Re: [RFC PATCH v1 07/37] KVM: Introduce KVM_SET_MEMORY_ATTRIBUTES2

[Ackerley Tng]

Sean Christopherson <seanjc@xxxxxxxxxx> writes:

> On Fri, Oct 24, 2025, Ackerley Tng wrote:
>> Sean Christopherson <seanjc@xxxxxxxxxx> writes:
>> >> 
>> >> [...snip...]
>> >> 
>> 
>> I've been thinking more about this:
>> 
>>   #ifdef CONFIG_KVM_VM_MEMORY_ATTRIBUTES
>>   	case KVM_CAP_MEMORY_ATTRIBUTES2:
>>   	case KVM_CAP_MEMORY_ATTRIBUTES:
>>   		if (!vm_memory_attributes)
>>   			return 0;
>>   
>>   		return kvm_supported_mem_attributes(kvm);
>>   #endif
>> 
>> And the purpose of adding KVM_CAP_MEMORY_ATTRIBUTES2 is that
>> KVM_CAP_MEMORY_ATTRIBUTES2 tells userspace that
>> KVM_SET_MEMORY_ATTRIBUTES2 is available iff there are valid
>> attributes.
>> 
>> (So there's still a purpose)
>> 
>> Without valid attributes, userspace can't tell if it should use
>> KVM_SET_MEMORY_ATTRIBUTES or the 2 version.
>
> To do what?  If there are no attributes, userspace can't do anything useful anyways.
>
>> I also added KVM_CAP_GUEST_MEMFD_MEMORY_ATTRIBUTES, which tells
>> userspace the valid attributes when calling KVM_SET_MEMORY_ATTRIBUTES2
>> on a guest_memfd:
>
> Ya, and that KVM_SET_MEMORY_ATTRIBUTES2 is supported on guest_memfd.
>
>>   #ifdef CONFIG_KVM_GUEST_MEMFD
>>   	case KVM_CAP_GUEST_MEMFD:
>>   		return 1;
>>   	case KVM_CAP_GUEST_MEMFD_FLAGS:
>>   		return kvm_gmem_get_supported_flags(kvm);
>>   	case KVM_CAP_GUEST_MEMFD_MEMORY_ATTRIBUTES:
>>   		if (vm_memory_attributes)
>>   			return 0;
>>   
>>   		return kvm_supported_mem_attributes(kvm);
>>   #endif
>>   
>> So to set memory attributes, userspace should
>
> Userspace *can*.  User could also decide it only wants to support guest_memfd
> attributes, e.g. because the platform admins controls the entire stack and built
> their entire operation around in-place conversion.
>
>>   if (kvm_check_cap(KVM_CAP_GUEST_MEMFD_MEMORY_ATTRIBUTES) > 0)
>> 	use KVM_SET_MEMORY_ATTRIBUTES2 with guest_memfd
>>   else if (kvm_check_cap(KVM_CAP_MEMORY_ATTRIBUTES2) > 0)
>>         use KVM_SET_MEMORY_ATTRIBUTES2 with VM fd
>>   else if (kvm_check_cap(KVM_CAP_MEMORY_ATTRIBUTES) > 0)
>> 	use KVM_SET_MEMORY_ATTRIBUTES with VM fd
>>   else
>> 	can't set memory attributes
>> 
>> Something like that?
>
> More or else, ya.
>
>> In selftests there's this, when KVM_SET_USER_MEMORY_REGION2 was
>> introduced:
>> 
>>   #define TEST_REQUIRE_SET_USER_MEMORY_REGION2()			\
>> 	__TEST_REQUIRE(kvm_has_cap(KVM_CAP_USER_MEMORY2),	\
>> 		       "KVM selftests now require KVM_SET_USER_MEMORY_REGION2 (introduced in v6.8)")
>> 
>> But looks like there's no direct equivalent for the introduction of
>> KVM_SET_MEMORY_ATTRIBUTES2?
>
> KVM_CAP_USER_MEMORY2 is the equivalent.
>
> There's was no need to enumerate anything beyond yes/no, because
> SET_USER_MEMORY_REGION2 didn't introduce new flags, it expanded the size of the
> structure passed in from userspace so that KVM_CAP_GUEST_MEMFD could be introduced
> without breaking backwards compatibility.
>
>> The closest would be to add a TEST_REQUIRE_VALID_ATTRIBUTES() which
>> checks KVM_CAP_MEMORY_ATTRIBUTES2 or
>> KVM_CAP_GUEST_MEMFD_MEMORY_ATTRIBUTES before making the vm or
>> guest_memfd ioctl respsectively.
>
> Yes.  This is what I did in my (never posted, but functional) version:
>
> @@ -486,6 +488,7 @@ struct kvm_vm *__vm_create(struct vm_shape shape, uint32_t nr_runnable_vcpus,
>         }
>         guest_rng = new_guest_random_state(guest_random_seed);
>         sync_global_to_guest(vm, guest_rng);
> +       sync_global_to_guest(vm, kvm_has_gmem_attributes);

I ported this [1] except for syncing this value to the guest, because I
think the guest shouldn't need to know this information, the host should
decide what to do. I think, if the guests really need to know this, the
test itself can do the syncing.

[1] https://lore.kernel.org/all/5656d432df1217c08da0cc2694fd79948bfd686f.1760731772.git.ackerleytng@xxxxxxxxxx/

>  
>         kvm_arch_vm_post_create(vm, nr_runnable_vcpus);
>  
> @@ -2319,6 +2333,8 @@ void __attribute((constructor)) kvm_selftest_init(void)
>         guest_random_seed = last_guest_seed = random();
>         pr_info("Random seed: 0x%x\n", guest_random_seed);
>  
> +       kvm_has_gmem_attributes = kvm_has_cap(KVM_CAP_GUEST_MEMFD_MEMORY_ATTRIBUTES);
> +
>         kvm_selftest_arch_init();
>  }
>  
> That way the core library code can pivot on gmem vs. VM attributes without having
> to rely on tests to define anything.  E.g.
>
> static inline void vm_mem_set_memory_attributes(struct kvm_vm *vm, uint64_t gpa,
> 						uint64_t size, uint64_t attrs)
> {
> 	if (kvm_has_gmem_attributes) {
> 		off_t fd_offset;
> 		uint64_t len;
> 		int fd;
>
> 		fd = kvm_gpa_to_guest_memfd(vm, gpa, &fd_offset, &len);
> 		TEST_ASSERT(len >= size, "Setting attributes beyond the length of a guest_memfd");
> 		gmem_set_memory_attributes(fd, fd_offset, size, attrs);
> 	} else {
> 		vm_set_memory_attributes(vm, gpa, size, attrs);
> 	}
> }