probably nobody has a firm idea whether overally all of the effects are linear or not, but why shouldn't they be?

Incidence of some cancers is described fairly well by a multistage model of carcinogenesis, which posits that a cell has to go through multiple pre-cancerous stages before it becomes a cancer. Suppose the model is true. If smoking accelerates the transition at multiple steps on the path to carcinogenesis, then smoking's effect on cumulative cancer incidence can be super-linear.

Not my clever idea, sadly, I got it from another paper.