Re: [PATCH v2 3/3] writeback: use a per-sb counter to drain inode wb switches at umount

[Christian Brauner]

On Sun, 17 May 2026 22:21:32 +0800, Baokun Li <libaokun@xxxxxxxxxxxxxxxxx> wrote:
> diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
> index 9ae290547eb2..4282fcfe027b 100644
> --- a/fs/fs-writeback.c
> +++ b/fs/fs-writeback.c
> @@ -1215,38 +1209,30 @@ int cgroup_writeback_by_id(u64 bdi_id, int memcg_id,
> [ ... skip 38 lines ... ]
> -		 * in-flight switcher that already passed the SB_ACTIVE check
> -		 * to finish queueing its work, so flush_workqueue() below
> -		 * will then drain it.
> -		 */
> -		synchronize_rcu();
> +	while (atomic_read(&sb->s_isw_nr_in_flight)) {

Would probably be a bit nicer to just wait until the count is zero?
We have that pattern for inode_dio_wait(). So something like (untested):

static bool cgroup_writeback_finished(const struct super_block *sb)
{
	if (atomic_read(&sb->s_isw_nr_in_flight) == 0)
		return true;

   	flush_workqueue(isw_wq);
	return false;
}

void cgroup_writeback_wait(const struct super_block *sb)
{
	wait_var_event(&sb->s_isw_nr_in_flight, cgroup_writeback_finished(sb));
}

void cgroup_writeback_finish(const struct super_block *sb)
{
	if (atomic_dec_and_test(&sb->s_isw_nr_in_flight))
		wake_up_var(&sb->s_isw_nr_in_flight);
}

-- 
Christian Brauner <brauner@xxxxxxxxxx>