On Thu, Jul 25, 2024 at 05:23:48PM -0700, James Houghton wrote:
> On Thu, Jul 25, 2024 at 3:41 PM Peter Xu <peterx@xxxxxxxxxx> wrote:
> >
> > On Thu, Jul 25, 2024 at 11:29:49AM -0700, James Houghton wrote:
> > > > - pages += change_pmd_range(tlb, vma, pud, addr, next, newprot,
> > > > +
> > > > + if (pud_leaf(pud)) {
> > > > + if ((next - addr != PUD_SIZE) ||
> > > > + pgtable_split_needed(vma, cp_flags)) {
> > > > + __split_huge_pud(vma, pudp, addr);
> > > > + goto again;
> > >
> > > IIUC, most of the time, we're just going to end up clearing the PUD in
> > > this case. __split_huge_pud() will just clear it, then we goto again
> > > and `continue` to the next pudp. Is that ok?
> > >
> > > (I think it's ok as long as: you never map an anonymous page with a
> > > PUD,
> >
> > I think this is true.
> >
> > > and that uffd-wp is not usable with non-hugetlb PUD mappings of
> > > user memory (which I think is only DAX?).
> >
> > Uffd-wp has the async mode that can even work with dax puds.. even though I
> > don't think anyone should be using it. Just like I'm more sure that nobody
> > is using mprotect() too with dax pud, and it further justifies why nobody
> > cared this much..
> >
> > What uffd-wp would do in this case is it'll make pgtable_split_needed()
> > returns true on this PUD, the PUD got wiped out, goto again, then
> > change_prepare() will populate this pud with a pgtable page. Then it goes
> > downwards, install PMD pgtable, then probably start installing pte markers
> > ultimately if it's a wr-protect operation.
>
> Ah, I didn't understand this patch correctly. You're right, you'll
> install PMDs underneath -- I thought we'd just find pud_none() and
> bail out. So this all makes sense, thanks!
>
> >
> > > So it seems ok today...?)
> >
> > Yes I think it's ok so far, unless you think it's not. :)
> >
> > >
> > > Also, does the comment in pgtable_split_needed() need updating?
> >
> > /*
> > * Return true if we want to split THPs into PTE mappings in change
> > * protection procedure, false otherwise.
> > */
> >
> > It looks to me it's ok for now to me? THP can represents PUD in dax, and we
> > indeed want to break THPs (no matter PUD/PMD) finally into PTE mappings.
>
> Oh, heh I was thinking of the other comment:
>
> /*
> * pte markers only resides in pte level, if we need pte markers,
> * we need to split. We cannot wr-protect shmem thp because file
> * thp is handled differently when split by erasing the pmd so far.
> */
>
> I guess this is fine too, it just kind of reads as if this function is
> only called for PMDs. *shrug*
Ah ok, yeah it looks mostly fine here too. Let's make this easy by keeping
it as-is, but if there'll be a new version I can touch it up if it helps
the readings.
>
> >
> > >
> > > Somewhat related question: change_huge_pmd() is very careful not to
> > > clear the PMD before writing the new value. Yet change_pmd_range(),
> > > when it calls into __split_huge_pmd(), will totally clear the PMD and
> > > then populate the PTEs underneath (in some cases at least), seemingly
> > > reintroducing the MADV_DONTNEED concern. But your PUD version, because
> > > it never re-populates the PUD (or PMDs/PTEs underneath) does not have
> > > this issue. WDYT?
>
> I guess I'm wrong about this -- your PUD version is the same as the
> PMD version in this respect: both clear the PUD/PMD and then install
> lower page table entries.
>
> >
> > Could you elaborate more on the DONTNEED issue you're mentioning here?
>
> In change_huge_pmd(), there is a comment about not clearing the pmd
> before setting the new value. I guess the issue is: if a user is
> MADV_DONTNEEDing some memory and happens to see a cleared PMD, it will
> just move on. But in this case, if change_huge_pmd() temporarily
> cleared a PMD, then MADV_DONTNEED saw it and moved on, and then
> change_huge_pmd() installed a new PMD, the MADV_DONTNEED has basically
> skipped over the PMD, probably not what the user wanted. It seems like
> we have the same issue when a PMD is cleared when we're splitting it.
>
> But yeah, given that your PUD version is apparently no different from
> the PMD version in this respect, maybe this question isn't very
> interesting. :)
Ah that one, so yeah that's why I introduced pudp_invalidate() in this
series to make sure that issue isn't there. Then I assume we're good.
Thanks,
--
Peter Xu
