Date: Mon, 5 Dec 2005 15:47:14 -0500
From: Russ Cox <rsc@swtch.com>
To: 9fans <9fans@cse.psu.edu>
Subject: [9fans] replacing 9load

For quite a few years, various of us have talked about replacing 9load. When Jim and I replaced b.com with the combination of ld.com and 9load, we considered using a stock Plan 9 kernel instead. Sean Quinlan took a few steps toward making this possible: he added the configuration environment #ec, the Dev.shutdown routines, and /dev/reboot. With these, one can use the kernel as an impromptu boot loader already. I believe Ron Minnich does just that in some of his machines.

There are a few things 9load does that the kernel depends on. The main one is loading plan9.ini from somewhere and leaving it in memory. A secondary one is setting up the APM BIOS interface, and in the original VESA code written for Inferno, 9load was responsible for setting up the VGA. With the new realmode code (or better, if someone would write it, VM86 code) in the kernel, the APM and VESA can be done by the kernel proper. That leaves plan9.ini.

Plan9.ini serves two related but different purposes. The first is to give the kernel hints about hardware recognition. The second is to specify a few aspects of the software configuration. Ideally, it would be nice if plan9.ini could become optional.

Hardware recognition was a much bigger problem in the old ISA days. Plan9.ini was, at the time, an excellent alternative to recompiling the kernel to hard-wire all the constants about your machine. Now that we have PCI and its device ids, you don't need to tell the kernel what hardware you have anymore. It knows, usually better than you.

That leaves the software configuration: mouse, vga, serial console, kernel location, root file system location, boot menu, etc. The mouse and vga could possibly be auto-detected, and the new vga resize code (if it works) makes it possible to correct a bad guess about the user's desired screen size.

The other options in plan9.ini -- the serial console, the kernel and root file system, the boot menu -- can't just be guessed. Jim has been playing with loading plan9.ini via PXE/TFTP, which is great for a real single-file-server Plan 9 setup, though it won't satisfy the single-machine users. I don't expect this part of plan9.ini to go away, though one could make 9load do without it in the general case (scan the local 9fats for files named 9pc*, for example).

When I put the new mmu code into the kernels, I didn't feel like putting it into 9load too. Instead I wrote a small binary that turns a gzipped kernel into a 9load binary that the PBS bootstrap could load into memory. You run

cat load 9pcload.gz >9load

to produce a 9load from a kernel. I have put this in /n/sources/contrib/rsc/load.

This new program means that you can build yourself a special pcload kernel with some 9load-like user-level program installed as /boot/boot instead of the usual boot. It would find plan9.ini, present any prompts if needed, find and load a kernel, and then /dev/reboot into it. You could put dossrv, ext2srv, fossil, venti, kfs, factotum, mount, tftp, ftpfs, hget, etc., into the root file system, to load kernels off of any conceivable medium.

While a new 9load using the real kernel could be great, it's not very high on my or Jim's to do list, since the current 9load suffices, and we're more interested in what happens once the kernel is in memory than the mundane and now all-too-familiar details of the x86 boot process.

I encourage those of you grumbling for a new 9load to attack one of the following problems:

• write vm86 code for the kernel to replace the realmode code. vm86 code should be safer and more reliable and will make the vesa and apm code run on more machines.

• write a mouse auto-detector: scan the usb, try ps2, scan the serial ports. this should be almost trivial.

• write a vga size auto-detector: try aux/vga -m vesa -p and fetch and parse the edid information from the monitor. aux/vga does most of this already.

• write a user-level program to be the `9load' in a 9pcload kernel. the program should have generic "get me a plan9.ini" and "get me a kernel" interfaces, like the current 9load, so that many boot methods can be plugged in.

Ironically, once this is all done we'll have moved full circle to where Plan 9 was many years ago, when the hobbit boot loader was just a Plan 9 kernel with a different load address and a special boot program that let you set environment variables, examine memory, attach to file servers, run programs that were downloaded, and load and boot a kernel.

Russ