RE: cciss patches for 2.4.21-rc1, 4 of 4

RE: cciss patches for 2.4.21-rc1, 4 of 4

[Cameron, Steve]

Arjan van de Ven wrote:
> Cameron, lots of people told you numerous times that pci_alloc_consitent
> is guaranteed to return 32 bit addresses by the API, see
> Documentation/DMA-mapping.txt for the API definition.

Ok.  You won't have to remind me again, since, as it happens,
today is officially my last day at HP doing anything in the linux kernel.
(I'm still at HP, just doing something else now.)

Thanks,

-- steve

RE: cciss patches for 2.4.21-rc1, 4 of 4

[Arjan van de Ven]

[-- Attachment #1: Type: text/plain, Size: 1158 bytes --]

On Fri, 2003-04-25 at 15:48, Miller, Mike (OS Dev) wrote:
> I haven't seen any issues (yet) on ia64. I'm running with 5GB RAM.
> 
> mikem
> 
> -----Original Message-----
> From: Cameron, Steve 
> Sent: Friday, April 25, 2003 8:25 AM
> Cc: linux-kernel@vger.kernel.org; Miller, Mike (OS Dev)
> Subject: RE: cciss patches for 2.4.21-rc1, 4 of 4
> 
> 
> 
> Mike Miller wrote:
> 
> > Changes:
> >	1. Sets the DMA mask to 64 bits. Removes RH's code for the DMA mask.
> 
> In order for this to work, it depends on pci_alloc_consistent always
> returning memory with physical addresses that fit in 32 bits, 
> regardless of the DMA mask, since the cciss device's command register 
> is 32 bits, and the command buffer addresses must fit in there.  If 
> that's the case, this is fine.  Otherwise, this may fail if pci_alloc_consistent
> returns memory above 4GB.  (on x86, I think this is not a problem, not
> sure of other archs, e.g. alpha, ia64)

Cameron, lots of people told you numerous times that pci_alloc_consitent
is guaranteed to return 32 bit addresses by the API, see
Documentation/DMA-mapping.txt for the API definition.


[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

RE: cciss patches for 2.4.21-rc1, 4 of 4

[Miller, Mike (OS Dev)]

I haven't seen any issues (yet) on ia64. I'm running with 5GB RAM.

mikem

-----Original Message-----
From: Cameron, Steve 
Sent: Friday, April 25, 2003 8:25 AM
Cc: linux-kernel@vger.kernel.org; Miller, Mike (OS Dev)
Subject: RE: cciss patches for 2.4.21-rc1, 4 of 4



Mike Miller wrote:

> Changes:
>	1. Sets the DMA mask to 64 bits. Removes RH's code for the DMA mask.

In order for this to work, it depends on pci_alloc_consistent always
returning memory with physical addresses that fit in 32 bits, 
regardless of the DMA mask, since the cciss device's command register 
is 32 bits, and the command buffer addresses must fit in there.  If 
that's the case, this is fine.  Otherwise, this may fail if pci_alloc_consistent
returns memory above 4GB.  (on x86, I think this is not a problem, not
sure of other archs, e.g. alpha, ia64)

Note the cciss devices *are* capable of 64 bit addressing for DMA, provided
you can find a way to tell it a 64 bit address.  For data, it's no problem.
Only for command buffers is it a problem.

-- steve

RE: cciss patches for 2.4.21-rc1, 4 of 4

[Alan Cox]

On Gwe, 2003-04-25 at 14:48, Miller, Mike (OS Dev) wrote:
> I haven't seen any issues (yet) on ia64. I'm running with 5GB RAM.

That doesn't make it correct. This same problem occurs in other drivers
and the usual trick is to set the pci mask to 32bit, allocate those
command buffers ready, then flip back to 64bit. Just be sure one thread
doesn't change it to 64bit while another is allocating commands. The
reverse is fine since an accidental odd bounce is no big deal