| Message ID | 20230907100931.1186690-5-maz@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/5] KVM: arm64: Simplify kvm_vcpu_get_mpidr_aff() | expand |
Hi Marc, On 2023/9/7 18:09, Marc Zyngier wrote: > As we're about to change the way SGIs are sent, start by splitting > out some of the basic functionnality: instead of intermingling functionality > the broadcast and non-broadcast cases with the actual SGI generation, > perform the following cleanups: > > - move the SGI queuing into its own helper > - split the broadcast code from the affinity-driven code > - replace the mask/shift combinations with FIELD_GET() > > The result is much more readable, and paves the way for further > optimisations. Indeed! > @@ -1070,19 +1102,30 @@ void vgic_v3_dispatch_sgi(struct kvm_vcpu *vcpu, u64 reg, bool allow_group1) > { > struct kvm *kvm = vcpu->kvm; > struct kvm_vcpu *c_vcpu; > - u16 target_cpus; > + unsigned long target_cpus; > u64 mpidr; > - int sgi; > - int vcpu_id = vcpu->vcpu_id; > - bool broadcast; > - unsigned long c, flags; > - > - sgi = (reg & ICC_SGI1R_SGI_ID_MASK) >> ICC_SGI1R_SGI_ID_SHIFT; > - broadcast = reg & BIT_ULL(ICC_SGI1R_IRQ_ROUTING_MODE_BIT); > - target_cpus = (reg & ICC_SGI1R_TARGET_LIST_MASK) >> ICC_SGI1R_TARGET_LIST_SHIFT; > + u32 sgi; > + unsigned long c; > + > + sgi = FIELD_GET(ICC_SGI1R_SGI_ID_MASK, reg); > + > + /* Broadcast */ > + if (unlikely(reg & BIT_ULL(ICC_SGI1R_IRQ_ROUTING_MODE_BIT))) { > + kvm_for_each_vcpu(c, c_vcpu, kvm) { > + /* Don't signal the calling VCPU */ > + if (c_vcpu == vcpu) > + continue; > + > + vgic_v3_queue_sgi(c_vcpu, sgi, allow_group1); > + } > + > + return; > + } > + > mpidr = SGI_AFFINITY_LEVEL(reg, 3); > mpidr |= SGI_AFFINITY_LEVEL(reg, 2); > mpidr |= SGI_AFFINITY_LEVEL(reg, 1); > + target_cpus = FIELD_GET(ICC_SGI1R_TARGET_LIST_MASK, reg); > > /* > * We iterate over all VCPUs to find the MPIDRs matching the request. > @@ -1091,54 +1134,19 @@ void vgic_v3_dispatch_sgi(struct kvm_vcpu *vcpu, u64 reg, bool allow_group1) > * VCPUs when most of the times we just signal a single VCPU. > */ > kvm_for_each_vcpu(c, c_vcpu, kvm) { > - struct vgic_irq *irq; > + int level0; > > /* Exit early if we have dealt with all requested CPUs */ > - if (!broadcast && target_cpus == 0) > + if (target_cpus == 0) > break; > - > - /* Don't signal the calling VCPU */ > - if (broadcast && c == vcpu_id) Unrelated to this patch, but it looks that we were comparing the value of *vcpu_idx* and vcpu_id to skip the calling VCPU. Is there a rule in KVM that userspace should invoke KVM_CREATE_VCPU with sequential "vcpu id"s, starting at 0, so that the user-provided vcpu_id always equals to the KVM-internal vcpu_idx for a given VCPU? I asked because it seems that in kvm/arm64 we always use kvm_get_vcpu(kvm, i) to obtain the kvm_vcpu pointer, even if *i* is sometimes essentially provided by userspace.. Besides, the refactor itself looks good to me. Thanks, Zenghui
On Sun, 10 Sep 2023 17:25:36 +0100, Zenghui Yu <zenghui.yu@linux.dev> wrote: > > Hi Marc, > > On 2023/9/7 18:09, Marc Zyngier wrote: > > As we're about to change the way SGIs are sent, start by splitting > > out some of the basic functionnality: instead of intermingling > > functionality > > > the broadcast and non-broadcast cases with the actual SGI generation, > > perform the following cleanups: > > > > - move the SGI queuing into its own helper > > - split the broadcast code from the affinity-driven code > > - replace the mask/shift combinations with FIELD_GET() > > > > The result is much more readable, and paves the way for further > > optimisations. > > Indeed! > > > @@ -1070,19 +1102,30 @@ void vgic_v3_dispatch_sgi(struct kvm_vcpu *vcpu, u64 reg, bool allow_group1) > > { > > struct kvm *kvm = vcpu->kvm; > > struct kvm_vcpu *c_vcpu; > > - u16 target_cpus; > > + unsigned long target_cpus; > > u64 mpidr; > > - int sgi; > > - int vcpu_id = vcpu->vcpu_id; > > - bool broadcast; > > - unsigned long c, flags; > > - > > - sgi = (reg & ICC_SGI1R_SGI_ID_MASK) >> ICC_SGI1R_SGI_ID_SHIFT; > > - broadcast = reg & BIT_ULL(ICC_SGI1R_IRQ_ROUTING_MODE_BIT); > > - target_cpus = (reg & ICC_SGI1R_TARGET_LIST_MASK) >> ICC_SGI1R_TARGET_LIST_SHIFT; > > + u32 sgi; > > + unsigned long c; > > + > > + sgi = FIELD_GET(ICC_SGI1R_SGI_ID_MASK, reg); > > + > > + /* Broadcast */ > > + if (unlikely(reg & BIT_ULL(ICC_SGI1R_IRQ_ROUTING_MODE_BIT))) { > > + kvm_for_each_vcpu(c, c_vcpu, kvm) { > > + /* Don't signal the calling VCPU */ > > + if (c_vcpu == vcpu) > > + continue; > > + > > + vgic_v3_queue_sgi(c_vcpu, sgi, allow_group1); > > + } > > + > > + return; > > + } > > + > > mpidr = SGI_AFFINITY_LEVEL(reg, 3); > > mpidr |= SGI_AFFINITY_LEVEL(reg, 2); > > mpidr |= SGI_AFFINITY_LEVEL(reg, 1); > > + target_cpus = FIELD_GET(ICC_SGI1R_TARGET_LIST_MASK, reg); > > /* > > * We iterate over all VCPUs to find the MPIDRs matching the request. > > @@ -1091,54 +1134,19 @@ void vgic_v3_dispatch_sgi(struct kvm_vcpu *vcpu, u64 reg, bool allow_group1) > > * VCPUs when most of the times we just signal a single VCPU. > > */ > > kvm_for_each_vcpu(c, c_vcpu, kvm) { > > - struct vgic_irq *irq; > > + int level0; > > /* Exit early if we have dealt with all requested CPUs > > */ > > - if (!broadcast && target_cpus == 0) > > + if (target_cpus == 0) > > break; > > - > > - /* Don't signal the calling VCPU */ > > - if (broadcast && c == vcpu_id) > > Unrelated to this patch, but it looks that we were comparing the value > of *vcpu_idx* and vcpu_id to skip the calling VCPU. Huh, well caught. That was definitely a bug that was there for ever, and only you spotted it. Guess I should flag it as a stable candidate. > Is there a rule in KVM that userspace should invoke KVM_CREATE_VCPU > with sequential "vcpu id"s, starting at 0, so that the user-provided > vcpu_id always equals to the KVM-internal vcpu_idx for a given VCPU? I don't think there is any such rule. As far as I can tell, any number will do as long as it is within the range [0, max_vcpu_id). Of course, max_vcpu_id doesn't even exist on arm64. From what I can tell, this is just some random number between 0 and 511 for us (GICv2 notwithstanding). > I asked because it seems that in kvm/arm64 we always use > kvm_get_vcpu(kvm, i) to obtain the kvm_vcpu pointer, even if *i* is > sometimes essentially provided by userspace.. Huh, this is incredibly dodgy. I had a go at a few occurrences (see below), but this is hardly a complete list. > Besides, the refactor itself looks good to me. Cool, thanks! M. diff --git a/arch/arm64/kvm/arch_timer.c b/arch/arm64/kvm/arch_timer.c index 6dcdae4d38cb..e32c867e7b48 100644 --- a/arch/arm64/kvm/arch_timer.c +++ b/arch/arm64/kvm/arch_timer.c @@ -458,7 +458,7 @@ static void kvm_timer_update_irq(struct kvm_vcpu *vcpu, bool new_level, timer_ctx->irq.level); if (!userspace_irqchip(vcpu->kvm)) { - ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id, + ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_idx, timer_irq(timer_ctx), timer_ctx->irq.level, timer_ctx); diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c index a3b13281d38a..1f7b074b81df 100644 --- a/arch/arm64/kvm/arm.c +++ b/arch/arm64/kvm/arm.c @@ -439,9 +439,9 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu) * We might get preempted before the vCPU actually runs, but * over-invalidation doesn't affect correctness. */ - if (*last_ran != vcpu->vcpu_id) { + if (*last_ran != vcpu->vcpu_idx) { kvm_call_hyp(__kvm_flush_cpu_context, mmu); - *last_ran = vcpu->vcpu_id; + *last_ran = vcpu->vcpu_idx; } vcpu->cpu = cpu; @@ -1207,7 +1207,7 @@ int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level, if (vcpu_idx >= nrcpus) return -EINVAL; - vcpu = kvm_get_vcpu(kvm, vcpu_idx); + vcpu = kvm_get_vcpu_by_id(kvm, vcpu_idx); if (!vcpu) return -EINVAL; @@ -1222,14 +1222,14 @@ int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level, if (vcpu_idx >= nrcpus) return -EINVAL; - vcpu = kvm_get_vcpu(kvm, vcpu_idx); + vcpu = kvm_get_vcpu_by_id(kvm, vcpu_idx); if (!vcpu) return -EINVAL; if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS) return -EINVAL; - return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL); + return kvm_vgic_inject_irq(kvm, vcpu->vcpu_idx, irq_num, level, NULL); case KVM_ARM_IRQ_TYPE_SPI: if (!irqchip_in_kernel(kvm)) return -ENXIO; diff --git a/arch/arm64/kvm/pmu-emul.c b/arch/arm64/kvm/pmu-emul.c index 6b066e04dc5d..4448940b6d79 100644 --- a/arch/arm64/kvm/pmu-emul.c +++ b/arch/arm64/kvm/pmu-emul.c @@ -348,7 +348,7 @@ static void kvm_pmu_update_state(struct kvm_vcpu *vcpu) pmu->irq_level = overflow; if (likely(irqchip_in_kernel(vcpu->kvm))) { - int ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id, + int ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_idx, pmu->irq_num, overflow, pmu); WARN_ON(ret); } diff --git a/arch/arm64/kvm/vgic/vgic-debug.c b/arch/arm64/kvm/vgic/vgic-debug.c index 07aa0437125a..85606a531dc3 100644 --- a/arch/arm64/kvm/vgic/vgic-debug.c +++ b/arch/arm64/kvm/vgic/vgic-debug.c @@ -166,7 +166,7 @@ static void print_header(struct seq_file *s, struct vgic_irq *irq, if (vcpu) { hdr = "VCPU"; - id = vcpu->vcpu_id; + id = vcpu->vcpu_idx; } seq_printf(s, "\n"); @@ -212,7 +212,7 @@ static void print_irq_state(struct seq_file *s, struct vgic_irq *irq, " %2d " "\n", type, irq->intid, - (irq->target_vcpu) ? irq->target_vcpu->vcpu_id : -1, + (irq->target_vcpu) ? irq->target_vcpu->vcpu_idx : -1, pending, irq->line_level, irq->active, @@ -224,7 +224,7 @@ static void print_irq_state(struct seq_file *s, struct vgic_irq *irq, irq->mpidr, irq->source, irq->priority, - (irq->vcpu) ? irq->vcpu->vcpu_id : -1); + (irq->vcpu) ? irq->vcpu->vcpu_idx : -1); } static int vgic_debug_show(struct seq_file *s, void *v) diff --git a/arch/arm64/kvm/vgic/vgic-kvm-device.c b/arch/arm64/kvm/vgic/vgic-kvm-device.c index 212b73a715c1..82b264ad68c4 100644 --- a/arch/arm64/kvm/vgic/vgic-kvm-device.c +++ b/arch/arm64/kvm/vgic/vgic-kvm-device.c @@ -345,7 +345,7 @@ int vgic_v2_parse_attr(struct kvm_device *dev, struct kvm_device_attr *attr, if (cpuid >= atomic_read(&dev->kvm->online_vcpus)) return -EINVAL; - reg_attr->vcpu = kvm_get_vcpu(dev->kvm, cpuid); + reg_attr->vcpu = kvm_get_vcpu_by_id(dev->kvm, cpuid); reg_attr->addr = attr->attr & KVM_DEV_ARM_VGIC_OFFSET_MASK; return 0;
On 2023/9/11 02:18, Marc Zyngier wrote: > On Sun, 10 Sep 2023 17:25:36 +0100, > Zenghui Yu <zenghui.yu@linux.dev> wrote: >> >> Hi Marc, >> >> I asked because it seems that in kvm/arm64 we always use >> kvm_get_vcpu(kvm, i) to obtain the kvm_vcpu pointer, even if *i* is >> sometimes essentially provided by userspace.. > > Huh, this is incredibly dodgy. I had a go at a few occurrences (see > below), but this is hardly a complete list. Another case is all kvm_get_vcpu(kvm, target_addr) in the vgic-its emulation code. As we expose GITS_TYPER.PTA=0 to guest, which indicates that the target address corresponds to the PE number specified by GICR_TYPER.Processor_Number, which is now encoded as vcpu->vcpu_id. Thanks, Zenghui
On Mon, 11 Sep 2023 16:57:39 +0100, Zenghui Yu <zenghui.yu@linux.dev> wrote: > > On 2023/9/11 02:18, Marc Zyngier wrote: > > On Sun, 10 Sep 2023 17:25:36 +0100, > > Zenghui Yu <zenghui.yu@linux.dev> wrote: > >> > >> Hi Marc, > >> > >> I asked because it seems that in kvm/arm64 we always use > >> kvm_get_vcpu(kvm, i) to obtain the kvm_vcpu pointer, even if *i* is > >> sometimes essentially provided by userspace.. > > > > Huh, this is incredibly dodgy. I had a go at a few occurrences (see > > below), but this is hardly a complete list. > > Another case is all kvm_get_vcpu(kvm, target_addr) in the vgic-its > emulation code. As we expose GITS_TYPER.PTA=0 to guest, which indicates > that the target address corresponds to the PE number specified by > GICR_TYPER.Processor_Number, which is now encoded as vcpu->vcpu_id. Yup, that's indeed missing. I'm going to hack kvmtool to generate stupid vcpu_ids and see what explodes... Thanks, M.
diff --git a/arch/arm64/kvm/vgic/vgic-mmio-v3.c b/arch/arm64/kvm/vgic/vgic-mmio-v3.c index 188d2187eede..88b8d4524854 100644 --- a/arch/arm64/kvm/vgic/vgic-mmio-v3.c +++ b/arch/arm64/kvm/vgic/vgic-mmio-v3.c @@ -1052,6 +1052,38 @@ static int match_mpidr(u64 sgi_aff, u16 sgi_cpu_mask, struct kvm_vcpu *vcpu) ((((reg) & ICC_SGI1R_AFFINITY_## level ##_MASK) \ >> ICC_SGI1R_AFFINITY_## level ##_SHIFT) << MPIDR_LEVEL_SHIFT(level)) +static void vgic_v3_queue_sgi(struct kvm_vcpu *vcpu, u32 sgi, bool allow_group1) +{ + struct vgic_irq *irq = vgic_get_irq(vcpu->kvm, vcpu, sgi); + unsigned long flags; + + raw_spin_lock_irqsave(&irq->irq_lock, flags); + + /* + * An access targeting Group0 SGIs can only generate + * those, while an access targeting Group1 SGIs can + * generate interrupts of either group. + */ + if (!irq->group || allow_group1) { + if (!irq->hw) { + irq->pending_latch = true; + vgic_queue_irq_unlock(vcpu->kvm, irq, flags); + } else { + /* HW SGI? Ask the GIC to inject it */ + int err; + err = irq_set_irqchip_state(irq->host_irq, + IRQCHIP_STATE_PENDING, + true); + WARN_RATELIMIT(err, "IRQ %d", irq->host_irq); + raw_spin_unlock_irqrestore(&irq->irq_lock, flags); + } + } else { + raw_spin_unlock_irqrestore(&irq->irq_lock, flags); + } + + vgic_put_irq(vcpu->kvm, irq); +} + /** * vgic_v3_dispatch_sgi - handle SGI requests from VCPUs * @vcpu: The VCPU requesting a SGI @@ -1070,19 +1102,30 @@ void vgic_v3_dispatch_sgi(struct kvm_vcpu *vcpu, u64 reg, bool allow_group1) { struct kvm *kvm = vcpu->kvm; struct kvm_vcpu *c_vcpu; - u16 target_cpus; + unsigned long target_cpus; u64 mpidr; - int sgi; - int vcpu_id = vcpu->vcpu_id; - bool broadcast; - unsigned long c, flags; - - sgi = (reg & ICC_SGI1R_SGI_ID_MASK) >> ICC_SGI1R_SGI_ID_SHIFT; - broadcast = reg & BIT_ULL(ICC_SGI1R_IRQ_ROUTING_MODE_BIT); - target_cpus = (reg & ICC_SGI1R_TARGET_LIST_MASK) >> ICC_SGI1R_TARGET_LIST_SHIFT; + u32 sgi; + unsigned long c; + + sgi = FIELD_GET(ICC_SGI1R_SGI_ID_MASK, reg); + + /* Broadcast */ + if (unlikely(reg & BIT_ULL(ICC_SGI1R_IRQ_ROUTING_MODE_BIT))) { + kvm_for_each_vcpu(c, c_vcpu, kvm) { + /* Don't signal the calling VCPU */ + if (c_vcpu == vcpu) + continue; + + vgic_v3_queue_sgi(c_vcpu, sgi, allow_group1); + } + + return; + } + mpidr = SGI_AFFINITY_LEVEL(reg, 3); mpidr |= SGI_AFFINITY_LEVEL(reg, 2); mpidr |= SGI_AFFINITY_LEVEL(reg, 1); + target_cpus = FIELD_GET(ICC_SGI1R_TARGET_LIST_MASK, reg); /* * We iterate over all VCPUs to find the MPIDRs matching the request. @@ -1091,54 +1134,19 @@ void vgic_v3_dispatch_sgi(struct kvm_vcpu *vcpu, u64 reg, bool allow_group1) * VCPUs when most of the times we just signal a single VCPU. */ kvm_for_each_vcpu(c, c_vcpu, kvm) { - struct vgic_irq *irq; + int level0; /* Exit early if we have dealt with all requested CPUs */ - if (!broadcast && target_cpus == 0) + if (target_cpus == 0) break; - - /* Don't signal the calling VCPU */ - if (broadcast && c == vcpu_id) + level0 = match_mpidr(mpidr, target_cpus, c_vcpu); + if (level0 == -1) continue; - if (!broadcast) { - int level0; - - level0 = match_mpidr(mpidr, target_cpus, c_vcpu); - if (level0 == -1) - continue; - - /* remove this matching VCPU from the mask */ - target_cpus &= ~BIT(level0); - } + /* remove this matching VCPU from the mask */ + target_cpus &= ~BIT(level0); - irq = vgic_get_irq(vcpu->kvm, c_vcpu, sgi); - - raw_spin_lock_irqsave(&irq->irq_lock, flags); - - /* - * An access targeting Group0 SGIs can only generate - * those, while an access targeting Group1 SGIs can - * generate interrupts of either group. - */ - if (!irq->group || allow_group1) { - if (!irq->hw) { - irq->pending_latch = true; - vgic_queue_irq_unlock(vcpu->kvm, irq, flags); - } else { - /* HW SGI? Ask the GIC to inject it */ - int err; - err = irq_set_irqchip_state(irq->host_irq, - IRQCHIP_STATE_PENDING, - true); - WARN_RATELIMIT(err, "IRQ %d", irq->host_irq); - raw_spin_unlock_irqrestore(&irq->irq_lock, flags); - } - } else { - raw_spin_unlock_irqrestore(&irq->irq_lock, flags); - } - - vgic_put_irq(vcpu->kvm, irq); + vgic_v3_queue_sgi(c_vcpu, sgi, allow_group1); } }
As we're about to change the way SGIs are sent, start by splitting out some of the basic functionnality: instead of intermingling the broadcast and non-broadcast cases with the actual SGI generation, perform the following cleanups: - move the SGI queuing into its own helper - split the broadcast code from the affinity-driven code - replace the mask/shift combinations with FIELD_GET() The result is much more readable, and paves the way for further optimisations. Signed-off-by: Marc Zyngier <maz@kernel.org> --- arch/arm64/kvm/vgic/vgic-mmio-v3.c | 110 ++++++++++++++++------------- 1 file changed, 59 insertions(+), 51 deletions(-)