On Fri, 2011-01-21 at 02:38 +0900, KyongHo Cho wrote:

Actually, as long as a bank in meminfo only resides in a pgdat, no
problem happens
because there is no restriction of size of area in a pgdat.
That's why I just considered about sparsemem.

Ahh, so "banks" are always underneath a single pgdat, and a "bank" is
always contiguous?  That's handy.

I worried that pfn_to_page() in sparsemem is a bit slower than that in flatmem.
Moreover, the previous one didn't use pfn_to_page() but page++ for the
performance.
Nevertheless, I also think that pfn_to_page() make the code neat.

The sparsemem_vmemmap pfn_to_page() is just arithmetic.  The table-based
sparsemem requires lookups and is a _bit_ slower, but the tables have
very nice CPU cache properties and shouldn't miss the L1 very often in a
loop like that.

show_mem() isn't exactly a performance-critical path, either, right?
It's just an exception or error path.

If it turns out that doing pfn_to_page() *is* too slow, there are a
couple more alternatives.  pfn_to_section_nr() is just a bit shift and
is really cheap.  Should be just an instruction or two with either no
memory access, or just a load of the pfn from the stack.

We could make a generic function like this (Or I guess we could also
just make sure that pfn_to_section_nr() always returns 0 for
non-sparsemem configurations):

int pfns_same_section(unsigned long pfn1, unsigned long pfn2)
{
#ifdef CONFIG_SPARSEMEM
	return (pfn_to_section_nr(pfn1) == pfn_to_section_nr(pfn2));
#else
	return 1;
#endif
}

and use it in show_mem like so:

                do {
                        total++;
                        if (PageReserved(page))
                                reserved++;
                        else if (PageSwapCache(page))
                                cached++;
                        else if (PageSlab(page))
                                slab++;
                        else if (!page_count(page))
                                free++;
                        else
                                shared += page_count(page) - 1;
			pfn1++;
			/*
			 * Did we just cross a section boundary?
			 * If so, our pointer arithmetic is not
			 * valid, and we must re-run pfn_to_page()
			 */
			if (pfns_same_section(pfn1-1, pfn1)) {
	                        page++;
			} else {
				page = pfn_to_page(pfn1);
			}
                } while (page < end);

We can do basically the same thing, but instead checking to see if we
crossed a MAX_ORDER boundary.  That would keep us from having to refer
to sparsemem@all.  The buddy allocator relies on that guarantee, so
it's pretty set in stone.

-- Dave