Fixes: 644e3c5d81 ("missing vmx features for Skylake-Server and Cascadelake-Server")
Signed-off-by: Han Han <hhan@redhat.com>
Reviewed-by: Chenyi Qiang <chenyi.qiang@intel.com>
Reviewed-by: Michael Tokarev <mjt@tls.msk.ru>
Signed-off-by: Michael Tokarev <mjt@tls.msk.ru>
(cherry picked from commit 93dcc9390e5ad0696ae7e9b7b3a5b08c2d1b6de6)
Signed-off-by: Michael Tokarev <mjt@tls.msk.ru>
Since commit 5286c36622 ("target/i386: properly reset TSC on reset")
QEMU writes the special value of "1" to each online vCPU TSC on VM reset
to reset it.
However parked vCPUs don't get that handling and due to that their TSCs
get desynchronized when the VM gets reset.
This in turn causes KVM to turn off PVCLOCK_TSC_STABLE_BIT in its exported
PV clock.
Note that KVM has no understanding of vCPU being currently parked.
Without PVCLOCK_TSC_STABLE_BIT the sched clock is marked unstable in
the guest's kvm_sched_clock_init().
This causes a performance regressions to show in some tests.
Fix this issue by writing the special value of "1" also to TSCs of parked
vCPUs on VM reset.
Reproducing the issue:
1) Boot a VM with "-smp 2,maxcpus=3" or similar
2) device_add host-x86_64-cpu,id=vcpu,node-id=0,socket-id=0,core-id=2,thread-id=0
3) Wait a few seconds
4) device_del vcpu
5) Inside the VM run:
# echo "t" >/proc/sysrq-trigger; dmesg | grep sched_clock_stable
Observe the sched_clock_stable() value is 1.
6) Reboot the VM
7) Once the VM boots once again run inside it:
# echo "t" >/proc/sysrq-trigger; dmesg | grep sched_clock_stable
Observe the sched_clock_stable() value is now 0.
Fixes: 5286c36622 ("target/i386: properly reset TSC on reset")
Signed-off-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Link: https://lore.kernel.org/r/5a605a88e9a231386dc803c60f5fed9b48108139.1734014926.git.maciej.szmigiero@oracle.com
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
(cherry picked from commit 3f2a05b31ee9ce2ddb6c75a9bc3f5e7f7af9a76f)
Signed-off-by: Michael Tokarev <mjt@tls.msk.ru>
Building without CONFIG_HYPERV is currently broken due to a missing
symbol 'hyperv_syndbg_query_options'. Add it to the stubs
that exist for that very reasons.
Reported-by: Michael Tokarev <mjt@tls.msk.ru>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The handling for CPUID function 0xD (supported XSAVE features) was
improved in a recent patch. Unfortunately, this appears to have
introduced undefined behaviour for cases where ecx > 30, as the result
of (1 << idx) is undefined if idx > 30.
Per Intel SDM section 13.2, the behaviour for ecx values up to and
including 62 are specified. This change therefore specifically sets
all registers returned by the CPUID instruction to 0 for 63 and higher.
Furthermore, the bit shift uses uint64_t, where behaviour for the entire
range of 2..62 is safe and correct.
Signed-off-by: Phil Dennis-Jordan <phil@philjordan.eu>
Link: https://lore.kernel.org/r/20241105155800.5461-3-phil@philjordan.eu
Reviewed-by: Roman Bolshakov <r.bolshakov@yadro.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Support for x2APIC mode was recently introduced in the software emulated
APIC implementation for TCG. Enabling it when using macOS’s hvf
accelerator is useful and significantly helps performance, as Qemu
currently uses the emulated APIC when running on hvf as well.
This change wires up the read & write operations for the MSR VM exits
and allow-lists the CPUID flag in the x86 hvf runtime.
Signed-off-by: Phil Dennis-Jordan <phil@philjordan.eu>
Link: https://lore.kernel.org/r/20241105155800.5461-2-phil@philjordan.eu
Reviewed-by: Roman Bolshakov <r.bolshakov@yadro.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Commit b56617bbcb ("target/i386: Walk NPT in guest real mode") added
logic to run the page table walker even in real mode if we are in NPT
mode. That function then determined whether real mode or paging is
active based on whether the pg_mode variable was 0.
Unfortunately pg_mode is 0 in two situations:
1) Paging is disabled (real mode)
2) Paging is in 2-level paging mode (32bit without PAE)
That means the walker now assumed that 2-level paging mode was real
mode, breaking NetBSD as well as Windows XP.
To fix that, this patch adds a new PG flag to pg_mode which indicates
whether paging is active at all and uses that to determine whether we
are in real mode or not.
Cc: qemu-stable@nongnu.org
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2654
Fixes: b56617bbcb ("target/i386: Walk NPT in guest real mode")
Signed-off-by: Alexander Graf <graf@amazon.com>
Reported-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Link: https://lore.kernel.org/r/20241106154329.67218-1-graf@amazon.com
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Cache topology needs to be defined based on CPU topology levels. Thus,
define CPU topology enumeration in qapi/machine.json to make it generic
for all architectures.
To match the general topology naming style, rename CPU_TOPO_LEVEL_* to
CPU_TOPOLOGY_LEVEL_*, and rename SMT and package levels to thread and
socket.
Also, enumerate additional topology levels for non-i386 arches, and add
a CPU_TOPOLOGY_LEVEL_DEFAULT to help future smp-cache object to work
with compatibility requirement of arch-specific cache topology models.
Signed-off-by: Zhao Liu <zhao1.liu@intel.com>
Tested-by: Yongwei Ma <yongwei.ma@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Acked-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-ID: <20241101083331.340178-3-zhao1.liu@intel.com>
Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
In the follow-up change, the CPU topology enumeration will be moved to
QAPI. And considerring "invalid" should not be exposed to QAPI as an
unsettable item, so, as a preparation for future changes, remove
"invalid" level from the current CPU topology enumeration structure
and define it by a macro instead.
Due to the removal of the enumeration of "invalid", bit 0 of
CPUX86State.avail_cpu_topo bitmap will no longer correspond to "invalid"
level, but will start at the SMT level. Therefore, to honor this change,
update the encoding rule for CPUID[0x1F].
Signed-off-by: Zhao Liu <zhao1.liu@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Message-ID: <20241101083331.340178-2-zhao1.liu@intel.com>
Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Set the NaN propagation rule explicitly for the float_status words
used in the x86 target.
This is a no-behaviour-change commit, so we retain the existing
behaviour of using the x87-style "prefer QNaN over SNaN, then prefer
the NaN with the larger significand" for MMX and SSE. This is
however not the documented hardware behaviour, so we leave a TODO
note about what we should be doing instead.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20241025141254.2141506-16-peter.maydell@linaro.org
Since the highest supported vector length for a processor implies that
all lesser vector lengths are also supported, add the dependencies of
the supported vector lengths. If all vector lengths aren't supported,
clear AVX10 enable bit as well.
Note that the order of AVX10 related dependencies should be kept as:
CPUID_24_0_EBX_AVX10_128 -> CPUID_24_0_EBX_AVX10_256,
CPUID_24_0_EBX_AVX10_256 -> CPUID_24_0_EBX_AVX10_512,
CPUID_24_0_EBX_AVX10_VL_MASK -> CPUID_7_1_EDX_AVX10,
CPUID_7_1_EDX_AVX10 -> CPUID_24_0_EBX,
so that prevent user from setting weird CPUID combinations, e.g. 256-bits
and 512-bits are supported but 128-bits is not, no vector lengths are
supported but AVX10 enable bit is still set.
Since AVX10_128 will be reserved as 1, adding these dependencies has the
bonus that when user sets -cpu host,-avx10-128, CPUID_7_1_EDX_AVX10 and
CPUID_24_0_EBX will be disabled automatically.
Tested-by: Xuelian Guo <xuelian.guo@intel.com>
Signed-off-by: Tao Su <tao1.su@linux.intel.com>
Link: https://lore.kernel.org/r/20241028024512.156724-5-tao1.su@linux.intel.com
Reviewed-by: Zhao Liu <zhao1.liu@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Link: https://lore.kernel.org/r/20241031085233.425388-7-tao1.su@linux.intel.com
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Right now, QEMU is using the "feature" and "bits" fields of ExtSaveArea
to query the accelerator for the support status of extended save areas.
This is a problem for AVX10, which attaches two feature bits (AVX512F
and AVX10) to the same extended save states.
To keep the AVX10 hacks to the minimum, limit usage of esa->features
and esa->bits. Instead, just query the accelerator for the 0xD leaf.
Do it in common code and clear esa->size if an extended save state is
unsupported.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Zhao Liu <zhao1.liu@intel.com>
Link: https://lore.kernel.org/r/20241031085233.425388-3-tao1.su@linux.intel.com
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This gives greater opportunity for reassociation on x86 targets,
since addition can use the LEA instruction.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If the operands of the arithmetic instruction fit within a half-register,
it's easiest to use a comparison instruction to compute the carry.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This makes it easier for the compiler to understand which bits are set,
and it also removes "cltq" instructions to canonicalize the output value
as 32-bit signed.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Mostly used for TEST+JG and TEST+JLE, but it is easy to cover
also JBE/JA and JL/JGE; shaves about 0.5% TCG ops.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Most uses of CC_OP_DYNAMIC are for CMP/JB/JE or similar sequences.
We can optimize many of them to avoid computation of the flags.
This eliminates both TCG ops to set up the new cc_op, and helper
instructions because evaluating just ZF is much cheaper.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Just use CC_OP_EFLAGS; it is not that likely that the flags computed by
CC_OP_CLR survive the end of the basic block, in which case there is no
need to spill cc_op_src.
cc_op_src now does need spilling if the XOR is followed by a memory
operation, but this only costs 0.2% extra TCG ops. They will be recouped
by simplifications in how QEMU evaluates ZF at runtime, which are even
greater with this change.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The call to xgetbv() is passing the ecx value for cpuid function 0xD,
index 0. The xgetbv call thus returns false (OSXSAVE is bit 27, which is
well out of the range of CPUID[0xD,0].ECX) and eax is not modified. While
fixing it, cache the whole computation of supported XCR0 bits since it
will be used for more than just CPUID leaf 0xD.
Furthermore, unsupported subleafs of CPUID 0xD (including all those
corresponding to zero bits in host's XCR0) must be hidden; if OSXSAVE
is not set at all, the whole of CPUID leaf 0xD plus the XSAVE bit must
be hidden.
Finally, unconditionally drop XSTATE_BNDREGS_MASK and XSTATE_BNDCSR_MASK;
real hardware will only show them if the MPX bit is set in CPUID;
this is never the case for hvf_get_supported_cpuid() because QEMU's
Hypervisor.framework support does not handle the VMX fields related to
MPX (even in the unlikely possibility that the host has MPX enabled).
So hide those bits in the new cache_host_xcr0().
Cc: Phil Dennis-Jordan <lists@philjordan.eu>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Newer AMD CPUs support ERAPS (Enhanced Return Address Prediction Security)
feature that enables the auto-clear of RSB entries on a TLB flush, context
switches and VMEXITs. The number of default RSP entries is reflected in
RapSize.
Add the feature bit and feature word to support these features.
CPUID_Fn80000021_EAX
Bits Feature Description
24 ERAPS:
Indicates support for enhanced return address predictor security.
CPUID_Fn80000021_EBX
Bits Feature Description
31-24 Reserved
23:16 RapSize:
Return Address Predictor size. RapSize x 8 is the minimum number
of CALL instructions software needs to execute to flush the RAP.
15-00 MicrocodePatchSize. Read-only.
Reports the size of the Microcode patch in 16-byte multiples.
If 0, the size of the patch is at most 5568 (15C0h) bytes.
Link: https://www.amd.com/content/dam/amd/en/documents/epyc-technical-docs/programmer-references/57238.zip
Signed-off-by: Babu Moger <babu.moger@amd.com>
Link: https://lore.kernel.org/r/7c62371fe60af1e9bbd853f5f8e949bf2d908bd0.1729807947.git.babu.moger@amd.com
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
CPUID leaf 0x80000022, i.e. ExtPerfMonAndDbg, advertises new performance
monitoring features for AMD processors. Bit 0 of EAX indicates support
for Performance Monitoring Version 2 (PerfMonV2) features. If found to
be set during PMU initialization, the EBX bits can be used to determine
the number of available counters for different PMUs. It also denotes the
availability of global control and status registers.
Add the required CPUID feature word and feature bit to allow guests to
make use of the PerfMonV2 features.
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
Signed-off-by: Babu Moger <babu.moger@amd.com>
Reviewed-by: Zhao Liu <zhao1.liu@intel.com>
Link: https://lore.kernel.org/r/a96f00ee2637674c63c61e9fc4dee343ea818053.1729807947.git.babu.moger@amd.com
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The check of cpu->phys_bits to be in range between
[32, TARGET_PHYS_ADDR_SPACE_BITS] in host_cpu_realizefn()
is duplicated with check in x86_cpu_realizefn().
Since the ckeck in x86_cpu_realizefn() is called later and can cover all
the x86 cases. Remove the one in host_cpu_realizefn().
Opportunistically adjust cpu->phys_bits directly in
host_cpu_adjust_phys_bits(), which matches more with the function name.
Signed-off-by: Xiaoyao Li <xiaoyao.li@intel.com>
Reviewed-by: Igor Mammedov <imammedo@redhat.com>
Reviewed-by: Zhao Liu <zhao1.liu@intel.com>
Link: https://lore.kernel.org/r/20240929085747.2023198-1-xiaoyao.li@intel.com
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
LFENCE and SFENCE were introduced with the original SSE instruction set;
marking them incorrectly as cpuid(SSE2) causes failures for CPU models
that lack SSE2, for example pentium3.
Reported-by: Guenter Roeck <linux@roeck-us.net>
Tested-by: Guenter Roeck <linux@roeck-us.net>
Cc: qemu-stable@nongnu.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Zhao Liu
Han Han
Maciej S. Szmigiero
Alexander Graf
Pierrick Bouvier
Peter Maydell
Paolo Bonzini
Phil Dennis-Jordan
Alexander Graf
Tao Su
Richard Henderson
Babu Moger
Sandipan Das
Xiaoyao Li