On Wed, Jun 18, 2025 at 04:37:36PM -0700, Song Liu wrote:
> Existing kernfs_xattr_get() locks iattr_mutex, so it cannot be used in
> RCU critical sections. Introduce __kernfs_xattr_get(), which reads xattr
> under RCU read lock. This can be used by BPF programs to access cgroupfs
> xattrs.
>
> Signed-off-by: Song Liu <song@xxxxxxxxxx>
> ---
> fs/kernfs/inode.c | 14 ++++++++++++++
> include/linux/kernfs.h | 2 ++
> 2 files changed, 16 insertions(+)
>
> diff --git a/fs/kernfs/inode.c b/fs/kernfs/inode.c
> index b83054da68b3..0ca231d2012c 100644
> --- a/fs/kernfs/inode.c
> +++ b/fs/kernfs/inode.c
> @@ -302,6 +302,20 @@ int kernfs_xattr_get(struct kernfs_node *kn, const char *name,
> return simple_xattr_get(&attrs->xattrs, name, value, size);
> }
>
> +int __kernfs_xattr_get(struct kernfs_node *kn, const char *name,
> + void *value, size_t size)
> +{
> + struct kernfs_iattrs *attrs;
> +
> + WARN_ON_ONCE(!rcu_read_lock_held());
> +
> + attrs = rcu_dereference(kn->iattr);
> + if (!attrs)
> + return -ENODATA;
Hm, that looks a bit silly. Which isn't your fault. I'm looking at the
kernfs code that does the xattr allocations and I think that's the
origin of the silliness. It uses a single global mutex for all kernfs
users thus serializing all allocations for kernfs->iattr. That seems
crazy but maybe I'm missing a good reason.
I'm appending a patch to remove that mutex. @Greg, @Tejun, can you take
a look whether that makes sense to you. Then I can take that patch and
you can build yours on top of the series and I'll pick it all up in one
go.
You should then just use READ_ONCE(kn->iattr) or the
kernfs_iattrs_noalloc(kn) helper in your kfunc.
