sis190
Quoted 2 lines
Message du 01/07/05 01:39
De : "Francois Romieu" <[email protected]>
[...]
Quoted 4 lines
There is an updated version at
http://www.zoreil.com/~romieu/sis190/20050630-2.6.13-rc1-sis190-test.patch
Quoted 8 lines
It would be nice to know how it behaves wrt preempt (no need to experiment with the media management), especially if you can describe the freeze more specifically.
--
Ueimor
François, it's really incredible!
A few lines diff, and now the driver is very stable with or without preempted kernel...
I'll be very happy if you can tell me where is the trick.
I tried it carefully : console, X11 (without nvidia), X11 (with nvidia), IRQ sharing between sis190/nvidia, full load : it worked perfectly.
# uname -a Linux local 2.6.12.1 #2 Thu Jun 23 11:05:57 CEST 2005 x86_64 x86_64 x86_64 GNU/Linux
# modinfo sis190 filename: /lib/modules/2.6.12.1/kernel/drivers/net/sis190.ko [...] vermagic: 2.6.12.1 preempt gcc-3.4
# cat /proc/interrupts
CPU0
0: 4403978 XT-PIC timer
1: 4594 XT-PIC i8042
2: 0 XT-PIC cascade
3: 1 XT-PIC ohci_hcd:usb3
5: 1 XT-PIC ohci_hcd:usb4
7: 7 XT-PIC ohci_hcd:usb2
8: 0 XT-PIC rtc
9: 0 XT-PIC acpi
10: 1 XT-PIC ehci_hcd:usb1
11: 18607531 XT-PIC SiS SI7012, nvidia, eth0
12: 139167 XT-PIC i8042
14: 196392 XT-PIC ide0
15: 39224 XT-PIC ide1
NMI: 2849
LOC: 4403221
ERR: 0
MIS: 0I executed simultaneously :
- a loop within 2 TX 700MB and 1 RX 700MB
- a loop within /usr/src/linux/make clean;make
- an audio stream
- an openglx demo
- my usual tasks (mozilla, xemacs)
# ifconfig
eth0 Link encap:Ethernet HWaddr 00:11:2F:E9:42:70
inet addr:10.169.21.21 Bcast:10.169.23.255 Mask:255.255.252.0
inet6 addr: fe80::211:2fff:fee9:4270/64 Scope:Link
UP BROADCAST RUNNING MULTICAST MTU:1500 Metric:1
RX packets:9922052 errors:0 dropped:0 overruns:0 frame:0
TX packets:17974399 errors:0 dropped:0 overruns:0 carrier:0
collisions:0 txqueuelen:1000
RX bytes:8064017712 (7.5 GiB) TX bytes:23410613404 (21.8 GiB)
Interrupt:11 Base address:0xdeadThe average speed of the transfers was about 10MB/s under full load without hurting too much CPU resource.
I retried the previous patch in the same conditions : it frooze the box after a few seconds, not only with X11 (no nvidia), but also with an audio stream (at the console).
I also retried (last patch) 10Mb/full autoneg off on the r8169 link partner (it prevously froze the box) : it worked perfectly.
The TX performances are also improved (r8169 LP):
- 10 full autoneg off : 1,2 MB both directions
- 100 full autoneg off : 11.6MB/s both directions
It still remains a problem in half duplex modes.
Perhaps somebody else want to try the patch with different link partners (switchs, 10Mb cards, ...)
I can not make other tries before Monday, but i'll take a look at the media management after...
BTW, can you remove the following printks from the patch ? The printks in interrupt functions make dmesg unusuable, and the stuff in sis190_get_drvinfo triggers a kernel oops when the module is loaded (null pointer assignment).
# diff -puN sis190-20050630.c sis190.c
--- sis190-20050630.c 2005-07-02 09:07:19.000000000 +0200
+++ sis190.c 2005-07-02 10:49:52.000000000 +0200
@@ -456,8 +456,6 @@ static inline int sis190_try_rx_copy(str
skb = dev_alloc_skb(pkt_size + NET_IP_ALIGN);
if (skb) {
skb_reserve(skb, NET_IP_ALIGN);
- printk(KERN_INFO "sk_buff[0]->tail = %p\n",
- sk_buff[0]->tail);
eth_copy_and_sum(skb, sk_buff[0]->tail, pkt_size, 0);
*sk_buff = skb;
sis190_give_to_asic(desc, rx_buf_sz);
@@ -483,15 +481,12 @@ static int sis190_rx_interrupt(struct ne
u32 status;rmb();
- printk(KERN_INFO "%s: Rx status = %08x\n", dev->name,
- desc->status);
if (desc->status & OWNbit)
break;
status = le32_to_cpu(desc->PSize);- printk(KERN_INFO "%s: Rx PSize = %08x\n", dev->name, status);
if (status & RxCRC) {
printk(KERN_INFO "%s: crc error. status = %08x\n",@@ -638,7 +633,6 @@ static irqreturn_t sis190_interrupt(int
SIS_W32(IntrStatus, status);
- printk(KERN_INFO "%s: status = %08x\n", dev->name, status);
if ((status & LinkChange) && netif_running(dev)) {
printk(KERN_INFO "%s: link change\n", dev->name);
@@ -1230,31 +1224,11 @@ static void sis190_get_drvinfo(struct ne
struct ethtool_drvinfo *info)
{
struct sis190_private *tp = netdev_priv(dev);- unsigned int i;
- u32 *u;
strcpy(info->driver, DRV_NAME);
strcpy(info->version, DRV_VERSION);
strcpy(info->bus_info, pci_name(tp->pci_dev));- printk(KERN_INFO "%s: dirty_rx=%ld cur_rx=%ld\n",
- dev->name, tp->dirty_rx, tp->cur_rx);
- u = (void *) tp->RxDescRing;
- printk(KERN_INFO " PSize status addr size PSize status addr size\n");
- for (i = 0; i < (NUM_RX_DESC / 2); i++) {- printk(KERN_INFO "%02d:%08x %08x %08x %08x %08x %08x %08x %08x\n",
- i, u[0], u[1], u[2], u[3], u[4], u[5], u[6], u[7]);
- u += 8;
- }
- printk(KERN_INFO "%s: dirty_tx=%ld cur_tx=%ld\n",
- dev->name, tp->dirty_tx, tp->cur_tx);
- u = (void *) tp->TxDescRing;
- printk(KERN_INFO " PSize status addr size PSize status addr size\n");
- for (i = 0; i < (NUM_TX_DESC / 2); i++) {- printk(KERN_INFO "%02d %08x %08x %08x %08x %08x %08x %08x %08x\n",
- i, u[0], u[1], u[2], u[3], u[4], u[5], u[6], u[7]);
- u += 8;
- }
}
static struct ethtool_ops sis190_ethtool_ops = {
Regards
Pascal
Pascal CHAPPERON <[email protected]> :
[...]
Quoted 4 lines
A few lines diff, and now the driver is very stable with or without preempted kernel...
I'll be very happy if you can tell me where is the trick.
Probably:
- when it filled the last Tx descriptor sis190_start_xmit() issued a netif_stop_queue and returned NETDEV_TX_BUSY;
- the asic completed DMAing the packet and acked it;
- sis190_tx_interrupt later released the descriptor and freed the skb;
-> since NETDEV_TX_BUSY assumes that the driver does not play with the skb,
one gets interesting results.
[...]
Quoted 2 lines
I tried it carefully : console, X11 (without nvidia), X11 (with nvidia), IRQ sharing between sis190/nvidia, full load : it worked perfectly.
Megateuf Wayne !
[...]
Quoted 4 lines
BTW, can you remove the following printks from the patch ? The printks in interrupt functions make dmesg unusuable, and the stuff in
sis190_get_drvinfotriggers a kernel oops when the module is loaded (null pointer assignment).
I'll polish the thing and sprinkle a few netif_msg_xxx() later today.
--
Ueimor
Pascal CHAPPERON <[email protected]> :
[...]
Quoted 7 lines
I can not make other tries before Monday, but i'll take a look at the media management after...
BTW, can you remove the following printks from the patch ? The printks in interrupt functions make dmesg unusuable, and the stuff in
sis190_get_drvinfotriggers a kernel oops when the module is loaded (null pointer assignment).
Done.
Can you check if there is a regression in sis190-000.patch available at
http://www.zoreil.com/~romieu/sis190/20050704-2.6.13-rc1/patches ?
If it works and you want some entertainment, you can apply sis190-010.patch
+ sis190-020.patch and experience with ethtool/mii-tool.
There is a tarball as well:
http://www.zoreil.com/~romieu/sis190/20050704-2.6.13-rc1.tar.bz2
--
Ueimor