Re: [PATCH] hyperv: Add CONFIG_MSHV_ROOT to gate hv_root_partition checks

[Nuno Das Neves <nunodasneves@linux.microsoft.com>]

On 2/12/2025 3:25 PM, Easwar Hariharan wrote:
> On 2/12/2025 3:01 PM, Nuno Das Neves wrote:
>> On 2/11/2025 9:47 PM, Easwar Hariharan wrote:
>>> On 2/11/2025 2:21 PM, Nuno Das Neves wrote:
>>>> Introduce CONFIG_MSHV_ROOT as a tristate to enable root partition
>>>> booting and future mshv driver functionality.
>>>>
>>>> Change hv_root_partition into a function which always returns false
>>>> if CONFIG_MSHV_ROOT=n.
>>>>
>>>> Introduce hv_current_partition_type to store the type of partition
>>>> (guest, root, or other kinds in future), and hv_identify_partition_type()
>>>> to it up early in Hyper-V initialization.
>>>
>>> ...to *set* it up early?
>>>
>> Yep! Thanks for catching that
>>
>>>>
>>>> Signed-off-by: Nuno Das Neves <nunodasneves@linux.microsoft.com>
>>>> ---
>>>> Depends on
>>>> https://lore.kernel.org/linux-hyperv/1738955002-20821-3-git-send-email-nunodasneves@linux.microsoft.com/
>>>>
>>>>  arch/arm64/hyperv/mshyperv.c       |  2 ++
>>>>  arch/x86/hyperv/hv_init.c          | 10 ++++----
>>>>  arch/x86/kernel/cpu/mshyperv.c     | 24 ++----------------
>>>>  drivers/clocksource/hyperv_timer.c |  4 +--
>>>>  drivers/hv/Kconfig                 | 12 +++++++++
>>>>  drivers/hv/Makefile                |  3 ++-
>>>>  drivers/hv/hv.c                    | 10 ++++----
>>>>  drivers/hv/hv_common.c             | 32 +++++++++++++++++++-----
>>>>  drivers/hv/vmbus_drv.c             |  2 +-
>>>>  drivers/iommu/hyperv-iommu.c       |  4 +--
>>>>  include/asm-generic/mshyperv.h     | 39 +++++++++++++++++++++++++-----
>>>>  11 files changed, 92 insertions(+), 50 deletions(-)
>>>>
>>>
>>> <snip>
>>>
>>>> +
>>>> +void hv_identify_partition_type(void)
>>>> +{
>>>> +	/*
>>>> +	 * Check partition creation and cpu management privileges
>>>> +	 *
>>>> +	 * Hyper-V should never specify running as root and as a Confidential
>>>> +	 * VM. But to protect against a compromised/malicious Hyper-V trying
>>>> +	 * to exploit root behavior to expose Confidential VM memory, ignore
>>>> +	 * the root partition setting if also a Confidential VM.
>>>> +	 */
>>>> +	if ((ms_hyperv.priv_high & HV_CREATE_PARTITIONS) &&
>>>> +	    (ms_hyperv.priv_high & HV_CPU_MANAGEMENT) &&
>>>> +	    !(ms_hyperv.priv_high & HV_ISOLATION)) {
>>>> +		hv_current_partition_type = HV_PARTITION_TYPE_ROOT;
>>>> +		pr_info("Hyper-V: running as root partition\n");
>>>> +	} else {
>>>> +		hv_current_partition_type = HV_PARTITION_TYPE_GUEST;
>>>> +	}
>>>> +}
>>>
>>> This should assume GUEST as default and modify to ROOT if all the checks pass.
>>>
>> It is doing that, isn't it?
>>
>> In fact the 'else' branch here is redundant and just there for additional clarity.
>>
>> hv_current_partition_type is zeroed (so GUEST) by default, but I could make that explicit
>> if you prefer:
> 
> Yes, explicit is better, but see comment below.
> 
>> +enum hv_partition_type hv_current_partition_type = HV_PARTITION_TYPE_GUEST;
>>
>> How does that sound? Am I misunderstanding something here?
> 
> I'd suggest centralizing that in this function, instead of having it spread in 2 places.
> Since your commit message hints at future partition types, it's ideal to have this function be
> a central clearing house, which I suppose is the intent. The preferred pattern in general, and what I'm
> suggesting, is something like this:
> 
> void hv_identify_partition_type(void)
> {
> 	/* Assume guest role */
> 	hv_current_partition_type = HV_PARTITION_TYPE_GUEST;
> 
> 	/*
> 	 * Check partition creation and cpu management privileges
> 	 *
> 	 * Hyper-V should never specify running as root and as a Confidential
> 	 * VM. But to protect against a compromised/malicious Hyper-V trying
> 	 * to exploit root behavior to expose Confidential VM memory, ignore
> 	 * the root partition setting if also a Confidential VM.
> 	 */
> 	if ((ms_hyperv.priv_high & HV_CREATE_PARTITIONS) &&
> 	    (ms_hyperv.priv_high & HV_CPU_MANAGEMENT) &&
> 	    !(ms_hyperv.priv_high & HV_ISOLATION)) {
> 		hv_current_partition_type = HV_PARTITION_TYPE_ROOT;
> 		pr_info("Hyper-V: running as root partition\n");
> 	}
> }
> 
Fair enough, happy to do it this way.

>>
>>> <snip>
>>>
>>>> +static inline int hv_call_deposit_pages(int node, u64 partition_id, u32 num_pages)
>>>> +{
>>>> +	return hv_result(U64_MAX);
>>>> +}
>>>
>>> Is there value in perhaps #defining hv_result_<whatever this is> as U64_MAX and returning that for documentation?
>>> For e.g. assuming this is something like EOPNOTSUPP
>>>
>>> #define HV_RESULT_NOT_SUPP U64_MAX
>>>
>>> static inline int hv_call_deposit_pages(int node, u64 partition_id, u32 num_pages)
>>> { return hv_result(HV_RESULT_NOT_SUPP); }
>>>
>> The idea here was to copy what hv_do_hypercall does returning U64_MAX in case the hypercall
>> page is missing, which will hv_result() into an invalid status code. A special value for
>> that status could make this pattern clearer.
> 
> Agreed, having a name for that status would be helpful, but we don't want to diverge too much from the hypervisor
> definitions, especially if we're going to change it later again anyway.
> 
>> I'd want to call out that this isn't a "real"
>> Hyper-V status code somehow. HV_STATUS's are 16 bits, so it would look more like:
>>
>> /* "LINUX" because this isn't really a status from the hypervisor.. */
>> #define HV_STATUS_LINUX_FAIL 0xFFFF
>> static inline int hv_call_deposit_pages(int node, u64 partition_id, u32 num_pages)
>> { return HV_STATUS_LINUX_FAIL; }
>>
>> Another option: there is another patch coming (which you know of) which maps hypercall
>> status codes to regular Linux errors like -EOPNOTSUPP. I could simply merge that patch
>> with this one (or make this a series for v2), and that would result in less churn.
>> (And leave alone the current use of U64_MAX in hv_do_hypercall, for now).
>>
> 
> I think that second option is a good idea. The hypervisor status should remain restricted to the functions that are
> hv_do_hypercall() or call it directly, while the rest of the code uses standard errno values. I'd suggest making
> it a series so each commit does 1 thing.
> 
Ok I'll do that, thanks.

Nuno

> Thanks,
> Easwar (he/him)