Eric Dumazet [off-list ref] wrote on 25/03/2009 15:04:26:

Joakim Tjernlund a =E9crit :

quoted

quoted

From 1c2f23b1f37f4818c0fd0217b93eb38ab6564840 Mon Sep 17 00:00:00 2001

From: Joakim Tjernlund <redacted>
Date: Tue, 24 Mar 2009 10:19:27 +0100
Subject: [PATCH] ucc=5Fgeth: Move freeing of TX packets to NAPI context.
 Also increase NAPI weight somewhat.
 This will make the system alot more responsive while
 ping flooding the ucc=5Fgeth ethernet interaface.
=20
=20
Signed-off-by: Joakim Tjernlund <redacted>
---
 drivers/net/ucc=5Fgeth.c |   30 +++++++++++-------------------
 drivers/net/ucc=5Fgeth.h |    2 +-
 2 files changed, 12 insertions(+), 20 deletions(-)
=20

diff --git a/drivers/net/ucc=5Fgeth.c b/drivers/net/ucc=5Fgeth.c
index 097aed8..7d5d110 100644
--- a/drivers/net/ucc=5Fgeth.c
+++ b/drivers/net/ucc=5Fgeth.c

@@ -3214,7 +3214,7 @@ static int ucc=5Fgeth=5Ftx(struct net=5Fdevice *d=

ev,=20
u8 txQ)

quoted

       dev->stats.tx=5Fpackets++;
=20
       /* Free the sk buffer associated with this TxBD */
-      dev=5Fkfree=5Fskb=5Firq(ugeth->
+      dev=5Fkfree=5Fskb(ugeth->
               tx=5Fskbuff[txQ][ugeth->skb=5Fdirtytx[txQ]]);
       ugeth->tx=5Fskbuff[txQ][ugeth->skb=5Fdirtytx[txQ]] =3D NULL;
       ugeth->skb=5Fdirtytx[txQ] =3D

@@ -3248,9 +3248,16 @@ static int ucc=5Fgeth=5Fpoll(struct napi=5Fstruc=

t=20
*napi, int budget)

quoted

    for (i =3D 0; i < ug=5Finfo->numQueuesRx; i++)
       howmany +=3D ucc=5Fgeth=5Frx(ugeth, i, budget - howmany);
=20

=20
Cant you test (ucce & UCCE=5FTX=5FEVENTS) or something here to avoid
taking lock and checking queues if not necessary ?

Probably, but I want this patch as simple as possible. There
are lots of optimizations left to do in this driver.

=20

quoted

+   /* Tx event processing */
+   spin=5Flock(&ugeth->lock);
+   for (i =3D 0; i < ug=5Finfo->numQueuesTx; i++) {
+      ucc=5Fgeth=5Ftx(ugeth->dev, i);
+   }
+   spin=5Funlock(&ugeth->lock);
+

=20
Why tx completions dont change "howmany" ?
It seems strange you changed UCC=5FGETH=5FDEV=5FWEIGHT if not taking into=

=20
account tx event above...

This is unclear and last I checked not very common amongst other drivers=20
in the
tree.

UCC=5FGETH=5FDEV=5FWEIGHT needs to be a bit bigger than the number of RX HW=
=20
buffers avaliable,
otherwise one won't be able to drain the whole queue in one go. Changing
weight to something bigger made a big difference.

 Jocke