From mboxrd@z Thu Jan  1 00:00:00 1970
From: jnc@mercury.lcs.mit.edu (Noel Chiappa)
Date: Thu,  1 Feb 2018 08:03:11 -0500 (EST)
Subject: [TUHS] Odd behaviour on V6 startup, etc
Message-ID: <20180201130311.4BAB118C0B2@mercury.lcs.mit.edu>

As I mentioned recently, Dave Bridgham and I are doing an RK11 emulator (and
soon an RP11, which just will take editing the Verilog a tiny bit) using an
SD card for storage, for heritage computer collectors who want to run their
PDP-11's but don't want to risk a head crash (since replacement heads are now
un-obtainium) - and also for people who have an -11, but no mass storage.

So the last stage was bringing up V6 on a hardware PDP-11/23, using the
emulated RK11, and we were having an issue (which turned out to be partial
block reads/writes not working properly; this has since been fixed, and we
have successfullly booted V6 on the -11/23 with only an SD card).


In the process of debugging that issue, I added logging to the RK driver, and
noticed that in the process of attempting to boot single-user, the system was
trying to do some swapping - two writes, and two reads - before it even tried
doing the fork() to create the shell process, followed by opening /dev/tty8
and doing the exec() of /bin/sh.

This confused me, since both newproc() and expand() have code which uses a
memory-memory copy if there is enough free memory - and on first booting,
there definitely was. Why was it swapping?


So I added some more logging, and the cause turned out that at one point, I
had re-compiled /etc/init - and without thinking about it too hard, had made
it a pure-text program.

And that's what did it: in exec() (called when process 1 tries to exec()
/etc/init), it calls xalloc(), which i) if the text was not already
available, gets together the pure text, and puts a copy on the swap device,
and ii) if the text is not already in core, swaps the rest of the process
out, because, as the code explains:

 * if the calling process
 * is misplaced in core the text image might not fit.
 * Quite possibly the code after "out:" could check to
 * see if the text does fit and simply swap it in. 

If you're looking at the code, the process doesn't have a data segment at
that point, just the U area; the data will be set up later in the exec()
code, after the call to xalloc() returns, after the process is swapped back
in.

In fact, the process will _never_ have anything other than a U area when in
this code; the only call to xalloc() in the system is immediately preceeded
by an "expand(USIZE)" which throws away everything except the U area.


So I'm contemplating doing what the comment suggests - adding code to check
to see if there's enough memory available to hold the pure text, and if so,
avoiding the swap-out/in of the process' U-only data segment.

If done 'right', it should also be possible to allow avoiding the reading-in
of the text from the swap device, when first exec'ing a pure text...

Probably not as interesting a project as doing the splice() system call was,
but no doubt it will teach me about a corner of the system I don't know very
well (back in the day, doing networking stuff, I was mostly looking at I/O
code).

Although there is potentially a certain amount of amusement in an additional
enhancement, which is trying to make sure there's enough room for both the
text and the data; immediately upon the return from xalloc(), the process is
expanded to have room for the data. Which might involve _another_ swap-out/in
sequence...  So perhaps the U area should simply be moved out of the way via a
direct memory-memory copy, rather than swapping it out/in, just before doing
it again!

You are not expected to understand this... :-)

I guess I should first take a look at the PWB1 pure text code, which is
heavily modified from the stock V6, to see what it does - this all may have
already been done there.

	Noel