20 Sep
2015
20 Sep
'15
6:39 p.m.
From: Peter Jeremy <peter(a)rulingia.com>
Why were the original read(2) and write(2) system
calls written to
offer synchronous I/O only?
A very interesting question (to me, particularly, see below). I don't think
any of the Unix papers answer this question?
It's relatively easy to create synchronous I/O
functions given
asynchronous I/O primitives but it's impossible to do the opposite.
Indeed, and I've seen operating systems (e.g. a real-time PDP-11 OS I worked
with a lot called MOS) that did that.
I actually did add asynchronous I/O to V6 UNIX, for use with very early
Internet networking software being done at MIT (in a user process). Actually,
it wasn't just asynchronous, it was _raw_ asynchronous I/O! (The networking
device was DMA, and the s/w did DMA directly into the user process' memory.)
The code also allowed more than one outstanding I/O request, too. (So the
input could be re-enabled on the device ASAP, without having to wake up a
process, have it run, do a new read call, etc.)
We didn't redo the whole Unix I/O system, to support/use asyn I/O throughout,
though; I just kind of warted it onto the side. (IIRC, it notified the user
process via a signal that the I/O had completed; the user software then had
to do an sgtty() call to get the transfer status, size, etc.)
Anyway, back to the original topic: I don't want to speculate (although I
could :-); perhaps someone who was around 'back then' can offer some insight?
If not, time for speculation! :-)
Noel