Re: Scalability problem (kmap_lock) with -aa kernels

[Hugh Dickins]

My guess: persistent kmaps are okay, kmapped high pagetables are okay,
persistent kmapped high pagetables are okay.  What's wrong is how we
flush_all_zero_pkmaps on all cpus, synchronously while holding the
kmap_lock everyone needs to get a new kmap (and hopefully more often,
just inc or dec the pkmap_count of kmap already got).  That's what
cries out for redesign: it's served us well but should now be improved.

Hugh