On Fri, May 16, 2014 at 09:59:25PM +0200, Paolo Bonzini wrote:
CS.RPL is not equal to the CPL in the few instructions between
setting CR0.PE and reloading CS. We get this right in the common
case, because writes to CR0 do not modify the CPL, but it would
not be enough if an SMI comes exactly during that brief period.
Were this to happen, the RSM instruction would erroneously set
CPL to the low two bits of the real-mode selector; and if they are
not 00, the next instruction fetch cannot access the code segment
and causes a triple fault.
However, SS.DPL *is* always equal to the CPL. In real processors
(AMD only) there is a weird case of SYSRET setting SS.DPL=SS.RPL
from the STAR register while forcing CPL=3, but we do not emulate
that.
I was in the process of testing something else, when I encountered a
problem with an old MSDOS 6.22 floppy I had. I tracked it down to an
error in one of the commits I did in this series (I sent a fix in a
separate email for it).
Unfortunately, after I fixed the problem in my patch, your patch above
breaks it again. I think it's another VM86 thing.
Steps to reproduce:
1 - grab the DOS 6.22 floppy from: http://bootdisk.com/bootdisk.htm
2 - boot it up and add emm386.exe to config.sys ("edit config.sys" and
add "DEVICE=EMM386.EXE" on the second line of the file).
3 - reboot with modified config.sys