@@ -53,6 +53,8 @@
#include "cvmx-gmxx-defs.h"
+#define CVM_OCT_SKB_CB(skb) ((u64 *)((skb)->cb))
+
/*
* You can define GET_SKBUFF_QOS() to override how the skbuff output
* function determines which output queue is used. The default
@@ -121,6 +123,7 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
uint64_t old_scratch;
uint64_t old_scratch2;
int qos;
+ int i;
enum {QUEUE_CORE, QUEUE_HW, QUEUE_DROP} queue_type;
struct octeon_ethernet *priv = netdev_priv(dev);
struct sk_buff *to_free_list;
@@ -171,6 +174,28 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
}
/*
+ * We have space for 6 segment pointers, If there will be more
+ * than that, we must linearize.
+ */
+ if (unlikely(skb_shinfo(skb)->nr_frags > 5)) {
+ if (unlikely(__skb_linearize(skb))) {
+ queue_type = QUEUE_DROP;
+ if (USE_ASYNC_IOBDMA) {
+ /* Get the number of skbuffs in use by the hardware */
+ CVMX_SYNCIOBDMA;
+ skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
+ } else {
+ /* Get the number of skbuffs in use by the hardware */
+ skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
+ MAX_SKB_TO_FREE);
+ }
+ skb_to_free = cvm_oct_adjust_skb_to_free(skb_to_free, priv->fau + qos * 4);
+ spin_lock_irqsave(&priv->tx_free_list[qos].lock, flags);
+ goto skip_xmit;
+ }
+ }
+
+ /*
* The CN3XXX series of parts has an errata (GMX-401) which
* causes the GMX block to hang if a collision occurs towards
* the end of a <68 byte packet. As a workaround for this, we
@@ -198,13 +223,6 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
}
}
- /* Build the PKO buffer pointer */
- hw_buffer.u64 = 0;
- hw_buffer.s.addr = cvmx_ptr_to_phys(skb->data);
- hw_buffer.s.pool = 0;
- hw_buffer.s.size =
- (unsigned long)skb_end_pointer(skb) - (unsigned long)skb->head;
-
/* Build the PKO command */
pko_command.u64 = 0;
pko_command.s.n2 = 1; /* Don't pollute L2 with the outgoing packet */
@@ -215,6 +233,31 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
pko_command.s.dontfree = 1;
pko_command.s.reg0 = priv->fau + qos * 4;
+
+ /* Build the PKO buffer pointer */
+ hw_buffer.u64 = 0;
+ if (skb_shinfo(skb)->nr_frags == 0) {
+ hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
+ hw_buffer.s.pool = 0;
+ hw_buffer.s.size = skb->len;
+ } else {
+ hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)skb->data);
+ hw_buffer.s.pool = 0;
+ hw_buffer.s.size = skb_headlen(skb);
+ CVM_OCT_SKB_CB(skb)[0] = hw_buffer.u64;
+ for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
+ struct skb_frag_struct *fs = skb_shinfo(skb)->frags + i;
+ hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)(page_address(fs->page) + fs->page_offset));
+ hw_buffer.s.size = fs->size;
+ CVM_OCT_SKB_CB(skb)[i + 1] = hw_buffer.u64;
+ }
+ hw_buffer.s.addr = XKPHYS_TO_PHYS((u64)CVM_OCT_SKB_CB(skb));
+ hw_buffer.s.size = skb_shinfo(skb)->nr_frags + 1;
+ pko_command.s.segs = skb_shinfo(skb)->nr_frags + 1;
+ pko_command.s.gather = 1;
+ goto dont_put_skbuff_in_hw;
+ }
+
/*
* See if we can put this skb in the FPA pool. Any strange
* behavior from the Linux networking stack will most likely
@@ -484,8 +484,11 @@ int cvm_oct_common_init(struct net_device *dev)
&& (always_use_pow || strstr(pow_send_list, dev->name)))
priv->queue = -1;
- if (priv->queue != -1 && USE_HW_TCPUDP_CHECKSUM)
- dev->features |= NETIF_F_IP_CSUM;
+ if (priv->queue != -1) {
+ dev->features |= NETIF_F_SG;
+ if (USE_HW_TCPUDP_CHECKSUM)
+ dev->features |= NETIF_F_IP_CSUM;
+ }
/* We do our own locking, Linux doesn't need to */
dev->features |= NETIF_F_LLTX;
Octeon ethernet hardware can handle NETIF_F_SG, so we enable it. A gather list of up to six fragments will fit in the SKB's CB structure, so no extra memory is required. If a SKB has more than six fragments, we must linearize it. Signed-off-by: David Daney <ddaney@caviumnetworks.com> --- drivers/staging/octeon/ethernet-tx.c | 57 +++++++++++++++++++++++++++++---- drivers/staging/octeon/ethernet.c | 7 +++- 2 files changed, 55 insertions(+), 9 deletions(-)