bnx2x: avoid two atomic ops per page on x86

Commit 4cace675 ("bnx2x: Alloc 4k fragment for each rx ring buffer
element") added extra put_page() and get_page() calls on arches where
PAGE_SIZE=4K like x86

Reorder things to avoid this overhead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com>
Cc: Yuval Mintz <Yuval.Mintz@cavium.com>
Cc: Ariel Elior <ariel.elior@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

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

@@ -549,14 +549,7 @@ static int bnx2x_alloc_rx_sge(struct bnx2x *bp, struct bnx2x_fastpath *fp,
 	struct bnx2x_alloc_pool *pool = &fp->page_pool;
 	dma_addr_t mapping;
 
-	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);
-
+	if (!pool->page) {
 		pool->page = alloc_pages(gfp_mask, PAGES_PER_SGE_SHIFT);
 		if (unlikely(!pool->page))
 			return -ENOMEM;
@@ -571,7 +564,6 @@ static int bnx2x_alloc_rx_sge(struct bnx2x *bp, struct bnx2x_fastpath *fp,
 		return -ENOMEM;
 	}
 
-	get_page(pool->page);
 	sw_buf->page = pool->page;
 	sw_buf->offset = pool->offset;
 
@@ -581,7 +573,10 @@ static int bnx2x_alloc_rx_sge(struct bnx2x *bp, struct bnx2x_fastpath *fp,
 	sge->addr_lo = cpu_to_le32(U64_LO(mapping));
 
 	pool->offset += SGE_PAGE_SIZE;
-
+	if (PAGE_SIZE - pool->offset >= SGE_PAGE_SIZE)
+		get_page(pool->page);
+	else
+		pool->page = NULL;
 	return 0;
 }