On 11/23/22 19:17, Huanyu ZHAI wrote:
Subject: [PATCH] cpu/i386: update xsave components after CPUID filtering
On i386 platform, CPUID data are setup through three consecutive steps:
CPU model definition, expansion and filtering.
XSAVE components are enabled during the expansion stage, by checking if
they are enabled in CPUID. However, it is still
probable that some XSAVE features will be enabled/disabled during the
filtering stage and the XSAVE components left unchanged.
Inconsistency between XSAVE features and enabled XSAVE components can
lead to problems on some Linux guests in the absence of
The patch is correct, but I think you can also remove the existing call
to x86_cpu_enable_xsave_components(). Can you test that?
Thanks,
Paolo
the following patch in the kernel:
https://www.mail-archive.com/[email protected]/msg1452368.html
<https://www.mail-archive.com/[email protected]/msg1452368.html>
A simple case to reproduce this problem is to start a SUSE 12 SP3 guest
with cpu model set to Skylake-Server:
$ qemu-system-x86_64 -cpu Skylake-Server ...
In the SUSE 12 SP3 guest, one can observe that PKRU will be enabled
without Intel PKU's presence.
That's because on platform with Skylake-Server cpus, Intel PKU is
disabled during x86_cpu_filter_features(),
but the XSAVE PKRU bit was enabled by x86_cpu_expand_features().
Signed-off-by: Huanyu ZHAI [email protected]
<mailto:[email protected]>
Signed-off-by: Xin Wang [email protected]
<mailto:[email protected]>
---
target/i386/cpu.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/target/i386/cpu.c b/target/i386/cpu.c
index 22b681ca37..2ee574cf05 100644
--- a/target/i386/cpu.c
+++ b/target/i386/cpu.c
@@ -6362,6 +6362,9 @@ static void x86_cpu_filter_features(X86CPU *cpu,
bool verbose)
mark_unavailable_features(cpu, FEAT_7_0_EBX,
CPUID_7_0_EBX_INTEL_PT, prefix);
}
}
+
+ /* Update XSAVE components again based on the filtered CPU feature
flags */
+ x86_cpu_enable_xsave_components(cpu);
}
static void x86_cpu_hyperv_realize(X86CPU *cpu)
--
2.27.0
