i40e/i40evf: use pages correctly in Rx

Refactor the packet split Rx code to properly use half-pages for
receives. The previous code was doing way more mapping and unmapping
than it needed to, and wasn't properly using half-pages.

Increment the page use count each time we give a half-page to an skb,
knowing that the stack will probably process and release the page before
we need it again. Only free and reallocate pages if the count shows that
both half-pages are in use. Add counters to track reallocations and page
reuse.

Change-ID: I534b299196036b64be82b4861a0a4036310a8f22
Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

drivers/net/ethernet/intel/i40e/i40e_debugfs.c

@@ -535,6 +535,11 @@ static void i40e_dbg_dump_vsi_seid(struct i40e_pf *pf, int seid)
 			 i,
 			 rx_ring->rx_stats.alloc_page_failed,
 			 rx_ring->rx_stats.alloc_buff_failed);
+		dev_info(&pf->pdev->dev,
+			 "    rx_rings[%i]: rx_stats: realloc_count = %lld, page_reuse_count = %lld\n",
+			 i,
+			 rx_ring->rx_stats.realloc_count,
+			 rx_ring->rx_stats.page_reuse_count);
 		dev_info(&pf->pdev->dev,
 			 "    rx_rings[%i]: size = %i, dma = 0x%08lx\n",
 			 i, rx_ring->size,

drivers/net/ethernet/intel/i40e/i40e_txrx.c

@@ -1060,7 +1060,7 @@ void i40e_clean_rx_ring(struct i40e_ring *rx_ring)
 			if (rx_bi->page_dma) {
 				dma_unmap_page(dev,
 					       rx_bi->page_dma,
-					       PAGE_SIZE / 2,
+					       PAGE_SIZE,
 					       DMA_FROM_DEVICE);
 				rx_bi->page_dma = 0;
 			}
@@ -1203,6 +1203,7 @@ bool i40e_alloc_rx_buffers_ps(struct i40e_ring *rx_ring, u16 cleaned_count)
 	u16 i = rx_ring->next_to_use;
 	union i40e_rx_desc *rx_desc;
 	struct i40e_rx_buffer *bi;
+	const int current_node = numa_node_id();
 
 	/* do nothing if no valid netdev defined */
 	if (!rx_ring->netdev || !cleaned_count)
@@ -1214,39 +1215,50 @@ bool i40e_alloc_rx_buffers_ps(struct i40e_ring *rx_ring, u16 cleaned_count)
 
 		if (bi->skb) /* desc is in use */
 			goto no_buffers;
+
+	/* If we've been moved to a different NUMA node, release the
+	 * page so we can get a new one on the current node.
+	 */
+		if (bi->page &&  page_to_nid(bi->page) != current_node) {
+			dma_unmap_page(rx_ring->dev,
+				       bi->page_dma,
+				       PAGE_SIZE,
+				       DMA_FROM_DEVICE);
+			__free_page(bi->page);
+			bi->page = NULL;
+			bi->page_dma = 0;
+			rx_ring->rx_stats.realloc_count++;
+		} else if (bi->page) {
+			rx_ring->rx_stats.page_reuse_count++;
+		}
+
 		if (!bi->page) {
 			bi->page = alloc_page(GFP_ATOMIC);
 			if (!bi->page) {
 				rx_ring->rx_stats.alloc_page_failed++;
 				goto no_buffers;
 			}
-		}
-
-		if (!bi->page_dma) {
-			/* use a half page if we're re-using */
-			bi->page_offset ^= PAGE_SIZE / 2;
 			bi->page_dma = dma_map_page(rx_ring->dev,
 						    bi->page,
-						    bi->page_offset,
-						    PAGE_SIZE / 2,
+						    0,
+						    PAGE_SIZE,
 						    DMA_FROM_DEVICE);
-			if (dma_mapping_error(rx_ring->dev,
-					      bi->page_dma)) {
+			if (dma_mapping_error(rx_ring->dev, bi->page_dma)) {
 				rx_ring->rx_stats.alloc_page_failed++;
+				__free_page(bi->page);
+				bi->page = NULL;
 				bi->page_dma = 0;
+				bi->page_offset = 0;
 				goto no_buffers;
 			}
+			bi->page_offset = 0;
 		}
 
-		dma_sync_single_range_for_device(rx_ring->dev,
-						 rx_ring->rx_bi[0].dma,
-						 i * rx_ring->rx_hdr_len,
-						 rx_ring->rx_hdr_len,
-						 DMA_FROM_DEVICE);
 		/* Refresh the desc even if buffer_addrs didn't change
 		 * because each write-back erases this info.
 		 */
-		rx_desc->read.pkt_addr = cpu_to_le64(bi->page_dma);
+		rx_desc->read.pkt_addr =
+				cpu_to_le64(bi->page_dma + bi->page_offset);
 		rx_desc->read.hdr_addr = cpu_to_le64(bi->dma);
 		i++;
 		if (i == rx_ring->count)
@@ -1527,14 +1539,14 @@ static int i40e_clean_rx_irq_ps(struct i40e_ring *rx_ring, const int budget)
 	unsigned int total_rx_bytes = 0, total_rx_packets = 0;
 	u16 rx_packet_len, rx_header_len, rx_sph, rx_hbo;
 	u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
-	const int current_node = numa_mem_id();
 	struct i40e_vsi *vsi = rx_ring->vsi;
 	u16 i = rx_ring->next_to_clean;
 	union i40e_rx_desc *rx_desc;
 	u32 rx_error, rx_status;
 	bool failure = false;
 	u8 rx_ptype;
 	u64 qword;
+	u32 copysize;
 
 	if (budget <= 0)
 		return 0;
@@ -1565,6 +1577,12 @@ static int i40e_clean_rx_irq_ps(struct i40e_ring *rx_ring, const int budget)
 		 * DD bit is set.
 		 */
 		dma_rmb();
+		/* sync header buffer for reading */
+		dma_sync_single_range_for_cpu(rx_ring->dev,
+					      rx_ring->rx_bi[0].dma,
+					      i * rx_ring->rx_hdr_len,
+					      rx_ring->rx_hdr_len,
+					      DMA_FROM_DEVICE);
 		if (i40e_rx_is_programming_status(qword)) {
 			i40e_clean_programming_status(rx_ring, rx_desc);
 			I40E_RX_INCREMENT(rx_ring, i);
@@ -1606,9 +1624,16 @@ static int i40e_clean_rx_irq_ps(struct i40e_ring *rx_ring, const int budget)
 
 		rx_ptype = (qword & I40E_RXD_QW1_PTYPE_MASK) >>
 			   I40E_RXD_QW1_PTYPE_SHIFT;
-		prefetch(rx_bi->page);
+		/* sync half-page for reading */
+		dma_sync_single_range_for_cpu(rx_ring->dev,
+					      rx_bi->page_dma,
+					      rx_bi->page_offset,
+					      PAGE_SIZE / 2,
+					      DMA_FROM_DEVICE);
+		prefetch(page_address(rx_bi->page) + rx_bi->page_offset);
 		rx_bi->skb = NULL;
 		cleaned_count++;
+		copysize = 0;
 		if (rx_hbo || rx_sph) {
 			int len;
 
@@ -1619,38 +1644,45 @@ static int i40e_clean_rx_irq_ps(struct i40e_ring *rx_ring, const int budget)
 			memcpy(__skb_put(skb, len), rx_bi->hdr_buf, len);
 		} else if (skb->len == 0) {
 			int len;
+			unsigned char *va = page_address(rx_bi->page) +
+					    rx_bi->page_offset;
 
-			len = (rx_packet_len > skb_headlen(skb) ?
-				skb_headlen(skb) : rx_packet_len);
-			memcpy(__skb_put(skb, len),
-			       rx_bi->page + rx_bi->page_offset,
-			       len);
-			rx_bi->page_offset += len;
+			len = min(rx_packet_len, rx_ring->rx_hdr_len);
+			memcpy(__skb_put(skb, len), va, len);
+			copysize = len;
 			rx_packet_len -= len;
 		}
-
 		/* Get the rest of the data if this was a header split */
 		if (rx_packet_len) {
-			skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
-					   rx_bi->page,
-					   rx_bi->page_offset,
-					   rx_packet_len);
-
-			skb->len += rx_packet_len;
-			skb->data_len += rx_packet_len;
-			skb->truesize += rx_packet_len;
-
-			if ((page_count(rx_bi->page) == 1) &&
-			    (page_to_nid(rx_bi->page) == current_node))
-				get_page(rx_bi->page);
-			else
+			skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
+					rx_bi->page,
+					rx_bi->page_offset + copysize,
+					rx_packet_len, I40E_RXBUFFER_2048);
+
+			get_page(rx_bi->page);
+			/* switch to the other half-page here; the allocation
+			 * code programs the right addr into HW. If we haven't
+			 * used this half-page, the address won't be changed,
+			 * and HW can just use it next time through.
+			 */
+			rx_bi->page_offset ^= PAGE_SIZE / 2;
+			/* If the page count is more than 2, then both halves
+			 * of the page are used and we need to free it. Do it
+			 * here instead of in the alloc code. Otherwise one
+			 * of the half-pages might be released between now and
+			 * then, and we wouldn't know which one to use.
+			 */
+			if (page_count(rx_bi->page) > 2) {
+				dma_unmap_page(rx_ring->dev,
+					       rx_bi->page_dma,
+					       PAGE_SIZE,
+					       DMA_FROM_DEVICE);
+				__free_page(rx_bi->page);
 				rx_bi->page = NULL;
+				rx_bi->page_dma = 0;
+				rx_ring->rx_stats.realloc_count++;
+			}
 
-			dma_unmap_page(rx_ring->dev,
-				       rx_bi->page_dma,
-				       PAGE_SIZE / 2,
-				       DMA_FROM_DEVICE);
-			rx_bi->page_dma = 0;
 		}
 		I40E_RX_INCREMENT(rx_ring, i);
 

drivers/net/ethernet/intel/i40e/i40e_txrx.h

@@ -209,6 +209,8 @@ struct i40e_rx_queue_stats {
 	u64 non_eop_descs;
 	u64 alloc_page_failed;
 	u64 alloc_buff_failed;
+	u64 page_reuse_count;
+	u64 realloc_count;
 };
 
 enum i40e_ring_state_t {