From mboxrd@z Thu Jan  1 00:00:00 1970
From: Miklos Szeredi <miklos@szeredi.hu>
Date: Wed, 25 Mar 2020 16:54:43 +0100
Subject: [Cluster-devel] [PATCH v10 24/25] fuse: Convert from readpages
	to readahead
In-Reply-To: <20200325153228.GB22483@bombadil.infradead.org>
References: <20200323202259.13363-1-willy@infradead.org>
	<20200323202259.13363-25-willy@infradead.org>
	<CAJfpegu7EFcWrg3bP+-2BX_kb52RrzBCo_U3QKYzUkZfe4EjDA@mail.gmail.com>
	<20200325120254.GA22483@bombadil.infradead.org>
	<CAJfpegshssCJiA8PBcq2XvBj3mR8dufHb0zWRFvvKKv82VQYsw@mail.gmail.com>
	<20200325153228.GB22483@bombadil.infradead.org>
Message-ID: <CAJfpegtrhGamoSqD-3Svfj3-iTdAbfD8TP44H_o+HE+g+CAnCA@mail.gmail.com>
List-Id: <cluster-devel.redhat.com>
To: cluster-devel.redhat.com
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: 7bit

On Wed, Mar 25, 2020 at 4:32 PM Matthew Wilcox <willy@infradead.org> wrote:
>
> On Wed, Mar 25, 2020 at 03:43:02PM +0100, Miklos Szeredi wrote:
> > >
> > > -       while ((page = readahead_page(rac))) {
> > > -               if (fuse_readpages_fill(&data, page) != 0)
> > > +               nr_pages = min(readahead_count(rac), fc->max_pages);
> >
> > Missing fc->max_read clamp.
>
> Yeah, I realised that.  I ended up doing ...
>
> +       unsigned int i, max_pages, nr_pages = 0;
> ...
> +       max_pages = min(fc->max_pages, fc->max_read / PAGE_SIZE);
>
> > > +               ia = fuse_io_alloc(NULL, nr_pages);
> > > +               if (!ia)
> > >                         return;
> > > +               ap = &ia->ap;
> > > +               __readahead_batch(rac, ap->pages, nr_pages);
> >
> > nr_pages = __readahead_batch(...)?
>
> That's the other bug ... this was designed for btrfs which has a fixed-size
> buffer.  But you want to dynamically allocate fuse_io_args(), so we need to
> figure out the number of pages beforehand, which is a little awkward.  I've
> settled on this for the moment:
>
>         for (;;) {
>                struct fuse_io_args *ia;
>                 struct fuse_args_pages *ap;
>
>                 nr_pages = readahead_count(rac) - nr_pages;
>                 if (nr_pages > max_pages)
>                         nr_pages = max_pages;
>                 if (nr_pages == 0)
>                         break;
>                 ia = fuse_io_alloc(NULL, nr_pages);
>                 if (!ia)
>                         return;
>                 ap = &ia->ap;
>                 __readahead_batch(rac, ap->pages, nr_pages);
>                 for (i = 0; i < nr_pages; i++) {
>                         fuse_wait_on_page_writeback(inode,
>                                                     readahead_index(rac) + i);
>                         ap->descs[i].length = PAGE_SIZE;
>                 }
>                 ap->num_pages = nr_pages;
>                 fuse_send_readpages(ia, rac->file);
>         }
>
> but I'm not entirely happy with that either.  Pondering better options.

I think that's fine.   Note how the original code possibly
over-allocates the the page array, because it doesn't know the batch
size beforehand.  So this is already better.

>
> > This will give consecutive pages, right?
>
> readpages() was already being called with consecutive pages.  Several
> filesystems had code to cope with the pages being non-consecutive, but
> that wasn't how the core code worked; if there was a discontiguity it
> would send off the pages that were consecutive and start a new batch.
>
> __readahead_batch() can't return fewer than nr_pages, so you don't need
> to check for that.

That's far from obvious.

I'd put a WARN_ON at least to make document the fact.

Thanks,
Miklos