[PATCH] ARM: mm: Regarding section when dealing with meminfo

[Russell King - ARM Linux]

On Mon, Jan 24, 2011 at 08:52:17AM -0800, Dave Hansen wrote:
> On Sun, 2011-01-23 at 18:05 +0000, Russell King - ARM Linux wrote:
> > On Thu, Jan 20, 2011 at 10:11:27AM -0800, Dave Hansen wrote:
> > > On Thu, 2011-01-20 at 18:01 +0000, Russell King - ARM Linux wrote:
> > > > > The x86 version of show_mem() actually manages to do this without any
> > > > > #ifdefs, and works for a ton of configuration options.  It uses
> > > > > pfn_valid() to tell whether it can touch a given pfn.
> > > > 
> > > > x86 memory layout tends to be very simple as it expects memory to
> > > > start at the beginning of every region described by a pgdat and extend
> > > > in one contiguous block.  I wish ARM was that simple.
> > > 
> > > x86 memory layouts can be pretty funky and have been that way for a long
> > > time.  That's why we *have* to handle holes in x86's show_mem().  My
> > > laptop even has a ~1GB hole in its ZONE_DMA32:
> > 
> > If x86 is soo funky, I suggest you try the x86 version of show_mem()
> > on an ARM platform with memory holes.  Make sure you try it with
> > sparsemem as well...
> 
> x86 uses the generic lib/ show_mem().  It works for any holes, as long
> as they're expressed in one of the memory models so that pfn_valid()
> notices them.

I think that's what I said.

> ARM looks like its pfn_valid() is backed up by searching the (ASM
> arch-specific) memblocks.  That looks like it would be fairly slow
> compared to the other pfn_valid() implementations and I can see why it's
> being avoided in show_mem().

Wrong.  For flatmem, we have a pfn_valid() which is backed by doing a
one, two or maybe rarely three compare search of the memblocks.  Short
of having a bitmap of every page in the 4GB memory space, you can't
get more efficient than that.

For sparsemem, sparsemem provides its own pfn_valid() which is _far_
from what we require:

static inline int pfn_valid(unsigned long pfn)
{
        if (pfn_to_section_nr(pfn) >= NR_MEM_SECTIONS)
                return 0;
        return valid_section(__nr_to_section(pfn_to_section_nr(pfn)));
}

> Maybe we should add either the MAX_ORDER or section_nr() trick to the
> lib/ implementation.  I bet that would use pfn_valid() rarely enough to
> meet any performance concerns.

No.  I think it's entirely possible that on some platforms we have holes
within sections.  Like I said, ours is more funky than x86.

The big problem we have on ARM is that the kernel sparsemem stuff doesn't
*actually* support sparse memory.  What it supports is fully populated
blocks of memory of fixed size (known at compile time) where some blocks
may be contiguous.  Blocks are assumed to be populated from physical
address zero.

It doesn't actually support partially populated blocks.

So, if your memory granule size is 4MB, and your memory starts at
0xc0000000 physical, you're stuck with 768 unused sparsemem blocks
at the beginning of memory.  If it's 1MB, then you have 3072 unused
sparsemem blocks.  Each mem_section structure is 8 bytes, so that
could be 24K of zeros.

What we actually need is infrastructure in the kernel which can properly
handle sparse memory efficiently without causing such wastage.  If your
platform has four memory chunks, eg at 0xc0000000, 0xc4000000, 0xc8000000,
and 0xcc000000, then you want to build the kernel to tell it "there may
be four chunks with a 64MB offset, each chunk may be partially populated."

It seems that Sparsemem can't do that efficiently.