If the frequencies are different, the phase is "rolling."
No, it detects and indicates phase. The result is some DC level if the frequencies are truly equal, or a "rolling dc" level (which is typically a sine or triangle wave) if they are really different.
Classic (simple) phase detectors detect phase difference, not frequency difference. So if the two input frequencies are, say, 10 Hz apart, the pd output is a sine or triangle wave at 10 Hz. If the signals are the same frequency but differ in phase (timing) the pd will give a corresponding DC out.
Such a classic pd (a multiplier, diode mixer, xor gate, d-flipflop etc) when fed into the oscillator control loop will result in lock only if the phase change is relatively slow, which only happens if the frequencies are close. If the frequencies are too far apart, the pd output is a high-frequency waveform with no dc component that just confuses the vco. So a simple pll has a limited "acquisition range", even though it may track a wide range *once it gets into lock*.
There are complex phase detectors that are smart enough to realize that they are way out of lock, in which case they force the pd output in the right direction until they're close enough to work in true phase-detect mode. That solves the acquisition problem.
If a loop had a true frequency detector, any tiny analog error would result in the vco frequency being not exactly the input frequency. But a loop with a phase detector has zero longterm frequency error; the waveforms are locked.
John