KVM: Add a kvm_run flag to signal need for completion
Subsystems (9)
Sean Christopherson
maintainercover KVM: Add a kvm_run flag to signal need for completion
Add a flag to kvm_run, KVM_RUN_NEEDS_COMPLETION, that is set by KVM to inform userspace that KVM_RUN needs to be re-invoked prior to state save/restore. The current approach of relying on KVM developers to update documentation, and on VMM developers to read said documentation, is brittle and error prone.
Note, this series is *very* lightly tested (borderline RFC).
Sean Christopherson (5):
KVM: x86: Document that KVM_EXIT_HYPERCALL requires completion
KVM: Clear vcpu->run->flags at start of KVM_RUN for all architectures
KVM: Add a common kvm_run flag to communicate an exit needs completion
KVM: selftests: Provide separate helper for KVM_RUN with
immediate_exit
KVM: selftests: Rely on KVM_RUN_NEEDS_COMPLETION to complete userspace
exits
Documentation/virt/kvm/api.rst | 77 ++++++++++++------- arch/arm64/kvm/arm.c | 1 - arch/arm64/kvm/handle_exit.c | 2 +- arch/powerpc/kvm/book3s_emulate.c | 1 + arch/powerpc/kvm/book3s_hv.c | 5 +- arch/powerpc/kvm/book3s_pr.c | 2 + arch/powerpc/kvm/booke.c | 1 + arch/x86/include/uapi/asm/kvm.h | 7 +- arch/x86/kvm/x86.c | 3 +- include/uapi/linux/kvm.h | 3 + .../testing/selftests/kvm/include/kvm_util.h | 13 +++- tools/testing/selftests/kvm/lib/kvm_util.c | 6 +- .../testing/selftests/kvm/lib/ucall_common.c | 2 +- .../testing/selftests/kvm/lib/x86/processor.c | 8 +- .../kvm/x86/nested_exceptions_test.c | 3 +- .../kvm/x86/triple_fault_event_test.c | 3 +- virt/kvm/kvm_main.c | 4 + 17 files changed, 90 insertions(+), 51 deletions(-)
base-commit: 10b2c8a67c4b8ec15f9d07d177f63b563418e948
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
Sean Christopherson
maintainerPATCH 1/5 KVM: x86: Document that KVM_EXIT_HYPERCALL requires completion
Update KVM's documentation to call out that KVM_EXIT_HYPERCALL requires userspace to do KVM_RUN before state save/restore, so that KVM can skip the hypercall instruction, otherwise resuming the vCPU after restore will restart the instruction and potentially lead to a spurious MAP_GPA_RANGE.
--- Documentation/virt/kvm/api.rst | 39 +++++++++++++++++----------------- 1 file changed, 20 insertions(+), 19 deletions(-)
Documentation/virt/kvm/api.rst +20 -19
@@ -6615,13 +6615,29 @@ The 'data' member contains, in its first 'len' bytes, the value as it would appear if the VCPU performed a load or store of the appropriate width directly to the byte array. +It is strongly recommended that userspace use ``KVM_EXIT_IO`` (x86) or +``KVM_EXIT_MMIO`` (all except s390) to implement functionality that +requires a guest to interact with host userspace. + +.. note:: KVM_EXIT_IO is significantly faster than KVM_EXIT_MMIO. + + /* KVM_EXIT_HYPERCALL */ + struct { + __u64 nr; + __u64 args[6]; + __u64 ret; + __u64 flags; + } hypercall; + + .. note:: For KVM_EXIT_IO, KVM_EXIT_MMIO, KVM_EXIT_OSI, KVM_EXIT_PAPR, KVM_EXIT_XEN, - KVM_EXIT_EPR, KVM_EXIT_X86_RDMSR and KVM_EXIT_X86_WRMSR the corresponding - operations are complete (and guest state is consistent) only after userspace - has re-entered the kernel with KVM_RUN. The kernel side will first finish - incomplete operations and then check for pending signals. + KVM_EXIT_EPR, KVM_EXIT_X86_RDMSR, KVM_EXIT_X86_WRMSR, and KVM_EXIT_HYPERCALL + the corresponding operations are complete (and guest state is consistent) + only after userspace has re-entered the kernel with KVM_RUN. The kernel + side will first finish incomplete operations and then check for pending + signals. The pending state of the operation is not preserved in state which is visible to userspace, thus userspace should ensure that the operation is
@@ -6632,21 +6648,6 @@ to the byte array. :: - /* KVM_EXIT_HYPERCALL */ - struct { - __u64 nr; - __u64 args[6]; - __u64 ret; - __u64 flags; - } hypercall; - - -It is strongly recommended that userspace use ``KVM_EXIT_IO`` (x86) or -``KVM_EXIT_MMIO`` (all except s390) to implement functionality that -requires a guest to interact with host userspace. - -.. note:: KVM_EXIT_IO is significantly faster than KVM_EXIT_MMIO. - For arm64: ----------
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
Sean Christopherson
maintainerPATCH 2/5 KVM: Clear vcpu->run->flags at start of KVM_RUN for all architectures
In reply to Sean ChristophersonClear kvm_run.flags at the start of KVM_RUN for all architectures to minimize the probability of leaving a stale flag set.
--- arch/arm64/kvm/arm.c | 1 - arch/arm64/kvm/handle_exit.c | 2 +- arch/powerpc/kvm/book3s_hv.c | 4 +--- arch/x86/kvm/x86.c | 1 - virt/kvm/kvm_main.c | 3 +++ 5 files changed, 5 insertions(+), 6 deletions(-)
arch/arm64/kvm/arm.c +0 -1
@@ -1128,7 +1128,6 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) ret =1 ; run->exit_reason =KVM_EXIT_UNKNOWN ; - run->flags =0 ;while (ret >0 ) {/* * Check conditions before entering the guest
arch/arm64/kvm/handle_exit.c +1 -1
@@ -186,7 +186,7 @@ static int kvm_handle_guest_debug(struct kvm_vcpu *vcpu) run->exit_reason =KVM_EXIT_DEBUG ; run->debug.arch.hsr =lower_32_bits (esr); run->debug.arch.hsr_high =upper_32_bits (esr); - run->flags =KVM_DEBUG_ARCH_HSR_HIGH_VALID ; + run->flags |=KVM_DEBUG_ARCH_HSR_HIGH_VALID ;switch (ESR_ELx_EC (esr)) {case ESR_ELx_EC_WATCHPT_LOW:
arch/powerpc/kvm/book3s_hv.c +1 -3
@@ -1704,9 +1704,7 @@ static int kvmppc_handle_exit_hv(struct kvm_vcpu *vcpu,/* Exit to guest with KVM_EXIT_NMI as exit reason */ run->exit_reason =KVM_EXIT_NMI ; run->hw.hardware_exit_reason = vcpu->arch.trap; -/* Clear out the old NMI status from run->flags */ - run->flags &= ~KVM_RUN_PPC_NMI_DISP_MASK ; -/* Now set the NMI status */ +/* Note, run->flags is cleared at the start of KVM_RUN. */ if (vcpu->arch.mce_evt.disposition ==MCE_DISPOSITION_RECOVERED ) run->flags |=KVM_RUN_PPC_NMI_DISP_FULLY_RECOV ;else
arch/x86/kvm/x86.c +0 -1
@@ -11465,7 +11465,6 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)vcpu_load (vcpu);kvm_sigset_activate (vcpu); - kvm_run->flags =0 ;kvm_load_guest_fpu (vcpu);kvm_vcpu_srcu_read_lock (vcpu);
virt/kvm/kvm_main.c +3 -0
@@ -4336,6 +4336,9 @@ static long kvm_vcpu_ioctl(struct file *filp,put_pid (oldpid); } + + vcpu->run->flags =0 ; + vcpu->wants_to_run = !READ_ONCE (vcpu->run->immediate_exit__unsafe); r =kvm_arch_vcpu_ioctl_run (vcpu); vcpu->wants_to_run = false;
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
Sean Christopherson
maintainerPATCH 3/5 KVM: Add a common kvm_run flag to communicate an exit needs completion
In reply to Sean ChristophersonAdd a kvm_run flag, KVM_RUN_NEEDS_COMPLETION, to communicate to userspace that KVM_RUN needs to be re-executed prior to save/restore in order to complete the instruction/operation that triggered the userspace exit.
KVM's current approach of adding notes in the Documentation is beyond brittle, e.g. there is at least one known case where a KVM developer added a new userspace exit type, and then that same developer forgot to handle completion when adding userspace support.
On x86, there are multiple exits that need completion, but they are all conveniently funneled through a single callback, i.e. in theory, this is a one-time thing for KVM x86 (and other architectures could follow suit with additional refactoring).
--- Documentation/virt/kvm/api.rst | 48 ++++++++++++++++++++++--------- arch/powerpc/kvm/book3s_emulate.c | 1 + arch/powerpc/kvm/book3s_hv.c | 1 + arch/powerpc/kvm/book3s_pr.c | 2 ++ arch/powerpc/kvm/booke.c | 1 + arch/x86/include/uapi/asm/kvm.h | 7 +++-- arch/x86/kvm/x86.c | 2 ++ include/uapi/linux/kvm.h | 3 ++ virt/kvm/kvm_main.c | 1 + 9 files changed, 49 insertions(+), 17 deletions(-)
Documentation/virt/kvm/api.rst +34 -14
@@ -6505,7 +6505,7 @@ local APIC is not used. __u16 flags; -More architecture-specific flags detailing state of the VCPU that may +Common and architecture-specific flags detailing state of the VCPU that may affect the device's behavior. Current defined flags:: /* x86, set if the VCPU is in system management mode */
@@ -6518,6 +6518,8 @@ affect the device's behavior. Current defined flags:: /* arm64, set for KVM_EXIT_DEBUG */ #define KVM_DEBUG_ARCH_HSR_HIGH_VALID (1 << 0) + /* all architectures, set when the exit needs completion (via KVM_RUN) */ + #define KVM_RUN_NEEDS_COMPLETION (1 << 15) :: /* in (pre_kvm_run), out (post_kvm_run) */
@@ -6632,19 +6634,10 @@ requires a guest to interact with host userspace. .. note:: - For KVM_EXIT_IO, KVM_EXIT_MMIO, KVM_EXIT_OSI, KVM_EXIT_PAPR, KVM_EXIT_XEN, - KVM_EXIT_EPR, KVM_EXIT_X86_RDMSR, KVM_EXIT_X86_WRMSR, and KVM_EXIT_HYPERCALL - the corresponding operations are complete (and guest state is consistent) - only after userspace has re-entered the kernel with KVM_RUN. The kernel - side will first finish incomplete operations and then check for pending - signals. + For some exits, userspace must re-enter the kernel with KVM_RUN to + complete the exit and ensure guest state is consistent. - The pending state of the operation is not preserved in state which is - visible to userspace, thus userspace should ensure that the operation is - completed before performing a live migration. Userspace can re-enter the - guest with an unmasked signal pending or with the immediate_exit field set - to complete pending operations without allowing any further instructions - to be executed. + See KVM_CAP_NEEDS_COMPLETION for details. ::
@@ -8239,7 +8232,7 @@ Note: Userspace is responsible for correctly configuring CPUID 0x15, a.k.a. the core crystal clock frequency, if a non-zero CPUID 0x15 is exposed to the guest. 7.36 KVM_CAP_X86_GUEST_MODE ------------------------------- +--------------------------- :Architectures: x86 :Returns: Informational only, -EINVAL on direct KVM_ENABLE_CAP.
@@ -8252,6 +8245,33 @@ KVM exits with the register state of either the L1 or L2 guest depending on which executed at the time of an exit. Userspace must take care to differentiate between these cases. +7.37 KVM_CAP_NEEDS_COMPLETION +----------------------------- + +:Architectures: all +:Returns: Informational only, -EINVAL on direct KVM_ENABLE_CAP. + +The presence of this capability indicates that KVM_RUN will set +KVM_RUN_NEEDS_COMPLETION in kvm_run.flags if KVM requires userspace to re-enter +the kernel KVM_RUN to complete the exit. + +For select exits, userspace must re-enter the kernel with KVM_RUN to complete +the corresponding operation, only after which is guest state guaranteed to be +consistent. On such a KVM_RUN, the kernel side will first finish incomplete +operations and then check for pending signals. + +The pending state of the operation for such exits is not preserved in state +which is visible to userspace, thus userspace should ensure that the operation +is completed before performing state save/restore, e.g. for live migration. +Userspace can re-enter the guest with an unmasked signal pending or with the +immediate_exit field set to complete pending operations without allowing any +further instructions to be executed. + +Without KVM_CAP_NEEDS_COMPLETION, KVM_RUN_NEEDS_COMPLETION will never be set +and userspace must assume that exits of type KVM_EXIT_IO, KVM_EXIT_MMIO, +KVM_EXIT_OSI, KVM_EXIT_PAPR, KVM_EXIT_XEN, KVM_EXIT_EPR, KVM_EXIT_X86_RDMSR, +KVM_EXIT_X86_WRMSR, and KVM_EXIT_HYPERCALL require completion. + 8. Other capabilities. ======================
arch/powerpc/kvm/book3s_emulate.c +1 -0
@@ -374,6 +374,7 @@ int kvmppc_core_emulate_op_pr(struct kvm_vcpu *vcpu, } vcpu->run->exit_reason =KVM_EXIT_PAPR_HCALL ; + vcpu->run->flags |=KVM_RUN_NEEDS_COMPLETION ; vcpu->arch.hcall_needed =1 ; emulated =EMULATE_EXIT_USER ;break ;
arch/powerpc/kvm/book3s_hv.c +1 -0
@@ -1767,6 +1767,7 @@ static int kvmppc_handle_exit_hv(struct kvm_vcpu *vcpu,for (i =0 ; i <9 ; ++i) run->papr_hcall.args[i] =kvmppc_get_gpr (vcpu,4 + i); run->exit_reason =KVM_EXIT_PAPR_HCALL ; + run->flags |=KVM_RUN_NEEDS_COMPLETION ; vcpu->arch.hcall_needed =1 ; r =RESUME_HOST ;break ;
arch/powerpc/kvm/book3s_pr.c +2 -0
@@ -1310,6 +1310,7 @@ int kvmppc_handle_exit_pr(struct kvm_vcpu *vcpu, unsigned int exit_nr) run->papr_hcall.args[i] = gpr; } run->exit_reason =KVM_EXIT_PAPR_HCALL ; + run->flags |=KVM_RUN_NEEDS_COMPLETION ; vcpu->arch.hcall_needed =1 ; r =RESUME_HOST ; } else if (vcpu->arch.osi_enabled &&
@@ -1320,6 +1321,7 @@ int kvmppc_handle_exit_pr(struct kvm_vcpu *vcpu, unsigned int exit_nr)int i; run->exit_reason =KVM_EXIT_OSI ; + run->flags |=KVM_RUN_NEEDS_COMPLETION ;for (i =0 ; i <32 ; i++) gprs[i] =kvmppc_get_gpr (vcpu, i); vcpu->arch.osi_needed =1 ;
arch/powerpc/kvm/booke.c +1 -0
@@ -751,6 +751,7 @@ int kvmppc_core_check_requests(struct kvm_vcpu *vcpu) vcpu->run->epr.epr =0 ; vcpu->arch.epr_needed = true; vcpu->run->exit_reason =KVM_EXIT_EPR ; + vcpu->run->flags |=KVM_RUN_NEEDS_COMPLETION ; r =0 ; }
arch/x86/include/uapi/asm/kvm.h +4 -3
@@ -104,9 +104,10 @@ struct kvm_ioapic_state {#define KVM_IRQCHIP_IOAPIC 2#define KVM_NR_IRQCHIPS 3 -#define KVM_RUN_X86_SMM (1 << 0) -#define KVM_RUN_X86_BUS_LOCK (1 << 1) -#define KVM_RUN_X86_GUEST_MODE (1 << 2) +#define KVM_RUN_X86_SMM (1 << 0) +#define KVM_RUN_X86_BUS_LOCK (1 << 1) +#define KVM_RUN_X86_GUEST_MODE (1 << 2) +#define KVM_RUN_X86_NEEDS_COMPLETION (1 << 2)/* for KVM_GET_REGS and KVM_SET_REGS */ struct kvm_regs {
arch/x86/kvm/x86.c +2 -0
@@ -10154,6 +10154,8 @@ static void post_kvm_run_save(struct kvm_vcpu *vcpu) kvm_run->flags |=KVM_RUN_X86_SMM ;if (is_guest_mode (vcpu)) kvm_run->flags |=KVM_RUN_X86_GUEST_MODE ; +if (vcpu->arch.complete_userspace_io) + kvm_run->flags |=KVM_RUN_NEEDS_COMPLETION ; }static void update_cr8_intercept (struct kvm_vcpu *vcpu)
include/uapi/linux/kvm.h +3 -0
@@ -192,6 +192,8 @@ struct kvm_xen_exit {/* Flags that describe what fields in emulation_failure hold valid data. */ #define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) +#define KVM_RUN_NEEDS_COMPLETION (1 << 15) +/* * struct kvm_run can be modified by userspace at any time, so KVM must be * careful to avoid TOCTOU bugs. In order to protect KVM, HINT_UNSAFE_IN_KVM()
@@ -933,6 +935,7 @@ struct kvm_enable_cap {#define KVM_CAP_PRE_FAULT_MEMORY 236#define KVM_CAP_X86_APIC_BUS_CYCLES_NS 237#define KVM_CAP_X86_GUEST_MODE 238 +#define KVM_CAP_NEEDS_COMPLETION 239struct kvm_irq_routing_irqchip {__u32 irqchip;
virt/kvm/kvm_main.c +1 -0
@@ -4746,6 +4746,7 @@ static int kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)case KVM_CAP_CHECK_EXTENSION_VM :case KVM_CAP_ENABLE_CAP_VM :case KVM_CAP_HALT_POLL : +case KVM_CAP_NEEDS_COMPLETION :return 1 ;#ifdef CONFIG_KVM_MMIO case KVM_CAP_COALESCED_MMIO :
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
Marc Zyngier
maintainerRe: [PATCH 3/5] KVM: Add a common kvm_run flag to communicate an exit needs completion
Quoted 9 lines · Sean Christopherson
Add a
kvm_runflag,KVM_RUN_NEEDS_COMPLETION, to communicate to userspace thatKVM_RUNneeds to be re-executed prior to save/restore in order to complete the instruction/operation that triggered the userspace exit.KVM's current approach of adding notes in the Documentation is beyond brittle, e.g. there is at least one known case where a KVM developer added a new userspace exit type, and then that same developer forgot to handle completion when adding userspace support.
Is this going to fix anything? If they couldn't be bothered to read the documentation, let alone update it, how is that going to be improved by extra rules and regulations?
I don't see how someone ignoring the documented behaviour of a given exit reason is, all of a sudden, have an epiphany and take a *new* flag into account.
Quoted 93 lines
On x86, there are multiple exits that need completion, but they are all conveniently funneled through a single callback, i.e. in theory, this is a one-time thing for KVM x86 (and other architectures could follow suit with additional refactoring).
Signed-off-by Sean Christopherson <[email protected]>--- Documentation/virt/kvm/api.rst | 48 ++++++++++++++++++++++--------- arch/powerpc/kvm/book3s_emulate.c | 1 + arch/powerpc/kvm/book3s_hv.c | 1 + arch/powerpc/kvm/book3s_pr.c | 2 ++ arch/powerpc/kvm/booke.c | 1 + arch/x86/include/uapi/asm/kvm.h | 7 +++-- arch/x86/kvm/x86.c | 2 ++ include/uapi/linux/kvm.h | 3 ++ virt/kvm/kvm_main.c | 1 + 9 files changed, 49 insertions(+), 17 deletions(-)Documentation/virt/kvm/api.rst +23
-14@@ -6505,7 +6505,7 @@ local APIC is not used. __u16 flags; -More architecture-specific flags detailing state of the VCPU that may +Common and architecture-specific flags detailing state of the VCPU that may affect the device's behavior. Current defined flags:: /* x86, set if the VCPU is in system management mode */@@ -6518,6 +6518,8 @@ affect the device's behavior. Current defined flags:: /* arm64, set for KVM_EXIT_DEBUG */ #define KVM_DEBUG_ARCH_HSR_HIGH_VALID (1 << 0) + /* all architectures, set when the exit needs completion (via KVM_RUN) */ + #define KVM_RUN_NEEDS_COMPLETION (1 << 15) :: /* in (pre_kvm_run), out (post_kvm_run) */@@ -6632,19 +6634,10 @@ requires a guest to interact with host userspace. .. note:: - For KVM_EXIT_IO, KVM_EXIT_MMIO, KVM_EXIT_OSI, KVM_EXIT_PAPR, KVM_EXIT_XEN, - KVM_EXIT_EPR, KVM_EXIT_X86_RDMSR, KVM_EXIT_X86_WRMSR, and KVM_EXIT_HYPERCALL - the corresponding operations are complete (and guest state is consistent) - only after userspace has re-entered the kernel with KVM_RUN. The kernel - side will first finish incomplete operations and then check for pending - signals. + For some exits, userspace must re-enter the kernel with KVM_RUN to + complete the exit and ensure guest state is consistent. - The pending state of the operation is not preserved in state which is - visible to userspace, thus userspace should ensure that the operation is - completed before performing a live migration. Userspace can re-enter the - guest with an unmasked signal pending or with the immediate_exit field set - to complete pending operations without allowing any further instructions - to be executed. + See KVM_CAP_NEEDS_COMPLETION for details. ::@@ -8239,7 +8232,7 @@ Note: Userspace is responsible for correctly configuring CPUID 0x15, a.k.a. the core crystal clock frequency, if a non-zero CPUID 0x15 is exposed to the guest. 7.36 KVM_CAP_X86_GUEST_MODE ------------------------------- +--------------------------- :Architectures: x86 :Returns: Informational only, -EINVAL on direct KVM_ENABLE_CAP.@@ -8252,6 +8245,33 @@ KVM exits with the register state of either the L1 or L2 guest depending on which executed at the time of an exit. Userspace must take care to differentiate between these cases. +7.37 KVM_CAP_NEEDS_COMPLETION +----------------------------- + +:Architectures: all +:Returns: Informational only, -EINVAL on direct KVM_ENABLE_CAP. + +The presence of this capability indicates that KVM_RUN will set +KVM_RUN_NEEDS_COMPLETION in kvm_run.flags if KVM requires userspace to re-enter +the kernel KVM_RUN to complete the exit. + +For select exits, userspace must re-enter the kernel with KVM_RUN to complete +the corresponding operation, only after which is guest state guaranteed to be +consistent. On such a KVM_RUN, the kernel side will first finish incomplete +operations and then check for pending signals. + +The pending state of the operation for such exits is not preserved in state
Documentation/virt/kvm/api.rst KVM exits with the register state of either the L1 or L2 guest · line 8273
Quoted 10 lines
Documentation/virt/kvm/api.rst +10
-0+which is visible to userspace, thus userspace should ensure that the operation +is completed before performing state save/restore, e.g. for live migration. +Userspace can re-enter the guest with an unmasked signal pending or with the +immediate_exit field set to complete pending operations without allowing any +further instructions to be executed. + +Without KVM_CAP_NEEDS_COMPLETION, KVM_RUN_NEEDS_COMPLETION will never be set +and userspace must assume that exits of type KVM_EXIT_IO, KVM_EXIT_MMIO, +KVM_EXIT_OSI, KVM_EXIT_PAPR, KVM_EXIT_XEN, KVM_EXIT_EPR, KVM_EXIT_X86_RDMSR, +KVM_EXIT_X86_WRMSR, and KVM_EXIT_HYPERCALL require completion.
So once you advertise KVM_CAP_NEEDS_COMPLETION, the completion flag must be present for all of these exits, right? And from what I can tell, this capability is unconditionally advertised.
Yet, you don't amend arm64 to publish that flag. Not that I think this causes any issue (even if you save the state at that point without reentering the guest, it will be still be consistent), but that directly contradicts the documentation (isn't that ironic? ;-).
Or is your intent to *relax* the requirements on arm64 (and anything else but x86 and POWER)?
M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
Sean Christopherson
maintainerRe: [PATCH 3/5] KVM: Add a common kvm_run flag to communicate an exit needs completion
Quoted 12 lines · Marc Zyngier
Quoted 9 lines · Sean Christopherson
Add a
kvm_runflag,KVM_RUN_NEEDS_COMPLETION, to communicate to userspace thatKVM_RUNneeds to be re-executed prior to save/restore in order to complete the instruction/operation that triggered the userspace exit.KVM's current approach of adding notes in the Documentation is beyond brittle, e.g. there is at least one known case where a KVM developer added a new userspace exit type, and then that same developer forgot to handle completion when adding userspace support.
Quoted 7 lines
Is this going to fix anything? If they couldn't be bothered to read the documentation, let alone update it, how is that going to be improved by extra rules and regulations?
I don't see how someone ignoring the documented behaviour of a given exit reason is, all of a sudden, have an epiphany and take a *new* flag into account.
The idea is to reduce the probability of introducing bugs, in KVM or userspace, every time KVM attaches a completion callback. Yes, userspace would need to be updated to handle KVM_RUN_NEEDS_COMPLETION, but once that flag is merged, neither KVM's documentation nor userspace would never need to be updated again. And if all architectures took an approach of handling completion via function callback, I'm pretty sure we'd never need to manually update KVM itself either.
Quoted 31 lines
Quoted 16 lines
+7.37
KVM_CAP_NEEDS_COMPLETION
+-----------------------------Documentation/virt/kvm/api.rst +14
-0+ +:Architectures: all +:Returns: Informational only, -EINVAL on direct KVM_ENABLE_CAP. + +The presence of this capability indicates that KVM_RUN will set +KVM_RUN_NEEDS_COMPLETION in kvm_run.flags if KVM requires userspace to re-enter +the kernel KVM_RUN to complete the exit. + +For select exits, userspace must re-enter the kernel with KVM_RUN to complete +the corresponding operation, only after which is guest state guaranteed to be +consistent. On such a KVM_RUN, the kernel side will first finish incomplete +operations and then check for pending signals. + +The pending state of the operation for such exits is not preserved in stateDocumentation/virt/kvm/api.rst KVM exits with the register state of either the L1 or L2 guest · line 8273
Quoted 10 lines
Documentation/virt/kvm/api.rst +10
-0+which is visible to userspace, thus userspace should ensure that the operation +is completed before performing state save/restore, e.g. for live migration. +Userspace can re-enter the guest with an unmasked signal pending or with the +immediate_exit field set to complete pending operations without allowing any +further instructions to be executed. + +Without KVM_CAP_NEEDS_COMPLETION, KVM_RUN_NEEDS_COMPLETION will never be set +and userspace must assume that exits of type KVM_EXIT_IO, KVM_EXIT_MMIO, +KVM_EXIT_OSI, KVM_EXIT_PAPR, KVM_EXIT_XEN, KVM_EXIT_EPR, KVM_EXIT_X86_RDMSR, +KVM_EXIT_X86_WRMSR, and KVM_EXIT_HYPERCALL require completion.So once you advertise
KVM_CAP_NEEDS_COMPLETION, the completion flag must be present for all of these exits, right? And from what I can tell, this capability is unconditionally advertised.
Quoted 4 lines
Yet, you don't amend arm64 to publish that flag. Not that I think this causes any issue (even if you save the state at that point without reentering the guest, it will be still be consistent), but that directly contradicts the documentation (isn't that ironic? ;-).
It does cause issues, I missed this code in kvm_arch_vcpu_ioctl_run():
if (run->exit_reason == KVM_EXIT_MMIO) {
ret = kvm_handle_mmio_return(vcpu);
if (ret <= 0)
return ret;
}Quoted 2 lines
Or is your intent to *relax* the requirements on arm64 (and anything else but x86 and POWER)?
Marc Zyngier
maintainerRe: [PATCH 3/5] KVM: Add a common kvm_run flag to communicate an exit needs completion
Quoted 22 lines · Sean Christopherson
Quoted 12 lines · Marc Zyngier
Quoted 9 lines · Sean Christopherson
Add a
kvm_runflag,KVM_RUN_NEEDS_COMPLETION, to communicate to userspace thatKVM_RUNneeds to be re-executed prior to save/restore in order to complete the instruction/operation that triggered the userspace exit.KVM's current approach of adding notes in the Documentation is beyond brittle, e.g. there is at least one known case where a KVM developer added a new userspace exit type, and then that same developer forgot to handle completion when adding userspace support.
Quoted 7 lines
Is this going to fix anything? If they couldn't be bothered to read the documentation, let alone update it, how is that going to be improved by extra rules and regulations?
I don't see how someone ignoring the documented behaviour of a given exit reason is, all of a sudden, have an epiphany and take a *new* flag into account.
Quoted 6 lines
The idea is to reduce the probability of introducing bugs, in KVM or userspace, every time KVM attaches a completion callback. Yes, userspace would need to be updated to handle
KVM_RUN_NEEDS_COMPLETION, but once that flag is merged, neither KVM's documentation nor userspace would never need to be updated again. And if all architectures took an approach of handling completion via function callback, I'm pretty sure we'd never need to manually update KVM itself either.
You are assuming that we need this completion, and I dispute this assertion.
Quoted 37 lines
Quoted 32 lines
Quoted 17 lines
+7.37
KVM_CAP_NEEDS_COMPLETION
+-----------------------------Documentation/virt/kvm/api.rst +14
-0+ +:Architectures: all +:Returns: Informational only, -EINVAL on direct KVM_ENABLE_CAP. + +The presence of this capability indicates that KVM_RUN will set +KVM_RUN_NEEDS_COMPLETION in kvm_run.flags if KVM requires userspace to re-enter +the kernel KVM_RUN to complete the exit. + +For select exits, userspace must re-enter the kernel with KVM_RUN to complete +the corresponding operation, only after which is guest state guaranteed to be +consistent. On such a KVM_RUN, the kernel side will first finish incomplete +operations and then check for pending signals. + +The pending state of the operation for such exits is not preserved in stateDocumentation/virt/kvm/api.rst KVM exits with the register state of either the L1 or L2 guest · line 8273
Quoted 10 lines
Documentation/virt/kvm/api.rst +10
-0+which is visible to userspace, thus userspace should ensure that the operation +is completed before performing state save/restore, e.g. for live migration. +Userspace can re-enter the guest with an unmasked signal pending or with the +immediate_exit field set to complete pending operations without allowing any +further instructions to be executed. + +Without KVM_CAP_NEEDS_COMPLETION, KVM_RUN_NEEDS_COMPLETION will never be set +and userspace must assume that exits of type KVM_EXIT_IO, KVM_EXIT_MMIO, +KVM_EXIT_OSI, KVM_EXIT_PAPR, KVM_EXIT_XEN, KVM_EXIT_EPR, KVM_EXIT_X86_RDMSR, +KVM_EXIT_X86_WRMSR, and KVM_EXIT_HYPERCALL require completion.So once you advertise
KVM_CAP_NEEDS_COMPLETION, the completion flag must be present for all of these exits, right? And from what I can tell, this capability is unconditionally advertised.Quoted 4 lines
Yet, you don't amend arm64 to publish that flag. Not that I think this causes any issue (even if you save the state at that point without reentering the guest, it will be still be consistent), but that directly contradicts the documentation (isn't that ironic? ;-).
Quoted 7 lines
It does cause issues, I missed this code in
kvm_arch_vcpu_ioctl_run():if (run->exit_reason == KVM_EXIT_MMIO) { ret = kvm_handle_mmio_return(vcpu); if (ret <= 0) return ret; }
That's satisfying a load from the guest forwarded to userspace. If the VMM did a save of the guest at this stage, restored and resumed it, *nothing* bad would happen, as PC still points to the instruction that got forwarded. You'll see the same load again.
As for all arm64 synchronous exceptions, they are idempotent, and can be repeated as often as you want without side effects.
M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
Sean Christopherson
maintainerRe: [PATCH 3/5] KVM: Add a common kvm_run flag to communicate an exit needs completion
Quoted 31 lines · Marc Zyngier
Quoted 22 lines · Sean Christopherson
Quoted 12 lines · Marc Zyngier
Quoted 9 lines · Sean Christopherson
Add a
kvm_runflag,KVM_RUN_NEEDS_COMPLETION, to communicate to userspace thatKVM_RUNneeds to be re-executed prior to save/restore in order to complete the instruction/operation that triggered the userspace exit.KVM's current approach of adding notes in the Documentation is beyond brittle, e.g. there is at least one known case where a KVM developer added a new userspace exit type, and then that same developer forgot to handle completion when adding userspace support.
Quoted 7 lines
Is this going to fix anything? If they couldn't be bothered to read the documentation, let alone update it, how is that going to be improved by extra rules and regulations?
I don't see how someone ignoring the documented behaviour of a given exit reason is, all of a sudden, have an epiphany and take a *new* flag into account.
Quoted 6 lines
The idea is to reduce the probability of introducing bugs, in KVM or userspace, every time KVM attaches a completion callback. Yes, userspace would need to be updated to handle
KVM_RUN_NEEDS_COMPLETION, but once that flag is merged, neither KVM's documentation nor userspace would never need to be updated again. And if all architectures took an approach of handling completion via function callback, I'm pretty sure we'd never need to manually update KVM itself either.
Quoted 2 lines
You are assuming that we need this completion, and I dispute this assertion.
Ah, gotcha.
Quoted 23 lines
Quoted 23 lines
Quoted 16 lines
Documentation/virt/kvm/api.rst KVM exits with the register state of either the L1 or L2 guest · line 8273
Quoted 11 lines
Documentation/virt/kvm/api.rst +11
-0+The pending state of the operation for such exits is not preserved in state +which is visible to userspace, thus userspace should ensure that the operation +is completed before performing state save/restore, e.g. for live migration. +Userspace can re-enter the guest with an unmasked signal pending or with the +immediate_exit field set to complete pending operations without allowing any +further instructions to be executed. + +Without KVM_CAP_NEEDS_COMPLETION, KVM_RUN_NEEDS_COMPLETION will never be set +and userspace must assume that exits of type KVM_EXIT_IO, KVM_EXIT_MMIO, +KVM_EXIT_OSI, KVM_EXIT_PAPR, KVM_EXIT_XEN, KVM_EXIT_EPR, KVM_EXIT_X86_RDMSR, +KVM_EXIT_X86_WRMSR, and KVM_EXIT_HYPERCALL require completion.So once you advertise
KVM_CAP_NEEDS_COMPLETION, the completion flag must be present for all of these exits, right? And from what I can tell, this capability is unconditionally advertised.Quoted 4 lines
Yet, you don't amend arm64 to publish that flag. Not that I think this causes any issue (even if you save the state at that point without reentering the guest, it will be still be consistent), but that directly contradicts the documentation (isn't that ironic? ;-).
It does cause issues, I missed this code in
kvm_arch_vcpu_ioctl_run():
Quoted 7 lines
Quoted 5 lines
if (run->exit_reason == KVM_EXIT_MMIO) { ret = kvm_handle_mmio_return(vcpu); if (ret <= 0) return ret; }That's satisfying a load from the guest forwarded to userspace.
And MMIO stores, no? I.e. PC needs to be incremented on stores as well.
Quoted 3 lines
If the VMM did a save of the guest at this stage, restored and resumed it, *nothing* bad would happen, as PC still points to the instruction that got forwarded. You'll see the same load again.
But replaying an MMIO store could cause all kinds of problems, and even MMIO loads could theoretically be problematic, e.g. if there are side effects in the device that trigger on access to a device register.
Marc Zyngier
maintainerRe: [PATCH 3/5] KVM: Add a common kvm_run flag to communicate an exit needs completion
Quoted 67 lines · Sean Christopherson
Quoted 31 lines · Marc Zyngier
Quoted 22 lines · Sean Christopherson
Quoted 12 lines · Marc Zyngier
Quoted 9 lines · Sean Christopherson
Add a
kvm_runflag,KVM_RUN_NEEDS_COMPLETION, to communicate to userspace thatKVM_RUNneeds to be re-executed prior to save/restore in order to complete the instruction/operation that triggered the userspace exit.KVM's current approach of adding notes in the Documentation is beyond brittle, e.g. there is at least one known case where a KVM developer added a new userspace exit type, and then that same developer forgot to handle completion when adding userspace support.
Quoted 7 lines
Is this going to fix anything? If they couldn't be bothered to read the documentation, let alone update it, how is that going to be improved by extra rules and regulations?
I don't see how someone ignoring the documented behaviour of a given exit reason is, all of a sudden, have an epiphany and take a *new* flag into account.
Quoted 6 lines
The idea is to reduce the probability of introducing bugs, in KVM or userspace, every time KVM attaches a completion callback. Yes, userspace would need to be updated to handle
KVM_RUN_NEEDS_COMPLETION, but once that flag is merged, neither KVM's documentation nor userspace would never need to be updated again. And if all architectures took an approach of handling completion via function callback, I'm pretty sure we'd never need to manually update KVM itself either.Quoted 2 lines
You are assuming that we need this completion, and I dispute this assertion.
Ah, gotcha.
Quoted 29 lines
Quoted 28 lines
Quoted 16 lines
Documentation/virt/kvm/api.rst KVM exits with the register state of either the L1 or L2 guest · line 8273
Quoted 11 lines
Documentation/virt/kvm/api.rst +11
-0+The pending state of the operation for such exits is not preserved in state +which is visible to userspace, thus userspace should ensure that the operation +is completed before performing state save/restore, e.g. for live migration. +Userspace can re-enter the guest with an unmasked signal pending or with the +immediate_exit field set to complete pending operations without allowing any +further instructions to be executed. + +Without KVM_CAP_NEEDS_COMPLETION, KVM_RUN_NEEDS_COMPLETION will never be set +and userspace must assume that exits of type KVM_EXIT_IO, KVM_EXIT_MMIO, +KVM_EXIT_OSI, KVM_EXIT_PAPR, KVM_EXIT_XEN, KVM_EXIT_EPR, KVM_EXIT_X86_RDMSR, +KVM_EXIT_X86_WRMSR, and KVM_EXIT_HYPERCALL require completion.So once you advertise
KVM_CAP_NEEDS_COMPLETION, the completion flag must be present for all of these exits, right? And from what I can tell, this capability is unconditionally advertised.Quoted 4 lines
Yet, you don't amend arm64 to publish that flag. Not that I think this causes any issue (even if you save the state at that point without reentering the guest, it will be still be consistent), but that directly contradicts the documentation (isn't that ironic? ;-).
It does cause issues, I missed this code in
kvm_arch_vcpu_ioctl_run():if (run->exit_reason == KVM_EXIT_MMIO) { ret = kvm_handle_mmio_return(vcpu); if (ret <= 0) return ret; }
Quoted 3 lines
Quoted 1 line
That's satisfying a load from the guest forwarded to userspace.
And MMIO stores, no? I.e. PC needs to be incremented on stores as well.
Yes, *after* the store as completed. If you replay the instruction, the same store comes out.
Quoted 8 lines
Quoted 4 lines
If the VMM did a save of the guest at this stage, restored and resumed it, *nothing* bad would happen, as PC still points to the instruction that got forwarded. You'll see the same load again.
But replaying an MMIO store could cause all kinds of problems, and even MMIO loads could theoretically be problematic, e.g. if there are side effects in the device that trigger on access to a device register.
But that's the VMM's problem. If it has modified its own state and doesn't return to the guest to complete the instruction, that's just as bad as a load, which *do* have side effects as well.
Overall, the guest state exposed by KVM is always correct, and replaying the instruction is not going to change that. It is if the VMM is broken that things turn ugly *for the VMM itself*, and I claim that no amount of flag being added is going to help that.
M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
Sean Christopherson
maintainerRe: [PATCH 3/5] KVM: Add a common kvm_run flag to communicate an exit needs completion
Quoted 32 lines · Marc Zyngier
Quoted 19 lines · Sean Christopherson
Quoted 17 lines · Marc Zyngier
Quoted 14 lines · Sean Christopherson
Quoted 4 lines · Marc Zyngier
Yet, you don't amend arm64 to publish that flag. Not that I think this causes any issue (even if you save the state at that point without reentering the guest, it will be still be consistent), but that directly contradicts the documentation (isn't that ironic? ;-).
It does cause issues, I missed this code in
kvm_arch_vcpu_ioctl_run():if (run->exit_reason == KVM_EXIT_MMIO) { ret = kvm_handle_mmio_return(vcpu); if (ret <= 0) return ret; }Quoted 3 lines
Quoted 1 line
That's satisfying a load from the guest forwarded to userspace.
And MMIO stores, no? I.e. PC needs to be incremented on stores as well.
Yes, *after* the store as completed. If you replay the instruction, the same store comes out.
Quoted 4 lines
Quoted 3 lines
If the VMM did a save of the guest at this stage, restored and resumed it, *nothing* bad would happen, as PC still points to the instruction that got forwarded. You'll see the same load again.
Quoted 7 lines
Quoted 3 lines
But replaying an MMIO store could cause all kinds of problems, and even MMIO loads could theoretically be problematic, e.g. if there are side effects in the device that trigger on access to a device register.
But that's the VMM's problem. If it has modified its own state and doesn't return to the guest to complete the instruction, that's just as bad as a load, which *do* have side effects as well.
Agreed, just wanted to make sure I wasn't completely misunderstanding something about arm64.
Quoted 4 lines
Overall, the guest state exposed by KVM is always correct, and replaying the instruction is not going to change that. It is if the VMM is broken that things turn ugly *for the VMM itself*, and I claim that no amount of flag being added is going to help that.
On x86 at least, adding KVM_RUN_NEEDS_COMPLETION reduces the chances for human error. x86 has had bugs in both KVM (patch 1) and userspace (Google's VMM when handling MSR exits) that would have been avoided if KVM_RUN_NEEDS_COMPLETION existed. Unless the VMM is doing something decidely odd, userspace needs to write code once (maybe even just once for all architectures). For KVM, the flag is set based on whether or not the vCPU has a valid completion callback, i.e. will be correct so long as the underlying KVM code is correct.
Contrast that with the current approach, where the KVM developer needs to get the KVM code correct and remember to update KVM's documentation. Documentation is especially problematic, because in practice it can't be tested, i.e. is much more likely to be missed by the developer and the maintainer. The VMM either needs to blindly redo KVM_RUN (as selftests do, and apparently as QEMU does), or the developer adding VMM support needs to be diligent in reading KVM's documentation. And like KVM documentation, testing that the VMM is implemented to KVM's "spec" is effectively impossible in practice, because 99.9999% of the time userspace exits and save/restore will work just fine.
I do agree that the VMM is likely going to run into problems sooner or later if the developers/maintainers don't fundamentally understand the need to redo KVM_RUN, but I also think there's significant value in reducing the chances for simple human error to result in broken VMs.
Re: [PATCH 3/5] KVM: Add a common kvm_run flag to communicate an exit needs completion
Quoted 4 lines · Sean Christopherson
I do agree that the VMM is likely going to run into problems sooner or later if the developers/maintainers don't fundamentally understand the need to redo
KVM_RUN, but I also think there's significant value in reducing the chances for simple human error to result in broken VMs.
I agree that it makes sense for the kernel to expose this flag, and keep it in sync with the exit reasons that need it as they accumulate over time, rather than having to keep the VMM in sync.
Coincidentally, I did an audit of a VMM only a few weeks ago, adding a userspace flag for 'exit_response_pending' and making sure it does one last run of the vCPU with an immediate exit requested, before pausing for migration. I added that flag for exactly the same KVM_EXIT_xxx reasons, matching the kernel... for now.
(I just said in a separate thread that you missed KVM_EXIT_XEN, but on closer inspection I was wrong because kvm_xen_hypercall() sets the arch.complete_userspace_io flag so it's fine; I just didn't see the Xen-specific setup in my first look.)
Re: [PATCH 3/5] KVM: Add a common kvm_run flag to communicate an exit needs completion
In reply to Sean ChristophersonQuoted 8 lines · Sean Christopherson
Add a
kvm_runflag,KVM_RUN_NEEDS_COMPLETION, to communicate to userspace thatKVM_RUNneeds to be re-executed prior to save/restore in order to complete the instruction/operation that triggered the userspace exit.KVM's current approach of adding notes in the Documentation is beyond brittle, e.g. there is at least one known case where a KVM developer added a new userspace exit type, and then that same developer forgot to handle completion when adding userspace support.
This answers one question I had:
https://lore.kernel.org/kvm/[email protected]/
So, it is the VMM's (i.e., QEMU's) responsibility to re-execute KVM_RUN in this case.
Btw, can this flag be used to address the issue [*] with steal time accounting? We can set the new flag for each vCPU in the PM notifier and we need to change the re-execution to handle steal time accounting (not just IO completion).
[*]: https://lore.kernel.org/kvm/[email protected]/
one nit below,
arch/x86/include/uapi/asm/kvm.h struct kvm_ioapic_state { · line 110
Quoted 13 lines
arch/x86/include/uapi/asm/kvm.h +4
-3@@ -104,9 +104,10 @@ struct kvm_ioapic_state {#define KVM_IRQCHIP_IOAPIC 2#define KVM_NR_IRQCHIPS 3 -#define KVM_RUN_X86_SMM (1 << 0) -#define KVM_RUN_X86_BUS_LOCK (1 << 1) -#define KVM_RUN_X86_GUEST_MODE (1 << 2) +#define KVM_RUN_X86_SMM (1 << 0) +#define KVM_RUN_X86_BUS_LOCK (1 << 1) +#define KVM_RUN_X86_GUEST_MODE (1 << 2) +#define KVM_RUN_X86_NEEDS_COMPLETION (1 << 2)
This X86_NEEDS_COMPLETION should be dropped. It is never used.
Re: [PATCH 3/5] KVM: Add a common kvm_run flag to communicate an exit needs completion
Quoted 11 lines · Chao Gao
Quoted 8 lines · Sean Christopherson
Add a
kvm_runflag,KVM_RUN_NEEDS_COMPLETION, to communicate to userspace thatKVM_RUNneeds to be re-executed prior to save/restore in order to complete the instruction/operation that triggered the userspace exit.KVM's current approach of adding notes in the Documentation is beyond brittle, e.g. there is at least one known case where a KVM developer added a new userspace exit type, and then that same developer forgot to handle completion when adding userspace support.
This answers one question I had:
https://lore.kernel.org/kvm/[email protected]/
In current QEMU code, it always returns back to KVM via KVM_RUN after it successfully handled a KVM exit reason, no matter what the exit reason is. The complete_userspace_io() callback will be called if it has been setup. So if a new kvm exit reason is added in QEMU, it seems QEMU doesn't need special handing to make the complete_userspace_io() callback be called.
However, QEMU is not the only userspace VMM that supports KVM, it makes sense to make the solution generic and clear for different userspace VMMs.
Regarding the support of MapGPA for TDX when live migration is considered, since a big range will be split into 2MB chunks, in order the status is right after TD live migration, it needs to set the return code to retry with the next_gpa in the complete_userspace_io() callback if vcpu->wants_to_run is false or vcpu->run->immediate_exit__unsafe is set, otherwise, TDX guest will see return code as successful and think the whole range has been converted successfully.
@@ -1093,7 +1093,8 @@ static int tdx_complete_vmcall_map_gpa(struct kvm_vcpu *vcpu)
* immediately after STI or MOV/POP SS.
*/
if (pi_has_pending_interrupt(vcpu) ||
- kvm_test_request(KVM_REQ_NMI, vcpu) || vcpu->arch.nmi_pending) {
+ kvm_test_request(KVM_REQ_NMI, vcpu) || vcpu->arch.nmi_pending ||+ !vcpu->wants_to_run) { tdvmcall_set_return_code(vcpu, TDVMCALL_STATUS_RETRY); tdx->vp_enter_args.r11 = tdx->map_gpa_next; return 1;
Of course, it can be addressed later when TD live migration is supported.
Quoted 27 lines
So, it is the VMM's (i.e., QEMU's) responsibility to re-execute
KVM_RUNin this case.Btw, can this flag be used to address the issue [*] with steal time accounting? We can set the new flag for each vCPU in the PM notifier and we need to change the re-execution to handle steal time accounting (not just IO completion).
[*]: https://lore.kernel.org/kvm/[email protected]/
one nit below,
arch/x86/include/uapi/asm/kvm.h struct kvm_ioapic_state { · line 110
Quoted 13 lines
arch/x86/include/uapi/asm/kvm.h +4
-3@@ -104,9 +104,10 @@ struct kvm_ioapic_state {#define KVM_IRQCHIP_IOAPIC 2#define KVM_NR_IRQCHIPS 3 -#define KVM_RUN_X86_SMM (1 << 0) -#define KVM_RUN_X86_BUS_LOCK (1 << 1) -#define KVM_RUN_X86_GUEST_MODE (1 << 2) +#define KVM_RUN_X86_SMM (1 << 0) +#define KVM_RUN_X86_BUS_LOCK (1 << 1) +#define KVM_RUN_X86_GUEST_MODE (1 << 2) +#define KVM_RUN_X86_NEEDS_COMPLETION (1 << 2)This
X86_NEEDS_COMPLETIONshould be dropped. It is never used.
Sean Christopherson
maintainerRe: [PATCH 3/5] KVM: Add a common kvm_run flag to communicate an exit needs completion
In reply to Chao GaoQuoted 10 lines · Chao Gao
Quoted 8 lines · Sean Christopherson
Add a
kvm_runflag,KVM_RUN_NEEDS_COMPLETION, to communicate to userspace thatKVM_RUNneeds to be re-executed prior to save/restore in order to complete the instruction/operation that triggered the userspace exit.KVM's current approach of adding notes in the Documentation is beyond brittle, e.g. there is at least one known case where a KVM developer added a new userspace exit type, and then that same developer forgot to handle completion when adding userspace support.
Quoted 5 lines
This answers one question I had:
https://lore.kernel.org/kvm/[email protected]/So, it is the VMM's (i.e., QEMU's) responsibility to re-execute
KVM_RUNin this case.
Yep.
Quoted 5 lines
Btw, can this flag be used to address the issue [*] with steal time accounting? We can set the new flag for each vCPU in the PM notifier and we need to change the re-execution to handle steal time accounting (not just IO completion).
Uh, hmm. Partially? And not without creating new, potentially worse problems.
I like the idea, but (a) there's no guarantee a vCPU would be "in" KVM_RUN at the time of suspend, and (b) KVM would need to take vcpu->mutex in the PM notifier in order to avoid clobbering the current completion callback, which is definitely a net negative (hello, deadlocks).
E.g. if a vCPU task is in userspace processing emulated MMIO at the time of suspend+resume, KVM's completion callback will be non-zero and must be preserved. And if a vCPU task is in userspace processing an exit that _doesn't_ require completion, setting KVM_RUN_NEEDS_COMPLETION would likely be missed by userspace, e.g. if userspace checks the flag only after regaining control from KVM_RUN.
In general, I think setting KVM_RUN_NEEDS_COMPLETION outside of KVM_RUN would add too much complexity.
Quoted 8 lines
one nit below,
Quoted 6 lines
arch/x86/include/uapi/asm/kvm.h +0
-0@@ -104,9 +104,10 @@ struct kvm_ioapic_state {#define KVM_IRQCHIP_IOAPIC 2#define KVM_NR_IRQCHIPS 3
Quoted 9 lines
arch/x86/include/uapi/asm/kvm.h struct kvm_ioapic_state { · line 110
Quoted 7 lines
arch/x86/include/uapi/asm/kvm.h +4
-3-#define KVM_RUN_X86_SMM (1 << 0) -#define KVM_RUN_X86_BUS_LOCK (1 << 1) -#define KVM_RUN_X86_GUEST_MODE (1 << 2) +#define KVM_RUN_X86_SMM (1 << 0) +#define KVM_RUN_X86_BUS_LOCK (1 << 1) +#define KVM_RUN_X86_GUEST_MODE (1 << 2) +#define KVM_RUN_X86_NEEDS_COMPLETION (1 << 2)This
X86_NEEDS_COMPLETIONshould be dropped. It is never used.
Gah, thanks!
Sean Christopherson
maintainerPATCH 4/5 KVM: selftests: Provide separate helper for KVM_RUN with immediate_exit
In reply to Sean ChristophersonTurn vcpu_run_complete_io() into a wrapper for a dedicated helper for doing KVM_RUN with immediate_exit = true, so that a future patch can do userspace exit completion if and only if it's actually necessary, whereas x86's nested exceptions test wants to unconditionally do KVM_RUN with an immediate exit.
No functional change intended.
--- tools/testing/selftests/kvm/include/kvm_util.h | 9 ++++++++- tools/testing/selftests/kvm/lib/kvm_util.c | 2 +- tools/testing/selftests/kvm/x86/nested_exceptions_test.c | 3 +-- 3 files changed, 10 insertions(+), 4 deletions(-)
tools/testing/selftests/kvm/include/kvm_util.h +8 -1
@@ -631,7 +631,14 @@ static inline int __vcpu_run(struct kvm_vcpu *vcpu)return __vcpu_ioctl (vcpu,KVM_RUN ,NULL ); } -void vcpu_run_complete_io (struct kvm_vcpu *vcpu); + +void vcpu_run_immediate_exit (struct kvm_vcpu *vcpu); + +static inline void vcpu_run_complete_io (struct kvm_vcpu *vcpu) +{ +vcpu_run_immediate_exit (vcpu); +} +struct kvm_reg_list *vcpu_get_reg_list (struct kvm_vcpu *vcpu); static inline void vcpu_enable_cap(struct kvm_vcpu *vcpu, uint32_t cap,
tools/testing/selftests/kvm/lib/kvm_util.c +1 -1
@@ -1665,7 +1665,7 @@ void vcpu_run(struct kvm_vcpu *vcpu)TEST_ASSERT (!ret,KVM_IOCTL_ERROR (KVM_RUN , ret)); } -void vcpu_run_complete_io (struct kvm_vcpu *vcpu) +void vcpu_run_immediate_exit (struct kvm_vcpu *vcpu) {int ret;
tools/testing/selftests/kvm/x86/nested_exceptions_test.c +1 -2
@@ -238,8 +238,7 @@ int main(int argc, char *argv[])/* Pend #SS and request immediate exit. #SS should still be pending. */ queue_ss_exception (vcpu, false); - vcpu->run->immediate_exit = true; -vcpu_run_complete_io (vcpu); +vcpu_run_immediate_exit (vcpu);/* Verify the pending events comes back out the same as it went in. */ vcpu_events_get (vcpu, &events);
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
Sean Christopherson
maintainerPATCH 5/5 KVM: selftests: Rely on KVM_RUN_NEEDS_COMPLETION to complete userspace exits
In reply to Sean ChristophersonAdd selftests coverage for KVM_RUN_NEEDS_COMPLETION by redoing KVM_RUN if and only if KVM states that completion is required.
Opportunistically rename the helper to replace "io" with "exit", as exits that require completion are no longer limited to I/O.
--- tools/testing/selftests/kvm/include/kvm_util.h | 8 ++++++-- tools/testing/selftests/kvm/lib/kvm_util.c | 4 ++++ tools/testing/selftests/kvm/lib/ucall_common.c | 2 +- tools/testing/selftests/kvm/lib/x86/processor.c | 8 +------- tools/testing/selftests/kvm/x86/triple_fault_event_test.c | 3 +-- 5 files changed, 13 insertions(+), 12 deletions(-)
tools/testing/selftests/kvm/include/kvm_util.h +6 -2
@@ -29,6 +29,8 @@#define NSEC_PER_SEC 1000000000L +extern bool kvm_has_needs_completion; +struct userspace_mem_region {struct kvm_userspace_memory_region2 region;struct sparsebit *unused_phy_pages;
@@ -634,9 +636,11 @@ static inline int __vcpu_run(struct kvm_vcpu *vcpu)void vcpu_run_immediate_exit (struct kvm_vcpu *vcpu); -static inline void vcpu_run_complete_io (struct kvm_vcpu *vcpu) +static inline void vcpu_run_complete_exit (struct kvm_vcpu *vcpu) { -vcpu_run_immediate_exit (vcpu); +if (!kvm_has_needs_completion || + (vcpu->run->flags &KVM_RUN_NEEDS_COMPLETION )) +vcpu_run_immediate_exit (vcpu); }struct kvm_reg_list *vcpu_get_reg_list (struct kvm_vcpu *vcpu);
tools/testing/selftests/kvm/lib/kvm_util.c +4 -0
@@ -19,6 +19,8 @@#define KVM_UTIL_MIN_PFN 2 +bool kvm_has_needs_completion; +uint32_t guest_random_seed;struct guest_random_state guest_rng;static uint32_t last_guest_seed;
@@ -2253,6 +2255,8 @@ void __attribute((constructor)) kvm_selftest_init(void)/* Tell stdout not to buffer its content. */ setbuf (stdout,NULL ); + kvm_has_needs_completion =kvm_check_cap (KVM_CAP_NEEDS_COMPLETION ); + guest_random_seed = last_guest_seed =random ();pr_info ("Random seed: 0x%x\n" , guest_random_seed);
tools/testing/selftests/kvm/lib/ucall_common.c +1 -1
@@ -154,7 +154,7 @@ uint64_t get_ucall(struct kvm_vcpu *vcpu, struct ucall *uc) "Guest failed to allocate ucall struct");memcpy (uc, addr,sizeof (*uc)); -vcpu_run_complete_io (vcpu); +vcpu_run_complete_exit (vcpu); }else {memset (uc,0 ,sizeof (*uc)); }
tools/testing/selftests/kvm/lib/x86/processor.c +1 -7
@@ -1077,13 +1077,7 @@ struct kvm_x86_state *vcpu_save_state(struct kvm_vcpu *vcpu) nested_size, sizeof(state->nested_)); } -/* -* When KVM exits to userspace with KVM_EXIT_IO, KVM guarantees -* guest state is consistent only after userspace re-enters the -* kernel with KVM_RUN. Complete IO prior to migrating state -* to a new VM. -*/ -vcpu_run_complete_io (vcpu); +vcpu_run_complete_exit (vcpu); state =malloc (sizeof (*state) + msr_list->nmsrs *sizeof (state->msrs.entries[0 ]));TEST_ASSERT (state,"-ENOMEM when allocating kvm state" );
tools/testing/selftests/kvm/x86/triple_fault_event_test.c +1 -2
@@ -97,8 +97,7 @@ int main(void) events.flags |=KVM_VCPUEVENT_VALID_TRIPLE_FAULT ; events.triple_fault.pending = true;vcpu_events_set (vcpu, &events); - run->immediate_exit = true; -vcpu_run_complete_io (vcpu); +vcpu_run_complete_exit (vcpu);vcpu_events_get (vcpu, &events); TEST_ASSERT(events.flags & KVM_VCPUEVENT_VALID_TRIPLE_FAULT,
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog