Wink Saville (wink@saville.com) on 5/15/06 wrote:
>
>Doesn't Linux use messaging, I would say it does, for
>instance in the block IO subsystem?

Not really.

Yes, the block IO subsystem can really be seen as a
"message queue", but it's definitely not messages in the
ukernel sense. It's very much a "pointers to data" model,
where you wouldn't send the data as part of the message
queue, you'd just have a pointer to it.

The message itself only contains the command information,
generally (ie the message will have a small data area
associated directly with it that describes what the
message is all about, but the actual payload for the
command is not directly attached).

Putting the data itself into the message would simply not
be acceptable from a performance standpoint, since we
want to try to avoid copying it around as much as humanly
possible.

So the block queues very much rely on the "one global
address space" thing, although for obvious reasons there
are the actual hardware-imposed issues of the "bus address
space" vs "virtual CPU address space" which come up as a
complicating factor.

A DMA controller cares about the bus address space, while a
PIO interface cares about the CPU sw-visible virtual
addressing, and the actual command data usually will be
accessed by both methods.

There are some subsystems that look more like real message
queues, but they are for things like mouse/keyboard data,
where the message data is usually very small, and the
performance requirements are obviously on a totally
different scale.

Socket buffers look more like messages too - there the
actual transport is obviously all about sending packets of
data..

Linus