[Fwd: [patch] fix slot clash in sch_sfq.c]

[Miguel Freitas <miguel@xxxxxxxxxxxxxxxx>]

I better send a copy of this one to the mailing list...

please cc me, i'm not subscribed.

regards,

Miguel

--- Begin Message ---

To:	kuznet@xxxxxxxxxxxxx
Subject:	[patch] fix slot clash in sch_sfq.c
From:	Miguel Freitas <miguel@xxxxxxxxxxxxxxxx>
Date:	Wed, 11 Feb 2004 12:13:45 -0200
Cc:	davem@xxxxxxxxxx

Hi Alexey, Hi David,

I've found a slot "leak" in sch_sfq.c that occurs with current 2.4 and
2.6 kernels. After dequeueing the last skb from a given slot, this slot
might be reused by a different hash (say new_hash). However the old hash
(q->ht[old_hash]) still points to the (now) reused slot. When new skb's
arrive with old_hash they will be enqueued at the same slot as new_hash
skb's (that is, q->ht[old_hash] == q->ht[new_hash]).

This produce a "slot clash" (opposed to a "hash clash"), severely
affecting the effectiveness of SFQ to provide fairness.

The attached one line patch fixes it, freeing the slot when last skb is
dequeued.

Note: I've noticed this problem while testing a fixed hash function of
the IP. the number of active slots didn't matched the number of active
IP clients. I have created a small standalone application, using
sch_sfq.c, to reproduce and debug this problem. I may provide this
application and a small data set (sequence of enqueues/dequeues) that
trigger the problem, let me know if you need it.

regards,

Miguel Freitas

sfq-slot-clash-fix.patch
Description: Text Data

--- End Message ---