Re: [RFC PATCH 14/15] PCI/P2PDMA: Introduce pci_mmap_p2pmem()

[Jason Gunthorpe]

On Fri, Nov 06, 2020 at 10:53:45AM -0700, Logan Gunthorpe wrote:
> 
> 
> On 2020-11-06 10:42 a.m., Jason Gunthorpe wrote:
> > On Fri, Nov 06, 2020 at 10:28:00AM -0700, Logan Gunthorpe wrote:
> >>
> >>
> >> On 2020-11-06 10:22 a.m., Jason Gunthorpe wrote:
> >>> On Fri, Nov 06, 2020 at 10:00:35AM -0700, Logan Gunthorpe wrote:
> >>>> Introduce pci_mmap_p2pmem() which is a helper to allocate and mmap
> >>>> a hunk of p2pmem into userspace.
> >>>>
> >>>> Signed-off-by: Logan Gunthorpe <logang@xxxxxxxxxxxx>
> >>>>  drivers/pci/p2pdma.c       | 104 +++++++++++++++++++++++++++++++++++++
> >>>>  include/linux/pci-p2pdma.h |   6 +++
> >>>>  2 files changed, 110 insertions(+)
> >>>>
> >>>> diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
> >>>> index 9961e779f430..8eab53ac59ae 100644
> >>>> +++ b/drivers/pci/p2pdma.c
> >>>> @@ -16,6 +16,7 @@
> >>>>  #include <linux/genalloc.h>
> >>>>  #include <linux/memremap.h>
> >>>>  #include <linux/percpu-refcount.h>
> >>>> +#include <linux/pfn_t.h>
> >>>>  #include <linux/random.h>
> >>>>  #include <linux/seq_buf.h>
> >>>>  #include <linux/xarray.h>
> >>>> @@ -1055,3 +1056,106 @@ ssize_t pci_p2pdma_enable_show(char *page, struct pci_dev *p2p_dev,
> >>>>  	return sprintf(page, "%s\n", pci_name(p2p_dev));
> >>>>  }
> >>>>  EXPORT_SYMBOL_GPL(pci_p2pdma_enable_show);
> >>>> +
> >>>> +struct pci_p2pdma_map {
> >>>> +	struct kref ref;
> >>>> +	struct pci_dev *pdev;
> >>>> +	void *kaddr;
> >>>> +	size_t len;
> >>>> +};
> >>>
> >>> Why have this at all? Nothing uses it and no vm_operations ops are
> >>> implemented?
> >>
> >> It's necessary to free the allocated p2pmem when the mapping is torn down.
> > 
> > That's suspicious.. Once in a VMA the lifetime of the page must be
> > controlled by the page refcount, it can't be put back into the genpool
> > just because the vma was destroed.
> 
> Ah, hmm, yes. I guess the pages have to be hooked and returned to the
> genalloc through free_devmap_managed_page(). 

That sounds about right, but in this case it doesn't need the VMA
operations.

> Seems like it might be doable... but it will complicate things for
> users that don't want to use the genpool (though no such users exist
> upstream).

I would like to use this stuff in RDMA pretty much immediately and the
genpool is harmful for those cases, so please don't make decisions
that are tying thing to genpool

Jason