There was Cross Validated post on this issue a while back. As my brother likes to joke (without malice, actually a kinda deep thought)
Perfection offends the Buddha
And the message indicates that the model produced a "perfect separation." According to the post, model results are still produced, but coefficients are inflated. This may arise either from a sampling artifact or the population characteristics itself. If the former, penalized methods can be used; if the former, then it is what it is.
I haven't got yet to the chapter in Applied Logistic Regression 3rd Edition by David W. Hosmer Jr., Stanley Lemeshow and Rodney X. Sturdivant (2009) that discusses this yet. Cross Validated is probably a better forum for such a relatively advanced question. Even so, it may catch the eye of one or more of the real statisticians who appear from time to time.