dpaa_eth: Remove dpa_bp_count from percpu_priv structure

Instead just rely on the percpu field in dpa_bp structure.
The performance lost due to the extra indirection is compensated
by the smaller size of the percpu_priv structure.

This way we begin to eliminate the assumption that private interfaces
only have one default buffer pool, which may not remain true for long.

Signed-off-by: Ioana Radulescu <ruxandra.radulescu@freescale.com>
Change-Id: I0b2c199c3f8883fa7d5fa1a06d2940f8170fc959
Reviewed-on: http://git.am.freescale.net:8181/3655
Reviewed-by: Bucur Madalin-Cristian-B32716 <madalin.bucur@freescale.com>
Reviewed-by: Hamciuc Bogdan-BHAMCIU1 <bogdan.hamciuc@freescale.com>
Reviewed-by: Fleming Andrew-AFLEMING <AFLEMING@freescale.com>
Tested-by: Fleming Andrew-AFLEMING <AFLEMING@freescale.com>

6 files changed, 25 insertions, 29 deletions

diff --git a/drivers/net/ethernet/freescale/dpa/dpaa_debugfs.c b/drivers/net/ethernet/freescale/dpa/dpaa_debugfs.c
index dee4384..f84b19e 100644
--- a/drivers/net/ethernet/freescale/dpa/dpaa_debugfs.c
+++ b/drivers/net/ethernet/freescale/dpa/dpaa_debugfs.c
@@ -78,8 +78,8 @@ static int dpa_debugfs_show(struct seq_file *file, void *offset)
 	for_each_online_cpu(i) {
 		percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
 
-		if (percpu_priv->dpa_bp_count)
-			dpa_bp_count = *percpu_priv->dpa_bp_count;
+		if (dpa_bp->percpu_count)
+			dpa_bp_count = *(per_cpu_ptr(dpa_bp->percpu_count, i));
 
 		total.in_interrupt += percpu_priv->in_interrupt;
 		total.stats.rx_packets += percpu_priv->stats.rx_packets;
diff --git a/drivers/net/ethernet/freescale/dpa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpa/dpaa_eth.c
index 6ac088c..d175cdc 100644
--- a/drivers/net/ethernet/freescale/dpa/dpaa_eth.c
+++ b/drivers/net/ethernet/freescale/dpa/dpaa_eth.c
@@ -529,12 +529,8 @@ static int __cold dpa_eth_priv_start(struct net_device *net_dev)
 	}
 	for_each_online_cpu(i) {
 		percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
-		if (!percpu_priv->dpa_bp) {
+		if (!percpu_priv->dpa_bp)
 			percpu_priv->dpa_bp = priv->dpa_bp;
-			percpu_priv->dpa_bp_count =
-				per_cpu_ptr(priv->dpa_bp->percpu_count,
-					    i);
-		}
 	}
 
 	dpaa_eth_napi_enable(priv);
diff --git a/drivers/net/ethernet/freescale/dpa/dpaa_eth.h b/drivers/net/ethernet/freescale/dpa/dpaa_eth.h
index 900bf11..d19991b 100644
--- a/drivers/net/ethernet/freescale/dpa/dpaa_eth.h
+++ b/drivers/net/ethernet/freescale/dpa/dpaa_eth.h
@@ -390,12 +390,6 @@ struct dpa_ern_cnt {
 
 struct dpa_percpu_priv_s {
 	struct net_device *net_dev;
-	/*
-	 * Pointer to the percpu_count of the shared buffer pool
-	 * used for the private ports; this assumes there is only
-	 * one bpool used
-	 */
-	int *dpa_bp_count;
 	struct dpa_bp *dpa_bp;
 	struct napi_struct napi;
 	u64 in_interrupt;
diff --git a/drivers/net/ethernet/freescale/dpa/dpaa_eth_non_sg.c b/drivers/net/ethernet/freescale/dpa/dpaa_eth_non_sg.c
index b7bfc3e..f2c8ced 100644
--- a/drivers/net/ethernet/freescale/dpa/dpaa_eth_non_sg.c
+++ b/drivers/net/ethernet/freescale/dpa/dpaa_eth_non_sg.c
@@ -140,9 +140,9 @@ void dpa_make_private_pool(struct dpa_bp *dpa_bp)
  */
 int dpaa_eth_refill_bpools(struct dpa_percpu_priv_s *percpu_priv)
 {
-	int *countptr = percpu_priv->dpa_bp_count;
-	int count = *countptr;
 	const struct dpa_bp *dpa_bp = percpu_priv->dpa_bp;
+	int *countptr = __this_cpu_ptr(dpa_bp->percpu_count);
+	int count = *countptr;
 	/* this function is called in softirq context;
 	 * no need to protect smp_processor_id() on RT kernel
 	 */
@@ -256,8 +256,9 @@ static int dpa_process_one(struct dpa_percpu_priv_s *percpu_priv,
 	unsigned long skb_addr = virt_to_phys(skb->head);
 	u32 pad = fd_addr - skb_addr;
 	unsigned int data_start;
+	int *countptr = __this_cpu_ptr(bp->percpu_count);
 
-	(*percpu_priv->dpa_bp_count)--;
+	(*countptr)--;
 
 	/* The skb is currently pointed at head + headroom. The packet
 	 * starts at skb->head + pad + fd offset.
@@ -468,6 +469,7 @@ static int skb_to_contig_fd(struct dpa_priv_s *priv,
 	bool can_recycle = false;
 	int offset, extra_offset;
 	int err;
+	int *countptr = __this_cpu_ptr(dpa_bp->percpu_count);
 
 	/* We are guaranteed that we have at least tx_headroom bytes.
 	 * Buffers we allocated are padded to improve cache usage. In order
@@ -490,7 +492,7 @@ static int skb_to_contig_fd(struct dpa_priv_s *priv,
 	if (likely(skb_is_recycleable(skb, dpa_bp->size) &&
 		   (skb_end_pointer(skb) - skb->head <=
 			DPA_RECYCLE_MAX_SIZE) &&
-		   (*percpu_priv->dpa_bp_count < dpa_bp->target_count))) {
+		   (*countptr < dpa_bp->target_count))) {
 		/* Compute the minimum necessary fd offset */
 		offset = dpa_bp->size - skb->len - skb_tailroom(skb);
 
@@ -585,6 +587,7 @@ int __hot dpa_tx(struct sk_buff *skb, struct net_device *net_dev)
 	struct rtnl_link_stats64 *percpu_stats;
 	int queue_mapping;
 	int err;
+	int *countptr;
 
 	/* If there is a Tx hook, run it. */
 	if (dpaa_eth_hooks.tx &&
@@ -595,6 +598,7 @@ int __hot dpa_tx(struct sk_buff *skb, struct net_device *net_dev)
 	priv = netdev_priv(net_dev);
 	percpu_priv = per_cpu_ptr(priv->percpu_priv, smp_processor_id());
 	percpu_stats = &percpu_priv->stats;
+	countptr = __this_cpu_ptr(priv->dpa_bp->percpu_count);
 
 	clear_fd(&fd);
 	queue_mapping = dpa_get_queue_mapping(skb);
@@ -665,7 +669,7 @@ int __hot dpa_tx(struct sk_buff *skb, struct net_device *net_dev)
 		 */
 		skb_recycle(skb);
 		skb = NULL;
-		(*percpu_priv->dpa_bp_count)++;
+		(*countptr)++;
 		percpu_priv->tx_returned++;
 	}
 
@@ -678,7 +682,7 @@ int __hot dpa_tx(struct sk_buff *skb, struct net_device *net_dev)
 
 xmit_failed:
 	if (fd.cmd & FM_FD_CMD_FCO) {
-		(*percpu_priv->dpa_bp_count)--;
+		(*countptr)--;
 		percpu_priv->tx_returned--;
 	}
 fd_create_failed:
diff --git a/drivers/net/ethernet/freescale/dpa/dpaa_eth_sg.c b/drivers/net/ethernet/freescale/dpa/dpaa_eth_sg.c
index bc90d29..cffc970 100644
--- a/drivers/net/ethernet/freescale/dpa/dpaa_eth_sg.c
+++ b/drivers/net/ethernet/freescale/dpa/dpaa_eth_sg.c
@@ -157,9 +157,9 @@ void dpa_make_private_pool(struct dpa_bp *dpa_bp)
  */
 int dpaa_eth_refill_bpools(struct dpa_percpu_priv_s *percpu_priv)
 {
-	int *countptr = percpu_priv->dpa_bp_count;
-	int count = *countptr;
 	const struct dpa_bp *dpa_bp = percpu_priv->dpa_bp;
+	int *countptr = __this_cpu_ptr(percpu_priv->dpa_bp->percpu_count);
+	int count = *countptr;
 	int new_bufs;
 
 	/* Add pages to the buffer pool */
@@ -807,11 +807,13 @@ int __hot dpa_tx(struct sk_buff *skb, struct net_device *net_dev)
 	int err = 0;
 	const int queue_mapping = dpa_get_queue_mapping(skb);
 	const bool nonlinear = skb_is_nonlinear(skb);
+	int *countptr;
 
 	priv = netdev_priv(net_dev);
 	/* Non-migratable context, safe to use __this_cpu_ptr */
 	percpu_priv = __this_cpu_ptr(priv->percpu_priv);
 	percpu_stats = &percpu_priv->stats;
+	countptr = __this_cpu_ptr(percpu_priv->dpa_bp->percpu_count);
 
 	clear_fd(&fd);
 
@@ -891,7 +893,7 @@ int __hot dpa_tx(struct sk_buff *skb, struct net_device *net_dev)
 		 * of the page containing the recycled buffer to make sure it
 		 * doesn't get freed.
 		 */
-		(*percpu_priv->dpa_bp_count)++;
+		(*countptr)++;
 		get_page(virt_to_head_page(skb->head));
 		percpu_priv->tx_returned++;
 	}
@@ -909,7 +911,7 @@ int __hot dpa_tx(struct sk_buff *skb, struct net_device *net_dev)
 
 xmit_failed:
 	if (fd.cmd & FM_FD_CMD_FCO) {
-		(*percpu_priv->dpa_bp_count)--;
+		(*countptr)--;
 		put_page(virt_to_head_page(skb->head));
 		percpu_priv->tx_returned--;
 	}
diff --git a/drivers/net/ethernet/freescale/dpa/dpaa_eth_unit_test.c b/drivers/net/ethernet/freescale/dpa/dpaa_eth_unit_test.c
index 46e8808..5df53a5 100644
--- a/drivers/net/ethernet/freescale/dpa/dpaa_eth_unit_test.c
+++ b/drivers/net/ethernet/freescale/dpa/dpaa_eth_unit_test.c
@@ -87,6 +87,7 @@ static enum qman_cb_dqrr_result tx_unit_test_dqrr(
 	dma_addr_t addr;
 	unsigned char *startaddr;
 	struct dpa_percpu_priv_s *percpu_priv;
+	int *countptr;
 
 	tx_unit_test_passed = false;
 
@@ -126,9 +127,10 @@ static enum qman_cb_dqrr_result tx_unit_test_dqrr(
 		 * If we didn't recycle, but the buffer was big enough,
 		 * increment the counter to put it back
 		 */
+		countptr = __this_cpu_ptr(priv->dpa_bp->percpu_count);
 		if (skb_end_pointer(skb) - skb->head >=
 			dpa_get_max_frm())
-			(*percpu_priv->dpa_bp_count)++;
+			(*countptr)++;
 
 		/* If we didn't recycle, the data pointer should be good */
 		if (skb->data != startaddr + dpa_fd_offset(fd))
@@ -234,6 +236,8 @@ static int dpa_tx_unit_test(struct net_device *net_dev)
 		     headroom += 16) {
 			int ret;
 			struct sk_buff *skb;
+			int *countptr =
+				__this_cpu_ptr(priv->dpa_bp->percpu_count);
 
 			test_count++;
 
@@ -247,7 +251,7 @@ static int dpa_tx_unit_test(struct net_device *net_dev)
 
 			if (skb_end_pointer(skb) - skb->head >=
 					dpa_get_max_frm())
-				(*percpu_priv->dpa_bp_count)--;
+				(*countptr)--;
 
 			skb_put(skb, size + headroom);
 			skb_pull(skb, headroom);
@@ -368,7 +372,6 @@ void dpa_unit_test_drain_default_pool(struct net_device *net_dev)
 	for_each_online_cpu(i) {
 		percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
 		percpu_priv->dpa_bp = NULL;
-		percpu_priv->dpa_bp_count = NULL;
 	}
 }
 
@@ -387,9 +390,6 @@ void dpa_unit_test_seed_default_pool(struct net_device *net_dev)
 		percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
 		if (!percpu_priv->dpa_bp) {
 			percpu_priv->dpa_bp = priv->dpa_bp;
-			percpu_priv->dpa_bp_count =
-				per_cpu_ptr(priv->dpa_bp->percpu_count,
-						i);
 		}
 	}
 }