PCI: hv: Add support for protocol 1.3 and support PCI_BUS_RELATIONS2
authorLong Li <longli@microsoft.com>
Wed, 26 Feb 2020 05:06:08 +0000 (21:06 -0800)
committerLorenzo Pieralisi <lorenzo.pieralisi@arm.com>
Fri, 6 Mar 2020 10:18:10 +0000 (10:18 +0000)
Starting with Hyper-V PCI protocol version 1.3, the host VSP can send
PCI_BUS_RELATIONS2 and pass the vNUMA node information for devices on the
bus. The vNUMA node tells which guest NUMA node this device is on based
on guest VM configuration topology and physical device information.

Add code to negotiate v1.3 and process PCI_BUS_RELATIONS2.

Signed-off-by: Long Li <longli@microsoft.com>
Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
Reviewed-by: Michael Kelley <mikelley@microsoft.com>
drivers/pci/controller/pci-hyperv.c

index dea197f..98d3776 100644 (file)
@@ -63,6 +63,7 @@
 enum pci_protocol_version_t {
        PCI_PROTOCOL_VERSION_1_1 = PCI_MAKE_VERSION(1, 1),      /* Win10 */
        PCI_PROTOCOL_VERSION_1_2 = PCI_MAKE_VERSION(1, 2),      /* RS1 */
+       PCI_PROTOCOL_VERSION_1_3 = PCI_MAKE_VERSION(1, 3),      /* Vibranium */
 };
 
 #define CPU_AFFINITY_ALL       -1ULL
@@ -72,6 +73,7 @@ enum pci_protocol_version_t {
  * first.
  */
 static enum pci_protocol_version_t pci_protocol_versions[] = {
+       PCI_PROTOCOL_VERSION_1_3,
        PCI_PROTOCOL_VERSION_1_2,
        PCI_PROTOCOL_VERSION_1_1,
 };
@@ -119,6 +121,7 @@ enum pci_message_type {
        PCI_RESOURCES_ASSIGNED2         = PCI_MESSAGE_BASE + 0x16,
        PCI_CREATE_INTERRUPT_MESSAGE2   = PCI_MESSAGE_BASE + 0x17,
        PCI_DELETE_INTERRUPT_MESSAGE2   = PCI_MESSAGE_BASE + 0x18, /* unused */
+       PCI_BUS_RELATIONS2              = PCI_MESSAGE_BASE + 0x19,
        PCI_MESSAGE_MAXIMUM
 };
 
@@ -164,6 +167,26 @@ struct pci_function_description {
        u32     ser;    /* serial number */
 } __packed;
 
+enum pci_device_description_flags {
+       HV_PCI_DEVICE_FLAG_NONE                 = 0x0,
+       HV_PCI_DEVICE_FLAG_NUMA_AFFINITY        = 0x1,
+};
+
+struct pci_function_description2 {
+       u16     v_id;   /* vendor ID */
+       u16     d_id;   /* device ID */
+       u8      rev;
+       u8      prog_intf;
+       u8      subclass;
+       u8      base_class;
+       u32     subsystem_id;
+       union   win_slot_encoding win_slot;
+       u32     ser;    /* serial number */
+       u32     flags;
+       u16     virtual_numa_node;
+       u16     reserved;
+} __packed;
+
 /**
  * struct hv_msi_desc
  * @vector:            IDT entry
@@ -299,6 +322,12 @@ struct pci_bus_relations {
        struct pci_function_description func[0];
 } __packed;
 
+struct pci_bus_relations2 {
+       struct pci_incoming_message incoming;
+       u32 device_count;
+       struct pci_function_description2 func[0];
+} __packed;
+
 struct pci_q_res_req_response {
        struct vmpacket_descriptor hdr;
        s32 status;                     /* negative values are failures */
@@ -1415,6 +1444,7 @@ static void hv_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
                break;
 
        case PCI_PROTOCOL_VERSION_1_2:
+       case PCI_PROTOCOL_VERSION_1_3:
                size = hv_compose_msi_req_v2(&ctxt.int_pkts.v2,
                                        dest,
                                        hpdev->desc.win_slot.slot,
@@ -1813,6 +1843,27 @@ static void hv_pci_remove_slots(struct hv_pcibus_device *hbus)
        }
 }
 
+/*
+ * Set NUMA node for the devices on the bus
+ */
+static void hv_pci_assign_numa_node(struct hv_pcibus_device *hbus)
+{
+       struct pci_dev *dev;
+       struct pci_bus *bus = hbus->pci_bus;
+       struct hv_pci_dev *hv_dev;
+
+       list_for_each_entry(dev, &bus->devices, bus_list) {
+               hv_dev = get_pcichild_wslot(hbus, devfn_to_wslot(dev->devfn));
+               if (!hv_dev)
+                       continue;
+
+               if (hv_dev->desc.flags & HV_PCI_DEVICE_FLAG_NUMA_AFFINITY)
+                       set_dev_node(&dev->dev, hv_dev->desc.virtual_numa_node);
+
+               put_pcichild(hv_dev);
+       }
+}
+
 /**
  * create_root_hv_pci_bus() - Expose a new root PCI bus
  * @hbus:      Root PCI bus, as understood by this driver
@@ -1835,6 +1886,7 @@ static int create_root_hv_pci_bus(struct hv_pcibus_device *hbus)
 
        pci_lock_rescan_remove();
        pci_scan_child_bus(hbus->pci_bus);
+       hv_pci_assign_numa_node(hbus);
        pci_bus_assign_resources(hbus->pci_bus);
        hv_pci_assign_slots(hbus);
        pci_bus_add_devices(hbus->pci_bus);
@@ -2103,6 +2155,7 @@ static void pci_devices_present_work(struct work_struct *work)
                 */
                pci_lock_rescan_remove();
                pci_scan_child_bus(hbus->pci_bus);
+               hv_pci_assign_numa_node(hbus);
                hv_pci_assign_slots(hbus);
                pci_unlock_rescan_remove();
                break;
@@ -2207,6 +2260,46 @@ static void hv_pci_devices_present(struct hv_pcibus_device *hbus,
 }
 
 /**
+ * hv_pci_devices_present2() - Handle list of new children
+ * @hbus:      Root PCI bus, as understood by this driver
+ * @relations: Packet from host listing children
+ *
+ * This function is the v2 version of hv_pci_devices_present()
+ */
+static void hv_pci_devices_present2(struct hv_pcibus_device *hbus,
+                                   struct pci_bus_relations2 *relations)
+{
+       struct hv_dr_state *dr;
+       int i;
+
+       dr = kzalloc(offsetof(struct hv_dr_state, func) +
+                    (sizeof(struct hv_pcidev_description) *
+                     (relations->device_count)), GFP_NOWAIT);
+
+       if (!dr)
+               return;
+
+       dr->device_count = relations->device_count;
+       for (i = 0; i < dr->device_count; i++) {
+               dr->func[i].v_id = relations->func[i].v_id;
+               dr->func[i].d_id = relations->func[i].d_id;
+               dr->func[i].rev = relations->func[i].rev;
+               dr->func[i].prog_intf = relations->func[i].prog_intf;
+               dr->func[i].subclass = relations->func[i].subclass;
+               dr->func[i].base_class = relations->func[i].base_class;
+               dr->func[i].subsystem_id = relations->func[i].subsystem_id;
+               dr->func[i].win_slot = relations->func[i].win_slot;
+               dr->func[i].ser = relations->func[i].ser;
+               dr->func[i].flags = relations->func[i].flags;
+               dr->func[i].virtual_numa_node =
+                       relations->func[i].virtual_numa_node;
+       }
+
+       if (hv_pci_start_relations_work(hbus, dr))
+               kfree(dr);
+}
+
+/**
  * hv_eject_device_work() - Asynchronously handles ejection
  * @work:      Work struct embedded in internal device struct
  *
@@ -2319,6 +2412,7 @@ static void hv_pci_onchannelcallback(void *context)
        struct pci_response *response;
        struct pci_incoming_message *new_message;
        struct pci_bus_relations *bus_rel;
+       struct pci_bus_relations2 *bus_rel2;
        struct pci_dev_inval_block *inval;
        struct pci_dev_incoming *dev_message;
        struct hv_pci_dev *hpdev;
@@ -2386,6 +2480,21 @@ static void hv_pci_onchannelcallback(void *context)
                                hv_pci_devices_present(hbus, bus_rel);
                                break;
 
+                       case PCI_BUS_RELATIONS2:
+
+                               bus_rel2 = (struct pci_bus_relations2 *)buffer;
+                               if (bytes_recvd <
+                                   offsetof(struct pci_bus_relations2, func) +
+                                   (sizeof(struct pci_function_description2) *
+                                    (bus_rel2->device_count))) {
+                                       dev_err(&hbus->hdev->device,
+                                               "bus relations v2 too small\n");
+                                       break;
+                               }
+
+                               hv_pci_devices_present2(hbus, bus_rel2);
+                               break;
+
                        case PCI_EJECT:
 
                                dev_message = (struct pci_dev_incoming *)buffer;