Support QEMU cpu models in MSHV accelerator
v4 cover Support QEMU cpu models in MSHV accelerator
Hey all,
In the current MSHV accelerator code passing CPU features via the -cpu flag doesn't work as intended yet. When using the MSHV hypervisor we either silently discard the specified model/features and leave it up to the hypervisor to provide a sensible set of features or if the user selects -cpu host, the hypervisor might refuse to create a partition.
This changeset introduces a more comprehensive support for passing desired guest cpu features to the hypervisor. It's also a prerequisite for Live Migration support.
Changes since v3:
- Rebased onto current master
Changes since v2:
- move disabling of la57 to arch specific code
- assert MSHV_NUM_CPU_FEATURES_BANKS == 2
- dropped QEMU_PACKED changes to hv headers (address in individal patch)
Changes since v1:
- query hypervisor for supported processor features instead of the host OS
- disable la57 (5 level page tables)
best,
magnus
Magnus Kulke (9):
accel/mshv: use mshv_create_partition_v2 payload target/i386/mshv: fix cpuid propagation bug target/i386/mshv: fix various cpuid traversal bugs target/i386/mshv: change cpuid mask to UINT32_MAX target/i386/mshv: set cpu model name on -cpu host target/i386: query mshv accel for supported cpuids target/i386/mshv: populate xsave area offsets target/i386/mshv: filter out CET bits in cpuid accel/mshv: disable la57 (5lvl paging)
accel/mshv/mshv-all.c | 78 ++++++++++- include/hw/hyperv/hvgdk_mini.h | 2 + include/hw/hyperv/hvhdk.h | 195 +++++++++++++++++++++++++++ include/system/mshv.h | 3 + include/system/mshv_int.h | 2 + target/i386/cpu.c | 8 ++ target/i386/mshv/mshv-cpu.c | 233 +++++++++++++++++++++++++++------ 7 files changed, 472 insertions(+), 49 deletions(-)
Signature · 1 line
2.34.1
PATCH v4 1/9 accel/mshv: use mshv_create_partition_v2 payload
When using the extended request format we can instruct the hypervisor to provision enough space for requested XSAVE features. This is required for supporting QEMU models provided via the -cpu flag properly.
--- accel/mshv/mshv-all.c | 71 ++++++++++-- include/hw/hyperv/hvgdk_mini.h | 2 + include/hw/hyperv/hvhdk.h | 195 +++++++++++++++++++++++++++++++++ 3 files changed, 261 insertions(+), 7 deletions(-)
accel/mshv/mshv-all.c +64 -7
@@ -110,21 +110,78 @@ static int resume_vm(int vm_fd)return 0 ; } +static int get_host_partition_property (int mshv_fd,uint32_t property_code, +uint64_t *value) +{ +int ret; +struct hv_input_get_partition_property in = {0 }; +struct hv_output_get_partition_property out = {0 }; +struct mshv_root_hvcall args = {0 }; + + in.property_code = property_code; + + args.code =HVCALL_GET_PARTITION_PROPERTY ; + args.in_sz =sizeof (in); + args.in_ptr = (uint64_t )∈ + args.out_sz =sizeof (out); + args.out_ptr = (uint64_t )&out; + + ret =ioctl (mshv_fd,MSHV_ROOT_HVCALL , &args); +if (ret <0 ) { +error_report ("Failed to get host partition property bank: %s" , +strerror (errno)); +return -1 ; + } + + *value = out.property_value; +return 0 ; +} +static int create_partition (int mshv_fd,int *vm_fd) {int ret; -struct mshv_create_partition args = {0 }; +uint64_t pt_flags, host_proc_features; +union hv_partition_processor_xsave_features disabled_xsave_features; +struct mshv_create_partition_v2 args = {0 }; + +QEMU_BUILD_BUG_ON (MSHV_NUM_CPU_FEATURES_BANKS !=2 );/* Initialize pt_flags with the desired features */ -uint64_t pt_flags = (1ULL <<MSHV_PT_BIT_LAPIC ) | - (1ULL <<MSHV_PT_BIT_X2APIC ) | - (1ULL <<MSHV_PT_BIT_GPA_SUPER_PAGES ); + pt_flags = (1ULL <<MSHV_PT_BIT_LAPIC ) | + (1ULL <<MSHV_PT_BIT_X2APIC ) | + (1ULL <<MSHV_PT_BIT_GPA_SUPER_PAGES ) | + (1ULL <<MSHV_PT_BIT_CPU_AND_XSAVE_FEATURES ); + +/* enable all */ + disabled_xsave_features.as_uint64 =0 ; + +/* +* query host for supported processor features and disable unsupported +* features: (0 means supported, 1 means disabled, hence the negation) +*/ + ret =get_host_partition_property (mshv_fd, +HV_PARTITION_PROPERTY_PROCESSOR_FEATURES0 , + &host_proc_features); +if (ret <0 ) { +error_report ("Failed to get host processor feature bank 0" ); +return -1 ; + } + args.pt_cpu_fbanks[0 ] = ~host_proc_features; -/* Set default isolation type */ -uint64_t pt_isolation =MSHV_PT_ISOLATION_NONE ; + ret =get_host_partition_property (mshv_fd, +HV_PARTITION_PROPERTY_PROCESSOR_FEATURES1 , + &host_proc_features); +if (ret <0 ) { +error_report ("Failed to get host processor feature bank 1" ); +return -1 ; + } + args.pt_cpu_fbanks[1 ] = ~host_proc_features; +/* populate args structure */ args.pt_flags = pt_flags; - args.pt_isolation = pt_isolation; + args.pt_isolation =MSHV_PT_ISOLATION_NONE ; + args.pt_disabled_xsave = disabled_xsave_features.as_uint64; + args.pt_num_cpu_fbanks =MSHV_NUM_CPU_FEATURES_BANKS ; ret =ioctl (mshv_fd,MSHV_CREATE_PARTITION , &args);if (ret <0 ) {
include/hw/hyperv/hvgdk_mini.h +2 -0
@@ -454,6 +454,8 @@ typedef struct hv_input_set_vp_registers {struct hv_register_assoc elements[]; }QEMU_PACKED hv_input_set_vp_registers; +#define MSHV_VP_MAX_REGISTERS 128 +union hv_interrupt_control {uint64_t as_uint64;struct {
include/hw/hyperv/hvhdk.h +195 -0
207 lines not on this page · raw message
Signature · 1 line
2.34.1
PATCH v4 2/9 target/i386/mshv: fix cpuid propagation bug
In reply to Magnus KulkeThe list argument was missing a pointer and hence e.g. -cpu qemu64 and -cpu EPYC would produce the same of cpuid leaves.
--- target/i386/mshv/mshv-cpu.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
target/i386/mshv/mshv-cpu.c +4 -4
@@ -441,7 +441,7 @@ int mshv_load_regs(CPUState *cpu)return 0 ; } -static void add_cpuid_entry (GList *cpuid_entries, +static void add_cpuid_entry (GList **cpuid_entries,uint32_t function,uint32_t index,uint32_t eax,uint32_t ebx,uint32_t ecx,uint32_t edx)
@@ -456,10 +456,10 @@ static void add_cpuid_entry(GList *cpuid_entries, entry->ecx = ecx; entry->edx = edx; - cpuid_entries =g_list_append (cpuid_entries, entry); + *cpuid_entries =g_list_append (*cpuid_entries, entry); } -static void collect_cpuid_entries (const CPUState *cpu,GList *cpuid_entries) +static void collect_cpuid_entries (const CPUState *cpu,GList **cpuid_entries) {X86CPU *x86_cpu =X86_CPU (cpu);CPUX86State *env = &x86_cpu->env;
@@ -615,7 +615,7 @@ static int set_cpuid2(const CPUState *cpu)struct hv_cpuid_entry *entry;GList *entries =NULL ; -collect_cpuid_entries (cpu, entries); +collect_cpuid_entries (cpu, &entries); n_entries =g_list_length (entries); cpuid_size =sizeof (struct hv_cpuid )
Signature · 1 line
2.34.1
PATCH v4 3/9 target/i386/mshv: fix various cpuid traversal bugs
In reply to Magnus Kulke- Hardcoded max_leaf was not accurate. We query leaf 0x0 and 0x80000000 to get the actual max leaves
- On all 0 zeroes on leaf 0x0d, we register 0-63 subleaves with zeros indicating XSAVE is disabled
- Subleaf 0 was hardcoded, so the Hypervisor returned defaults for other subleaves
- Subleaf 0 was hardcoded, so we were passing 0 instead of actual subleaf when adding entries. We now pass the correct subleaf value to add_cpuid_entry()
- Leaves 0x04,0x07,0d,0f,10 weren't marked as subleaf-specific
--- target/i386/mshv/mshv-cpu.c | 88 +++++++++++++++++++++++++------------ 1 file changed, 61 insertions(+), 27 deletions(-)
target/i386/mshv/mshv-cpu.c +61 -27
@@ -465,14 +465,17 @@ static void collect_cpuid_entries(const CPUState *cpu, GList **cpuid_entries)CPUX86State *env = &x86_cpu->env;uint32_t eax, ebx, ecx, edx;uint32_t leaf, subleaf; -size_t max_leaf =0x1F ; -size_t max_subleaf =0x20 ; - -uint32_t leaves_with_subleaves[] = {0x4 ,0x7 ,0xD ,0xF ,0x10 }; +uint32_t max_basic_leaf, max_extended_leaf; +uint32_t max_subleaf =0x20 ; +uint32_t leaves_with_subleaves[] = {0x04 ,0x07 ,0x0d ,0x0f ,0x10 };int n_subleaf_leaves =ARRAY_SIZE (leaves_with_subleaves); -/* Regular leaves without subleaves */ -for (leaf =0 ; leaf <= max_leaf; leaf++) { +/* Get maximum basic and and extended CPUID leaves */ +cpu_x86_cpuid (env,0 ,0 , &max_basic_leaf, &ebx, &ecx, &edx); +cpu_x86_cpuid (env,0x80000000 ,0 , &max_extended_leaf, &ebx, &ecx, &edx); + +/* Collect basic leaves (0x0 to max_basic_leaf) */ +for (leaf =0 ; leaf <= max_basic_leaf; leaf++) {bool has_subleaves = false;for (int i =0 ; i < n_subleaf_leaves; i++) {if (leaf == leaves_with_subleaves[i]) {
@@ -483,27 +486,40 @@ static void collect_cpuid_entries(const CPUState *cpu, GList **cpuid_entries)if (!has_subleaves) {cpu_x86_cpuid (env, leaf,0 , &eax, &ebx, &ecx, &edx); -if (eax ==0 && ebx ==0 && ecx ==0 && edx ==0 ) { -/* all zeroes indicates no more leaves */ -continue ; - } -add_cpuid_entry (cpuid_entries, leaf,0 , eax, ebx, ecx, edx);continue ; } +/* +* Valid XSAVE components can exist at a higher index se we need to set +* all subleaves for leaf 0x0d, even if we encounter an empty one. +*/ +if (leaf ==0x0d ) { +for (subleaf =0 ; subleaf <=63 ; subleaf++) { +cpu_x86_cpuid (env, leaf, subleaf, &eax, &ebx, &ecx, &edx); +add_cpuid_entry (cpuid_entries, leaf, subleaf, + eax, ebx, ecx, edx); + } +continue ; + } + subleaf =0 ;while (subleaf < max_subleaf) {cpu_x86_cpuid (env, leaf, subleaf, &eax, &ebx, &ecx, &edx);if (eax ==0 && ebx ==0 && ecx ==0 && edx ==0 ) { -/* all zeroes indicates no more leaves */ break ; } -add_cpuid_entry (cpuid_entries, leaf,0 , eax, ebx, ecx, edx); +add_cpuid_entry (cpuid_entries, leaf, subleaf, eax, ebx, ecx, edx); subleaf++; } } + +/* Collect extended leaves (0x80000000 to max_extended_leaf) */ +for (leaf =0x80000000 ; leaf <= max_extended_leaf; leaf++) { +cpu_x86_cpuid (env, leaf,0 , &eax, &ebx, &ecx, &edx); +add_cpuid_entry (cpuid_entries, leaf,0 , eax, ebx, ecx, edx); + } } static int register_intercept_result_cpuid_entry(const CPUState *cpu,
@@ -576,22 +592,40 @@ static int register_intercept_result_cpuid(const CPUState *cpu, subleaf_specific =0 ; always_override =1 ; -/* Intel */ -/* 0xb - Extended Topology Enumeration Leaf */ -/* 0x1f - V2 Extended Topology Enumeration Leaf */ -/* AMD */ -/* 0x8000_001e - Processor Topology Information */ -/* 0x8000_0026 - Extended CPU Topology */ -if (entry->function ==0xb - || entry->function ==0x1f - || entry->function ==0x8000001e - || entry->function ==0x80000026 ) { +/* +* Intel +* 0xb - Extended Topology Enumeration Leaf +* 0x1f - V2 Extended Topology Enumeration Leaf +* AMD +* 0x8000_001e - Processor Topology Information +* 0x8000_0026 - Extended CPU Topology +*/ +if (entry->function ==0xb || + entry->function ==0x1f || + entry->function ==0x8000001e || + entry->function ==0x80000026 ) { + subleaf_specific =1 ; + always_override =1 ; +/* +* Feature enumeration leaves (subleaf-specific) +* 0x04: Deterministic Cache Parameters +* 0x07: Structured Extended Feature Flags +* 0x0D: Processor Extended State Enumeration +* 0x0F: Platform QoS Monitoring +* 0x10: Platform QoS Enforcement +*/ + }else if (entry->function ==0x04 || + entry->function ==0x07 || + entry->function ==0x0d || + entry->function ==0x0f || + entry->function ==0x10 ) { subleaf_specific =1 ; always_override =1 ; - }else if (entry->function ==0x00000001 - || entry->function ==0x80000000 - || entry->function ==0x80000001 - || entry->function ==0x80000008 ) { +/* Basic feature leaves (no subleaves) */ + }else if (entry->function ==0x00000001 || + entry->function ==0x80000000 || + entry->function ==0x80000001 || + entry->function ==0x80000008 ) { subleaf_specific =0 ; always_override =1 ; }
Signature · 1 line
2.34.1
PATCH v4 4/9 target/i386/mshv: change cpuid mask to UINT32_MAX
In reply to Magnus KulkeThe current implementation would only override the feature bits that are enabled, however we also want to consider disabled features, hence all bits are set on the masks in the hypercall argument.
--- target/i386/mshv/mshv-cpu.c | 17 ++++++----------- 1 file changed, 6 insertions(+), 11 deletions(-)
target/i386/mshv/mshv-cpu.c +6 -11
@@ -538,22 +538,17 @@ static int register_intercept_result_cpuid_entry(const CPUState *cpu, .input.always_override = always_override, .input.padding = 0,/* -* With regard to masks - these are to specify bits to be overwritten -* The current CpuidEntry structure wouldn't allow to carry the masks -* in addition to the actual register values. For this reason, the -* masks are set to the exact values of the corresponding register bits -* to be registered for an overwrite. To view resulting values the -* hypervisor would return, HvCallGetVpCpuidValues hypercall can be -* used. +* Masks specify which bits to override. Set to 0xFFFFFFFF to +* override all bits with the values from the QEMU CPU model. */ .result.eax = entry->eax, - .result.eax_mask = entry->eax, + .result.eax_mask = 0xFFFFFFFF, .result.ebx = entry->ebx, - .result.ebx_mask = entry->ebx, + .result.ebx_mask = 0xFFFFFFFF, .result.ecx = entry->ecx, - .result.ecx_mask = entry->ecx, + .result.ecx_mask = 0xFFFFFFFF, .result.edx = entry->edx, - .result.edx_mask = entry->edx, + .result.edx_mask = 0xFFFFFFFF, };union hv_register_intercept_result_parameters parameters = { .cpuid = cpuid_params,
Signature · 1 line
2.34.1
PATCH v4 5/9 target/i386/mshv: set cpu model name on -cpu host
In reply to Magnus KulkeBy registering a cpu_accel_class for mshv, we can report an accurate model name on -cpu host.
--- target/i386/mshv/mshv-cpu.c | 31 +++++++++++++++++++++++++++++++ 1 file changed, 31 insertions(+)
target/i386/mshv/mshv-cpu.c +31 -0
@@ -24,10 +24,13 @@#include "hw/i386/apic_internal.h" #include "cpu.h" +#include "host-cpu.h" #include "emulate/x86_decode.h" #include "emulate/x86_emu.h" #include "emulate/x86_flags.h" +#include "accel/accel-cpu-target.h" +#include "trace-accel_mshv.h" #include "trace.h"
@@ -1675,3 +1678,31 @@ int mshv_arch_post_init_vm(int vm_fd)return ret; } + +static void mshv_cpu_instance_init (CPUState *cs) +{ +X86CPU *cpu =X86_CPU (cs); + +host_cpu_instance_init (cpu); +} + +static void mshv_cpu_accel_class_init (ObjectClass *oc,const void *data) +{ +AccelCPUClass *acc =ACCEL_CPU_CLASS (oc); + + acc->cpu_instance_init = mshv_cpu_instance_init; +} + +static const TypeInfo mshv_cpu_accel_type_info = { + .name =ACCEL_CPU_NAME ("mshv" ), + .parent =TYPE_ACCEL_CPU , + .class_init = mshv_cpu_accel_class_init, + .abstract = true, +}; + +static void mshv_cpu_accel_register_types (void ) +{ +type_register_static (&mshv_cpu_accel_type_info); +} + +type_init (mshv_cpu_accel_register_types);
Signature · 1 line
2.34.1
PATCH v4 6/9 target/i386: query mshv accel for supported cpuids
In reply to Magnus KulkeWe implement mshv_get_supported_cpuid() and invoke it in x86_cpu_get_supported_feature_word() retrieve the cpu features that the host is supporting. Initially we mask the virtualization capabilitities potentially we might need to mask more in the future.
--- include/system/mshv.h | 3 +++ target/i386/cpu.c | 8 ++++++++ target/i386/mshv/mshv-cpu.c | 27 +++++++++++++++++++++++++++ 3 files changed, 38 insertions(+)
include/system/mshv.h +3 -0
@@ -60,4 +60,7 @@ int mshv_irqchip_add_irqfd_notifier_gsi(const EventNotifier *n, const EventNotifier *rn, int virq);int mshv_irqchip_remove_irqfd_notifier_gsi (const EventNotifier *n,int virq); +/* cpuid */ +uint32_t mshv_get_supported_cpuid (uint32_t func,uint32_t idx,int reg); +#endif
target/i386/cpu.c +8 -0
@@ -26,6 +26,7 @@#include "tcg/helper-tcg.h" #include "exec/translation-block.h" #include "system/hvf.h" +#include "system/mshv.h" #include "hvf/hvf-i386.h" #include "kvm/kvm_i386.h" #include "kvm/tdx.h"
@@ -8087,6 +8088,13 @@ uint64_t x86_cpu_get_supported_feature_word(X86CPU *cpu, FeatureWord w) r =hvf_get_supported_cpuid (wi->cpuid.eax, wi->cpuid.ecx, wi->cpuid.reg); + }else if (mshv_enabled ()) { +if (wi->type !=CPUID_FEATURE_WORD ) { +return 0 ; + } + r =mshv_get_supported_cpuid (wi->cpuid.eax, + wi->cpuid.ecx, + wi->cpuid.reg); }else if (tcg_enabled () ||qtest_enabled ()) { r = wi->tcg_features; }else {
target/i386/mshv/mshv-cpu.c +27 -0
@@ -1639,6 +1639,33 @@ void mshv_arch_destroy_vcpu(CPUState *cpu)g_clear_pointer (&env->emu_mmio_buf, g_free); } +uint32_t mshv_get_supported_cpuid (uint32_t func,uint32_t idx,int reg) +{ +uint32_t eax, ebx, ecx, edx; +uint32_t ret =0 ; + +host_cpuid (func, idx, &eax, &ebx, &ecx, &edx); +switch (reg) { +case R_EAX : + ret = eax;break ; +case R_EBX : + ret = ebx;break ; +case R_ECX : + ret = ecx;break ; +case R_EDX : + ret = edx;break ; + } + +/* Disable nested virtualization features not yet supported by MSHV */ +if (func ==0x80000001 && reg ==R_ECX ) { + ret &= ~CPUID_EXT3_SVM ; + } +if (func ==0x01 && reg ==R_ECX ) { + ret &= ~CPUID_EXT_VMX ; + } +return ret; +} +/* * Default Microsoft Hypervisor behavior for unimplemented MSR is to send a * fault to the guest if it tries to access it. It is possible to override
Signature · 1 line
2.34.1
PATCH v4 7/9 target/i386/mshv: populate xsave area offsets
In reply to Magnus KulkeThis function is copied verbatim from kvm_cpu_xsave_init(), for MSHV we need to populate the offsets similarly.
--- target/i386/mshv/mshv-cpu.c | 31 +++++++++++++++++++++++++++++++ 1 file changed, 31 insertions(+)
target/i386/mshv/mshv-cpu.c +31 -0
@@ -1706,11 +1706,42 @@ int mshv_arch_post_init_vm(int vm_fd)return ret; } +static void mshv_cpu_xsave_init (void ) +{ +static bool first = true; +uint32_t eax, ebx, ecx, edx; +int i; + +if (!first) { +return ; + } + first = false; + +/* x87 and SSE states are in the legacy region of the XSAVE area. */ + x86_ext_save_areas[XSTATE_FP_BIT ].offset =0 ; + x86_ext_save_areas[XSTATE_SSE_BIT ].offset =0 ; + +for (i =XSTATE_SSE_BIT +1 ; i <XSAVE_STATE_AREA_COUNT ; i++) { +ExtSaveArea *esa = &x86_ext_save_areas[i]; + +if (!esa->size) { +continue ; + } +host_cpuid (0xd , i, &eax, &ebx, &ecx, &edx); +if (eax !=0 ) { +assert (esa->size == eax); + esa->offset = ebx; + esa->ecx = ecx; + } + } +} +static void mshv_cpu_instance_init (CPUState *cs) {X86CPU *cpu =X86_CPU (cs);host_cpu_instance_init (cpu); +mshv_cpu_xsave_init (); }static void mshv_cpu_accel_class_init (ObjectClass *oc,const void *data)
Signature · 1 line
2.34.1
PATCH v4 8/9 target/i386/mshv: filter out CET bits in cpuid
In reply to Magnus Kulkee46dcef1d0 introduced CET_U/CET_S xstate bits in the ExtSaveArea. The response in EBX of a 0x4,1 cpuid query will contain a size contingent on what the guest enabled in the IA32_XSS MSR (compacted xsave size).
Since the MSHV accelerator currently responds with static answers to 0x4d cpuid queries, we'll have to disable those bits, even if the host supports them.
--- target/i386/mshv/mshv-cpu.c | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+)
target/i386/mshv/mshv-cpu.c +19 -0
@@ -1663,6 +1663,25 @@ uint32_t mshv_get_supported_cpuid(uint32_t func, uint32_t idx, int reg)if (func ==0x01 && reg ==R_ECX ) { ret &= ~CPUID_EXT_VMX ; } + +/* +* MSHV currently uses static CPUID intercept results for leaf 0xD. +* However there are feature-responses that are dynamic based on what a +* guest enables in XCR0 and XSS, such as CET shadow stack. +* +* A guest which doesn't know about those features yet would encounter an +* unexpcted CPUID[0xD,1].EBX (compactes XSAVE size) and either fail +* or gracefully degrade by not using XSAVE at all. +* +* To avoid this, we filter out supervisor xstate features. +*/ +if (func ==0x07 && idx ==0 && reg ==R_ECX ) { + ret &= ~CPUID_7_0_ECX_CET_SHSTK ; + } +if (func ==0x07 && idx ==0 && reg ==R_EDX ) { + ret &= ~CPUID_7_0_EDX_CET_IBT ; + } +return ret; }
Signature · 1 line
2.34.1
Re: [PATCH v4 8/9] target/i386/mshv: filter out CET bits in cpuid
Quoted 24 lines · Magnus Kulke
e46dcef1d0introduced CET_U/CET_S xstate bits in the ExtSaveArea. The response in EBX of a 0x4,1 cpuid query will contain a size contingent on what the guest enabled in theIA32_XSSMSR (compacted xsave size).Since the MSHV accelerator currently responds with static answers to
0x4dcpuid queries, we'll have to disable those bits, even if the host supports them.Signed-off-by Magnus Kulke <[email protected]>--- target/i386/mshv/mshv-cpu.c | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+)target/i386/mshv/mshv-cpu.c +2
-0@@ -1663,6 +1663,25 @@ uint32_t mshv_get_supported_cpuid(uint32_t func, uint32_t idx, int reg)if (func ==0x01 && reg ==R_ECX ) { ret &= ~CPUID_EXT_VMX ; } + +/*
target/i386/mshv/mshv-cpu.c uint32_t mshv_get_supported_cpuid(uint32_t func, uint32_t idx, int reg) · line 1677
Quoted 10 lines
target/i386/mshv/mshv-cpu.c +10
-0+* MSHV currently uses static CPUID intercept results for leaf 0xD. +* However there are feature-responses that are dynamic based on what a +* guest enables in XCR0 and XSS, such as CET shadow stack. +* +* A guest which doesn't know about those features yet would encounter an +* unexpcted CPUID[0xD,1].EBX (compactes XSAVE size) and either fail +* or gracefully degrade by not using XSAVE at all. +* +* To avoid this, we filter out supervisor xstate features. +*/
I don't understand this, but it seems to be a Linux bug. Don't *all* xsave features cause changes to CPUID[0xD,1].EBX?
Paolo
Quoted 10 lines
target/i386/mshv/mshv-cpu.c +7
-0+if (func ==0x07 && idx ==0 && reg ==R_ECX ) { + ret &= ~CPUID_7_0_ECX_CET_SHSTK ; + } +if (func ==0x07 && idx ==0 && reg ==R_EDX ) { + ret &= ~CPUID_7_0_EDX_CET_IBT ; + } +return ret; }
Re: [PATCH v4 8/9] target/i386/mshv: filter out CET bits in cpuid
Quoted 2 lines · Paolo Bonzini
I don't understand this, but it seems to be a Linux bug. Don't *all* xsave features cause changes to CPUID[0xD,1].EBX?
oh, you're right, thx. the CPUID[0xD,1].EBX responses (also CPUID[0xD,0]) returned wrong responses in all cases the introduction of CET in QEMU probably just triggered some overflow, but it was broken before.
We cannot statically register size responses for xsave features that a guest can dynamically enable/disable.
I was planning to implement a dynamic handler for the xsave size responses anyway, but maybe there's an easier way: We can provide an ebx mask = 0 in the payload when we register the cpuid responses at the hypervisor. If we do that for CPUID[0xD,{0,1}], the hypervisor should return the answer based on the guest state (XCR0,XSS).
I'll test this approach and probably post a new revision with one of those options.
Re: [PATCH v4 8/9] target/i386/mshv: filter out CET bits in cpuid
Quoted 8 lines · Magnus Kulke
Quoted 2 lines · Paolo Bonzini
I don't understand this, but it seems to be a Linux bug. Don't *all* xsave features cause changes to CPUID[0xD,1].EBX?
Quoted 6 lines
oh, you're right, thx. the CPUID[0xD,1].EBX responses (also CPUID[0xD,0]) returned wrong responses in all cases the introduction of CET in QEMU probably just triggered some overflow, but it was broken before.
We cannot statically register size responses for xsave features that a guest can dynamically enable/disable.
Hello,
Is CPUID performance such a problem?
Quoted 10 lines
I was planning to implement a dynamic handler for the xsave size responses anyway, but maybe there's an easier way: We can provide an ebx mask = 0 in the payload when we register the cpuid responses at the hypervisor. If we do that for CPUID[0xD,{0,1}], the hypervisor should return the answer based on the guest state (XCR0,XSS).
I'll test this approach and probably post a new revision with one of those options.
What I have from WHPX on this - where for now CPUID vmexits are used instead of this fancy interface* - although that might change in the future:
Those have been seen to vary dynamically for a vCPU:
CPUID[1:EDX].APICCPUID_7_0_EDX_CET_IBTCPUID_7_0_ECX_CET_SHSTKCPUID_7_0_ECX_OSPKECPUID_EXT_OSXSAVE
The whole 0xD leave
I used the default result to override with the Hyper-V stock response for those bits, although masking as provided by the new interface would work too.
* not using it yet because Windows 10 Hyper-V didn’t have it, and the perf impact for CPUIDs didn’t sound like an actual problem so trapped as much leaves as it could make sense
Re: [PATCH v4 8/9] target/i386/mshv: filter out CET bits in cpuid
Quoted 14 lines · Mohamed Mediouni
Is CPUID performance such a problem?
What I have from WHPX on this - where for now CPUID vmexits are used instead of this fancy interface* - although that might change in the future:
Those have been seen to vary dynamically for a vCPU:
CPUID[1:EDX].APIC
CPUID_7_0_EDX_CET_IBTCPUID_7_0_ECX_CET_SHSTKCPUID_7_0_ECX_OSPKECPUID_EXT_OSXSAVE
The whole0xDleave
Quoted 6 lines
I used the default result to override with the Hyper-V stock response for those bits, although masking as provided by the new interface would work too.
* not using it yet because Windows 10 Hyper-V didn’t have it, and the perf impact for CPUIDs didn’t sound like an actual problem so trapped as much leaves as it could make sense
Hey Mohamed. yeah, it's arguable.
I would prefer not to introduce a VM exit. You are probably right that it doesn't matter much, however I can imagine that there are workloads that perform feature detection in a hot path (e.g. right before some AVX computation) and trapping it to the vmm would impact perf for those.
But yeah, if it turns out that the static responses and the mask do not work, we have to consider this.
best,
magnus
PATCH v4 9/9 accel/mshv: disable la57 (5lvl paging)
In reply to Magnus KulkeThis change disable la57 paging on the mshv hypervisor on both the mshv processor feature bitmap and mask the cpuid feature leaf to the guest.
Since the removal of hypervisor-assisted gva=>gpa translation in 1c85a4a3d7 we have seen MMIO errors in guests on la57-enabled hw. We will have to investigate and test this further.
--- accel/mshv/mshv-all.c | 7 +++++++ include/system/mshv_int.h | 2 ++ target/i386/mshv/mshv-cpu.c | 12 ++++++++++++ 3 files changed, 21 insertions(+)
accel/mshv/mshv-all.c +7 -0
@@ -142,6 +142,8 @@ static int create_partition(int mshv_fd, int *vm_fd)int ret;uint64_t pt_flags, host_proc_features;union hv_partition_processor_xsave_features disabled_xsave_features; +union hv_partition_processor_features disabled_partition_features = {0 }; +struct mshv_create_partition_v2 args = {0 };QEMU_BUILD_BUG_ON (MSHV_NUM_CPU_FEATURES_BANKS !=2 );
@@ -177,6 +179,11 @@ static int create_partition(int mshv_fd, int *vm_fd) } args.pt_cpu_fbanks[1 ] = ~host_proc_features; +/* arch-specific features we disable regardless of host support */ +mshv_arch_disable_partition_proc_features (&disabled_partition_features); + args.pt_cpu_fbanks[0 ] |= disabled_partition_features.as_uint64[0 ]; + args.pt_cpu_fbanks[1 ] |= disabled_partition_features.as_uint64[1 ]; +/* populate args structure */ args.pt_flags = pt_flags; args.pt_isolation =MSHV_PT_ISOLATION_NONE ;
include/system/mshv_int.h +2 -0
@@ -94,6 +94,8 @@ void mshv_arch_init_vcpu(CPUState *cpu);void mshv_arch_destroy_vcpu (CPUState *cpu);void mshv_arch_amend_proc_features (union hv_partition_synthetic_processor_features *features); +void mshv_arch_disable_partition_proc_features ( +union hv_partition_processor_features *disabled_features);int mshv_arch_post_init_vm (int vm_fd);typedef struct mshv_root_hvcall mshv_root_hvcall ;
target/i386/mshv/mshv-cpu.c +12 -0
@@ -1111,6 +1111,12 @@ void mshv_arch_amend_proc_features( features->access_guest_idle_reg =1 ; } +void mshv_arch_disable_partition_proc_features ( +union hv_partition_processor_features *disabled_features) +{ + disabled_features->la57_support =1 ; +} +static int set_memory_info (const struct hyperv_message *msg,struct hv_x64_memory_intercept_message *info) {
@@ -1677,6 +1683,12 @@ uint32_t mshv_get_supported_cpuid(uint32_t func, uint32_t idx, int reg)*/ if (func ==0x07 && idx ==0 && reg ==R_ECX ) { ret &= ~CPUID_7_0_ECX_CET_SHSTK ; +/* +* LA57 (5-level paging) causes incorrect GVA=>GPA translations +* in the instruction decoder/emulator. Disable until page table +* walk in x86_mmu.c works w/ 5-level paging. +*/ + ret &= ~CPUID_7_0_ECX_LA57 ; }if (func ==0x07 && idx ==0 && reg ==R_EDX ) { ret &= ~CPUID_7_0_EDX_CET_IBT ;
Signature · 1 line
2.34.1
Re: [PATCH v4 9/9] accel/mshv: disable la57 (5lvl paging)
Quoted 10 lines · Magnus Kulke
This change disable la57 paging on the mshv hypervisor on both the mshv processor feature bitmap and mask the cpuid feature leaf to the guest.
Since the removal of hypervisor-assisted gva=>gpa translation in
1c85a4a3d7we have seen MMIO errors in guests on la57-enabled hw. We will have to investigate and test this further.Signed-off-by Magnus Kulke <[email protected]>