Re: Regression - locking (all from 2.6.28)

[Dave Hansen <dave@linux.vnet.ibm.com>]

On Tue, 2009-03-03 at 15:01 +0000, Catalin Marinas wrote:
> > +     /* mem_map scanning */
> > +     for_each_online_node(i) {
> > +             struct page *page, *end;
> > +
> > +             page = NODE_MEM_MAP(i);
> > +             end  = page + NODE_DATA(i)->node_spanned_pages;
> > +
> > +             scan_block(page, end, NULL);
> > +     }
> > 
> > The alternative is to inform kmemleak about the page structures returned
> > from __alloc_pages_internal() but there would be problems with recursive
> > calls into kmemleak when it allocates its own data structures.
> > 
> > I'll look at re-adding the hunk above, maybe with some extra checks like
> > pfn_valid().
> 
> Looking again at this, the node_mem_map is always contiguous and the
> code above only scans the node_mem_map, not the memory represented by
> the node (which may not be contiguous). So I think it is a valid code
> sequence.

The above is *not* a valid code sequence.

It is valid with discontig, but isn't valid for sparsemem.  You simply
can't expect to do math on 'struct page' pointers for any granularity
larger than MAX_ORDER_NR_PAGES.

Also, we don't even define NODE_MEM_MAP() for all configurations so that
code snippet won't even compile.  We would be smart to kill that macro.

One completely unoptimized thing you can do which will scan a 'struct
page' at a time is this:

	for_each_online_node(i) {
		unsigned long pfn;
		for (pfn = node_start_pfn(i); pfn < node_end_pfn(i); pfn++) {
			struct page *page;
			if (!pfn_valid(pfn))
				continue;
			page = pfn_to_page(pfn);
			scan_block(page, page+1, NULL);
		}
	}
		
The way to optimize it would be to call scan_block() only once for each
MAX_ORDER_NR_PAGES that you encounter.  The other option would be to use
the active_regions functions to walk the memory.  

Is there a requirement to reduce the number of calls to scan_block()
here?

-- Dave