On Thu, Jan 14, 2021 at 10:47:03PM +0300, yaroshchuk2000@gmail.com wrote:
> From: Vladislav Yaroshchuk <yaroshchuk2000@gmail.com>
>
> For `-accel hvf` cpu_x86_cpuid() is wrapped with hvf_cpu_x86_cpuid() to
> add paravirtualization cpuid leaf 0x40000010
> https://lkml.org/lkml/2008/10/1/246
>
> Leaf 0x40000010, Timing Information:
> EAX: (Virtual) TSC frequency in kHz.
> EBX: (Virtual) Bus (local apic timer) frequency in kHz.
> ECX, EDX: RESERVED (Per above, reserved fields are set to zero).
>
> On macOS TSC and APIC Bus frequencies can be readed by sysctl call with
> names `machdep.tsc.frequency` and `hw.busfrequency`
>
> This options is required for Darwin-XNU guest to be synchronized with
> host
>
> Signed-off-by: Vladislav Yaroshchuk <yaroshchuk2000@gmail.com>
> ---
> target/i386/hvf/hvf.c | 90 ++++++++++++++++++++++++++++++++++++++++++-
> 1 file changed, 89 insertions(+), 1 deletion(-)
>
> diff --git a/target/i386/hvf/hvf.c b/target/i386/hvf/hvf.c
> index ed9356565c..a5daafe202 100644
> --- a/target/i386/hvf/hvf.c
> +++ b/target/i386/hvf/hvf.c
> @@ -65,6 +65,7 @@
>
> #include <Hypervisor/hv.h>
> #include <Hypervisor/hv_vmx.h>
> +#include <sys/sysctl.h>
>
> #include "exec/address-spaces.h"
> #include "hw/i386/apic_internal.h"
> @@ -456,6 +457,48 @@ static void dummy_signal(int sig)
> {
> }
>
> +static void init_tsc_freq(CPUX86State *env)
> +{
> + size_t length;
> + uint64_t tsc_freq;
> +
> + if (env->tsc_khz != 0) {
> + return;
> + }
> +
> + length = sizeof(uint64_t);
> + if (sysctlbyname("machdep.tsc.frequency", &tsc_freq, &length, NULL, 0)) {
> + return;
> + }
> + env->tsc_khz = tsc_freq / 1000; /* Hz to KHz */
> +}
> +
> +static void init_apic_bus_freq(CPUX86State *env)
> +{
> + size_t length;
> + uint64_t bus_freq;
> +
> + if (env->apic_bus_freq != 0) {
> + return;
> + }
> +
> + length = sizeof(uint64_t);
> + if (sysctlbyname("hw.busfrequency", &bus_freq, &length, NULL, 0)) {
> + return;
> + }
> + env->apic_bus_freq = bus_freq;
> +}
> +
> +static inline bool tsc_is_known(CPUX86State *env)
> +{
> + return env->tsc_khz != 0;
> +}
> +
> +static inline bool apic_bus_freq_is_known(CPUX86State *env)
> +{
> + return env->apic_bus_freq != 0;
> +}
> +
> int hvf_init_vcpu(CPUState *cpu)
> {
>
> @@ -480,6 +523,15 @@ int hvf_init_vcpu(CPUState *cpu)
> hvf_state->hvf_caps = g_new0(struct hvf_vcpu_caps, 1);
> env->hvf_mmio_buf = g_new(char, 4096);
>
> + if (x86cpu->vmware_cpuid_freq) {
> + init_tsc_freq(env);
> + init_apic_bus_freq(env);
> +
> + if (!tsc_is_known(env) || !apic_bus_freq_is_known(env)) {
> + error_report("vmware-cpuid-freq: feature couldn't be enabled");
> + }
> + }
> +
> r = hv_vcpu_create((hv_vcpuid_t *)&cpu->hvf_fd, HV_VCPU_DEFAULT);
> cpu->vcpu_dirty = 1;
> assert_hvf_ok(r);
> @@ -597,6 +649,42 @@ static void hvf_store_events(CPUState *cpu, uint32_t ins_len, uint64_t idtvec_in
> }
> }
>
> +static void hvf_cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,
> + uint32_t *eax, uint32_t *ebx,
> + uint32_t *ecx, uint32_t *edx)
> +{
> + /*
> + * A wrapper extends cpu_x86_cpuid with 0x40000000 and 0x40000010 leafs
> + * Provides vmware-cpuid-freq support to hvf
> + */
> +
> + uint32_t signature[3];
> +
> + if (!tsc_is_known(env) || !apic_bus_freq_is_known(env)) {
> + cpu_x86_cpuid(env, index, count, eax, ebx, ecx, edx);
> + return;
> + }
> +
> + switch (index) {
> + case 0x40000000:
> + memcpy(signature, "TCGTCGTCGTCG", 12); /* QEMU Signature */
Hi Vladislav,
TCG belongs to TCG accel identification, for HVF it should be
HVFHVFHVFHVF.
> + *eax = 0x40000010; /* Max available cpuid leaf */
> + *ebx = signature[0];
> + *ecx = signature[1];
> + *edx = signature[2];
TCG and KVM don't report their identity unless kvm or tcg-cpuid
properties are set. I wonder if we need to guard it likewise?
But as of now QEMU is not consistent in that regard. Two parameters are
needed for KVM - kvm=on,vmware-cpuid-freq=on. vmware-cpuid-freq is
sufficient for WHPX but WHPX doesn't expose itself (ebx=ecx=edx=0). TCG
doesn't seem to support vmware-cpuid-freq but reports it's name only if
tcg-cpuid property is set.
Ok, I understood the mistake. HVFHVFHVFHVF sounds more suitable, but
hypervisor signature is not a necessary thing for vmware-cpuid-freq
(especially unknown to anyone `HVFHVFHVFHVF`). Seems it can be dropped
by filling ebx=ecx=edx=0, as WHPX does. No reason to expose the HVF in
this case. If needed, it can be added as another feature.
> + break;
CPUID for not implemented hypervisor-specific leafs from 0x40000001 up
to 0x4000000f should be all zeroes but cpu_x86_cpuid() only returns zero
values for 0x40000001. Likely, you need to reset return values for the
leafs here or in cpu_x86_cpuid(). In the latter case you'll also fix a
similar bug in WHPX accel.
I'll fix it in-place, resetting values in hvf_cpu_x86_cpuid()
Thank you for the review!
Sincerely,
Vladislav Yaroshchuk
Otherwise, looks good.
Thanks,
Roman
> + case 0x40000010:
> + *eax = env->tsc_khz;
> + *ebx = env->apic_bus_freq / 1000; /* Hz to KHz */
> + *ecx = 0;
> + *edx = 0;
> + break;
> + default:
> + cpu_x86_cpuid(env, index, count, eax, ebx, ecx, edx);
> + break;
> + }
> +}
> +
> int hvf_vcpu_exec(CPUState *cpu)
> {
> X86CPU *x86_cpu = X86_CPU(cpu);
> @@ -734,7 +822,7 @@ int hvf_vcpu_exec(CPUState *cpu)
> uint32_t rcx = (uint32_t)rreg(cpu->hvf_fd, HV_X86_RCX);
> uint32_t rdx = (uint32_t)rreg(cpu->hvf_fd, HV_X86_RDX);
>
> - cpu_x86_cpuid(env, rax, rcx, &rax, &rbx, &rcx, &rdx);
> + hvf_cpu_x86_cpuid(env, rax, rcx, &rax, &rbx, &rcx, &rdx);
>
> wreg(cpu->hvf_fd, HV_X86_RAX, rax);
> wreg(cpu->hvf_fd, HV_X86_RBX, rbx);
> --
> 2.28.0
>