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@xxxxxxxxxxxxx>
---
drivers/hv/channel_mgmt.c | 11 ++++++-----
drivers/hv/hv.c | 9 +++++++++
drivers/hv/hyperv_vmbus.h | 5 +++++
include/linux/hyperv.h | 1 -
4 files changed, 20 insertions(+), 6 deletions(-)
diff --git a/drivers/hv/channel_mgmt.c b/drivers/hv/channel_mgmt.c
index 4506a66..f52373a 100644
--- a/drivers/hv/channel_mgmt.c
+++ b/drivers/hv/channel_mgmt.c
@@ -391,6 +391,7 @@ 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,
@@ -408,7 +409,6 @@ 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;
@@ -433,21 +433,22 @@ 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];
diff --git a/drivers/hv/hv.c b/drivers/hv/hv.c
index d3943bc..85498a7 100644
--- a/drivers/hv/hv.c
+++ b/drivers/hv/hv.c
@@ -329,6 +329,13 @@ 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) {
@@ -342,6 +349,7 @@ 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] =
@@ -390,6 +398,7 @@ void hv_synic_free(void)
{
int cpu;
+ kfree(hv_context.hv_numa_map);
for_each_online_cpu(cpu)
hv_synic_free_cpu(cpu);
}
diff --git a/drivers/hv/hyperv_vmbus.h b/drivers/hv/hyperv_vmbus.h
index cddc0c9..7bdef66 100644
--- a/drivers/hv/hyperv_vmbus.h
+++ b/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;
diff --git a/include/linux/hyperv.h b/include/linux/hyperv.h
index 30d3a1f..34cf811 100644
--- a/include/linux/hyperv.h
+++ b/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,
--
1.7.4.1
_______________________________________________
devel mailing list
devel@xxxxxxxxxxxxxxxxxxxxxx
http://driverdev.linuxdriverproject.org/mailman/listinfo/driverdev-devel
