net: add gro_cells infrastructure
authorEric Dumazet <edumazet@google.com>
Thu, 27 Sep 2012 19:29:05 +0000 (19:29 +0000)
committerDavid S. Miller <davem@davemloft.net>
Mon, 1 Oct 2012 21:01:46 +0000 (17:01 -0400)
This adds a new include file (include/net/gro_cells.h), to bring GRO
(Generic Receive Offload) capability to tunnels, in a modular way.

Because tunnels receive path is lockless, and GRO adds a serialization
using a napi_struct, I chose to add an array of up to
DEFAULT_MAX_NUM_RSS_QUEUES cells, so that multi queue devices wont be
slowed down because of GRO layer.

skb_get_rx_queue() is used as selector.

In the future, we might add optional fanout capabilities, using rxhash
for example.

With help from Ben Hutchings who reminded me
netif_get_num_default_rss_queues() function.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/gro_cells.h [new file with mode: 0644]
net/core/dev.c

diff --git a/include/net/gro_cells.h b/include/net/gro_cells.h
new file mode 100644 (file)
index 0000000..4fd8a4b
--- /dev/null
@@ -0,0 +1,103 @@
+#ifndef _NET_GRO_CELLS_H
+#define _NET_GRO_CELLS_H
+
+#include <linux/skbuff.h>
+#include <linux/slab.h>
+#include <linux/netdevice.h>
+
+struct gro_cell {
+       struct sk_buff_head     napi_skbs;
+       struct napi_struct      napi;
+} ____cacheline_aligned_in_smp;
+
+struct gro_cells {
+       unsigned int            gro_cells_mask;
+       struct gro_cell         *cells;
+};
+
+static inline void gro_cells_receive(struct gro_cells *gcells, struct sk_buff *skb)
+{
+       unsigned long flags;
+       struct gro_cell *cell = gcells->cells;
+       struct net_device *dev = skb->dev;
+
+       if (!cell || skb_cloned(skb) || !(dev->features & NETIF_F_GRO)) {
+               netif_rx(skb);
+               return;
+       }
+
+       if (skb_rx_queue_recorded(skb))
+               cell += skb_get_rx_queue(skb) & gcells->gro_cells_mask;
+
+       if (skb_queue_len(&cell->napi_skbs) > netdev_max_backlog) {
+               atomic_long_inc(&dev->rx_dropped);
+               kfree_skb(skb);
+               return;
+       }
+
+       spin_lock_irqsave(&cell->napi_skbs.lock, flags);
+
+       __skb_queue_tail(&cell->napi_skbs, skb);
+       if (skb_queue_len(&cell->napi_skbs) == 1)
+               napi_schedule(&cell->napi);
+
+       spin_unlock_irqrestore(&cell->napi_skbs.lock, flags);
+}
+
+static inline int gro_cell_poll(struct napi_struct *napi, int budget)
+{
+       struct gro_cell *cell = container_of(napi, struct gro_cell, napi);
+       struct sk_buff *skb;
+       int work_done = 0;
+
+       while (work_done < budget) {
+               skb = skb_dequeue(&cell->napi_skbs);
+               if (!skb)
+                       break;
+
+               napi_gro_receive(napi, skb);
+               work_done++;
+       }
+
+       if (work_done < budget)
+               napi_complete(napi);
+       return work_done;
+}
+
+static inline int gro_cells_init(struct gro_cells *gcells, struct net_device *dev)
+{
+       int i;
+
+       gcells->gro_cells_mask = roundup_pow_of_two(netif_get_num_default_rss_queues()) - 1;
+       gcells->cells = kcalloc(sizeof(struct gro_cell),
+                               gcells->gro_cells_mask + 1,
+                               GFP_KERNEL);
+       if (!gcells->cells)
+               return -ENOMEM;
+
+       for (i = 0; i <= gcells->gro_cells_mask; i++) {
+               struct gro_cell *cell = gcells->cells + i;
+
+               skb_queue_head_init(&cell->napi_skbs);
+               netif_napi_add(dev, &cell->napi, gro_cell_poll, 64);
+               napi_enable(&cell->napi);
+       }
+       return 0;
+}
+
+static inline void gro_cells_destroy(struct gro_cells *gcells)
+{
+       struct gro_cell *cell = gcells->cells;
+       int i;
+
+       if (!cell)
+               return;
+       for (i = 0; i <= gcells->gro_cells_mask; i++,cell++) {
+               netif_napi_del(&cell->napi);
+               skb_queue_purge(&cell->napi_skbs);
+       }
+       kfree(gcells->cells);
+       gcells->cells = NULL;
+}
+
+#endif
index 3e645f3..ba4bb11 100644 (file)
@@ -2645,6 +2645,8 @@ EXPORT_SYMBOL(dev_queue_xmit);
   =======================================================================*/
 
 int netdev_max_backlog __read_mostly = 1000;
+EXPORT_SYMBOL(netdev_max_backlog);
+
 int netdev_tstamp_prequeue __read_mostly = 1;
 int netdev_budget __read_mostly = 300;
 int weight_p __read_mostly = 64;            /* old backlog weight */