Re: [PATCH v3 2/3] s390/pci: store DMA offset in bus_dma_region

[Niklas Schnelle]

On Fri, 2025-01-24 at 15:17 -0500, Matthew Rosato wrote:
> PCI devices on s390 have a DMA offset that is reported via CLP.  In
> preparation for allowing identity domains, setup the bus_dma_region
> for all PCI devices using the reported CLP value.
> 
> Signed-off-by: Matthew Rosato <mjrosato@xxxxxxxxxxxxx>
> ---
>  arch/s390/pci/pci_bus.c | 18 ++++++++++++++++++
>  1 file changed, 18 insertions(+)
> 
> diff --git a/arch/s390/pci/pci_bus.c b/arch/s390/pci/pci_bus.c
> index d5ace00d10f0..51fa993b64fc 100644
> --- a/arch/s390/pci/pci_bus.c
> +++ b/arch/s390/pci/pci_bus.c
> @@ -19,6 +19,7 @@
>  #include <linux/jump_label.h>
>  #include <linux/pci.h>
>  #include <linux/printk.h>
> +#include <linux/dma-direct.h>
>  
>  #include <asm/pci_clp.h>
>  #include <asm/pci_dma.h>
> @@ -284,10 +285,27 @@ static struct zpci_bus *zpci_bus_alloc(int topo, bool topo_is_tid)
>  	return zbus;
>  }
>  
> +static void pci_dma_range_setup(struct pci_dev *pdev)
> +{
> +	struct zpci_dev *zdev = to_zpci(pdev);
> +	struct bus_dma_region *map;
> +
> +	map = kzalloc(sizeof(*map), GFP_KERNEL);
> +	if (!map)
> +		return;
> +
> +	map->cpu_start = 0;
> +	map->dma_start = PAGE_ALIGN(zdev->start_dma);
> +	map->size = zdev->end_dma - zdev->start_dma + 1;

Shouldn't the size calculation also use PAGE_ALIGN(zdev->start_dma) in
case that actually aligns? And for zdev->end_dma I think we want
PAGE_ALIGN_DOWN(zdev->end_dma + 1). There is also the, hopefully
theoretical, case that zdev->end_dma could be less than PAGE_SIZE
larger than zdev->start_dma. I guess in that case we should just ensure
that size is 0, maybe using max(…, 0) around the expression.

> +	pdev->dev.dma_range_map = map;
> +}
> +
>  void pcibios_bus_add_device(struct pci_dev *pdev)
>  {
>  	struct zpci_dev *zdev = to_zpci(pdev);
>  
> +	pci_dma_range_setup(pdev);
> +
>  	/*
>  	 * With pdev->no_vf_scan the common PCI probing code does not
>  	 * perform PF/VF linking.

I was initially wondering where the map is freed, then found the
kfree()s in common code in device_unbind_cleanup() and
device_release().

Despite the thoughts above I already tested this on my setup and it
worked great, so you can already add my:

Tested-by: Niklas Schnelle <schnelle@xxxxxxxxxxxxx>