Michael Kelley
2021-Feb-04 16:41 UTC
[PATCH v5 16/16] iommu/hyperv: setup an IO-APIC IRQ remapping domain for root partition
From: Wei Liu <wei.liu at kernel.org> Sent: Wednesday, February 3, 2021 4:47 AM> > On Wed, Jan 27, 2021 at 05:47:08AM +0000, Michael Kelley wrote: > > From: Wei Liu <wei.liu at kernel.org> Sent: Wednesday, January 20, 2021 4:01 AM > > > > > > Just like MSI/MSI-X, IO-APIC interrupts are remapped by Microsoft > > > Hypervisor when Linux runs as the root partition. Implement an IRQ > > > domain to handle mapping and unmapping of IO-APIC interrupts. > > > > > > Signed-off-by: Wei Liu <wei.liu at kernel.org> > > > --- > > > arch/x86/hyperv/irqdomain.c | 54 ++++++++++ > > > arch/x86/include/asm/mshyperv.h | 4 + > > > drivers/iommu/hyperv-iommu.c | 179 +++++++++++++++++++++++++++++++- > > > 3 files changed, 233 insertions(+), 4 deletions(-) > > > > > > diff --git a/arch/x86/hyperv/irqdomain.c b/arch/x86/hyperv/irqdomain.c > > > index 19637cd60231..8e2b4e478b70 100644 > > > --- a/arch/x86/hyperv/irqdomain.c > > > +++ b/arch/x86/hyperv/irqdomain.c > > > @@ -330,3 +330,57 @@ struct irq_domain * __init hv_create_pci_msi_domain(void) > > > } > > > > > > #endif /* CONFIG_PCI_MSI */ > > > + > > > +int hv_unmap_ioapic_interrupt(int ioapic_id, struct hv_interrupt_entry *entry) > > > +{ > > > + union hv_device_id device_id; > > > + > > > + device_id.as_uint64 = 0; > > > + device_id.device_type = HV_DEVICE_TYPE_IOAPIC; > > > + device_id.ioapic.ioapic_id = (u8)ioapic_id; > > > + > > > + return hv_unmap_interrupt(device_id.as_uint64, entry) & > HV_HYPERCALL_RESULT_MASK; > > > > The masking is already done in hv_unmap_interrupt. > > Fixed. > > > > > > +} > > > +EXPORT_SYMBOL_GPL(hv_unmap_ioapic_interrupt); > > > + > > > +int hv_map_ioapic_interrupt(int ioapic_id, bool level, int vcpu, int vector, > > > + struct hv_interrupt_entry *entry) > > > +{ > > > + unsigned long flags; > > > + struct hv_input_map_device_interrupt *input; > > > + struct hv_output_map_device_interrupt *output; > > > + union hv_device_id device_id; > > > + struct hv_device_interrupt_descriptor *intr_desc; > > > + u16 status; > > > + > > > + device_id.as_uint64 = 0; > > > + device_id.device_type = HV_DEVICE_TYPE_IOAPIC; > > > + device_id.ioapic.ioapic_id = (u8)ioapic_id; > > > + > > > + local_irq_save(flags); > > > + input = *this_cpu_ptr(hyperv_pcpu_input_arg); > > > + output = *this_cpu_ptr(hyperv_pcpu_output_arg); > > > + memset(input, 0, sizeof(*input)); > > > + intr_desc = &input->interrupt_descriptor; > > > + input->partition_id = hv_current_partition_id; > > > + input->device_id = device_id.as_uint64; > > > + intr_desc->interrupt_type = HV_X64_INTERRUPT_TYPE_FIXED; > > > + intr_desc->target.vector = vector; > > > + intr_desc->vector_count = 1; > > > + > > > + if (level) > > > + intr_desc->trigger_mode = HV_INTERRUPT_TRIGGER_MODE_LEVEL; > > > + else > > > + intr_desc->trigger_mode = HV_INTERRUPT_TRIGGER_MODE_EDGE; > > > + > > > + __set_bit(vcpu, (unsigned long *)&intr_desc->target.vp_mask); > > > + > > > + status = hv_do_rep_hypercall(HVCALL_MAP_DEVICE_INTERRUPT, 0, 0, input, > output) & > > > + HV_HYPERCALL_RESULT_MASK; > > > + local_irq_restore(flags); > > > + > > > + *entry = output->interrupt_entry; > > > + > > > + return status; > > > > As a cross-check, I was comparing this code against hv_map_msi_interrupt(). They are > > mostly parallel, though some of the assignments are done in a different order. It's a nit, > > but making them as parallel as possible would be nice. :-) > > > > Indeed. I will see about factoring out a function.If factoring out a separate helper function is clumsy, just having the parallel code in the two functions be as similar as possible makes it easier to see what's the same and what's different.> > > Same 64 vCPU comment applies here as well. > > > > This is changed to use vpset instead. Took me a bit of time to get it > working because document is a bit lacking. > > > > > > +} > > > +EXPORT_SYMBOL_GPL(hv_map_ioapic_interrupt); > > > diff --git a/arch/x86/include/asm/mshyperv.h b/arch/x86/include/asm/mshyperv.h > > > index ccc849e25d5e..345d7c6f8c37 100644 > > > --- a/arch/x86/include/asm/mshyperv.h > > > +++ b/arch/x86/include/asm/mshyperv.h > > > @@ -263,6 +263,10 @@ static inline void hv_set_msi_entry_from_desc(union > > > hv_msi_entry *msi_entry, > > > > > > struct irq_domain *hv_create_pci_msi_domain(void); > > > > > > +int hv_map_ioapic_interrupt(int ioapic_id, bool level, int vcpu, int vector, > > > + struct hv_interrupt_entry *entry); > > > +int hv_unmap_ioapic_interrupt(int ioapic_id, struct hv_interrupt_entry *entry); > > > + > > > #else /* CONFIG_HYPERV */ > > > static inline void hyperv_init(void) {} > > > static inline void hyperv_setup_mmu_ops(void) {} > > > diff --git a/drivers/iommu/hyperv-iommu.c b/drivers/iommu/hyperv-iommu.c > > > index b7db6024e65c..6d35e4c303c6 100644 > > > --- a/drivers/iommu/hyperv-iommu.c > > > +++ b/drivers/iommu/hyperv-iommu.c > > > @@ -116,30 +116,43 @@ static const struct irq_domain_ops hyperv_ir_domain_ops = { > > > .free = hyperv_irq_remapping_free, > > > }; > > > > > > +static const struct irq_domain_ops hyperv_root_ir_domain_ops; > > > static int __init hyperv_prepare_irq_remapping(void) > > > { > > > struct fwnode_handle *fn; > > > int i; > > > + const char *name; > > > + const struct irq_domain_ops *ops; > > > > > > if (!hypervisor_is_type(X86_HYPER_MS_HYPERV) || > > > x86_init.hyper.msi_ext_dest_id() || > > > - !x2apic_supported() || hv_root_partition) > > > + !x2apic_supported()) > > > > Any reason that the check for hv_root_partition was added > > in patch #4 of this series, and then removed here? Could > > patch #4 just be dropped? > > > > Before v5 (or v4?) IO-APIC was not handled via Hyper-V IOMMU. Now it is. > > Patch 4 has become redundant with that change. I already dropped patch 4 > in the v6 branch I have locally. > > > > return -ENODEV; > > > > > > - fn = irq_domain_alloc_named_id_fwnode("HYPERV-IR", 0); > > > + if (hv_root_partition) { > > > + name = "HYPERV-ROOT-IR"; > > > + ops = &hyperv_root_ir_domain_ops; > > > + } else { > > > + name = "HYPERV-IR"; > > > + ops = &hyperv_ir_domain_ops; > > > + } > > > + > [...] > > > +static void > > > +hyperv_root_ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg) > > > +{ > > > + u16 status; > > > + u32 vector; > > > + struct irq_cfg *cfg; > > > + int ioapic_id; > > > + struct cpumask *affinity; > > > + int cpu, vcpu; > > > + struct hv_interrupt_entry entry; > > > + struct hyperv_root_ir_data *data = irq_data->chip_data; > > > + struct IO_APIC_route_entry e; > > > + > > > + cfg = irqd_cfg(irq_data); > > > + affinity = irq_data_get_effective_affinity_mask(irq_data); > > > + cpu = cpumask_first_and(affinity, cpu_online_mask); > > > + vcpu = hv_cpu_number_to_vp_number(cpu); > > > + > > > + vector = cfg->vector; > > > + ioapic_id = data->ioapic_id; > > > + > > > + if (data->entry.source == HV_DEVICE_TYPE_IOAPIC > > > > Does 'data' need to be checked to be non-NULL? The parallel code in > > hv_irq_compose_msi_msg() makes such a check. > > The usage of irq_data->chip_data is different in these two functions. > > In this function, we're sure it is correctly allocated by > hyperv_root_ir_remapping_alloc at some point before. > > In hv_irq_compose_msi_msg, irq_data->chip_data is instead used as a > temporary place to stash some state that is controlled solely by the > said function. > > Once we get to the point of introducing a paravirtualized IOMMU for the > root partition, we can then unify these two paths.OK, thanks for the explanation.> > > > > > + && data->entry.ioapic_rte.as_uint64) { > > > + entry = data->entry; > > > + > > > + status = hv_unmap_ioapic_interrupt(ioapic_id, &entry); > > > + > > > + if (status != HV_STATUS_SUCCESS) > > > + pr_debug("%s: unexpected unmap status %d\n", __func__, > status); > > > + > > > + data->entry.ioapic_rte.as_uint64 = 0; > > > + data->entry.source = 0; /* Invalid source */ > > > > Again comparing, hv_irq_compose_msi_msg() frees the old > > entry, and then allocates a new one. This code reuses the old entry. > > Any reason for the difference? > > > > See above. > > I can perhaps tweak the logic a bit to reuse the same entry, but the > overall design won't change. I opted to always reallocate because that > looked more straight-forward to me. > > Let me know if you feel strongly about reusing.I don't feel strongly about reusing. I was just comparing/contrasting the two functions.> > > > + } > > > + > > > + > > > + status = hv_map_ioapic_interrupt(ioapic_id, data->is_level, vcpu, > > > + vector, &entry); > > > + > > > + if (status != HV_STATUS_SUCCESS) { > > > + pr_err("%s: map hypercall failed, status %d\n", __func__, status); > > > + return; > > > + } > > > + > > > + data->entry = entry; > > > + > > > + /* Turn it into an IO_APIC_route_entry, and generate MSI MSG. */ > > > + e.w1 = entry.ioapic_rte.low_uint32; > > > + e.w2 = entry.ioapic_rte.high_uint32; > > > + > > > + memset(msg, 0, sizeof(*msg)); > > > + msg->arch_data.vector = e.vector; > > > + msg->arch_data.delivery_mode = e.delivery_mode; > > > + msg->arch_addr_lo.dest_mode_logical = e.dest_mode_logical; > > > + msg->arch_addr_lo.dmar_format = e.ir_format; > > > + msg->arch_addr_lo.dmar_index_0_14 = e.ir_index_0_14; > > > +} > > > > Having this whole function be more parallel to hv_irq_compose_msi_msg() > > would be nice. :-) > > > > Unlike hv_map_ioapic_interrupt and hv_map_msi_interrupt, which can > benefit from unifying now, this and hv_irq_compose_msi_msg will need to > wait till we have an IOMMU for the reason I stated above.OK. Just having the code in the two functions be more parallel where possible would make it easier to see similarities and differences. But it's not a big deal.> > > > + > > > +static int hyperv_root_ir_set_affinity(struct irq_data *data, > > > + const struct cpumask *mask, bool force) > > > +{ > > > + struct irq_data *parent = data->parent_data; > > > + struct irq_cfg *cfg = irqd_cfg(data); > > > + int ret; > > > + > > > + ret = parent->chip->irq_set_affinity(parent, mask, force); > > > + if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE) > > > + return ret; > > > + > > > + send_cleanup_vector(cfg); > > > + > > > + return 0; > > > +} > > > + > [...] > > > + > > > +static void hyperv_root_irq_remapping_free(struct irq_domain *domain, > > > + unsigned int virq, unsigned int nr_irqs) > > > +{ > > > + struct irq_data *irq_data; > > > + struct hyperv_root_ir_data *data; > > > + struct hv_interrupt_entry *e; > > > + int i; > > > + > > > + for (i = 0; i < nr_irqs; i++) { > > > + irq_data = irq_domain_get_irq_data(domain, virq + i); > > > + > > > + if (irq_data && irq_data->chip_data) { > > > + data = irq_data->chip_data; > > > > Set irq_data->chip_data to NULL? That seems to be done in other > > similar places in your code. > > There is no need to do that. By the time this function returns, irq_data > will be gone too -- freed by irq_domain_free_irqs_common.OK> > > > > > + e = &data->entry; > > > + > > > + if (e->source == HV_DEVICE_TYPE_IOAPIC > > > + && e->ioapic_rte.as_uint64) > > > + hv_unmap_ioapic_interrupt(data->ioapic_id, > > > + &data->entry); > > > + > > > + kfree(data); > > > + } > > > + } > > > + > > > + irq_domain_free_irqs_common(domain, virq, nr_irqs); > > > +} > > > + > > > +static const struct irq_domain_ops hyperv_root_ir_domain_ops = { > > > + .select = hyperv_irq_remapping_select, > > > + .alloc = hyperv_root_irq_remapping_alloc, > > > + .free = hyperv_root_irq_remapping_free, > > > +}; > > > + > > > #endif > > > -- > > > 2.20.1 > >