Re: [PATCH v2 4/6] xen_pvdev: Do not assume Dom0 when creating a directory

[Volodymyr Babchuk]

Hi Paul,

Paul Durrant <xadimgnik@gmail.com> writes:

> On 23/11/2023 00:07, Volodymyr Babchuk wrote:
>> Hi,
>> Volodymyr Babchuk <volodymyr_babchuk@epam.com> writes:
>> 
>>> Hi Stefano,
>>>
>>> Stefano Stabellini <sstabellini@kernel.org> writes:
>>>
>>>> On Wed, 22 Nov 2023, David Woodhouse wrote:
>>>>> On Wed, 2023-11-22 at 15:09 -0800, Stefano Stabellini wrote:
>>>>>> On Wed, 22 Nov 2023, David Woodhouse wrote:
>>>>>>> On Wed, 2023-11-22 at 14:29 -0800, Stefano Stabellini wrote:
>>>>>>>> On Wed, 22 Nov 2023, Paul Durrant wrote:
>>>>>>>>> On 21/11/2023 22:10, Volodymyr Babchuk wrote:
>>>>>>>>>> From: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>
>>>>>>>>>>
>>>>>>>>>> Instead of forcing the owner to domid 0, use XS_PRESERVE_OWNER to
>>>>>>>>>> inherit the owner of the directory.
>>>>>>>>>
>>>>>>>>> Ah... so that's why the previous patch is there.
>>>>>>>>>
>>>>>>>>> This is not the right way to fix it. The QEMU Xen support is 
>>>>>>>>> *assuming* that
>>>>>>>>> QEMU is either running in, or emulating, dom0. In the emulation case 
>>>>>>>>> this is
>>>>>>>>> probably fine, but the 'real Xen' case it should be using the correct 
>>>>>>>>> domid
>>>>>>>>> for node creation. I guess this could either be supplied on the 
>>>>>>>>> command line
>>>>>>>>> or discerned by reading the local domain 'domid' node.
>>>>>>>>
>>>>>>>> yes, it should be passed as command line option to QEMU
>>>>>>>
>>>>>>> I'm not sure I like the idea of a command line option for something
>>>>>>> which QEMU could discover for itself.
>>>>>>
>>>>>> That's fine too. I meant to say "yes, as far as I know the toolstack
>>>>>> passes the domid to QEMU as a command line option today".
>>>>>
>>>>> The -xen-domid argument on the QEMU command line today is the *guest*
>>>>> domain ID, not the domain ID in which QEMU itself is running.
>>>>>
>>>>> Or were you thinking of something different?
>>>>
>>>> Ops, you are right and I understand your comment better now. The backend
>>>> domid is not on the command line but it should be discoverable (on
>>>> xenstore if I remember right).
>>>
>>> Yes, it is just "~/domid". I'll add a function that reads it.
>> Just a quick question to QEMU folks: is it better to add a global
>> variable where we will store own Domain ID or it will be okay to read
>> domid from Xenstore every time we need it?
>> If global variable variant is better, what is proffered place to
>> define
>> this variable? system/globals.c ?
>> 
>
> Actually... is it possible for QEMU just to use a relative path for
> the backend nodes? That way it won't need to know it's own domid, will
> it?

Well, it is possible to use relative path, AFAIK, linux-based backends
are doing exactly this. But problem is with xenstore_mkdir() function,
which requires domain id to correctly set owner when it causes call to
set_permissions().

As David said, architecturally it will be better to get rid of
xenstore_mkdir() usage, because it is used by legacy backends only. But
to do this, someone needs to convert legacy backends to use newer API. I
have no capacity to do this right now, so I implemented a contained
solution:

static int xenstore_read_own_domid(unsigned int *domid)

in xen_pvdev.c. I believe, this new function will be removed along with
whole xen_pvdev.c when there will be no legacy backends left.

-- 
WBR, Volodymyr