Re: [PATCH 17/31] fuse: use an unrestricted backing device with iomap pagecache io

["Darrick J. Wong" <djwong@kernel.org>]

On Tue, Jan 27, 2026 at 10:04:28AM -0800, Joanne Koong wrote:
> On Mon, Jan 26, 2026 at 6:09 PM Darrick J. Wong <djwong@kernel.org> wrote:
> >
> > On Mon, Jan 26, 2026 at 05:35:05PM -0800, Joanne Koong wrote:
> > > On Mon, Jan 26, 2026 at 3:55 PM Darrick J. Wong <djwong@kernel.org> wrote:
> > > >
> > > > On Mon, Jan 26, 2026 at 02:03:35PM -0800, Joanne Koong wrote:
> > > > > On Tue, Oct 28, 2025 at 5:49 PM Darrick J. Wong <djwong@kernel.org> wrote:
> > > > > >
> > > > > > From: Darrick J. Wong <djwong@kernel.org>
> > > > > >
> > > > > > With iomap support turned on for the pagecache, the kernel issues
> > > > > > writeback to directly to block devices and we no longer have to push all
> > > > > > those pages through the fuse device to userspace.  Therefore, we don't
> > > > > > need the tight dirty limits (~1M) that are used for regular fuse.  This
> > > > > > dramatically increases the performance of fuse's pagecache IO.
> > > > > >
> > > > > > Signed-off-by: "Darrick J. Wong" <djwong@kernel.org>
> > > > > > ---
> > > > > >  fs/fuse/file_iomap.c |   21 +++++++++++++++++++++
> > > > > >  1 file changed, 21 insertions(+)
> > > > > >
> > > > > >
> > > > > > diff --git a/fs/fuse/file_iomap.c b/fs/fuse/file_iomap.c
> > > > > > index 0bae356045638b..a9bacaa0991afa 100644
> > > > > > --- a/fs/fuse/file_iomap.c
> > > > > > +++ b/fs/fuse/file_iomap.c
> > > > > > @@ -713,6 +713,27 @@ const struct fuse_backing_ops fuse_iomap_backing_ops = {
> > > > > >  void fuse_iomap_mount(struct fuse_mount *fm)
> > > > > >  {
> > > > > >         struct fuse_conn *fc = fm->fc;
> > > > > > +       struct super_block *sb = fm->sb;
> > > > > > +       struct backing_dev_info *old_bdi = sb->s_bdi;
> > > > > > +       char *suffix = sb->s_bdev ? "-fuseblk" : "-fuse";
> > > > > > +       int res;
> > > > > > +
> > > > > > +       /*
> > > > > > +        * sb->s_bdi points to the initial private bdi.  However, we want to
> > > > > > +        * redirect it to a new private bdi with default dirty and readahead
> > > > > > +        * settings because iomap writeback won't be pushing a ton of dirty
> > > > > > +        * data through the fuse device.  If this fails we fall back to the
> > > > > > +        * initial fuse bdi.
> > > > > > +        */
> > > > > > +       sb->s_bdi = &noop_backing_dev_info;
> > > > > > +       res = super_setup_bdi_name(sb, "%u:%u%s.iomap", MAJOR(fc->dev),
> > > > > > +                                  MINOR(fc->dev), suffix);
> > > > > > +       if (res) {
> > > > > > +               sb->s_bdi = old_bdi;
> > > > > > +       } else {
> > > > > > +               bdi_unregister(old_bdi);
> > > > > > +               bdi_put(old_bdi);
> > > > > > +       }
> > > > >
> > > > > Maybe I'm missing something here, but isn't sb->s_bdi already set to
> > > > > noop_backing_dev_info when fuse_iomap_mount() is called?
> > > > > fuse_fill_super() -> fuse_fill_super_common() -> fuse_bdi_init() does
> > > > > this already before the fuse_iomap_mount() call, afaict.
> > > >
> > > > Right.
> > > >
> > > > > I think what we need to do is just unset BDI_CAP_STRICTLIMIT and
> > > > > adjust the bdi max ratio?
> > > >
> > > > That's sufficient to undo the effects of fuse_bdi_init, yes.  However
> > > > the BDI gets created with the name "$major:$minor{-fuseblk}" and there
> > > > are "management" scripts that try to tweak fuse BDIs for better
> > > > performance.
> > > >
> > > > I don't want some dumb script to mismanage a fuse-iomap filesystem
> > > > because it can't tell the difference, so I create a new bdi with the
> > > > name "$major:$minor.iomap" to make it obvious.  But super_setup_bdi_name
> > > > gets cranky if s_bdi isn't set to noop and we don't want to fail a mount
> > > > here due to ENOMEM so ... I implemented this weird switcheroo code.
> > >
> > > I see. It might be useful to copy/paste this into the commit message
> > > just for added context. I don't see a better way of doing it than what
> > > you have in this patch then since we rely on the init reply to know
> > > whether iomap should be used or not...
> >
> > I'll do that.  I will also add that as soon as any BDI is created, it
> > will be exposed to userspace in sysfs.  That means that running the code
> > from fuse_bdi_init in reverse will not necessarily produce the same
> > results as a freshly created BDI.
> >
> > > If the new bdi setup fails, I wonder if the mount should just fail
> > > entirely then. That seems better to me than letting it succeed with
> >
> > Err, which new bdi setup?  If fuse-iomap can't create a new BDI, it will
> > set s_bdi back to the old one and move on.  You'll get degraded
> > performance, but that's not the end of the world.
> 
> I was thinking from the user POV, I'd rather the whole mount fail
> (which it seems like would only be a transient failure, eg running out
> of memory) and I retry, than it work but have writes potentially run
> 10x slower (10x comes from the benchmarks Jingbo saw in [1])

Hrmm.  The difficulty of preallocating the iomap bdi is that I think
you'd have to do it in fuse_bdi_init, which occurs before the kernel has
even seen the reply to FUSE_INIT and therefore knows if the fuse server
even cares about iomap.

> > > strictlimiting enforced, especially since large folios will be enabled
> > > for fuse iomap. [1] has some numbers for the performance degradations
> > > I saw for writes with strictlimiting on and large folios enabled.
> >
> > If fuse_bdi_init can't set up a bdi it will fail the mount.
> >
> > That said... from reading [1], if strictlimiting is enabled with large
> > folios, then can we figure out what is the effective max folio size and
> > lower it to that?
> 
> I'm not really sure how we figure that out, unless I guess we try to
> do it experimentally? The throttling logic for this is in
> balance_dirty_pages().

Oh, I see, it's a dynamic limit.  I don't know how to deal with that
either.

> > > Speaking of strictlimiting though, from a policy standpoint if we
> > > think strictlimiting is needed in general in fuse (there's a thread
> > > from last year [1] about removing strict limiting), then I think that
> >
> > (did you mean [2] here?)
> 
> Ah yes sorry, I had meant [2].
> >
> > > would need to apply to iomap as well, at least for unprivileged
> > > servers.
> >
> > iomap requires a privileged server, FWIW.
> 
> Oh right, I forgot iomap only runs with privileges enabled. In that
> case, I think that makes the whole strictlimiting thing a lot simpler
> then. imo for privileged servers, we should get rid of strictlimiting
> entirely. Though I'm not sure how MIklos feels about that.

<nod>

--D

> Thanks,
> Joanne
> 
> >
> > > [1] https://lore.kernel.org/linux-fsdevel/CAJnrk1bwat_r4+pmhaWH-ThAi+zoAJFwmJG65ANj1Zv0O0s4_A@mail.gmail.com/
> > > [2] https://lore.kernel.org/linux-fsdevel/20251010150113.GC6174@frogsfrogsfrogs/T/#ma34ff5ae338a83f8b2e946d7e5332ea835fa0ff6
> > >
> > > >
> > > > > This is more of a nit, but I think it'd also be nice if we
> > > > > swapped the ordering of this patch with the previous one enabling
> > > > > large folios, so that large folios gets enabled only when all the bdi
> > > > > stuff for it is ready.
> > > >
> > > > Will do, thanks for reading these patches!
> > > >
> > > > Also note that I've changed this part of the patchset quite a lot since
> > > > this posting; iomap configuration is now a completely separate fuse
> > > > command that gets triggered after the FUSE_INIT reply is received.
> > >
> > > Great, I'll look at your upstream tree then for this part.
> >
> > Ok.
> >
> > --D
> >
> > > Thanks,
> > > Joanne
> > >
> > > >
> > > > --D
> > > >
> > > > > Thanks,
> > > > > Joanne
> > > > >
> > > > > >
> > > > > >         /*
> > > > > >          * Enable syncfs for iomap fuse servers so that we can send a final
> > > > > >
> > > > >
> > >
>