diff --git a/drivers/infiniband/hw/mlx5/cq.c b/drivers/infiniband/hw/mlx5/cq.c
index 367a71bc5f4b628b97f5bc70e0e1efd910b100ba..3dec3de903b751054c68bd2f407d3c44bd1c4ef6 100644
--- a/drivers/infiniband/hw/mlx5/cq.c
+++ b/drivers/infiniband/hw/mlx5/cq.c
@@ -330,6 +330,22 @@ static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev,
 		dump_cqe(dev, cqe);
 }
 
+static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64,
+			   u16 tail, u16 head)
+{
+	u16 idx;
+
+	do {
+		idx = tail & (qp->sq.wqe_cnt - 1);
+		if (idx == head)
+			break;
+
+		tail = qp->sq.w_list[idx].next;
+	} while (1);
+	tail = qp->sq.w_list[idx].next;
+	qp->sq.last_poll = tail;
+}
+
 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf)
 {
 	mlx5_frag_buf_free(dev->mdev, &buf->frag_buf);
@@ -368,7 +384,7 @@ static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
 }
 
 static void sw_comp(struct mlx5_ib_qp *qp, int num_entries, struct ib_wc *wc,
-		    int *npolled, int is_send)
+		    int *npolled, bool is_send)
 {
 	struct mlx5_ib_wq *wq;
 	unsigned int cur;
@@ -383,10 +399,16 @@ static void sw_comp(struct mlx5_ib_qp *qp, int num_entries, struct ib_wc *wc,
 		return;
 
 	for (i = 0;  i < cur && np < num_entries; i++) {
-		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
+		unsigned int idx;
+
+		idx = (is_send) ? wq->last_poll : wq->tail;
+		idx &= (wq->wqe_cnt - 1);
+		wc->wr_id = wq->wrid[idx];
 		wc->status = IB_WC_WR_FLUSH_ERR;
 		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
 		wq->tail++;
+		if (is_send)
+			wq->last_poll = wq->w_list[idx].next;
 		np++;
 		wc->qp = &qp->ibqp;
 		wc++;
@@ -473,6 +495,7 @@ static int mlx5_poll_one(struct mlx5_ib_cq *cq,
 		wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
 		idx = wqe_ctr & (wq->wqe_cnt - 1);
 		handle_good_req(wc, cqe64, wq, idx);
+		handle_atomics(*cur_qp, cqe64, wq->last_poll, idx);
 		wc->wr_id = wq->wrid[idx];
 		wq->tail = wq->wqe_head[idx] + 1;
 		wc->status = IB_WC_SUCCESS;
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index bb78142bca5eaf8a137785d144520ed391a3bbf4..f3bdbd5e50964efda778bbd69e0663ac55389de2 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -288,6 +288,7 @@ struct mlx5_ib_wq {
 	unsigned		head;
 	unsigned		tail;
 	u16			cur_post;
+	u16			last_poll;
 	void			*cur_edge;
 };
 
diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index 957f3a52589bef9026fadcee8c5601fc6aa0c2d8..c2967982f02c6863cddf4364a743932421230056 100644
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -3775,6 +3775,7 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
 		qp->sq.cur_post = 0;
 		if (qp->sq.wqe_cnt)
 			qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
+		qp->sq.last_poll = 0;
 		qp->db.db[MLX5_RCV_DBR] = 0;
 		qp->db.db[MLX5_SND_DBR] = 0;
 	}