Why do we have both CONFIG_SOFTMMU and CONFIG_USER_ONLY?

Why do we have both CONFIG_SOFTMMU and CONFIG_USER_ONLY?

[Claudio Fontana]

As in Subject,

am I understanding correctly that the one or the other is redundant?

Should we keep only one of them?

Ciao,

Claudio

-- 
Claudio Fontana
Engineering Manager Virtualization, SUSE Labs Core

SUSE Software Solutions Italy Srl

Re: Why do we have both CONFIG_SOFTMMU and CONFIG_USER_ONLY?

[Paolo Bonzini]

On 07/12/20 12:19, Claudio Fontana wrote:
> As in Subject,
> 
> am I understanding correctly that the one or the other is redundant?
> 
> Should we keep only one of them?

In the code it's not strictly necessary.  However in both Makefiles and 
meson.build it made/makes things a bit clearer.

Paolo

Re: Why do we have both CONFIG_SOFTMMU and CONFIG_USER_ONLY?

[Claudio Fontana]

On 12/7/20 12:26 PM, Paolo Bonzini wrote:
> On 07/12/20 12:19, Claudio Fontana wrote:
>> As in Subject,
>>
>> am I understanding correctly that the one or the other is redundant?
>>
>> Should we keep only one of them?
> 
> In the code it's not strictly necessary.  However in both Makefiles and 
> meson.build it made/makes things a bit clearer.
> 
> Paolo
> 

K.. thanks a lot, maybe we should document all of this stuff in docs/devel/...

Ciao,

Claudio

Re: Why do we have both CONFIG_SOFTMMU and CONFIG_USER_ONLY?

[Peter Maydell]

On Mon, 7 Dec 2020 at 11:39, Claudio Fontana <cfontana@suse.de> wrote:
>
> As in Subject,
>
> am I understanding correctly that the one or the other is redundant?
>
> Should we keep only one of them?

I think that perhaps the idea at one point was that we
might have a version of linux-user which used a softmmu
(this would allow better control of the guest's view of
its address space, so guest mmap() to fixed addresses
would work better, for instance). But nobody's ever actually
tried to implement that, so I imagine that if we ever did
we'd find that some CONFIG_SOFTMMU and some CONFIG_USER_ONLY
defines were the wrong way around...

thanks
-- PMM

Re: Why do we have both CONFIG_SOFTMMU and CONFIG_USER_ONLY?

[Philippe Mathieu-Daudé]

On 12/7/20 12:27 PM, Claudio Fontana wrote:
> On 12/7/20 12:26 PM, Paolo Bonzini wrote:
>> On 07/12/20 12:19, Claudio Fontana wrote:
>>> As in Subject,
>>>
>>> am I understanding correctly that the one or the other is redundant?
>>>
>>> Should we keep only one of them?

CONFIG_SOFTMMU is poisoned while CONFIG_USER_ONLY isn't.
I doubt this is a design decision, probably an unfinished
transition? Anyhow we ended using CONFIG_USER_ONLY where
CONFIG_SOFTMMU triggers a poisoned error.

>>
>> In the code it's not strictly necessary.  However in both Makefiles and 
>> meson.build it made/makes things a bit clearer.
>>
>> Paolo
>>
> 
> K.. thanks a lot, maybe we should document all of this stuff in docs/devel/...
> 
> Ciao,
> 
> Claudio
> 

Re: Why do we have both CONFIG_SOFTMMU and CONFIG_USER_ONLY?

[Claudio Fontana]

On 12/7/20 12:50 PM, Peter Maydell wrote:
> On Mon, 7 Dec 2020 at 11:39, Claudio Fontana <cfontana@suse.de> wrote:
>>
>> As in Subject,
>>
>> am I understanding correctly that the one or the other is redundant?
>>
>> Should we keep only one of them?
> 
> I think that perhaps the idea at one point was that we
> might have a version of linux-user which used a softmmu
> (this would allow better control of the guest's view of
> its address space, so guest mmap() to fixed addresses
> would work better, for instance). But nobody's ever actually
> tried to implement that, so I imagine that if we ever did
> we'd find that some CONFIG_SOFTMMU and some CONFIG_USER_ONLY
> defines were the wrong way around...
> 
> thanks
> -- PMM
> 

Hi Peter,

thanks for the background,

indeed I am seeing some of these cases, target/XXX/cpu.c is protecting code with #ifndef CONFIG_USER_ONLY,

but the header files in include/hw/core/cpu.h and others use #ifdef CONFIG_SOFTMMU.

Ciao,

CLaudio

Re: Why do we have both CONFIG_SOFTMMU and CONFIG_USER_ONLY?

[Richard Henderson]

On 12/7/20 5:56 AM, Claudio Fontana wrote:
> On 12/7/20 12:50 PM, Peter Maydell wrote:
>> On Mon, 7 Dec 2020 at 11:39, Claudio Fontana <cfontana@suse.de> wrote:
>>>
>>> As in Subject,
>>>
>>> am I understanding correctly that the one or the other is redundant?
>>>
>>> Should we keep only one of them?
>>
>> I think that perhaps the idea at one point was that we
>> might have a version of linux-user which used a softmmu
>> (this would allow better control of the guest's view of
>> its address space, so guest mmap() to fixed addresses
>> would work better, for instance). But nobody's ever actually
>> tried to implement that, so I imagine that if we ever did
>> we'd find that some CONFIG_SOFTMMU and some CONFIG_USER_ONLY
>> defines were the wrong way around...
>>
>> thanks
>> -- PMM
>>
> 
> Hi Peter,
> 
> thanks for the background,
> 
> indeed I am seeing some of these cases, target/XXX/cpu.c is protecting code with #ifndef CONFIG_USER_ONLY,
> 
> but the header files in include/hw/core/cpu.h and others use #ifdef CONFIG_SOFTMMU.

In an ideal future in which linux-user could use softmmu, I would have softmmu
be a runtime option rather than a compile-time option.  The option would merely
affect how TCG generates code.

So while in that ideal future only CONFIG_USER_ONLY would remain, in an ideal
present CONFIG_SOFTMMU would mark those places where a runtime check should be
added.

But the present is not ideal, and system-ness (or non-user-only-ness) and
actual mmu emulation are often confused.  Cleanups welcome.  ;-)


r~