On Mon, Apr 20, 2015 at 04:28:03PM +0200, Roger Pau Monné wrote:
Hello,
El 16/04/15 a les 19.51, Kevin O'Connor ha escrit:
On Thu, Apr 16, 2015 at 06:37:29PM +0200, Roger Pau Monné wrote:
El 16/04/15 a les 17.52, Kevin O'Connor ha escrit:
Seems like the same problem. You wont be able to set a gdb breakpoint for the freebsd call because freebsd isn't calling the bios - it's attempting to interpret the bios code.
Does the seabios patch below fix the problem for you?
Seems to kind of fix it, but it's hard to tell.
Most of the time the original SeaBIOS binary works without problems. There's sometimes were the int 0x15 call with ah=0xc0 returns what seem to be valid values in ah and flg, but the values in es and bx are corrupted so when freebsd tries to access this region (es << 4 + bx) it gets a page fault.
This is what I see now with the patch applied:
atkbdc0: <Keyboard controller (i8042)> port 0x60,0x64 irq 1 on acpi0 atkbd0: <AT Keyboard> irq 1 on atkbdc0 Calling INT 0x15 (ax=0xc000 bx=0x0000 cx=0x0000 dx=0x0000 es=0x0000 di=0x0000) Exiting INT 0x15 (ax=0xf9c0 bx=0xf9c0 cx=0xf99e dx=0xdf80 es=0x0000 di=0x0000) kbd0 at atkbd0 atkbd0: [GIANT-LOCKED]
Ah, looks like the freebsd code isn't even checking if x86emu exited abnormally.
Yes, this is something that can be solved without much work AFAICT, so that we know if the emulator exited correctly or not. However this is only a side-effect of what's actually happening.
If changing freebsd code, I would change it to set the repeat rate to a standard default irrespective of what the bios uses, and thus not call the bios at all.
To summarize, this looks to be the same problem that I investigated two years ago:
http://lists.nongnu.org/archive/html/qemu-devel/2013-03/msg01311.html
Basically, freebsd is attempting to interpret the x86 bios code, but it is using an incomplete interpreter that misinterprets some x86 instructions. That broken interpreter could cause page faults, loop forever, or return bogus values.
I've added a little bit more debug to the FreeBSD kernel and x86emu in order to see what's going on. It seems like SeaBIOS contains VERR/VERW instructions (or x86emu in FreeBSD thinks so, but maybe this is just the fail over of some badly emulated instructions), which x86emu doesn't know how to handle, can this be the case?
Yes. It is known that x86emu does not properly emulate the calll, retl, leavel, and leal instructions (32bit forms of call, ret, leave and lea). There are other unsupported instructions as well (eg, smsww, outsb, enterl), but these are less problamatic because gcc doesn't emit them.
Because x86emu is used so frequently with vga bios, SeaVGABIOS has code to fixup the gcc assembler to avoid the above instructions - see scripts/vgafixup.py .
-Kevin