[netdrvr] sfc: Add TSO support
authorBen Hutchings <bhutchings@solarflare.com>
Wed, 7 May 2008 11:51:12 +0000 (12:51 +0100)
committerJeff Garzik <jgarzik@redhat.com>
Tue, 13 May 2008 05:31:40 +0000 (01:31 -0400)
The SFC4000 controller does not have hardware support for TSO, and the
core GSO code incurs a high cost in allocating and freeing skbs.  This
TSO implementation uses lightweight packet header structures and is
substantially faster.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
drivers/net/sfc/efx.c
drivers/net/sfc/ethtool.c
drivers/net/sfc/net_driver.h
drivers/net/sfc/tx.c

index 59edcf793c19253c4be6d5bceeae730327d87c78..418f2e53a95b6709df9a4ce2d3328902c22ba3d4 100644 (file)
@@ -1873,6 +1873,7 @@ static int efx_init_struct(struct efx_nic *efx, struct efx_nic_type *type,
                tx_queue->queue = i;
                tx_queue->buffer = NULL;
                tx_queue->channel = &efx->channel[0]; /* for safety */
+               tx_queue->tso_headers_free = NULL;
        }
        for (i = 0; i < EFX_MAX_RX_QUEUES; i++) {
                rx_queue = &efx->rx_queue[i];
@@ -2071,7 +2072,8 @@ static int __devinit efx_pci_probe(struct pci_dev *pci_dev,
        net_dev = alloc_etherdev(sizeof(*efx));
        if (!net_dev)
                return -ENOMEM;
-       net_dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_HIGHDMA;
+       net_dev->features |= (NETIF_F_IP_CSUM | NETIF_F_SG |
+                             NETIF_F_HIGHDMA | NETIF_F_TSO);
        if (lro)
                net_dev->features |= NETIF_F_LRO;
        efx = net_dev->priv;
index ad541badbd98bd94392111ec9c81c23c96d80839..b756840e2a16ba5636d617a797d9471b1c9cde65 100644 (file)
@@ -272,6 +272,22 @@ static void efx_ethtool_get_stats(struct net_device *net_dev,
        }
 }
 
+static int efx_ethtool_set_tso(struct net_device *net_dev, u32 enable)
+{
+       int rc;
+
+       /* Our TSO requires TX checksumming, so force TX checksumming
+        * on when TSO is enabled.
+        */
+       if (enable) {
+               rc = efx_ethtool_set_tx_csum(net_dev, 1);
+               if (rc)
+                       return rc;
+       }
+
+       return ethtool_op_set_tso(net_dev, enable);
+}
+
 static int efx_ethtool_set_tx_csum(struct net_device *net_dev, u32 enable)
 {
        struct efx_nic *efx = net_dev->priv;
@@ -283,6 +299,15 @@ static int efx_ethtool_set_tx_csum(struct net_device *net_dev, u32 enable)
 
        efx_flush_queues(efx);
 
+       /* Our TSO requires TX checksumming, so disable TSO when
+        * checksumming is disabled
+        */
+       if (!enable) {
+               rc = efx_ethtool_set_tso(net_dev, 0);
+               if (rc)
+                       return rc;
+       }
+
        return 0;
 }
 
@@ -451,6 +476,8 @@ struct ethtool_ops efx_ethtool_ops = {
        .set_tx_csum            = efx_ethtool_set_tx_csum,
        .get_sg                 = ethtool_op_get_sg,
        .set_sg                 = ethtool_op_set_sg,
+       .get_tso                = ethtool_op_get_tso,
+       .set_tso                = efx_ethtool_set_tso,
        .get_flags              = ethtool_op_get_flags,
        .set_flags              = ethtool_op_set_flags,
        .get_strings            = efx_ethtool_get_strings,
index c505482c25203d1fa8de92195c404cd4bce8683a..6ffa7116336004bf315a77aa747bb893c061fb19 100644 (file)
@@ -134,6 +134,8 @@ struct efx_special_buffer {
  *     Set only on the final fragment of a packet; %NULL for all other
  *     fragments.  When this fragment completes, then we can free this
  *     skb.
+ * @tsoh: The associated TSO header structure, or %NULL if this
+ *     buffer is not a TSO header.
  * @dma_addr: DMA address of the fragment.
  * @len: Length of this fragment.
  *     This field is zero when the queue slot is empty.
@@ -144,6 +146,7 @@ struct efx_special_buffer {
  */
 struct efx_tx_buffer {
        const struct sk_buff *skb;
+       struct efx_tso_header *tsoh;
        dma_addr_t dma_addr;
        unsigned short len;
        unsigned char continuation;
@@ -187,6 +190,13 @@ struct efx_tx_buffer {
  *     variable indicates that the queue is full.  This is to
  *     avoid cache-line ping-pong between the xmit path and the
  *     completion path.
+ * @tso_headers_free: A list of TSO headers allocated for this TX queue
+ *     that are not in use, and so available for new TSO sends. The list
+ *     is protected by the TX queue lock.
+ * @tso_bursts: Number of times TSO xmit invoked by kernel
+ * @tso_long_headers: Number of packets with headers too long for standard
+ *     blocks
+ * @tso_packets: Number of packets via the TSO xmit path
  */
 struct efx_tx_queue {
        /* Members which don't change on the fast path */
@@ -206,6 +216,10 @@ struct efx_tx_queue {
        unsigned int insert_count ____cacheline_aligned_in_smp;
        unsigned int write_count;
        unsigned int old_read_count;
+       struct efx_tso_header *tso_headers_free;
+       unsigned int tso_bursts;
+       unsigned int tso_long_headers;
+       unsigned int tso_packets;
 };
 
 /**
index fbb866b2185ebdd5e37317f9f18341e23ed684e6..9b436f5b48889148792134f397830ea4716a17fb 100644 (file)
@@ -82,6 +82,46 @@ static inline void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
        }
 }
 
+/**
+ * struct efx_tso_header - a DMA mapped buffer for packet headers
+ * @next: Linked list of free ones.
+ *     The list is protected by the TX queue lock.
+ * @dma_unmap_len: Length to unmap for an oversize buffer, or 0.
+ * @dma_addr: The DMA address of the header below.
+ *
+ * This controls the memory used for a TSO header.  Use TSOH_DATA()
+ * to find the packet header data.  Use TSOH_SIZE() to calculate the
+ * total size required for a given packet header length.  TSO headers
+ * in the free list are exactly %TSOH_STD_SIZE bytes in size.
+ */
+struct efx_tso_header {
+       union {
+               struct efx_tso_header *next;
+               size_t unmap_len;
+       };
+       dma_addr_t dma_addr;
+};
+
+static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
+                              const struct sk_buff *skb);
+static void efx_fini_tso(struct efx_tx_queue *tx_queue);
+static void efx_tsoh_heap_free(struct efx_tx_queue *tx_queue,
+                              struct efx_tso_header *tsoh);
+
+static inline void efx_tsoh_free(struct efx_tx_queue *tx_queue,
+                                struct efx_tx_buffer *buffer)
+{
+       if (buffer->tsoh) {
+               if (likely(!buffer->tsoh->unmap_len)) {
+                       buffer->tsoh->next = tx_queue->tso_headers_free;
+                       tx_queue->tso_headers_free = buffer->tsoh;
+               } else {
+                       efx_tsoh_heap_free(tx_queue, buffer->tsoh);
+               }
+               buffer->tsoh = NULL;
+       }
+}
+
 
 /*
  * Add a socket buffer to a TX queue
@@ -114,6 +154,9 @@ static inline int efx_enqueue_skb(struct efx_tx_queue *tx_queue,
 
        EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
 
+       if (skb_shinfo((struct sk_buff *)skb)->gso_size)
+               return efx_enqueue_skb_tso(tx_queue, skb);
+
        /* Get size of the initial fragment */
        len = skb_headlen(skb);
 
@@ -166,6 +209,8 @@ static inline int efx_enqueue_skb(struct efx_tx_queue *tx_queue,
                        insert_ptr = (tx_queue->insert_count &
                                      efx->type->txd_ring_mask);
                        buffer = &tx_queue->buffer[insert_ptr];
+                       efx_tsoh_free(tx_queue, buffer);
+                       EFX_BUG_ON_PARANOID(buffer->tsoh);
                        EFX_BUG_ON_PARANOID(buffer->skb);
                        EFX_BUG_ON_PARANOID(buffer->len);
                        EFX_BUG_ON_PARANOID(buffer->continuation != 1);
@@ -432,6 +477,9 @@ void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
 
        efx_release_tx_buffers(tx_queue);
 
+       /* Free up TSO header cache */
+       efx_fini_tso(tx_queue);
+
        /* Release queue's stop on port, if any */
        if (tx_queue->stopped) {
                tx_queue->stopped = 0;
@@ -450,3 +498,619 @@ void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
 }
 
 
+/* Efx TCP segmentation acceleration.
+ *
+ * Why?  Because by doing it here in the driver we can go significantly
+ * faster than the GSO.
+ *
+ * Requires TX checksum offload support.
+ */
+
+/* Number of bytes inserted at the start of a TSO header buffer,
+ * similar to NET_IP_ALIGN.
+ */
+#if defined(__i386__) || defined(__x86_64__)
+#define TSOH_OFFSET    0
+#else
+#define TSOH_OFFSET    NET_IP_ALIGN
+#endif
+
+#define TSOH_BUFFER(tsoh)      ((u8 *)(tsoh + 1) + TSOH_OFFSET)
+
+/* Total size of struct efx_tso_header, buffer and padding */
+#define TSOH_SIZE(hdr_len)                                     \
+       (sizeof(struct efx_tso_header) + TSOH_OFFSET + hdr_len)
+
+/* Size of blocks on free list.  Larger blocks must be allocated from
+ * the heap.
+ */
+#define TSOH_STD_SIZE          128
+
+#define PTR_DIFF(p1, p2)  ((u8 *)(p1) - (u8 *)(p2))
+#define ETH_HDR_LEN(skb)  (skb_network_header(skb) - (skb)->data)
+#define SKB_TCP_OFF(skb)  PTR_DIFF(tcp_hdr(skb), (skb)->data)
+#define SKB_IPV4_OFF(skb) PTR_DIFF(ip_hdr(skb), (skb)->data)
+
+/**
+ * struct tso_state - TSO state for an SKB
+ * @remaining_len: Bytes of data we've yet to segment
+ * @seqnum: Current sequence number
+ * @packet_space: Remaining space in current packet
+ * @ifc: Input fragment cursor.
+ *     Where we are in the current fragment of the incoming SKB.  These
+ *     values get updated in place when we split a fragment over
+ *     multiple packets.
+ * @p: Parameters.
+ *     These values are set once at the start of the TSO send and do
+ *     not get changed as the routine progresses.
+ *
+ * The state used during segmentation.  It is put into this data structure
+ * just to make it easy to pass into inline functions.
+ */
+struct tso_state {
+       unsigned remaining_len;
+       unsigned seqnum;
+       unsigned packet_space;
+
+       struct {
+               /* DMA address of current position */
+               dma_addr_t dma_addr;
+               /* Remaining length */
+               unsigned int len;
+               /* DMA address and length of the whole fragment */
+               unsigned int unmap_len;
+               dma_addr_t unmap_addr;
+               struct page *page;
+               unsigned page_off;
+       } ifc;
+
+       struct {
+               /* The number of bytes of header */
+               unsigned int header_length;
+
+               /* The number of bytes to put in each outgoing segment. */
+               int full_packet_size;
+
+               /* Current IPv4 ID, host endian. */
+               unsigned ipv4_id;
+       } p;
+};
+
+
+/*
+ * Verify that our various assumptions about sk_buffs and the conditions
+ * under which TSO will be attempted hold true.
+ */
+static inline void efx_tso_check_safe(const struct sk_buff *skb)
+{
+       EFX_BUG_ON_PARANOID(skb->protocol != htons(ETH_P_IP));
+       EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
+                           skb->protocol);
+       EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
+       EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
+                            + (tcp_hdr(skb)->doff << 2u)) >
+                           skb_headlen(skb));
+}
+
+
+/*
+ * Allocate a page worth of efx_tso_header structures, and string them
+ * into the tx_queue->tso_headers_free linked list. Return 0 or -ENOMEM.
+ */
+static int efx_tsoh_block_alloc(struct efx_tx_queue *tx_queue)
+{
+
+       struct pci_dev *pci_dev = tx_queue->efx->pci_dev;
+       struct efx_tso_header *tsoh;
+       dma_addr_t dma_addr;
+       u8 *base_kva, *kva;
+
+       base_kva = pci_alloc_consistent(pci_dev, PAGE_SIZE, &dma_addr);
+       if (base_kva == NULL) {
+               EFX_ERR(tx_queue->efx, "Unable to allocate page for TSO"
+                       " headers\n");
+               return -ENOMEM;
+       }
+
+       /* pci_alloc_consistent() allocates pages. */
+       EFX_BUG_ON_PARANOID(dma_addr & (PAGE_SIZE - 1u));
+
+       for (kva = base_kva; kva < base_kva + PAGE_SIZE; kva += TSOH_STD_SIZE) {
+               tsoh = (struct efx_tso_header *)kva;
+               tsoh->dma_addr = dma_addr + (TSOH_BUFFER(tsoh) - base_kva);
+               tsoh->next = tx_queue->tso_headers_free;
+               tx_queue->tso_headers_free = tsoh;
+       }
+
+       return 0;
+}
+
+
+/* Free up a TSO header, and all others in the same page. */
+static void efx_tsoh_block_free(struct efx_tx_queue *tx_queue,
+                               struct efx_tso_header *tsoh,
+                               struct pci_dev *pci_dev)
+{
+       struct efx_tso_header **p;
+       unsigned long base_kva;
+       dma_addr_t base_dma;
+
+       base_kva = (unsigned long)tsoh & PAGE_MASK;
+       base_dma = tsoh->dma_addr & PAGE_MASK;
+
+       p = &tx_queue->tso_headers_free;
+       while (*p != NULL)
+               if (((unsigned long)*p & PAGE_MASK) == base_kva)
+                       *p = (*p)->next;
+               else
+                       p = &(*p)->next;
+
+       pci_free_consistent(pci_dev, PAGE_SIZE, (void *)base_kva, base_dma);
+}
+
+static struct efx_tso_header *
+efx_tsoh_heap_alloc(struct efx_tx_queue *tx_queue, size_t header_len)
+{
+       struct efx_tso_header *tsoh;
+
+       tsoh = kmalloc(TSOH_SIZE(header_len), GFP_ATOMIC | GFP_DMA);
+       if (unlikely(!tsoh))
+               return NULL;
+
+       tsoh->dma_addr = pci_map_single(tx_queue->efx->pci_dev,
+                                       TSOH_BUFFER(tsoh), header_len,
+                                       PCI_DMA_TODEVICE);
+       if (unlikely(pci_dma_mapping_error(tsoh->dma_addr))) {
+               kfree(tsoh);
+               return NULL;
+       }
+
+       tsoh->unmap_len = header_len;
+       return tsoh;
+}
+
+static void
+efx_tsoh_heap_free(struct efx_tx_queue *tx_queue, struct efx_tso_header *tsoh)
+{
+       pci_unmap_single(tx_queue->efx->pci_dev,
+                        tsoh->dma_addr, tsoh->unmap_len,
+                        PCI_DMA_TODEVICE);
+       kfree(tsoh);
+}
+
+/**
+ * efx_tx_queue_insert - push descriptors onto the TX queue
+ * @tx_queue:          Efx TX queue
+ * @dma_addr:          DMA address of fragment
+ * @len:               Length of fragment
+ * @skb:               Only non-null for end of last segment
+ * @end_of_packet:     True if last fragment in a packet
+ * @unmap_addr:                DMA address of fragment for unmapping
+ * @unmap_len:         Only set this in last segment of a fragment
+ *
+ * Push descriptors onto the TX queue.  Return 0 on success or 1 if
+ * @tx_queue full.
+ */
+static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
+                              dma_addr_t dma_addr, unsigned len,
+                              const struct sk_buff *skb, int end_of_packet,
+                              dma_addr_t unmap_addr, unsigned unmap_len)
+{
+       struct efx_tx_buffer *buffer;
+       struct efx_nic *efx = tx_queue->efx;
+       unsigned dma_len, fill_level, insert_ptr, misalign;
+       int q_space;
+
+       EFX_BUG_ON_PARANOID(len <= 0);
+
+       fill_level = tx_queue->insert_count - tx_queue->old_read_count;
+       /* -1 as there is no way to represent all descriptors used */
+       q_space = efx->type->txd_ring_mask - 1 - fill_level;
+
+       while (1) {
+               if (unlikely(q_space-- <= 0)) {
+                       /* It might be that completions have happened
+                        * since the xmit path last checked.  Update
+                        * the xmit path's copy of read_count.
+                        */
+                       ++tx_queue->stopped;
+                       /* This memory barrier protects the change of
+                        * stopped from the access of read_count. */
+                       smp_mb();
+                       tx_queue->old_read_count =
+                               *(volatile unsigned *)&tx_queue->read_count;
+                       fill_level = (tx_queue->insert_count
+                                     - tx_queue->old_read_count);
+                       q_space = efx->type->txd_ring_mask - 1 - fill_level;
+                       if (unlikely(q_space-- <= 0))
+                               return 1;
+                       smp_mb();
+                       --tx_queue->stopped;
+               }
+
+               insert_ptr = tx_queue->insert_count & efx->type->txd_ring_mask;
+               buffer = &tx_queue->buffer[insert_ptr];
+               ++tx_queue->insert_count;
+
+               EFX_BUG_ON_PARANOID(tx_queue->insert_count -
+                                   tx_queue->read_count >
+                                   efx->type->txd_ring_mask);
+
+               efx_tsoh_free(tx_queue, buffer);
+               EFX_BUG_ON_PARANOID(buffer->len);
+               EFX_BUG_ON_PARANOID(buffer->unmap_len);
+               EFX_BUG_ON_PARANOID(buffer->skb);
+               EFX_BUG_ON_PARANOID(buffer->continuation != 1);
+               EFX_BUG_ON_PARANOID(buffer->tsoh);
+
+               buffer->dma_addr = dma_addr;
+
+               /* Ensure we do not cross a boundary unsupported by H/W */
+               dma_len = (~dma_addr & efx->type->tx_dma_mask) + 1;
+
+               misalign = (unsigned)dma_addr & efx->type->bug5391_mask;
+               if (misalign && dma_len + misalign > 512)
+                       dma_len = 512 - misalign;
+
+               /* If there is enough space to send then do so */
+               if (dma_len >= len)
+                       break;
+
+               buffer->len = dma_len; /* Don't set the other members */
+               dma_addr += dma_len;
+               len -= dma_len;
+       }
+
+       EFX_BUG_ON_PARANOID(!len);
+       buffer->len = len;
+       buffer->skb = skb;
+       buffer->continuation = !end_of_packet;
+       buffer->unmap_addr = unmap_addr;
+       buffer->unmap_len = unmap_len;
+       return 0;
+}
+
+
+/*
+ * Put a TSO header into the TX queue.
+ *
+ * This is special-cased because we know that it is small enough to fit in
+ * a single fragment, and we know it doesn't cross a page boundary.  It
+ * also allows us to not worry about end-of-packet etc.
+ */
+static inline void efx_tso_put_header(struct efx_tx_queue *tx_queue,
+                                     struct efx_tso_header *tsoh, unsigned len)
+{
+       struct efx_tx_buffer *buffer;
+
+       buffer = &tx_queue->buffer[tx_queue->insert_count &
+                                  tx_queue->efx->type->txd_ring_mask];
+       efx_tsoh_free(tx_queue, buffer);
+       EFX_BUG_ON_PARANOID(buffer->len);
+       EFX_BUG_ON_PARANOID(buffer->unmap_len);
+       EFX_BUG_ON_PARANOID(buffer->skb);
+       EFX_BUG_ON_PARANOID(buffer->continuation != 1);
+       EFX_BUG_ON_PARANOID(buffer->tsoh);
+       buffer->len = len;
+       buffer->dma_addr = tsoh->dma_addr;
+       buffer->tsoh = tsoh;
+
+       ++tx_queue->insert_count;
+}
+
+
+/* Remove descriptors put into a tx_queue. */
+static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
+{
+       struct efx_tx_buffer *buffer;
+
+       /* Work backwards until we hit the original insert pointer value */
+       while (tx_queue->insert_count != tx_queue->write_count) {
+               --tx_queue->insert_count;
+               buffer = &tx_queue->buffer[tx_queue->insert_count &
+                                          tx_queue->efx->type->txd_ring_mask];
+               efx_tsoh_free(tx_queue, buffer);
+               EFX_BUG_ON_PARANOID(buffer->skb);
+               buffer->len = 0;
+               buffer->continuation = 1;
+               if (buffer->unmap_len) {
+                       pci_unmap_page(tx_queue->efx->pci_dev,
+                                      buffer->unmap_addr,
+                                      buffer->unmap_len, PCI_DMA_TODEVICE);
+                       buffer->unmap_len = 0;
+               }
+       }
+}
+
+
+/* Parse the SKB header and initialise state. */
+static inline void tso_start(struct tso_state *st, const struct sk_buff *skb)
+{
+       /* All ethernet/IP/TCP headers combined size is TCP header size
+        * plus offset of TCP header relative to start of packet.
+        */
+       st->p.header_length = ((tcp_hdr(skb)->doff << 2u)
+                              + PTR_DIFF(tcp_hdr(skb), skb->data));
+       st->p.full_packet_size = (st->p.header_length
+                                 + skb_shinfo(skb)->gso_size);
+
+       st->p.ipv4_id = ntohs(ip_hdr(skb)->id);
+       st->seqnum = ntohl(tcp_hdr(skb)->seq);
+
+       EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
+       EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
+       EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
+
+       st->packet_space = st->p.full_packet_size;
+       st->remaining_len = skb->len - st->p.header_length;
+}
+
+
+/**
+ * tso_get_fragment - record fragment details and map for DMA
+ * @st:                        TSO state
+ * @efx:               Efx NIC
+ * @data:              Pointer to fragment data
+ * @len:               Length of fragment
+ *
+ * Record fragment details and map for DMA.  Return 0 on success, or
+ * -%ENOMEM if DMA mapping fails.
+ */
+static inline int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
+                                  int len, struct page *page, int page_off)
+{
+
+       st->ifc.unmap_addr = pci_map_page(efx->pci_dev, page, page_off,
+                                         len, PCI_DMA_TODEVICE);
+       if (likely(!pci_dma_mapping_error(st->ifc.unmap_addr))) {
+               st->ifc.unmap_len = len;
+               st->ifc.len = len;
+               st->ifc.dma_addr = st->ifc.unmap_addr;
+               st->ifc.page = page;
+               st->ifc.page_off = page_off;
+               return 0;
+       }
+       return -ENOMEM;
+}
+
+
+/**
+ * tso_fill_packet_with_fragment - form descriptors for the current fragment
+ * @tx_queue:          Efx TX queue
+ * @skb:               Socket buffer
+ * @st:                        TSO state
+ *
+ * Form descriptors for the current fragment, until we reach the end
+ * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
+ * space in @tx_queue.
+ */
+static inline int tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
+                                               const struct sk_buff *skb,
+                                               struct tso_state *st)
+{
+
+       int n, end_of_packet, rc;
+
+       if (st->ifc.len == 0)
+               return 0;
+       if (st->packet_space == 0)
+               return 0;
+
+       EFX_BUG_ON_PARANOID(st->ifc.len <= 0);
+       EFX_BUG_ON_PARANOID(st->packet_space <= 0);
+
+       n = min(st->ifc.len, st->packet_space);
+
+       st->packet_space -= n;
+       st->remaining_len -= n;
+       st->ifc.len -= n;
+       st->ifc.page_off += n;
+       end_of_packet = st->remaining_len == 0 || st->packet_space == 0;
+
+       rc = efx_tx_queue_insert(tx_queue, st->ifc.dma_addr, n,
+                                st->remaining_len ? NULL : skb,
+                                end_of_packet, st->ifc.unmap_addr,
+                                st->ifc.len ? 0 : st->ifc.unmap_len);
+
+       st->ifc.dma_addr += n;
+
+       return rc;
+}
+
+
+/**
+ * tso_start_new_packet - generate a new header and prepare for the new packet
+ * @tx_queue:          Efx TX queue
+ * @skb:               Socket buffer
+ * @st:                        TSO state
+ *
+ * Generate a new header and prepare for the new packet.  Return 0 on
+ * success, or -1 if failed to alloc header.
+ */
+static inline int tso_start_new_packet(struct efx_tx_queue *tx_queue,
+                                      const struct sk_buff *skb,
+                                      struct tso_state *st)
+{
+       struct efx_tso_header *tsoh;
+       struct iphdr *tsoh_iph;
+       struct tcphdr *tsoh_th;
+       unsigned ip_length;
+       u8 *header;
+
+       /* Allocate a DMA-mapped header buffer. */
+       if (likely(TSOH_SIZE(st->p.header_length) <= TSOH_STD_SIZE)) {
+               if (tx_queue->tso_headers_free == NULL)
+                       if (efx_tsoh_block_alloc(tx_queue))
+                               return -1;
+               EFX_BUG_ON_PARANOID(!tx_queue->tso_headers_free);
+               tsoh = tx_queue->tso_headers_free;
+               tx_queue->tso_headers_free = tsoh->next;
+               tsoh->unmap_len = 0;
+       } else {
+               tx_queue->tso_long_headers++;
+               tsoh = efx_tsoh_heap_alloc(tx_queue, st->p.header_length);
+               if (unlikely(!tsoh))
+                       return -1;
+       }
+
+       header = TSOH_BUFFER(tsoh);
+       tsoh_th = (struct tcphdr *)(header + SKB_TCP_OFF(skb));
+       tsoh_iph = (struct iphdr *)(header + SKB_IPV4_OFF(skb));
+
+       /* Copy and update the headers. */
+       memcpy(header, skb->data, st->p.header_length);
+
+       tsoh_th->seq = htonl(st->seqnum);
+       st->seqnum += skb_shinfo(skb)->gso_size;
+       if (st->remaining_len > skb_shinfo(skb)->gso_size) {
+               /* This packet will not finish the TSO burst. */
+               ip_length = st->p.full_packet_size - ETH_HDR_LEN(skb);
+               tsoh_th->fin = 0;
+               tsoh_th->psh = 0;
+       } else {
+               /* This packet will be the last in the TSO burst. */
+               ip_length = (st->p.header_length - ETH_HDR_LEN(skb)
+                            + st->remaining_len);
+               tsoh_th->fin = tcp_hdr(skb)->fin;
+               tsoh_th->psh = tcp_hdr(skb)->psh;
+       }
+       tsoh_iph->tot_len = htons(ip_length);
+
+       /* Linux leaves suitable gaps in the IP ID space for us to fill. */
+       tsoh_iph->id = htons(st->p.ipv4_id);
+       st->p.ipv4_id++;
+
+       st->packet_space = skb_shinfo(skb)->gso_size;
+       ++tx_queue->tso_packets;
+
+       /* Form a descriptor for this header. */
+       efx_tso_put_header(tx_queue, tsoh, st->p.header_length);
+
+       return 0;
+}
+
+
+/**
+ * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
+ * @tx_queue:          Efx TX queue
+ * @skb:               Socket buffer
+ *
+ * Context: You must hold netif_tx_lock() to call this function.
+ *
+ * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
+ * @skb was not enqueued.  In all cases @skb is consumed.  Return
+ * %NETDEV_TX_OK or %NETDEV_TX_BUSY.
+ */
+static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
+                              const struct sk_buff *skb)
+{
+       int frag_i, rc, rc2 = NETDEV_TX_OK;
+       struct tso_state state;
+       skb_frag_t *f;
+
+       /* Verify TSO is safe - these checks should never fail. */
+       efx_tso_check_safe(skb);
+
+       EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
+
+       tso_start(&state, skb);
+
+       /* Assume that skb header area contains exactly the headers, and
+        * all payload is in the frag list.
+        */
+       if (skb_headlen(skb) == state.p.header_length) {
+               /* Grab the first payload fragment. */
+               EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
+               frag_i = 0;
+               f = &skb_shinfo(skb)->frags[frag_i];
+               rc = tso_get_fragment(&state, tx_queue->efx,
+                                     f->size, f->page, f->page_offset);
+               if (rc)
+                       goto mem_err;
+       } else {
+               /* It may look like this code fragment assumes that the
+                * skb->data portion does not cross a page boundary, but
+                * that is not the case.  It is guaranteed to be direct
+                * mapped memory, and therefore is physically contiguous,
+                * and so DMA will work fine.  kmap_atomic() on this region
+                * will just return the direct mapping, so that will work
+                * too.
+                */
+               int page_off = (unsigned long)skb->data & (PAGE_SIZE - 1);
+               int hl = state.p.header_length;
+               rc = tso_get_fragment(&state, tx_queue->efx,
+                                     skb_headlen(skb) - hl,
+                                     virt_to_page(skb->data), page_off + hl);
+               if (rc)
+                       goto mem_err;
+               frag_i = -1;
+       }
+
+       if (tso_start_new_packet(tx_queue, skb, &state) < 0)
+               goto mem_err;
+
+       while (1) {
+               rc = tso_fill_packet_with_fragment(tx_queue, skb, &state);
+               if (unlikely(rc))
+                       goto stop;
+
+               /* Move onto the next fragment? */
+               if (state.ifc.len == 0) {
+                       if (++frag_i >= skb_shinfo(skb)->nr_frags)
+                               /* End of payload reached. */
+                               break;
+                       f = &skb_shinfo(skb)->frags[frag_i];
+                       rc = tso_get_fragment(&state, tx_queue->efx,
+                                             f->size, f->page, f->page_offset);
+                       if (rc)
+                               goto mem_err;
+               }
+
+               /* Start at new packet? */
+               if (state.packet_space == 0 &&
+                   tso_start_new_packet(tx_queue, skb, &state) < 0)
+                       goto mem_err;
+       }
+
+       /* Pass off to hardware */
+       falcon_push_buffers(tx_queue);
+
+       tx_queue->tso_bursts++;
+       return NETDEV_TX_OK;
+
+ mem_err:
+       EFX_ERR(tx_queue->efx, "Out of memory for TSO headers, or PCI mapping"
+               " error\n");
+       dev_kfree_skb_any((struct sk_buff *)skb);
+       goto unwind;
+
+ stop:
+       rc2 = NETDEV_TX_BUSY;
+
+       /* Stop the queue if it wasn't stopped before. */
+       if (tx_queue->stopped == 1)
+               efx_stop_queue(tx_queue->efx);
+
+ unwind:
+       efx_enqueue_unwind(tx_queue);
+       return rc2;
+}
+
+
+/*
+ * Free up all TSO datastructures associated with tx_queue. This
+ * routine should be called only once the tx_queue is both empty and
+ * will no longer be used.
+ */
+static void efx_fini_tso(struct efx_tx_queue *tx_queue)
+{
+       unsigned i;
+
+       if (tx_queue->buffer)
+               for (i = 0; i <= tx_queue->efx->type->txd_ring_mask; ++i)
+                       efx_tsoh_free(tx_queue, &tx_queue->buffer[i]);
+
+       while (tx_queue->tso_headers_free != NULL)
+               efx_tsoh_block_free(tx_queue, tx_queue->tso_headers_free,
+                                   tx_queue->efx->pci_dev);
+}