On Wed, 2012-06-06 at 12:04 +0200, Jean-Michel Hautbois wrote:

Well, well, well, after having tested several configurations, several
drivers, I have a big difference between an old 2.6.26 kernel and a
newer one (I tried 3.2 and 3.4).

Here is my stream : UDP packets (multicast), 4000 bytes length, MTU
set to 4096. I am sending packets only, nothing on RX.
I send from 1Gbps upto 2.4Gbps and I see no drops in tc with 2.6.26
kernel, but a lot of drops with a newer kernel.
So, I don't know if I missed something in my kernel configuration, but
I have used the 2.6.26 one as a reference, in order to set the same
options (DMA related, etc).

I easily reproduce this problem and setting a bigger txqueuelen solves
it partially.
1Gbps requires a txqueulen of 9000, 2.4Gbps requires more than 20000 !

If you have any idea, I am interested, as this is a big issue for my use case.

Yep.

This driver wants to limit number of tx completions, thats just wrong.

Fix and dirty patch:

diff --git a/drivers/net/ethernet/emulex/benet/be.h b/drivers/net/ethernet/emulex/benet/be.h
index c5c4c0e..1e8f8a6 100644
--- a/drivers/net/ethernet/emulex/benet/be.h
+++ b/drivers/net/ethernet/emulex/benet/be.h

@@ -105,7 +105,7 @@ static inline char *nic_name(struct pci_dev *pdev)
 #define MAX_TX_QS		8
 #define MAX_ROCE_EQS		5
 #define MAX_MSIX_VECTORS	(MAX_RSS_QS + MAX_ROCE_EQS) /* RSS qs + RoCE */
-#define BE_TX_BUDGET		256
+#define BE_TX_BUDGET		65535
 #define BE_NAPI_WEIGHT		64
 #define MAX_RX_POST		BE_NAPI_WEIGHT /* Frags posted at a time */
 #define RX_FRAGS_REFILL_WM	(RX_Q_LEN - MAX_RX_POST)