kvm: x86: Allow userspace to handle emulation errors
Subsystems (5)
PATCH v2 1/2 kvm: x86: Allow userspace to handle emulation errors
Add a fallback mechanism to the in-kernel instruction emulator that allows userspace the opportunity to process an instruction the emulator was unable to. When the in-kernel instruction emulator fails to process an instruction it will either inject a #UD into the guest or exit to userspace with exit reason KVM_INTERNAL_ERROR. This is because it does not know how to proceed in an appropriate manner. This feature lets userspace get involved to see if it can figure out a better path forward.
--- Documentation/virt/kvm/api.rst | 19 ++++++++++++++++++ arch/x86/include/asm/kvm_host.h | 6 ++++++ arch/x86/kvm/x86.c | 35 +++++++++++++++++++++++++++++---- include/uapi/linux/kvm.h | 23 ++++++++++++++++++++++ tools/include/uapi/linux/kvm.h | 23 ++++++++++++++++++++++ 5 files changed, 102 insertions(+), 4 deletions(-)
Documentation/virt/kvm/api.rst +19 -0
@@ -6233,6 +6233,25 @@ KVM_RUN_BUS_LOCK flag is used to distinguish between them. This capability can be used to check / enable 2nd DAWR feature provided by POWER10 processor. +7.24 KVM_CAP_EXIT_ON_EMULATION_FAILURE +-------------------------------------- + +:Architectures: x86 +:Parameters: args[0] whether the feature should be enabled or not + +With this capability enabled, the in-kernel instruction emulator packs the exit +struct of KVM_INTERNAL_ERROR with the instruction length and instruction bytes +when an error occurs while emulating an instruction. This allows userspace to +then take a look at the instruction and see if it is able to handle it more +gracefully than the in-kernel emulator. + +When this capability is enabled use the emulation_failure struct instead of the +internal struct for the exit struct. They have the same layout, but the +emulation_failure struct matches the content better. + +It is noted that KVM still exits on certain types (skip) even if this capability +is not enabled, and KVM will never exit if VMware #GP emulation fails. + 8. Other capabilities. ======================
arch/x86/include/asm/kvm_host.h +6 -0
@@ -1049,6 +1049,12 @@ struct kvm_arch {bool exception_payload_enabled;bool bus_lock_detection_enabled; +/* +* If exit_on_emulation_error is set, and the in-kernel instruction +* emulator fails to emulate an instruction, allow userspace +* the opportunity to look at it. +*/ +bool exit_on_emulation_error;/* Deflect RDMSR and WRMSR to user space when they trigger a #GP */ u32 user_space_msr_mask;
arch/x86/kvm/x86.c +31 -4
@@ -3771,6 +3771,7 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)case KVM_CAP_X86_USER_SPACE_MSR :case KVM_CAP_X86_MSR_FILTER :case KVM_CAP_ENFORCE_PV_FEATURE_CPUID : +case KVM_CAP_EXIT_ON_EMULATION_FAILURE : r =1 ;break ;#ifdef CONFIG_KVM_XEN
@@ -5357,6 +5358,10 @@ int kvm_vm_ioctl_enable_cap(struct kvm *kvm, kvm->arch.bus_lock_detection_enabled = true; r =0 ;break ; +case KVM_CAP_EXIT_ON_EMULATION_FAILURE : + kvm->arch.exit_on_emulation_error = cap->args[0 ]; + r =0 ; +break ;default : r = -EINVAL ;break ;
@@ -7119,8 +7124,31 @@ void kvm_inject_realmode_interrupt(struct kvm_vcpu *vcpu, int irq, int inc_eip) }EXPORT_SYMBOL_GPL (kvm_inject_realmode_interrupt); +static void prepare_emulation_failure_exit (struct kvm_vcpu *vcpu) +{ +struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt; +u64 insn_size = ctxt->fetch.end - ctxt->fetch.data; +struct kvm_run *run = vcpu->run; + + + run->exit_reason =KVM_EXIT_INTERNAL_ERROR ; + run->emulation_failure.suberror =KVM_INTERNAL_ERROR_EMULATION ; + run->emulation_failure.ndata =0 ; + run->emulation_failure.flags =0 ; +if (insn_size) { + run->emulation_failure.ndata =3 ; + run->emulation_failure.flags |= +KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES ; + run->emulation_failure.insn_size = insn_size; +memcpy (run->emulation_failure.insn_bytes, + ctxt->fetch.data,sizeof (ctxt->fetch.data)); + } +} +static int handle_emulation_failure (struct kvm_vcpu *vcpu,int emulation_type) { +struct kvm *kvm = vcpu->kvm; + ++vcpu->stat.insn_emulation_fail;trace_kvm_emulate_insn_failed (vcpu);
@@ -7129,10 +7157,9 @@ static int handle_emulation_failure(struct kvm_vcpu *vcpu, int emulation_type)return 1 ; } -if (emulation_type &EMULTYPE_SKIP ) { - vcpu->run->exit_reason =KVM_EXIT_INTERNAL_ERROR ; - vcpu->run->internal.suberror =KVM_INTERNAL_ERROR_EMULATION ; - vcpu->run->internal.ndata =0 ; +if (kvm->arch.exit_on_emulation_error || + (emulation_type &EMULTYPE_SKIP )) { +prepare_emulation_failure_exit (vcpu);return 0 ; }
include/uapi/linux/kvm.h +23 -0
@@ -279,6 +279,9 @@ struct kvm_xen_exit {/* Encounter unexpected vm-exit reason */ #define KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON 4 +/* Flags that describe what fields in emulation_failure hold valid data */ +#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) +/* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ struct kvm_run {/* in */
@@ -382,6 +385,25 @@ struct kvm_run {__u32 ndata;__u64 data[16 ]; } internal; +/* +* KVM_INTERNAL_ERROR_EMULATION +* +* "struct emulation_failure" is an overlay of "struct internal" +* that is used for the KVM_INTERNAL_ERROR_EMULATION sub-type of +* KVM_EXIT_INTERNAL_ERROR. Note, unlike other internal error +* sub-types, this struct is ABI! It also needs to be backwards +* compabile with "struct internal". Take special care that +* "ndata" is correct, that new fields are enumerated in "flags", +* and that each flag enumerates fields that are 64-bit aligned +* and sized (so that ndata+internal.data[] is valid/accurate). +*/ +struct { +__u32 suberror; +__u32 ndata; +__u64 flags; +__u8 insn_size; +__u8 insn_bytes[15 ]; + } emulation_failure;/* KVM_EXIT_OSI */ struct {__u64 gprs[32 ];
@@ -1078,6 +1100,7 @@ struct kvm_ppc_resize_hpt {#define KVM_CAP_DIRTY_LOG_RING 192#define KVM_CAP_X86_BUS_LOCK_EXIT 193#define KVM_CAP_PPC_DAWR1 194 +#define KVM_CAP_EXIT_ON_EMULATION_FAILURE 195#ifdef KVM_CAP_IRQ_ROUTING
tools/include/uapi/linux/kvm.h +23 -0
@@ -279,6 +279,9 @@ struct kvm_xen_exit {/* Encounter unexpected vm-exit reason */ #define KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON 4 +/* Flags that describe what fields in emulation_failure hold valid data */ +#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) +/* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ struct kvm_run {/* in */
@@ -382,6 +385,25 @@ struct kvm_run {__u32 ndata;__u64 data[16 ]; } internal; +/* +* KVM_INTERNAL_ERROR_EMULATION +* +* "struct emulation_failure" is an overlay of "struct internal" +* that is used for the KVM_INTERNAL_ERROR_EMULATION sub-type of +* KVM_EXIT_INTERNAL_ERROR. Note, unlike other internal error +* sub-types, this struct is ABI! It also needs to be backwards +* compabile with "struct internal". Take special care that +* "ndata" is correct, that new fields are enumerated in "flags", +* and that each flag enumerates fields that are 64-bit aligned +* and sized (so that ndata+internal.data[] is valid/accurate). +*/ +struct { +__u32 suberror; +__u32 ndata; +__u64 flags; +__u8 insn_size; +__u8 insn_bytes[15 ]; + } emulation_failure;/* KVM_EXIT_OSI */ struct {__u64 gprs[32 ];
@@ -1078,6 +1100,7 @@ struct kvm_ppc_resize_hpt {#define KVM_CAP_DIRTY_LOG_RING 192#define KVM_CAP_X86_BUS_LOCK_EXIT 193#define KVM_CAP_PPC_DAWR1 194 +#define KVM_CAP_EXIT_ON_EMULATION_FAILURE 195#ifdef KVM_CAP_IRQ_ROUTING
Signature · 1 line
2.31.1.498.g6c1eba8ee3d-goog
PATCH v2 2/2 selftests: kvm: Allows userspace to handle emulation errors.
This test exercises the feature KVM_CAP_EXIT_ON_EMULATION_FAILURE. When enabled, errors in the in-kernel instruction emulator are forwarded to userspace with the instruction bytes stored in the exit struct for KVM_EXIT_INTERNAL_ERROR. So, when the guest attempts to emulate an 'flds' instruction, which isn't able to be emulated in KVM, instead of failing, KVM sends the instruction to userspace to handle.
For this test to work properly the module parameter 'allow_smaller_maxphyaddr' has to be set.
--- tools/testing/selftests/kvm/.gitignore | 1 + tools/testing/selftests/kvm/Makefile | 1 + .../selftests/kvm/include/x86_64/processor.h | 3 + .../selftests/kvm/lib/x86_64/processor.c | 58 +++++ .../kvm/x86_64/emulator_error_test.c | 209 ++++++++++++++++++ 5 files changed, 272 insertions(+) create mode 100644 tools/testing/selftests/kvm/x86_64/emulator_error_test.c
tools/testing/selftests/kvm/.gitignore +1 -0
@@ -7,6 +7,7 @@ /x86_64/cr4_cpuid_sync_test /x86_64/debug_regs /x86_64/evmcs_test +/x86_64/emulator_error_test /x86_64/get_cpuid_test /x86_64/get_msr_index_features /x86_64/kvm_pv_test
tools/testing/selftests/kvm/Makefile +1 -0
@@ -41,6 +41,7 @@ LIBKVM_s390x = lib/s390x/processor.c lib/s390x/ucall.c lib/s390x/diag318_test_ha TEST_GEN_PROGS_x86_64 = x86_64/cr4_cpuid_sync_test TEST_GEN_PROGS_x86_64 += x86_64/get_msr_index_features TEST_GEN_PROGS_x86_64 += x86_64/evmcs_test +TEST_GEN_PROGS_x86_64 += x86_64/emulator_error_test TEST_GEN_PROGS_x86_64 += x86_64/get_cpuid_test TEST_GEN_PROGS_x86_64 += x86_64/hyperv_clock TEST_GEN_PROGS_x86_64 += x86_64/hyperv_cpuid
tools/testing/selftests/kvm/include/x86_64/processor.h +3 -0
@@ -394,6 +394,9 @@ void vcpu_init_descriptor_tables(struct kvm_vm *vm, uint32_t vcpuid);void vm_handle_exception (struct kvm_vm *vm,int vector,void (*handler)(struct ex_regs *)); +uint64_t vm_get_page_table_entry (struct kvm_vm *vm,uint64_t vaddr); +void vm_set_page_table_entry (struct kvm_vm *vm,uint64_t vaddr,uint64_t mask); +/* * set_cpuid() - overwrites a matching cpuid entry with the provided value. * matches based on ent->function && ent->index. returns true
tools/testing/selftests/kvm/lib/x86_64/processor.c +58 -0
64 lines not on this page · raw message
tools/testing/selftests/kvm/x86_64/emulator_error_test.c new file +209 -0
209 lines not on this page · raw message
Signature · 1 line
2.31.1.498.g6c1eba8ee3d-goog
Quoted 27 lines · Aaron Lewis
Add a fallback mechanism to the in-kernel instruction emulator that allows userspace the opportunity to process an instruction the emulator was unable to. When the in-kernel instruction emulator fails to process an instruction it will either inject a #UD into the guest or exit to userspace with exit reason
KVM_INTERNAL_ERROR. This is because it does not know how to proceed in an appropriate manner. This feature lets userspace get involved to see if it can figure out a better path forward.Signed-off-by Aaron Lewis <[email protected]>--- Documentation/virt/kvm/api.rst | 19 ++++++++++++++++++ arch/x86/include/asm/kvm_host.h | 6 ++++++ arch/x86/kvm/x86.c | 35 +++++++++++++++++++++++++++++---- include/uapi/linux/kvm.h | 23 ++++++++++++++++++++++ tools/include/uapi/linux/kvm.h | 23 ++++++++++++++++++++++ 5 files changed, 102 insertions(+), 4 deletions(-)Documentation/virt/kvm/api.rst +1
-0@@ -6233,6 +6233,25 @@ KVM_RUN_BUS_LOCK flag is used to distinguish between them. This capability can be used to check / enable 2nd DAWR feature provided by POWER10 processor. +7.24 KVM_CAP_EXIT_ON_EMULATION_FAILURE
Documentation/virt/kvm/api.rst KVM_RUN_BUS_LOCK flag is used to distinguish between them. · line 6246
Quoted 10 lines
Documentation/virt/kvm/api.rst +10
-0+-------------------------------------- + +:Architectures: x86 +:Parameters: args[0] whether the feature should be enabled or not + +With this capability enabled, the in-kernel instruction emulator packs the exit +struct of KVM_INTERNAL_ERROR with the instruction length and instruction bytes +when an error occurs while emulating an instruction. This allows userspace to +then take a look at the instruction and see if it is able to handle it more +gracefully than the in-kernel emulator.
It will do so without KVM_CAP_EXIT_ON_EMULATION_FAILURE enabled in this set of changes.
That is, the payload can be present irrespective of the capability being enabled.
Quoted 51 lines
Documentation/virt/kvm/api.rst +7
-0+When this capability is enabled use the emulation_failure struct instead of the +internal struct for the exit struct. They have the same layout, but the +emulation_failure struct matches the content better. + +It is noted that KVM still exits on certain types (skip) even if this capability +is not enabled, and KVM will never exit if VMware #GP emulation fails. + 8. Other capabilities. ======================arch/x86/include/asm/kvm_host.h +6
-0@@ -1049,6 +1049,12 @@ struct kvm_arch {bool exception_payload_enabled;bool bus_lock_detection_enabled; +/* +* If exit_on_emulation_error is set, and the in-kernel instruction +* emulator fails to emulate an instruction, allow userspace +* the opportunity to look at it. +*/ +bool exit_on_emulation_error;/* Deflect RDMSR and WRMSR to user space when they trigger a #GP */ u32 user_space_msr_mask;arch/x86/kvm/x86.c +5
-0@@ -3771,6 +3771,7 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)case KVM_CAP_X86_USER_SPACE_MSR :case KVM_CAP_X86_MSR_FILTER :case KVM_CAP_ENFORCE_PV_FEATURE_CPUID : +case KVM_CAP_EXIT_ON_EMULATION_FAILURE : r =1 ;break ;#ifdef CONFIG_KVM_XEN @@ -5357,6 +5358,10 @@ int kvm_vm_ioctl_enable_cap(struct kvm *kvm, kvm->arch.bus_lock_detection_enabled = true; r =0 ;break ; +case KVM_CAP_EXIT_ON_EMULATION_FAILURE : + kvm->arch.exit_on_emulation_error = cap->args[0 ]; + r =0 ; +break ;default : r = -EINVAL ;break ;
arch/x86/kvm/x86.c void kvm_inject_realmode_interrupt(struct kvm_vcpu *vcpu, int irq, int inc_eip) · line 7133
Quoted 10 lines
arch/x86/kvm/x86.c +7
-0@@ -7119,8 +7124,31 @@ void kvm_inject_realmode_interrupt(struct kvm_vcpu *vcpu, int irq, int inc_eip) }EXPORT_SYMBOL_GPL (kvm_inject_realmode_interrupt); +static void prepare_emulation_failure_exit (struct kvm_vcpu *vcpu) +{ +struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt; +u64 insn_size = ctxt->fetch.end - ctxt->fetch.data; +struct kvm_run *run = vcpu->run; + +
Unnecessary extra blank line here...
arch/x86/kvm/x86.c void kvm_inject_realmode_interrupt(struct kvm_vcpu *vcpu, int irq, int inc_eip) · line 7137
Quoted 4 lines
arch/x86/kvm/x86.c +4
-0+ run->exit_reason =KVM_EXIT_INTERNAL_ERROR ; + run->emulation_failure.suberror =KVM_INTERNAL_ERROR_EMULATION ; + run->emulation_failure.ndata =0 ; + run->emulation_failure.flags =0 ;
...that I would move here, but your call :-)
arch/x86/kvm/x86.c void kvm_inject_realmode_interrupt(struct kvm_vcpu *vcpu, int irq, int inc_eip) · line 7144
Quoted 7 lines
arch/x86/kvm/x86.c +7
-0+if (insn_size) { + run->emulation_failure.ndata =3 ; + run->emulation_failure.flags |= +KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES ; + run->emulation_failure.insn_size = insn_size; +memcpy (run->emulation_failure.insn_bytes, + ctxt->fetch.data,sizeof (ctxt->fetch.data));
We're relying on the fact that insn_bytes is at least as large as fetch.data, which is fine, but worth an assertion?
"Leaking" irrelevant bytes here also seems bad, but I can't immediately see a problem as a result.
Quoted 77 lines
arch/x86/kvm/x86.c +8
-4+ } +} +static int handle_emulation_failure (struct kvm_vcpu *vcpu,int emulation_type) { +struct kvm *kvm = vcpu->kvm; + ++vcpu->stat.insn_emulation_fail;trace_kvm_emulate_insn_failed (vcpu);@@ -7129,10 +7157,9 @@ static int handle_emulation_failure(struct kvm_vcpu *vcpu, int emulation_type)return 1 ; } -if (emulation_type &EMULTYPE_SKIP ) { - vcpu->run->exit_reason =KVM_EXIT_INTERNAL_ERROR ; - vcpu->run->internal.suberror =KVM_INTERNAL_ERROR_EMULATION ; - vcpu->run->internal.ndata =0 ; +if (kvm->arch.exit_on_emulation_error || + (emulation_type &EMULTYPE_SKIP )) { +prepare_emulation_failure_exit (vcpu);return 0 ; }include/uapi/linux/kvm.h +23
-0@@ -279,6 +279,9 @@ struct kvm_xen_exit {/* Encounter unexpected vm-exit reason */ #define KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON 4 +/* Flags that describe what fields in emulation_failure hold valid data */ +#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) +/* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ struct kvm_run {/* in */ @@ -382,6 +385,25 @@ struct kvm_run {__u32 ndata;__u64 data[16 ]; } internal; +/* +* KVM_INTERNAL_ERROR_EMULATION +* +* "struct emulation_failure" is an overlay of "struct internal" +* that is used for the KVM_INTERNAL_ERROR_EMULATION sub-type of +* KVM_EXIT_INTERNAL_ERROR. Note, unlike other internal error +* sub-types, this struct is ABI! It also needs to be backwards +* compabile with "struct internal". Take special care that +* "ndata" is correct, that new fields are enumerated in "flags", +* and that each flag enumerates fields that are 64-bit aligned +* and sized (so that ndata+internal.data[] is valid/accurate). +*/ +struct { +__u32 suberror; +__u32 ndata; +__u64 flags; +__u8 insn_size; +__u8 insn_bytes[15 ]; + } emulation_failure;/* KVM_EXIT_OSI */ struct {__u64 gprs[32 ];@@ -1078,6 +1100,7 @@ struct kvm_ppc_resize_hpt {#define KVM_CAP_DIRTY_LOG_RING 192#define KVM_CAP_X86_BUS_LOCK_EXIT 193#define KVM_CAP_PPC_DAWR1 194 +#define KVM_CAP_EXIT_ON_EMULATION_FAILURE 195#ifdef KVM_CAP_IRQ_ROUTING tools/include/uapi/linux/kvm.h +0
-0
tools/include/uapi/linux/kvm.h struct kvm_xen_exit { · line 282
Quoted 4 lines
tools/include/uapi/linux/kvm.h +1
-0@@ -279,6 +279,9 @@ struct kvm_xen_exit {/* Encounter unexpected vm-exit reason */ #define KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON 4 +/* Flags that describe what fields in emulation_failure hold valid data */
Extra space at the end.
Quoted 41 lines
tools/include/uapi/linux/kvm.h +22
-0+#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) +/* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ struct kvm_run {/* in */ @@ -382,6 +385,25 @@ struct kvm_run {__u32 ndata;__u64 data[16 ]; } internal; +/* +* KVM_INTERNAL_ERROR_EMULATION +* +* "struct emulation_failure" is an overlay of "struct internal" +* that is used for the KVM_INTERNAL_ERROR_EMULATION sub-type of +* KVM_EXIT_INTERNAL_ERROR. Note, unlike other internal error +* sub-types, this struct is ABI! It also needs to be backwards +* compabile with "struct internal". Take special care that +* "ndata" is correct, that new fields are enumerated in "flags", +* and that each flag enumerates fields that are 64-bit aligned +* and sized (so that ndata+internal.data[] is valid/accurate). +*/ +struct { +__u32 suberror; +__u32 ndata; +__u64 flags; +__u8 insn_size; +__u8 insn_bytes[15 ]; + } emulation_failure;/* KVM_EXIT_OSI */ struct {__u64 gprs[32 ];@@ -1078,6 +1100,7 @@ struct kvm_ppc_resize_hpt {#define KVM_CAP_DIRTY_LOG_RING 192#define KVM_CAP_X86_BUS_LOCK_EXIT 193#define KVM_CAP_PPC_DAWR1 194 +#define KVM_CAP_EXIT_ON_EMULATION_FAILURE 195#ifdef KVM_CAP_IRQ_ROUTING Signature · 1 line
2.31.1.498.g6c1eba8ee3d-goog
dme.
Signature · 1 line
Everyone I know, goes away in the end.
Quoted 8 lines
arch/x86/kvm/x86.c void kvm_inject_realmode_interrupt(struct kvm_vcpu *vcpu, int irq, int inc_eip) · line 7144
Quoted 7 lines
arch/x86/kvm/x86.c +7
-0+if (insn_size) { + run->emulation_failure.ndata =3 ; + run->emulation_failure.flags |= +KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES ; + run->emulation_failure.insn_size = insn_size; +memcpy (run->emulation_failure.insn_bytes, + ctxt->fetch.data,sizeof (ctxt->fetch.data));
Quoted 6 lines
We're relying on the fact that
insn_bytesis at least as large as fetch.data, which is fine, but worth an assertion?"Leaking" irrelevant bytes here also seems bad, but I can't immediately see a problem as a result.
I don't think this is a problem because the instruction bytes stream has irrelevant bytes in it anyway. In the test attached I verify that it receives an flds instruction in userspace that was emulated in the guest. In the stream that comes through insn_size is set to 15 and the instruction is only 2 bytes long, so the stream has irrelevant bytes in it as far as this instruction is concerned.
Quoted 3 lines
Quoted 3 lines
arch/x86/kvm/x86.c +3
-0+ } +} +
Quoted 15 lines · Aaron Lewis
Quoted 8 lines
arch/x86/kvm/x86.c void kvm_inject_realmode_interrupt(struct kvm_vcpu *vcpu, int irq, int inc_eip) · line 7144
Quoted 7 lines
arch/x86/kvm/x86.c +7
-0+if (insn_size) { + run->emulation_failure.ndata =3 ; + run->emulation_failure.flags |= +KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES ; + run->emulation_failure.insn_size = insn_size; +memcpy (run->emulation_failure.insn_bytes, + ctxt->fetch.data,sizeof (ctxt->fetch.data));Quoted 6 lines
We're relying on the fact that
insn_bytesis at least as large as fetch.data, which is fine, but worth an assertion?"Leaking" irrelevant bytes here also seems bad, but I can't immediately see a problem as a result.
Quoted 6 lines
I don't think this is a problem because the instruction bytes stream has irrelevant bytes in it anyway. In the test attached I verify that it receives an flds instruction in userspace that was emulated in the guest. In the stream that comes through
insn_sizeis set to 15 and the instruction is only 2 bytes long, so the stream has irrelevant bytes in it as far as this instruction is concerned.
As an experiment I added[1] reporting of the exit reason using flag 2. On emulation failure (without the instruction bytes flag enabled), one run of QEMU reported:
Quoted 6 lines
KVM internal error. Suberror: 1
extra data[0]: 2
extra data[1]: 4
extra data[2]: 0
extra data[3]: 31
emulation failure
data[1] and data[2] are not indicated as valid, but it seems unfortunate that I got (not really random) garbage there.
Admittedly, with only your patches applied ndata will never skip past any bytes, as there is only one flag. As soon as I add another, is it my job to zero out those unused bytes? Maybe we should be clearing all of the payload at the top of prepare_emulation_failure_exit().
Footnotes:
[1] https://disaster-area.hh.sledj.net/tmp/dme-581090/
dme.
Signature · 1 line
Music has magic, it's good clear syncopation.
Quoted 19 lines
Quoted 7 lines
I don't think this is a problem because the instruction bytes stream has irrelevant bytes in it anyway. In the test attached I verify that it receives an flds instruction in userspace that was emulated in the guest. In the stream that comes through
insn_sizeis set to 15 and the instruction is only 2 bytes long, so the stream has irrelevant bytes in it as far as this instruction is concerned.As an experiment I added[1] reporting of the exit reason using flag 2. On emulation failure (without the instruction bytes flag enabled), one run of QEMU reported:
Quoted 6 lines
KVM internal error. Suberror: 1
extra data[0]: 2
extra data[1]: 4
extra data[2]: 0
extra data[3]: 31
emulation failure
Quoted 8 lines
data[1] and data[2] are not indicated as valid, but it seems unfortunate that I got (not really random) garbage there.
Admittedly, with only your patches applied ndata will never skip past any bytes, as there is only one flag. As soon as I add another, is it my job to zero out those unused bytes? Maybe we should be clearing all of the payload at the top of
prepare_emulation_failure_exit().
Clearing the bytes at the top of prepare_emulation_failure_exit() sounds good to me. That will keep the data more deterministic. Though, I will say that I don't think that is required. If the first flag isn't set the data shouldn't be read, no?
Quoted 6 lines
Footnotes:
[1] https://disaster-area.hh.sledj.net/tmp/dme-581090/dme.
--
Music has magic, it's good clear syncopation.
Quoted 28 lines · Aaron Lewis
Quoted 19 lines
Quoted 7 lines
I don't think this is a problem because the instruction bytes stream has irrelevant bytes in it anyway. In the test attached I verify that it receives an flds instruction in userspace that was emulated in the guest. In the stream that comes through
insn_sizeis set to 15 and the instruction is only 2 bytes long, so the stream has irrelevant bytes in it as far as this instruction is concerned.As an experiment I added[1] reporting of the exit reason using flag 2. On emulation failure (without the instruction bytes flag enabled), one run of QEMU reported:
Quoted 6 lines
KVM internal error. Suberror: 1
extra data[0]: 2
extra data[1]: 4
extra data[2]: 0
extra data[3]: 31
emulation failureQuoted 8 lines
data[1] and data[2] are not indicated as valid, but it seems unfortunate that I got (not really random) garbage there.
Admittedly, with only your patches applied ndata will never skip past any bytes, as there is only one flag. As soon as I add another, is it my job to zero out those unused bytes? Maybe we should be clearing all of the payload at the top of
prepare_emulation_failure_exit().
Quoted 4 lines
Clearing the bytes at the top of
prepare_emulation_failure_exit()sounds good to me. That will keep the data more deterministic. Though, I will say that I don't think that is required. If the first flag isn't set the data shouldn't be read, no?
Agreed. As Jim indicated in his other reply, there should be no new data leaked by not zeroing the bytes.
For now at least, this is not a performance critical path, so clearing the payload doesn't seem too onerous.
Quoted 6 lines
Quoted 6 lines
Footnotes:
[1] https://disaster-area.hh.sledj.net/tmp/dme-581090/dme.
--
Music has magic, it's good clear syncopation.
dme.
Signature · 1 line
Don't you know you're never going to get to France.
Sean Christopherson
maintainerQuoted 35 lines · David Edmondson
Quoted 28 lines · Aaron Lewis
Quoted 19 lines
Quoted 7 lines
I don't think this is a problem because the instruction bytes stream has irrelevant bytes in it anyway. In the test attached I verify that it receives an flds instruction in userspace that was emulated in the guest. In the stream that comes through
insn_sizeis set to 15 and the instruction is only 2 bytes long, so the stream has irrelevant bytes in it as far as this instruction is concerned.As an experiment I added[1] reporting of the exit reason using flag 2. On emulation failure (without the instruction bytes flag enabled), one run of QEMU reported:
Quoted 6 lines
KVM internal error. Suberror: 1
extra data[0]: 2
extra data[1]: 4
extra data[2]: 0
extra data[3]: 31
emulation failureQuoted 8 lines
data[1] and data[2] are not indicated as valid, but it seems unfortunate that I got (not really random) garbage there.
Admittedly, with only your patches applied ndata will never skip past any bytes, as there is only one flag. As soon as I add another, is it my job to zero out those unused bytes? Maybe we should be clearing all of the payload at the top of
prepare_emulation_failure_exit().Quoted 4 lines
Clearing the bytes at the top of
prepare_emulation_failure_exit()sounds good to me. That will keep the data more deterministic. Though, I will say that I don't think that is required. If the first flag isn't set the data shouldn't be read, no?
Quoted 5 lines
Agreed. As Jim indicated in his other reply, there should be no new data leaked by not zeroing the bytes.
For now at least, this is not a performance critical path, so clearing the payload doesn't seem too onerous.
I feel quite strongly that KVM should _not_ touch the unused bytes. As Jim pointed out, a stream of 0x0 0x0 0x0 ... is not benign, it will decode to one or more ADD instructions. Arguably 0x90, 0xcc, or an undending stream of prefixes would be more appropriate so that it's less likely for userspace to decode a bogus instruction.
I don't see any reason why unused insn bytes should be treated any differently than unused mmio.data[], or unused internal.data[], etc...
IMO, the better option is to do nothing and let userspace initialize vcpu->run before KVM_RUN if they want to avoid consuming stale data.
Quoted 36 lines · Sean Christopherson
Quoted 35 lines · David Edmondson
Quoted 32 lines · Aaron Lewis
Quoted 19 lines
Quoted 7 lines
I don't think this is a problem because the instruction bytes stream has irrelevant bytes in it anyway. In the test attached I verify that it receives an flds instruction in userspace that was emulated in the guest. In the stream that comes through
insn_sizeis set to 15 and the instruction is only 2 bytes long, so the stream has irrelevant bytes in it as far as this instruction is concerned.As an experiment I added[1] reporting of the exit reason using flag 2. On emulation failure (without the instruction bytes flag enabled), one run of QEMU reported:
Quoted 6 lines
KVM internal error. Suberror: 1
extra data[0]: 2
extra data[1]: 4
extra data[2]: 0
extra data[3]: 31
emulation failureQuoted 8 lines
data[1] and data[2] are not indicated as valid, but it seems unfortunate that I got (not really random) garbage there.
Admittedly, with only your patches applied ndata will never skip past any bytes, as there is only one flag. As soon as I add another, is it my job to zero out those unused bytes? Maybe we should be clearing all of the payload at the top of
prepare_emulation_failure_exit().Clearing the bytes at the top of
prepare_emulation_failure_exit()sounds good to me. That will keep the data more deterministic. Though, I will say that I don't think that is required. If the first flag isn't set the data shouldn't be read, no?
Quoted 7 lines
Quoted 5 lines
Agreed. As Jim indicated in his other reply, there should be no new data leaked by not zeroing the bytes.
For now at least, this is not a performance critical path, so clearing the payload doesn't seem too onerous.
I feel quite strongly that KVM should _not_ touch the unused bytes.
I'm fine with that, but...
Quoted 4 lines
As Jim pointed out, a stream of
0x00x00x0... is not benign, it will decode to one or more ADD instructions. Arguably0x90,0xcc, or an undending stream of prefixes would be more appropriate so that it's less likely for userspace to decode a bogus instruction.
...I don't understand this position. If the user-level instruction decoder starts interpreting bytes that the kernel did *not* indicate as valid (by setting insn_size to include them), it's broken.
Quoted 5 lines
I don't see any reason why unused insn bytes should be treated any differently than unused mmio.data[], or unused internal.data[], etc...
IMO, the better option is to do nothing and let userspace initialize
vcpu->runbeforeKVM_RUNif they want to avoid consuming stale data.
dme.
Signature · 1 line
I've still got sand in my shoes.
Sean Christopherson
maintainerQuoted 13 lines · David Edmondson
Quoted 8 lines · Sean Christopherson
Quoted 5 lines · David Edmondson
Agreed. As Jim indicated in his other reply, there should be no new data leaked by not zeroing the bytes.
For now at least, this is not a performance critical path, so clearing the payload doesn't seem too onerous.
I feel quite strongly that KVM should _not_ touch the unused bytes.
I'm fine with that, but...
Quoted 8 lines
Quoted 4 lines
As Jim pointed out, a stream of
0x00x00x0... is not benign, it will decode to one or more ADD instructions. Arguably0x90,0xcc, or an undending stream of prefixes would be more appropriate so that it's less likely for userspace to decode a bogus instruction....I don't understand this position. If the user-level instruction decoder starts interpreting bytes that the kernel did *not* indicate as valid (by setting
insn_sizeto include them), it's broken.
Yes, so what's the point of clearing the unused bytes? Doing so won't magically fix a broken userspace. That's why I argue that 0x90 or 0xcc would be more appropriate; there's at least a non-zero chance that it will help userspace avoid doing something completely broken.
On the other hand, userspace can guard against a broken _KVM_ by initializing vcpu->run with a known pattern and logging if KVM exits to userspace with seemingly bogus data. Crushing the unused bytes to zero defeats userspace's sanity check, e.g. if the actual memcpy() of the instruction bytes copies the wrong number of bytes, then userspace's magic pattern will be lost and debugging the KVM bug will be that much harder.
This is very much not a theoretical problem, I have debugged two separate KVM bugs in the last few months where KVM completely failed to set vcpu->run->exit_reason before exiting to userspace. The exit_reason is a bit of a special case because it's disturbingly easy for KVM to get confused over return values and unintentionally exit to userspace, but it's not a big stretch to imagine a bug where KVM provides incomplete data.
Quoted 19 lines · Sean Christopherson
Quoted 18 lines · David Edmondson
Quoted 8 lines · Sean Christopherson
Quoted 5 lines · David Edmondson
Agreed. As Jim indicated in his other reply, there should be no new data leaked by not zeroing the bytes.
For now at least, this is not a performance critical path, so clearing the payload doesn't seem too onerous.
I feel quite strongly that KVM should _not_ touch the unused bytes.
I'm fine with that, but...
Quoted 4 lines
As Jim pointed out, a stream of
0x00x00x0... is not benign, it will decode to one or more ADD instructions. Arguably0x90,0xcc, or an undending stream of prefixes would be more appropriate so that it's less likely for userspace to decode a bogus instruction.
Quoted 5 lines
Quoted 3 lines
...I don't understand this position. If the user-level instruction decoder starts interpreting bytes that the kernel did *not* indicate as valid (by setting
insn_sizeto include them), it's broken.Yes, so what's the point of clearing the unused bytes?
Given that it doesn't prevent any known leakage, it's purely aesthetic, which is why I'm happy not to bother.
Quoted 4 lines
Doing so won't magically fix a broken userspace. That's why I argue that
0x90or0xccwould be more appropriate; there's at least a non-zero chance that it will help userspace avoid doing something completely broken.
Perhaps an invalid instruction would be more useful in this respect, but INT03 fills a similar purpose.
Quoted 7 lines
On the other hand, userspace can guard against a broken _KVM_ by initializing
vcpu->runwith a known pattern and logging if KVM exits to userspace with seemingly bogus data. Crushing the unused bytes to zero defeats userspace's sanity check, e.g. if the actualmemcpy()of the instruction bytes copies the wrong number of bytes, then userspace's magic pattern will be lost and debugging the KVM bug will be that much harder.
Quoted 6 lines
This is very much not a theoretical problem, I have debugged two separate KVM bugs in the last few months where KVM completely failed to set
vcpu->run->exit_reasonbefore exiting to userspace. Theexit_reasonis a bit of a special case because it's disturbingly easy for KVM to get confused over return values and unintentionally exit to userspace, but it's not a big stretch to imagine a bug where KVM provides incomplete data.
Understood.
So is the conclusion that KVM should copy only insn_size bytes rather than the full 15?
dme.
Signature · 1 line
But they'll laugh at you in Jackson, and I'll be dancin' on a Pony Keg.
Quoted 53 lines · David Edmondson
Quoted 19 lines · Sean Christopherson
Quoted 18 lines · David Edmondson
Quoted 8 lines · Sean Christopherson
Quoted 5 lines · David Edmondson
Agreed. As Jim indicated in his other reply, there should be no new data leaked by not zeroing the bytes.
For now at least, this is not a performance critical path, so clearing the payload doesn't seem too onerous.
I feel quite strongly that KVM should _not_ touch the unused bytes.
I'm fine with that, but...
Quoted 4 lines
As Jim pointed out, a stream of
0x00x00x0... is not benign, it will decode to one or more ADD instructions. Arguably0x90,0xcc, or an undending stream of prefixes would be more appropriate so that it's less likely for userspace to decode a bogus instruction.Quoted 5 lines
Quoted 3 lines
...I don't understand this position. If the user-level instruction decoder starts interpreting bytes that the kernel did *not* indicate as valid (by setting
insn_sizeto include them), it's broken.Yes, so what's the point of clearing the unused bytes?
Given that it doesn't prevent any known leakage, it's purely aesthetic, which is why I'm happy not to bother.
Quoted 4 lines
Doing so won't magically fix a broken userspace. That's why I argue that
0x90or0xccwould be more appropriate; there's at least a non-zero chance that it will help userspace avoid doing something completely broken.Perhaps an invalid instruction would be more useful in this respect, but INT03 fills a similar purpose.
Quoted 7 lines
On the other hand, userspace can guard against a broken _KVM_ by initializing
vcpu->runwith a known pattern and logging if KVM exits to userspace with seemingly bogus data. Crushing the unused bytes to zero defeats userspace's sanity check, e.g. if the actualmemcpy()of the instruction bytes copies the wrong number of bytes, then userspace's magic pattern will be lost and debugging the KVM bug will be that much harder.Quoted 6 lines
This is very much not a theoretical problem, I have debugged two separate KVM bugs in the last few months where KVM completely failed to set
vcpu->run->exit_reasonbefore exiting to userspace. Theexit_reasonis a bit of a special case because it's disturbingly easy for KVM to get confused over return values and unintentionally exit to userspace, but it's not a big stretch to imagine a bug where KVM provides incomplete data.
Quoted 4 lines
Understood.
So is the conclusion that KVM should copy only
insn_sizebytes rather than the full 15?
Insn_size should almost always be 15. It will only be less when the emulator hits a page crossing before fetching 15 bytes and it can't fetch from the second page.
Quoted 3 lines
dme.
--
But they'll laugh at you in Jackson, and I'll be dancin' on a Pony Keg.
Quoted 55 lines · Jim Mattson
Quoted 53 lines · David Edmondson
Quoted 19 lines · Sean Christopherson
Quoted 18 lines · David Edmondson
Quoted 8 lines · Sean Christopherson
Quoted 5 lines · David Edmondson
Agreed. As Jim indicated in his other reply, there should be no new data leaked by not zeroing the bytes.
For now at least, this is not a performance critical path, so clearing the payload doesn't seem too onerous.
I feel quite strongly that KVM should _not_ touch the unused bytes.
I'm fine with that, but...
Quoted 4 lines
As Jim pointed out, a stream of
0x00x00x0... is not benign, it will decode to one or more ADD instructions. Arguably0x90,0xcc, or an undending stream of prefixes would be more appropriate so that it's less likely for userspace to decode a bogus instruction.Quoted 5 lines
Quoted 3 lines
...I don't understand this position. If the user-level instruction decoder starts interpreting bytes that the kernel did *not* indicate as valid (by setting
insn_sizeto include them), it's broken.Yes, so what's the point of clearing the unused bytes?
Given that it doesn't prevent any known leakage, it's purely aesthetic, which is why I'm happy not to bother.
Quoted 4 lines
Doing so won't magically fix a broken userspace. That's why I argue that
0x90or0xccwould be more appropriate; there's at least a non-zero chance that it will help userspace avoid doing something completely broken.Perhaps an invalid instruction would be more useful in this respect, but INT03 fills a similar purpose.
Quoted 13 lines
On the other hand, userspace can guard against a broken _KVM_ by initializing
vcpu->runwith a known pattern and logging if KVM exits to userspace with seemingly bogus data. Crushing the unused bytes to zero defeats userspace's sanity check, e.g. if the actualmemcpy()of the instruction bytes copies the wrong number of bytes, then userspace's magic pattern will be lost and debugging the KVM bug will be that much harder.This is very much not a theoretical problem, I have debugged two separate KVM bugs in the last few months where KVM completely failed to set
vcpu->run->exit_reasonbefore exiting to userspace. Theexit_reasonis a bit of a special case because it's disturbingly easy for KVM to get confused over return values and unintentionally exit to userspace, but it's not a big stretch to imagine a bug where KVM provides incomplete data.
Quoted 8 lines
Quoted 4 lines
Understood.
So is the conclusion that KVM should copy only
insn_sizebytes rather than the full 15?
Insn_sizeshould almost always be 15. It will only be less when the emulator hits a page crossing before fetching 15 bytes and it can't fetch from the second page.
Oh, or if the CS limit is reached. (cf. AMD's APM, volume 2, section 15.8.4: Nested and intercepted #PF).
Quoted 3 lines
Quoted 3 lines
dme.
--
But they'll laugh at you in Jackson, and I'll be dancin' on a Pony Keg.
Quoted 10 lines
Quoted 7 lines
Quoted 3 lines
So is the conclusion that KVM should copy only
insn_sizebytes rather than the full 15?
Insn_sizeshould almost always be 15. It will only be less when the emulator hits a page crossing before fetching 15 bytes and it can't fetch from the second page.Oh, or if the CS limit is reached. (cf. AMD's APM, volume 2, section 15.8.4: Nested and intercepted #PF).
To sum this up as I understand it. I'm _not_ going to clear 'run->internal.data' to zero. I'll leave it to userspace to clear vcpu->run. I'll copy over 'insn_size' bytes rather than 'sizeof(ctxt->fetch.data)' bytes to 'run->emulation_failure.insn_bytes', and if 'insn_size' < 15, I'll stamp the remaining bytes with 0x90.
Let me know if I missed anything.
Quoted 2 lines
Quoted 11 lines · Aaron Lewis
Add a fallback mechanism to the in-kernel instruction emulator that allows userspace the opportunity to process an instruction the emulator was unable to. When the in-kernel instruction emulator fails to process an instruction it will either inject a #UD into the guest or exit to userspace with exit reason
KVM_INTERNAL_ERROR. This is because it does not know how to proceed in an appropriate manner. This feature lets userspace get involved to see if it can figure out a better path forward.Signed-off-by Aaron Lewis <[email protected]>
The instruction bytes are a good start, but in many cases, they aren't sufficient context to decode the next instruction. Should we eagerly provide that information in this exit, or should we just let userspace gather it via subsequent ioctls if necessary?
Quoted 12 lines · Jim Mattson
Quoted 10 lines · Aaron Lewis
Add a fallback mechanism to the in-kernel instruction emulator that allows userspace the opportunity to process an instruction the emulator was unable to. When the in-kernel instruction emulator fails to process an instruction it will either inject a #UD into the guest or exit to userspace with exit reason
KVM_INTERNAL_ERROR. This is because it does not know how to proceed in an appropriate manner. This feature lets userspace get involved to see if it can figure out a better path forward.
Quoted 6 lines
Quoted 1 line
Signed-off-by Aaron Lewis <[email protected]>The instruction bytes are a good start, but in many cases, they aren't sufficient context to decode the next instruction. Should we eagerly provide that information in this exit, or should we just let userspace gather it via subsequent ioctls if necessary?
Why is there a concern for the next instruction? This patch is about userspace helping with the current instruction being emulated. I'm not sure why we would be concerned about the next one.
Quoted 16 lines · Aaron Lewis
Quoted 14 lines · Jim Mattson
Quoted 11 lines · Aaron Lewis
Add a fallback mechanism to the in-kernel instruction emulator that allows userspace the opportunity to process an instruction the emulator was unable to. When the in-kernel instruction emulator fails to process an instruction it will either inject a #UD into the guest or exit to userspace with exit reason
KVM_INTERNAL_ERROR. This is because it does not know how to proceed in an appropriate manner. This feature lets userspace get involved to see if it can figure out a better path forward.Signed-off-by Aaron Lewis <[email protected]>
Quoted 8 lines
Quoted 4 lines
The instruction bytes are a good start, but in many cases, they aren't sufficient context to decode the next instruction. Should we eagerly provide that information in this exit, or should we just let userspace gather it via subsequent ioctls if necessary?
Why is there a concern for the next instruction? This patch is about userspace helping with the current instruction being emulated. I'm not sure why we would be concerned about the next one.
Sorry; I should have said, "The instruction bytes are a good start, but in many cases, they aren't sufficient context to decode *the* instruction." For example, suppose the first byte is 0x48. Without more context, we don't know if this is a DEC or a REX prefix.