Re: [PATCH v2] x86/xen/64: Rearrange the SYSCALL entries

[Andy Lutomirski]

On Sun, Aug 13, 2017 at 7:44 PM, Brian Gerst <brgerst@xxxxxxxxx> wrote:
> On Mon, Aug 7, 2017 at 11:59 PM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>>  /* Normal 64-bit system call target */
>>  ENTRY(xen_syscall_target)
>> -       undo_xen_syscall
>> -       jmp entry_SYSCALL_64_after_swapgs
>> +       popq %rcx
>> +       popq %r11
>> +       jmp entry_SYSCALL_64_after_hwframe
>>  ENDPROC(xen_syscall_target)
>>
>>  #ifdef CONFIG_IA32_EMULATION
>>
>>  /* 32-bit compat syscall target */
>>  ENTRY(xen_syscall32_target)
>> -       undo_xen_syscall
>> -       jmp entry_SYSCALL_compat
>> +       popq %rcx
>> +       popq %r11
>> +       jmp entry_SYSCALL_compat_after_hwframe
>>  ENDPROC(xen_syscall32_target)
>>
>>  /* 32-bit compat sysenter target */
>>  ENTRY(xen_sysenter_target)
>> -       undo_xen_syscall
>> +       mov 0*8(%rsp), %rcx
>> +       mov 1*8(%rsp), %r11
>> +       mov 5*8(%rsp), %rsp
>>         jmp entry_SYSENTER_compat
>>  ENDPROC(xen_sysenter_target)
>
> This patch causes the iopl_32 and ioperm_32 self-tests to fail on a
> 64-bit PV kernel.  The 64-bit versions pass. It gets a seg fault after
> "parent: write to 0x80 (should fail)", and the fault isn't caught by
> the signal handler.  It just dumps back to the shell.  The tests pass
> after reverting this.

I can reproduce it if I emulate an AMD machine.  I can "fix" it like this:

diff --git a/arch/x86/xen/xen-asm_64.S b/arch/x86/xen/xen-asm_64.S
index a8a4f4c460a6..6255e00f425e 100644
--- a/arch/x86/xen/xen-asm_64.S
+++ b/arch/x86/xen/xen-asm_64.S
@@ -97,6 +97,9 @@ ENDPROC(xen_syscall_target)
 ENTRY(xen_syscall32_target)
        popq %rcx
        popq %r11
+       movq $__USER32_DS, 4*8(%rsp)
+       movq $__USER32_CS, 1*8(%rsp)
+       movq %r11, 2*8(%rsp)
        jmp entry_SYSCALL_compat_after_hwframe
 ENDPROC(xen_syscall32_target)

but I haven't tried to diagnose precisely what's going on.

Xen seems to be putting the 0xe0?? values in ss and cs, which oughtn't
to be a problem, but it kills opportunistic sysretl.  Maybe that's
triggering a preexisting bug?