IPoIB: Add send gather support
authorEli Cohen <eli@mellanox.co.il>
Wed, 30 Jan 2008 16:30:53 +0000 (18:30 +0200)
committerRoland Dreier <rolandd@cisco.com>
Fri, 8 Feb 2008 22:32:37 +0000 (14:32 -0800)
This patch acts as a preparation for using checksum offload for IB
devices capable of inserting/verifying checksum in IP packets.  The
patch does not actaully turn on NETIF_F_SG - we defer that to the
patches adding checksum offload capabilities.

We only add support for send gathers for datagram mode, since existing
HW does not support checksum offload on connected QPs.

Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il>
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/ulp/ipoib/ipoib.h
drivers/infiniband/ulp/ipoib/ipoib_cm.c
drivers/infiniband/ulp/ipoib/ipoib_ib.c
drivers/infiniband/ulp/ipoib/ipoib_verbs.c

index fe250c60607d9a4a700ff0e6461cf586c6d547e5..f9b7caa541439666e4379b1c8b2cff00c60ed181 100644 (file)
@@ -143,7 +143,7 @@ struct ipoib_rx_buf {
 
 struct ipoib_tx_buf {
        struct sk_buff *skb;
-       u64             mapping;
+       u64             mapping[MAX_SKB_FRAGS + 1];
 };
 
 struct ib_cm_id;
@@ -296,7 +296,7 @@ struct ipoib_dev_priv {
        struct ipoib_tx_buf *tx_ring;
        unsigned             tx_head;
        unsigned             tx_tail;
-       struct ib_sge        tx_sge;
+       struct ib_sge        tx_sge[MAX_SKB_FRAGS + 1];
        struct ib_send_wr    tx_wr;
        unsigned             tx_outstanding;
 
index 1818f958c25027543aa2b36890b2ae0c5464aa85..7dd2ec473d24e297950aefffee9609959301b51f 100644 (file)
@@ -634,8 +634,8 @@ static inline int post_send(struct ipoib_dev_priv *priv,
 {
        struct ib_send_wr *bad_wr;
 
-       priv->tx_sge.addr       = addr;
-       priv->tx_sge.length     = len;
+       priv->tx_sge[0].addr          = addr;
+       priv->tx_sge[0].length        = len;
 
        priv->tx_wr.wr_id       = wr_id | IPOIB_OP_CM;
 
@@ -676,7 +676,7 @@ void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_
                return;
        }
 
-       tx_req->mapping = addr;
+       tx_req->mapping[0] = addr;
 
        if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
                               addr, skb->len))) {
@@ -715,7 +715,7 @@ void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 
        tx_req = &tx->tx_ring[wr_id];
 
-       ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
+       ib_dma_unmap_single(priv->ca, tx_req->mapping[0], tx_req->skb->len, DMA_TO_DEVICE);
 
        /* FIXME: is this right? Shouldn't we only increment on success? */
        ++dev->stats.tx_packets;
@@ -1110,7 +1110,7 @@ timeout:
 
        while ((int) p->tx_tail - (int) p->tx_head < 0) {
                tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
-               ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
+               ib_dma_unmap_single(priv->ca, tx_req->mapping[0], tx_req->skb->len,
                                    DMA_TO_DEVICE);
                dev_kfree_skb_any(tx_req->skb);
                ++p->tx_tail;
index 52bc2bd5799ae227a66046bfd2d79e3f4b1028b2..9d3e778dc56d62f41582521d03682b6739a0c930 100644 (file)
@@ -239,6 +239,54 @@ repost:
                           "for buf %d\n", wr_id);
 }
 
+static int ipoib_dma_map_tx(struct ib_device *ca,
+                           struct ipoib_tx_buf *tx_req)
+{
+       struct sk_buff *skb = tx_req->skb;
+       u64 *mapping = tx_req->mapping;
+       int i;
+
+       mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
+                                      DMA_TO_DEVICE);
+       if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
+               return -EIO;
+
+       for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
+               skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
+               mapping[i + 1] = ib_dma_map_page(ca, frag->page,
+                                                frag->page_offset, frag->size,
+                                                DMA_TO_DEVICE);
+               if (unlikely(ib_dma_mapping_error(ca, mapping[i + 1])))
+                       goto partial_error;
+       }
+       return 0;
+
+partial_error:
+       ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
+
+       for (; i > 0; --i) {
+               skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
+               ib_dma_unmap_page(ca, mapping[i], frag->size, DMA_TO_DEVICE);
+       }
+       return -EIO;
+}
+
+static void ipoib_dma_unmap_tx(struct ib_device *ca,
+                              struct ipoib_tx_buf *tx_req)
+{
+       struct sk_buff *skb = tx_req->skb;
+       u64 *mapping = tx_req->mapping;
+       int i;
+
+       ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
+
+       for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
+               skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
+               ib_dma_unmap_page(ca, mapping[i + 1], frag->size,
+                                 DMA_TO_DEVICE);
+       }
+}
+
 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 {
        struct ipoib_dev_priv *priv = netdev_priv(dev);
@@ -257,8 +305,7 @@ static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 
        tx_req = &priv->tx_ring[wr_id];
 
-       ib_dma_unmap_single(priv->ca, tx_req->mapping,
-                           tx_req->skb->len, DMA_TO_DEVICE);
+       ipoib_dma_unmap_tx(priv->ca, tx_req);
 
        ++dev->stats.tx_packets;
        dev->stats.tx_bytes += tx_req->skb->len;
@@ -341,16 +388,23 @@ void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 static inline int post_send(struct ipoib_dev_priv *priv,
                            unsigned int wr_id,
                            struct ib_ah *address, u32 qpn,
-                           u64 addr, int len)
+                           u64 *mapping, int headlen,
+                           skb_frag_t *frags,
+                           int nr_frags)
 {
        struct ib_send_wr *bad_wr;
+       int i;
 
-       priv->tx_sge.addr             = addr;
-       priv->tx_sge.length           = len;
-
-       priv->tx_wr.wr_id             = wr_id;
-       priv->tx_wr.wr.ud.remote_qpn  = qpn;
-       priv->tx_wr.wr.ud.ah          = address;
+       priv->tx_sge[0].addr         = mapping[0];
+       priv->tx_sge[0].length       = headlen;
+       for (i = 0; i < nr_frags; ++i) {
+               priv->tx_sge[i + 1].addr = mapping[i + 1];
+               priv->tx_sge[i + 1].length = frags[i].size;
+       }
+       priv->tx_wr.num_sge          = nr_frags + 1;
+       priv->tx_wr.wr_id            = wr_id;
+       priv->tx_wr.wr.ud.remote_qpn = qpn;
+       priv->tx_wr.wr.ud.ah         = address;
 
        return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 }
@@ -360,7 +414,6 @@ void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 {
        struct ipoib_dev_priv *priv = netdev_priv(dev);
        struct ipoib_tx_buf *tx_req;
-       u64 addr;
 
        if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
                ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
@@ -383,20 +436,19 @@ void ipoib_send(struct net_device *dev, struct sk_buff *skb,
         */
        tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
        tx_req->skb = skb;
-       addr = ib_dma_map_single(priv->ca, skb->data, skb->len,
-                                DMA_TO_DEVICE);
-       if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
+       if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
                ++dev->stats.tx_errors;
                dev_kfree_skb_any(skb);
                return;
        }
-       tx_req->mapping = addr;
 
        if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
-                              address->ah, qpn, addr, skb->len))) {
+                              address->ah, qpn,
+                              tx_req->mapping, skb_headlen(skb),
+                              skb_shinfo(skb)->frags, skb_shinfo(skb)->nr_frags))) {
                ipoib_warn(priv, "post_send failed\n");
                ++dev->stats.tx_errors;
-               ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
+               ipoib_dma_unmap_tx(priv->ca, tx_req);
                dev_kfree_skb_any(skb);
        } else {
                dev->trans_start = jiffies;
@@ -615,10 +667,7 @@ int ipoib_ib_dev_stop(struct net_device *dev, int flush)
                        while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
                                tx_req = &priv->tx_ring[priv->tx_tail &
                                                        (ipoib_sendq_size - 1)];
-                               ib_dma_unmap_single(priv->ca,
-                                                   tx_req->mapping,
-                                                   tx_req->skb->len,
-                                                   DMA_TO_DEVICE);
+                               ipoib_dma_unmap_tx(priv->ca, tx_req);
                                dev_kfree_skb_any(tx_req->skb);
                                ++priv->tx_tail;
                                --priv->tx_outstanding;
index 433e99ac227b85a2bd6e8594c69fe47f3dc6e274..a3aeb911f024f6b5d62ebd19fedd2a82251e6518 100644 (file)
@@ -157,6 +157,7 @@ int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
        };
 
        int ret, size;
+       int i;
 
        priv->pd = ib_alloc_pd(priv->ca);
        if (IS_ERR(priv->pd)) {
@@ -191,6 +192,9 @@ int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
        init_attr.send_cq = priv->cq;
        init_attr.recv_cq = priv->cq;
 
+       if (dev->features & NETIF_F_SG)
+               init_attr.cap.max_send_sge = MAX_SKB_FRAGS + 1;
+
        priv->qp = ib_create_qp(priv->pd, &init_attr);
        if (IS_ERR(priv->qp)) {
                printk(KERN_WARNING "%s: failed to create QP\n", ca->name);
@@ -201,11 +205,11 @@ int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
        priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
        priv->dev->dev_addr[3] = (priv->qp->qp_num      ) & 0xff;
 
-       priv->tx_sge.lkey       = priv->mr->lkey;
+       for (i = 0; i < MAX_SKB_FRAGS + 1; ++i)
+               priv->tx_sge[i].lkey = priv->mr->lkey;
 
        priv->tx_wr.opcode      = IB_WR_SEND;
-       priv->tx_wr.sg_list     = &priv->tx_sge;
-       priv->tx_wr.num_sge     = 1;
+       priv->tx_wr.sg_list     = priv->tx_sge;
        priv->tx_wr.send_flags  = IB_SEND_SIGNALED;
 
        return 0;