1328,6 +1364,8 @@ static int htb_delete(struct Qdisc *sch,
 	struct htb_sched *q = qdisc_priv(sch);
 	struct htb_class *cl = (struct htb_class *)arg;
 	unsigned int qlen;
+	struct Qdisc *new_q = NULL;
+	int last_child = 0;
 
 	// TODO: why don't allow to delete subtree ? references ? does
 	// tc subsys quarantee us that in htb_destroy it holds no class
@@ -1335,6 +1373,12 @@ static int htb_delete(struct Qdisc *sch,
 	if (!list_empty(&cl->children) || cl->filter_cnt)
 		return -EBUSY;
 
+	if (!cl->level && htb_parent_last_child(cl)) {
+		new_q = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
+						cl->parent->classid);
+		last_child = 1;
+	}
+
 	sch_tree_lock(sch);
 
 	/* delete from hash and active; remainder in destroy_class */
@@ -1349,6 +1393,9 @@ static int htb_delete(struct Qdisc *sch,
 	if (cl->prio_activity)
 		htb_deactivate(q, cl);
 
+	if (last_child)
+		htb_parent_to_leaf(cl, new_q);
+		
 	if (--cl->refcnt == 0)
 		htb_destroy_class(sch, cl);
 
@@ -1483,6 +1530,10 @@ static int htb_change_class(struct Qdisc
 			cl->un.leaf.quantum = hopt->quantum;
 		if ((cl->un.leaf.prio = hopt->prio) >= TC_HTB_NUMPRIO)
 			cl->un.leaf.prio = TC_HTB_NUMPRIO - 1;
+
+		/* backup for htb_parent_to_leaf */
+		cl->quantum = cl->un.leaf.quantum;
+		cl->prio = cl->un.leaf.prio;
 	}
 
 	cl->buffer = hopt->buffer;
#4 lore raw
Quoted 6 lines · Jarek Poplawski

[NET_SCHED] sch_htb:

[PATCH 2.6.19-rc6 with "Fix endless loops" set of patches]

- turn intermediate classes into leaves again when their last child is deleted (struct htb_class changed)

Looks good to me too, but it still seems to be missing class level adjustment after deletion. The classification function refuses to queue packets to classes with level > 0.

#5 lore raw
Quoted 11 lines · Patrick McHardy
Quoted 6 lines · Jarek Poplawski

[NET_SCHED] sch_htb:

[PATCH 2.6.19-rc6 with "Fix endless loops" set of patches]

- turn intermediate classes into leaves again when their last child is deleted (struct htb_class changed)

Looks good to me too, but it still seems to be missing class level adjustment after deletion. The classification function refuses to queue packets to classes with level > 0.

+static void htb_parent_to_leaf(struct htb_class *cl, struct Qdisc *new_q) +{
+ struct htb_class *parent = cl->parent;
+
+ BUG_TRAP(!cl->level && cl->un.leaf.q && !cl->prio_activity);
+
+ parent->level = 0;

I've thought this is enough, but probably you mean something else?

Jarek P.

#6 lore raw
Quoted 24 lines · Jarek Poplawski
Quoted 12 lines · Patrick McHardy
Quoted 6 lines · Jarek Poplawski

[NET_SCHED] sch_htb:

[PATCH 2.6.19-rc6 with "Fix endless loops" set of patches]

- turn intermediate classes into leaves again when their last child is deleted (struct htb_class changed)

Looks good to me too, but it still seems to be missing class level adjustment after deletion. The classification function refuses to queue packets to classes with level > 0.

net/sched/sch_htb.c +7 -0
+static void htb_parent_to_leaf(struct htb_class *cl, struct Qdisc *new_q)
+{
+	struct htb_class *parent = cl->parent;
+
+	BUG_TRAP(!cl->level && cl->un.leaf.q && !cl->prio_activity);
+
+	parent->level = 0;
Quoted 2 lines

I've thought this is enough, but probably you mean something else?

I missed that, thanks.

maintainer
#7 lore raw

From: Patrick McHardy <[email protected]>
Date: Thu, 30 Nov 2006 14:12:32 +0100

Quoted 25 lines
Quoted 24 lines · Jarek Poplawski
Quoted 12 lines · Patrick McHardy
Quoted 6 lines · Jarek Poplawski

[NET_SCHED] sch_htb:

[PATCH 2.6.19-rc6 with "Fix endless loops" set of patches]

- turn intermediate classes into leaves again when their last child is deleted (struct htb_class changed)

Looks good to me too, but it still seems to be missing class level adjustment after deletion. The classification function refuses to queue packets to classes with level > 0.

net/sched/sch_htb.c +7 -0
+static void htb_parent_to_leaf(struct htb_class *cl, struct Qdisc *new_q)
+{
+	struct htb_class *parent = cl->parent;
+
+	BUG_TRAP(!cl->level && cl->un.leaf.q && !cl->prio_activity);
+
+	parent->level = 0;
Quoted 4 lines
Quoted 2 lines

I've thought this is enough, but probably you mean something else?

I missed that, thanks.

Patch applied, thanks again.