[PATCH v11 22/26] virt: gunyah: Add proxy-scheduled vCPUs

[Elliot Berman]

On 4/18/2023 10:31 AM, Alex Elder wrote:
> On 4/18/23 12:18 PM, Elliot Berman wrote:
>>
>>
>> On 4/17/2023 3:41 PM, Elliot Berman wrote:
>>>
>>>
>>> On 3/31/2023 7:27 AM, Alex Elder wrote:
>>>> On 3/3/23 7:06 PM, Elliot Berman wrote:
>>>
>>> [snip]
>>>
>>>>> diff --git a/include/uapi/linux/gunyah.h b/include/uapi/linux/gunyah.h
>>>>> index caeb3b3a3e9a..e52265fa5715 100644
>>>>> --- a/include/uapi/linux/gunyah.h
>>>>> +++ b/include/uapi/linux/gunyah.h
>>>>> @@ -62,8 +62,32 @@ struct gh_vm_dtb_config {
>>>>>   #define GH_VM_START        _IO(GH_IOCTL_TYPE, 0x3)
>>>>> +/**
>>>>> + * GH_FN_VCPU - create a vCPU instance to control a vCPU
>>>>> + *
>>>>> + * gh_fn_desc is filled with &struct gh_fn_vcpu_arg
>>>>> + *
>>>>> + * The vcpu type will register with the VM Manager to expect to 
>>>>> control
>>>>> + * vCPU number `vcpu_id`. It returns a file descriptor allowing 
>>>>> interaction with
>>>>> + * the vCPU. See the Gunyah vCPU API description sections for 
>>>>> interacting with
>>>>> + * the Gunyah vCPU file descriptors.
>>>>> + *
>>>>> + * Return: file descriptor to manipulate the vcpu. See GH_VCPU_* 
>>>>> ioctls
>>>>> + */
>>>>> +#define GH_FN_VCPU         1
>>>>
>>>> I think you should define GH_VN_VCPU, GN_FN_IRQFD, and GN_FN_IOEVENTFD
>>>> in an enumerated type.  Each has a type associated with it, and you can
>>>> add the explanation for the function in the kernel-doc comments above
>>>> thosse type definitions.
>>>>
>>>
>>> I'd like to enumify the GH_FN_* macros, but one challenge I'm facing 
>>> is that it breaks the module alias implementation in patch 19.
>>>
>>> MODULE_ALIAS("ghfunc:"__stringify(_type))
>>>
>>> When the GH_FN_* are regular preprocessor macros backed by an 
>>> integer, the preprocessor will make the module alias ghfunc:0 (or 
>>> ghfunc:1, etc). This works well because I can do
>>>
>>> request_module("ghfunc:%d", type);
>>>
>>> If the function hasn't been registered and then gunyah_vcpu.ko gets 
>>> loaded automatically.
>>>
>>> With enum, compiler knows the value of GH_FN_VCPU and preprocessor 
>>> will make the module alias like ghfunc:GH_FN_VCPU.
>>>
>>
>> I still like the idea of having enum for documentation and clarity. I 
>> noticed that nfnetlink.h saw the same problem for NFNL_SUBSYS_*.
>>
>> Is this compromise terrible and I should give up on the enum?
> 
> You know, I've seen this pattern in the kernel and never thought
> too much about why it was done.  Maybe this is exactly the reason.
> 
> It sure *seems* like there might be some macro magic that might
> cause the enum symbol's numeric value to be used but I think the
> problem is that enums are C tokens, which are not evaluated at
> preprocessor time.
> 
> You could probably skip the leading underscore, and do this as
> it's done for nfnetlink_groups in that same header file.
> 
> Maybe somebody else can confirm, or has a better suggestion.
> 

In the preprocessor macro case, the preprocessor macro GH_FN_VCPU 
expands to the GH_FN_VCPU enum value and stuck back as if I didn't have 
the preprocessor macro in first place. I'm not sure why the preprocessor 
macros are done for nfnetlink_groups. I saw one case where enum 
kvm_device_type does the same, but that might be done because it was 
converting preprocessor macro to enum.

Just a guess -- maybe the preprocessor macro was preserved to support 
userspace code doing this?

#ifdef KVM_DEV_TYPE_FSL_MPIC_20
...
#endif

>                      -Alex
> 
> 
>> enum gh_fn_type {
>> /* _GH_FN_* macro required for MODULE_ALIAS, otherwise __stringify() 
>> trick
>>   * won't work anymore */
>> #define _GH_FN_VCPU        1
>>      GH_FN_VCPU        = _GH_FN_VCPU,
>> #define _GH_FN_IRQFD        2
>>      GH_FN_IRQFD        = _GH_FN_IRQFD,
>> #define _GH_FN_IOEVENTFD    3
>>      GH_FN_IOEVENTFD        = _GH_FN_IOEVENTFD,
>> };
>>
>>> [snip]
>>>
>>>>> +
>>>>> +/*
>>>>> + * Gunyah presently sends max 4 bytes of exit_reason.
>>>>> + * If that changes, this macro can be safely increased without 
>>>>> breaking
>>>>> + * userspace so long as struct gh_vcpu_run < PAGE_SIZE.
>>>>
>>>> Is PAGE_SIZE allowed to be anything other than 4096 bytes?  Do you
>>>> expect this driver to work properly if the page size were configured
>>>> to be 16384 bytes?  In other words, is this a Gunyah constant, or
>>>> is it *really* the page size configured for Linux?
>>>>
>>>
>>> Our implementations are only doing 4096 bytes. I expect the driver to 
>>> work properly when using 16k pages. This really is a Linux page. It's 
>>> a reflection of the alloc_page in gh_vcpu_bind().
>>>
>>> The exit reason is copied from hypervisor into field accessible by 
>>> userspace directly. Gunyah makes the exit reason size dynamic -- 
>>> there's no architectural limitation preventing the exit reason from 
>>> being a string or some lengthy data.
>>>
>>> As I was writing this response, I realized that I should be able to 
>>> make this a zero-length array and ensure that reason[] doesn't 
>>> overflow PAGE_SIZE...
>>>
>>> The comment was trying to explain that Linux itself imposes a 
>>> limitation on the maximum exit reason size. If we need to support 
>>> longer exit reason, we're OK to do so long as the total size doesn't 
>>> overrun PAGE_SIZE. There aren't any plans to need longer exit reasons 
>>> than the 8 bytes mentioned today.
>>>
>>> Thanks,
>>> Elliot
>