[PATCH 1/2] blk-mq: allocate blk_mq_tags and requests in correct node

[PATCH 1/2] blk-mq: allocate blk_mq_tags and requests in correct node

[Shaohua Li]

blk_mq_tags/requests of specific hardware queue are mostly used in
specific cpus, which might not be in the same numa node as disk. For
example, a nvme card is in node 0. half hardware queue will be used by
node 0, the other node 1.

Signed-off-by: Shaohua Li <shli@fb.com>
---
 block/blk-mq.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index c3400b5..90a37ba 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -1594,9 +1594,13 @@ static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
 	struct blk_mq_tags *tags;
 	unsigned int i, j, entries_per_page, max_order = 4;
 	size_t rq_size, left;
+	int node;
+
+	node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
+	if (node == NUMA_NO_NODE)
+		node = set->numa_node;
 
-	tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
-				set->numa_node,
+	tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, node,
 				BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
 	if (!tags)
 		return NULL;
@@ -1605,7 +1609,7 @@ static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
 
 	tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
 				 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
-				 set->numa_node);
+				 node);
 	if (!tags->rqs) {
 		blk_mq_free_tags(tags);
 		return NULL;
@@ -1629,7 +1633,7 @@ static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
 			this_order--;
 
 		do {
-			page = alloc_pages_node(set->numa_node,
+			page = alloc_pages_node(node,
 				GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
 				this_order);
 			if (page)
@@ -1660,7 +1664,7 @@ static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
 			if (set->ops->init_request) {
 				if (set->ops->init_request(set->driver_data,
 						tags->rqs[i], hctx_idx, i,
-						set->numa_node)) {
+						node)) {
 					tags->rqs[i] = NULL;
 					goto fail;
 				}
-- 
2.9.3

[PATCH 2/2] nvme: allocate nvme_queue in correct node

[Shaohua Li]

nvme_queue is per-cpu queue (mostly). Allocating it in node where blk-mq
will use it.

Signed-off-by: Shaohua Li <shli@fb.com>
---
 drivers/nvme/host/pci.c | 19 +++++++++++++++----
 1 file changed, 15 insertions(+), 4 deletions(-)

diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
index 3faefab..f81c0ed 100644
--- a/drivers/nvme/host/pci.c
+++ b/drivers/nvme/host/pci.c
@@ -1041,9 +1041,10 @@ static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
 }
 
 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
-							int depth)
+							int depth, int node)
 {
-	struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL);
+	struct nvme_queue *nvmeq = kzalloc_node(sizeof(*nvmeq), GFP_KERNEL,
+							node);
 	if (!nvmeq)
 		return NULL;
 
@@ -1219,7 +1220,8 @@ static int nvme_configure_admin_queue(struct nvme_dev *dev)
 
 	nvmeq = dev->queues[0];
 	if (!nvmeq) {
-		nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
+		nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH,
+						dev_to_node(dev->dev));
 		if (!nvmeq)
 			return -ENOMEM;
 	}
@@ -1309,9 +1311,18 @@ static int nvme_create_io_queues(struct nvme_dev *dev)
 {
 	unsigned i, max;
 	int ret = 0;
+	const struct cpumask *mask;
 
 	for (i = dev->queue_count; i <= dev->max_qid; i++) {
-		if (!nvme_alloc_queue(dev, i, dev->q_depth)) {
+		int node = dev_to_node(dev->dev);
+
+		mask = pci_irq_get_affinity(to_pci_dev(dev->dev), i);
+		if (mask) {
+			node = cpu_to_node(cpumask_first(mask));
+			node = local_memory_node(node);
+		}
+
+		if (!nvme_alloc_queue(dev, i, dev->q_depth, node)) {
 			ret = -ENOMEM;
 			break;
 		}
-- 
2.9.3

Re: [PATCH 1/2] blk-mq: allocate blk_mq_tags and requests in correct node

[Christoph Hellwig]

Hi Shaohua,

the code your patching has changed a lot in Jens' tree, so I think
you'll have to respin it.  But the idea looks fine.

Re: [PATCH 2/2] nvme: allocate nvme_queue in correct node

[Christoph Hellwig]

> +		mask = pci_irq_get_affinity(to_pci_dev(dev->dev), i);
> +		if (mask) {
> +			node = cpu_to_node(cpumask_first(mask));
> +			node = local_memory_node(node);
> +		}

Can you move this to a PCI-layer helper, e.g. something like:

int pci_irq_get_node(struct pci_dev *dev, unsigned vec)
{
	const struct cpumask *mask = pci_irq_get_affinity(dev), i);
	if (mask)
		return local_memory_node(cpu_to_node(cpumask_first(mask)));
	return dev_to_node(&dev->dev);
}

Otherwise this looks fine.