Regarding the implementation question, maybe you could tell us more about your hardware: ARM family (cortex M3 or M4?), clock rate...
A fixed-point implementation of the Goertzel algorithm is a dozen instructions, so this would indeed be doable with a sample rate of 1 MHz on a low-end ARM micro (say STM32F1 or LPC1343) clocked at 72 MHz (assuming you don't have much overhead for reading your ADC, eg. by letting the DMA subsystem handle this).
- You might hit the upper range of your ADC sample rate, and/or will have little margin to design your anti-aliasing filter (say 105kHz is the upper frequency of interest... You can sample at 210kHz but will need the steepest anti-aliasing filter... Or at 800kHz and have more margin for the transition band of your anti-aliasing filter).
- This still looks like a waste of CPU cycles, especially if you are only interested in a narrow band of the signal. Keep in mind that if the only thing that interests you is the 99kHz to 101kHz band, you actually need a sample rate of just 4kHz to capture all the useful information - provided you have a mean of modulating the input signal by a 100kHz sine wave.
In such situations, and if you control the target hardware, a bit of pre-processing in the analog domain could be handy, such as:
- Frequency shifting your signal by 100kHz so the band of interest becomes the base-band and so that you can sample it at a low sample-rate (requires a sine oscillator, a multiplier, and a low-pass filter).
- Applying a narrow band-pass filter in the area of interest + rectifier + low-pass filter - this would measure the energy of the frequency band of interest.