Re: CVS commit: src/sys/arch/i386/i386

[Andrew Doran <ad%NetBSD.org@localhost>]

On Mon, May 04, 2009 at 03:41:16PM +0200, Manuel Bouyer wrote:

> On Mon, May 04, 2009 at 11:47:29AM +0000, Andrew Doran wrote:
> > Module Name:        src
> > Committed By:       ad
> > Date:               Mon May  4 11:47:29 UTC 2009
> > 
> > Modified Files:
> >     src/sys/arch/i386/i386: locore.S
> > 
> > Log Message:
> > PR kern/41342: BSDi binaries cause panic
> > 
> > XXX Manuel, please have a look as I am not sure what to do for XEN here!
> 
> I tried the following code instead:
> IDTVEC(osyscall)
>         CLI(%eax)               # must be first instruction
>       pushfl                  # set eflags in trap frame
>       popl    8(%esp)
>       orl     $PSL_I,(%esp)   # re-enable ints on return to user
>       pushl   $7              # size of instruction for restart
>       jmp     syscall1
> 
> but the test binary from the PR segfaults:
> truc# kdump |less                                                             
>  
> 34      0 ktrace   EMUL  "netbsd"
> 34      0 ktrace   RET   ktrace 0
> 34      0 ktrace   CALL  execve(0xbf7ffc02,0xbf7feb3c,0xbf7feb44)
> 34      0 ktrace   NAMI  "./architextIndex"
> 34      0 architextIndex EMUL  "netbsd"
> 34      0 architextIndex RET   syscall JUSTRETURN
> 34      0 architextIndex PSIG  SIGSEGV SIG_DFL: code=SEGV_ACCERR, addr=0xacb 
> 94, trap=4)
> 34      0 architextIndex NAMI  "architextIndex.core"
> 
> On Xen CLI(%eax) expands to:
>    movl    CPUVAR(VCPU),%eax ;
>    movb $1,EVTCHN_UPCALL_MASK(%eax)

At this point the segment registers won't be set up. And %eax contains the
syscall number.
 
> I guess this is a problem. Is there a way to account for this somewhere ?

It is difficult to avoid the LDT/segreg problems without having interrupts
disabled instantly on entry.

Maybe we could add really ugly logic to compensate for it in trap() since
oosyscall is the only place where we enter with interupts on (I don't know
how interrupts/traps are set up on xen currently).

xen isn't as vulnerable to the LDT/segreg problem as native x86 because
it's not MP and doesn't do kernel preemption. For the time being I guess
it would suffice to #ifdef the 'cli'.