diff --git a/block/blk-merge.c b/block/blk-merge.c
index 7efa8c3e2b72..8054410536fd 100644
--- a/block/blk-merge.c
+++ b/block/blk-merge.c
@@ -7,6 +7,8 @@
 #include <linux/bio.h>
 #include <linux/blkdev.h>
 #include <linux/scatterlist.h>
+#include <linux/blkdev.h>
+#include <linux/blk-cgroup.h>
 
 #include <trace/events/block.h>
 
@@ -486,6 +488,9 @@ static inline int ll_new_hw_segment(struct request_queue *q,
 	if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
 		goto no_merge;
 
+	if (!blk_cgroup_mergeable(req, bio))
+		goto no_merge;
+
 	if (blk_integrity_merge_bio(q, req, bio) == false)
 		goto no_merge;
 
@@ -609,6 +614,9 @@ static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
 	if (total_phys_segments > queue_max_segments(q))
 		return 0;
 
+	if (!blk_cgroup_mergeable(req, next->bio))
+		return 0;
+
 	if (blk_integrity_merge_rq(q, req, next) == false)
 		return 0;
 
@@ -843,6 +851,10 @@ bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
 	if (rq->rq_disk != bio->bi_disk || req_no_special_merge(rq))
 		return false;
 
+	/* don't merge across cgroup boundaries */
+	if (!blk_cgroup_mergeable(rq, bio))
+		return false;
+
 	/* only merge integrity protected bio into ditto rq */
 	if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
 		return false;
diff --git a/include/linux/blk-cgroup.h b/include/linux/blk-cgroup.h
index 6d766a19f2bb..8f1be8b49391 100644
--- a/include/linux/blk-cgroup.h
+++ b/include/linux/blk-cgroup.h
@@ -21,6 +21,7 @@
 #include <linux/blkdev.h>
 #include <linux/atomic.h>
 #include <linux/kthread.h>
+#include <linux/blkdev.h>
 
 /* percpu_counter batch for blkg_[rw]stats, per-cpu drift doesn't matter */
 #define BLKG_STAT_CPU_BATCH	(INT_MAX / 2)
@@ -844,6 +845,21 @@ static inline void blkcg_use_delay(struct blkcg_gq *blkg)
 		atomic_inc(&blkg->blkcg->css.cgroup->congestion_count);
 }
 
+/**
+ * blk_cgroup_mergeable - Determine whether to allow or disallow merges
+ * @rq: request to merge into
+ * @bio: bio to merge
+ *
+ * @bio and @rq should belong to the same cgroup and their issue_as_root should
+ * match. The latter is necessary as we don't want to throttle e.g. a metadata
+ * update because it happens to be next to a regular IO.
+ */
+static inline bool blk_cgroup_mergeable(struct request *rq, struct bio *bio)
+{
+	return rq->bio->bi_blkg == bio->bi_blkg &&
+		bio_issue_as_root_blkg(rq->bio) == bio_issue_as_root_blkg(bio);
+}
+
 static inline int blkcg_unuse_delay(struct blkcg_gq *blkg)
 {
 	int old = atomic_read(&blkg->use_delay);
@@ -947,6 +963,7 @@ static inline struct request_list *blk_rq_rl(struct request *rq) { return &rq->q
 
 static inline bool blkcg_bio_issue_check(struct request_queue *q,
 					 struct bio *bio) { return true; }
+static inline bool blk_cgroup_mergeable(struct request *rq, struct bio *bio) { return true; }
 
 #define blk_queue_for_each_rl(rl, q)	\
 	for ((rl) = &(q)->root_rl; (rl); (rl) = NULL)