Re: [PATCH] xen/x86: Adjust stack pointer in xen_sysexit

From: Andrew Cooper <andrew.cooper3@citrix.com>
To: Andy Lutomirski <luto@amacapital.net>
Cc: "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	xen-devel <xen-devel@lists.xen.org>,
	Borislav Petkov <bp@alien8.de>,
	David Vrabel <david.vrabel@citrix.com>,
	"H. Peter Anvin" <hpa@zytor.com>,
	Boris Ostrovsky <boris.ostrovsky@oracle.com>
Subject: Re: [PATCH] xen/x86: Adjust stack pointer in xen_sysexit
Date: Tue, 17 Nov 2015 19:29:15 +0000	[thread overview]
Message-ID: <564B800B.1060006__47593.984186142$1447788668$gmane$org@citrix.com> (raw)
In-Reply-To: <CALCETrXv2=JUJvcx0_wmd8bQQ8UUAOyhX+_LTKch8RaJFspPcw@mail.gmail.com>

On 17/11/15 19:16, Andy Lutomirski wrote:
> On Tue, Nov 17, 2015 at 11:12 AM, Andrew Cooper
> <andrew.cooper3@citrix.com> wrote:
>> On 17/11/15 18:49, Andy Lutomirski wrote:
>>> On Nov 17, 2015 6:40 AM, "Boris Ostrovsky" <boris.ostrovsky@oracle.com> wrote:
>>>> On 11/16/2015 04:55 PM, H. Peter Anvin wrote:
>>>>> On 11/16/15 12:22, Borislav Petkov wrote:
>>>>>> Huh, so what's wrong with a jump:
>>>>>>
>>>>>>         jmp 1f
>>>>>>         swapgs
>>>>>>         1:
>>>>>>
>>>>> What is the point of that jump?
>>>>>
>>>>>>> If it would make you feel better, it could be X86_BUG_XENPV :-p
>>>>>> That doesn't matter - I just don't want to open the flood gates on
>>>>>> pseudo feature bits.
>>>>>>
>>>>>> hpa, what do you think?
>>>>> Pseudo feature bits are fine, we already have plenty of them.  They make
>>>>> sense as they let us reuse a lot of infrastructure.
>>>>
>>>> So how about something like this? And then I think we can remove usergs_sysret32 and irq_enable_sysexit pv ops completely as noone will use them (lguest doesn't set them)
>>>>
>>> Looks good to me.  Does Xen have any sysexit/sysret32 equivalent to
>>> return to 32-bit user mode?  If so, it could be worth trying to wire
>>> it up by patching the jz instead of the test instruction.
>> From the guests point of view, there is only hypercall_iret.
> Doesn't hypercall_iret have flags that ask for different behavior,
> though?  (VG_syscall or whatever for the 64-bit case?)

The one and only flag is VGCF_in_syscall

Xen has its own logic for choosing between sysretq/sysretl if
VGCF_in_syscall, but will end up on an iret path in all other
circumstances. 

There is definitely some room for optimisation here, but in in some
copious free time before that, I want to see about brining most of our
asm code up into C.  The vast majority of it doesn't need to be written
in asm.

>
>>> Also, I'd prefer X86_FEATURE_XENPV.  IMO "PV" means too many things to
>>> too many people.
>> I agree - PV on its own is too generic.
>>
>> An alternative might be X86_FEATURE_XEN_PV_GUEST which is very clear an
>> unambiguous, although rather longer.
> Works for me, too, although seeing "xen_pv_host" in the Linux cpu
> features would be very strange indeed :)

This makes me wonders whether the `insmod xen` project has managed to
gain any traction ;)

~Andrew