I usually work more with structural than functional analyses, so probably some other people on this list are better qualified to correct me and/or point to specific pipelines, etc. However, usually I think you would want to co-register the subject functional to the subject anatomical (i.e., opposite of first line) because FLIRT tends to work better with a hi-res image as the reference. You can always invert the matrix with convert_xfm to get the transform to go the other way. 

Then, I'm not sure what your images are filename-wise, but you would want to use the subject anatomical image (not the functional image) to co-register to MNI space. To ultimately get the subject functional into MNI space, you can use convert_xfm to combine the transforms.

So, overall, I think you'd want to do something like the following:

> flirt -in SUBJ_CENTRAL_FUNC_IMAGE -ref SUBJ_ANAT_IMAGE -out F2S -omat F2S.mat -dof 6

> flirt -in SUBJ_ANAT_IMAGE -ref MNI_2mm_Template -out SA2T -omat SA2T.mat

> convert_xfm -omat SF2T.mat -concat SA2T.mat F2S.mat

> flirt -in SUBJ_CENTRAL_FUNC_IMAGE -ref MNI_2mm_Template -out SF2T -applyxfm -init SF2T.mat

That being said, you should probably use the BBR cost function for FLIRT in the first step if your functional data is EPI (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FLIRT_BBR). Someone else would have to help with the specifics of that, though.

All the best,

Mike