Re: [PATCH 32/32] NFS: Convert readpage to readahead and use netfs_readahead for fscache

[David Wysochanski]

On Mon, Jan 25, 2021 at 8:37 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
>
> For Subject: s/readpage/readpages/
>
Fixed

> On Mon, Jan 25, 2021 at 09:37:29PM +0000, David Howells wrote:
> > +int __nfs_readahead_from_fscache(struct nfs_readdesc *desc,
> > +                              struct readahead_control *rac)
>
> I thought you wanted it called ractl instead of rac?  That's what I've
> been using in new code.
>
Fixed

> > -     dfprintk(FSCACHE, "NFS: nfs_getpages_from_fscache (0x%p/%u/0x%p)\n",
> > -              nfs_i_fscache(inode), npages, inode);
> > +     dfprintk(FSCACHE, "NFS: nfs_readahead_from_fscache (0x%p/0x%p)\n",
> > +              nfs_i_fscache(inode), inode);
>
> We do have readahead_count() if this is useful information to be logging.
>
Right, I used it elsewhere so I'll add here as well.

> > +static inline int nfs_readahead_from_fscache(struct nfs_readdesc *desc,
> > +                                          struct readahead_control *rac)
> >  {
> > -     if (NFS_I(inode)->fscache)
> > -             return __nfs_readpages_from_fscache(ctx, inode, mapping, pages,
> > -                                                 nr_pages);
> > +     if (NFS_I(rac->mapping->host)->fscache)
> > +             return __nfs_readahead_from_fscache(desc, rac);
> >       return -ENOBUFS;
> >  }
>
> Not entirely sure that it's worth having the two functions separated any more.
>
Yeah it's questionable so I'll collapse.  I'll also do that with
nfs_readpage_from_fscache().

> >       /* attempt to read as many of the pages as possible from the cache
> >        * - this returns -ENOBUFS immediately if the cookie is negative
> >        */
> > -     ret = nfs_readpages_from_fscache(desc.ctx, inode, mapping,
> > -                                      pages, &nr_pages);
> > +     ret = nfs_readahead_from_fscache(&desc, rac);
> >       if (ret == 0)
> >               goto read_complete; /* all pages were read */
> >
> >       nfs_pageio_init_read(&desc.pgio, inode, false,
> >                            &nfs_async_read_completion_ops);
> >
> > -     ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
> > +     while ((page = readahead_page(rac))) {
> > +             ret = readpage_async_filler(&desc, page);
> > +             put_page(page);
> > +     }
>
> I thought with the new API we didn't need to do this kind of thing
> any more?  ie no matter whether fscache is configured in or not, it'll
> submit the I/Os.
>

We don't. This patchset was only intended as a stepping stone to get the
netfs API accepted with minimal invasiveness in NFS.

I have another patch which will unconditionally call netfs API but I
didn't post it. Since I'm not an NFS maintainer, and maintainer's didn't
weigh in on the approach, I opted to go with the least invasive approach.

There's an NFS "remote bakeathon" coming up at the end of Feb.
That would probably be a good time to get further testing on NFS
unconditionally calling the netfs API, and we should be able to
cover things like any performance concerns, etc.