drm/radeon: add golden register settings for Hainan (v2)
[deliverable/linux.git] / drivers / hv / channel_mgmt.c
index 2f84c5cff8d4f95a026a42034eb1fac960ec6e6a..ff1be167eb040f76950f708a308186d30615ac6f 100644 (file)
@@ -257,6 +257,70 @@ static void vmbus_process_offer(struct work_struct *work)
        }
 }
 
+enum {
+       IDE = 0,
+       SCSI,
+       NIC,
+       MAX_PERF_CHN,
+};
+
+/*
+ * This is an array of device_ids (device types) that are performance critical.
+ * We attempt to distribute the interrupt load for these devices across
+ * all available CPUs.
+ */
+static const struct hv_vmbus_device_id hp_devs[] = {
+       /* IDE */
+       { HV_IDE_GUID, },
+       /* Storage - SCSI */
+       { HV_SCSI_GUID, },
+       /* Network */
+       { HV_NIC_GUID, },
+};
+
+
+/*
+ * We use this state to statically distribute the channel interrupt load.
+ */
+static u32  next_vp;
+
+/*
+ * Starting with Win8, we can statically distribute the incoming
+ * channel interrupt load by binding a channel to VCPU. We
+ * implement here a simple round robin scheme for distributing
+ * the interrupt load.
+ * We will bind channels that are not performance critical to cpu 0 and
+ * performance critical channels (IDE, SCSI and Network) will be uniformly
+ * distributed across all available CPUs.
+ */
+static u32 get_vp_index(uuid_le *type_guid)
+{
+       u32 cur_cpu;
+       int i;
+       bool perf_chn = false;
+       u32 max_cpus = num_online_cpus();
+
+       for (i = IDE; i < MAX_PERF_CHN; i++) {
+               if (!memcmp(type_guid->b, hp_devs[i].guid,
+                                sizeof(uuid_le))) {
+                       perf_chn = true;
+                       break;
+               }
+       }
+       if ((vmbus_proto_version == VERSION_WS2008) ||
+           (vmbus_proto_version == VERSION_WIN7) || (!perf_chn)) {
+               /*
+                * Prior to win8, all channel interrupts are
+                * delivered on cpu 0.
+                * Also if the channel is not a performance critical
+                * channel, bind it to cpu 0.
+                */
+               return 0;
+       }
+       cur_cpu = (++next_vp % max_cpus);
+       return cur_cpu;
+}
+
 /*
  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
  *
@@ -275,6 +339,35 @@ static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
                return;
        }
 
+       /*
+        * By default we setup state to enable batched
+        * reading. A specific service can choose to
+        * disable this prior to opening the channel.
+        */
+       newchannel->batched_reading = true;
+
+       /*
+        * Setup state for signalling the host.
+        */
+       newchannel->sig_event = (struct hv_input_signal_event *)
+                               (ALIGN((unsigned long)
+                               &newchannel->sig_buf,
+                               HV_HYPERCALL_PARAM_ALIGN));
+
+       newchannel->sig_event->connectionid.asu32 = 0;
+       newchannel->sig_event->connectionid.u.id = VMBUS_EVENT_CONNECTION_ID;
+       newchannel->sig_event->flag_number = 0;
+       newchannel->sig_event->rsvdz = 0;
+
+       if (vmbus_proto_version != VERSION_WS2008) {
+               newchannel->is_dedicated_interrupt =
+                               (offer->is_dedicated_interrupt != 0);
+               newchannel->sig_event->connectionid.u.id =
+                               offer->connection_id;
+       }
+
+       newchannel->target_vp = get_vp_index(&offer->offer.if_type);
+
        memcpy(&newchannel->offermsg, offer,
               sizeof(struct vmbus_channel_offer_channel));
        newchannel->monitor_grp = (u8)offer->monitorid / 32;
This page took 0.024774 seconds and 5 git commands to generate.