Wireless systems must be resilient to jamming attacks. Existing mitigation methods based on multi-antenna processing require knowledge of the jammer's transmit characteristics that may be difficult to acquire, especially for smart jammers that evade mitigation by transmitting only at specific instants. We propose a novel method to mitigate smart jamming attacks on the massive multi-user multiple-input multiple-output (MU-MIMO) uplink which does not require the jammer to be active at any specific instant. By formulating an optimization problem that unifies jammer estimation and mitigation, channel estimation, and data detection, we exploit that a jammer cannot change its subspace within a coherence interval. Theoretical results for our problem formulation show that its solution is guaranteed to recover the users' data symbols under certain conditions. We develop two efficient iterative algorithms for approximately solving the proposed problem formulation: MAED, a parameter-free algorithm which uses forward-backward splitting with a box symbol prior, and SO-MAED, which replaces the prior of MAED with soft-output symbol estimates that exploit the discrete transmit constellation and which uses deep unfolding to optimize algorithm parameters. We use simulations to demonstrate that the proposed algorithms effectively mitigate a wide range of smart jammers without a priori knowledge about the attack type.