amd-xgbe: Add NUMA affinity support for memory allocations

Add support to perform memory allocations on the node of the device. The
original allocation or the ring structure and Tx/Rx queues allocated all
of the memory at once and then carved it up for each channel and queue.
To best ensure that we get as much memory from the NUMA node as we can,
break the channel and ring allocations into individual allocations.

Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/amd/xgbe/xgbe-desc.c

@@ -176,43 +176,69 @@ static void xgbe_free_ring_resources(struct xgbe_prv_data *pdata)
 
 	DBGPR("-->xgbe_free_ring_resources\n");
 
-	channel = pdata->channel;
-	for (i = 0; i < pdata->channel_count; i++, channel++) {
+	for (i = 0; i < pdata->channel_count; i++) {
+		channel = pdata->channel[i];
 		xgbe_free_ring(pdata, channel->tx_ring);
 		xgbe_free_ring(pdata, channel->rx_ring);
 	}
 
 	DBGPR("<--xgbe_free_ring_resources\n");
 }
 
+static void *xgbe_alloc_node(size_t size, int node)
+{
+	void *mem;
+
+	mem = kzalloc_node(size, GFP_KERNEL, node);
+	if (!mem)
+		mem = kzalloc(size, GFP_KERNEL);
+
+	return mem;
+}
+
+static void *xgbe_dma_alloc_node(struct device *dev, size_t size,
+				 dma_addr_t *dma, int node)
+{
+	void *mem;
+	int cur_node = dev_to_node(dev);
+
+	set_dev_node(dev, node);
+	mem = dma_alloc_coherent(dev, size, dma, GFP_KERNEL);
+	set_dev_node(dev, cur_node);
+
+	if (!mem)
+		mem = dma_alloc_coherent(dev, size, dma, GFP_KERNEL);
+
+	return mem;
+}
+
 static int xgbe_init_ring(struct xgbe_prv_data *pdata,
 			  struct xgbe_ring *ring, unsigned int rdesc_count)
 {
-	DBGPR("-->xgbe_init_ring\n");
+	size_t size;
 
 	if (!ring)
 		return 0;
 
 	/* Descriptors */
+	size = rdesc_count * sizeof(struct xgbe_ring_desc);
+
 	ring->rdesc_count = rdesc_count;
-	ring->rdesc = dma_alloc_coherent(pdata->dev,
-					 (sizeof(struct xgbe_ring_desc) *
-					  rdesc_count), &ring->rdesc_dma,
-					 GFP_KERNEL);
+	ring->rdesc = xgbe_dma_alloc_node(pdata->dev, size, &ring->rdesc_dma,
+					  ring->node);
 	if (!ring->rdesc)
 		return -ENOMEM;
 
 	/* Descriptor information */
-	ring->rdata = kcalloc(rdesc_count, sizeof(struct xgbe_ring_data),
-			      GFP_KERNEL);
+	size = rdesc_count * sizeof(struct xgbe_ring_data);
+
+	ring->rdata = xgbe_alloc_node(size, ring->node);
 	if (!ring->rdata)
 		return -ENOMEM;
 
 	netif_dbg(pdata, drv, pdata->netdev,
-		  "rdesc=%p, rdesc_dma=%pad, rdata=%p\n",
-		  ring->rdesc, &ring->rdesc_dma, ring->rdata);
-
-	DBGPR("<--xgbe_init_ring\n");
+		  "rdesc=%p, rdesc_dma=%pad, rdata=%p, node=%d\n",
+		  ring->rdesc, &ring->rdesc_dma, ring->rdata, ring->node);
 
 	return 0;
 }
@@ -223,10 +249,8 @@ static int xgbe_alloc_ring_resources(struct xgbe_prv_data *pdata)
 	unsigned int i;
 	int ret;
 
-	DBGPR("-->xgbe_alloc_ring_resources\n");
-
-	channel = pdata->channel;
-	for (i = 0; i < pdata->channel_count; i++, channel++) {
+	for (i = 0; i < pdata->channel_count; i++) {
+		channel = pdata->channel[i];
 		netif_dbg(pdata, drv, pdata->netdev, "%s - Tx ring:\n",
 			  channel->name);
 
@@ -250,8 +274,6 @@ static int xgbe_alloc_ring_resources(struct xgbe_prv_data *pdata)
 		}
 	}
 
-	DBGPR("<--xgbe_alloc_ring_resources\n");
-
 	return 0;
 
 err_ring:
@@ -261,21 +283,33 @@ static int xgbe_alloc_ring_resources(struct xgbe_prv_data *pdata)
 }
 
 static int xgbe_alloc_pages(struct xgbe_prv_data *pdata,
-			    struct xgbe_page_alloc *pa, gfp_t gfp, int order)
+			    struct xgbe_page_alloc *pa, int alloc_order,
+			    int node)
 {
 	struct page *pages = NULL;
 	dma_addr_t pages_dma;
-	int ret;
+	gfp_t gfp;
+	int order, ret;
+
+again:
+	order = alloc_order;
 
 	/* Try to obtain pages, decreasing order if necessary */
-	gfp |= __GFP_COLD | __GFP_COMP | __GFP_NOWARN;
+	gfp = GFP_ATOMIC | __GFP_COLD | __GFP_COMP | __GFP_NOWARN;
 	while (order >= 0) {
-		pages = alloc_pages(gfp, order);
+		pages = alloc_pages_node(node, gfp, order);
 		if (pages)
 			break;
 
 		order--;
 	}
+
+	/* If we couldn't get local pages, try getting from anywhere */
+	if (!pages && (node != NUMA_NO_NODE)) {
+		node = NUMA_NO_NODE;
+		goto again;
+	}
+
 	if (!pages)
 		return -ENOMEM;
 
@@ -327,14 +361,14 @@ static int xgbe_map_rx_buffer(struct xgbe_prv_data *pdata,
 	int ret;
 
 	if (!ring->rx_hdr_pa.pages) {
-		ret = xgbe_alloc_pages(pdata, &ring->rx_hdr_pa, GFP_ATOMIC, 0);
+		ret = xgbe_alloc_pages(pdata, &ring->rx_hdr_pa, 0, ring->node);
 		if (ret)
 			return ret;
 	}
 
 	if (!ring->rx_buf_pa.pages) {
-		ret = xgbe_alloc_pages(pdata, &ring->rx_buf_pa, GFP_ATOMIC,
-				       PAGE_ALLOC_COSTLY_ORDER);
+		ret = xgbe_alloc_pages(pdata, &ring->rx_buf_pa,
+				       PAGE_ALLOC_COSTLY_ORDER, ring->node);
 		if (ret)
 			return ret;
 	}
@@ -362,8 +396,8 @@ static void xgbe_wrapper_tx_descriptor_init(struct xgbe_prv_data *pdata)
 
 	DBGPR("-->xgbe_wrapper_tx_descriptor_init\n");
 
-	channel = pdata->channel;
-	for (i = 0; i < pdata->channel_count; i++, channel++) {
+	for (i = 0; i < pdata->channel_count; i++) {
+		channel = pdata->channel[i];
 		ring = channel->tx_ring;
 		if (!ring)
 			break;
@@ -403,8 +437,8 @@ static void xgbe_wrapper_rx_descriptor_init(struct xgbe_prv_data *pdata)
 
 	DBGPR("-->xgbe_wrapper_rx_descriptor_init\n");
 
-	channel = pdata->channel;
-	for (i = 0; i < pdata->channel_count; i++, channel++) {
+	for (i = 0; i < pdata->channel_count; i++) {
+		channel = pdata->channel[i];
 		ring = channel->rx_ring;
 		if (!ring)
 			break;