I don't think that adding 20,000 covariates is a good idea. You should do a PCA analysis on the CNV data and then use the first X components (where X is a number between 2 and 10, let's say) as covariates. PCA can used to correct for population structure (doing PCA on the genotype data) or to incorporate unaccounted bias in genotypes (e.g. I also suggest that you use as a covariate the PCA of your expression data). There is lot of discussion on the use of PCA in GWAS. You will find useful references here.