Re: [PATCH 5/7] khugepaged: Allow to collapse PTE-mapped compound pages

[Yang Shi]

On Sat, Mar 28, 2020 at 5:33 AM Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
>
> On Fri, Mar 27, 2020 at 09:17:00PM -0400, Zi Yan wrote:
> > > The compound page may be locked here if the function called for the first
> > > time for the page and not locked after that (becouse we've unlocked it we
> > > saw it the first time). The same with LRU.
> > >
> >
> > For the first time, the compound page is locked and not on LRU, so this VM_BUG_ON passes.
> > For the second time and so on, the compound page is unlocked and on the LRU,
> > so this VM_BUG_ON still passes.
> >
> > For base page, VM_BUG_ON passes.
> >
> > Other unexpected situation (a compound page is locked and on LRU) triggers the VM_BU_ON,
> > but your VM_BUG_ON will not detect this situation, right?
>
> Right. I will rework this code. I've just realized it is racy: after
> unlock and putback on LRU the page can be locked by somebody else and this
> code can unlock it which completely borken.

I'm wondering if we shall skip putting the page back to LRU. Since the
page is about to be freed soon, so as I mentioned in the other patch
it sounds not very productive to put it back to LRU again.

>
> I'll pass down compound_pagelist to release_pte_pages() and handle the
> situation there.
>
> > >>>     if (likely(writable)) {
> > >>>             if (likely(referenced)) {
> > >>
> > >> Do we need a list here? There should be at most one compound page we will see here, right?
> > >
> > > Um? It's outside the pte loop. We get here once per PMD range.
> > >
> > > 'page' argument to trace_mm_collapse_huge_page_isolate() is misleading:
> > > it's just the last page handled in the loop.
> > >
> >
> > Throughout the pte loop, we should only see at most one compound page, right?
>
> No. mremap(2) opens a possibility for HPAGE_PMD_NR compound pages for
> single PMD range.
>
>
> --
>  Kirill A. Shutemov
>