When trying to match really huge amount of files (several hundred of thousands), glob()'s stack overflows because it uses alloca() to allocate the entry records temporarily. The attached patch (which is not totally trivial so you might want to postpone that post glibc-2.5) aims to fix that.
Created attachment 1320 [details] The proposed patch
That patch is not acceptable. We cannot sacrifize speed just for the sake of some insane people. Yes, we shouldn't crash but this doesn't mean that reasonable requests should be slowed down. The correct solution would use alloca only if the limit determined by __libc_alloca_cutoff is exceeded.
Should be fixed in cvs with a different patch.