Re: [PATCH 2/4] mm: khugepaged: remove mm when all memory has been collapsed

[David Hildenbrand (Red Hat)]

On 12/19/25 09:35, Vernon Yang wrote:
> On Thu, Dec 18, 2025 at 10:29:18AM +0100, David Hildenbrand (Red Hat) wrote:
> > On 12/15/25 10:04, Vernon Yang wrote:
> > > The following data is traced by bpftrace on a desktop system. After
> > > the system has been left idle for 10 minutes upon booting, a lot of
> > > SCAN_PMD_MAPPED or SCAN_PMD_NONE are observed during a full scan by
> > > khugepaged.
> > >
> > > @scan_pmd_status[1]: 1           ## SCAN_SUCCEED
> > > @scan_pmd_status[4]: 158         ## SCAN_PMD_MAPPED
> > > @scan_pmd_status[3]: 174         ## SCAN_PMD_NONE
> > > total progress size: 701 MB
> > > Total time         : 440 seconds ## include khugepaged_scan_sleep_millisecs
> > >
> > > The khugepaged_scan list save all task that support collapse into hugepage,
> > > as long as the take is not destroyed, khugepaged will not remove it from
> > > the khugepaged_scan list. This exist a phenomenon where task has already
> > > collapsed all memory regions into hugepage, but khugepaged continues to
> > > scan it, which wastes CPU time and invalid, and due to
> > > khugepaged_scan_sleep_millisecs (default 10s) causes a long wait for
> > > scanning a large number of invalid task, so scanning really valid task
> > > is later.
> > >
> > > After applying this patch, when all memory is either SCAN_PMD_MAPPED or
> > > SCAN_PMD_NONE, the mm is automatically removed from khugepaged's scan
> > > list. If the page fault or MADV_HUGEPAGE again, it is added back to
> > > khugepaged.
> >
> > I don't like that, as it assumes that memory within such a process would be
> > rather static, which is easily not the case (e.g., allocators just doing
> > MADV_DONTNEED to free memory).
> >
> > If most stuff is collapsed to PMDs already, can't we just skip over these
> > regions a bit faster?
>
> I have a flash of inspiration and came up with a good idea.
>
> If these regions have already been collapsed into hugepage, rechecking
> them would be very fast. Due to the khugepaged_pages_to_scan can also
> represent the number of VMAs to skip, we can extend its semantics as
> follows:
>
> 	/*
> 	 * default scan 8*HPAGE_PMD_NR ptes, pmd_mapped, no_pte_table or vmas
> 	 * every 10 second.
> 	 */
> 	static unsigned int khugepaged_pages_to_scan __read_mostly;
>
> 	switch (*result) {
> 	case SCAN_NO_PTE_TABLE:
> 	case SCAN_PMD_MAPPED:
> 	case SCAN_PTE_MAPPED_HUGEPAGE:
> 		progress++; // here
> 		break;
> 	case SCAN_SUCCEED:
> 		++khugepaged_pages_collapsed;
> 		fallthrough;
> 	default:
> 		progress += HPAGE_PMD_NR;
> 	}
>
> This way can achieve our goal. David, do you like it?

I'd have to see the full patch, but IMHO we should rather focus on on 
"how many pte/pmd entries did we check" and not "how many PMD areas did 
we check".

Maybe there is a history to this, but conceptually I think we wanted to 
limit the work we do in one operation to something reasonable. Reading a 
single PMD is obviously faster than 512 PTEs.

--
Cheers

David