On Mon, Jul 04, 2016 at 10:38:49AM +0200, Christoph Hellwig wrote:

On Mon, Jul 04, 2016 at 10:15:41AM +0200, Alexander Gordeev wrote:

quoted

On Tue, Jun 14, 2016 at 09:59:04PM +0200, Christoph Hellwig wrote:

quoted

+static int blk_mq_create_mq_map(struct blk_mq_tag_set *set,
+		const struct cpumask *affinity_mask)
+{
+	int queue = -1, cpu = 0;
+
+	set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
+			GFP_KERNEL, set->numa_node);
+	if (!set->mq_map)
+		return -ENOMEM;
+
+	if (!affinity_mask)
+		return 0;	/* map all cpus to queue 0 */
+
+	/* If cpus are offline, map them to first hctx */
+	for_each_online_cpu(cpu) {
+		if (cpumask_test_cpu(cpu, affinity_mask))
+			queue++;

CPUs missing in an affinity mask are mapped to hctxs. Is that intended?

Yes - each CPU needs to be mapped to some hctx, otherwise we can't
submit I/O from that CPU.

quoted

quoted

+		if (queue > 0)

Why this check?

quoted

+			set->mq_map[cpu] = queue;

mq_map is initialized to zero already, so we don't really need the
assignment for queue 0.  The reason why this check exists is because
we start with queue = -1 and we never want to assignment -1 to mq_map.

Would this read better then?

	int queue = 0;

	...

	/* If cpus are offline, map them to first hctx */
	for_each_online_cpu(cpu) {
		set->mq_map[cpu] = queue;
		if (cpumask_test_cpu(cpu, affinity_mask))
			queue++;
	}