On Fri, Feb 18, 2022 at 02:20:45PM -0500, Felix Kuehling wrote:
> Am 2022-02-17 um 19:19 schrieb Jason Gunthorpe:
> > On Thu, Feb 17, 2022 at 04:12:20PM -0500, Felix Kuehling wrote:
> >
> > > I'm thinking of a more theoretical approach: Instead of auditing all
> > > users,
> > > I'd ask, what are the invariants that a vm_normal_page should have. Then
> > > check, whether our DEVICE_COHERENT pages satisfy them. But maybe the
> > > concept
> > > of a vm_normal_page isn't defined clearly enough for that.
> > I would say the expectation is that only 'page cache and anon' pages
> > are returned - ie the first union in struct page
> >
> > This is because the first file in your list I looked at:
> >
> > static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
> > unsigned long addr, unsigned long end,
> > struct mm_walk *walk)
> >
> > {
> > page = vm_normal_page(vma, addr, ptent);
> > [..]
> > if (pageout) {
> > if (!isolate_lru_page(page)) {
> >
> > Uses the LRU field, so this is incompatible with all the other page
> > types.
> >
> > One mitigation of this might be to formally make vm_normal_page() ==
> > 'pte to page cache and anon page' and add a new function that is 'pte
> > to any struct page'
> >
> > Then go through and sort callers as appropriate.
> >
> > The 'pte to page cache and anon page' can detect ZONE_DEVICE by
> > calling is_zone_device_page() insted of pte_devmap() and then continue
> > to return NULL. This same trick will fix GUP_fast.
>
> Sounds good to me. What about vm_normal_page_pmd? Should we remove the
> pmd_devmap check from that function as well. I'm not even sure what a huge
> zone_device page would look like, but maybe that's a worthwhile future
> optimization for our driver.
IIRC there are other problems here as PMDs are currently wired to THPs
and not general at all..
We have huge zone_device pages, it is just any compound page of
contiguous pfns - you should be aggregating any contiguous string of
logical PFNs together into a folio for performance. If the folio is
stuffed into a PMD or not is a different question..
> I'd propose the function names vm_normal_page and vm_normal_or_device_page
> for the two functions you described.
I wouldn't say device_page, it should be any type of page - though
device_page is the only other option ATM, AFIAK.
> current vm_normal_page with the pte_devmap check removed. vm_normal_page
> could be implemented as a wrapper around vm_normal_or_device_page, which
> just adds the !is_zone_device_page() check.
Yes
Jason
