My opinion is that it should be completely acceptable. Of course, a very specific question might be interesting to only a very few people, in which case it's not likely to attract more than a few votes. But I wouldn't take that as a sign that it's not acceptable.

If it is well motivated and enough background/reference is given, and if the paper is of research level, then such a question should be acceptable. If it looks like a homework question in disguise without motivation, there will be objections.

I think a good example looks like the following: "I am researching X using the paper Y, where on page 6 the author says Z. I am surprised because I thought Z' was true, and it seems that the context (specifically the assumption of W), would prohibit Z because (something W implies that is far from Z). Can someone help clear this up for me?"

The example above can be modified to show what work you have done in attempting to resolve your own difficulty, and where you got stuck. You might include pertinent online references that are freely available if the paper itself is not.

Gerhard "Ask Me About System Design" Paseman, 2011.11.15

I think asking a question about something you don't understand in a paper is perfectly fine. What I don't like, however, is a question that relies on the reader having a copy of the paper. If you ask such a question, it's fine to mention that your question is arises from a specific point in a paper but you should try to write the question so that it is self-contained and can be understood and answered by someone without having to look at the paper at all. I concede that this isn't always possible, but you should try to come as close as you can to doing this.

When questions are unacceptable, they get closed. If your previous questions haven't been closed, that's prima facie evidence that they are acceptable, regardless of the number of votes they get.

I think one of the interesting differences between MathOverflow and StackOverflow is that, on MathOverflow, the most appropriate questions are the ones that don't get very many upvotes, because one of the features of mathematics is that important technical questions are only understood by a few people.