Hi,
(I originally only posted this to -workstation)
I've been thinking about bootloaders. Particularly with respect to PReP
machines.
The current method of making the kernel a binary loaded by the firmware
is a bit of a kluge. It means we have to jump through hoops to change
kernel parameters etc. It seems to me there are two or three obvious
approaches, but first I need some information.
1) What if any services does the PReP firmware provide once it loads an
image? I'm guessing that it isn't a lot, just the residual data to tell
us what hardware we have. I could be wrong here though - can anybody
tell me where to find softcopy PReP documentation?
2) What does the ARC bootloader goop for NT provide in the way of
services? I'm guessing rather more. In particular I suspect we have a
way or reading files, by name, from a FAT16-formatted partition, and
possibly passing them arguments, maybe stored in nvram. This is the way
some Alpha systems do it; you set up an ARC boot entry to run a little
executable (ldmilo.exe) which loads a file called 'milo' from the same
directory and executes it. Milo then takes over and loads the kernel.
Do all the PReP machines have ARC available for them? I would guess most
do, but I could be wrong.
IMHO milo itself is overkill; it actually contains an awful lot of the
kernel code (basically the SCSI drivers etc) so that once loaded it can
load the kernel from any device that linux knows about, even if the
machine firmware and ARC don't know about it. The current PReP boot code
covers that eventuality even if it is a bit of a kluge - as long as the
kernel can be loaded by the firmware (even from floppy) it will work.
Assuming that more PReP machines have ARC images available, I'd be
interested in creating a bootloader which, once loaded from ARC, would
be able to load a kernel image from a device ARC could read, using ARC
services. Anybody got any documentation?
Oh, we could also have the bootloader do a lot of fixing up of things
like the PCI spaces and interrupts - I note that NT and Linux on the
IBM 850 have different ideas of where things are and what interrupts
they use. Seems to work though.
Ideally we would get to the point where all the platform-specific fixup
code etc was moved into the bootloaders, and then do away with the
CHRP/PReP/whatever compile option.
Any comments?
Cheers,
David
** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/
On Mon, 6 Mar 2000, David Monro wrote:
I've been thinking about bootloaders. Particularly with respect to PReP
machines.
Me too...
The current method of making the kernel a binary loaded by the firmware
is a bit of a kluge. It means we have to jump through hoops to change
kernel parameters etc. It seems to me there are two or three obvious
approaches, but first I need some information.
1) What if any services does the PReP firmware provide once it loads an
image? I'm guessing that it isn't a lot, just the residual data to tell
us what hardware we have. I could be wrong here though - can anybody
tell me where to find softcopy PReP documentation?
It's somewhere on IBM sites if it's still there...
2) What does the ARC bootloader goop for NT provide in the way of
services? I'm guessing rather more. In particular I suspect we have a
way or reading files, by name, from a FAT16-formatted partition, and
possibly passing them arguments, maybe stored in nvram. This is the way
some Alpha systems do it; you set up an ARC boot entry to run a little
executable (ldmilo.exe) which loads a file called 'milo' from the same
directory and executes it. Milo then takes over and loads the kernel.
Do all the PReP machines have ARC available for them? I would guess most
do, but I could be wrong.
No, I've never seen a single machine with ARC myself...
IMHO milo itself is overkill; it actually contains an awful lot of the
kernel code (basically the SCSI drivers etc) so that once loaded it can
load the kernel from any device that linux knows about, even if the
machine firmware and ARC don't know about it. The current PReP boot code
covers that eventuality even if it is a bit of a kluge - as long as the
kernel can be loaded by the firmware (even from floppy) it will work.
Assuming that more PReP machines have ARC images available, I'd be
interested in creating a bootloader which, once loaded from ARC, would
be able to load a kernel image from a device ARC could read, using ARC
services. Anybody got any documentation?
The problem is that ARC has some services, PPCBUG used on Motorola boards
has very likely different ones...
Oh, we could also have the bootloader do a lot of fixing up of things
like the PCI spaces and interrupts - I note that NT and Linux on the
IBM 850 have different ideas of where things are and what interrupts
they use. Seems to work though.
For starting point see my prepboot bootloader which you can find at
ftp://vlab1.iram.es/linux-2.[23]
it is included in the mvme patches.
Basically:
- for now it adds a new directory (for now), arch/ppc/preepboot and a
CONFIG option to enable it.
- it reorganizes PCI I/O and memory space, actually for my machines based
on Raven/Falcon or Hawk Motorola chipsets, it remaps them completely
to something closer to CHRP
- it initializes video boards with an x86 ROM BIOS emulator if neceessary
- it still needs some work to handle the NVRAM, I want too to get rid of
most kludges in the early part of the kernel
- I have code which correctly gets the interrupt routing from residual
data
- it includes its own virtual memory management system, simple, primitive
and not resistant at all to fragmentation, but anyway it's thrown away
almost immediately,
- it is completely relocatable so that it can run whatever the memory
map of the system is ((except for interrupt vectors which are fixed by
HW). Actually it moves itself high so that it can decompress the kernel
in low memory.
Gabriel.
** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/
Gabriel Paubert wrote:
On Mon, 6 Mar 2000, David Monro wrote:
quoted
[..]
quoted
us what hardware we have. I could be wrong here though - can anybody
tell me where to find softcopy PReP documentation?
It's somewhere on IBM sites if it's still there...
Got it (and the nvram and residual data specs). Looks like a PReP
machine with Open Firmware would possibly have enough smarts to do
something sensible, but I don't think the 850 and the like implement OF,
so that isn't very useful. Do we know what machines do implement OF,
which ones have ARC available, which ones have PPCbug, and are there any
other PReP machines with none-of-the-above?
[..]
quoted
Do all the PReP machines have ARC available for them? I would guess most
do, but I could be wrong.
No, I've never seen a single machine with ARC myself...
Which machines do you deal with - did they not have an NT port? I could
be wrong, but I thought ARC was pretty much always used to boot NT on
non-intel machines...
[..]
The problem is that ARC has some services, PPCBUG used on Motorola boards
has very likely different ones...
Very true. Ugh.
quoted
Oh, we could also have the bootloader do a lot of fixing up of things
like the PCI spaces and interrupts - I note that NT and Linux on the
IBM 850 have different ideas of where things are and what interrupts
they use. Seems to work though.
For starting point see my prepboot bootloader which you can find at
ftp://vlab1.iram.es/linux-2.[23]
it is included in the mvme patches.
I've had a look at this before - didn't work for me at the time, but
that was probably for other reasons. I'll have to give it another shot.
Ahh - I guess there probably wasn't an NT port to a VME bus system :-)
Last mvme box I played with was I think the 131 - 68020 with 68851
running sysv/68k r2v3. Long time ago :-)
It is still part of the kernel image though, isn't it? What I would like
is an interactive standalone bootloader which knows quite a bit about
the individual systems, which is then capable of loading the kernel and
booting it - like milo on the alpha.
Basically:
- for now it adds a new directory (for now), arch/ppc/preepboot and a
CONFIG option to enable it.
- it reorganizes PCI I/O and memory space, actually for my machines based
on Raven/Falcon or Hawk Motorola chipsets, it remaps them completely
to something closer to CHRP
Hmm. Not sure what mine in the 850 is; lspci mentions a motorola MCP105
Eagle as the host bridge.
- it initializes video boards with an x86 ROM BIOS emulator if neceessary
Nice idea. It would still be kind of hard to access SMS (system
management services, basically the tools to select what device you plan
to boot off etc) if the firwmare doesn't support your video card though.
I don't think the 850 supports serial console at the firmware level
unfortunately :-)
- it still needs some work to handle the NVRAM, I want too to get rid of
most kludges in the early part of the kernel
Good..
- I have code which correctly gets the interrupt routing from residual
data
- it includes its own virtual memory management system, simple, primitive
and not resistant at all to fragmentation, but anyway it's thrown away
almost immediately,
- it is completely relocatable so that it can run whatever the memory
map of the system is ((except for interrupt vectors which are fixed by
HW). Actually it moves itself high so that it can decompress the kernel
in low memory.
Yeah, I noticed that. Keep up the good work!
btw - which tree are your patches on vlab1 against - Linus', vger or
fsmlabs?
Cheers,
David
** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/
On Mon, 6 Mar 2000, David Monro wrote:
Got it (and the nvram and residual data specs). Looks like a PReP
machine with Open Firmware would possibly have enough smarts to do
something sensible, but I don't think the 850 and the like implement OF,
so that isn't very useful. Do we know what machines do implement OF,
which ones have ARC available, which ones have PPCbug, and are there any
other PReP machines with none-of-the-above?
I had OF on my early MVME2600 but then Motoroola switched back to PPCBUG.
The OF was so buggy and lacking (no interrupt routing, no RTAS, even told
me that the OpenPIC was at bus 0 device 31 when it was device 0) that it
was useless.
Which machines do you deal with - did they not have an NT port? I could
be wrong, but I thought ARC was pretty much always used to boot NT on
non-intel machines...
I have seen VME boards from several manufacturers, and might see more
soon. All have different firmware, actually I liked very much the first
one which had a decent OF but it failed after a few weeks of operation.
Right now I only have Motorola to test (26xx, 24xx, hopefully soon
51xx but I know people who use my kernels on 23xx and 27xx).
quoted
The problem is that ARC has some services, PPCBUG used on Motorola boards
has very likely different ones...
Very true. Ugh.
Oh, and BTW the PPCBUG services can't be used with the cache or the MMU
enabled. Even sending a character to the console does not work. Don't
believe the documentation, it is so obviously plain wrong in this respect
that it's not even funny.
I checked on 603e boards, it seems the firmware forces the use of the high
interrupt entry points in ROM by setting MSR[IP] which do not provide
TLB reload handlers and prevents replacing MMU handlers.
In short I can't call PPCBUG services under my own memory management :-(,
it might work if I provided all possible translations that the firmware
use through BATS, but there are only 4 of them.
quoted
For starting point see my prepboot bootloader which you can find at
ftp://vlab1.iram.es/linux-2.[23]
it is included in the mvme patches.
I've had a look at this before - didn't work for me at the time, but
that was probably for other reasons. I'll have to give it another shot.
Ahh - I guess there probably wasn't an NT port to a VME bus system :-)
Last mvme box I played with was I think the 131 - 68020 with 68851
running sysv/68k r2v3. Long time ago :-)
It is still part of the kernel image though, isn't it? What I would like
is an interactive standalone bootloader which knows quite a bit about
the individual systems, which is then capable of loading the kernel and
booting it - like milo on the alpha.
Well it is part of the kernel image but it needs not be actually. You can
look at it as an MS-DOS like system. It would be fairly easy to add simple
drivers:it already provides simple memory management, console I/O and
basic PCI configuration services for a single task which may be anything
you want (right now it is decompressing an embedded kernel image +
setting bootinfo + moving initrd around). But it could also load another
image from disk or network, actually I wanted to do it but the PPCBUG
problems make it painful (include a simple fs which can read the kernel
image and pass it the correct parameters).
However I would not like to add interrupt management since polling should
be sufficient in a bootloader. Better clock and timing management could be
added if it simplifies the handling of timeouts however. As long as it's
mostly my code, I'll try to prevent needless bloat, however.
Hmm. Not sure what mine in the 850 is; lspci mentions a motorola MCP105
Eagle as the host bridge.
Fine, it's a standard PreP bridge, old but working. It can't be used for
CHRP like mappings though AFAIR (unlike the 106).
quoted
- it initializes video boards with an x86 ROM BIOS emulator if neceessary
Nice idea. It would still be kind of hard to access SMS (system
management services, basically the tools to select what device you plan
to boot off etc) if the firwmare doesn't support your video card though.
Indeed, unless you rewrite the firmware...
I don't think the 850 supports serial console at the firmware level
unfortunately :-)
PPCBUG does, actually most of my boxes are headless (don't even have a
keyboard interface or USB, which makes it somewhat awkward to plug in
a standard keyboard).
AIX firmware certainly had some support for serial console. Can't you put
a different firmware in your box (I know I'm lucky with the MVME boards,
they have 2 separate Flash ROMS and a jumper to select the one used at
reset so I can test a new firmware version and easily switch back to the
previous one if it fails).
quoted
- it is completely relocatable so that it can run whatever the memory
map of the system is ((except for interrupt vectors which are fixed by
HW). Actually it moves itself high so that it can decompress the kernel
in low memory.
Yeah, I noticed that. Keep up the good work!
btw - which tree are your patches on vlab1 against - Linus', vger or
fsmlabs?
Linus', I might try fsmlabs when BK it becomes official but I was so
severely burnt the day vger was closed to the public (had to modify
something like 600 files to get my systems to boot again), that I'm going
to be very reluctant to try anything else. Besides there are mirrors of
Linus'tree everywhere, but from here a checkout of vger does not always
finish overnight :-(
Cheers,
Gabriel.
** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/