Re: [PATCH] mm/page_alloc: Make alloc_gigantic_page() available for general use

[Michal Hocko]

On Wed 16-10-19 10:56:16, David Hildenbrand wrote:
> On 16.10.19 10:51, Michal Hocko wrote:
> > On Wed 16-10-19 10:08:21, David Hildenbrand wrote:
> > > On 16.10.19 09:34, Anshuman Khandual wrote:
> > [...]
> > > > +static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
> > > > +				   unsigned long nr_pages)
> > > > +{
> > > > +	unsigned long i, end_pfn = start_pfn + nr_pages;
> > > > +	struct page *page;
> > > > +
> > > > +	for (i = start_pfn; i < end_pfn; i++) {
> > > > +		page = pfn_to_online_page(i);
> > > > +		if (!page)
> > > > +			return false;
> > > > +
> > > > +		if (page_zone(page) != z)
> > > > +			return false;
> > > > +
> > > > +		if (PageReserved(page))
> > > > +			return false;
> > > > +
> > > > +		if (page_count(page) > 0)
> > > > +			return false;
> > > > +
> > > > +		if (PageHuge(page))
> > > > +			return false;
> > > > +	}
> > > 
> > > We might still try to allocate a lot of ranges that contain unmovable data
> > > (we could avoid isolating a lot of page blocks in the first place). I'd love
> > > to see something like pfn_range_movable() (similar, but different to
> > > is_mem_section_removable(), which uses has_unmovable_pages()).
> > 
> > Just to make sure I understand. Do you want has_unmovable_pages to be
> > called inside pfn_range_valid_contig?
> 
> I think this requires more thought, as has_unmovable_pages() works on
> pageblocks only AFAIK. If you try to allocate < MAX_ORDER - 1, you could get
> a lot of false positives.
> 
> E.g., if a free "MAX_ORDER - 1" page spans two pageblocks and you only test
> the second pageblock, you might detect "unmovable" if not taking proper care
> of the "bigger" free page. (alloc_contig_range() properly works around that
> issue)

OK, I see your point. You are right that false positives are possible. I
would deal with those in a separate patch though.

> > [...]
> > > > +struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
> > > > +				int nid, nodemask_t *nodemask)
> > > > +{
> > > > +	unsigned long ret, pfn, flags;
> > > > +	struct zonelist *zonelist;
> > > > +	struct zone *zone;
> > > > +	struct zoneref *z;
> > > > +
> > > > +	zonelist = node_zonelist(nid, gfp_mask);
> > > > +	for_each_zone_zonelist_nodemask(zone, z, zonelist,
> > > > +					gfp_zone(gfp_mask), nodemask) {
> > > 
> > > One important part is to never use the MOVABLE zone here (otherwise
> > > unmovable data would end up on the movable zone). But I guess the caller is
> > > responsible for that (not pass GFP_MOVABLE) like gigantic pages do.
> > 
> > Well, if the caller uses GFP_MOVABLE then the movability should be
> > implemented in some form. If that is not the case then it is a bug on
> > the caller behalf.
> > 
> > > > +		spin_lock_irqsave(&zone->lock, flags);
> > > > +
> > > > +		pfn = ALIGN(zone->zone_start_pfn, nr_pages);
> > > 
> > > This alignment does not make too much sense when allowing passing in !power
> > > of two orders. Maybe the caller should specify the requested alignment
> > > instead? Or should we enforce this to be aligned to make our life easier for
> > > now?
> > 
> > Are there any usecases that would require than the page alignment?
> 
> Gigantic pages have to be aligned AFAIK.

Aligned to what? I do not see any guarantee like that in the existing
code.

-- 
Michal Hocko
SUSE Labs