On 2020-11-06 5:14 p.m., Jason Gunthorpe wrote:

On Fri, Nov 06, 2020 at 01:03:26PM -0700, Logan Gunthorpe wrote:

quoted

I don't think a function like that will work for the p2pmem use case. In
order to implement proper page freeing I expect I'll need a loop around
the allocator and vm_insert_mixed()... Something roughly like:

for (addr = vma->vm_start; addr < vma->vm_end; addr += PAGE_SIZE) {
        vaddr = pci_alloc_p2pmem(pdev, PAGE_SIZE);
	ret = vmf_insert_mixed(vma, addr,
  		       __pfn_to_pfn_t(virt_to_pfn(vaddr), PFN_DEV | PFN_MAP));
}

That way we can call pci_free_p2pmem() when a page's ref count goes to
zero. I suspect your use case will need to do something similar.

Yes, but I would say the pci_alloc_p2pmem() layer should be able to
free pages on a page-by-page basis so you don't have to do stuff like
the above.

There is often a lot of value in having physical contiguous addresses,
so allocating page by page as well seems poor.

Agreed. But I'll have to dig to see if genalloc supports freeing blocks
in different sizes than the allocations.

Logan