[TUHS] Re: [COFF] Re: Supervisor mode on ye olde PDP-11

[TUHS] Re: [COFF] Re: Supervisor mode on ye olde PDP-11

[Noel Chiappa]

    > From: Warner Losh

    > 2.11BSD used a mode between kernel and user for the TCP stack to get
    > more effective address space...

Is there a document for 2.11 which explains in detail why they did that? I
suspect it's actually a little more complicated than just "more address
space".

The thing is that PDP-11 Unix had been using overlays in the kernel for quite
a while to provide more address space. I forget where they first came in (I
suspect there were a number of local hacks, before everyone started using the
BSD approach), but by 2.9 BSD they were a standard part of the system. (See:

  https://minnie.tuhs.org/cgi-bin/utree.pl?file=2.9BSD/usr/src/sys/conf/Ovmakefile

for some clues about how this works. There is unfortunately no documentation
that I know of which explains clearly how it works; if anyone knows of any,
can you please let me know? Otherwise you'll have to read the sources.)

I can think of two possible reasons they started using supervisor mode: i)
There were a limited number of the 2.9-type overlays, and they were not
large; trying to support all the networking code with the existing overlay
system may have been too hard. ii) I think this one is unlikely, but I'll
list it as a possibility. Switching overlays took a certain amount of
overhead (since mapping registers had to be re-loaded); if all the networking
code ran in supervisor mode, the supervisor mode mapping registers could be
loaded with the right thing and just left.

	Noel

[TUHS] Re: [COFF] Re: Supervisor mode on ye olde PDP-11

[Phil Budne]

JNC wrote:
> Is there a document for 2.11 which explains in detail why they did that? I
> suspect it's actually a little more complicated than just "more address
> space".
...
> ... Switching overlays took a certain amount of
> overhead (since mapping registers had to be re-loaded); if all the networking
> code ran in supervisor mode, the supervisor mode mapping registers could be
> loaded with the right thing and just left.

That's my understanding... It allows mbufs to be mapped only
in supervisor mode...

https://minnie.tuhs.org/PUPS/Setup/2.11bsd_setup.html says:

       The networking in 2.11BSD, runs in supervisor mode, separate
       from the mainstream kernel. There is room without overlaying to
       hold both a SL/IP and ethernet driver. This is a major win, as
       it allows the networking to maintain its mbufs in normal data
       space, among other things. The networking portion of the kernel
       resides in ``/netnix'', and is loaded after the kernel is
       running. Since the kernel only looks for the file ``/netnix'',
       it will not run if it is unable to load ``/netnix'' , sites
       should build and keep a non-networking kernel in ``/'' at all
       times, as a backup. NOTE: The ``/unix'' and ``/netnix''
       imagines must have been created at the same time, do not
       attempt to use mismatched images. The ability to have boot tell
       the kernel which network image to load is on the wish list (had
       to have something take the place of wishing for disklabels
       ;-)).

https://wfjm.github.io/home/ouxr/ shows the code path for the socket(2) syscall

[TUHS] Re: [COFF] Re: Supervisor mode on ye olde PDP-11

[Warner Losh]

[-- Attachment #1: Type: text/plain, Size: 2276 bytes --]

On Wed, Jun 19, 2024 at 10:56 AM Phil Budne <phil@ultimate.com> wrote:

> JNC wrote:
> > Is there a document for 2.11 which explains in detail why they did that?
> I
> > suspect it's actually a little more complicated than just "more address
> > space".
> ...
> > ... Switching overlays took a certain amount of
> > overhead (since mapping registers had to be re-loaded); if all the
> networking
> > code ran in supervisor mode, the supervisor mode mapping registers could
> be
> > loaded with the right thing and just left.
>
> That's my understanding... It allows mbufs to be mapped only
> in supervisor mode...
>

Yea. that's a much better explanation than my glossed over 'more address
space'. It's both to get more text space (overlays weren't infinite, and
being a
separate image allowed more selective communication across the interface
boundary) and to provide some separation and allow for more data to be
around easily (the BSD kernel didn't overlay data, which is technically
possible
on PDP-11, but the linker didn't support it).


> https://minnie.tuhs.org/PUPS/Setup/2.11bsd_setup.html says:
>
>        The networking in 2.11BSD, runs in supervisor mode, separate
>        from the mainstream kernel. There is room without overlaying to
>        hold both a SL/IP and ethernet driver. This is a major win, as
>        it allows the networking to maintain its mbufs in normal data
>        space, among other things. The networking portion of the kernel
>        resides in ``/netnix'', and is loaded after the kernel is
>        running. Since the kernel only looks for the file ``/netnix'',
>        it will not run if it is unable to load ``/netnix'' , sites
>        should build and keep a non-networking kernel in ``/'' at all
>        times, as a backup. NOTE: The ``/unix'' and ``/netnix''
>        imagines must have been created at the same time, do not
>        attempt to use mismatched images. The ability to have boot tell
>        the kernel which network image to load is on the wish list (had
>        to have something take the place of wishing for disklabels
>        ;-)).
>
> https://wfjm.github.io/home/ouxr/ shows the code path for the socket(2)
> syscall
>

Oh, that's nice.

Warner

[-- Attachment #2: Type: text/html, Size: 3239 bytes --]