Re: [PATCH]: PCI: GART iommu alignment fixes [v2]

[Joerg Roedel <joro@8bytes.org>]

On Thu, Jul 24, 2008 at 07:34:35PM +0900, FUJITA Tomonori wrote:
> On Thu, 24 Jul 2008 06:09:31 -0400
> Prarit Bhargava <prarit@redhat.com> wrote:
> >         skge->mem = pci_alloc_consistent(hw->pdev, skge->mem_size, 
> > &skge->dma);
> >         if (!skge->mem)
> >                 return -ENOMEM;
> > 
> >         BUG_ON(skge->dma & 7);
> > 
> >         if ((u64)skge->dma >> 32 != ((u64) skge->dma + skge->mem_size) 
> >  >> 32) {
> >                 printk(KERN_ERR PFX "pci_alloc_consistent region crosses 
> > 4G boundary\n");
> >                 err = -EINVAL;
> >                 goto free_pci_mem;
> >         }
> > 
> > 
> > If pci_alloc_consistent did the "right" thing, we should *never* see 
> > that warning message.
> 
> Well, I think that this is not releated with the pci_alloc_consistent
> alignment problem that you talk about.
> 
> I think that the driver tries to avoid 4GB boundary crossing
> problem. You can find some work to avoid this, for example:
> 
> http://www.ussg.iu.edu/hypermail/linux/kernel/0712.0/2206.html
> 
> pci_device_add() has the following code to avoid this:
> 
> pci_set_dma_seg_boundary(dev, 0xffffffff);
> 
> I suspect that the problem you talk about, alloc_consistent doesn't
> return the reqeuested size aligned memory, breaks anything.

But I think Prarit is right with this change. If the interface defines
this behavior the IOMMU drivers have to implement it. I am just
wondering that the problem never showed up before. The GART driver is a
few years old now.

Joerg