Re: Review of Thomas's >2GB ext2fs proposal

[Neal H. Walfield]

At Tue, 17 Aug 2004 06:07:02 -0400,
Neal H. Walfield <neal@cs.uml.edu> wrote:
> 
> At 17 Aug 2004 02:51:37 -0700,
> Thomas Bushnell BSG wrote:
> > 
> > "Neal H. Walfield" <neal@cs.uml.edu> writes:
> > 
> > > > Just a small techical correction: we remove mapping only when we are 
> > > > remapping page to contain another block.  So when page is evicted, it 
> > > > continues to map the same disk block until ext2fs decides that we want 
> > > > this page to hold another block (only if page is still evicted, of
> > > > course).
> > > 
> > > Well, I think we ought to drain the mapping cache when the page is
> > > evicted.  This keeps memory usage lower.  Further, relative to the
> > > cost of reading from disk, adding an entry to the hashes, etc. is
> > > cheap.
> > 
> > I'm confused.  It sounded like Ognyan was saying that the page is
> > still assigned to the same area.  But it is necessary to vm_free when
> > the kernel tells you to page out.  It is unfriendly (and very bad for
> > performance) to hold on to the page.

I was slightly confused as well.  The page is and remains being
unmapped in my proposal by an munmap in libpager/data-return.c.  What
Ogi has not done, I think, is to remove the association between the
virtual address and the disk blocks which I think ought to be done.

Neal