On 2019-07-25 1:29 p.m., Jason Gunthorpe wrote:
> On Thu, Jul 25, 2019 at 01:17:02PM -0600, Logan Gunthorpe wrote:
>>
>>
>> On 2019-07-25 12:58 p.m., Jason Gunthorpe wrote:
>>> On Mon, Jul 22, 2019 at 05:08:56PM -0600, Logan Gunthorpe wrote:
>>>> Any requests that traverse the host bridge will need to be mapped into
>>>> the IOMMU, so call dma_map_sg() inside pci_p2pdma_map_sg() when
>>>> appropriate.
>>>>
>>>> Similarly, call dma_unmap_sg() inside pci_p2pdma_unmap_sg().
>>>>
>>>> Signed-off-by: Logan Gunthorpe <logang@xxxxxxxxxxxx>
>>>> drivers/pci/p2pdma.c | 31 ++++++++++++++++++++++++++++++-
>>>> 1 file changed, 30 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
>>>> index 5f43f92f9336..76f51678342c 100644
>>>> +++ b/drivers/pci/p2pdma.c
>>>> @@ -830,8 +830,22 @@ int pci_p2pdma_map_sg_attrs(struct device *dev, struct scatterlist *sg,
>>>> int nents, enum dma_data_direction dir, unsigned long attrs)
>>>> {
>>>> struct dev_pagemap *pgmap = sg_page(sg)->pgmap;
>>>> + struct pci_dev *client;
>>>> + int dist;
>>>> +
>>>> + client = find_parent_pci_dev(dev);
>>>> + if (WARN_ON_ONCE(!client))
>>>> + return 0;
>>>>
>>>> - return __pci_p2pdma_map_sg(pgmap, dev, sg, nents);
>>>> + dist = upstream_bridge_distance(pgmap->pci_p2pdma_provider,
>>>> + client, NULL);
>
> Isn't is a bit of a leap to assume that the pgmap is uniform across
> all the sgs?
This is definitely a wart but there's not much we can do about it.
Currently we can't support mixing p2p pages with regular pages. In the
same way we can't support mixing p2p pages from different sources. No
current users do that and it would be weird for them to want to at this
point.
>>>> + if (WARN_ON_ONCE(dist & P2PDMA_NOT_SUPPORTED))
>>>> + return 0;
>>>> +
>>>> + if (dist & P2PDMA_THRU_HOST_BRIDGE)
>>>> + return dma_map_sg_attrs(dev, sg, nents, dir, attrs);
>>>
>>> IIRC at this point the SG will have struct page references to the BAR
>>> memory - so (all?) the IOMMU drivers are able to handle P2P setup in
>>> this case?
>>
>> Yes. The IOMMU drivers refer to the physical address for BAR which they
>> can get from the struct page. And this works fine today.
>
> Interesting.
>
> So, for the places where we already map BAR memory to userspace, if I
> were to make struct pages for those BARs and use vm_insert_page()
> instead of io_remap_pfn_range(), then the main thing missing in RDMA
> to actually do P2P DMA is a way to get those struct pages out of
> get_user_pages and know to call the pci_p2pdma_map_sg version (ie in
> ib_umem_get())?
Yes, we've been doing that for a long time with hacky code that would
never get upstream.
Essentially, if you expose those pages to userspace we also need to
ensure that all other users of GUP either reject those pages or map them
correctly.
Logan
