[openib-general] mthca FMR correctness (and memory windows)

[Roland Dreier]

    Thomas> Really? The idea of FMR, I thought, was to preallocate the
    Thomas> TPT entry up front (alloc_fmr), and then populate it with
    Thomas> the lightweight fmr api (map_phys_fmr).

Yes, you do save that cost with FMRs.  However, the other advantage of
FMRs is that the cost of pipeline flushing to invalidate cached
translations can be amortized over multiple FMRs at once (by lazily
unmapping a list of FMRs).

 - R.