Hi Grant,

Le Sunday 31 May 2009 03:40:36 Grant Grundler, vous avez écrit :

Florian,
Summary: proposed patch below for you to test (I've not tested it yet)
and some explanation on what I think is happening.


On Fri, May 01, 2009 at 12:38:57PM +0200, Florian Fainelli wrote:
...

quoted

Linux Tulip driver version 1.1.15-NAPI (Feb 27, 2007)

...

quoted

PCI: Enabling device 0000:00:0c.0 (0005 -> 0007)
tulip1: Old format EEPROM on 'Cobalt Microserver' board.  Using
substitute media control info.
tulip1:  EEPROM default media type Autosense.
tulip1:  Index #0 - Media MII (#11) described by a 21142 MII PHY (3)
block. tulip1:  MII transceiver #1 config 1000 status 7809 advertising
01e1. eth1: Digital DS21142/43 Tulip rev 65 at MMIO 0x12082400,
00:10:e0:00:88:b9, IRQ 20.
[snip]
0000:00:07.0: tulip_stop_rxtx() failed (CSR5 0xf0660000 CSR6 0xb20e2202)

Looking up these bits in the publicly available manual:
    ftp://download.intel.com/design/network/manuals/27807401.pdf

"Rev 65" == 0x41 == 21143-PD or 21143-TD (page 3-7)

Operation Mode Register (CSR6­Offset 30H)
----
bit : Val
31  : 1 Special Capture Effect Enable
30  : 1 Receive All
13  : 1 ST - Start Transmission
9   : 1 Full Duplex
1   : 1 SR - Start Receive

Status Register (CSR5­Offset 28H)
----
28-31 : reserved
23-25 : 0x6 TX State = Suspended--Transmit FIFO underflow, or an
unavailable transmit descriptor 20-22 : 0x3 RX State = Running--Waiting for
receive packet
1     : 0x0 TX isn't stopped
0     : 0x0 No TX Interrupt pending

The RX/TX engines are in a wedged state to begin with. :(

I suppose this is due to the Bootloader, either CoLo or the original Cobalt 
microservers bootloader.

The normal calling path here is tulip_init() to register the driver
callbacks and tulip_init_one() gets called immediately by pci subsystem.
tulip_up() gets called after netdev registration when someone ifconfig's
the device (ifconfig up). I don't know when else tulip_up() is called.

I have two ideas on how to fix this:
1) reset the RX/TX engines in tulip_init_one() before tulip_stop_rxtx().
2) reset the RX/TX engines in tulip_stop_rxtx() if they are "wedged".
3) remove tulip_stop_rxtx() call in tulip_init_one()


(1) seems like a reasonable thing to do at init time anyway.
(2) feels like a pretty big hammer and I don't know all the side effects.
(3) tulip_up() will reset the RX/TX engine and call tulip_stop_rxtx()
   when the NIC is opene/ifconfig'd. Calling pci_set_master()
   will allow the device to scribble into Host Memory...probably
   should move that call to tulip_up() as well. But I don't see anything
   in tulip_init_one() that requires DMA (and thus pci_set_master()).

Is there any reason for NOT implementing (1) and (3)?

I still need to work out where to call pci_set_master() in
tulip_up(). If that's feasible,

hrm. WTF. tulip_stop_rxtx() is called again from tulip_up() right before
CSR6 is written. Earlier in tulip_up() we reset the chip.  Removed.

Can you test the patch below?

I have not tested or even compiled this patch...will do so on parisc/ia64
machines once I get some feedback on this patch.

And I just noticed pci_clear_master() is not called *anywhere*. :(
Need to add such a call after tulip_stop_rxtx() some place (many places?).
This patch is just RFC and not suitable for merging upstream.

The patch below does not help on my Qube2, I am still having the same message 
appearing.

-- 
Best regards, Florian Fainelli
Email : florian@openwrt.org
http://openwrt.org
-------------------------------