[openib-general] Re: [PATCH][RFC][0/4] InfiniBand userspace verbs implementation

[Roland Dreier]

    Timur> If you look at the Infiniband code that was recently
    Timur> submitted, I think you'll see it does exactly that: after
    Timur> calling mlock(), the driver calls get_user_pages(), and it
    Timur> stores the page mappings for future use.

    Andrew> Where?

The code isn't merged yet.  I sent a version to lkml for review -- in
fact it was this very thread that we're in now.  The code in question
is in http://lkml.org/lkml/2005/4/4/266

This implements a "userspace verbs" character device that memory
registration goes through.  This means the kernel has a device node
that will be closed when a process dies, and so the memory can be
cleaned up.

 - R.