Re: [PATCH v6] page_alloc: allow migration of smaller hugepages during contig_alloc

[Gregory Price <gourry@gourry.net>]

On Fri, Dec 19, 2025 at 03:46:25PM -0500, Zi Yan wrote:
> On 19 Dec 2025, at 9:26, Gregory Price wrote:
> 
> > On Fri, Dec 19, 2025 at 12:08:00AM +0000, Wei Yang wrote:
> >>> +
> >>> +			page = compound_head(page);
> >>> +			order = compound_order(page);
> >>
> >> The order is get from head page.
> >>
> >>> +			if ((order >= MAX_FOLIO_ORDER) ||
> >>> +			    (nr_pages <= (1 << order)))
> >>> +				return false;
> >>> +
> >>> +			/* No need to check the pfns for this page */
> >>> +			i += (1 << order) - 1;
> >>
> >> So this advance should based on "head page" instead of original page, right?
> >>
> >
> > hm, I think the thought here was that since we're moving forward from
> > start of an aligned chunk, we'd never hit a non-head page - but this
> > may not be true.
> >
> > Will think about this for a bit.
> 
> The sole caller of pfn_range_valid_contig(), alloc_contig_pages_noprof(),
> scans from the beginning of a zone to the end. pfn_range_valid_contig()
> should see head pages all the time, except it scans in the middle of
> a 1GB hugetlb when alloc_contig_pages_noprof() is asking for a smaller
> nr_pages, like 2MB. But in that case, the if above i += (1 << order) - 1
> would return false without reaching it. Basically, to get to
> i += ..., pfn_range_valid_contig() needs to search for nr_pages larger
> than PageHuge(page) and nr_pages is always power of two based on
> alloc_contig_pages_noprof() requirement, but that means
> pfn_range_valid_contig() always sees such PageHuge pages as a whole
> within nr_pages range, thus cannot see a tail PageHuge page at the
> point of i += ....
> 

Thinking about this a bit more, it might be worthwhile to detect this
condiition and just skip that hugepage in the external code.

while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
	if (pfn_range_valid_contig(zone, pfn, nr_pages,
				   skip_hugetlb,
				   &skipped_hugetlb)) {
		... snip ...
	}
	pfn += nr_pages;
	/*
	 * TODO: If the last scanned page was a hugepage that caused
	 *       the zone to be invalid, skip the rest of that page
	 *       (e.g. if we hit a 1GB page trying to allocate a 2MB
	 *       page, skip the entire 1GB instead of scanning the
	 *       same page 1GB/2MB times).
	 */
	 ...
}

But this solves a different problem than this patch, so i will defer.

~Gregory