Most of your definitions are correct. Linkage studies are performed when you have pedigrees of related individals and a phenotype (such as breast cancer) that is present in some but not all of the family members. These individuals could be humans or animals; linkage in humans is studied using existing families, so no breeding is involved. For each locus, you tabulate cases where parents and children who do or don't show the phenotype also have the same allele. Linkage studies are the most powerful approach when studying highly penetrant phenotypes, which means that if you have the allele you have a strong probability of exhibiting the phenotype. They can identiy rare alleles that are present in small numbers of families, usualy due to a founder mutation. Linkage is how you find an allele such as the mutations in BRCA1 associated with breast cancer.

Association studies are used when you don't have pedigrees; here the statistical test is a logistic regression or a related test for trends. They work when the phenotype has much lower penetrance; they are in fact more powerful than linkage analysis in those cases, provided you have enough informative cases and matched controls. Association studies are how you find common, low penetrance alleles such as the variations in FGFR2 that confer small increases in breast cancer susceptibility.

In The Old Days, neither association tests nor linkage tests were "genome-wide"; there wasn't a technically feasable or affordable way to test the whole genome at once. Studies were often performed at various levels of resolution as the locus associated with the phenotype was refined. Studies were often performed with a small number of loci chosen because of prior knowledge or hunches. Now the most common way to perform these studies in humans is to use SNP chips that measure hundreds of thousands of loci spread across the whole genome, thus the name GWAS. The reason you're testing "the whole genome" without sequencing the whole genome of each case and control is an important point that is a separate topic; if you don't yet know how this works, start with the concept of Linkage Disequilibrium. I haven't encountered the term GWLS myself, but I think it's safe to say that this is just a way to indicate that the whole genome was queried for linkage to a phenotype.

Just to complement David's answer:

Genomic Convergence of Genome-wide Investigations for Complex Traits

A lot of good examples in the references of this paper.

But, I disagree on the pedigree part. Neither type of studies needs a pedigree to proceed. Both can use a pedigree in calculations to help sort out the correlations with the phenotype Anyway, it's much harder to use GWLS in the same way GWAS because of genetic drift, demographic stochasticity (including migration), selection and genetic linkage. You could have a highly penetrant allele with essentially no LD, just by the action of drift. This is quite probable in loci with low effetive population size (Ne). GWAS somewhat mitigates this effect by using an Ne-independent definition of SNPs and related polymorfisms. It's also important to stress that the differences posted by David apply very well when the background scenario is Mendelian, assumption that is becoming increasingly complicated.

I can say that GWLS and GWAS address pretty different questions. GWLS if properly conducted can access the genetic architeture of a trait and answer if a given configuration of alleles is related to a trait. You also can answer questions about selection and drift. So, variants don't need to be independent. On the other hand, you must be quite precise about your haplotypes and their blocks. By the way, math can be very complicated in this case.

GWAS are much more simple to conduct. You simply count and test each variant independently, then in pairs, etc. No need to worry with genome structure, population genetics and similar stuff.

So, both approaches do have an overlap in applications. But, it's clear that you can use them in very different problems.