Re: [E1000-devel] Lost Ticks and ICH8

[Jason G. <jgu...@ob...>]

On Fri, Oct 20, 2006 at 02:26:43PM -0700, Brandeburg, Jesse wrote:

> any luck?  I am expecting at the very least the BIOS changed behavior.

Ok, I have upgraded to the latest BIOS (since I last posted yet
another one has been released..) dmidecode says the version is now
CO96510J.86A.5434.2006.1016.1710

With this bios I now have HPET working:
 ACPI: HPET id: 0x8086a201 base:0xfed00000
 Using 14.318180 MHz WALL HPET GTOD HPET/TSC timer.

This configuration still gives lost tick messages similar to before
when the port is in gigabit mode.

ethtool behaves the same way as before.

I turned on CONFIG_PRINTK_TIME and added printks to trace
the problem:

[  143.106472] in e1000_update_stats 3179
[  143.106501] e1000_update_stats 3255
[  143.106509] e1000_update_stats 3308
[  143.106511] e1000_update_stats 3321
[  143.106512] media_type 0 link_speed 1000
[  143.106556] in e1000_read_phy_reg 3363
[  143.106557] in e1000_swfw_sync_acquire 3271 0x2
[  143.106559] out e1000_swfw_sync_acquire 3275
[  143.206750] out e1000_get_software_flag 8385 extcnf_ctrl=0x20000a
[  143.206752] out e1000_read_phy_reg 3372
[  143.206754] e1000_update_stats 3336
[  143.206760] time.c: Lost 24 timer tick(s)! rip _spin_unlock_irqrestore+0x8/0x10)
[..]
[  143.208397] out e1000_update_stats 3338

I've included the patch I am using so you can correlate the line
numbers.

The upshot is that the e1000_get_software_flag is timing out. That is
called as part of e1000_read_phy_reg that is executed in the tail end of
the update_stats function, while holding a spin lock.

Thanks,
Jason