Re: [PATCH] ALSA update [6/10] - 2002/07/20

[Jaroslav Kysela]

On 30 Sep 2002, Alan Cox wrote:

> On Mon, 2002-09-30 at 01:53, David S. Miller wrote:
> >    From: Jaroslav Kysela <[e-mail]>
> >    Date: Sun, 29 Sep 2002 22:34:51 +0200 (CEST)
> >     
> >    -	if (hwdev == NULL || ((u32)hwdev->dma_mask != 0xffffffff))
> >    +	if (hwdev == NULL || (u32)hwdev->dma_mask <= 0x00ffffff)
> >     		gfp |= GFP_DMA;
> > 
> > So alan, why is this really broken?
> > 
> > EISA/ISA DMA is defined as using a hwdev of NULL or requiring
> > <16MB address, he is preserving GFP_DMA in those cases.
> 
> Firstly the DMA mask on x86 can't be below 24bits, we don't support
> allocation from a smaller zone. Secondly what about PCI for 25-31bits -
> there we do need to force gfp_dma to have any chance of getting the
> right pages

Not really. The test in original code - if the page is in right area - was 
bellow first allocation. Sure, it's dirty hack.

> Giving the page allocator a mask argument does sound a lot nicer

Right, my aim was to point to a bug. Let's go to create proper DMA 
allocation for broken hardware. I still wonder why hardware vendors are 
lazy to support full bus addressing. Perhaps, saving few coins per 
chip make them happy.

						Jaroslav

-----
Jaroslav Kysela <[e-mail]>
Linux Kernel Sound Maintainer
ALSA Project  http://www.alsa-project.org
SuSE Linux    http://www.suse.com