[RFC PATCH 3/5] mm/vma: add support for peer to peer to device vma

Logan Gunthorpe logang at deltatee.com
Wed Jan 30 22:52:13 UTC 2019



On 2019-01-30 2:50 p.m., Jason Gunthorpe wrote:
> On Wed, Jan 30, 2019 at 02:01:35PM -0700, Logan Gunthorpe wrote:
> 
>> And I feel the GUP->SGL->DMA flow should still be what we are aiming
>> for. Even if we need a special GUP for special pages, and a special DMA
>> map; and the SGL still has to be homogenous....
> 
> *shrug* so what if the special GUP called a VMA op instead of
> traversing the VMA PTEs today? Why does it really matter? It could
> easily change to a struct page flow tomorrow..

Well it's so that it's composable. We want the SGL->DMA side to work for
APIs from kernel space and not have to run a completely different flow
for kernel drivers than from userspace memory.

For GUP to do a special VMA traversal it would now need to return
something besides struct pages which means no SGL and it means a
completely different DMA mapping call.
> Would you feel better if this also came along with a:
> 
>   struct dma_sg_table *sgl_dma_map_user(struct device *dma_device, 
>              void __user *prt, size_t len)

That seems like a nice API. But certainly the implementation would need
to use existing dma_map or pci_p2pdma_map calls, or whatever as part of
it...

,
> flow which returns a *DMA MAPPED* sgl that does not have struct page
> pointers as another interface?
> 
> We can certainly call an API like this from RDMA for non-ODP MRs.
> 
> Eliminating the page pointers also eliminates the __iomem
> problem. However this sgl object is not copyable or accessible from
> the CPU, so the caller must be sure it doesn't need CPU access when
> using this API. 

We actually stopped caring about the __iomem problem. We are working
under the assumption that pages returned by devm_memremap_pages() can be
accessed as normal RAM and does not need the __iomem designation. The
main problem now is that code paths need to know to use pci_p2pdma_map
or not. And in theory this could be pushed into regular dma_map
implementations but we'd have to get it into all of them which is a pain.

Logan


More information about the dri-devel mailing list