Re: [PATCH 1/9] iommu/pages: Add support for a incoherent IOMMU page walker

[Jason Gunthorpe <jgg@nvidia.com>]

On Mon, Jul 21, 2025 at 04:41:03PM +0800, Baolu Lu wrote:
> On 7/17/2025 3:57 AM, Jason Gunthorpe wrote:
> > @@ -98,4 +101,36 @@ static inline void *iommu_alloc_pages_sz(gfp_t gfp, size_t size)
> >   	return iommu_alloc_pages_node_sz(NUMA_NO_NODE, gfp, size);
> >   }
> > -#endif	/* __IOMMU_PAGES_H */
> > +int iommu_pages_start_incoherent(void *virt, struct device *dma_dev);
> > +int iommu_pages_start_incoherent_list(struct iommu_pages_list *list,
> > +                                      struct device *dma_dev);
> > +
> > +#ifdef CONFIG_X86
> > +#include <asm/cacheflush.h>
> > +
> > +static inline void iommu_pages_flush_incoherent(struct device *dma_dev,
> > +						void *virt, size_t offset,
> > +						size_t len)
> > +{
> > +	clflush_cache_range(virt + offset, len);
> > +}
> > +static inline void
> > +iommu_pages_stop_incoherent_list(struct iommu_pages_list *list,
> > +				 struct device *dma_dev)
> > +{
> 
> Do we need to clear iopt->incoherent for X86, given that
> iopt->incoherent is set in the start path? 

No.. When I wrote this I felt we may as well keep x86 fast since it
doesn't use the DMA API and did it like this:

static void __iommu_free_desc(struct ioptdesc *iopt)
{
[..]
	if (!IS_ENABLED(CONFIG_X86))
		WARN_ON_ONCE(iopt->incoherent);

To largely ignore the incoherent flag.

So:

	/*
	 * For performance leave the incoherent flag alone which turns this into
	 * a NOP. For X86 the rest of the stop/free flow ignores the flag.
	 */

Thanks,
Jason