netxen: cut-through rx changes
NX3031 supports cut-through operation where ingress packets are
directly dma'ed into host buffers to reduce latency.
This requires larger dma buffers (2kb) and different alignemnt.
The buffer posting logic is changed a bit. The free rx buffers
are maintained in linked list, since the received reference
handles can be out of order. However rx descriptors are still
posted sequentially, indexed by producer.
Signed-off-by: Dhananjay Phadke <dhananjay@netxen.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
diff --git a/drivers/net/netxen/netxen_nic_init.c b/drivers/net/netxen/netxen_nic_init.c
index 7b51240..01ab31b3 100644
--- a/drivers/net/netxen/netxen_nic_init.c
+++ b/drivers/net/netxen/netxen_nic_init.c
@@ -262,17 +262,30 @@
rds_ring->max_rx_desc_count =
adapter->max_rx_desc_count;
rds_ring->flags = RCV_DESC_NORMAL;
- rds_ring->dma_size = RX_DMA_MAP_LEN;
- rds_ring->skb_size = MAX_RX_BUFFER_LENGTH;
+ if (adapter->ahw.cut_through) {
+ rds_ring->dma_size =
+ NX_CT_DEFAULT_RX_BUF_LEN;
+ rds_ring->skb_size =
+ NX_CT_DEFAULT_RX_BUF_LEN;
+ } else {
+ rds_ring->dma_size = RX_DMA_MAP_LEN;
+ rds_ring->skb_size =
+ MAX_RX_BUFFER_LENGTH;
+ }
break;
case RCV_DESC_JUMBO:
rds_ring->max_rx_desc_count =
adapter->max_jumbo_rx_desc_count;
rds_ring->flags = RCV_DESC_JUMBO;
- rds_ring->dma_size = RX_JUMBO_DMA_MAP_LEN;
+ if (NX_IS_REVISION_P3(adapter->ahw.revision_id))
+ rds_ring->dma_size =
+ NX_P3_RX_JUMBO_BUF_MAX_LEN;
+ else
+ rds_ring->dma_size =
+ NX_P2_RX_JUMBO_BUF_MAX_LEN;
rds_ring->skb_size =
- MAX_RX_JUMBO_BUFFER_LENGTH;
+ rds_ring->dma_size + NET_IP_ALIGN;
break;
case RCV_RING_LRO:
@@ -294,6 +307,7 @@
goto err_out;
}
memset(rds_ring->rx_buf_arr, 0, RCV_BUFFSIZE);
+ INIT_LIST_HEAD(&rds_ring->free_list);
rds_ring->begin_alloc = 0;
/*
* Now go through all of them, set reference handles
@@ -302,6 +316,8 @@
num_rx_bufs = rds_ring->max_rx_desc_count;
rx_buf = rds_ring->rx_buf_arr;
for (i = 0; i < num_rx_bufs; i++) {
+ list_add_tail(&rx_buf->list,
+ &rds_ring->free_list);
rx_buf->ref_handle = i;
rx_buf->state = NETXEN_BUFFER_FREE;
rx_buf++;
@@ -1137,15 +1153,47 @@
return 0;
}
+static struct sk_buff *netxen_process_rxbuf(struct netxen_adapter *adapter,
+ struct nx_host_rds_ring *rds_ring, u16 index, u16 cksum)
+{
+ struct netxen_rx_buffer *buffer;
+ struct sk_buff *skb;
+
+ buffer = &rds_ring->rx_buf_arr[index];
+
+ pci_unmap_single(adapter->pdev, buffer->dma, rds_ring->dma_size,
+ PCI_DMA_FROMDEVICE);
+
+ skb = buffer->skb;
+ if (!skb)
+ goto no_skb;
+
+ if (likely(adapter->rx_csum && cksum == STATUS_CKSUM_OK)) {
+ adapter->stats.csummed++;
+ skb->ip_summed = CHECKSUM_UNNECESSARY;
+ } else
+ skb->ip_summed = CHECKSUM_NONE;
+
+ skb->dev = adapter->netdev;
+
+ buffer->skb = NULL;
+
+no_skb:
+ buffer->state = NETXEN_BUFFER_FREE;
+ buffer->lro_current_frags = 0;
+ buffer->lro_expected_frags = 0;
+ list_add_tail(&buffer->list, &rds_ring->free_list);
+ return skb;
+}
+
/*
* netxen_process_rcv() send the received packet to the protocol stack.
* and if the number of receives exceeds RX_BUFFERS_REFILL, then we
* invoke the routine to send more rx buffers to the Phantom...
*/
static void netxen_process_rcv(struct netxen_adapter *adapter, int ctxid,
- struct status_desc *desc)
+ struct status_desc *desc, struct status_desc *frag_desc)
{
- struct pci_dev *pdev = adapter->pdev;
struct net_device *netdev = adapter->netdev;
u64 sts_data = le64_to_cpu(desc->status_desc_data);
int index = netxen_get_sts_refhandle(sts_data);
@@ -1154,8 +1202,8 @@
struct sk_buff *skb;
u32 length = netxen_get_sts_totallength(sts_data);
u32 desc_ctx;
+ u16 pkt_offset = 0, cksum;
struct nx_host_rds_ring *rds_ring;
- int ret;
desc_ctx = netxen_get_sts_type(sts_data);
if (unlikely(desc_ctx >= NUM_RCV_DESC_RINGS)) {
@@ -1191,41 +1239,52 @@
}
}
- pci_unmap_single(pdev, buffer->dma, rds_ring->dma_size,
- PCI_DMA_FROMDEVICE);
+ cksum = netxen_get_sts_status(sts_data);
- skb = (struct sk_buff *)buffer->skb;
+ skb = netxen_process_rxbuf(adapter, rds_ring, index, cksum);
+ if (!skb)
+ return;
- if (likely(adapter->rx_csum &&
- netxen_get_sts_status(sts_data) == STATUS_CKSUM_OK)) {
- adapter->stats.csummed++;
- skb->ip_summed = CHECKSUM_UNNECESSARY;
- } else
- skb->ip_summed = CHECKSUM_NONE;
-
- skb->dev = netdev;
if (desc_ctx == RCV_DESC_LRO_CTXID) {
/* True length was only available on the last pkt */
skb_put(skb, buffer->lro_length);
} else {
- skb_put(skb, length);
+ if (length > rds_ring->skb_size)
+ skb_put(skb, rds_ring->skb_size);
+ else
+ skb_put(skb, length);
+
+ pkt_offset = netxen_get_sts_pkt_offset(sts_data);
+ if (pkt_offset)
+ skb_pull(skb, pkt_offset);
}
skb->protocol = eth_type_trans(skb, netdev);
- ret = netif_receive_skb(skb);
- netdev->last_rx = jiffies;
-
/*
- * We just consumed one buffer so post a buffer.
+ * rx buffer chaining is disabled, walk and free
+ * any spurious rx buffer chain.
*/
- buffer->skb = NULL;
- buffer->state = NETXEN_BUFFER_FREE;
- buffer->lro_current_frags = 0;
- buffer->lro_expected_frags = 0;
+ if (frag_desc) {
+ u16 i, nr_frags = desc->nr_frags;
- adapter->stats.no_rcv++;
- adapter->stats.rxbytes += length;
+ dev_kfree_skb_any(skb);
+ for (i = 0; i < nr_frags; i++) {
+ index = frag_desc->frag_handles[i];
+ skb = netxen_process_rxbuf(adapter,
+ rds_ring, index, cksum);
+ if (skb)
+ dev_kfree_skb_any(skb);
+ }
+ adapter->stats.rxdropped++;
+ } else {
+
+ netif_receive_skb(skb);
+ netdev->last_rx = jiffies;
+
+ adapter->stats.no_rcv++;
+ adapter->stats.rxbytes += length;
+ }
}
/* Process Receive status ring */
@@ -1233,9 +1292,11 @@
{
struct netxen_recv_context *recv_ctx = &(adapter->recv_ctx[ctxid]);
struct status_desc *desc_head = recv_ctx->rcv_status_desc_head;
- struct status_desc *desc; /* used to read status desc here */
+ struct status_desc *desc, *frag_desc;
u32 consumer = recv_ctx->status_rx_consumer;
int count = 0, ring;
+ u64 sts_data;
+ u16 opcode;
while (count < max) {
desc = &desc_head[consumer];
@@ -1244,9 +1305,26 @@
netxen_get_sts_owner(desc));
break;
}
- netxen_process_rcv(adapter, ctxid, desc);
+
+ sts_data = le64_to_cpu(desc->status_desc_data);
+ opcode = netxen_get_sts_opcode(sts_data);
+ frag_desc = NULL;
+ if (opcode == NETXEN_NIC_RXPKT_DESC) {
+ if (desc->nr_frags) {
+ consumer = get_next_index(consumer,
+ adapter->max_rx_desc_count);
+ frag_desc = &desc_head[consumer];
+ netxen_set_sts_owner(frag_desc,
+ STATUS_OWNER_PHANTOM);
+ }
+ }
+
+ netxen_process_rcv(adapter, ctxid, desc, frag_desc);
+
netxen_set_sts_owner(desc, STATUS_OWNER_PHANTOM);
- consumer = (consumer + 1) & (adapter->max_rx_desc_count - 1);
+
+ consumer = get_next_index(consumer,
+ adapter->max_rx_desc_count);
count++;
}
for (ring = 0; ring < adapter->max_rds_rings; ring++)
@@ -1348,36 +1426,31 @@
int index = 0;
netxen_ctx_msg msg = 0;
dma_addr_t dma;
+ struct list_head *head;
rds_ring = &recv_ctx->rds_rings[ringid];
producer = rds_ring->producer;
index = rds_ring->begin_alloc;
- buffer = &rds_ring->rx_buf_arr[index];
+ head = &rds_ring->free_list;
+
/* We can start writing rx descriptors into the phantom memory. */
- while (buffer->state == NETXEN_BUFFER_FREE) {
+ while (!list_empty(head)) {
+
skb = dev_alloc_skb(rds_ring->skb_size);
if (unlikely(!skb)) {
- /*
- * TODO
- * We need to schedule the posting of buffers to the pegs.
- */
rds_ring->begin_alloc = index;
- DPRINTK(ERR, "netxen_post_rx_buffers: "
- " allocated only %d buffers\n", count);
break;
}
+ buffer = list_entry(head->next, struct netxen_rx_buffer, list);
+ list_del(&buffer->list);
+
count++; /* now there should be no failure */
pdesc = &rds_ring->desc_head[producer];
-#if defined(XGB_DEBUG)
- *(unsigned long *)(skb->head) = 0xc0debabe;
- if (skb_is_nonlinear(skb)) {
- printk("Allocated SKB @%p is nonlinear\n");
- }
-#endif
- skb_reserve(skb, 2);
+ if (!adapter->ahw.cut_through)
+ skb_reserve(skb, 2);
/* This will be setup when we receive the
* buffer after it has been filled FSL TBD TBD
* skb->dev = netdev;
@@ -1395,7 +1468,6 @@
producer =
get_next_index(producer, rds_ring->max_rx_desc_count);
index = get_next_index(index, rds_ring->max_rx_desc_count);
- buffer = &rds_ring->rx_buf_arr[index];
}
/* if we did allocate buffers, then write the count to Phantom */
if (count) {
@@ -1439,32 +1511,29 @@
struct netxen_rx_buffer *buffer;
int count = 0;
int index = 0;
+ struct list_head *head;
rds_ring = &recv_ctx->rds_rings[ringid];
producer = rds_ring->producer;
index = rds_ring->begin_alloc;
- buffer = &rds_ring->rx_buf_arr[index];
+ head = &rds_ring->free_list;
/* We can start writing rx descriptors into the phantom memory. */
- while (buffer->state == NETXEN_BUFFER_FREE) {
+ while (!list_empty(head)) {
+
skb = dev_alloc_skb(rds_ring->skb_size);
if (unlikely(!skb)) {
- /*
- * We need to schedule the posting of buffers to the pegs.
- */
rds_ring->begin_alloc = index;
- DPRINTK(ERR, "netxen_post_rx_buffers_nodb: "
- " allocated only %d buffers\n", count);
break;
}
+
+ buffer = list_entry(head->next, struct netxen_rx_buffer, list);
+ list_del(&buffer->list);
+
count++; /* now there should be no failure */
pdesc = &rds_ring->desc_head[producer];
- skb_reserve(skb, 2);
- /*
- * This will be setup when we receive the
- * buffer after it has been filled
- * skb->dev = netdev;
- */
+ if (!adapter->ahw.cut_through)
+ skb_reserve(skb, 2);
buffer->skb = skb;
buffer->state = NETXEN_BUFFER_BUSY;
buffer->dma = pci_map_single(pdev, skb->data,