[PATCH] net/sched: drr: reseed active class deficit after quantum changes

[PATCH] net/sched: drr: reseed active class deficit after quantum changes

[Samuel Moelius]

Changing the quantum of an active DRR class leaves the old deficit in
place.  The next scheduling round can therefore use credit accumulated
under a different quantum.

This can be observed by making a class active, changing its quantum, and
then dequeuing with the old deficit still present.

When an active class quantum changes, reseed its deficit from the new
quantum so the changed class weight is reflected immediately.

Assisted-by: Codex:gpt-5.5-cyber-preview
Signed-off-by: Samuel Moelius <sam.moelius@trailofbits.com>
---
 net/sched/sch_drr.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/net/sched/sch_drr.c b/net/sched/sch_drr.c
index 01335a49e091..b40df8725ba6 100644
--- a/net/sched/sch_drr.c
+++ b/net/sched/sch_drr.c
@@ -98,8 +98,11 @@ static int drr_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
 		}
 
 		sch_tree_lock(sch);
-		if (tb[TCA_DRR_QUANTUM])
+		if (tb[TCA_DRR_QUANTUM]) {
 			cl->quantum = quantum;
+			if (cl_is_active(cl))
+				cl->deficit = quantum;
+		}
 		sch_tree_unlock(sch);
 
 		return 0;
-- 
2.43.0

Re: [PATCH] net/sched: drr: reseed active class deficit after quantum changes

[Jakub Kicinski]

On Tue,  9 Jun 2026 00:36:18 +0000 Samuel Moelius wrote:
> Subject: [PATCH] net/sched: drr: reseed active class deficit after quantum changes

If the change is not for a serious bug it needs to be generated against
net-next. This was generated against Linus's tree I guess and doesn't
apply to -next.

> Changing the quantum of an active DRR class leaves the old deficit in
> place.  The next scheduling round can therefore use credit accumulated
> under a different quantum.
> 
> This can be observed by making a class active, changing its quantum, and
> then dequeuing with the old deficit still present.
> 
> When an active class quantum changes, reseed its deficit from the new
> quantum so the changed class weight is reflected immediately.

TBH the current implementation is how I would expect DRR to work.
quantum is the "refill" value, it should not reset the state of 
the current round? It wouldn't in an ASIC.
-- 
pw-bot: cr