Do anyone of implemented methods where one can permute genotype data to gain an experiment wide P value but also incorporate Eigenvectors to control for population structure. So keep the eigenvectors with the genotype data but shuffle the phenotype label.
Supposing you are going to perform a typical case-control association study, you could include the effect of population structure in different ways.
I propose a couple:
i) do PCA analysis then include eigenvectors as covariates in logistic regression models, and finally compute your empirical family-wise P val permuting 1000-10000 times your regressions (so you randomize only your phenotypes and preserve genotype and pop structure !)
ii) instead of PCA compute Fst (a measure of population structure related to heterozygosity) and use the lambda value to correct the significance of your associations. You can do this with or without adoption of permutation testing