On 04/01/18 03:13 PM, Jason Gunthorpe wrote:
On Thu, Jan 04, 2018 at 12:52:24PM -0700, Logan Gunthorpe wrote:
> We tried things like this in an earlier iteration which assumed the SG
> was homogenous (all P2P or all regular memory). This required serious
> ugliness to try and ensure SGs were in fact homogenous.
I'm confused, these patches already assume the sg is homogenous,
right? Sure looks that way. So  is just debugging??
Yes, but it's a bit different to expect that someone calling
pci_p2pmem_map_sg() will know what they're doing and provide a
homogenous SG. It is relatively clear by convention that the entire SG
must be homogenous given they're calling a pci_p2pmem function. Where
as, allowing P2P SGs into the core DMA code means all we can do is hope
that future developers don't screw it up and allow P2P pages to mix in
with regular pages.
It's also very difficult to add similar functionality to dma_map_page
seeing dma_unmap_page won't have any way to know what it's dealing with.
It just seems confusing to support P2P in the SG version and not the
What I meant is to rely on the fact it is already homogenous and
create a function like this:
static inline bool sg_is_table_p2p(struct scatterlist *sg)
And then insert
Yes, this is almost identical to the earlier patch I referenced...
Then we don't need to patch RDMA because RDMA is not special when
comes to P2P. P2P should work with everything.
Yes, I agree this would be very nice. But it's challenging at this time
to do it safely. And something like this can still be done in future
work after this patch set gets merged. The changes in this set to the
RDMA and block layers are fairly small and the majority of the block
layer changes would still be required anyway to ensure a given
queue/driver supports P2P requests.
>> The signature of pci_p2pmem_map_sg also doesn't look very
>> should mirror the usual dma_map_sg, otherwise it needs more revision
>> to extend this to do P2P through a root complex.
> Generally, the feedback that I get is to not make changes that try to
> anticipate the future. It would be easy enough to add those arguments when
> the code for going through the RC is made (which I expect will be very
> involved anyway).
Yes, but DMA mapping fundamentally requires the arguments to
dma_map_sg, so it is hard to accept an API called dma_map without
them.. Maybe just a naming issue.
I don't agree that this is a fundamental requirement. But if no one else
objects to adding unused arguments I'd be fine with adding them.