| Message ID | 1474548769-195643-1-git-send-email-agraf@suse.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Sep 22, 2016 at 02:52:49PM +0200, Alexander Graf wrote: > We have 2 modes for dealing with interrupts in the ARM world. We can either > handle them all using hardware acceleration through the vgic or we can emulate > a gic in user space and only drive CPU IRQ pins from there. > > Unfortunately, when driving IRQs from user space, we never tell user space > about timer events that may result in interrupt line state changes, so we > lose out on timer events if we run with user space gic emulation. > > This patch fixes that by syncing user space's view of the vtimer irq line > with the kvm view of that same line. > > With this patch I can successfully run edk2 and Linux with user space gic > emulation. > > Signed-off-by: Alexander Graf <agraf@suse.de> > > --- > > v1 -> v2: > > - Add back curly brace that got lost > > v2 -> v3: > > - Split into patch set > > v3 -> v4: > > - Improve documentation > > v4 -> v5: > > - Rewrite to use pending state sync in sregs (marc) > - Remove redundant checks of vgic_initialized() > - qemu tree to try this out: https://github.com/agraf/u-boot.git no-kvm-irqchip-for-v5 huh, qemu=u-boot? > --- > Documentation/virtual/kvm/api.txt | 26 ++++++++ > arch/arm/include/uapi/asm/kvm.h | 3 + > arch/arm/kvm/arm.c | 14 ++--- > arch/arm64/include/uapi/asm/kvm.h | 3 + > include/kvm/arm_arch_timer.h | 2 +- > include/uapi/linux/kvm.h | 6 ++ > virt/kvm/arm/arch_timer.c | 129 ++++++++++++++++++++++++++------------ > 7 files changed, 134 insertions(+), 49 deletions(-) > > diff --git a/Documentation/virtual/kvm/api.txt b/Documentation/virtual/kvm/api.txt > index 739db9a..8049327 100644 > --- a/Documentation/virtual/kvm/api.txt > +++ b/Documentation/virtual/kvm/api.txt > @@ -3928,3 +3928,29 @@ In order to use SynIC, it has to be activated by setting this > capability via KVM_ENABLE_CAP ioctl on the vcpu fd. Note that this > will disable the use of APIC hardware virtualization even if supported > by the CPU, as it's incompatible with SynIC auto-EOI behavior. > + > +8.3 KVM_CAP_ARM_TIMER > + > +Architectures: arm, arm64 > +This capability, if KVM_CHECK_EXTENSION indicates that it is available and no > +in-kernel interrupt controller is in use, means that that the kernel populates > +the vcpu's run->s.regs.kernel_timer_pending field with timers that are currently > +considered pending by kvm. Be careful with the word 'pending' here. I think this could be misleading, because pending is a state in the GIC, but not really something I can find specific to the timer. It would be more descriptive to say that the kernel maintained generic timer's output signal is asserted. > + > +If active, it also allows user space to propagate its own pending state of timer > +interrupt lines using run->s.regs.user_timer_pending. If those two fields > +mismatch during CPU execution, kvm will exit to user space to give it a chance I don't quite understand the semantics here. The only entity that knows what the level state of the output of the timer is, is the kernel, which emulates the timer. Userspace knows interrupt controller state, but if it has a different view of the timer state than the kernel, it's because the kernel failed to notify userspace of a change or userspace failed to listen? > +to update its own interrupt pending status. This usually involves triggering > +an interrupt line on a user space emulated interrupt controller. To me it feels like the semantics should be that userspace can always derive the status of the timer and the level of the output signal from the timer by simply looking at kvm_run structure. The remaining two problems are: (1) when should the kernel trigger exits to userspace? Presumably on any change in the timer's output level, because this change has to be propagated to the userspace interrupt controller. (2) the kernel needs to somehow mask the underlying hardware timer interrupt signal when it's active, because otherwise the guest won't proceed. If we simply mask the hardware signal after telling userspace the output signal is asserted and until the output signal ever becomes deasserted, why do we need to listen to anything userspace has to say? > + > +The fields run->s.regs.kernel_timer_pending and run->s.regs.user_timer_pending > +are available independent of run->kvm_valid_regs or run->kvm_dirty_regs bits. > +If no in-kernel interrupt controller is used and the capability exists, they > +will always be available and used. > + > +Currently the following bits are defined for both bitmaps: > + > + KVM_ARM_TIMER_VTIMER - virtual timer > + > +Future versions of kvm may implement additional timer events. These will get > +indicated by additional KVM_CAP extensions. > diff --git a/arch/arm/include/uapi/asm/kvm.h b/arch/arm/include/uapi/asm/kvm.h > index a2b3eb3..caad81d 100644 > --- a/arch/arm/include/uapi/asm/kvm.h > +++ b/arch/arm/include/uapi/asm/kvm.h > @@ -105,6 +105,9 @@ struct kvm_debug_exit_arch { > }; > > struct kvm_sync_regs { > + /* Used with KVM_CAP_ARM_TIMER */ > + u8 kernel_timer_pending; > + u8 user_timer_pending; > }; > > struct kvm_arch_memory_slot { > diff --git a/arch/arm/kvm/arm.c b/arch/arm/kvm/arm.c > index 75f130e..dc19221 100644 > --- a/arch/arm/kvm/arm.c > +++ b/arch/arm/kvm/arm.c > @@ -187,6 +187,7 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext) > case KVM_CAP_ARM_PSCI_0_2: > case KVM_CAP_READONLY_MEM: > case KVM_CAP_MP_STATE: > + case KVM_CAP_ARM_TIMER: > r = 1; > break; > case KVM_CAP_COALESCED_MMIO: > @@ -474,13 +475,7 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu) > return ret; > } > > - /* > - * Enable the arch timers only if we have an in-kernel VGIC > - * and it has been properly initialized, since we cannot handle > - * interrupts from the virtual timer with a userspace gic. > - */ > - if (irqchip_in_kernel(kvm) && vgic_initialized(kvm)) > - ret = kvm_timer_enable(vcpu); > + ret = kvm_timer_enable(vcpu); > > return ret; > } > @@ -588,7 +583,10 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run) > */ > preempt_disable(); > kvm_pmu_flush_hwstate(vcpu); > - kvm_timer_flush_hwstate(vcpu); > + if (kvm_timer_flush_hwstate(vcpu)) { > + ret = -EINTR; > + run->exit_reason = KVM_EXIT_INTR; > + } > kvm_vgic_flush_hwstate(vcpu); > > local_irq_disable(); > diff --git a/arch/arm64/include/uapi/asm/kvm.h b/arch/arm64/include/uapi/asm/kvm.h > index 3051f86..9aac860 100644 > --- a/arch/arm64/include/uapi/asm/kvm.h > +++ b/arch/arm64/include/uapi/asm/kvm.h > @@ -143,6 +143,9 @@ struct kvm_debug_exit_arch { > #define KVM_GUESTDBG_USE_HW (1 << 17) > > struct kvm_sync_regs { > + /* Used with KVM_CAP_ARM_TIMER */ > + u8 kernel_timer_pending; > + u8 user_timer_pending; > }; > > struct kvm_arch_memory_slot { > diff --git a/include/kvm/arm_arch_timer.h b/include/kvm/arm_arch_timer.h > index dda39d8..8cd7240 100644 > --- a/include/kvm/arm_arch_timer.h > +++ b/include/kvm/arm_arch_timer.h > @@ -63,7 +63,7 @@ void kvm_timer_init(struct kvm *kvm); > int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu, > const struct kvm_irq_level *irq); > void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu); > -void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu); > +int kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu); > void kvm_timer_sync_hwstate(struct kvm_vcpu *vcpu); > void kvm_timer_vcpu_terminate(struct kvm_vcpu *vcpu); > > diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h > index 300ef25..1fc02d7 100644 > --- a/include/uapi/linux/kvm.h > +++ b/include/uapi/linux/kvm.h > @@ -870,6 +870,7 @@ struct kvm_ppc_smmu_info { > #define KVM_CAP_S390_USER_INSTR0 130 > #define KVM_CAP_MSI_DEVID 131 > #define KVM_CAP_PPC_HTM 132 > +#define KVM_CAP_ARM_TIMER 133 > > #ifdef KVM_CAP_IRQ_ROUTING > > @@ -1327,4 +1328,9 @@ struct kvm_assigned_msix_entry { > #define KVM_X2APIC_API_USE_32BIT_IDS (1ULL << 0) > #define KVM_X2APIC_API_DISABLE_BROADCAST_QUIRK (1ULL << 1) > > +/* Available with KVM_CAP_ARM_TIMER */ > + > +/* Bits for run->s.regs.{user,kernel}_timer_pending */ > +#define KVM_ARM_TIMER_VTIMER (1 << 0) > + > #endif /* __LINUX_KVM_H */ > diff --git a/virt/kvm/arm/arch_timer.c b/virt/kvm/arm/arch_timer.c > index 4309b60..0c6fc38 100644 > --- a/virt/kvm/arm/arch_timer.c > +++ b/virt/kvm/arm/arch_timer.c > @@ -166,21 +166,36 @@ bool kvm_timer_should_fire(struct kvm_vcpu *vcpu) > return cval <= now; > } > > +/* > + * Synchronize the timer IRQ state with the interrupt controller. > + */ > static void kvm_timer_update_irq(struct kvm_vcpu *vcpu, bool new_level) > { > int ret; > struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu; > > - BUG_ON(!vgic_initialized(vcpu->kvm)); > - > timer->active_cleared_last = false; > timer->irq.level = new_level; > - trace_kvm_timer_update_irq(vcpu->vcpu_id, timer->irq.irq, > + trace_kvm_timer_update_irq(vcpu->vcpu_id, host_vtimer_irq, > timer->irq.level); > - ret = kvm_vgic_inject_mapped_irq(vcpu->kvm, vcpu->vcpu_id, > - timer->irq.irq, > - timer->irq.level); > - WARN_ON(ret); > + > + if (irqchip_in_kernel(vcpu->kvm)) { > + BUG_ON(!vgic_initialized(vcpu->kvm)); > + > + /* Fire the timer in the VGIC */ > + ret = kvm_vgic_inject_mapped_irq(vcpu->kvm, vcpu->vcpu_id, > + timer->irq.irq, > + timer->irq.level); > + > + WARN_ON(ret); > + } else { > + struct kvm_sync_regs *regs = &vcpu->run->s.regs; > + > + /* Populate the timer bitmap for user space */ > + regs->kernel_timer_pending &= ~KVM_ARM_TIMER_VTIMER; > + if (new_level) > + regs->kernel_timer_pending |= KVM_ARM_TIMER_VTIMER; I think if you got here, it means you have to exit to userspace to update it of the new state. If you don't want to propagate a return value from here, I think you should just not do anything an then later compare timer->irq.level with whatever was last written to run->kernel_timer_pending (which should be named something else than pending). > + } > } > > /* > @@ -197,7 +212,8 @@ static int kvm_timer_update_state(struct kvm_vcpu *vcpu) > * because the guest would never see the interrupt. Instead wait > * until we call this function from kvm_timer_flush_hwstate. > */ > - if (!vgic_initialized(vcpu->kvm) || !timer->enabled) > + if ((irqchip_in_kernel(vcpu->kvm) && !vgic_initialized(vcpu->kvm)) || > + !timer->enabled) > return -ENODEV; > > if (kvm_timer_should_fire(vcpu) != timer->irq.level) > @@ -248,15 +264,20 @@ void kvm_timer_unschedule(struct kvm_vcpu *vcpu) > * > * Check if the virtual timer has expired while we were running in the host, > * and inject an interrupt if that was the case. > + * > + * Returns: > + * > + * 0 - success > + * 1 - need exit to user space this is opposite to all other exit-related APIs we have. Why not just return -EINTR? > */ > -void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu) > +int kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu) > { > struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu; > bool phys_active; > int ret; > > if (kvm_timer_update_state(vcpu)) > - return; > + return 0; > > /* > * If we enter the guest with the virtual input level to the VGIC > @@ -275,38 +296,61 @@ void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu) > * to ensure that hardware interrupts from the timer triggers a guest > * exit. > */ > - phys_active = timer->irq.level || > - kvm_vgic_map_is_active(vcpu, timer->irq.irq); > - > - /* > - * We want to avoid hitting the (re)distributor as much as > - * possible, as this is a potentially expensive MMIO access > - * (not to mention locks in the irq layer), and a solution for > - * this is to cache the "active" state in memory. > - * > - * Things to consider: we cannot cache an "active set" state, > - * because the HW can change this behind our back (it becomes > - * "clear" in the HW). We must then restrict the caching to > - * the "clear" state. > - * > - * The cache is invalidated on: > - * - vcpu put, indicating that the HW cannot be trusted to be > - * in a sane state on the next vcpu load, > - * - any change in the interrupt state > - * > - * Usage conditions: > - * - cached value is "active clear" > - * - value to be programmed is "active clear" > - */ > - if (timer->active_cleared_last && !phys_active) > - return; > - > - ret = irq_set_irqchip_state(host_vtimer_irq, > - IRQCHIP_STATE_ACTIVE, > - phys_active); > - WARN_ON(ret); > + if (irqchip_in_kernel(vcpu->kvm)) { > + phys_active = timer->irq.level || > + kvm_vgic_map_is_active(vcpu, timer->irq.irq); > + > + /* > + * We want to avoid hitting the (re)distributor as much as > + * possible, as this is a potentially expensive MMIO access > + * (not to mention locks in the irq layer), and a solution for > + * this is to cache the "active" state in memory. > + * > + * Things to consider: we cannot cache an "active set" state, > + * because the HW can change this behind our back (it becomes > + * "clear" in the HW). We must then restrict the caching to > + * the "clear" state. > + * > + * The cache is invalidated on: > + * - vcpu put, indicating that the HW cannot be trusted to be > + * in a sane state on the next vcpu load, > + * - any change in the interrupt state > + * > + * Usage conditions: > + * - cached value is "active clear" > + * - value to be programmed is "active clear" > + */ > + if (timer->active_cleared_last && !phys_active) > + return 0; > + > + ret = irq_set_irqchip_state(host_vtimer_irq, > + IRQCHIP_STATE_ACTIVE, > + phys_active); > + WARN_ON(ret); > + } else { > + struct kvm_sync_regs *regs = &vcpu->run->s.regs; > + > + /* > + * User space handles timer events, so we need to check whether > + * its view of the world is in sync with ours. > + */ > + if (regs->kernel_timer_pending != regs->user_timer_pending) { > + /* Return to user space */ > + return 1; > + } Maybe I'm misunderstanding and user_timer_pending is just a cached verison of what you said last, but as I said above, I think you can just compare timer->irq.level with the last value the kvm_run struct, and if something changed, you have to exit. > + > + /* > + * As long as user space is aware that the timer is pending, > + * we do not need to get new host timer events. > + */ yes, correct, but I don't think this concept was clearly reflected in your API text above. > + if (timer->irq.level) > + disable_percpu_irq(host_vtimer_irq); > + else > + enable_percpu_irq(host_vtimer_irq, 0); > + } could we move these two blocks into their own functions instead? That would also give nice names to the huge chunk of complicated functionality, e.g. flush_timer_state_to_user() and flush_timer_state_to_vgic(). > > timer->active_cleared_last = !phys_active; > + return 0; > } > > /** > @@ -479,6 +523,10 @@ int kvm_timer_enable(struct kvm_vcpu *vcpu) > if (timer->enabled) > return 0; > > + /* No need to route physical IRQs when we don't use the vgic */ > + if (!irqchip_in_kernel(vcpu->kvm)) > + goto no_vgic; > + > /* > * Find the physical IRQ number corresponding to the host_vtimer_irq > */ > @@ -502,6 +550,7 @@ int kvm_timer_enable(struct kvm_vcpu *vcpu) > if (ret) > return ret; > > +no_vgic: > > /* > * There is a potential race here between VCPUs starting for the first > -- > 1.8.5.6 > Thanks, -Christoffer
On 22.09.16 23:28, Christoffer Dall wrote: > On Thu, Sep 22, 2016 at 02:52:49PM +0200, Alexander Graf wrote: >> We have 2 modes for dealing with interrupts in the ARM world. We can either >> handle them all using hardware acceleration through the vgic or we can emulate >> a gic in user space and only drive CPU IRQ pins from there. >> >> Unfortunately, when driving IRQs from user space, we never tell user space >> about timer events that may result in interrupt line state changes, so we >> lose out on timer events if we run with user space gic emulation. >> >> This patch fixes that by syncing user space's view of the vtimer irq line >> with the kvm view of that same line. >> >> With this patch I can successfully run edk2 and Linux with user space gic >> emulation. >> >> Signed-off-by: Alexander Graf <agraf@suse.de> >> >> --- >> >> v1 -> v2: >> >> - Add back curly brace that got lost >> >> v2 -> v3: >> >> - Split into patch set >> >> v3 -> v4: >> >> - Improve documentation >> >> v4 -> v5: >> >> - Rewrite to use pending state sync in sregs (marc) >> - Remove redundant checks of vgic_initialized() >> - qemu tree to try this out: https://github.com/agraf/u-boot.git no-kvm-irqchip-for-v5 > > huh, qemu=u-boot? Bleks, qemu.git of course. > >> --- >> Documentation/virtual/kvm/api.txt | 26 ++++++++ >> arch/arm/include/uapi/asm/kvm.h | 3 + >> arch/arm/kvm/arm.c | 14 ++--- >> arch/arm64/include/uapi/asm/kvm.h | 3 + >> include/kvm/arm_arch_timer.h | 2 +- >> include/uapi/linux/kvm.h | 6 ++ >> virt/kvm/arm/arch_timer.c | 129 ++++++++++++++++++++++++++------------ >> 7 files changed, 134 insertions(+), 49 deletions(-) >> >> diff --git a/Documentation/virtual/kvm/api.txt b/Documentation/virtual/kvm/api.txt >> index 739db9a..8049327 100644 >> --- a/Documentation/virtual/kvm/api.txt >> +++ b/Documentation/virtual/kvm/api.txt >> @@ -3928,3 +3928,29 @@ In order to use SynIC, it has to be activated by setting this >> capability via KVM_ENABLE_CAP ioctl on the vcpu fd. Note that this >> will disable the use of APIC hardware virtualization even if supported >> by the CPU, as it's incompatible with SynIC auto-EOI behavior. >> + >> +8.3 KVM_CAP_ARM_TIMER >> + >> +Architectures: arm, arm64 >> +This capability, if KVM_CHECK_EXTENSION indicates that it is available and no >> +in-kernel interrupt controller is in use, means that that the kernel populates >> +the vcpu's run->s.regs.kernel_timer_pending field with timers that are currently >> +considered pending by kvm. > > Be careful with the word 'pending' here. I think this could be > misleading, because pending is a state in the GIC, but not really > something I can find specific to the timer. It would be more > descriptive to say that the kernel maintained generic timer's output > signal is asserted. Sure, asserted works for me. Or maybe istatus? > >> + >> +If active, it also allows user space to propagate its own pending state of timer >> +interrupt lines using run->s.regs.user_timer_pending. If those two fields >> +mismatch during CPU execution, kvm will exit to user space to give it a chance > > I don't quite understand the semantics here. The only entity that knows > what the level state of the output of the timer is, is the kernel, which > emulates the timer. Userspace knows interrupt controller state, but if > it has a different view of the timer state than the kernel, it's because > the kernel failed to notify userspace of a change or userspace failed to > listen? Right, and the reason we have 2 fields is to get us exit-less (and easier) updates whenever we can. In most cases, taking the assertion down will coincide with an MMIO exit to user space for an EOI for example. Somewhere in between in the development I had a version that explicitly triggered a KVM_EXIT for every state change of the timer. But that gets messy very quickly. You need to update the state change before an MMIO for example, otherwise a sequence like * set cval to future (which in turns sets istatus=0) * read gic pending state will give you bogus results, as the mmio read to user space did not yet have the timer status updated. And I really didn't want to wrap my head around restarting MMIO exits ;). So having this side channel where user space potentially expects a timer status change on every exit is much cleaner. > >> +to update its own interrupt pending status. This usually involves triggering >> +an interrupt line on a user space emulated interrupt controller. > > To me it feels like the semantics should be that userspace can always > derive the status of the timer and the level of the output signal from > the timer by simply looking at kvm_run structure. Yes, it can, and it does. That's what the "kernel_timer_pending" field is. The kernel however also needs to know whether user space's view is in sync, because we don't know whether there was an exit between our internal state change and the guest entry. That's what the "user_timer_pending" field is for. > > The remaining two problems are: > > (1) when should the kernel trigger exits to userspace? Presumably on > any change in the timer's output level, because this change has to be > propagated to the userspace interrupt controller. Yes, but if we can we want to piggy-back on an existing exit. > (2) the kernel needs to somehow mask the underlying hardware timer > interrupt signal when it's active, because otherwise the guest won't > proceed. If we simply mask the hardware signal after telling userspace > the output signal is asserted and until the output signal ever becomes > deasserted, why do we need to listen to anything userspace has to say? Because we need to make sure that the cpu IRQ line is updated before we enter the guest. Otherwise we might get spurious interrupts. >> + >> +The fields run->s.regs.kernel_timer_pending and run->s.regs.user_timer_pending >> +are available independent of run->kvm_valid_regs or run->kvm_dirty_regs bits. >> +If no in-kernel interrupt controller is used and the capability exists, they >> +will always be available and used. >> + >> +Currently the following bits are defined for both bitmaps: >> + >> + KVM_ARM_TIMER_VTIMER - virtual timer >> + >> +Future versions of kvm may implement additional timer events. These will get >> +indicated by additional KVM_CAP extensions. >> diff --git a/arch/arm/include/uapi/asm/kvm.h b/arch/arm/include/uapi/asm/kvm.h >> index a2b3eb3..caad81d 100644 >> --- a/arch/arm/include/uapi/asm/kvm.h >> +++ b/arch/arm/include/uapi/asm/kvm.h >> @@ -105,6 +105,9 @@ struct kvm_debug_exit_arch { >> }; >> >> struct kvm_sync_regs { >> + /* Used with KVM_CAP_ARM_TIMER */ >> + u8 kernel_timer_pending; >> + u8 user_timer_pending; >> }; >> >> struct kvm_arch_memory_slot { >> diff --git a/arch/arm/kvm/arm.c b/arch/arm/kvm/arm.c >> index 75f130e..dc19221 100644 >> --- a/arch/arm/kvm/arm.c >> +++ b/arch/arm/kvm/arm.c >> @@ -187,6 +187,7 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext) >> case KVM_CAP_ARM_PSCI_0_2: >> case KVM_CAP_READONLY_MEM: >> case KVM_CAP_MP_STATE: >> + case KVM_CAP_ARM_TIMER: >> r = 1; >> break; >> case KVM_CAP_COALESCED_MMIO: >> @@ -474,13 +475,7 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu) >> return ret; >> } >> >> - /* >> - * Enable the arch timers only if we have an in-kernel VGIC >> - * and it has been properly initialized, since we cannot handle >> - * interrupts from the virtual timer with a userspace gic. >> - */ >> - if (irqchip_in_kernel(kvm) && vgic_initialized(kvm)) >> - ret = kvm_timer_enable(vcpu); >> + ret = kvm_timer_enable(vcpu); >> >> return ret; >> } >> @@ -588,7 +583,10 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run) >> */ >> preempt_disable(); >> kvm_pmu_flush_hwstate(vcpu); >> - kvm_timer_flush_hwstate(vcpu); >> + if (kvm_timer_flush_hwstate(vcpu)) { >> + ret = -EINTR; >> + run->exit_reason = KVM_EXIT_INTR; >> + } >> kvm_vgic_flush_hwstate(vcpu); >> >> local_irq_disable(); >> diff --git a/arch/arm64/include/uapi/asm/kvm.h b/arch/arm64/include/uapi/asm/kvm.h >> index 3051f86..9aac860 100644 >> --- a/arch/arm64/include/uapi/asm/kvm.h >> +++ b/arch/arm64/include/uapi/asm/kvm.h >> @@ -143,6 +143,9 @@ struct kvm_debug_exit_arch { >> #define KVM_GUESTDBG_USE_HW (1 << 17) >> >> struct kvm_sync_regs { >> + /* Used with KVM_CAP_ARM_TIMER */ >> + u8 kernel_timer_pending; >> + u8 user_timer_pending; >> }; >> >> struct kvm_arch_memory_slot { >> diff --git a/include/kvm/arm_arch_timer.h b/include/kvm/arm_arch_timer.h >> index dda39d8..8cd7240 100644 >> --- a/include/kvm/arm_arch_timer.h >> +++ b/include/kvm/arm_arch_timer.h >> @@ -63,7 +63,7 @@ void kvm_timer_init(struct kvm *kvm); >> int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu, >> const struct kvm_irq_level *irq); >> void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu); >> -void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu); >> +int kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu); >> void kvm_timer_sync_hwstate(struct kvm_vcpu *vcpu); >> void kvm_timer_vcpu_terminate(struct kvm_vcpu *vcpu); >> >> diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h >> index 300ef25..1fc02d7 100644 >> --- a/include/uapi/linux/kvm.h >> +++ b/include/uapi/linux/kvm.h >> @@ -870,6 +870,7 @@ struct kvm_ppc_smmu_info { >> #define KVM_CAP_S390_USER_INSTR0 130 >> #define KVM_CAP_MSI_DEVID 131 >> #define KVM_CAP_PPC_HTM 132 >> +#define KVM_CAP_ARM_TIMER 133 >> >> #ifdef KVM_CAP_IRQ_ROUTING >> >> @@ -1327,4 +1328,9 @@ struct kvm_assigned_msix_entry { >> #define KVM_X2APIC_API_USE_32BIT_IDS (1ULL << 0) >> #define KVM_X2APIC_API_DISABLE_BROADCAST_QUIRK (1ULL << 1) >> >> +/* Available with KVM_CAP_ARM_TIMER */ >> + >> +/* Bits for run->s.regs.{user,kernel}_timer_pending */ >> +#define KVM_ARM_TIMER_VTIMER (1 << 0) >> + >> #endif /* __LINUX_KVM_H */ >> diff --git a/virt/kvm/arm/arch_timer.c b/virt/kvm/arm/arch_timer.c >> index 4309b60..0c6fc38 100644 >> --- a/virt/kvm/arm/arch_timer.c >> +++ b/virt/kvm/arm/arch_timer.c >> @@ -166,21 +166,36 @@ bool kvm_timer_should_fire(struct kvm_vcpu *vcpu) >> return cval <= now; >> } >> >> +/* >> + * Synchronize the timer IRQ state with the interrupt controller. >> + */ >> static void kvm_timer_update_irq(struct kvm_vcpu *vcpu, bool new_level) >> { >> int ret; >> struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu; >> >> - BUG_ON(!vgic_initialized(vcpu->kvm)); >> - >> timer->active_cleared_last = false; >> timer->irq.level = new_level; >> - trace_kvm_timer_update_irq(vcpu->vcpu_id, timer->irq.irq, >> + trace_kvm_timer_update_irq(vcpu->vcpu_id, host_vtimer_irq, >> timer->irq.level); >> - ret = kvm_vgic_inject_mapped_irq(vcpu->kvm, vcpu->vcpu_id, >> - timer->irq.irq, >> - timer->irq.level); >> - WARN_ON(ret); >> + >> + if (irqchip_in_kernel(vcpu->kvm)) { >> + BUG_ON(!vgic_initialized(vcpu->kvm)); >> + >> + /* Fire the timer in the VGIC */ >> + ret = kvm_vgic_inject_mapped_irq(vcpu->kvm, vcpu->vcpu_id, >> + timer->irq.irq, >> + timer->irq.level); >> + >> + WARN_ON(ret); >> + } else { >> + struct kvm_sync_regs *regs = &vcpu->run->s.regs; >> + >> + /* Populate the timer bitmap for user space */ >> + regs->kernel_timer_pending &= ~KVM_ARM_TIMER_VTIMER; >> + if (new_level) >> + regs->kernel_timer_pending |= KVM_ARM_TIMER_VTIMER; > > I think if you got here, it means you have to exit to userspace to > update it of the new state. If you don't want to propagate a return Yes, but we can't exit straight away with our own exit reason because we might be inside an MMIO exit path here which already occupies the exit_reason. > value from here, I think you should just not do anything an then later > compare timer->irq.level with whatever was last written to > run->kernel_timer_pending (which should be named something else than > pending). Hm, so you're saying we should just update kernel_timer_pending in flush_hwstate()? That way we miss out on the piggy backing, no? > >> + } >> } >> >> /* >> @@ -197,7 +212,8 @@ static int kvm_timer_update_state(struct kvm_vcpu *vcpu) >> * because the guest would never see the interrupt. Instead wait >> * until we call this function from kvm_timer_flush_hwstate. >> */ >> - if (!vgic_initialized(vcpu->kvm) || !timer->enabled) >> + if ((irqchip_in_kernel(vcpu->kvm) && !vgic_initialized(vcpu->kvm)) || >> + !timer->enabled) >> return -ENODEV; >> >> if (kvm_timer_should_fire(vcpu) != timer->irq.level) >> @@ -248,15 +264,20 @@ void kvm_timer_unschedule(struct kvm_vcpu *vcpu) >> * >> * Check if the virtual timer has expired while we were running in the host, >> * and inject an interrupt if that was the case. >> + * >> + * Returns: >> + * >> + * 0 - success >> + * 1 - need exit to user space > > this is opposite to all other exit-related APIs we have. Why not just > return -EINTR? All functions in the arch timer file are in normal C API notion (0=success). I actually had it all in exit-related API notion at first and it became very awkward and hard to read to convert exit codes in between. Doing it just once in kvm.c felt much cleaner. > >> */ >> -void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu) >> +int kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu) >> { >> struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu; >> bool phys_active; >> int ret; >> >> if (kvm_timer_update_state(vcpu)) >> - return; >> + return 0; >> >> /* >> * If we enter the guest with the virtual input level to the VGIC >> @@ -275,38 +296,61 @@ void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu) >> * to ensure that hardware interrupts from the timer triggers a guest >> * exit. >> */ >> - phys_active = timer->irq.level || >> - kvm_vgic_map_is_active(vcpu, timer->irq.irq); >> - >> - /* >> - * We want to avoid hitting the (re)distributor as much as >> - * possible, as this is a potentially expensive MMIO access >> - * (not to mention locks in the irq layer), and a solution for >> - * this is to cache the "active" state in memory. >> - * >> - * Things to consider: we cannot cache an "active set" state, >> - * because the HW can change this behind our back (it becomes >> - * "clear" in the HW). We must then restrict the caching to >> - * the "clear" state. >> - * >> - * The cache is invalidated on: >> - * - vcpu put, indicating that the HW cannot be trusted to be >> - * in a sane state on the next vcpu load, >> - * - any change in the interrupt state >> - * >> - * Usage conditions: >> - * - cached value is "active clear" >> - * - value to be programmed is "active clear" >> - */ >> - if (timer->active_cleared_last && !phys_active) >> - return; >> - >> - ret = irq_set_irqchip_state(host_vtimer_irq, >> - IRQCHIP_STATE_ACTIVE, >> - phys_active); >> - WARN_ON(ret); >> + if (irqchip_in_kernel(vcpu->kvm)) { >> + phys_active = timer->irq.level || >> + kvm_vgic_map_is_active(vcpu, timer->irq.irq); >> + >> + /* >> + * We want to avoid hitting the (re)distributor as much as >> + * possible, as this is a potentially expensive MMIO access >> + * (not to mention locks in the irq layer), and a solution for >> + * this is to cache the "active" state in memory. >> + * >> + * Things to consider: we cannot cache an "active set" state, >> + * because the HW can change this behind our back (it becomes >> + * "clear" in the HW). We must then restrict the caching to >> + * the "clear" state. >> + * >> + * The cache is invalidated on: >> + * - vcpu put, indicating that the HW cannot be trusted to be >> + * in a sane state on the next vcpu load, >> + * - any change in the interrupt state >> + * >> + * Usage conditions: >> + * - cached value is "active clear" >> + * - value to be programmed is "active clear" >> + */ >> + if (timer->active_cleared_last && !phys_active) >> + return 0; >> + >> + ret = irq_set_irqchip_state(host_vtimer_irq, >> + IRQCHIP_STATE_ACTIVE, >> + phys_active); >> + WARN_ON(ret); >> + } else { >> + struct kvm_sync_regs *regs = &vcpu->run->s.regs; >> + >> + /* >> + * User space handles timer events, so we need to check whether >> + * its view of the world is in sync with ours. >> + */ >> + if (regs->kernel_timer_pending != regs->user_timer_pending) { >> + /* Return to user space */ >> + return 1; >> + } > > Maybe I'm misunderstanding and user_timer_pending is just a cached > verison of what you said last, but as I said above, I think you can just > compare timer->irq.level with the last value the kvm_run struct, and if > something changed, you have to exit. So how would user space know whether the line went up or down? Or didn't change at all (if we coalesce with an MMIO exit)? > >> + >> + /* >> + * As long as user space is aware that the timer is pending, >> + * we do not need to get new host timer events. >> + */ > > yes, correct, but I don't think this concept was clearly reflected in > your API text above. > >> + if (timer->irq.level) >> + disable_percpu_irq(host_vtimer_irq); >> + else >> + enable_percpu_irq(host_vtimer_irq, 0); >> + } > > could we move these two blocks into their own functions instead? That > would also give nice names to the huge chunk of complicated > functionality, e.g. flush_timer_state_to_user() and > flush_timer_state_to_vgic(). That's probably a very useful cleanup, yes :). Alex
On 23/09/2016 09:14, Alexander Graf wrote: >>> +/* >>> + * Synchronize the timer IRQ state with the interrupt controller. >>> + */ >>> static void kvm_timer_update_irq(struct kvm_vcpu *vcpu, bool new_level) >>> { >>> int ret; >>> struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu; >>> >>> timer->active_cleared_last = false; >>> timer->irq.level = new_level; >>> - trace_kvm_timer_update_irq(vcpu->vcpu_id, timer->irq.irq, >>> + trace_kvm_timer_update_irq(vcpu->vcpu_id, host_vtimer_irq, >>> timer->irq.level); >>> + [...] >>> + struct kvm_sync_regs *regs = &vcpu->run->s.regs; >>> + >>> + /* Populate the timer bitmap for user space */ >>> + regs->kernel_timer_pending &= ~KVM_ARM_TIMER_VTIMER; >>> + if (new_level) >>> + regs->kernel_timer_pending |= KVM_ARM_TIMER_VTIMER; >> >> I think if you got here, it means you have to exit to userspace to >> update it of the new state. If you don't want to propagate a return > > Yes, but we can't exit straight away with our own exit reason because we > might be inside an MMIO exit path here which already occupies the > exit_reason. So the idea is that whenever you're here you have one of the following cases: - are coming from kvm_timer_flush_hwstate, and then you exit immediately with KVM_EXIT_INTR if needed - you are coming from the kvm_timer_sync_hwstate just before handle_exit. Then if there's a vmexit you have already set regs->kernel_timer_pending, if not you'll do a kvm_timer_flush_hwstate soon. - you are coming from the kvm_timer_sync_hwstate in the middle of kvm_arch_vcpu_ioctl_run, and then "continue" will either exit the loop immediately (if ret <= 0) or go to kvm_timer_flush_hwstate as in the previous case Right? >> Maybe I'm misunderstanding and user_timer_pending is just a cached >> verison of what you said last, but as I said above, I think you can just >> compare timer->irq.level with the last value the kvm_run struct, and if >> something changed, you have to exit. > > So how would user space know whether the line went up or down? Or didn't > change at all (if we coalesce with an MMIO exit)? It would save the status of the line somewhere in its own variable, without introducing a relatively delicate API between kernel and user. I agree that you cannot update kernel_timer_pending only in flush_hwstate; otherwise you miss on case (2) when there is a vmexit. That has to stay in kvm_timer_sync_hwstate (or it could become a new function called at the very end of kvm_arch_vcpu_ioctl_run). However, I'm not sure why user_timer_pending is necessary. If it is just the value you assigned to regs->kernel_timer_pending at the time of the previous vmexit, can kvm_timer_flush_hwstate just do if (timer->prev_kernel_timer_pending != regs->kernel_timer_pending) { timer->prev_kernel_timer_pending = regs->kernel_timer_pending; return 1; } ? Or even if (timer->prev_irq_level != timer->irq.level) { timer->prev_irq_level = regs->irq.level; return 1; } so that regs->kernel_timer_pending remains exclusively kvm_timer_sync_hwstate's business? Thanks, Paolo >> >>> + >>> + /* >>> + * As long as user space is aware that the timer is pending, >>> + * we do not need to get new host timer events. >>> + */ >> >> yes, correct, but I don't think this concept was clearly reflected in >> your API text above. >> >>> + if (timer->irq.level) >>> + disable_percpu_irq(host_vtimer_irq); >>> + else >>> + enable_percpu_irq(host_vtimer_irq, 0); >>> + } >> >> could we move these two blocks into their own functions instead? That >> would also give nice names to the huge chunk of complicated >> functionality, e.g. flush_timer_state_to_user() and >> flush_timer_state_to_vgic(). > > That's probably a very useful cleanup, yes :). > > > Alex > -- > To unsubscribe from this list: send the line "unsubscribe kvm" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html >
On 23.09.16 10:57, Paolo Bonzini wrote: > > > On 23/09/2016 09:14, Alexander Graf wrote: >>>> +/* >>>> + * Synchronize the timer IRQ state with the interrupt controller. >>>> + */ >>>> static void kvm_timer_update_irq(struct kvm_vcpu *vcpu, bool new_level) >>>> { >>>> int ret; >>>> struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu; >>>> >>>> timer->active_cleared_last = false; >>>> timer->irq.level = new_level; >>>> - trace_kvm_timer_update_irq(vcpu->vcpu_id, timer->irq.irq, >>>> + trace_kvm_timer_update_irq(vcpu->vcpu_id, host_vtimer_irq, >>>> timer->irq.level); >>>> + [...] >>>> + struct kvm_sync_regs *regs = &vcpu->run->s.regs; >>>> + >>>> + /* Populate the timer bitmap for user space */ >>>> + regs->kernel_timer_pending &= ~KVM_ARM_TIMER_VTIMER; >>>> + if (new_level) >>>> + regs->kernel_timer_pending |= KVM_ARM_TIMER_VTIMER; >>> >>> I think if you got here, it means you have to exit to userspace to >>> update it of the new state. If you don't want to propagate a return >> >> Yes, but we can't exit straight away with our own exit reason because we >> might be inside an MMIO exit path here which already occupies the >> exit_reason. > > So the idea is that whenever you're here you have one of the following > cases: > > - are coming from kvm_timer_flush_hwstate, and then you exit immediately > with KVM_EXIT_INTR if needed > > - you are coming from the kvm_timer_sync_hwstate just before > handle_exit. Then if there's a vmexit you have already set > regs->kernel_timer_pending, if not you'll do a kvm_timer_flush_hwstate soon. > > - you are coming from the kvm_timer_sync_hwstate in the middle of > kvm_arch_vcpu_ioctl_run, and then "continue" will either exit the loop > immediately (if ret <= 0) or go to kvm_timer_flush_hwstate as in the > previous case > > Right? Yup :) > >>> Maybe I'm misunderstanding and user_timer_pending is just a cached >>> verison of what you said last, but as I said above, I think you can just >>> compare timer->irq.level with the last value the kvm_run struct, and if >>> something changed, you have to exit. >> >> So how would user space know whether the line went up or down? Or didn't >> change at all (if we coalesce with an MMIO exit)? > > It would save the status of the line somewhere in its own variable, > without introducing a relatively delicate API between kernel and user. > > I agree that you cannot update kernel_timer_pending only in > flush_hwstate; otherwise you miss on case (2) when there is a vmexit. > That has to stay in kvm_timer_sync_hwstate (or it could become a new > function called at the very end of kvm_arch_vcpu_ioctl_run). The beauty of having it in the timer update function is that it gets called from flush_hwstate() as well. That way we catch the update also in cases where we need it before we enter the vcpu. > However, I'm not sure why user_timer_pending is necessary. If it is > just the value you assigned to regs->kernel_timer_pending at the time of > the previous vmexit, can kvm_timer_flush_hwstate just do > > if (timer->prev_kernel_timer_pending != regs->kernel_timer_pending) { > timer->prev_kernel_timer_pending = regs->kernel_timer_pending; > return 1; > } > > ? Or even > > if (timer->prev_irq_level != timer->irq.level) { > timer->prev_irq_level = regs->irq.level; > return 1; > } > > so that regs->kernel_timer_pending remains exclusively > kvm_timer_sync_hwstate's business? We could do that too, yes. But it puts more burden on user space - it would have to ensure that it *always* checks for the pending timer status. With this approach, user space may opt to only check for timer state changes on -EINTR and things would still work. Alex
On 23/09/2016 11:10, Alexander Graf wrote: >>>> Maybe I'm misunderstanding and user_timer_pending is just a cached >>>> verison of what you said last, but as I said above, I think you can just >>>> compare timer->irq.level with the last value the kvm_run struct, and if >>>> something changed, you have to exit. >>> >>> So how would user space know whether the line went up or down? Or didn't >>> change at all (if we coalesce with an MMIO exit)? >> >> It would save the status of the line somewhere in its own variable, >> without introducing a relatively delicate API between kernel and user. >> >> I agree that you cannot update kernel_timer_pending only in >> flush_hwstate; otherwise you miss on case (2) when there is a vmexit. >> That has to stay in kvm_timer_sync_hwstate (or it could become a new >> function called at the very end of kvm_arch_vcpu_ioctl_run). > > The beauty of having it in the timer update function is that it gets > called from flush_hwstate() as well. That way we catch the update also > in cases where we need it before we enter the vcpu. Yeah, the timer update function is ideal because it is called from both flush_hwstate and sync_hwstate. >> However, I'm not sure why user_timer_pending is necessary. If it is >> just the value you assigned to regs->kernel_timer_pending at the time of >> the previous vmexit, can kvm_timer_flush_hwstate just do >> >> if (timer->prev_kernel_timer_pending != regs->kernel_timer_pending) { >> timer->prev_kernel_timer_pending = regs->kernel_timer_pending; >> return 1; >> } >> >> ? Or even >> >> if (timer->prev_irq_level != timer->irq.level) { >> timer->prev_irq_level = regs->irq.level; >> return 1; >> } >> >> so that regs->kernel_timer_pending remains exclusively >> kvm_timer_sync_hwstate's business? > > We could do that too, yes. But it puts more burden on user space - it > would have to ensure that it *always* checks for the pending timer > status. With this approach, user space may opt to only check for timer > state changes on -EINTR and things would still work. That would be overloading EINTR a bit though. But that's a digression, I don't think it matters much. On the other hand, what happens if you run new QEMU with old userspace? With user_timer_pending you'd get an infinite stream of vmexits the first time the timer fires, wouldn't you? Whereas if you keep it in the kernel, userspace would simply not get the interrupt (because it doesn't know about kernel_timer_pending) and think it got a spurious vmexit. The kernel's IRQ would stay masked and everything would just (not) work like before your patch? Thanks, Paolo
On 23.09.16 11:15, Paolo Bonzini wrote: > > > On 23/09/2016 11:10, Alexander Graf wrote: >>>>> Maybe I'm misunderstanding and user_timer_pending is just a cached >>>>> verison of what you said last, but as I said above, I think you can just >>>>> compare timer->irq.level with the last value the kvm_run struct, and if >>>>> something changed, you have to exit. >>>> >>>> So how would user space know whether the line went up or down? Or didn't >>>> change at all (if we coalesce with an MMIO exit)? >>> >>> It would save the status of the line somewhere in its own variable, >>> without introducing a relatively delicate API between kernel and user. >>> >>> I agree that you cannot update kernel_timer_pending only in >>> flush_hwstate; otherwise you miss on case (2) when there is a vmexit. >>> That has to stay in kvm_timer_sync_hwstate (or it could become a new >>> function called at the very end of kvm_arch_vcpu_ioctl_run). >> >> The beauty of having it in the timer update function is that it gets >> called from flush_hwstate() as well. That way we catch the update also >> in cases where we need it before we enter the vcpu. > > Yeah, the timer update function is ideal because it is called from both > flush_hwstate and sync_hwstate. > >>> However, I'm not sure why user_timer_pending is necessary. If it is >>> just the value you assigned to regs->kernel_timer_pending at the time of >>> the previous vmexit, can kvm_timer_flush_hwstate just do >>> >>> if (timer->prev_kernel_timer_pending != regs->kernel_timer_pending) { >>> timer->prev_kernel_timer_pending = regs->kernel_timer_pending; >>> return 1; >>> } >>> >>> ? Or even >>> >>> if (timer->prev_irq_level != timer->irq.level) { >>> timer->prev_irq_level = regs->irq.level; >>> return 1; >>> } >>> >>> so that regs->kernel_timer_pending remains exclusively >>> kvm_timer_sync_hwstate's business? >> >> We could do that too, yes. But it puts more burden on user space - it >> would have to ensure that it *always* checks for the pending timer >> status. With this approach, user space may opt to only check for timer >> state changes on -EINTR and things would still work. > > That would be overloading EINTR a bit though. But that's a digression, > I don't think it matters much. > > On the other hand, what happens if you run new QEMU with old userspace? > With user_timer_pending you'd get an infinite stream of vmexits the > first time the timer fires, wouldn't you? Whereas if you keep it in the > kernel, userspace would simply not get the interrupt (because it doesn't > know about kernel_timer_pending) and think it got a spurious vmexit. > The kernel's IRQ would stay masked and everything would just (not) work > like before your patch? Yes, we'd definitely stay more compatible by tracking it only in the kernel. I'm not fully convinced that it's the better interface, but since both Christoffer and you seem to choke on that part, I'll give it a stab ;). Alex
On 23/09/2016 11:17, Alexander Graf wrote: >> > >> > On the other hand, what happens if you run new QEMU with old userspace? >> > With user_timer_pending you'd get an infinite stream of vmexits the >> > first time the timer fires, wouldn't you? Whereas if you keep it in the >> > kernel, userspace would simply not get the interrupt (because it doesn't >> > know about kernel_timer_pending) and think it got a spurious vmexit. >> > The kernel's IRQ would stay masked and everything would just (not) work >> > like before your patch? > Yes, we'd definitely stay more compatible by tracking it only in the > kernel. I'm not fully convinced that it's the better interface, but > since both Christoffer and you seem to choke on that part, I'll give it > a stab ;). Cool! FWIW my suggestion for kernel_timer_pending's name would be timer_irq_level (nicely matching timer->irq.level in the kernel). Paolo
On Fri, Sep 23, 2016 at 11:10:46AM +0200, Alexander Graf wrote: > > > On 23.09.16 10:57, Paolo Bonzini wrote: > > > > > > On 23/09/2016 09:14, Alexander Graf wrote: > >>>> +/* > >>>> + * Synchronize the timer IRQ state with the interrupt controller. > >>>> + */ > >>>> static void kvm_timer_update_irq(struct kvm_vcpu *vcpu, bool new_level) > >>>> { > >>>> int ret; > >>>> struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu; > >>>> > >>>> timer->active_cleared_last = false; > >>>> timer->irq.level = new_level; > >>>> - trace_kvm_timer_update_irq(vcpu->vcpu_id, timer->irq.irq, > >>>> + trace_kvm_timer_update_irq(vcpu->vcpu_id, host_vtimer_irq, > >>>> timer->irq.level); > >>>> + [...] > >>>> + struct kvm_sync_regs *regs = &vcpu->run->s.regs; > >>>> + > >>>> + /* Populate the timer bitmap for user space */ > >>>> + regs->kernel_timer_pending &= ~KVM_ARM_TIMER_VTIMER; > >>>> + if (new_level) > >>>> + regs->kernel_timer_pending |= KVM_ARM_TIMER_VTIMER; > >>> > >>> I think if you got here, it means you have to exit to userspace to > >>> update it of the new state. If you don't want to propagate a return > >> > >> Yes, but we can't exit straight away with our own exit reason because we > >> might be inside an MMIO exit path here which already occupies the > >> exit_reason. > > > > So the idea is that whenever you're here you have one of the following > > cases: > > > > - are coming from kvm_timer_flush_hwstate, and then you exit immediately > > with KVM_EXIT_INTR if needed > > > > - you are coming from the kvm_timer_sync_hwstate just before > > handle_exit. Then if there's a vmexit you have already set > > regs->kernel_timer_pending, if not you'll do a kvm_timer_flush_hwstate soon. > > > > - you are coming from the kvm_timer_sync_hwstate in the middle of > > kvm_arch_vcpu_ioctl_run, and then "continue" will either exit the loop > > immediately (if ret <= 0) or go to kvm_timer_flush_hwstate as in the > > previous case > > > > Right? > > Yup :) > > > > >>> Maybe I'm misunderstanding and user_timer_pending is just a cached > >>> verison of what you said last, but as I said above, I think you can just > >>> compare timer->irq.level with the last value the kvm_run struct, and if > >>> something changed, you have to exit. > >> > >> So how would user space know whether the line went up or down? Or didn't > >> change at all (if we coalesce with an MMIO exit)? > > > > It would save the status of the line somewhere in its own variable, > > without introducing a relatively delicate API between kernel and user. > > > > I agree that you cannot update kernel_timer_pending only in > > flush_hwstate; otherwise you miss on case (2) when there is a vmexit. > > That has to stay in kvm_timer_sync_hwstate (or it could become a new > > function called at the very end of kvm_arch_vcpu_ioctl_run). > > The beauty of having it in the timer update function is that it gets > called from flush_hwstate() as well. That way we catch the update also > in cases where we need it before we enter the vcpu. > > > However, I'm not sure why user_timer_pending is necessary. If it is > > just the value you assigned to regs->kernel_timer_pending at the time of > > the previous vmexit, can kvm_timer_flush_hwstate just do > > > > if (timer->prev_kernel_timer_pending != regs->kernel_timer_pending) { > > timer->prev_kernel_timer_pending = regs->kernel_timer_pending; > > return 1; > > } > > > > ? Or even > > > > if (timer->prev_irq_level != timer->irq.level) { > > timer->prev_irq_level = regs->irq.level; > > return 1; > > } > > > > so that regs->kernel_timer_pending remains exclusively > > kvm_timer_sync_hwstate's business? > > We could do that too, yes. But it puts more burden on user space - it > would have to ensure that it *always* checks for the pending timer > status. With this approach, user space may opt to only check for timer > state changes on -EINTR and things would still work. > Huh? I thought the whole point was that you could piggy back on an MMIO exit with a change in the line state, for example. So I don't understand this. In any case, exposing internal historical state only used by the kernel to figure out whether or not it should force an exit if there's not already one happening, to user space, just feels weird to me, and my emphasis is on having a clean timer emulation component in the kernel where the semantics are clear. Have a look at my proposal in the other mail on this thread. -Christoffer
diff --git a/Documentation/virtual/kvm/api.txt b/Documentation/virtual/kvm/api.txt index 739db9a..8049327 100644 --- a/Documentation/virtual/kvm/api.txt +++ b/Documentation/virtual/kvm/api.txt @@ -3928,3 +3928,29 @@ In order to use SynIC, it has to be activated by setting this capability via KVM_ENABLE_CAP ioctl on the vcpu fd. Note that this will disable the use of APIC hardware virtualization even if supported by the CPU, as it's incompatible with SynIC auto-EOI behavior. + +8.3 KVM_CAP_ARM_TIMER + +Architectures: arm, arm64 +This capability, if KVM_CHECK_EXTENSION indicates that it is available and no +in-kernel interrupt controller is in use, means that that the kernel populates +the vcpu's run->s.regs.kernel_timer_pending field with timers that are currently +considered pending by kvm. + +If active, it also allows user space to propagate its own pending state of timer +interrupt lines using run->s.regs.user_timer_pending. If those two fields +mismatch during CPU execution, kvm will exit to user space to give it a chance +to update its own interrupt pending status. This usually involves triggering +an interrupt line on a user space emulated interrupt controller. + +The fields run->s.regs.kernel_timer_pending and run->s.regs.user_timer_pending +are available independent of run->kvm_valid_regs or run->kvm_dirty_regs bits. +If no in-kernel interrupt controller is used and the capability exists, they +will always be available and used. + +Currently the following bits are defined for both bitmaps: + + KVM_ARM_TIMER_VTIMER - virtual timer + +Future versions of kvm may implement additional timer events. These will get +indicated by additional KVM_CAP extensions. diff --git a/arch/arm/include/uapi/asm/kvm.h b/arch/arm/include/uapi/asm/kvm.h index a2b3eb3..caad81d 100644 --- a/arch/arm/include/uapi/asm/kvm.h +++ b/arch/arm/include/uapi/asm/kvm.h @@ -105,6 +105,9 @@ struct kvm_debug_exit_arch { }; struct kvm_sync_regs { + /* Used with KVM_CAP_ARM_TIMER */ + u8 kernel_timer_pending; + u8 user_timer_pending; }; struct kvm_arch_memory_slot { diff --git a/arch/arm/kvm/arm.c b/arch/arm/kvm/arm.c index 75f130e..dc19221 100644 --- a/arch/arm/kvm/arm.c +++ b/arch/arm/kvm/arm.c @@ -187,6 +187,7 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext) case KVM_CAP_ARM_PSCI_0_2: case KVM_CAP_READONLY_MEM: case KVM_CAP_MP_STATE: + case KVM_CAP_ARM_TIMER: r = 1; break; case KVM_CAP_COALESCED_MMIO: @@ -474,13 +475,7 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu) return ret; } - /* - * Enable the arch timers only if we have an in-kernel VGIC - * and it has been properly initialized, since we cannot handle - * interrupts from the virtual timer with a userspace gic. - */ - if (irqchip_in_kernel(kvm) && vgic_initialized(kvm)) - ret = kvm_timer_enable(vcpu); + ret = kvm_timer_enable(vcpu); return ret; } @@ -588,7 +583,10 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run) */ preempt_disable(); kvm_pmu_flush_hwstate(vcpu); - kvm_timer_flush_hwstate(vcpu); + if (kvm_timer_flush_hwstate(vcpu)) { + ret = -EINTR; + run->exit_reason = KVM_EXIT_INTR; + } kvm_vgic_flush_hwstate(vcpu); local_irq_disable(); diff --git a/arch/arm64/include/uapi/asm/kvm.h b/arch/arm64/include/uapi/asm/kvm.h index 3051f86..9aac860 100644 --- a/arch/arm64/include/uapi/asm/kvm.h +++ b/arch/arm64/include/uapi/asm/kvm.h @@ -143,6 +143,9 @@ struct kvm_debug_exit_arch { #define KVM_GUESTDBG_USE_HW (1 << 17) struct kvm_sync_regs { + /* Used with KVM_CAP_ARM_TIMER */ + u8 kernel_timer_pending; + u8 user_timer_pending; }; struct kvm_arch_memory_slot { diff --git a/include/kvm/arm_arch_timer.h b/include/kvm/arm_arch_timer.h index dda39d8..8cd7240 100644 --- a/include/kvm/arm_arch_timer.h +++ b/include/kvm/arm_arch_timer.h @@ -63,7 +63,7 @@ void kvm_timer_init(struct kvm *kvm); int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu, const struct kvm_irq_level *irq); void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu); -void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu); +int kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu); void kvm_timer_sync_hwstate(struct kvm_vcpu *vcpu); void kvm_timer_vcpu_terminate(struct kvm_vcpu *vcpu); diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h index 300ef25..1fc02d7 100644 --- a/include/uapi/linux/kvm.h +++ b/include/uapi/linux/kvm.h @@ -870,6 +870,7 @@ struct kvm_ppc_smmu_info { #define KVM_CAP_S390_USER_INSTR0 130 #define KVM_CAP_MSI_DEVID 131 #define KVM_CAP_PPC_HTM 132 +#define KVM_CAP_ARM_TIMER 133 #ifdef KVM_CAP_IRQ_ROUTING @@ -1327,4 +1328,9 @@ struct kvm_assigned_msix_entry { #define KVM_X2APIC_API_USE_32BIT_IDS (1ULL << 0) #define KVM_X2APIC_API_DISABLE_BROADCAST_QUIRK (1ULL << 1) +/* Available with KVM_CAP_ARM_TIMER */ + +/* Bits for run->s.regs.{user,kernel}_timer_pending */ +#define KVM_ARM_TIMER_VTIMER (1 << 0) + #endif /* __LINUX_KVM_H */ diff --git a/virt/kvm/arm/arch_timer.c b/virt/kvm/arm/arch_timer.c index 4309b60..0c6fc38 100644 --- a/virt/kvm/arm/arch_timer.c +++ b/virt/kvm/arm/arch_timer.c @@ -166,21 +166,36 @@ bool kvm_timer_should_fire(struct kvm_vcpu *vcpu) return cval <= now; } +/* + * Synchronize the timer IRQ state with the interrupt controller. + */ static void kvm_timer_update_irq(struct kvm_vcpu *vcpu, bool new_level) { int ret; struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu; - BUG_ON(!vgic_initialized(vcpu->kvm)); - timer->active_cleared_last = false; timer->irq.level = new_level; - trace_kvm_timer_update_irq(vcpu->vcpu_id, timer->irq.irq, + trace_kvm_timer_update_irq(vcpu->vcpu_id, host_vtimer_irq, timer->irq.level); - ret = kvm_vgic_inject_mapped_irq(vcpu->kvm, vcpu->vcpu_id, - timer->irq.irq, - timer->irq.level); - WARN_ON(ret); + + if (irqchip_in_kernel(vcpu->kvm)) { + BUG_ON(!vgic_initialized(vcpu->kvm)); + + /* Fire the timer in the VGIC */ + ret = kvm_vgic_inject_mapped_irq(vcpu->kvm, vcpu->vcpu_id, + timer->irq.irq, + timer->irq.level); + + WARN_ON(ret); + } else { + struct kvm_sync_regs *regs = &vcpu->run->s.regs; + + /* Populate the timer bitmap for user space */ + regs->kernel_timer_pending &= ~KVM_ARM_TIMER_VTIMER; + if (new_level) + regs->kernel_timer_pending |= KVM_ARM_TIMER_VTIMER; + } } /* @@ -197,7 +212,8 @@ static int kvm_timer_update_state(struct kvm_vcpu *vcpu) * because the guest would never see the interrupt. Instead wait * until we call this function from kvm_timer_flush_hwstate. */ - if (!vgic_initialized(vcpu->kvm) || !timer->enabled) + if ((irqchip_in_kernel(vcpu->kvm) && !vgic_initialized(vcpu->kvm)) || + !timer->enabled) return -ENODEV; if (kvm_timer_should_fire(vcpu) != timer->irq.level) @@ -248,15 +264,20 @@ void kvm_timer_unschedule(struct kvm_vcpu *vcpu) * * Check if the virtual timer has expired while we were running in the host, * and inject an interrupt if that was the case. + * + * Returns: + * + * 0 - success + * 1 - need exit to user space */ -void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu) +int kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu) { struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu; bool phys_active; int ret; if (kvm_timer_update_state(vcpu)) - return; + return 0; /* * If we enter the guest with the virtual input level to the VGIC @@ -275,38 +296,61 @@ void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu) * to ensure that hardware interrupts from the timer triggers a guest * exit. */ - phys_active = timer->irq.level || - kvm_vgic_map_is_active(vcpu, timer->irq.irq); - - /* - * We want to avoid hitting the (re)distributor as much as - * possible, as this is a potentially expensive MMIO access - * (not to mention locks in the irq layer), and a solution for - * this is to cache the "active" state in memory. - * - * Things to consider: we cannot cache an "active set" state, - * because the HW can change this behind our back (it becomes - * "clear" in the HW). We must then restrict the caching to - * the "clear" state. - * - * The cache is invalidated on: - * - vcpu put, indicating that the HW cannot be trusted to be - * in a sane state on the next vcpu load, - * - any change in the interrupt state - * - * Usage conditions: - * - cached value is "active clear" - * - value to be programmed is "active clear" - */ - if (timer->active_cleared_last && !phys_active) - return; - - ret = irq_set_irqchip_state(host_vtimer_irq, - IRQCHIP_STATE_ACTIVE, - phys_active); - WARN_ON(ret); + if (irqchip_in_kernel(vcpu->kvm)) { + phys_active = timer->irq.level || + kvm_vgic_map_is_active(vcpu, timer->irq.irq); + + /* + * We want to avoid hitting the (re)distributor as much as + * possible, as this is a potentially expensive MMIO access + * (not to mention locks in the irq layer), and a solution for + * this is to cache the "active" state in memory. + * + * Things to consider: we cannot cache an "active set" state, + * because the HW can change this behind our back (it becomes + * "clear" in the HW). We must then restrict the caching to + * the "clear" state. + * + * The cache is invalidated on: + * - vcpu put, indicating that the HW cannot be trusted to be + * in a sane state on the next vcpu load, + * - any change in the interrupt state + * + * Usage conditions: + * - cached value is "active clear" + * - value to be programmed is "active clear" + */ + if (timer->active_cleared_last && !phys_active) + return 0; + + ret = irq_set_irqchip_state(host_vtimer_irq, + IRQCHIP_STATE_ACTIVE, + phys_active); + WARN_ON(ret); + } else { + struct kvm_sync_regs *regs = &vcpu->run->s.regs; + + /* + * User space handles timer events, so we need to check whether + * its view of the world is in sync with ours. + */ + if (regs->kernel_timer_pending != regs->user_timer_pending) { + /* Return to user space */ + return 1; + } + + /* + * As long as user space is aware that the timer is pending, + * we do not need to get new host timer events. + */ + if (timer->irq.level) + disable_percpu_irq(host_vtimer_irq); + else + enable_percpu_irq(host_vtimer_irq, 0); + } timer->active_cleared_last = !phys_active; + return 0; } /** @@ -479,6 +523,10 @@ int kvm_timer_enable(struct kvm_vcpu *vcpu) if (timer->enabled) return 0; + /* No need to route physical IRQs when we don't use the vgic */ + if (!irqchip_in_kernel(vcpu->kvm)) + goto no_vgic; + /* * Find the physical IRQ number corresponding to the host_vtimer_irq */ @@ -502,6 +550,7 @@ int kvm_timer_enable(struct kvm_vcpu *vcpu) if (ret) return ret; +no_vgic: /* * There is a potential race here between VCPUs starting for the first
We have 2 modes for dealing with interrupts in the ARM world. We can either handle them all using hardware acceleration through the vgic or we can emulate a gic in user space and only drive CPU IRQ pins from there. Unfortunately, when driving IRQs from user space, we never tell user space about timer events that may result in interrupt line state changes, so we lose out on timer events if we run with user space gic emulation. This patch fixes that by syncing user space's view of the vtimer irq line with the kvm view of that same line. With this patch I can successfully run edk2 and Linux with user space gic emulation. Signed-off-by: Alexander Graf <agraf@suse.de> --- v1 -> v2: - Add back curly brace that got lost v2 -> v3: - Split into patch set v3 -> v4: - Improve documentation v4 -> v5: - Rewrite to use pending state sync in sregs (marc) - Remove redundant checks of vgic_initialized() - qemu tree to try this out: https://github.com/agraf/u-boot.git no-kvm-irqchip-for-v5 --- Documentation/virtual/kvm/api.txt | 26 ++++++++ arch/arm/include/uapi/asm/kvm.h | 3 + arch/arm/kvm/arm.c | 14 ++--- arch/arm64/include/uapi/asm/kvm.h | 3 + include/kvm/arm_arch_timer.h | 2 +- include/uapi/linux/kvm.h | 6 ++ virt/kvm/arm/arch_timer.c | 129 ++++++++++++++++++++++++++------------ 7 files changed, 134 insertions(+), 49 deletions(-)