[openib-general] Question about pinning memory

[Roland Dreier]

    Gleb> I think the better approach would be to send signal to
    Gleb> process whenever vma list changes and let application verify
    Gleb> whether cache entries should be freed.

It's probably better to use a file descriptor for kernel upcalls to
userspace.  Then you don't have to go through the expensive signal
delivery path.  Also you're not limited to the info you can squeeze
into the siginfo.

In this case I'm not sure whether it's better to have things driven by
userspace or the kernel.  Having the kernel push new info up to
userspace seems more efficient, since the kernel knows exactly when
things have changed, but this seems to lead to synchronization
problems where userspace may post a work request using an obsolete
mapping before it gets the notification that the mapping changed.

 - R.