Re: [PATCH v5 16/16] iommu/hyperv: setup an IO-APIC IRQ remapping domain for root partition

On Wed, Jan 27, 2021 at 05:47:08AM +0000, Michael Kelley wrote:

From: Wei Liu <wei.liu@kernel.org> Sent: Wednesday, January 20, 2021 4:01 AM

quoted

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@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.

quoted

+}
+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.

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.

quoted

+}
+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.

quoted

 		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;
+	}
+

[...]

quoted

+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.

quoted

+	    && 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.

quoted

+	}
+
+
+	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.

quoted

+
+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;
+}
+

[...]

quoted

+
+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.

quoted

+			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