Re: [PATCH] drm: Drop HPAGE_PMD_SIZE dependency in dma_iova_try_alloc calls

[Matthew Brost <matthew.brost@intel.com>]

On Thu, May 07, 2026 at 11:16:06AM +0200, Francois Dugast wrote:
> The phys argument to dma_iova_try_alloc() is used only to compute
> the sub-granule offset (phys & (granule - 1)). Since HPAGE_PMD_SIZE
> is a power of two larger than any IOMMU granule, this expression
> always evaluates to zero.

I looks like I screwed up how this interface is supposed to be used. I
thought the phys argument to dma_iova_try_alloc was really the
alignment, but it looks I was mistaken. I believe the above analysis is
correct.

> 
> Replace the ternary expressions with a plain 0, which is what the
> API documentation recommends for callers doing PAGE_SIZE-aligned
> transfers. This also removes the dependency on HPAGE_PMD_SIZE and
> thus on CONFIG_PGTABLE_HAS_HUGE_LEAVES / HAVE_ARCH_TRANSPARENT_HUGEPAGE,
> fixing build failures on architectures such as arm32 that lack that
> config.
> 

Fixes tag so this gets in drm-misc-fixes cycle?

> Signed-off-by: Francois Dugast <francois.dugast@intel.com>
> Assisted-by: GitHub Copilot:claude-sonnet-4.6 # Documentation
> Link: https://lore.kernel.org/intel-xe/c36e7dfb-cf62-4d21-a3b1-f54cb43e0832@infradead.org/
> ---
>  drivers/gpu/drm/drm_gpusvm.c  | 5 +----
>  drivers/gpu/drm/drm_pagemap.c | 5 +----
>  2 files changed, 2 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/gpu/drm/drm_gpusvm.c b/drivers/gpu/drm/drm_gpusvm.c
> index 365a9c0b522a..958cb605aedd 100644
> --- a/drivers/gpu/drm/drm_gpusvm.c
> +++ b/drivers/gpu/drm/drm_gpusvm.c
> @@ -1556,10 +1556,7 @@ int drm_gpusvm_get_pages(struct drm_gpusvm *gpusvm,
>  
>  			if (!i)
>  				dma_iova_try_alloc(gpusvm->drm->dev, state,
> -						   npages * PAGE_SIZE >=
> -						   HPAGE_PMD_SIZE ?
> -						   HPAGE_PMD_SIZE : 0,
> -						   npages * PAGE_SIZE);
> +						   0, npages * PAGE_SIZE);

One thing to do would be to confirm via a printk that, when npages ==
512 on Xe, the IOVA is 2M-aligned. If not, we likely need to tweak the
dma_iova_try_alloc interface—but that would be out of scope here, I
think. From my quick read of dma_iova_try_alloc, npages == 512 should
produce a 2M-aligned IOVA.

Anyways with a Fixes tag:
Reviewed-by: Matthew Brost <matthew.brost@intel.com>

>  
>  			if (dma_use_iova(state)) {
>  				err = dma_iova_link(gpusvm->drm->dev, state,
> diff --git a/drivers/gpu/drm/drm_pagemap.c b/drivers/gpu/drm/drm_pagemap.c
> index d82ea7ccb8da..15c78eca180b 100644
> --- a/drivers/gpu/drm/drm_pagemap.c
> +++ b/drivers/gpu/drm/drm_pagemap.c
> @@ -347,10 +347,7 @@ drm_pagemap_migrate_map_system_pages(struct device *dev,
>  
>  		if (!try_alloc) {
>  			dma_iova_try_alloc(dev, &state->dma_state,
> -					   (npages - i) * PAGE_SIZE >=
> -					   HPAGE_PMD_SIZE ?
> -					   HPAGE_PMD_SIZE : 0,
> -					   npages * PAGE_SIZE);
> +					   0, npages * PAGE_SIZE);
>  			try_alloc = true;
>  		}
>  
> -- 
> 2.43.0
>