Re: [PATCH 1/2] mm: vmalloc: streamline vmalloc memory accounting

[Johannes Weiner <hannes@cmpxchg.org>]

On Mon, Feb 23, 2026 at 04:30:32PM +0100, Uladzislau Rezki wrote:
> On Fri, Feb 20, 2026 at 02:10:34PM -0500, Johannes Weiner wrote:
> > @@ -3655,6 +3649,8 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
> >  			continue;
> >  		}
> >  
> > +		mod_node_page_state(page, NR_VMALLOC, 1 << large_order);
> > +
> >  		split_page(page, large_order);
> >  		for (i = 0; i < (1U << large_order); i++)
> >  			pages[nr_allocated + i] = page + i;
> > @@ -3675,6 +3671,7 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
> >  	if (!order) {
> >  		while (nr_allocated < nr_pages) {
> >  			unsigned int nr, nr_pages_request;
> > +			int i;
> >  
> >  			/*
> >  			 * A maximum allowed request is hard-coded and is 100
> > @@ -3698,6 +3695,9 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
> >  							nr_pages_request,
> >  							pages + nr_allocated);
> >  
> > +			for (i = nr_allocated; i < nr_allocated + nr; i++)
> > +				inc_node_page_state(pages[i], NR_VMALLOC);
> > +
> >  			nr_allocated += nr;
> >  
> >  			/*
> > @@ -3722,6 +3722,8 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
> >  		if (unlikely(!page))
> >  			break;
> >  
> > +		mod_node_page_state(page, NR_VMALLOC, 1 << order);
> > +
> >  		/*
> Can we move *_node_page_stat() to the end of the vm_area_alloc_pages()?
> 
> Or mod_node_page_state in first place should be invoked on high-order
> page before split(to avoid of looping over small pages afterword)?
> 
> I mean it would be good to place to the one solid place. If it is possible
> of course.

Note that the top one in the fast path IS called before the
split. We're accounting in the same step size as the page allocator
can give us.

In the fallback paths (bulk allocator, and one-by-one loop), the issue
is that the individual pages could be coming from different nodes, so
they need to bump different counters. One possible solution would be
to remember the last node and accumulate until it differs, then flush:

fallback_loop() {
	page = alloc_pages();
	nid = page_to_nid(page);
	if (nid != last_nid) {
		if (node_count) {
			mod_node_page_state(...);
			node_count = 0;
		}
		last_nid = nid;
	}
}

if (node_count)
	mod_node_page_state(...);

But it IS the slow path, and these are fairly cheap per-cpu
counters. Especially compared to the cost of calling into the
allocator. So I'm not sure it's worth it... What do you think?