Re: [PATCH v3 1/2] vfs: add rcu-based find_inode variants for iget ops

[Dave Chinner]

On Tue, Jun 11, 2024 at 03:13:45PM +0200, Mateusz Guzik wrote:
> On Tue, Jun 11, 2024 at 3:04 PM Christian Brauner <brauner@xxxxxxxxxx> wrote:
> >
> > On Tue, Jun 11, 2024 at 01:40:37PM +0200, Mateusz Guzik wrote:
> > > On Tue, Jun 11, 2024 at 12:50:11PM +0200, Jan Kara wrote:
> > > > On Tue 11-06-24 12:16:31, Mateusz Guzik wrote:
> > > > > +/**
> > > > > + * ilookup5 - search for an inode in the inode cache
> > > >       ^^^ ilookup5_rcu
> > > >
> > >
> > > fixed in my branch
> > >
> > > > > + * @sb:          super block of file system to search
> > > > > + * @hashval:     hash value (usually inode number) to search for
> > > > > + * @test:        callback used for comparisons between inodes
> > > > > + * @data:        opaque data pointer to pass to @test
> > > > > + *
> > > > > + * This is equivalent to ilookup5, except the @test callback must
> > > > > + * tolerate the inode not being stable, including being mid-teardown.
> > > > > + */
> > > > ...
> > > > > +struct inode *ilookup5_nowait_rcu(struct super_block *sb, unsigned long hashval,
> > > > > +         int (*test)(struct inode *, void *), void *data);
> > > >
> > > > I'd prefer wrapping the above so that it fits into 80 columns.
> > > >
> > >
> > > the last comma is precisely at 80, but i can wrap it if you insist
> > >
> > > > Otherwise feel free to add:
> > > >
> > > > Reviewed-by: Jan Kara <jack@xxxxxxx>
> > > >
> > >
> > > thanks
> > >
> > > I'm going to wait for more feedback, tweak the commit message to stress
> > > that this goes from 2 hash lock acquires to 1, maybe fix some typos and
> > > submit a v4.
> > >
> > > past that if people want something faster they are welcome to implement
> > > or carry it over the finish line themselves.
> >
> > I'm generally fine with this but I would think that we shouldn't add all
> > these helpers without any users. I'm not trying to make this a chicken
> > and egg problem though. Let's get the blessing from Josef to convert
> > btrfs to that *_rcu variant and then we can add that helper. Additional
> > helpers can follow as needed? @Jan, thoughts?
> 
> That's basically v1 of the patch (modulo other changes like EXPORT_SYMBOL_GPL).
> 
> It only has iget5_locked_rcu for btrfs and ilookup5_rcu for bcachefs,
> which has since turned out to not use it.
> 
> Jan wanted iget5_locked_rcu to follow the iget5_locked in style, hence
> I ended up with 3 helpers instead of 1.

We don't need any extra APIs if you just convert the inode cache to
using hash-bl rather than just converting lookups to RCU. Everyone
automatically gets the all extra scalability improvements and no new
interfaces are needed at all.

-Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx