[PATCH] KVM: arm64: Take a protected vCPU's TWI and TWE only at vCPU load

[PATCH] KVM: arm64: Take a protected vCPU's TWI and TWE only at vCPU load

[Sasha Levin]

Commit 06499de8f998 ("KVM: arm64: Skip fixed-feature state flush for
protected vCPUs") moved the copy of the host's TWI and TWE in
flush_hyp_vcpu() into the non-protected branch: a protected vCPU takes
them in __pkvm_vcpu_load() instead, and only VSE comes from the host on
each entry.

Commit 04447b95c62b ("KVM: arm64: Use the host's HCR_EL2 for
non-protected VMs in pKVM") was written against the old code, where
both VM types took TWI, TWE and VSE on each entry, and kept that for a
protected VM as PKVM_HCR_EL2_HOST_PVM. The merge of the two applies
PKVM_HCR_EL2_HOST_PVM in the common path, so a protected vCPU takes TWI
and TWE from the host's vCPU on every entry again.

That doesn't change what a protected guest runs with today: once the
vCPU is initialised, the host only changes TWI and TWE in
vcpu_set_wfx_traps(), which pkvm_vcpu_load() calls right before
__pkvm_vcpu_load(). But it undoes part of 06499de8f998, and has EL2
read a protected vCPU's trap configuration from host memory on every
entry, which nothing needs.

Make PKVM_HCR_EL2_HOST_PVM VSE alone and add TWI and TWE to
PKVM_HCR_EL2_HOST_NPVM, so that each mask again lists what EL2 takes
from the host on each entry for its VM type. A non-protected VM takes
the same bits as before. Say in the comments that a protected vCPU's
TWI and TWE come at vCPU load.

llminus, an LLM-assisted merge conflict resolution tool, found the
difference while resolving the same conflict in an integration tree.

Fixes: 439d8f9cb01c ("Merge tag 'kvmarm-7.4' of git://git.kernel.org/pub/scm/linux/kernel/git/kvmarm/kvmarm into HEAD")
Assisted-by: LLM
Signed-off-by: Sasha Levin <sashal@kernel.org>
---
With KVM's own host code this changes nothing a guest can observe: the
per-entry copy writes back the TWI/TWE value __pkvm_vcpu_load() already
set. It restores the split 06499de8f998 made.

Marc's resolution in kvmarm/fixes, aca73c1d23f1 ("Merge branch
kvmarm-7.4 into kvmarm-master/fixes"), is the same as the one in
kvm/next, so this applies to both.

Build-tested only: arm64 defconfig with gcc and clang, W=1, no new
warnings. The only code generation change is the protected vCPU's
per-entry mask, 0x6100 to 0x100. Not runtime-tested, as the in-tree
selftests can't create protected VMs.

 arch/arm64/kvm/hyp/include/nvhe/pkvm.h | 13 +++++++------
 arch/arm64/kvm/hyp/nvhe/pkvm.c         |  5 ++++-
 2 files changed, 11 insertions(+), 7 deletions(-)

diff --git a/arch/arm64/kvm/hyp/include/nvhe/pkvm.h b/arch/arm64/kvm/hyp/include/nvhe/pkvm.h
index 4373e9b63e87f..0261b2bc681f7 100644
--- a/arch/arm64/kvm/hyp/include/nvhe/pkvm.h
+++ b/arch/arm64/kvm/hyp/include/nvhe/pkvm.h
@@ -14,13 +14,14 @@
 #include <nvhe/spinlock.h>
 
 /*
- * HCR_EL2 bits EL2 takes from the host on each entry, per VM type. The rest
- * are EL2's own and nothing the host sets there reaches the guest.
+ * HCR_EL2 bits EL2 takes from the host on each entry, per VM type. A protected
+ * VM also takes TWI and TWE, at vCPU load. The rest are EL2's own and nothing
+ * the host sets there reaches the guest.
  */
-#define PKVM_HCR_EL2_HOST_PVM	(HCR_EL2_TWI | HCR_EL2_TWE | HCR_EL2_VSE)
-#define PKVM_HCR_EL2_HOST_NPVM	(PKVM_HCR_EL2_HOST_PVM | HCR_EL2_VI | HCR_EL2_VF |	\
-				 HCR_EL2_TVM | HCR_EL2_TID2 | HCR_EL2_TID4 |		\
-				 HCR_EL2_TID5 | HCR_EL2_TTLBOS)
+#define PKVM_HCR_EL2_HOST_PVM	HCR_EL2_VSE
+#define PKVM_HCR_EL2_HOST_NPVM	(PKVM_HCR_EL2_HOST_PVM | HCR_EL2_TWI | HCR_EL2_TWE |	\
+				 HCR_EL2_VI | HCR_EL2_VF | HCR_EL2_TVM | HCR_EL2_TID2 |	\
+				 HCR_EL2_TID4 | HCR_EL2_TID5 | HCR_EL2_TTLBOS)
 
 /*
  * Holds the relevant data for maintaining the vcpu state completely at hyp.
diff --git a/arch/arm64/kvm/hyp/nvhe/pkvm.c b/arch/arm64/kvm/hyp/nvhe/pkvm.c
index 237483d2442a5..fc0475f0877d1 100644
--- a/arch/arm64/kvm/hyp/nvhe/pkvm.c
+++ b/arch/arm64/kvm/hyp/nvhe/pkvm.c
@@ -38,7 +38,10 @@ unsigned int kvm_host_sve_max_vl;
  */
 static DEFINE_PER_CPU(struct pkvm_hyp_vcpu *, loaded_hyp_vcpu);
 
-/* The PKVM_HCR_EL2_HOST_{PVM,NPVM} bits of this value come from the host on each entry. */
+/*
+ * The PKVM_HCR_EL2_HOST_{PVM,NPVM} bits of this value come from the host on
+ * each entry, and a protected vCPU's TWI and TWE at vCPU load.
+ */
 static void pkvm_vcpu_reset_hcr(struct kvm_vcpu *vcpu)
 {
 	vcpu->arch.hcr_el2 = HCR_GUEST_FLAGS;

base-commit: 2ea3a929fe01ec5ae7313a94badaf9b70a8c69b1
-- 
2.53.0

Re: [PATCH] KVM: arm64: Take a protected vCPU's TWI and TWE only at vCPU load

[Fuad Tabba]

Thanks for fixing this Sasha.

On Sat, 17 Oct 2026 17:52:04 +0100, Sasha Levin <sashal@kernel.org> wrote:
[...]
> Build-tested only: arm64 defconfig with gcc and clang, W=1, no new
> warnings. The only code generation change is the protected vCPU's
> per-entry mask, 0x6100 to 0x100. Not runtime-tested, as the in-tree
> selftests can't create protected VMs.

I tested it on QEMU, with the host booted with
kvm-arm.wfi_trap_policy=notrap and a protected VM using kvmtool. The
protected VM doesn't take any WFI exits (4415 with the default
policy), so TWI still reaches it at vCPU load.

Reviewed-by: Fuad Tabba <fuad.tabba@linux.dev>
Tested-by: Fuad Tabba <fuad.tabba@linux.dev>

Cheers,
/fuad