rt.representation theory - Which is the correct universal enveloping algebra in positive characteristic?

This is an extension of this question about symmetric algebras in positive characteristic. The title is also a bit tongue-in-cheek, as I am sure that there are multiple "correct" answers.

Let $mathfrak g$ be a Lie algebra over $k$. One can define the universal enveloping algebra $Umathfrak g$ in terms of the adjunction: $text{Hom}_{rm LieAlg}(mathfrak g, A) = text{Hom}_{rm AsAlg}(Umathfrak g, A)$ for any associative algebra $A$. Then it's easy enough to check that $Umathfrak g$ is the quotient of the free tensor algebra generated by $mathfrak g$ by the ideal generated by elements of the form $xy - yx - [x,y]$. (At least, I'm sure of this when the characteristic is not $2$. I don't have a good grasp in characteristic $2$, though, because I've heard that the correct notion of "Lie algebra" is different.)

But there's another good algebra, which agrees with $Umathfrak g$ in characteristic $0$. Namely, if $mathfrak g$ is the Lie algebra of some algebraic group $G$, then I think that the algebra of left-invariant differential operators is some sort of "divided-power" version of $Umathfrak g$.

So, am I correct that this notions diverge in positive characteristic? If so, does the divided-power algebra have a nice generators-and-relations description? More importantly, which rings are used for what?