The general aim of the recommender system is to provide personalized suggestions to users, which is opposed to suggesting popular items. However, the normal training paradigm, i.e., fitting a recommender model to recover the user behavior data with pointwise or pairwise loss, makes the model biased towards popular items. This results in the terrible Matthew effect, making popular items be more frequently recommended and become even more popular. Existing work addresses this issue with Inverse Propensity Weighting (IPW), which decreases the impact of popular items on the training and increases the impact of long-tail items. Although theoretically sound, IPW methods are highly sensitive to the weighting strategy, which is notoriously difficult to tune. In this work, we explore the popularity bias issue from a novel and fundamental perspective -- cause-effect. We identify that popularity bias lies in the direct effect from the item node to the ranking score, such that an item's intrinsic property is the cause of mistakenly assigning it a higher ranking score. To eliminate popularity bias, it is essential to answer the counterfactual question that what the ranking score would be if the model only uses item property. To this end, we formulate a causal graph to describe the important cause-effect relations in the recommendation process. During training, we perform multi-task learning to achieve the contribution of each cause; during testing, we perform counterfactual inference to remove the effect of item popularity. Remarkably, our solution amends the learning process of recommendation which is agnostic to a wide range of models -- it can be easily implemented in existing methods. We demonstrate it on Matrix Factorization (MF) and LightGCN [20]. Experiments on five real-world datasets demonstrate the effectiveness of our method.