Re: [PATCH v1] erofs: enable SB_I_ALLOW_HSM to support fanotify pre access event

[Gao Xiang]

(+cc Amir, Jan)

On Wed, Oct 21, 2026 at 07:18:06AM +0000, Friendy.Su@xxxxxxxx wrote:
> Hi Gao,
> 
> Sorry for late reply.
> 
> We have considered fanotify in erofs-utils. I think it must be used with 
> file-backend mount. It actually utilizes FAN_PRE_ACCESS on ext4 or xfs. 
> Also file-backend mount can not support DAX mount.
> 
> But for our case, file-backend mount can not be used. Firstly, there is no 
> ext4 or xfs file system in actual environment to put a sparse erofs image 
> file. We have to mount a normal block device. Secondly, mount must be DAX.
> 
> The basic flow is file content is fetched from remote then filled into 
> local erofs file on the DAX mounted block device in FAN_PRE_ACCESS handler.
> 
> "/dev/loop0" is just an example, not our actual use case. The actual block 
> device is something like /dev/pmem0.
> 
> In order to make SB_I_ALLOW_HSM only set under block device backend, 
> we can do like this:
> 
> if (!erofs_is_fileio_mode(sbi)) {
>         /*
>          * Allow HSM to support fanotify pre access event
>          */

nit: single line /* Allow HSM to support fanotify pre access event */
     according to EROFS code style.

>         sb->s_iflags |= SB_I_ALLOW_HSM;
> }

I am fine with that as long as you limits SB_I_ALLOW_HSM **only if**
s_bdev is valid (IOWs, I hope we don't need to bother it with EROFS
SB_I_ALLOW_HSM + file-backed mounts; it aligns with other filesystems).

Also I think it would be better to add a testcase to check the basic
EROFS SB_I_ALLOW_HSM functionality in erofs-utils (see
experimental-tests). You don't need to add support for EROFS
SB_I_ALLOW_HSM directly in erofs-utils; but a basic functionality
testcase for this is needed.


BTW, for your use cases:

> mount must be DAX.

As for virtio-pmem, why not just use userfaultfd on the host for
virtio-pmem lazy pulling? for example,
https://github.com/firecracker-microvm/firecracker/issues/5740

Personally I still don't think monitoring FAN_PRE_ACCESS of EROFS for
this kind of usage is a good idea.

But anyway, it doesn't matter to much if it applies only to bdev/pmem.

> 
> How do you think?

Also please just adjust the patch subject to:

erofs: set SB_I_ALLOW_HSM for bdev-based mounts

Adjust the commit message accordingly to your use cases, and
explain file-backed mounts are not your cases in the commit message
as well.

And since it enables SB_I_ALLOW_HSM, it might be better to
Cc fsnotify maintainers in the following versions.

Thanks,
Gao Xiang

> 
> Best Regards
> Friendy
> 
> ________________________________________
> From: Gao Xiang <xiang@xxxxxxxxxx>
> Sent: Sunday, October 11, 2026 5:20
> To: Su, Friendy
> Cc: xiang@xxxxxxxxxx; chao@xxxxxxxxxx; linux-erofs@xxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; Mo, Yuezhang; Palmer, Daniel (SGC)
> Subject: Re: [PATCH v1] erofs: enable SB_I_ALLOW_HSM to support fanotify pre access event
> 
> Hi Friendy, On Sat, Oct 10, 2026 at 05: 40: 00PM +0800, Friendy Su wrote: > This patch makes erofs support a use case below: > > erofs file data are not available at local, but need be > filled at fanotify pre access event handler. 
> 
> 
> Hi Friendy,
> 
> On Sat, Oct 10, 2026 at 05:40:00PM +0800, Friendy Su wrote:
> > This patch makes erofs support a use case below:
> >
> > erofs file data are not available at local, but need be
> > filled at fanotify pre access event handler.
> >
> > E.g. when format erofs to blob mode, meta data image
> > is store locally, blob image size is big so it is stored remotely.
> > At erofs mount, a loop device based on a sparse image file is
> > used as blob device. The sparse image file size is equal to blob
> > image size. Like:
> >
> > $ sudo mount -o device=/dev/loop1 /dev/sdb1 /mnt/erofs/
> >
> > Then at file access, need retrieve data from remote, then fill
> > into sparse image at FAN_PRE_ACCESS event handler.
> >
> > Signed-off-by: Friendy Su <friendy.su@xxxxxxxx>
> > Reviewed-by: Yuezhang Mo <Yuezhang.Mo@xxxxxxxx>
> > Reviewed-by: Daniel Palmer <daniel.palmer@xxxxxxxx>
> > ---
> >  fs/erofs/super.c | 4 ++++
> >  1 file changed, 4 insertions(+)
> >
> > diff --git a/fs/erofs/super.c b/fs/erofs/super.c
> > index 0967aace9f7f..b92d3353319d 100644
> > --- a/fs/erofs/super.c
> > +++ b/fs/erofs/super.c
> > @@ -659,6 +659,10 @@ static int erofs_fc_fill_super(struct super_block *sb, struct fs_context *fc)
> >       sb->s_maxbytes = MAX_LFS_FILESIZE;
> >       sb->s_op = &erofs_sops;
> >
> > +     /*
> > +      * Allow HSM to support fanotify pre access event
> > +      */
> > +     sb->s_iflags |= SB_I_ALLOW_HSM;
> 
> Did you notice
> https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xiang/erofs-utils.git/tree/lib/backends/fanotify.c__;!!O7_YSHcmd9jp3hj_4dEAcyQ!3EhQW78Lg9hjdjMscHBj-V19jmMX6nlYdkPVufSmwLJEM5yM0uTCOLuqwilIL3MGLDqMV8LkNo1Chg$[git[.]kernel[.]org]
> ?
> 
> I am not sure why EROFS itself needs to set SB_I_ALLOW_HSM because EROFS
> is an immutable filesystem, its data should be considered as a golden
> one and it shouldn't be partially filled.
> 
> Instead, SB_I_ALLOW_HSM should be set on the backing filesystems (ext4,
> xfs, ...), and if I understand correctly, combining EROFS file-backed
> mounts (pls don't use loop for this) already implements what you said
> above.
> 
> And this functionality is already released in the latest erofs-utils
> and Nydus v3:
> https://urldefense.com/v3/__https://github.com/dragonflyoss/nydus/blob/v3/docs/fanotify.md__;!!O7_YSHcmd9jp3hj_4dEAcyQ!3EhQW78Lg9hjdjMscHBj-V19jmMX6nlYdkPVufSmwLJEM5yM0uTCOLuqwilIL3MGLDqMV8IXHHnVqQ$[github[.]com]
> 
> Thanks,
> Gao Xiang
> 
>