Re: [RFC] /dev/anon

[Andrew Morton <akpm@osdl.org>]

Davide Libenzi <davidel@xmailserver.org> wrote:
>
> On Wed, 14 Jan 2004, Jeff Dike wrote:
> 
> > > I thought your goal was to release memory 
> > > to the host, that's why I proposed sys_madvise(MADV_DONTNEED).
> > 
> > It is, I want memory released immediately as though it were clean, and
> > MADV_DONTNEED doesn't help.
> 
> Strange, I didn't notice this before. If you look at the comment in 
> mm/madvise.c:madvise_dontneed, it advertises that dirty pages are actually 
> thrown away (that would be what you're actually looking for). But if you 
> go down to zap_page_range -> unmap_vmas -> unmap_page_range -> 
> zap_pmd_range -> zap_pte_range, if the page is dirty, set_page_dirty -> 
> __set_page_dirty_buffers pushes the page into the mapping dirty pages list 
> and __mark_inode_dirty push the inode inside the superblock dirty list. So 
> the comment seems to be wrong (I also verified this with a simple program, 
> and pages are actually flushed).
> 

We cannot invalidate the pages due to MADV_DONTNEED: there may be
freshly-allocated, unwritten file blocks associated with them.  You'd have
to be playing games with write() amd MAP_SHARED to do this.