I highly recommend reading Howard Heys's Tutorial on Linear and Differential Cryptanalysis available online (google is your friend) for a detailed explanation for the SPN case.
In brief, if a linear expression $a\cdot x+ b\cdot S(x)$ is used then all the output wires that correspond to 1 bit in the vector $b$ turn on the Sboxes they are connected to in the next round. It's best to start with one active Sbox in the initial round to turn on as few Sboxes as possible in the next round (by choosing $a$ and $b$ so that $b$ has a minimum Hamming weight but at he linear expression has a large bias).
See also the answers to this question and this one.
AES turns on the maximum number of Sboxes in a different way, since there are no bit permutations but a diffusion matrix. See a nice explanation in this answer.