Re: [PATCH v19 05/20] KVM: arm64: Track the type of VM in kvm_arch

[Gavin Shan <gshan@redhat.com>]

On 9/24/26 6:48 PM, Suzuki K Poulose wrote:
> On 24/09/2026 02:11, Gavin Shan wrote:
>> On 9/24/26 7:37 AM, Suzuki K Poulose wrote:
>>> On 23/09/2026 17:27, Suzuki K Poulose wrote:
>>>> On 23/09/2026 14:23, Gavin Shan wrote:
>>>>> On 9/23/26 8:24 PM, Suzuki K Poulose wrote:
>>>>>> On 23/09/2026 07:19, Gavin Shan wrote:
>>>>>>> On 9/23/26 4:05 PM, Gavin Shan wrote:
>>>>>>>> On 9/21/26 7:28 AM, Suzuki K Poulose wrote:
>>>>>>>>> KVM arm64 has different types of VMs with all the different modes in which
>>>>>>>>> the hypervisor code can be run. e.g., VHE, nVHE, pKVM etc. Then there is
>>>>>>>>> protected VM and normal VMs with pKVM. We might soon add other types,
>>>>>>>>> e.g., Arm CCA Realm. So in an effort to make the handling of these
>>>>>>>>> different types of VMs a bit more friendly to the eyes, add a VM flavor to
>>>>>>>>> the kvm_arch and we could then add handlers for different operations based
>>>>>>>>> on the VM type.
>>>>>>>>>
>>>>>>>>> Keep the flavor initialisation at the beginning to allow for the detection
>>>>>>>>> early enough and fail out on any unsupported requests.
>>>>>>>>>
>>>>>>>>> With that, add wrappers for checking the "type" of a VM and replace the
>>>>>>>>> existing users with the new wrappers.
>>>>>>>>>
>>>>>>>>> Given we already have the construct of "kvm_vm_is_protected" in the core
>>>>>>>>> KVM code, use that for all confidential compute guests including Realms
>>>>>>>>> that we are about to add.
>>>>>>>>>
>>>>>>>>> Adds __VM_PROTECTED marker vm flavor to generalize kvm_vm_is_protected()
>>>>>>>>> to predicate all confidential guests running on KVM. In later patches, we
>>>>>>>>> would add Realm VMs, which would also be classified as protected.
>>>>>>>>>
>>>>>>>>> Add explicit helper to detect if a given VM is a "protected" VM under pKVM.
>>>>>>>>> Change the existing users that precisely want to check the VM type. These
>>>>>>>>> include :
>>>>>>>>>    - kvm_arch_prepare_memory_region - For preventing memslot changes after
>>>>>>>>>      pVM creation.
>>>>>>>>>
>>>>>>>>> All the others are retained as a wider check for confidential guest VMs.
>>>>>>>>> These are:
>>>>>>>>>   - kvm_vm_ioctl_set_counter_offset - For disallowing timer offset
>>>>>>>>>     configuration
>>>>>>>>>   - io_mem_abort for dabt handling without valid syndrome information
>>>>>>>>>
>>>>>>>>> Both of which are true for Realms too.
>>>>>>>>>
>>>>>>>>> Realms support is restricted to VHE host and thus "kvm_vm_is_protected()"
>>>>>>>>> checks in the pkvm hyp specific code doesn't need to change, as the only
>>>>>>>>> protected guests it deals with is "protected pKVM" guests. To tighten this
>>>>>>>>> init_pkvm_hyp_vm() restricts the hyp copy of the vm_flavor to the ones it
>>>>>>>>> supports.
>>>>>>>>>
>>>>>>>>> Suggested-by: Marc Zyngier <maz@kernel.org>
>>>>>>>>> Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com>
>>>>>>>>> ---
>>>>>>>>> Changes since v18:
>>>>>>>>>   - Merge the __VM_PROTECTED marker and the widening of kvm_vm_is_protected()
>>>>>>>>>     to this patch.
>>>>>>>>>   - Merge the use of kvm_vm_is_unprotected_pkvm() for ! kvm_vm_is_protected()
>>>>>>>>>     given the scope changes here.
>>>>>>>>>   - Drop Fuad's review tag, as this patch has multiple merges
>>>>>>>>>   - Restrict the VM flavors to the supported types in init_pkvm_hyp_vm().
>>>>>>>>>   - Drop kvm_vm_hyp_is_pkvm() and revert to is_protected_kvm_enabled()
>>>>>>>>>   - Use is_protected_kvm_enabled() to make the pKVM guest flavor checks.
>>>>>>>>>   - s/PKVM/pKVM for commit descriptions too
>>>>>>>>>
>>>>>>>>> Changes since v17:
>>>>>>>>>   * s/PKVM/pKVM for the comments
>>>>>>>>>   * Drop type argument for pkvm_init_host_vm and also drop protected variable.
>>>>>>>>>   * Add helpers for checking if the VM is running on pKVM (kvm_vm_hyp_is_pkvm())
>>>>>>>>>   * Use kvm_vm_hyp_is_pkvm() to replace is_protected_kvm_enabled() with valid
>>>>>>>>>     kvm instance
>>>>>>>>> ---
>>>>>>>>>   arch/arm64/include/asm/kvm_host.h | 22 +++++++++++++++++++---
>>>>>>>>>   arch/arm64/include/asm/kvm_pkvm.h |  4 ++--
>>>>>>>>>   arch/arm64/kvm/arm.c              | 31 +++++++++++++++++++++++ ++ +-----
>>>>>>>>>   arch/arm64/kvm/handle_exit.c      |  2 +-
>>>>>>>>>   arch/arm64/kvm/hyp/nvhe/pkvm.c    |  6 +++++-
>>>>>>>>>   arch/arm64/kvm/mmu.c              |  2 +-
>>>>>>>>>   arch/arm64/kvm/pkvm.c             |  6 ++----
>>>>>>>>>   7 files changed, 56 insertions(+), 17 deletions(-)
>>>>>>>>>
>>>>>>>>
>>>>>>>> This causes broken nVHE mode. I applied PATCH[01-05] to v7.3.rc4 whose head commit
>>>>>>>> is f0100363d8c3, and kselftests/kvm/guest_print causes host crash (as below). I don't
>>>>>>>> see the crash if only PATCH[01-04] are applied.
>>>>>>>>
>>>>>>>> host$ cat /proc/cmdline | grep kvm-arm\.mode
>>>>>>>> BOOT_IMAGE=(hd0,gpt2)/vmlinuz-7.3.0-rc4-gavin+ root=/dev/mapper/ rhel_nvidia--grace--hopper--01-root ro crashkernel=2G-4G:406M,4G-64G:470M,64G-:726M rd.lvm.lv=rhel_nvidia- grace-hopper-01/root rd.lvm.lv=rhel_nvidia-grace-hopper-01/swap video=simplefb:off kvm-arm.mode=nvhe
>>>>>>>>
>>>>>>>> host$ cd linux/tools/testing/selftests/kvm
>>>>>>>> host$ ./guest_print_test
>>>>>>>> Random seed: 0x193a0ed3
>>>>>>>> [  192.754328] kvm [6674]: nVHE hyp panic at: [<ffff8000811da044>] __kvm_nvhe___timer_enable_traps+0x4/0x160!
>>>>>>>> [  192.754338] kvm [6674]: nVHE call trace:
>>>>>>>> [  192.754339] kvm [6674]:  [<ffff8000811dcf34>] __kvm_nvhe_hyp_panic+0xb4/0xe0
>>>>>>>> [  192.754342] kvm [6674]:  [<ffff8000811dcba4>] __kvm_nvhe___kvm_vcpu_run+0x164/0x440
>>>>>>>> [  192.754344] kvm [6674]:  [<ffff8000811dfe90>] __kvm_nvhe_handle___kvm_vcpu_run+0x40/0x1f0
>>>>>>>> [  192.754346] kvm [6674]:  [<ffff8000811e0238>] __kvm_nvhe_handle_trap+0x158/0x280
>>>>>>>> [  192.754347] kvm [6674]:  [<ffff8000811dd8fc>] __kvm_nvhe___skip_pauth_save+0x4/0x4
>>>>>>>> [  192.754348] kvm [6674]: ---[ end nVHE call trace ]---
>>>>>>>> [  192.754350] Code: d2818002 17fffff7 d503201f f9400001 (b94a6821)
>>>>>>>> [  192.754350] kvm [6674]: Hyp Offset: 0xfffeb0d7fe2e0000
>>>>>>>> [  192.754351] Kernel panic - not syncing: HYP panic:
>>>>>>>> [  192.754351] PS:834003c9 PC:0000cf2882efa044 ESR:0000000096000004
>>>>>>>> [  192.754351] FAR:ffff00009b286a68 HPFAR:8000000000000000 PAR:1d00ec7edbadc8de
>>>>>>>> [  192.754351] VCPU:0000cf011848a350
>>>>>>>> [  192.845154] CPU: 12 UID: 0 PID: 6674 Comm: guest_print_tes Kdump: loaded Not tainted 7.3.0-rc4-gavin+ #7 PREEMPT(full)
>>>>>>>> [  192.856182] Hardware name:  GH200 P5042, BIOS 02.04.01 20250422
>>>>>>>> [  192.862231] Call trace:
>>>>>>>> [  192.864725]  show_stack+0x20/0x38 (C)
>>>>>>>> [  192.868471]  dump_stack_lvl+0x88/0xb8
>>>>>>>> [  192.872215]  dump_stack+0x18/0x30
>>>>>>>> [  192.875598]  vpanic+0x280/0x498
>>>>>>>> [  192.878806]  panic+0x68/0x70
>>>>>>>> [  192.881745]  nvhe_hyp_panic_handler+0x184/0x190
>>>>>>>> [  192.886372]  kvm_arm_vcpu_enter_exit+0x24/0x100
>>>>>>>> [  192.891003]  kvm_arch_vcpu_ioctl_run+0x254/0x7c0
>>>>>>>> [  192.895726]  kvm_vcpu_ioctl+0x174/0xb40
>>>>>>>> [  192.899645]  __arm64_sys_ioctl+0xb0/0x120
>>>>>>>> [  192.903745]  invoke_syscall.constprop.0+0xa8/0x100
>>>>>>>> [  192.908639]  do_el0_svc+0xb8/0xe0
>>>>>>>> [  192.912022]  el0_svc+0x48/0x1f8
>>>>>>>> [  192.915228]  el0t_64_sync_handler+0xa0/0xe8
>>>>>>>> [  192.919500]  el0t_64_sync+0x1ac/0x1b0
>>>>>>>> [  192.923243] SMP: stopping secondary CPUs
>>>>>>>> [  192.927653] Starting crashdump kernel...
>>>>>>>> [  192.931658] Bye!
>>>>>>>>
>>>>>>>
>>>>>>> With the following changes applied after PATCH[01-05] on top of v7.3.rc4, the crash
>>>>>>> is avoided.
>>>>>>>
>>>>>>> In arch/arm64/include/asm/kvm_host.h:
>>>>>>>
>>>>>>> -#define kvm_vm_is_protected(kvm)       ((kvm)->arch.vm_flavor >= __VM_PROTECTED)
>>>>>>> +#define kvm_vm_is_protected(kvm)       \
>>>>>>> +       (is_protected_kvm_enabled() && (kvm)->arch.vm_flavor >= __VM_PROTECTED)
>>>>>>
>>>>>> That only papers over the problem. We can't use the vcpu_is_*
>>>>>> constructs from nvhe hyp, without converting the vcpu->kvm to
>>>>>> the hyp address, before accessing it. The fix is a bit more
>>>>>> involved. One option is to define the vcpu_is_* helpers only
>>>>>> for the !NVHE hyp code in the kvm_host.h (to avoid pulling
>>>>>> in the asm/kvm_mmu.h in to kvm_host.h and then make a mess
>>>>>> with header dependencies) and define the NVHE version in
>>>>>> asm/kvm_hyp.h. Something like :
>>>>>>
>>>>>>
>>>>>> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/ include/ asm/kvm_host.h
>>>>>> index 46a7f6c1e426c..4016d09ed39c3 100644
>>>>>> --- a/arch/arm64/include/asm/kvm_host.h
>>>>>> +++ b/arch/arm64/include/asm/kvm_host.h
>>>>>> @@ -1545,11 +1545,14 @@ struct kvm *kvm_arch_alloc_vm(void);
>>>>>>   #define __KVM_HAVE_ARCH_FLUSH_REMOTE_TLBS_RANGE
>>>>>>
>>>>>>   #define kvm_vm_is_protected(kvm)       ((kvm)->arch.vm_flavor >= __VM_PROTECTED)
>>>>>> +
>>>>>> +#ifndef __KVM_NVHE_HYPERVISOR__
>>>>>>   #define vcpu_is_protected(vcpu) kvm_vm_is_protected((vcpu)->kvm)
>>>>>> +#define vcpu_is_protected_pkvm(vcpu) kvm_vm_is_protected_pkvm((vcpu)- >kvm)
>>>>>> +#endif
>>>>>>
>>>>>>   #define kvm_vm_is_protected_pkvm(kvm)          \
>>>>>>          (is_protected_kvm_enabled() && ((kvm)->arch.vm_flavor == VM_PROTECTED_PKVM))
>>>>>> -#define vcpu_is_protected_pkvm(vcpu) kvm_vm_is_protected_pkvm(vcpu- >kvm)
>>>>>>
>>>>>>   #define kvm_vm_is_unprotected_pkvm(kvm)                \
>>>>>>          (is_protected_kvm_enabled() && ((kvm)->arch.vm_flavor == VM_PKVM))
>>>>>> diff --git a/arch/arm64/include/asm/kvm_hyp.h b/arch/arm64/include/ asm/kvm_hyp.h
>>>>>> index 4974492744cc8..3bf87e81af430 100644
>>>>>> --- a/arch/arm64/include/asm/kvm_hyp.h
>>>>>> +++ b/arch/arm64/include/asm/kvm_hyp.h
>>>>>> @@ -137,6 +137,19 @@ int __pkvm_init(phys_addr_t phys, unsigned long size, unsigned long *per_cpu_bas
>>>>>>   void __noreturn __host_enter(struct kvm_cpu_context *host_ctxt);
>>>>>>   #endif
>>>>>>
>>>>>> +#ifdef __KVM_NVHE_HYPERVISOR__
>>>>>> +#define vcpu_is_protected(vcpu)        \
>>>>>> + ({                                                              \
>>>>>> +               struct kvm *__kvm = READ_ONCE((vcpu)- >kvm);             \
>>>>>> +               __kvm && kvm_vm_is_protected((kern_hyp_va(__kvm)));     \
>>>>>> +       })
>>>>>> +#define vcpu_is_protected_pkvm(vcpu)                                   \
>>>>>> + ({                                                              \
>>>>>> +               struct kvm *__kvm = READ_ONCE((vcpu)- >kvm);             \
>>>>>> +               __kvm && kvm_vm_is_protected_pkvm((kern_hyp_va(__kvm)));\
>>>>>> +       })
>>>>>> +#endif
>>>>>> +
>>>>>>
>>>>>>
>>>>>> Or explicitly convert all nvhe accessors to a new "nvhe_vcpu_is_protected"
>>>>>>
>>>>>> The second one sounds like a better option to me.
>>>>>>
>>>>>
>>>>> We just need to include "asm/kvm_mmu.h" to "arch/arm64/mmio.c". With it,
>>>>> we can have unified functions (macros) in kvm_host.h to accomodate all
>>>>> cases, like below.
>>>>>
>>>>> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/ asm/ kvm_host.h
>>>>> index 9b1cf9c59e81..b9c9a9f203f9 100644
>>>>> --- a/arch/arm64/include/asm/kvm_host.h
>>>>> +++ b/arch/arm64/include/asm/kvm_host.h
>>>>> @@ -1514,11 +1514,36 @@ struct kvm *kvm_arch_alloc_vm(void);
>>>>>   #define __KVM_HAVE_ARCH_FLUSH_REMOTE_TLBS_RANGE
>>>>>
>>>>>   #define kvm_vm_is_protected(kvm)       ((kvm)->arch.vm_flavor >= __VM_PROTECTED)
>>>>> -#define vcpu_is_protected(vcpu) kvm_vm_is_protected((vcpu)->kvm)
>>>>> +#define vcpu_is_protected(vcpu)                                                \
>>>>> + ({                                                              \
>>>>> +               struct kvm *__kvm = READ_ONCE((vcpu)- >kvm);             \
>>>>> +               bool __protected = false;                               \
>>>>> +                                                                       \
>>>>> +               if (__kvm) {                                            \
>>>>> +                       if (is_nvhe_hyp_code())                         \
>>>
>>> This needs to be :
>>>                if (is_nvhe_hyp_code() && !is_protected_kvm_enabled())
>>>
>>
>> Yes.
>>
>>>>> +                               __kvm = kern_hyp_va(__kvm);             \
>>>
>>> Otherwise, we unnecessarily convert the vcpu->kvm for a hyp_vcpu and
>>> that is not going to end well for pKVM.
>>>
>>
>> Yes. kern_hyp_va() isn't needed when is_protected_kvm_enabled() is true.
> 
> nit : It is not that it is needed, but it is incorrect to convert the vcpu->kvm which is already a HYP VA via kern_hyp_va() which is not
> idempotent and thus we end up accessing something junk.
> 

Yes.

>> Besides, vcpu_is_protected_pkvm() can be further simplifed by fully exploiting
>> the check (is_protected_kvm_enabled()) done in vcpu_is_protected_pkvm(),
>> see below.
> 
> Good point.
> 
>>
>>> Cheers
>>> Suzuki
>>>
>>>
>>>>> +                       __protected = kvm_vm_is_protected(__kvm);       \
>>>>> +               }                                                       \
>>>>> +                                                                       \
>>>>> + __protected;                                            \
>>>>> +        })
>>>>> +
>>>>>
>>>>>   #define kvm_vm_is_protected_pkvm(kvm)          \
>>>>>          (is_protected_kvm_enabled() && ((kvm)->arch.vm_flavor == VM_PROTECTED_PKVM))
>>>>> -#define vcpu_is_protected_pkvm(vcpu) kvm_vm_is_protected_pkvm(vcpu- >kvm)
>>>>> +#define vcpu_is_protected_pkvm(vcpu)                                   \
>>>>> + ({                                                              \
>>>>> +               struct kvm *__kvm = READ_ONCE((vcpu)- >kvm);             \
>>>>> +               bool __protected = false;                               \
>>>>> +                                                                       \
>>>>> +               if (__kvm) {                                            \
>>>>> +                       if (is_nvhe_hyp_code())                         \
>>>>> +                               __kvm = kern_hyp_va(__kvm);             \
>>>>> +                       __protected = kvm_vm_is_protected_pkvm(__kvm);  \
>>>>> +               }                                                       \
>>>>> +                                                                       \
>>>>> + __protected;                                            \
>>>>> +        })
>>
>> This can be further simplified by fully expliting the check (is_protected_kvm_enabled())
>> in kvm_vm_is_protected_pkvm(), as below. The point is that the check is_protected_kvm_enabled()
>> gurantees that kern_hyp_va() isn't needed.
>>
>> #define vcpu_is_protected_pkvm(vcpu)                                    \
>>          ({                                                              \
>>                  struct kvm *__kvm = READ_ONCE((vcpu)->kvm);             \
>>                  bool __protected = false;                               \
>>                                                                          \
>>                  if (__kvm)                                              \
>>                          __protected = kvm_vm_is_protected_pkvm(__kvm);  \
>>                                                                          \
>>                  __protected;                                            \
>>          })
>>
> 
> Ack, I have reduced this to:
> 
> 
> /*
>   * Rely on is_protected_kvm_enabled() check in kvm_vm_is_protected_pkvm() to
>   * make sure the vcpu->kvm is always valid VA in the context
>   */
> #define vcpu_is_protected_pkvm(vcpu)                                    \
>          ({                                                              \
>                  struct kvm *__kvm = READ_ONCE((vcpu)->kvm);             \
>                                                                          \
>                  (__kvm) && kvm_vm_is_protected_pkvm(__kvm));            \
>          })
> 

Parentheses for the local variable '__kvm' can be dropped. With this, this
looks good to me.

                    (__kvm && kvm_vm_is_protected_pkvm(__kvm));

Thanks,
Gavin

> 
> Cheers
> Suzuki
> 
>>>>>
>>>>>   #define kvm_vm_is_unprotected_pkvm(kvm)                \
>>>>>          (is_protected_kvm_enabled() && ((kvm)->arch.vm_flavor == VM_PKVM))
>>>>> diff --git a/arch/arm64/kvm/mmio.c b/arch/arm64/kvm/mmio.c
>>>>> index d1c3a352d5a2..ab1d2fef9a52 100644
>>>>> --- a/arch/arm64/kvm/mmio.c
>>>>> +++ b/arch/arm64/kvm/mmio.c
>>>>> @@ -6,6 +6,7 @@
>>>>>
>>>>>   #include <linux/kvm_host.h>
>>>>>   #include <asm/kvm_emulate.h>
>>>>> +#include <asm/kvm_mmu.h>
>>>>>   #include <trace/events/kvm.h>
>>>>>
>>>>>   #include "trace.h"
>>>>>
>>>>
>>>> Thanks Gavin, that works and looks neat. I will incorporate it.
>>>>
>>
>> Thanks,
>> Gavin
>>
>