[Intel-gfx] [PATCH hmm 5/5] mm/hmm: remove the customizable pfn format from hmm_range_fault

[Christoph Hellwig]

On Wed, Apr 22, 2020 at 09:39:11AM -0300, Jason Gunthorpe wrote:
> > Nice callchain from hell..  Unfortunately such "code listings" tend to
> > get out of date very quickly, so I'm not sure it is worth keeping in
> > the code.  What would be really worthile is consolidating the two
> > different sets of defines (NVIF_VMM_PFNMAP_V0_ vs NVKM_VMM_PFN_)
> > to make the code a little easier to follow.
> 
> I was mainly concerned that this function is using hmm properly,
> becuase it sure looks like it is just forming the CPU physical address
> into a HW specific data. But it turns out it is just an internal data
> for some other code and the dma_map is impossibly far away
> 
> It took forever to find, I figured I'd leave a hint for the next poor
> soul that has to look at this.. 
> 
> Also, I think it shows there is no 'performance' argument here, if
> this path needs more performance the above should be cleaned
> before we abuse hmm_range_fault.
> 
> Put it in the commit message instead?

Yes, the graph itself sounds reasonable for the commit log as a point
of time information.

> > >  	npages = (range->end - range->start) >> PAGE_SHIFT;
> > >  	for (i = 0; i < npages; ++i) {
> > >  		struct page *page;
> > >  
> > > +		if (!(range->hmm_pfns[i] & HMM_PFN_VALID)) {
> > > +			ioctl_addr[i] = 0;
> > >  			continue;
> > > +		}
> > 
> > Can't we rely on the caller pre-zeroing the array?
> 
> This ends up as args.phys in nouveau_svm_fault - I didn't see a
> zeroing?
> 
> I think it makes sense that this routine fully sets the output array
> and does not assume pre-initialize

Ok.