Re: (resend)WARNING: trying to isolate tail page in isolate_lru_page

[Minchan Kim <minchan@kernel.org>]

On Fri, Aug 26, 2022 at 11:19:37AM -0700, Yang Shi wrote:
> On Fri, Aug 26, 2022 at 10:53 AM Minchan Kim <minchan@kernel.org> wrote:
> >
> > On Fri, Aug 26, 2022 at 11:27:53AM -0600, Yu Zhao wrote:
> > > On Fri, Aug 26, 2022 at 11:15 AM Matthew Wilcox <willy@infradead.org> wrote:
> > > >
> > > > On Fri, Aug 26, 2022 at 11:20:58AM +0800, Yin, Fengwei wrote:
> > > > > On 8/26/2022 2:46 AM, Matthew Wilcox wrote:
> > > > > > There's a number of ways of solving this.  I'm interested in seeing
> > > > > > which one Minchan thinks is best.
> > > > >
> > > > > My understanding is:
> > > > > PageTransCompound() return false for compound page if THP is disabled
> > > > > in kernel config. Replacing PageTransCompound() with PageCompound()
> > > > > could work here. But for the long term, folio should be the answer. :).
> > > >
> > > > Yes, ultimately, isolate_lru_page() is going away as an interface
> > > > and one will have to call folio_isolate_lru().  But should
> > > > madvise_cold_or_pageout_pte_range() even be getting called for VMAs
> > > > which are mmaps of af_packet?  can_madv_lru_vma() rules out a number
> > > > of different types of VMA; should it also be ruling out af_packet VMAs?
> > >
> > > Agreed.
> > >
> > > > If so, how?
> > >
> > > We should add a reliable helper to tell whether a file VMA is
> > > reclaimable or not. I don't think we have one. Currently MGLRU checks
> > > mapping->a_ops->read_folio for file VMAs to determine whether they are
> > > reclaimable.
> > >
> >
> > Long term, that's better idea(For stable backport, I'd like to go with
> > simple PageLRU check).
> >
> > I wonder it's possible to mix LRU pages and non-struct pages together
> > in a VMA. Otherwise, could we reuse(abuse) VM_MIXEDMAP?
> 
> I don't think the mix is going to happen via regular mmap call.

At a quick glance of vm_insert_page, it could change vm_flags
into VM_MIXEDMAP dynamically. I am worry that a driver may
call the vm_insert_page in the vma which already has LRU pages
before so the VMA could have both LRU pages and non-LRU pages
in the VMA. If it's doable, we may miss LRU pages if we filter
VM_MIXEDMAP out in can_madv_lru_vma. Not sure it's common case.
Otherwise, yeah, we may introduce a new vm_flag, something like,
VM_NON_LRU. Just thought.