On Wed, 2022-07-06 at 14:05 -0300, Jason Gunthorpe wrote:
> On Tue, Jul 05, 2022 at 11:27:53PM -0700, Nicolin Chen wrote:
> > This driver is the only caller of vfio_pin/unpin_pages that might
> > pass
> > in a non-contiguous PFN list, but in many cases it has a contiguous
> > PFN
> > list to process. So letting VFIO API handle a non-contiguous PFN
> > list
> > is actually counterproductive.
> >
> > Add a pair of simple loops to pass in contiguous PFNs only, to have
> > an
> > efficient implementation in VFIO.
> >
> > Signed-off-by: Nicolin Chen <nicolinc@xxxxxxxxxx>
> > ---
> > drivers/s390/cio/vfio_ccw_cp.c | 70 +++++++++++++++++++++++++++---
> > ----
> > 1 file changed, 56 insertions(+), 14 deletions(-)
>
> I think this is fine as-is for this series, but someone who knows and
> can test ccw should go in and fix things so that pfn_array_alloc()
> doesn't exist. Allocating memory and filling it with consecutive
> integers is kind of silly given we can just call vfio_pin_pages()
> with
> pa_nr directly.
>
> pa->pa_iova_pfn[0] = pa->pa_iova >> PAGE_SHIFT;
> pa->pa_pfn[0] = -1ULL;
> for (i = 1; i < pa->pa_nr; i++) {
> pa->pa_iova_pfn[i] = pa->pa_iova_pfn[i - 1] + 1;
>
> It looks like only the 'ccw_is_idal' flow can actually create
> non-continuities.
This code is certainly not my favorite, but you're right that it's the
IDAL flow that generates the non-contiguous requests and the code you
reference is simply an initialization for the !IDAL case. As I have a
todo in this code anyway, I'll register your suggestion to see if they
can be untangled.
> Also the loop in copy_from_iova() should ideally be
> using the much faster 'rw' interface, and not a pin/unpin cycle just
> to memcpy.
I guess I missed when that was added. This looks like low hanging fruit
for some ollllld code regardless of the above. Will get to this once
I'm back. Thank you!
Eric
>
> If I guess right these changes would significantly speed this driver
> up.
>
> Anyhow,
>
> Reviewed-by: Jason Gunthorpe <jgg@xxxxxxxxxx>
>
> Jason
