On Fri, Mar 20, 2020 at 10:48:48AM -0700, Matthew Wilcox wrote:
> On Fri, Mar 20, 2020 at 10:37:34AM -0700, Eric Biggers wrote:
> > On Fri, Mar 20, 2020 at 07:22:26AM -0700, Matthew Wilcox wrote:
> > > From: "Matthew Wilcox (Oracle)" <willy@xxxxxxxxxxxxx>
> > >
> > > Use the new readahead operation in ext4
> > >
> > > Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> > > Reviewed-by: William Kucharski <william.kucharski@xxxxxxxxxx>
> > > ---
> > > fs/ext4/ext4.h | 3 +--
> > > fs/ext4/inode.c | 21 +++++++++------------
> > > fs/ext4/readpage.c | 22 ++++++++--------------
> > > 3 files changed, 18 insertions(+), 28 deletions(-)
> > >
> >
> > Reviewed-by: Eric Biggers <ebiggers@xxxxxxxxxx>
> >
> > > + if (rac) {
> > > + page = readahead_page(rac);
> > > prefetchw(&page->flags);
> > > - list_del(&page->lru);
> > > - if (add_to_page_cache_lru(page, mapping, page->index,
> > > - readahead_gfp_mask(mapping)))
> > > - goto next_page;
> > > }
> >
> > Maybe the prefetchw(&page->flags) should be included in readahead_page()?
> > Most of the callers do it.
>
> I did notice that a lot of callers do that. I wonder whether it (still)
> helps or whether it's just cargo-cult programming. It can't possibly
> have helped before because we did list_del(&page->lru) as the very next
> instruction after prefetchw(), and they're in the same cacheline. It'd
> be interesting to take it out and see what happens to performance.
Yeah, it does look like the list_del() made the prefetchw() useless, so it
should just be removed. The prefetchw() dates all the way back to
mpage_readpages() being added in 2002, but even then the list_del() was
immediately afterwards, and 'flags' and 'lru' were in the same cache line in
'struct page' even then (assuming at least a 32-byte cache line size), so...
- Eric
