On 5/15/24 10:30, Waiman Long wrote:
> During a cgroup_rstat_flush() call, the lowest level of nodes are flushed
> first before their parents. Since commit 3b8cc6298724 ("blk-cgroup:
> Optimize blkcg_rstat_flush()"), iostat propagation was still done to
> the parent. Grandparent, however, may not get the iostat update if the
> parent has no blkg_iostat_set queued in its lhead lockless list.
>
> Fix this iostat propagation problem by queuing the parent's global
> blkg->iostat into one of its percpu lockless lists to make sure that
> the delta will always be propagated up to the grandparent and so on
> toward the root blkcg.
>
> Note that successive calls to __blkcg_rstat_flush() are serialized by
> the cgroup_rstat_lock. So no special barrier is used in the reading
> and writing of blkg->iostat.lqueued.
>
> Fixes: 3b8cc6298724 ("blk-cgroup: Optimize blkcg_rstat_flush()")
> Reported-by: Dan Schatzberg <[email protected]>
> Closes: https://lore.kernel.org/lkml/ZkO6l%2FODzadSgdhC@dschatzberg-fedora-PF3DHTBV/
> Signed-off-by: Waiman Long <[email protected]>
> ---
> block/blk-cgroup.c | 19 ++++++++++++++++++-
> 1 file changed, 18 insertions(+), 1 deletion(-)
>
> diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
> index 059467086b13..2a7624c32a1a 100644
> --- a/block/blk-cgroup.c
> +++ b/block/blk-cgroup.c
> @@ -323,6 +323,7 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk,
> blkg->q = disk->queue;
> INIT_LIST_HEAD(&blkg->q_node);
> blkg->blkcg = blkcg;
> + blkg->iostat.blkg = blkg;
> #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
> spin_lock_init(&blkg->async_bio_lock);
> bio_list_init(&blkg->async_bios);
> @@ -1025,6 +1026,8 @@ static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu)
> unsigned int seq;
>
> WRITE_ONCE(bisc->lqueued, false);
> + if (bisc == &blkg->iostat)
> + goto propagate_up; /* propagate up to parent only */
>
> /* fetch the current per-cpu values */
> do {
> @@ -1034,10 +1037,24 @@ static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu)
>
> blkcg_iostat_update(blkg, &cur, &bisc->last);
>
> +propagate_up:
> /* propagate global delta to parent (unless that's root) */
> - if (parent && parent->parent)
> + if (parent && parent->parent) {
> blkcg_iostat_update(parent, &blkg->iostat.cur,
> &blkg->iostat.last);
> + /*
> + * Queue parent->iostat to its blkcg's lockless
> + * list to propagate up to the grandparent if the
> + * iostat hasn't been queued yet.
> + */
> + if (!parent->iostat.lqueued) {
> + struct llist_head *plhead;
> +
> + plhead = per_cpu_ptr(parent->blkcg->lhead, cpu);
> + llist_add(&parent->iostat.lnode, plhead);
> + parent->iostat.lqueued = true;
> + }
> + }
> }
> raw_spin_unlock_irqrestore(&blkg_stat_lock, flags);
> out:
Note that this patch will conflict with Ming's patch
(https://lore.kernel.org/linux-block/[email protected]/).
Will send out an updated version later on to synchronize with his patch.
Cheers,
Longman