[PATCH] 8139too Interframe Gap Time

[PATCH] 8139too Interframe Gap Time

[Rene Herman]

[-- Attachment #1: Type: text/plain, Size: 834 bytes --]

Hi Jeff,

in drivers/net/8139too.c, we have:

    /* Check this value: the documentation for IFG contradicts ifself. */
    RTL_W32 (TxConfig, rtl8139_tx_config);

I see that in older versions of the documentation there indeed were some
contradictions but the current version 1.4:

ftp://210.51.181.211/cn/nic/rtl8139abcd8130810xseries/rtl8139cspec_1.4.pdf

clears it up and it seems 8139too guessed wrong. I have verified that
realtek's own windows 9x driver agrees with the documentation.

No equipment to actually measure the IFG and no difference in network
behaviour shows up on a good 100Mbit-FD switch connected LAN, but it
does on an el-cheapo 10Mbit-HD hub connected LAN. Without attached
patch the packet and collision counters are basically the same; with
patch, the latter is only some 70% of the former. Hurray!

Rene.



[-- Attachment #2: linux-2.6.4-rc2-mm1_8139too-IFG.diff --]
[-- Type: text/plain, Size: 1494 bytes --]

--- linux-2.6.4-rc2-mm1/drivers/net/8139too.c.orig	2004-03-10 21:01:08.000000000 +0100
+++ linux-2.6.4-rc2-mm1/drivers/net/8139too.c	2004-03-10 21:01:14.000000000 +0100
@@ -383,8 +383,14 @@
 
 /* Bits in TxConfig. */
 enum tx_config_bits {
-	TxIFG1 = (1 << 25),	/* Interframe Gap Time */
-	TxIFG0 = (1 << 24),	/* Enabling these bits violates IEEE 802.3 */
+
+        /* Interframe Gap Time. Only TxIFG96 doesn't violate IEEE 802.3 */
+        TxIFGShift = 24,
+        TxIFG84 = (0 << TxIFGShift),    /* 8.4us / 840ns (10 / 100Mbps) */
+        TxIFG88 = (1 << TxIFGShift),    /* 8.8us / 880ns (10 / 100Mbps) */
+        TxIFG92 = (2 << TxIFGShift),    /* 9.2us / 920ns (10 / 100Mbps) */
+        TxIFG96 = (3 << TxIFGShift),    /* 9.6us / 960ns (10 / 100Mbps) */
+
 	TxLoopBack = (1 << 18) | (1 << 17), /* enable loopback test mode */
 	TxCRC = (1 << 16),	/* DISABLE appending CRC to end of Tx packets */
 	TxClearAbt = (1 << 0),	/* Clear abort (WO) */
@@ -716,7 +722,7 @@
 #endif
 
 static const unsigned int rtl8139_tx_config =
-	(TX_DMA_BURST << TxDMAShift) | (TX_RETRY << TxRetryShift);
+	TxIFG96 | (TX_DMA_BURST << TxDMAShift) | (TX_RETRY << TxRetryShift);
 
 static void __rtl8139_cleanup_dev (struct net_device *dev)
 {
@@ -1393,8 +1399,6 @@
 
 	tp->rx_config = rtl8139_rx_config | AcceptBroadcast | AcceptMyPhys;
 	RTL_W32 (RxConfig, tp->rx_config);
-
-	/* Check this value: the documentation for IFG contradicts ifself. */
 	RTL_W32 (TxConfig, rtl8139_tx_config);
 
 	tp->cur_rx = 0;

Re: [PATCH] 8139too Interframe Gap Time

[Jeff Garzik]

Rene Herman wrote:
> Hi Jeff,
> 
> in drivers/net/8139too.c, we have:
> 
>    /* Check this value: the documentation for IFG contradicts ifself. */
>    RTL_W32 (TxConfig, rtl8139_tx_config);
> 
> I see that in older versions of the documentation there indeed were some
> contradictions but the current version 1.4:
> 
> ftp://210.51.181.211/cn/nic/rtl8139abcd8130810xseries/rtl8139cspec_1.4.pdf
> 
> clears it up and it seems 8139too guessed wrong. I have verified that
> realtek's own windows 9x driver agrees with the documentation.
> 
> No equipment to actually measure the IFG and no difference in network
> behaviour shows up on a good 100Mbit-FD switch connected LAN, but it
> does on an el-cheapo 10Mbit-HD hub connected LAN. Without attached
> patch the packet and collision counters are basically the same; with
> patch, the latter is only some 70% of the former. Hurray!


Nice!

I'm glad somebody finally got around to doing this :)

I'll give this some testing (and put it in Andrew's -mm for that reason 
as well), and then apply it.

Thanks,

	Jeff