New Paravirtual front-end for PCI in a Hyper-V VM

drivers/pci/host/hv_pcifront.c new file +7 -0

+#include <linux/kernel.h>
+#include <linux/jiffies.h>
+#include <linux/mman.h>
+#include <linux/delay.h>
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/slab.h>

drivers/pci/host/hv_pcifront.c new file +5 -0

+/* Interrupt management hooks */
+
+/**
+ * hv_msi_free() - Free the MSI.
+ * @domain:	The interrupt domain pointer

drivers/pci/host/hv_pcifront.c new file +6 -0

+{
+	struct pci_delete_interrupt *int_pkt;
+	struct {
+		struct pci_packet pkt;
+		u8 buffer[sizeof(struct pci_delete_interrupt) -
+			  sizeof(struct pci_message)];

drivers/pci/host/hv_pcifront.c new file +43 -0

+{
+	struct irq_cfg *cfg = irqd_cfg(data);
+	struct hv_pcibus_device *hbus;
+	struct hv_pci_dev *hpdev;
+	struct pci_bus *pbus;
+	struct pci_create_interrupt *int_pkt;
+	struct compose_comp_ctxt comp;
+	struct tran_int_desc *int_desc;
+	struct irq_desc *irq_desc;
+	struct {
+		struct pci_packet pkt;
+		u8 buffer[sizeof(struct pci_create_interrupt) -
+			  sizeof(struct pci_message)];
+	} ctxt;
+	int cpu;
+	int ret;
+
+	pbus = data->msi_desc->dev->bus;
+	hbus = container_of(pbus->sysdata, struct hv_pcibus_device, sysdata);
+	hpdev = lookup_hv_dev(hbus, devfn_to_wslot(data->msi_desc->dev->devfn));
+
+	if (!hpdev)
+		goto return_null_message;
+
+	int_desc = kzalloc(sizeof(*int_desc), GFP_KERNEL);
+	if (!int_desc)
+		goto return_null_message;
+
+	memset(&ctxt, 0, sizeof(ctxt));
+	init_completion(&comp.comp_pkt.host_event);
+	ctxt.pkt.completion_func = hv_pci_compose_compl;
+	ctxt.pkt.compl_ctxt = ∁
+	int_pkt = (struct pci_create_interrupt *)&ctxt.pkt.message;
+	int_pkt->message_type.message_type = PCI_CREATE_INTERRUPT_MESSAGE;
+	int_pkt->wslot.slot = hpdev->desc.win_slot.slot;
+	int_pkt->int_desc.vector = cfg->vector;
+	int_pkt->int_desc.vector_count = 1;
+	int_pkt->int_desc.delivery_mode =
+		(apic->irq_delivery_mode == dest_LowestPrio) ? 1 : 0;
+
+	/*
+	 * Cut down interrupt targets to a single processor. Hyper-V
+	 * currently only supports picking the target of an MSI by

drivers/pci/host/hv_pcifront.c new file +11 -0

+	ret = vmbus_sendpacket(hpdev->hbus->hdev->channel, int_pkt,
+			       sizeof(*int_pkt), (unsigned long)&ctxt.pkt,
+			       VM_PKT_DATA_INBAND,
+			       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
+	if (!ret)
+		wait_for_completion(&comp.comp_pkt.host_event);
+
+	if (comp.comp_pkt.completion_status < 0) {
+		pr_err("Request for interrupt failed: 0x%x",
+		       comp.comp_pkt.completion_status);
+		goto return_null_message;

drivers/pci/host/hv_pcifront.c new file +16 -0

+};
+
+static irq_hw_number_t hv_msi_domain_ops_get_hwirq(struct msi_domain_info *info,
+						   msi_alloc_info_t *arg)
+{
+	return arg->msi_hwirq;
+}
+
+static int hv_msi_domain_ops_prepare(struct irq_domain *domain,
+				     struct device *dev, int nvec,
+				     msi_alloc_info_t *arg)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct msi_desc *desc = first_pci_msi_entry(pdev);
+
+	memset(arg, 0, sizeof(*arg));

-----Original Message-----
From: Thomas Gleixner [mailto:[email protected]]
Sent: Sunday, August 2, 2015 1:47 AM
To: Jake Oshins <[email protected]>
Cc: [email protected]; KY Srinivasan <[email protected]>; LKML <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; linux- [email protected]; [email protected]; [email protected]; Jiang Liu <[email protected]>
Subject: Re: [PATCH 4/4] drivers:pci:hv: New paravirtual PCI front-end for Hyper-V VMs

Quoted 8 lines

drivers/pci/host/hv_pcifront.c new file +8 -0

+static int hv_msi_domain_ops_prepare(struct irq_domain *domain,
+				     struct device *dev, int nvec,
+				     msi_alloc_info_t *arg)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct msi_desc *desc = first_pci_msi_entry(pdev);
+
+	memset(arg, 0, sizeof(*arg));

drivers/pci/host/hv_pcifront.c line 855

Quoted 10 lines

drivers/pci/host/hv_pcifront.c new file +10 -0

+	arg->msi_dev = pdev;
+	if (desc->msi_attrib.is_msix) {
+		arg->type = X86_IRQ_ALLOC_TYPE_MSIX;
+	} else {
+		arg->type = X86_IRQ_ALLOC_TYPE_MSI;
+		arg->flags |= X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
+	}
+
+	return 0;
+}

Pretty much a copy of the x86 code, right?

I wonder whether this MSI infrastructure code would be better seperated out and moved to arch/x86/hyperv/msi.c or arch/x86/kernel/apic/hvmsi.c. It's small enough to be built in. So all you'd need to export is hv_pcie_init_irq_domain and hv_pcie_free_irq_domain.

Thanks,

tglx

I do have a question about your last point, though. If I build this into the kernel, it will need to depend on sending and receiving messages through the hv_vmbus driver, which isn't built in. It

seemed like the indirection and glue code necessary to make that work would be almost as big as this entire implementation (which admittedly isn't very big.) If you prefer that, I'll do it. But it would make more sense to me to refactor this a bit more so that functions like the one above are exported rather than putting hv_pcie_init_irq_domain and hv_pcie_free_irq_domain into the kernel itself. Would that work?

drivers/pci/host/hv_pcifront.c new file +7 -0

 
+static void exit_hv_pci_drv(void)
+{
+	vmbus_driver_unregister(&hv_pci_drv);
+}
+
+static int __init init_hv_pci_drv(void)
+{

Quoted 10 lines · On zo, 2015-08-02 at 00:54 +0000, [email protected] wrote:

drivers/pci/host/hv_pcifront.c new file +7 -0

 
+static void exit_hv_pci_drv(void)
+{
+	vmbus_driver_unregister(&hv_pci_drv);
+}
+
+static int __init init_hv_pci_drv(void)
+{

drivers/pci/host/hv_pcifront.c line 2222

Quoted 10 lines

drivers/pci/host/hv_pcifront.c new file +10 -0

+	int ret;
+
+	/* Register this driver with VMBus. */
+	ret = vmbus_driver_register(&hv_pci_drv);
+
+	if (ret < 0)
+		exit_hv_pci_drv();
+
+	return ret;
+}

Wouldn't just

    static int __init init_hv_pci_drv(void)
    {
    	return vmbus_driver_register(&hv_pci_drv);
    }

suffice here?