Squeal Mode is fully scizophrenic operation of the 386+.
Neologisms used...
dismiss IRET, return with address, code segment and flags passed
dual 16 bit datum, INTeLese "word"
event user/invoker of IDT or legvecs
legvec 8086 interrupt vector
reflex device handler bottom end, event service
pull pop value keeping value
submit INT N, synchronous trap
core as in pullcore, A C D B SP BP SI and DI
This legmeg dir is for XT/AT stuff like BIOS calls. This is much of what a
PC is and I don't want to throw that info out. Proximity is the best
documentation. The ROMs are on your bus. I also don't want to write a lot
of 8086 code. Like floppy drivers. Although I don't even like the BIOS
keyboard driver. Options are good though, particularly when trying to get
a system going at all. I looked deep into the shiny yellow eyes of...
o spot-reassembling the BIOS to USE16 pmode
o VM86
o switching to real mode
Switching to real mode is what the BIOS will like best. And is the
simplest of the three, and I've done it before. And is least dependant on
the more inscrutable INTeLnesses. This way we worry about the PE flag and
not much other pmode bletchery. This minimizes the Wintendo game of trying
to find out how things work that don't work. This is more
Chuck-Moore-grade stack gymnastics. Molesting a gargoyle like the PC isn't
half bad if you leave the flag draped over it's faces.
Here's the plan ...
init ...
trivial & done
Copy the legvecs to $legvec.
well under way
Write vsplice, which goes to ring0 and submits an event
into a pmode reflex, returns to rmode, then dismisses
(back to rmode). "pigcatcher" might be a better name.
trivial
set 0<->1k to the call addresses of each of the calls to
vsplice in the ESA.
trivial
Fill the 0->1k with the above.
trivial
Hook the hot ones like NMI with an rmode reflex. In
addition to the pmode reflex.
Write a FIxNN that switches to real mode, calls the
BIOS from real mode, switches back to pmode.
Replicate the above for the 16 or so interesting BIOS INTs.
from ring0 ...
Use legacy BIOS calls at will via a fakeINT calls. Set up
the call as per BIOS register conventions, ignoring the
high duals of the core registers. Recieve BIOS return info
as per BIOS legacy usual.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Squeal Mode bimodal events
8086 legvecs @ 0
______
|vsplice
/|8086 \
/ |8086 \
living r / |vsplice-->-------> vspliceN
well in / |8086 / |
rmode *POW*/ |vsplice / |
or --->>>EVENT |vsplice / |
pmode \ |___ |
\ |
p \ switch to pmode ring 0
| r |
| /--------- submit the event number N
| | |
| | pmode event handler returns here
A V V /-> consuming submit frame
| | / |
| r | / switch to rmode
\ ----- | <------- | <-\ |
| | \ dismiss
\ ____ |r \_____/
\ | |
\ | I \ |
\ | \ |dismiss
\ | D \|
\| __/\
| T / \
| / \
/\ | / \
| |____ \p
\ p /
\_________________________/
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ring 0 wants to call a BIOS INT
There will be routines ala FIx10, FI0x11, FIx12... based on this, FIx10.
get your (386+) regs as per the particular
BIOS INT spec, but bigger
|
|
call FIx10 (pmode interface to BIOS INT 0x10)
|
switch to rmode
|
push an 8086 dismiss frame
by hand
|
jump far BIOS 0x10 entry point** -----> BIOS ------> dismiss
| /
BIOS dismisses to here *** <-------------------/
|
set up BIOS reg returns
for a pmode pullcore
|
switch to pmode
|
pullcore
|
return
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
** the far jump 16:16 pointers are selfmodded for each FIxnn to the
appropriate RAM ROM image at init time, based on the original bootup
legvecs+RAMROM shifts. Or just leave 'em in ROM. But still it's a self-mod
for more than one box.
You need about 16 instances of vsplice init-modded to the pertinent event
number. NMI will have a straight-8086 reflex in addition to the pmode
reflex, so you need vsplice for the IRQs. The 8086 exceptions can have
8086 ingnore reflexes.
*** part of the aforementioned 8086 IRET stack frame push.
SUMMARY
In order to flip back and forth from rmode to pmode and not kill
event-handling, i.e. to still run an OS, you have to leave events on. That
means you have to handle events in either mode roughly equivalently. I
want the option of calling the legacy BIOS INTs. Some of them probably
kill event-handling, but that's the breaks. I want the option of using the
BIOS only in the kernel though. That means I'm going from pmode to rmode,
BIOS, then back to pmode. I'm going to leave events on in those
circumstances. The BIOS probably varies from INT to INT on that. Worst
case, I doubt the BIOS turns off NMI except when asked to. Hope.
Given an rmode BIOS INT in what is otherwise a pmode OS, events occur in
rmode. Asynchronous events that may occur in such a scenario are NMI, 8086
exceptions, and IRQs. No protection exceptions, no synchronous traps. That
gives about 25 events that one has to prepare for in rmode. IRQs get
spliced back to pmode, dismiss, and return to rmode. They use the pmode
handlers. They need a splice. NMI will have an explicit rmode reflex in
addition to the pmode reflex, or the one NMI reflex will be coded
mode-indifferently. 8086 exceptions can be ignored in a kernel.
Thus to not throw out the BIOS and still have an OS one has to have a BIOS
INT wrapper (FIxNN in Ha3sm), and an IRQ modesplice (vsplice in Ha3sm).
Both are instantiated for each IRQ or INT in question to reduce problems
with the severe state restrictions on code handling events on a
scizophrenic machine.
So no, I'm not working on a "bootloader" in osimplay. Other than I'm
writing a femto-kernel. You may recall that my floppy "bootsector" is 15
bytes. Those 15 bytes apparently load one floppy cylinder to 0x7c00, which
18k or so I'm going to have a hard time consuming with the parts of a real
OS that interest me.
Rick Hohensee
Ha3sm is pronounced "haze 'em"
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/