CPUID[0xD,0].EBX is the size of the XSAVE area for the guest's current
XCR0, and CPUID[0xD,1].EBX the size for XCR0 | IA32_XSS: those are the
two values that depend on guest state. For sub-leaf 2 and up, EBX is the
component's offset in the standard format, a constant.
whpx_vcpu_run() forwards Hyper-V's DefaultResultRbx for sub-leaves 1 and
2 instead. Sub-leaf 2 is harmless: QEMU's own offset was read from
Hyper-V at vCPU init. Sub-leaf 0 is not: QEMU computes it from
env->xcr0, which WHPX refreshes only in whpx_get_registers(), never on
a CPUID exit, so after the guest's XSETBV the leaf still reports the
size for the reset XCR0, 576 bytes.
A Linux guest on a host without XSAVES takes that as its kernel XSAVE
size, fails the consistency check against the size it computed from the
component offsets, disables XSAVE and panics on the fallback path:
XSAVE consistency problem: size 832 != kernel_size 576
WARNING: CPU: 0 PID: 0 at arch/x86/kernel/fpu/xstate.c:638
fpu__init_system_xstate
x86/fpu: XSAVE disabled
BUG: kernel NULL pointer dereference, address: 0000000000000010
RIP: 0010:fpstate_reset+0x1a/0x90
Kernel panic - not syncing: Attempted to kill the idle task!
Every CPU model with XSAVE is affected there (-cpu max, Haswell,
SandyBridge); models without it (Westmere, Nehalem, qemu64) boot. A host
with XSAVES boots because Linux takes the compacted size from sub-leaf 1,
which is forwarded; sub-leaf 0 is stale there too.
Reproduced on an Intel Core i5-4440 (Haswell) under Windows 11 Home
build 26300 with QEMU 11.1.2 (the code is unchanged on master), guest
Linux 6.18. With this patch the guest enables xstate 0x7 at 832 bytes
and boots with -cpu max. An AMD Ryzen 9 7900X under build 26100, not
affected before, still boots with the patch, enabling 0xe7 (compacted,
2432 bytes).
Fixes: 79b3cb1f0145 ("whpx: i386: introduce proper cpuid support")
Signed-off-by: Kostas Chatzikokolakis <[email protected]>
---
target/i386/whpx/whpx-all.c | 10 +++++++---
1 file changed, 7 insertions(+), 3 deletions(-)
diff --git a/target/i386/whpx/whpx-all.c b/target/i386/whpx/whpx-all.c
index 113f79af..b51e9dfc 100644
--- a/target/i386/whpx/whpx-all.c
+++ b/target/i386/whpx/whpx-all.c
@@ -2618,10 +2618,14 @@ int whpx_vcpu_run(CPUState *cpu)
}
}
- /* CPUID[0xD,{1,2}].EBX are dynamic depending on guest features. */
+ /*
+ * CPUID[0xD,{0,1}].EBX are dynamic: the XSAVE area size for the
+ * guest's current XCR0, resp. XCR0 | IA32_XSS. For sub-leaf >= 2,
+ * EBX is the component's fixed offset, which QEMU already reports.
+ */
if (vcpu->exit_ctx.CpuidAccess.Rax == 0xd) {
- if (vcpu->exit_ctx.CpuidAccess.Rcx == 1
- || vcpu->exit_ctx.CpuidAccess.Rcx == 2) {
+ if (vcpu->exit_ctx.CpuidAccess.Rcx == 0
+ || vcpu->exit_ctx.CpuidAccess.Rcx == 1) {
reg_values[4].Reg64 =
vcpu->exit_ctx.CpuidAccess.DefaultResultRbx;
}
}
--
2.54.0
