Summary: The framework of budget-feasible mechanism design studies procurement auctions where the auctioneer (buyer) aims to maximize his valuation function subject to a hard budget constraint. We study the problem of designing truthful mechanisms that have good approximation guarantees and never pay the participating agents (sellers) more than the budget. We focus on the case of general (non-monotone) submodular valuation functions and derive the first truthful, budget-feasible, and \(O(1)\)-approximation mechanisms that run in polynomial time in the value query model, for both offline and online auctions. Prior to our work, the only \(O(1)\)-approximation mechanism known for non-monotone submodular objectives required an exponential number of value queries. At the heart of our approach lies a novel greedy algorithm for non-monotone submodular maximization under a knapsack constraint. Our algorithm builds two candidate solutions simultaneously (to achieve a good approximation), yet ensures that agents cannot jump from one solution to the other (to implicitly enforce truthfulness). The fact that in our mechanism the agents are not ordered according to their marginal value per cost allows us to appropriately adapt these ideas to the online setting as well. To further illustrate the applicability of our approach, we also consider the case where additional feasibility constraints are present, for example, at most \(k\) agents can be selected. We obtain \(O(p)\)-approximation mechanisms for both monotone and non-monotone submodular objectives, when the feasible solutions are independent sets of a \(p\)-system. With the exception of additive valuation functions, no mechanisms were known for this setting prior to our work. Finally, we provide lower bounds suggesting that, when one cares about nontrivial approximation guarantees in polynomial time, our results are, asymptotically, the best possible.