[openib-general] Re: Re: madvise MADV_DONTFORK/MADV_DOFORK

[Roland Dreier]

    Gleb> It had come to my attention that there is file memory.c in
    Gleb> libibverbs that implements refcounting for mlock. I think it
    Gleb> was meant to be used from reg_mr() (since interface is
    Gleb> hidden) back when mlock was needed for kernel bug
    Gleb> workaround. If it was good idea back than why not now?
    Gleb> Alternatively we can make this interface public for
    Gleb> application to use explicitly.

    Gleb> Roland can you please tell the history behind memory.c?

This was used for mlock() when that was required.  The reason I left
it around was that I always thought that the MADV_DONTFORK would be
handled the same way.

Probably the best thing to do would be to have libibverbs handle the
madvise() stuff transparently by default, but perhaps add some hooks
for more intelligent applications or MPI libraries to do their own thing.

 - R.