blk-cgroup: Properly propagate the iostat update up the hierarchy

During a cgroup_rstat_flush() call, the lowest level of nodes are flushed
first before their parents. Since commit 3b8cc62 ("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: 3b8cc62 ("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]>
Reviewed-by: Ming Lei <[email protected]>
Acked-by: Tejun Heo <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Jens Axboe <[email protected]>

Author: Waiman-Long

block/blk-cgroup.c

@@ -322,6 +322,7 @@ 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);
@@ -1046,6 +1047,8 @@ static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu)
 		smp_mb();
 
 		WRITE_ONCE(bisc->lqueued, false);
+		if (bisc == &blkg->iostat)
+			goto propagate_up; /* propagate up to parent only */
 
 		/* fetch the current per-cpu values */
 		do {
@@ -1055,10 +1058,24 @@ 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: