Re: 2.5.73-mm2

[Nikita Danilov]

William Lee Irwin III writes:
 > On Tue, 1 Jul 2003, William Lee Irwin III wrote:
 > >> Well, I was mostly looking for getting handed back 0 when lowmem is
 > >> empty; I actually did realize they didn't give entirely accurate counts
 > >> of free lowmem pages.
 > 
 > On Tue, Jul 01, 2003 at 12:08:03PM +0100, Hugh Dickins wrote:
 > > I'm not pleading for complete accuracy, but nr_free_buffer_pages()
 > > will never hand back 0 (if your system managed to boot).
 > > It's a static count of present_pages (adjusted), not of
 > > free pages.  Or am I misreading nr_free_zone_pages()?
 > 
 > You're right. Wow, that's even more worse than I suspected.
 > 

Another thing is that if one boots with mem=X, nr_free_pagecache_pages()
returns X. However part of X (occupied by kernel image, etc) is not part
of any zone. As a result, zone actually contains fewer pages than
reported by nr_free_pagecache_pages(). With X small enough (comparable
with kernel image size, for example) this can confuse
balance_dirty_pages() enough so that throttling would never start, and
VM will oom_kill().

 > 
 > -- wli

Nikita.