[TUHS] Mach for i386 / Mt Xinu or other
downing.nick at gmail.com
Wed Feb 22 09:10:05 AEST 2017
I'm just going to jump in about the philosophical "everything is a
file" Unix concept.
I reckon this works very well for things that are "file like" in
nature. The genius of Thompson and Ritchie was to realize that a file
is a sequence of characters, and so is a TTY (or at least one
direction of a TTY). And of course it is REALLY REALLY HANDY, to have
things like an ad-hoc scripting ability (redirect a file to a program
that's expecting a TTY) and all the other combinations that are
At the same time, files have an additional capability which is
seeking. So did they implement some crazy ad-hoc scheme where you have
to write an escape sequence to the file in order to effect a seek? No
they did not. They added a seek(), or later lseek() system call, which
sensibly returns an error code if the thing receiving the command is
not capable of doing a seek. That is, they made it object-oriented.
As well, terminals have additional capabilities, like setting the
baudrate, changing the interrupt character, etc. And, in the same
vein, they added an ioctl() system call, which sensibly returns an
error code (ENOTTY) if the thing receiving the command is not a TTY.
Same thing for pipes, except now another object-oriented feature comes
into the mix: Constructors. Once created, a pipe does not have any
interesting extra capabilities, indeed it acts like a base class of
files and terminals since it does not support lseek() OR ioctl(),
however the interesting thing is that pipes (ignoring named pipes) are
not created with open(), they have their own constructor function
The BSD sockets interface is nice because it exactly follows the same
philosophy: Sockets are like terminals except that instead of ioctl()
they have slightly different capabilities (setting buffer size,
checking for out-of-band data, etc), and they also have some
interesting new characteristics like being bidirectional. But the most
interesting thing with sockets is the elaborate set of
constructors/factories for them.
Now, along comes Plan 9 and what do they do? Completely reverse all
this and force everyone to go through open() in the filesystem, which
is NOT object-oriented, can you imagine C++ without constructors? And,
in my understanding at least (it's a while since I looked at this),
you do things like opening sockets by echoing strings to various
places in the filesystem. This is nice if you're shell scripting.
To explain why I think this is a bad thing I'm going to digress
slightly. Let's consider the AT command set for modems. Well it's nice
because it's human readable, and it was also a decent engineering
solution to a problem where a programmable device was connected
through a dumb terminal interface. Likewise, the ANSI command set for
terminals, pretty much the same thing -- a good idea at the time.
But what I most emphatically DO NOT agree with, is when a modem is
built into the system, and still emulates an AT command set. Or when a
terminal is built into the system (xterm, say), and still emulates an
ANSI command set. It IS nice in some respects: (1) compatibility, and
(2) shell scripting / expect type stuff. But as an engineering
solution it's really bad. It's wasteful since any other program, such
as say a dialer, is encoding the commands into AT commands, and then
they are immediately getting decoded and executed. Same thing with
xterm and the ANSI command sequences. It's nice to be able to put
escape sequences in your prompt to change the colour of your prompt,
but something like the curses library should NOT have to encode
everything into ANSI when it's then decoded immediately.
So, in my humble opinion the best way to handle such issues is, to
provide a programmatic API where you tell the device what you want it
to do, and then if it's something like an xterm or an internal modem,
it immediately does what you ask it to do, whereas if it's something
like an actual ANSI terminal or an external modem, the driver takes
care of encoding your commands into ANSI or AT commands and then
untangling the results, and reports to you in a machine-readable way
(rather than a human-readable way) whether it succeeded etc.
So, my philosophy is quite opposite to the Plan 9 philosophy, and
that's why I don't use Plan 9. I don't believe in trying to put square
pegs in round holes basically. What I would like to see is basically
every device to implement the common functionality like read() and
write() but ONLY IF THAT IS SENSIBLE, and to expose a programmatic
interface through ioctl() or similar. Note that I have actually come
to believe that overloading everything onto ioctl() is a bit
undesirable since ioctl() was originally intended only for terminals.
It would be better if each thing that a user process can hold a "fd"
to, could define its own syscalls, but ioctl() is okay as an
encapsulation for such syscalls.
I also don't really agree with things like the /proc filesystem. For
its intended and original purpose, it's fine -- it was supposed to be
a collection of files named for the pids of running processes in the
system, so that debuggers and such like, could read/write the
process's address space. Personally I think it should have its own
constructor, like say: fd = openaddressspace(pid, O_RDWR); which would
be equivalent to something like: sprintf(buf, "/proc/%d", pid); fd =
open(buf, O_RDWR); since there's no good reason to put it in the
filesystem in my opinion. Having said that, there's no real harm in
putting it in the filesystem, especially as it's hardly ever used.
What I DO NOT agree with, is all the pseudo-files, where you do
something like: echo 1 >/proc/sys/some/silly/function from the shell
to turn stuff on or off.
So now having explained my basic philosophy I will touch on the SCSI
driver and burning CDs/DVDs/BDs. We can look at it two ways, either we
can build a high level interface into the system so that writing
software doesn't need to know how the drive is connected. This is a
bit like my proposal for xterms vs. ANSI terminals and internal vs
external modems. Then if it HAPPENS to be a SCSI drive then the driver
should communicate with the drive in SCSI language, etc. BUT: This is
a huge amount of work, because the SCSI specification is very
complicated. Moreover, it's highly unlikely that any device will be
built in the foreseeable future that burns CDs/DVDs/BDs and does not
understand SCSI. (It might be iSCSI, it might be SCSI encapsulated
into IDE or SATA, or whatever, but it still understands SCSI). In my
opinion then, there is no need for a high level driver. It's like the
situation with ANSI terminals or external models BEFORE things like
xterms or internal modems were ever invented. So why would you need an
extra level of abstraction? Just provide access to the TTY.
On Wed, Feb 22, 2017 at 8:49 AM, Noel Chiappa <jnc at mercury.lcs.mit.edu> wrote:
> > From: Larry McVoy
> > The DOS file system, while stupid, was very robust in the face of
> > crashes
> I'm not sure it's so much the file system (in the sense of the on-disk
> format), as how the system _used_ it (although I suppose one could consider
> that part of the FS too).
> The duplicated FAT, and the way file sectors are linked using it, is I suppose
> a little more robust than other designs (e.g. the way early Unixes did it,
> with indirect blocks and free lists), but I think a lot of it must have been
> that DOS wrote stuff out quickly (as opposed to e.g. the delayed writes on
> early Unix FS's, etc). That probably appoximated the write-ordering of more
> designed-robust FS's.
More information about the TUHS