In the first part, we show an exponential algorithm for deciding circularity. In the second part, we show that this the problem is coNP-hard. In the third part, we show that every circular language is a union of languages of the form $r^+$ (here $r$ could be the empty regexp); the union is not necessarily disjoint. In the fourth part, we exhibit a circular language which cannot be written as a disjoint sum $\sum r_i^+$.
Edit: Incorporated some corrections following Mark's comments. In particular, my earlier claims that circularity is coNP-complete or NP-hard are corrected.
Edit: Corrected normal form from $\sum r_i^*$ to $\sum r_i^+$. Exhibited an "inherently ambiguous" language.
Continuing Peter Taylor's comment, here's how to decide (extremely inefficiently) whether a language is circular given its DFA. Construct a new DFA whose states are $n$-tuples of the old states. This new DFA runs $n$ copies of the old DFA in parallel.
If the language is not circular then there is a word $w$ such that if we run it through the DFA repeatedly, starting with the initial state $s_0$, then we get states $s_1,\ldots,s_n$ such that $s_1$ is accepting but one of the other ones is not accepting (if all of them are accepting then then the sequence $s_0,\ldots,s_n$ must cycle so that $w^*$ is always in the language). In other words, we have a path from $s_0,\ldots,s_{n-1}$ to $s_1,\ldots,s_n$ where $s_1$ is accepting but one of the others is not accepting. Conversely, if the language is circular then that cannot happen.
So we've reduced the problem to a simple directed reachability test (just check all possible "bad" $n$-tuples).
The problem of circularity is coNP-hard. Suppose we're given a 3SAT instance with $n$ variables $\vec{x}$ and $m$ clauses $C_1,\ldots,C_m$. We can assume that $n = m$ (add dummy variables) and that $n$ is prime (otherwise find a prime between $n$ and $2n$ using AKS primality testing, and add dummy variables and clauses).
Consider the following language: "the input is not of the form $\vec{x}_1 \cdots \vec{x}_n$ where $\vec{x}_i$ is a satisfying assignment for $C_i$". It is easy to construct an $O(n^2)$ DFA for this language. If the language is not circular then there is a word $w$ in the language, some power of which is not in the language. Since the only words not in the language have length $n^2$, $w$ must be of length $1$ or $n$. If it is of length $1$, consider $w^n$ instead (it is still in the language), so that $w$ is in the language and $w^n$ is not in the language. The fact that $w^n$ is not in the language means that $w$ is a satisfying assignment.
Conversely, any satisfying assignment translates to a word proving the non-circularity of the language: the satisfying assignment $w$ belongs to the language but $w^n$ does not. Thus the language is circular iff the 3SAT instance is unsatisfiable.
In this part, we discuss a normal form for circular languages. Consider some DFA for a circular language $L$. A sequence $C = C_0,\ldots$ is real if $C_0 = s$ (the initial state), all other states are accepting, and $C_i = C_j$ implies $C_{i+1} = C_{j+1}$. Thus every real sequence is eventually periodic, and there are only finitely many real sequences (since the DFA has finitely many states).
We say that a word behaves according to $C$ if the word takes the DFA from state $c_i$ to state $c_{i+1}$, for all $i$. The set of all such words $E(C)$ is regular (the argument is similar to the first part of this answer). Note that $E(C)$ is a subset of $L$.
Given a real sequence $C$, define $C^k$ to be the sequence $C^k(t) = C(kt)$. The sequence $C^k$ is also real. Since there are only finitely many different sequences $C^k$, the language $D(C)$ which is the union of all $E(C^k)$ is also regular.
We claim that $D(C)$ has the property that if $x,y \in D(C)$ then $xy \in D(C)$. Indeed, suppose that $x \in C^k$ and $y \in C^l$. Then $xy \in C^{k+l}$. Thus $D(C) = D(C)^+$ can be written in the form $r^+$ for some regular expression $r$.
Every word $w$ in the language corresponds to some real sequence $C$, i.e. there exists a real sequence $C$ that $w$ behaves according to. Thus $L$ is the union of $D(C)$ over all real sequence $C$. Therefore every circular language has a representation of the form $\sum r_i^+$. Conversely, every such language is circular (trivially).
Consider the circular language $L$ of all words over $a,b$ that contain either an even number or $a$'s or an even number of $b$'s (or both). We show that it cannot be written as a disjoint sum $\sum r_i^+$; by "disjoint" we mean that $r_i^+ \cap r_j^+ = \varnothing$.
Let $N_i$ be the size of the some DFA for $r_i^+$, and $N > \max N_i$ be some odd integer. Consider $x = a^N b^{N!}$. Since $x \in L$, $x \in r_i^+$ for some $i$. By the pumping lemma, we can pump a prefix of $x$ of length at most $N$. Thus $r_i^+$ generates $z = a^{N!} b^{N!}$. Similarly, $y = a^{N!} b^N$ is generated by some $r_j^+$, which also generates $z$. Note that $i \neq j$ since $xy \notin L$. Thus the representation cannot be disjoint.