Taking a VM fault is not expensive if the underlying page is already in core and can simply be mapped into the process, but it can become expensive if you take a whole lot of them on a regular basis. A good example of this is running a program such as ls(1) or ps(1) over and over again. If the program binary is mapped into memory but not mapped into the page table, then all the pages that will be accessed by the program will have to be faulted in every time the program is run. This is unnecessary when the pages in question are already in the VM Cache, so FreeBSD will attempt to pre-populate a process's page tables with those pages that are already in the VM Cache. One thing that FreeBSD does not yet do is pre-copy-on-write certain pages on exec. For example, if you run the ls(1) program while running vmstat 1 you will notice that it always takes a certain number of page faults, even when you run it over and over again. These are zero-fill faults, not program code faults (which were pre-faulted in already). Pre-copying pages on exec or fork is an area that could use more study.
A large percentage of page faults that occur are zero-fill faults. You can usually see this by observing the vmstat -s output. These occur when a process accesses pages in its BSS area. The BSS area is expected to be initially zero but the VM system does not bother to allocate any memory at all until the process actually accesses it. When a fault occurs the VM system must not only allocate a new page, it must zero it as well. To optimize the zeroing operation the VM system has the ability to pre-zero pages and mark them as such, and to request pre-zeroed pages when zero-fill faults occur. The pre-zeroing occurs whenever the CPU is idle but the number of pages the system pre-zeros is limited in order to avoid blowing away the memory caches. This is an excellent example of adding complexity to the VM system in order to optimize the critical path.