Re: Message passing in user-land

[Niels Möller]

rreale@iol.it writes:

> But isn't it possible to build a sort of mechanism based on shared memory,
> thus avoiding the passage across the kernel?

It might be possible to create an ipc mechanism for pure data using
only shared memory, even if it seems somewhat tricky to make it robust.

But IPC is more than just data. For Hurd IPC we also need to be able
to send memory pages (by reference, not copying them), and port
rights. On Mach, all that is in the kernel, on L4, management of port
rights will likely move to user space, but management of the ownership
of individual memory pages is a natural thing for the kernel to do,
and that's the case also for L4.

I suspect that the minimal kernel IPC mechanism that is needed for the
Hurd is something like "give this other process access to this
particular memory page". L4 provides slightly more than that, and
that's probably good for performance, in particular for ipc calls
where the data fit in registers.

Regards,
/Niels