On Tue, Oct 23, 2007 at 11:30:35AM +0200, Andi Kleen wrote:
> On Tue, Oct 23, 2007 at 05:04:14PM +1000, David Chinner wrote:
> > > You mean like vmap() could record the pages passed to it in the
> > > area->pages
> > > array, and we walk and release than in __vunmap() like it already does
> > > for vfree()?
> > >
> > > If we did this, it would probably be best to pass a page release function
> > > into the vmap or vunmap call - we'd need page_cache_release() called on
> > > the page rather than __free_page()....
> > >
> > > The solution belongs behind the vmap/vunmap interface, not in XFS....
> >
> > Lightly tested(*) patch that does this with lazy unmapping
> > below for comment.
>
> Thanks
>
> >
> > (*) a) kernel boots, b) made an XFS filesystem with 64k directory
> > blocks, created ~100,000 files in a directory to get a wide btree
> > (~1700 blocks, still only a single level) and run repeated finds
> > across it dropping caches in between. Each traversal maps and
> > unmaps every btree block.
>
> Hmm, the __free_page -> page_cache_release() change in vfree() would
> have been simpler wouldn't it?
Yes, it would, but I tried to leave vmalloc doing the same thing as
it was before. I think that it would be safe simply to call put_page()
directly in the __vunmap() code and drop all the release function
passing, but I don't know enough about that code to say for certain.
I'll spin up another patch tomorrow that does this and see how it goes.
> But if it works it is fine.
Seems to - it's passed XFSQA with 64k directories and a bunch of
dirstress workouts as well.
Nick, Jeremy, (others?) any objections to this approach to solve
the problem?
Cheers,
Dave.
--
Dave Chinner
Principal Engineer
SGI Australian Software Group
|