Re: [PATCH] Skip memory holes in FLATMEM when reading /proc/pagetypeinfo (resend)

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

On Thu, 2008-08-21 at 14:28 +0100, Mel Gorman wrote:
> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
> index 4b8acd2..70dba16 100644
> --- a/arch/arm/Kconfig
> +++ b/arch/arm/Kconfig
> @@ -810,6 +810,11 @@ config OABI_COMPAT
>  	  UNPREDICTABLE (in fact it can be predicted that it won't work
>  	  at all). If in doubt say Y.
>  
> +config ARCH_FLATMEM_HAS_HOLES
> +	bool
> +	default y
> +	depends on FLATMEM
> +
>  config ARCH_DISCONTIGMEM_ENABLE
>  	bool
>  	default (ARCH_LH7A40X && !LH7A40X_CONTIGMEM)
> diff --git a/mm/vmstat.c b/mm/vmstat.c
> index b0d08e6..98aa882 100644
> --- a/mm/vmstat.c
> +++ b/mm/vmstat.c
> @@ -516,9 +516,26 @@ static void pagetypeinfo_showblockcount_print(struct seq_file *m,
>  			continue;
>  
>  		page = pfn_to_page(pfn);
> +#ifdef CONFIG_ARCH_FLATMEM_HAS_HOLES
> +		/*
> +		 * Ordinarily, memory holes in flatmem still have a valid
> +		 * memmap for the PFN range. However, an architecture for
> +		 * embedded systems (e.g. ARM) can free up the memmap backing
> +		 * holes to save memory on the assumption the memmap is
> +		 * never used.

I'm not sure where this assumption is coming from.  Have you taken a
look at the ARM code?

> 		               The page_zone linkages are then broken even
> +		 * though pfn_valid() returns true. Skip the page if the
> +		 * linkages are broken. Even if this test passed, the impact
> +		 * is that the counters for the movable type are off but
> +		 * fragmentation monitoring is likely meaningless on small
> +		 * systems.
> +		 */
> +		if (page_zone(page) != zone)
> +			continue;
> +#endif
>  		mtype = get_pageblock_migratetype(page);
>  
> -		count[mtype]++;
> +		if (mtype < MIGRATE_TYPES)
> +			count[mtype]++;
>  	}

Is that really worth an #ifdef?  It is in code that isn't in a hot path,
and the page_zone() should be a repetitive operation on a structure that
is already (or will soon be) in the cpu cache.  

Ugh, but as I think about it, this is going to be a much more widespread
problem.  Looking at the arm code, the memory for the mem_map[]e *does*
get freed back into the bootmem allocator:

free_memmap(int node, unsigned long start_pfn, unsigned long end_pfn)
{
...
        if (pg < pgend)
                free_bootmem_node(NODE_DATA(node), pg, pgend - pg);
}

Doesn't that mean that we can get random gunk in the middle of the
mem_map[] when someone else allocates and uses this memory?

We have quite a few things across the kernel that iterate over pfn
ranges and require pfn_valid() to be working. 

Can we talk ARM into converting over to sparsemem?  A plain sparsemem
port for MIPS that got coded up recently was ~25 lines of arch code.  A
bit more if they decide to do vmemmap.  There's even a nice bit for
Documentation/ about what was done:

http://article.gmane.org/gmane.linux.ports.mips.general/21248/match=mips+sparsemem

-- Dave