bnx2x: Alloc 4k fragment for each rx ring buffer element

The driver allocates one page for each buffer on the rx ring, which is
too much on architectures like ppc64 and can cause unexpected allocation
failures when the system is under stress.  Now, we keep a memory pool
per queue, and if the architecture's PAGE_SIZE is greater than 4k, we
fragment pages and assign each 4k segment to a ring element, which
reduces the overall memory consumption on such architectures.  This
helps avoiding errors like the example below:

[bnx2x_alloc_rx_sge:435(eth1)]Can't alloc sge
[c00000037ffeb900] [d000000075eddeb4] .bnx2x_alloc_rx_sge+0x44/0x200 [bnx2x]
[c00000037ffeb9b0] [d000000075ee0b34] .bnx2x_fill_frag_skb+0x1ac/0x460 [bnx2x]
[c00000037ffebac0] [d000000075ee11f0] .bnx2x_tpa_stop+0x160/0x2e8 [bnx2x]
[c00000037ffebb90] [d000000075ee1560] .bnx2x_rx_int+0x1e8/0xc30 [bnx2x]
[c00000037ffebcd0] [d000000075ee2084] .bnx2x_poll+0xdc/0x3d8 [bnx2x] (unreliable)
Signed-off-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com>
Acked-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Reviewed-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/broadcom/bnx2x/bnx2x.h

@@ -357,6 +357,7 @@ struct sw_tx_bd {
 struct sw_rx_page {
 	struct page	*page;
 	DEFINE_DMA_UNMAP_ADDR(mapping);
+	unsigned int	offset;
 };
 
 union db_prod {
@@ -381,9 +382,10 @@ union db_prod {
 
 #define PAGES_PER_SGE_SHIFT	0
 #define PAGES_PER_SGE		(1 << PAGES_PER_SGE_SHIFT)
-#define SGE_PAGE_SIZE		PAGE_SIZE
-#define SGE_PAGE_SHIFT		PAGE_SHIFT
-#define SGE_PAGE_ALIGN(addr)	PAGE_ALIGN((typeof(PAGE_SIZE))(addr))
+#define SGE_PAGE_SHIFT		12
+#define SGE_PAGE_SIZE		(1 << SGE_PAGE_SHIFT)
+#define SGE_PAGE_MASK		(~(SGE_PAGE_SIZE - 1))
+#define SGE_PAGE_ALIGN(addr)	(((addr) + SGE_PAGE_SIZE - 1) & SGE_PAGE_MASK)
 #define SGE_PAGES		(SGE_PAGE_SIZE * PAGES_PER_SGE)
 #define TPA_AGG_SIZE		min_t(u32, (min_t(u32, 8, MAX_SKB_FRAGS) * \
 					    SGE_PAGES), 0xffff)
@@ -526,6 +528,12 @@ enum bnx2x_tpa_mode_t {
 	TPA_MODE_GRO
 };
 
+struct bnx2x_alloc_pool {
+	struct page	*page;
+	dma_addr_t	dma;
+	unsigned int	offset;
+};
+
 struct bnx2x_fastpath {
 	struct bnx2x		*bp; /* parent */
 
@@ -599,6 +607,8 @@ struct bnx2x_fastpath {
 	     4 (for the digits and to make it DWORD aligned) */
 #define FP_NAME_SIZE		(sizeof(((struct net_device *)0)->name) + 8)
 	char			name[FP_NAME_SIZE];
+
+	struct bnx2x_alloc_pool	page_pool;
 };
 
 #define bnx2x_fp(bp, nr, var)	((bp)->fp[(nr)].var)

drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.c

@@ -544,30 +544,49 @@ static void bnx2x_set_gro_params(struct sk_buff *skb, u16 parsing_flags,
 static int bnx2x_alloc_rx_sge(struct bnx2x *bp, struct bnx2x_fastpath *fp,
 			      u16 index, gfp_t gfp_mask)
 {
-	struct page *page = alloc_pages(gfp_mask, PAGES_PER_SGE_SHIFT);
 	struct sw_rx_page *sw_buf = &fp->rx_page_ring[index];
 	struct eth_rx_sge *sge = &fp->rx_sge_ring[index];
+	struct bnx2x_alloc_pool *pool = &fp->page_pool;
 	dma_addr_t mapping;
 
-	if (unlikely(page == NULL)) {
+	if (!pool->page || (PAGE_SIZE - pool->offset) < SGE_PAGE_SIZE) {
+
+		/* put page reference used by the memory pool, since we
+		 * won't be using this page as the mempool anymore.
+		 */
+		if (pool->page)
+			put_page(pool->page);
+
+		pool->page = alloc_pages(gfp_mask, PAGES_PER_SGE_SHIFT);
+		if (unlikely(!pool->page)) {
 			BNX2X_ERR("Can't alloc sge\n");
 			return -ENOMEM;
 		}
 
-	mapping = dma_map_page(&bp->pdev->dev, page, 0,
-			       SGE_PAGES, DMA_FROM_DEVICE);
-	if (unlikely(dma_mapping_error(&bp->pdev->dev, mapping))) {
-		__free_pages(page, PAGES_PER_SGE_SHIFT);
+		pool->dma = dma_map_page(&bp->pdev->dev, pool->page, 0,
+					 PAGE_SIZE, DMA_FROM_DEVICE);
+		if (unlikely(dma_mapping_error(&bp->pdev->dev,
+					       pool->dma))) {
+			__free_pages(pool->page, PAGES_PER_SGE_SHIFT);
+			pool->page = NULL;
 			BNX2X_ERR("Can't map sge\n");
 			return -ENOMEM;
 		}
+		pool->offset = 0;
+	}
+
+	get_page(pool->page);
+	sw_buf->page = pool->page;
+	sw_buf->offset = pool->offset;
 
-	sw_buf->page = page;
+	mapping = pool->dma + sw_buf->offset;
 	dma_unmap_addr_set(sw_buf, mapping, mapping);
 
 	sge->addr_hi = cpu_to_le32(U64_HI(mapping));
 	sge->addr_lo = cpu_to_le32(U64_LO(mapping));
 
+	pool->offset += SGE_PAGE_SIZE;
+
 	return 0;
 }
 
@@ -629,20 +648,22 @@ static int bnx2x_fill_frag_skb(struct bnx2x *bp, struct bnx2x_fastpath *fp,
 			return err;
 		}
 
-		/* Unmap the page as we're going to pass it to the stack */
-		dma_unmap_page(&bp->pdev->dev,
+		dma_unmap_single(&bp->pdev->dev,
 				 dma_unmap_addr(&old_rx_pg, mapping),
-			       SGE_PAGES, DMA_FROM_DEVICE);
+				 SGE_PAGE_SIZE, DMA_FROM_DEVICE);
 		/* Add one frag and update the appropriate fields in the skb */
 		if (fp->mode == TPA_MODE_LRO)
-			skb_fill_page_desc(skb, j, old_rx_pg.page, 0, frag_len);
+			skb_fill_page_desc(skb, j, old_rx_pg.page,
+					   old_rx_pg.offset, frag_len);
 		else { /* GRO */
 			int rem;
 			int offset = 0;
 			for (rem = frag_len; rem > 0; rem -= gro_size) {
 				int len = rem > gro_size ? gro_size : rem;
 				skb_fill_page_desc(skb, frag_id++,
-						   old_rx_pg.page, offset, len);
+						   old_rx_pg.page,
+						   old_rx_pg.offset + offset,
+						   len);
 				if (offset)
 					get_page(old_rx_pg.page);
 				offset += len;

drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.h

@@ -804,9 +804,13 @@ static inline void bnx2x_free_rx_sge(struct bnx2x *bp,
 	if (!page)
 		return;
 
-	dma_unmap_page(&bp->pdev->dev, dma_unmap_addr(sw_buf, mapping),
-		       SGE_PAGES, DMA_FROM_DEVICE);
-	__free_pages(page, PAGES_PER_SGE_SHIFT);
+	/* Since many fragments can share the same page, make sure to
+	 * only unmap and free the page once.
+	 */
+	dma_unmap_single(&bp->pdev->dev, dma_unmap_addr(sw_buf, mapping),
+			 SGE_PAGE_SIZE, DMA_FROM_DEVICE);
+
+	put_page(page);
 
 	sw_buf->page = NULL;
 	sge->addr_hi = 0;
@@ -964,6 +968,25 @@ static inline void bnx2x_set_fw_mac_addr(__le16 *fw_hi, __le16 *fw_mid,
 	((u8 *)fw_lo)[1]  = mac[4];
 }
 
+static inline void bnx2x_free_rx_mem_pool(struct bnx2x *bp,
+					  struct bnx2x_alloc_pool *pool)
+{
+	if (!pool->page)
+		return;
+
+	/* Page was not fully fragmented.  Unmap unused space */
+	if (pool->offset < PAGE_SIZE) {
+		dma_addr_t dma = pool->dma + pool->offset;
+		int size = PAGE_SIZE - pool->offset;
+
+		dma_unmap_single(&bp->pdev->dev, dma, size, DMA_FROM_DEVICE);
+	}
+
+	put_page(pool->page);
+
+	pool->page = NULL;
+}
+
 static inline void bnx2x_free_rx_sge_range(struct bnx2x *bp,
 					   struct bnx2x_fastpath *fp, int last)
 {
@@ -974,6 +997,8 @@ static inline void bnx2x_free_rx_sge_range(struct bnx2x *bp,
 
 	for (i = 0; i < last; i++)
 		bnx2x_free_rx_sge(bp, fp, i);
+
+	bnx2x_free_rx_mem_pool(bp, &fp->page_pool);
 }
 
 static inline void bnx2x_set_next_page_rx_bd(struct bnx2x_fastpath *fp)