Re: [PATCH 10/11] maps3: add /proc/kpagecount and /proc/kpageflags interfaces

[Matt Mackall <mpm@selenic.com>]

On Mon, Oct 15, 2007 at 04:34:57PM -0700, Dave Hansen wrote:
> On Mon, 2007-10-15 at 18:11 -0500, Matt Mackall wrote:
> > > Could we just have /proc/kpagereferenced?  Is there a legitimate need
> > > for other flags to be visible?
> > 
> > Referenced, dirty, uptodate, lru, active, slab, writeback, reclaim,
> > and buddy all look like they might be interesting to me from the point
> > of view of watching what's happening in the VM graphically in
> > real-time.
> 
> This is true, but it forces a lot of logic from the kernel to be run in
> userspace to figure out what is going on.  Looking at mainline today:
> 
> #define PG_reclaim              17      /* To be reclaimed asap */
> ...
> #define PG_readahead            PG_reclaim /* Reminder to do async read-ahead */
> 
> All of a sudden, to figure out which flag it actually is, we need to
> have all of the logic that the kernel does.  
> 
> Does this establish a fixed user<->kernel ABI that will keep us from
> doing this in the future:
> 
> -#define PG_slab                  7      /* slab debug (Suparna wants this) */
> +#define PG_slab                  14      /* slab debug (Suparna wants this) */
> 
> Or, even something like this:
> 
> -#define PageSlab(page)          test_bit(PG_slab, &(page)->flags)
> +#define PageSlab(page)          (!PageLRU(page) && !PageHighmem(page))

Yeah, there are a bunch of flags that aren't mutually exclusive and we
could probably recover a few.

> If we actually had several (or even still one file) that exposed this
> state, independent of the actual content of page->flags, I think we'd be
> better off.  I think that's the difference between a fun, super-useful
> debugging feature and one that can stay in mainline and have
> applications stay using it (without breaking) for a long time.
> 
> The flags you listed are things that I would imagine will always exist,
> logically.  But, we might not always have a specific page flag for pages
> under writeback or in the buddy list for that matter.  PG_buddy isn't
> that old.  Perhaps that would be better abstracted to something like
> page_in_main_allocator().

Perhaps we need something like:

flags = page->flags;
userflags = 
	  FLAG_BIT(USER_REFERENCED, flags & PG_referenced) |
	  ...

etc. for the flags we want to export. This will let us change to

	 FLAG_BIT(USER_SLAB, PageSlab(page)) |

if we make a virtual slab bit.

And it shows up in grep.

Unfortunately, i386 test_bit is an asm inline and not a macro so we
can't hope for the compiler to fold up a bunch of identity bit
mappings for us.


-- 
Mathematics is the supreme nostalgia of our time.