Drivers: hv: vmbus: Improve the CPU affiliation for channels

The current code tracks the assigned CPUs within a NUMA node in the context of
the primary channel. So, if we have a VM with a single NUMA node with 8 VCPUs, we may
end up unevenly distributing the channel load. Fix the issue by tracking affiliations
globally.

Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/hv/channel_mgmt.c

@@ -392,6 +392,7 @@ static void init_vp_index(struct vmbus_channel *channel, const uuid_le *type_gui
 	struct vmbus_channel *primary = channel->primary_channel;
 	int next_node;
 	struct cpumask available_mask;
+	struct cpumask *alloced_mask;
 
 	for (i = IDE; i < MAX_PERF_CHN; i++) {
 		if (!memcmp(type_guid->b, hp_devs[i].guid,
@@ -409,7 +410,6 @@ static void init_vp_index(struct vmbus_channel *channel, const uuid_le *type_gui
 		 * channel, bind it to cpu 0.
 		 */
 		channel->numa_node = 0;
-		cpumask_set_cpu(0, &channel->alloced_cpus_in_node);
 		channel->target_cpu = 0;
 		channel->target_vp = hv_context.vp_index[0];
 		return;
@@ -434,21 +434,22 @@ static void init_vp_index(struct vmbus_channel *channel, const uuid_le *type_gui
 		channel->numa_node = next_node;
 		primary = channel;
 	}
+	alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 
-	if (cpumask_weight(&primary->alloced_cpus_in_node) ==
+	if (cpumask_weight(alloced_mask) ==
 	    cpumask_weight(cpumask_of_node(primary->numa_node))) {
 		/*
 		 * We have cycled through all the CPUs in the node;
 		 * reset the alloced map.
 		 */
-		cpumask_clear(&primary->alloced_cpus_in_node);
+		cpumask_clear(alloced_mask);
 	}
 
-	cpumask_xor(&available_mask, &primary->alloced_cpus_in_node,
+	cpumask_xor(&available_mask, alloced_mask,
 		    cpumask_of_node(primary->numa_node));
 
 	cur_cpu = cpumask_next(-1, &available_mask);
-	cpumask_set_cpu(cur_cpu, &primary->alloced_cpus_in_node);
+	cpumask_set_cpu(cur_cpu, alloced_mask);
 
 	channel->target_cpu = cur_cpu;
 	channel->target_vp = hv_context.vp_index[cur_cpu];

drivers/hv/hv.c

@@ -332,6 +332,13 @@ int hv_synic_alloc(void)
 	size_t ced_size = sizeof(struct clock_event_device);
 	int cpu;
 
+	hv_context.hv_numa_map = kzalloc(sizeof(struct cpumask) * nr_node_ids,
+					 GFP_ATOMIC);
+	if (hv_context.hv_numa_map == NULL) {
+		pr_err("Unable to allocate NUMA map\n");
+		goto err;
+	}
+
 	for_each_online_cpu(cpu) {
 		hv_context.event_dpc[cpu] = kmalloc(size, GFP_ATOMIC);
 		if (hv_context.event_dpc[cpu] == NULL) {
@@ -345,6 +352,7 @@ int hv_synic_alloc(void)
 			pr_err("Unable to allocate clock event device\n");
 			goto err;
 		}
+
 		hv_init_clockevent_device(hv_context.clk_evt[cpu], cpu);
 
 		hv_context.synic_message_page[cpu] =
@@ -393,6 +401,7 @@ void hv_synic_free(void)
 {
 	int cpu;
 
+	kfree(hv_context.hv_numa_map);
 	for_each_online_cpu(cpu)
 		hv_synic_free_cpu(cpu);
 }

drivers/hv/hyperv_vmbus.h

@@ -551,6 +551,11 @@ struct hv_context {
 	 * Support PV clockevent device.
 	 */
 	struct clock_event_device *clk_evt[NR_CPUS];
+	/*
+	 * To manage allocations in a NUMA node.
+	 * Array indexed by numa node ID.
+	 */
+	struct cpumask *hv_numa_map;
 };
 
 extern struct hv_context hv_context;

include/linux/hyperv.h

@@ -699,7 +699,6 @@ struct vmbus_channel {
 	/*
 	 * State to manage the CPU affiliation of channels.
 	 */
-	struct cpumask alloced_cpus_in_node;
 	int numa_node;
 	/*
 	 * Support for sub-channels. For high performance devices,