Skip to contentSkip to conversation

i386/kvm: Expose ARCH_CAP_FB_CLEAR when invulnerable to MDS

PATCH Applied Jon Kohler 3 participants 3 messages Oct 8, 2025 – Oct 11, 2025 KVM

lore
Share & tools
lore

Jon Kohler

author
#0 lore raw

PATCH 1/1 i386/kvm: Expose ARCH_CAP_FB_CLEAR when invulnerable to MDS

Newer Intel hardware (Sapphire Rapids and higher) sets multiple MDS immunity bits in MSR_IA32_ARCH_CAPABILITIES but lacks the hardware-level MSR_ARCH_CAP_FB_CLEAR (bit 17):
ARCH_CAP_MDS_NO
ARCH_CAP_TAA_NO
ARCH_CAP_PSDP_NO
ARCH_CAP_FBSDP_NO
ARCH_CAP_SBDR_SSDP_NO

This prevents VMs with fb-clear=on from migrating from older hardware (Cascade Lake, Ice Lake) to newer hardware, limiting live migration capabilities. Note fb-clear was first introduced in v8.1.0 [1].

Expose MSR_ARCH_CAP_FB_CLEAR for MDS-invulnerable systems to enable seamless migration between hardware generations.

Note: There is no impact when a guest migrates to newer hardware as the existing bit combinations already mark the host as MMIO-immune and disable FB_CLEAR operations in the kernel (see Linux's arch_cap_mmio_immune() and vmx_update_fb_clear_dis()). See kernel side discussion for [2] for additional context.

[1] 22e1094ca82 ("target/i386: add support for FB_CLEAR feature")
[2] https://patchwork.kernel.org/project/kvm/patch/[email protected]/

---
 target/i386/kvm/kvm.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
target/i386/kvm/kvm.c +17 -0
index 6a3a1c1ed8..610d786310 100644
@@ -657,6 +657,23 @@ uint64_t kvm_arch_get_supported_msr_feature(KVMState *s, uint32_t index)
         must_be_one = (uint32_t)value;
         can_be_one = (uint32_t)(value >> 32);
         return can_be_one & ~must_be_one;
+    case MSR_IA32_ARCH_CAPABILITIES:
+        /*
+         * Special handling for fb-clear bit in ARCH_CAPABILITIES MSR.
+         * KVM will only report the bit if it is enabled in the host,
+         * but, for live migration capability purposes, we want to
+         * expose the bit to the guest even if it is disabled in the
+         * host, as long as the host itself is not vulnerable to
+         * the issue that the fb-clear bit is meant to mitigate.
+         */
+        if ((value & MSR_ARCH_CAP_MDS_NO) &&
+            (value & MSR_ARCH_CAP_TAA_NO) &&
+            (value & MSR_ARCH_CAP_SBDR_SSDP_NO) &&
+            (value & MSR_ARCH_CAP_FBSDP_NO) &&
+            (value & MSR_ARCH_CAP_PSDP_NO)) {
+                value |= MSR_ARCH_CAP_FB_CLEAR;
+        }
+        return value;
 
     default:
         return value;
Signature · 1 line
2.43.0
#1 lore raw
Quoted 9 lines · Jon Kohler

Newer Intel hardware (Sapphire Rapids and higher) sets multiple MDS immunity bits in MSR_IA32_ARCH_CAPABILITIES but lacks the hardware-level MSR_ARCH_CAP_FB_CLEAR (bit 17):
ARCH_CAP_MDS_NO
ARCH_CAP_TAA_NO
ARCH_CAP_PSDP_NO
ARCH_CAP_FBSDP_NO
ARCH_CAP_SBDR_SSDP_NO

Quoted 6 lines

This prevents VMs with fb-clear=on from migrating from older hardware (Cascade Lake, Ice Lake) to newer hardware, limiting live migration capabilities. Note fb-clear was first introduced in v8.1.0 [1].

Expose MSR_ARCH_CAP_FB_CLEAR for MDS-invulnerable systems to enable seamless migration between hardware generations.

LGTM, thanks!