Oh yes, and the "& extent threshold {voxels}" doesn't perform any statistical test, you can enter any arbitrary number if you want to. However, if you go with the default "0" you can then look at the bottom of the results table, FWEc tells you the size of the smallest, still sig. cluster, e.g. 112 voxels. You can then reload the SPM.mat, take the same uncorrected voxel threshold (e.g. .001), followed by "112" for "& extent threshold {voxels}". This way only sig. clusters are displayed in the glassbrain and reported in the results table.
If you go with correction on voxel/peak level there's no need for an additional correction on cluster level, but often, people go with an arbitrary extent threshold of a few voxels (e.g. 5, 10, 20) nonetheless. Their implicit assumption is that while the voxels are significant by itself (due to surviving the voxel threshold), they are only relevant if they form a "cluster" of several connected voxels.
For the statistical concepts behind voxel and cluster thresholds I'd suggest to look at the corresponding literature, but e.g. these sites http://mindhive.mit.edu/book/export/html/90 and http://support.brainvoyager.com/functional-analysis-statistics/40-cluster-tresholding/87-users-guide-the-multiple-comparisons-problem.html might give a first overview.
