Re: 答复: [PATCH V2] vfio dma_map/unmap: optimized for hugetlbfs pages

[Peter Xu]

On Wed, Aug 26, 2020 at 01:56:43PM +0000, Maoming (maoming, Cloud Infrastructure Service Product Dept.) wrote:
> > +	/*
> > +	 * Unlike THP, the splitting should not happen for hugetlb pages.
> > +	 * Since PG_reserved is not relevant for compound pages, and the pfn of
> > +	 * PAGE_SIZE page which in hugetlb pages is valid,
> > +	 * it is not necessary to check rsvd for hugetlb pages.
> > +	 * We do not need to alloc pages because of vaddr and we can finish all
> > +	 * work by a single operation to the head page.
> > +	 */
> > +	atomic_add(contiguous_npage, compound_pincount_ptr(head));
> > +	page_ref_add(head, contiguous_npage);
> > +	mod_node_page_state(page_pgdat(head), NR_FOLL_PIN_ACQUIRED, 
> > +contiguous_npage);
> 
> I think I asked this question in v1, but I didn't get any answer... So I'm trying again...
> 
> Could I ask why manual referencing of pages is done here rather than using
> pin_user_pages_remote() just like what we've done with vaddr_get_pfn(), and let
> try_grab_page() to do the page reference and accountings?
> 
> I feel like this at least is against the FOLL_PIN workflow of gup, because those FOLL_PIN paths were bypassed, afaict.
> 
> 
> Hi,
> My apologies for not answering your question.
> As I understand, pin_user_pages_remote() might spend much time.
> Because all PAGE_SIZE-pages in a hugetlb page are pinned one by one in pin_user_pages_remote() and try_grab_page().
> So I think maybe we can use these simple code to do all work.
> Am I wrong? And is there something else we can use? For example :pin_user_pages_fast()

Yeah I can understand your concern, however so far it's not about the perf but
correctness.  Documentation/core-api/pin_user_pages.rst tells us that we should
always use pin_user_page*() APIs to pin DMA pages (with FOLL_LONGTERM).  That's
something we should follow for now, otherwise the major logic of either
FOLL_PIN or FULL_LONGTERM could be bypassed without being noticed.

I'm not sure whether the perf issue is a big one.  So have you tried the pin
page APIs first and did some measurement?  There is indeed a tight loop in
follow_hugetlb_page() however not sure how much it'll affect VFIO_IOMMU_MAP_DMA
in general.  Even if we want to do something, it seems to be more suitable to
be done inside follow_hugetlb_page() rather than in vfio, imho.

Another comment is about the design of the whole patch - I think Alex commented
on that too on the awkwardness on appending the hugetlbfs logic to the end of
the existing logic.  Considering that current logic of vfio_pin_pages_remote()
is "let's pin some pages as long as continuous", not sure whether we can make
it into:

vfio_pin_pages_remote()
{
  if (PageHuge(first_page))
    vfio_pin_pages_hugetlbfs();
  else
    vfio_pin_pages_normal();
}

The thing is, if the 1st page is normal page, then the follow-up pages
shouldn't normally be hugetlbfs pages so they won't be physically continuous.
Vice versa.  In other words, each call to vfio_pin_pages_remote() should only
handle only one type of page after all.  So maybe we can diverge them at the
beginning of the call directly.

-- 
Peter Xu