mmap() scalability in the presence of the MAP_POPULATE flag

[Roman Dubtsov <dubtsov@gmail.com>]

Concurrent mmap() calls from the same process are serialized via downing
mm->mmap_sem for write. This means that operations like populating the
pages which do not alter vmas are also performed serially. Anecdotal
data from two machines I have access to is that populating pages by
touching them in a loop outside of mmap() improves performance of the
synthetic micro-benchmark by ~40% in the worst case. A crude patch that
modifies vm_mmap_pgoff() to call make_pages_present() outside of
do_mmap_pgoff() after upping the semaphore when MAP_POPULATE is present
brings identical performance improvement.

Is there an interest in fixing this or concurrent mmaps() from the same
process are too much of a corner case to worry about it?

Regards,
Roma