RE: [ECOS] Re: Port of 82559 ethernet driver for PC target (was: "RE: [ECOS] New bie Question!")

RE: [ECOS] Re: Port of 82559 ethernet driver for PC target (was: "RE: [ECOS] New bie Question!")

[Fabrice Gautier]

Hi,

> -----Original Message-----
> From: Andrew Lunn [ mailto:andrew.lunn@ascom.ch ]
> Sent: Wednesday, October 04, 2000 12:34 AM
> To: Fabrice Gautier
> Cc: eCos Disuss
> Subject: [ECOS] Re: Port of 82559 ethernet driver for PC target (was:
> "RE: [ECOS] New bie Question!")
> 
> 
> Well, on the EBSA you have to tell the MMU not to use the read cache
> or the write buffer for the memory in the PCI window. I presume the
> same applies for the 386-PC?

The good news is  that the i386 port of eCos doesn't seem to have the cache
enabled.
 
> It has to be in the PCI window and somewhere the DMA can get it. I
> know in olden days the i386 hardware was limmited on where the DMA
> could work. Does this apply? Its a long time since i've played with
> this sort of thing.

The DMA? I'm not sure but I don't see any programming of a DMA controller in
the existing driver.
Unless there is part of the RAM that is not "connected" to the PCI bus I
guess all of it is accessible, but why would it be so? Does the DMA thing is
what connect the PCI bus to the RAM ?

Anyway I think the DMA limitation is something like we can't do DMA
transfert to/from memory above 16 Meg - And I don't have more than 16 Meg on
my target.

And Anyway, I will assume there is nothing to worry about... and we will
see.

Regards,

-- 
Fabrice Gautier
fabrice_gautier@sdesigns.com 

Re: [ECOS] Re: Port of 82559 ethernet driver for PC target (was: "RE: [ECOS] New bie Question!")

[Andrew Lunn]

> The good news is  that the i386 port of eCos doesn't seem to have the cache
> enabled.

!?! That will be killing performance realy badly. I suggest someone
fixes this.

> The DMA? I'm not sure but I don't see any programming of a DMA controller in
> the existing driver.

You don't explicitly program the DMA, but the ethernet card has a DMA
in it. It DMAs the frame into the DRAM. 

> And Anyway, I will assume there is nothing to worry about... and we will
> see.

You are probably right.

        Andrew

RE: [ECOS] Re: Port of 82559 ethernet driver for PC target (was: "RE: [ECOS] New bie Question!")

[Fabrice Gautier]

> -----Original Message-----
> From: Andrew Lunn [ mailto:andrew.lunn@ascom.ch ]
> Sent: Thursday, October 05, 2000 12:48 AM
> To: Fabrice Gautier
> Cc: eCos Disuss
> Subject: Re: [ECOS] Re: Port of 82559 ethernet driver for PC target
> (was: "RE: [ECOS] New bie Question!")
> 
> 
> > The good news is  that the i386 port of eCos doesn't seem 
> to have the cache
> > enabled.
> 
> !?! That will be killing performance realy badly. I suggest someone
> fixes this.

It may kill performances, but anyway I think many eCos users use the PC
platform as an evaluation platform so performance is not really an issue.
But if someone wants to use eCos and a 386 as a final target I suggest they
eventually pay someone to work seriously on the PC platform.


A+
-- 
Fabrice Gautier
fabrice_gautier@sdesigns.com 

[ECOS] Re: Port of 82559 ethernet driver for PC target (was: "RE: [ECOS] New bie Question!")

[Andrew Lunn]

> > Decide where to put the PCI memory window. Change the MMU to support
> > this. Also watch out for the new heap/malloc code. It may try to put a
> > heap on the PCI window. 
> 
> Well, Afaik there is no MMU involved in the 386-PC target. It uses a simple
> 32 bits flat model.
> 
> So I don't think there is no need to setup a PCI memory window. I think the
> 3 BAR are setup by the BIOS. I read:

Well, on the EBSA you have to tell the MMU not to use the read cache
or the write buffer for the memory in the PCI window. I presume the
same applies for the 386-PC?
 
> And for the heap/malloc things, I don't understand why there is a special
> malloc stuff. Why not regular memory pools?. And it seems there is need to
> set up a region in the memory layout. Why?

You need some memory for receive and transmit bufferes. These go into
the PCI window. On the EBSA you can only fiddle the cache for 1Mbyte
pages so 1Mbytes has been allocated to the PCI window. This is in real
RAM. I think to stop the new heap code trying to use this piece of
memory a section is placed on top of it. I've not looked at the
details, i just know it works somehow. Maybe Hugo or Jifl can
clarrify?
 
> As I understand how the 82559 works, when we want to send a packet, we have
> to write the adress of some structure in the SCB General Pointer register
> (offset 04h) and then a transmit command in the SCB command register. Is
> that right? If so, the adress we pass can just be anywhere in memory right?

It has to be in the PCI window and somewhere the DMA can get it. I
know in olden days the i386 hardware was limmited on where the DMA
could work. Does this apply? Its a long time since i've played with
this sort of thing.

        Andrew