FUSE Passthrough Mode Merged For Linux 6.9

[trapexit]

Is it enabled by default?
Do we have to do something to use it?

You can't "enable it by default". It is a kernel feature that a FUSE server which happens to be a passthrough / union / overlay filesystem can leverage. It is only relevant for that usecase and requires the server to manage setting up the file descriptor to be passed into the kernel.

[trapexit]

FUSE is used for user-space filesystem implementations, it has nothing to do with implementations of ntfs in kernel-space.

Also... this passthrough feature couldn't be used by ntfs-3g anyway. People seem to have a fundamental misunderstanding of FUSE and/or this feature.

[intelfx]

Also... this passthrough feature couldn't be used by ntfs-3g anyway. People seem to have a fundamental misunderstanding of FUSE and/or this feature.

It's just the Phoronix syndrome. People see words "kernel" and "speedup" and immediately go "HOW CAN I USE DIS TO MAKE MY PEECEE GO FASTA". It isn't a misunderstanding because there was never any understanding to begin with.

(Please pardon the caps.)

[stiiixy]

.....

[Joe2021]

It won't effect actual filesystems in any way. What it does is pass through file handles from an underlying filesystem, such as squashfs or ext4fs, to the process opening the file from a FUSE filesystem mount, such as overlayfs. That allows process IO calls directly on VFS to the underlying filesystem without going through FUSE any more.
[...]
It will also have some security problems if it was used with containers, which is sad, as that would otherwise be a really awesome use case.

Thanks for the explanation. What implications are to expect with regards to virtiofsd? Do you have an estimation regarding that?

[boeroboy]

Interesting. I'm in a GCP VM having upgraded to the rawhide 6.9rc kernel testing with GCSFUSE and I don't think it's working by default.

$ uname -r
6.9.0-0.rc0.20240318gitf6cef5f8c37f.10.fc41.x86_64
$ mount | grep /mnt
MYBUCKET on /mnt type fuse.gcsfuse (rw,nosuid,nodev,relatime,user_id=1001,group_id=1002,default_permissions)

I see there's an IOCTL that must be passed to /dev/fuse but not sure if the fuse client needs to be updated for that or what. Performance difference is negligible and CPU load is still reasonably higher than it should be if FUSE Passthrough is working. Anyone test it in 6.9rc yet?

Kernel config definiteliy has it:

$ grep -ni passthrough /boot/config-6.9.0-0.rc0.20240318gitf6cef5f8c37f.10.fc41.x86_64  
8894:# CONFIG_IOMMU_DEFAULT_PASSTHROUGH is not set
9963:CONFIG_FUSE_PASSTHROUGH=y

​

EDIT: It seems most FUSE filesystems are inelegible as it requires passing a file descriptor to passthrough so that's a shame.