Re: [XEN PATCH v1] xen: mem_access: conditionally compile vm_event.c & monitor.c

[Jan Beulich <jbeulich@xxxxxxxx>]

On 06.01.2025 19:09, Tamas K Lengyel wrote:
> On Mon, Jan 6, 2025 at 10:10 AM Jan Beulich <jbeulich@xxxxxxxx> wrote:
>>
>> On 06.01.2025 15:05, Tamas K Lengyel wrote:
>>> On Mon, Jan 6, 2025 at 5:16 AM Jan Beulich <jbeulich@xxxxxxxx> wrote:
>>>>
>>>> On 30.12.2024 07:30, Sergiy Kibrik wrote:
>>>>> From: Stefano Stabellini <stefano.stabellini@xxxxxxx>
>>>>>
>>>>> Extend coverage of CONFIG_MEM_ACCESS option and make the build of VM 
>>>>> events
>>>>> and monitoring support optional.
>>>>
>>>> Yet doesn't this end up in things becoming misleading? Don't we rather 
>>>> need a
>>>> 2nd Kconfig option, with a dependency between the two? Or alternatively a
>>>> rename of the existing option (to describe the higher-level feature rather
>>>> than the lower level one)? Tamas, I'm particularly interested in knowing 
>>>> your
>>>> view here as well.
>>>
>>> Thanks Jan, I was thinking the same thing. The dependency of these
>>> subsystems is mem_access -> monitor -> vm_event. If the goal here is
>>> to disable all three levels the ideal way would be to have separate
>>> kconfig options for each level. It may be a bit too fine-grained
>>> though on ARM since there are only two types of events for monitor
>>> (SMC & mem_access) and only the monitor uses the vm_event channel (no
>>> mem-sharing/paging on ARM). So if doing separate kconfig for each
>>> individual feature is an overkill I would suggest using
>>> CONFIG_VM_EVENT that disables all three levels, including both
>>> mem_access & smc monitor hooks.
>>
>> Except that "disables all three levels" doesn't work, unless the other
>> option(s) are promptless (and selected). I'd have expected VM_EVENT to
>> maybe have a "depends on MEM_ACCESS", whereas a "select MEM_ACCESS"
>> wouldn't make much sense as long as MEM_ACCESS can be enabled
>> individually (with it being unclear to me whether such a configuration
>> is actually useful in any way).
> 
> Not sure I follow. None of these systems make sense to enable
> individually. Without vm_event monitor/mem_access are useless, that's
> why I would pick CONFIG_VM_EVENT as the option on ARM to disable all
> three levels if we don't want to start splitting it into multiple
> kconfig options (which I think may be an overkill here).

Oh, okay, you suggest to replace MEM_ACCESS by VM_EVENT at the Kconfig
level. That would be fine with me, so long as it's also appropriate on
(in particular) x86. Then, if there was ever a 2nd use of mem-access,
MEM_ACCESS could be re-introduced as a standalone option.

Jan