sfc: Implement firmware-assisted TSO for EF10
authorBen Hutchings <bhutchings@solarflare.com>
Fri, 8 Mar 2013 21:20:09 +0000 (21:20 +0000)
committerBen Hutchings <bhutchings@solarflare.com>
Fri, 20 Sep 2013 18:31:48 +0000 (19:31 +0100)
Segmentation remains in the driver, but we generate option descriptors
describing the required packet editing rather than making our own
copies.

Reduce tso_state::ipv4_id to 16 bits, so it doesn't overflow into the
TCP_FLAGS field of the option descriptor.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
drivers/net/ethernet/sfc/tx.c

index 9025356969289508196a3af15f25555bf93ca7b0..a23ba0dbdd1d72eb6df65f62d1c955bba961d716 100644 (file)
@@ -21,6 +21,7 @@
 #include "efx.h"
 #include "nic.h"
 #include "workarounds.h"
+#include "ef10_regs.h"
 
 static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
                               struct efx_tx_buffer *buffer,
@@ -83,8 +84,10 @@ unsigned int efx_tx_max_skb_descs(struct efx_nic *efx)
         */
        unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
 
-       /* Possibly one more per segment for the alignment workaround */
-       if (EFX_WORKAROUND_5391(efx))
+       /* Possibly one more per segment for the alignment workaround,
+        * or for option descriptors
+        */
+       if (EFX_WORKAROUND_5391(efx) || efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
                max_descs += EFX_TSO_MAX_SEGS;
 
        /* Possibly more for PCIe page boundaries within input fragments */
@@ -628,6 +631,9 @@ void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
  * @tcp_off: Offset of TCP header
  * @header_len: Number of bytes of header
  * @ip_base_len: IPv4 tot_len or IPv6 payload_len, before TCP payload
+ * @header_dma_addr: Header DMA address, when using option descriptors
+ * @header_unmap_len: Header DMA mapped length, or 0 if not using option
+ *     descriptors
  *
  * The state used during segmentation.  It is put into this data structure
  * just to make it easy to pass into inline functions.
@@ -636,7 +642,7 @@ struct tso_state {
        /* Output position */
        unsigned out_len;
        unsigned seqnum;
-       unsigned ipv4_id;
+       u16 ipv4_id;
        unsigned packet_space;
 
        /* Input position */
@@ -651,6 +657,8 @@ struct tso_state {
        unsigned int tcp_off;
        unsigned header_len;
        unsigned int ip_base_len;
+       dma_addr_t header_dma_addr;
+       unsigned int header_unmap_len;
 };
 
 
@@ -825,7 +833,10 @@ static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
 static int tso_start(struct tso_state *st, struct efx_nic *efx,
                     const struct sk_buff *skb)
 {
+       bool use_options = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
+       struct device *dma_dev = &efx->pci_dev->dev;
        unsigned int header_len, in_len;
+       dma_addr_t dma_addr;
 
        st->ip_off = skb_network_header(skb) - skb->data;
        st->tcp_off = skb_transport_header(skb) - skb->data;
@@ -848,22 +859,32 @@ static int tso_start(struct tso_state *st, struct efx_nic *efx,
 
        st->out_len = skb->len - header_len;
 
-       if (likely(in_len == 0)) {
-               st->unmap_len = 0;
+       if (!use_options) {
+               st->header_unmap_len = 0;
+
+               if (likely(in_len == 0)) {
+                       st->dma_flags = 0;
+                       st->unmap_len = 0;
+                       return 0;
+               }
+
+               dma_addr = dma_map_single(dma_dev, skb->data + header_len,
+                                         in_len, DMA_TO_DEVICE);
+               st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
+               st->dma_addr = dma_addr;
+               st->unmap_addr = dma_addr;
+               st->unmap_len = in_len;
+       } else {
+               dma_addr = dma_map_single(dma_dev, skb->data,
+                                         skb_headlen(skb), DMA_TO_DEVICE);
+               st->header_dma_addr = dma_addr;
+               st->header_unmap_len = skb_headlen(skb);
                st->dma_flags = 0;
-               return 0;
+               st->dma_addr = dma_addr + header_len;
+               st->unmap_len = 0;
        }
 
-       st->unmap_addr = dma_map_single(&efx->pci_dev->dev,
-                                       skb->data + header_len, in_len,
-                                       DMA_TO_DEVICE);
-       if (unlikely(dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr)))
-               return -ENOMEM;
-
-       st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
-       st->unmap_len = in_len;
-       st->dma_addr = st->unmap_addr;
-       return 0;
+       return unlikely(dma_mapping_error(dma_dev, dma_addr)) ? -ENOMEM : 0;
 }
 
 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
@@ -948,54 +969,97 @@ static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
 {
        struct efx_tx_buffer *buffer =
                &tx_queue->buffer[tx_queue->insert_count & tx_queue->ptr_mask];
-       struct tcphdr *tsoh_th;
-       unsigned ip_length;
-       u8 *header;
-       int rc;
-
-       /* Allocate and insert a DMA-mapped header buffer. */
-       header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len);
-       if (!header)
-               return -ENOMEM;
+       bool is_last = st->out_len <= skb_shinfo(skb)->gso_size;
+       u8 tcp_flags_clear;
 
-       tsoh_th = (struct tcphdr *)(header + st->tcp_off);
-
-       /* Copy and update the headers. */
-       memcpy(header, skb->data, st->header_len);
-
-       tsoh_th->seq = htonl(st->seqnum);
-       st->seqnum += skb_shinfo(skb)->gso_size;
-       if (st->out_len > skb_shinfo(skb)->gso_size) {
-               /* This packet will not finish the TSO burst. */
+       if (!is_last) {
                st->packet_space = skb_shinfo(skb)->gso_size;
-               tsoh_th->fin = 0;
-               tsoh_th->psh = 0;
+               tcp_flags_clear = 0x09; /* mask out FIN and PSH */
        } else {
-               /* This packet will be the last in the TSO burst. */
                st->packet_space = st->out_len;
-               tsoh_th->fin = tcp_hdr(skb)->fin;
-               tsoh_th->psh = tcp_hdr(skb)->psh;
+               tcp_flags_clear = 0x00;
        }
-       ip_length = st->ip_base_len + st->packet_space;
 
-       if (st->protocol == htons(ETH_P_IP)) {
-               struct iphdr *tsoh_iph = (struct iphdr *)(header + st->ip_off);
+       if (!st->header_unmap_len) {
+               /* Allocate and insert a DMA-mapped header buffer. */
+               struct tcphdr *tsoh_th;
+               unsigned ip_length;
+               u8 *header;
+               int rc;
 
-               tsoh_iph->tot_len = htons(ip_length);
+               header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len);
+               if (!header)
+                       return -ENOMEM;
 
-               /* Linux leaves suitable gaps in the IP ID space for us to fill. */
-               tsoh_iph->id = htons(st->ipv4_id);
-               st->ipv4_id++;
+               tsoh_th = (struct tcphdr *)(header + st->tcp_off);
+
+               /* Copy and update the headers. */
+               memcpy(header, skb->data, st->header_len);
+
+               tsoh_th->seq = htonl(st->seqnum);
+               ((u8 *)tsoh_th)[13] &= ~tcp_flags_clear;
+
+               ip_length = st->ip_base_len + st->packet_space;
+
+               if (st->protocol == htons(ETH_P_IP)) {
+                       struct iphdr *tsoh_iph =
+                               (struct iphdr *)(header + st->ip_off);
+
+                       tsoh_iph->tot_len = htons(ip_length);
+                       tsoh_iph->id = htons(st->ipv4_id);
+               } else {
+                       struct ipv6hdr *tsoh_iph =
+                               (struct ipv6hdr *)(header + st->ip_off);
+
+                       tsoh_iph->payload_len = htons(ip_length);
+               }
+
+               rc = efx_tso_put_header(tx_queue, buffer, header);
+               if (unlikely(rc))
+                       return rc;
        } else {
-               struct ipv6hdr *tsoh_iph =
-                       (struct ipv6hdr *)(header + st->ip_off);
+               /* Send the original headers with a TSO option descriptor
+                * in front
+                */
+               u8 tcp_flags = ((u8 *)tcp_hdr(skb))[13] & ~tcp_flags_clear;
+
+               buffer->flags = EFX_TX_BUF_OPTION;
+               buffer->len = 0;
+               buffer->unmap_len = 0;
+               EFX_POPULATE_QWORD_5(buffer->option,
+                                    ESF_DZ_TX_DESC_IS_OPT, 1,
+                                    ESF_DZ_TX_OPTION_TYPE,
+                                    ESE_DZ_TX_OPTION_DESC_TSO,
+                                    ESF_DZ_TX_TSO_TCP_FLAGS, tcp_flags,
+                                    ESF_DZ_TX_TSO_IP_ID, st->ipv4_id,
+                                    ESF_DZ_TX_TSO_TCP_SEQNO, st->seqnum);
+               ++tx_queue->insert_count;
 
-               tsoh_iph->payload_len = htons(ip_length);
+               /* We mapped the headers in tso_start().  Unmap them
+                * when the last segment is completed.
+                */
+               buffer = &tx_queue->buffer[tx_queue->insert_count &
+                                          tx_queue->ptr_mask];
+               buffer->dma_addr = st->header_dma_addr;
+               buffer->len = st->header_len;
+               if (is_last) {
+                       buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_MAP_SINGLE;
+                       buffer->unmap_len = st->header_unmap_len;
+                       /* Ensure we only unmap them once in case of a
+                        * later DMA mapping error and rollback
+                        */
+                       st->header_unmap_len = 0;
+               } else {
+                       buffer->flags = EFX_TX_BUF_CONT;
+                       buffer->unmap_len = 0;
+               }
+               ++tx_queue->insert_count;
        }
 
-       rc = efx_tso_put_header(tx_queue, buffer, header);
-       if (unlikely(rc))
-               return rc;
+       st->seqnum += skb_shinfo(skb)->gso_size;
+
+       /* Linux leaves suitable gaps in the IP ID space for us to fill. */
+       ++st->ipv4_id;
 
        ++tx_queue->tso_packets;
 
@@ -1091,6 +1155,11 @@ static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
                                       state.unmap_len, DMA_TO_DEVICE);
        }
 
+       /* Free the header DMA mapping, if using option descriptors */
+       if (state.header_unmap_len)
+               dma_unmap_single(&efx->pci_dev->dev, state.header_dma_addr,
+                                state.header_unmap_len, DMA_TO_DEVICE);
+
        efx_enqueue_unwind(tx_queue);
        return NETDEV_TX_OK;
 }