[Intel-gfx] [PATCH v9 5/7] vfio: Define vfio based dma-buf operations

[Alex Williamson]

On Fri, 23 Jun 2017 10:31:28 +0200
Gerd Hoffmann <kraxel at redhat.com> wrote:

> On Fri, 2017-06-23 at 15:49 +0800, Zhi Wang wrote:
> > Hi:
> >      Thanks for the discussions! If the userspace application has 
> > already maintained a LRU list, it looks like we don't need
> > generation 
> > anymore,  
> 
> generation isn't required, things are working just fine without that. 
> It is just a small optimization, userspace can skip the LRU lookup
> altogether if the generation didn't change.
> 
> But of couse that only pays off if the kernel doesn't has to put much
> effort into maintaining the generation id.  Something simple like
> increasing it each time the guest writes a register which affects
> plane_info.

But it seems like that simple management algorithm pretty much
guarantees that the kernel will never revisit a generation and
therefore caching dmabuf fds is pointless.  AIUI the optimization is to
allow userspace to 'at a glance' test one plane_info vs another.  The
user could also do this with a memcmp of the plane_info structs if
that's its only purpose.  A randomly incremented field within that
struct could actually be a hindrance to the user for such a
comparison.  Are there cases where the plane_info struct is otherwise
identical where the user would need to get a new dmabuf fd anyway?
Thanks,

Alex