On Thu, Feb 12, 2026 at 1:09 PM Waiman Long <[email protected]> wrote:
>
> When the audit subsystem is enabled, it can do a lot of get_fs_pwd()
> calls to get references to fs->pwd and then releasing those references
> back with path_put() later. That may cause a lot of spinlock contention
> on a single pwd's dentry lock because of the constant changes to the
> reference count when there are many processes on the same working
> directory actively doing open/close system calls. This can cause
> noticeable performance regresssion when compared with the case where
> the audit subsystem is turned off especially on systems with a lot of
> CPUs which is becoming more common these days.
>
> A simple and elegant solution to avoid this kind of performance
> regression is to add a common pool of extra fs->pwd references inside
> the fs_struct. When a caller needs a pwd reference, it can borrow one
> from pool, if available, to avoid an explicit path_get(). When it is
> time to release the reference, it can put it back into the common pool
> if fs->pwd isn't changed before without doing a path_put(). We still
> need to acquire the fs's spinlock, but fs_struct is more distributed
> and it is less common to have many tasks sharing a single fs_struct.
>
> A new set of get_fs_pwd_pool/put_fs_pwd_pool() APIs are introduced
> with this patch to enable other subsystems to acquire and release
> a pwd reference from the common pool without doing unnecessary
> path_get/path_put().
>
> Besides fs/fs_struct.c, the copy_mnt_ns() function of fs/namespace.c is
> also modified to properly handle the extra pwd references, if available.
>
> Signed-off-by: Waiman Long <[email protected]>
> ---
> fs/fs_struct.c | 26 +++++++++++++++++++++-----
> fs/namespace.c | 8 ++++++++
> include/linux/fs_struct.h | 30 +++++++++++++++++++++++++++++-
> 3 files changed, 58 insertions(+), 6 deletions(-)
...
> diff --git a/fs/namespace.c b/fs/namespace.c
> index c58674a20cad..a2323ba84d76 100644
> --- a/fs/namespace.c
> +++ b/fs/namespace.c
> @@ -40,6 +41,33 @@ static inline void get_fs_pwd(struct fs_struct *fs, struct
> path *pwd)
> read_sequnlock_excl(&fs->seq);
> }
>
> +/* Acquire a pwd reference from the pwd_refs pool, if available */
> +static inline void get_fs_pwd_pool(struct fs_struct *fs, struct path *pwd)
> +{
> + read_seqlock_excl(&fs->seq);
> + *pwd = fs->pwd;
> + if (fs->pwd_refs)
> + fs->pwd_refs--;
> + else
> + path_get(pwd);
> + read_sequnlock_excl(&fs->seq);
> +}
> +
> +/* Release a pwd reference back to the pwd_refs pool, if appropriate */
> +static inline void put_fs_pwd_pool(struct fs_struct *fs, struct path *pwd)
> +{
> + bool put = false;
> +
> + read_seqlock_excl(&fs->seq);
> + if ((fs->pwd.dentry == pwd->dentry) && (fs->pwd.mnt == pwd->mnt))
> + fs->pwd_refs++;
> + else
> + put = true;
> + read_sequnlock_excl(&fs->seq);
> + if (put)
> + path_put(pwd);
> +}
This is a nitpick, and perhaps I'm missing something, but I think you
could skip the local 'put' boolean by setting 'pwd' to NULL in the
pool case, e.g.
static inline void put_fs_pwd_pool(fs, pwd)
{
read_seqlock_excl(&fs)
if (fs == pwd) {
fs->pwd_refs++
pwd = NULL
}
read_sequnlock_excl(&fs)
if (pwd)
path_put(pwd)
}
--
paul-moore.com
