Assume that a regular cardinal $\kappa > \omega_1$ has the property that for every $\kappa$.c.c. forcing notion $\mathbb{P}$ and name of set of ordinals $x \in V^{\mathbb{P}}$ smaller than $\kappa$, there is a subforcing $\mathbb{Q}$ of cardinality less than $\kappa$ that already decides the value of $x$.
We will show that $\kappa$ is Mahlo. The proof shows that in order to conclude that $\kappa$ is Mahlo, we only need that for every $\kappa$.c.c. forcing that adding a real there is a regular subforcing of smaller cardinality that adds this real.
Assume, for a start, that at every uncountable non-weakly compact regular cardinal there is a Suslin tree.
Let start by showing that in this case $\kappa$ must be inaccessible. Assume that $\lambda$ is the first cardinal such that $2^\lambda \geq \kappa$ (so $2^{<\lambda} < \kappa$).
Let $T$ be a Suslin tree at $\kappa$ and let choose an almost disjoint family of subsets of $\lambda$ with size $\kappa$, $\mathcal{A} \in V$. Now, forcing a branch to $T$ is $\kappa$.c.c.. The forcing that codes it into a subset of $\lambda$ is $2^{<\lambda}$-centered and therefore $\kappa$.c.c. in $V^T$ (note that if $\lambda$ is singular, this forcing is not closed at all). So the iteration is $\kappa$.c.c., but we can't have a small sub-forcing that adds the code for the branch, since it must add the branch itself.
Now, let drop the assumption the there are many Suslin trees. Instead, we use the fact that for every successor cardinal $\lambda$ there are two $\lambda$.c.c. posets of size $\lambda$ such that their product that is not $\lambda$.c.c. (and replace "forcing with Suslin tree" by forcing with one of them and get that the small subforcing can't add a the subset that codes a large antichain in the other one). I think that Assaf Rinot proved that this holds for every successor of regular cardinal and Shelah proved that this holds for every successor of singular (see Complicated Colorings).
Now, let show that $\kappa$ must be Mahlo.
Assume that $\kappa$ is non-Mahlo inaccessible. Using the coding method of Jensen and Solovay (from "Some applications of almost disjoint sets") we can code a generic of Levi collapse $Col(\omega,<\kappa)$ into a real with some additional $c.c.c.$ forcing. The iteration is again $\kappa$.c.c., but since $V[x] \models \kappa = \aleph_1$, we can't add this real with any forcing smaller than $\kappa$.