Also, if the vm area is large, pages might be thrashing, requiring the kernel to do loads of work swapping things in and out and changing the pagetable struct.
I doubt that that's any different than allocating memory assuming your working set would be the same. Largely it should be unless the mmap design makes it easier to facilitate a random access pattern in your data structure that you would think twice about if you actually had to issue the reads yourself.
The bigger issue is that madvise & friends risk cross table shootdowns since Linux might not have relevant support for DONT_NEED other than unmapping (vs scheduling an unmapping to happen lazily). However, that isn't necessarily the case with Windows.
Linux just takes that as a hint to unmap immediately and this results in a cross tlb shoot down. Cross tlb shoot downs are crazy expensive. Windows might have better support for doing it lazily and not paying the cost.