Re: [PATCH] fuse: expose backing files in the daemon's fd table

[Amir Goldstein <amir73il@gmail.com>]

[removing irrelevant lists]

Hi Matan!

On Fri, Apr 24, 2026 at 02:45:53PM +0000, Matan Cohen wrote:
> FUSE passthrough backing files are currently not installed in the
> daemon's fd table, making them invisible to lsof(8) and /proc/<pid>/fd.
> This complicates debugging and is the reason passthrough currently
> requires CAP_SYS_ADMIN.
> 
> Install each backing file into the daemon's fd table on BACKING_OPEN
> and close it on BACKING_CLOSE.  Add an fd field to struct fuse_backing
> to carry the installed fd number.
> 
> BACKING_CLOSE always runs in the daemon's ioctl context, so close_fd()
> targets the correct fd table directly.
> 
> To handle fds that remain open at connection teardown (e.g. on umount
> or connection abort without an explicit BACKING_CLOSE), store a
> reference to the daemon's task_struct in fc->daemon_task on the first
> BACKING_OPEN.  During teardown, fuse_backing_files_free() schedules a
> task_work callback on the daemon task that calls close_fd() from the
> daemon's own context.  If the daemon has already exited (task_work_add
> returns -ESRCH), its fd table was already cleaned up on exit.
> 

I am sorry but both the approach and the implementation are wrong for
several different reasons, so I will not start to list them.

Please follow this discussion:
https://lore.kernel.org/linux-fsdevel/CAJfpegvaCTxS_wC6EGFfh3Gim5DEgOtuju=_=qCsouzkCRvJog@mail.gmail.com/

It does not end with a definite decision on the best way to export
backing files to lsof, but it does pave the general direction the
the solution should take.

> This is a prerequisite for relaxing the CAP_SYS_ADMIN requirement on
> FUSE passthrough; a follow-up patch will drop the capability check now
> that backing fds are accountable via /proc/<pid>/fd.

This is one prerequisite, not all of them.
See "Resource Accounting and Visibility" section in
Documentation/filesystems/fuse/fuse-passthrough.rst

Please CC me and the new fuse-devel list on future attempts to address
this issue.  There is no need to CC LKML and selftest maintainers.

Thanks,
Amir.