]> err.no Git - linux-2.6/commitdiff
[NET]: Generic Large Receive Offload for TCP traffic
authorJan-Bernd Themann <themann@de.ibm.com>
Thu, 9 Aug 2007 05:38:05 +0000 (22:38 -0700)
committerDavid S. Miller <davem@sunset.davemloft.net>
Wed, 10 Oct 2007 23:47:46 +0000 (16:47 -0700)
This patch provides generic Large Receive Offload (LRO) functionality
for IPv4/TCP traffic.

LRO combines received tcp packets to a single larger tcp packet and
passes them then to the network stack in order to increase performance
(throughput). The interface supports two modes: Drivers can either
pass SKBs or fragment lists to the LRO engine.

Signed-off-by: Jan-Bernd Themann <themann@de.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/inet_lro.h [new file with mode: 0644]
net/ipv4/Kconfig
net/ipv4/Makefile
net/ipv4/inet_lro.c [new file with mode: 0644]

diff --git a/include/linux/inet_lro.h b/include/linux/inet_lro.h
new file mode 100644 (file)
index 0000000..e1fc1d1
--- /dev/null
@@ -0,0 +1,177 @@
+/*
+ *  linux/include/linux/inet_lro.h
+ *
+ *  Large Receive Offload (ipv4 / tcp)
+ *
+ *  (C) Copyright IBM Corp. 2007
+ *
+ *  Authors:
+ *       Jan-Bernd Themann <themann@de.ibm.com>
+ *       Christoph Raisch <raisch@de.ibm.com>
+ *
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2, or (at your option)
+ * any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+#ifndef __INET_LRO_H_
+#define __INET_LRO_H_
+
+#include <net/ip.h>
+#include <net/tcp.h>
+
+/*
+ * LRO statistics
+ */
+
+struct net_lro_stats {
+       unsigned long aggregated;
+       unsigned long flushed;
+       unsigned long no_desc;
+};
+
+/*
+ * LRO descriptor for a tcp session
+ */
+struct net_lro_desc {
+       struct sk_buff *parent;
+       struct sk_buff *last_skb;
+       struct skb_frag_struct *next_frag;
+       struct iphdr *iph;
+       struct tcphdr *tcph;
+       struct vlan_group *vgrp;
+       __wsum  data_csum;
+       u32 tcp_rcv_tsecr;
+       u32 tcp_rcv_tsval;
+       u32 tcp_ack;
+       u32 tcp_next_seq;
+       u32 skb_tot_frags_len;
+       u16 ip_tot_len;
+       u16 tcp_saw_tstamp;             /* timestamps enabled */
+       u16 tcp_window;
+       u16 vlan_tag;
+       int pkt_aggr_cnt;               /* counts aggregated packets */
+       int vlan_packet;
+       int mss;
+       int active;
+};
+
+/*
+ * Large Receive Offload (LRO) Manager
+ *
+ * Fields must be set by driver
+ */
+
+struct net_lro_mgr {
+       struct net_device *dev;
+       struct net_lro_stats stats;
+
+       /* LRO features */
+       unsigned long features;
+#define LRO_F_NAPI            1  /* Pass packets to stack via NAPI */
+#define LRO_F_EXTRACT_VLAN_ID 2  /* Set flag if VLAN IDs are extracted
+                                   from received packets and eth protocol
+                                   is still ETH_P_8021Q */
+
+       u32 ip_summed;      /* Set in non generated SKBs in page mode */
+       u32 ip_summed_aggr; /* Set in aggregated SKBs: CHECKSUM_UNNECESSARY
+                            * or CHECKSUM_NONE */
+
+       int max_desc; /* Max number of LRO descriptors  */
+       int max_aggr; /* Max number of LRO packets to be aggregated */
+
+       struct net_lro_desc *lro_arr; /* Array of LRO descriptors */
+
+       /*
+        * Optimized driver functions
+        *
+        * get_skb_header: returns tcp and ip header for packet in SKB
+        */
+       int (*get_skb_header)(struct sk_buff *skb, void **ip_hdr,
+                             void **tcpudp_hdr, u64 *hdr_flags, void *priv);
+
+       /* hdr_flags: */
+#define LRO_IPV4 1 /* ip_hdr is IPv4 header */
+#define LRO_TCP  2 /* tcpudp_hdr is TCP header */
+
+       /*
+        * get_frag_header: returns mac, tcp and ip header for packet in SKB
+        *
+        * @hdr_flags: Indicate what kind of LRO has to be done
+        *             (IPv4/IPv6/TCP/UDP)
+        */
+       int (*get_frag_header)(struct skb_frag_struct *frag, void **mac_hdr,
+                              void **ip_hdr, void **tcpudp_hdr, u64 *hdr_flags,
+                              void *priv);
+};
+
+/*
+ * Processes a SKB
+ *
+ * @lro_mgr: LRO manager to use
+ * @skb: SKB to aggregate
+ * @priv: Private data that may be used by driver functions
+ *        (for example get_tcp_ip_hdr)
+ */
+
+void lro_receive_skb(struct net_lro_mgr *lro_mgr,
+                    struct sk_buff *skb,
+                    void *priv);
+
+/*
+ * Processes a SKB with VLAN HW acceleration support
+ */
+
+void lro_vlan_hwaccel_receive_skb(struct net_lro_mgr *lro_mgr,
+                                 struct sk_buff *skb,
+                                 struct vlan_group *vgrp,
+                                 u16 vlan_tag,
+                                 void *priv);
+
+/*
+ * Processes a fragment list
+ *
+ * This functions aggregate fragments and generate SKBs do pass
+ * the packets to the stack.
+ *
+ * @lro_mgr: LRO manager to use
+ * @frags: Fragment to be processed. Must contain entire header in first
+ *         element.
+ * @len: Length of received data
+ * @true_size: Actual size of memory the fragment is consuming
+ * @priv: Private data that may be used by driver functions
+ *        (for example get_tcp_ip_hdr)
+ */
+
+void lro_receive_frags(struct net_lro_mgr *lro_mgr,
+                      struct skb_frag_struct *frags,
+                      int len, int true_size, void *priv, __wsum sum);
+
+void lro_vlan_hwaccel_receive_frags(struct net_lro_mgr *lro_mgr,
+                                   struct skb_frag_struct *frags,
+                                   int len, int true_size,
+                                   struct vlan_group *vgrp,
+                                   u16 vlan_tag,
+                                   void *priv, __wsum sum);
+
+/*
+ * Forward all aggregated SKBs held by lro_mgr to network stack
+ */
+
+void lro_flush_all(struct net_lro_mgr *lro_mgr);
+
+void lro_flush_pkt(struct net_lro_mgr *lro_mgr,
+                  struct iphdr *iph, struct tcphdr *tcph);
+
+#endif
index fb7909774254f777636744bf086e8f152302e3e2..d894f616c3d63ce732073d90f185441f3411da75 100644 (file)
@@ -394,6 +394,14 @@ config INET_XFRM_MODE_BEET
 
          If unsure, say Y.
 
+config INET_LRO
+       tristate "Large Receive Offload (ipv4/tcp)"
+
+       ---help---
+         Support for Large Receive Offload (ipv4/tcp).
+
+         If unsure, say Y.
+
 config INET_DIAG
        tristate "INET: socket monitoring interface"
        default y
index fbf1674e0c2ab1e92cbd23df734b9240e6843841..a02c36d0a13e4581efd11ebbf9a77ab75f7f010f 100644 (file)
@@ -25,6 +25,7 @@ obj-$(CONFIG_INET_ESP) += esp4.o
 obj-$(CONFIG_INET_IPCOMP) += ipcomp.o
 obj-$(CONFIG_INET_XFRM_TUNNEL) += xfrm4_tunnel.o
 obj-$(CONFIG_INET_XFRM_MODE_BEET) += xfrm4_mode_beet.o
+obj-$(CONFIG_INET_LRO) += inet_lro.o
 obj-$(CONFIG_INET_TUNNEL) += tunnel4.o
 obj-$(CONFIG_INET_XFRM_MODE_TRANSPORT) += xfrm4_mode_transport.o
 obj-$(CONFIG_INET_XFRM_MODE_TUNNEL) += xfrm4_mode_tunnel.o
diff --git a/net/ipv4/inet_lro.c b/net/ipv4/inet_lro.c
new file mode 100644 (file)
index 0000000..20bc593
--- /dev/null
@@ -0,0 +1,600 @@
+/*
+ *  linux/net/ipv4/inet_lro.c
+ *
+ *  Large Receive Offload (ipv4 / tcp)
+ *
+ *  (C) Copyright IBM Corp. 2007
+ *
+ *  Authors:
+ *       Jan-Bernd Themann <themann@de.ibm.com>
+ *       Christoph Raisch <raisch@de.ibm.com>
+ *
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2, or (at your option)
+ * any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+
+#include <linux/module.h>
+#include <linux/if_vlan.h>
+#include <linux/inet_lro.h>
+
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Jan-Bernd Themann <themann@de.ibm.com>");
+MODULE_DESCRIPTION("Large Receive Offload (ipv4 / tcp)");
+
+#define TCP_HDR_LEN(tcph) (tcph->doff << 2)
+#define IP_HDR_LEN(iph) (iph->ihl << 2)
+#define TCP_PAYLOAD_LENGTH(iph, tcph) \
+       (ntohs(iph->tot_len) - IP_HDR_LEN(iph) - TCP_HDR_LEN(tcph))
+
+#define IPH_LEN_WO_OPTIONS 5
+#define TCPH_LEN_WO_OPTIONS 5
+#define TCPH_LEN_W_TIMESTAMP 8
+
+#define LRO_MAX_PG_HLEN 64
+
+#define LRO_INC_STATS(lro_mgr, attr) { lro_mgr->stats.attr++; }
+
+/*
+ * Basic tcp checks whether packet is suitable for LRO
+ */
+
+static int lro_tcp_ip_check(struct iphdr *iph, struct tcphdr *tcph,
+                           int len, struct net_lro_desc *lro_desc)
+{
+        /* check ip header: don't aggregate padded frames */
+       if (ntohs(iph->tot_len) != len)
+               return -1;
+
+       if (TCP_PAYLOAD_LENGTH(iph, tcph) == 0)
+               return -1;
+
+       if (iph->ihl != IPH_LEN_WO_OPTIONS)
+               return -1;
+
+       if (tcph->cwr || tcph->ece || tcph->urg || !tcph->ack
+           || tcph->rst || tcph->syn || tcph->fin)
+               return -1;
+
+       if (INET_ECN_is_ce(ipv4_get_dsfield(iph)))
+               return -1;
+
+       if (tcph->doff != TCPH_LEN_WO_OPTIONS
+           && tcph->doff != TCPH_LEN_W_TIMESTAMP)
+               return -1;
+
+       /* check tcp options (only timestamp allowed) */
+       if (tcph->doff == TCPH_LEN_W_TIMESTAMP) {
+               u32 *topt = (u32 *)(tcph + 1);
+
+               if (*topt != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
+                                  | (TCPOPT_TIMESTAMP << 8)
+                                  | TCPOLEN_TIMESTAMP))
+                       return -1;
+
+               /* timestamp should be in right order */
+               topt++;
+               if (lro_desc && after(ntohl(lro_desc->tcp_rcv_tsval),
+                                     ntohl(*topt)))
+                       return -1;
+
+               /* timestamp reply should not be zero */
+               topt++;
+               if (*topt == 0)
+                       return -1;
+       }
+
+       return 0;
+}
+
+static void lro_update_tcp_ip_header(struct net_lro_desc *lro_desc)
+{
+       struct iphdr *iph = lro_desc->iph;
+       struct tcphdr *tcph = lro_desc->tcph;
+       u32 *p;
+       __wsum tcp_hdr_csum;
+
+       tcph->ack_seq = lro_desc->tcp_ack;
+       tcph->window = lro_desc->tcp_window;
+
+       if (lro_desc->tcp_saw_tstamp) {
+               p = (u32 *)(tcph + 1);
+               *(p+2) = lro_desc->tcp_rcv_tsecr;
+       }
+
+       iph->tot_len = htons(lro_desc->ip_tot_len);
+
+       iph->check = 0;
+       iph->check = ip_fast_csum((u8 *)lro_desc->iph, iph->ihl);
+
+       tcph->check = 0;
+       tcp_hdr_csum = csum_partial((u8 *)tcph, TCP_HDR_LEN(tcph), 0);
+       lro_desc->data_csum = csum_add(lro_desc->data_csum, tcp_hdr_csum);
+       tcph->check = csum_tcpudp_magic(iph->saddr, iph->daddr,
+                                       lro_desc->ip_tot_len -
+                                       IP_HDR_LEN(iph), IPPROTO_TCP,
+                                       lro_desc->data_csum);
+}
+
+static __wsum lro_tcp_data_csum(struct iphdr *iph, struct tcphdr *tcph, int len)
+{
+       __wsum tcp_csum;
+       __wsum tcp_hdr_csum;
+       __wsum tcp_ps_hdr_csum;
+
+       tcp_csum = ~csum_unfold(tcph->check);
+       tcp_hdr_csum = csum_partial((u8 *)tcph, TCP_HDR_LEN(tcph), tcp_csum);
+
+       tcp_ps_hdr_csum = csum_tcpudp_nofold(iph->saddr, iph->daddr,
+                                            len + TCP_HDR_LEN(tcph),
+                                            IPPROTO_TCP, 0);
+
+       return csum_sub(csum_sub(tcp_csum, tcp_hdr_csum),
+                       tcp_ps_hdr_csum);
+}
+
+static void lro_init_desc(struct net_lro_desc *lro_desc, struct sk_buff *skb,
+                         struct iphdr *iph, struct tcphdr *tcph,
+                         u16 vlan_tag, struct vlan_group *vgrp)
+{
+       int nr_frags;
+       u32 *ptr;
+       u32 tcp_data_len = TCP_PAYLOAD_LENGTH(iph, tcph);
+
+       nr_frags = skb_shinfo(skb)->nr_frags;
+       lro_desc->parent = skb;
+       lro_desc->next_frag = &(skb_shinfo(skb)->frags[nr_frags]);
+       lro_desc->iph = iph;
+       lro_desc->tcph = tcph;
+       lro_desc->tcp_next_seq = ntohl(tcph->seq) + tcp_data_len;
+       lro_desc->tcp_ack = ntohl(tcph->ack_seq);
+       lro_desc->tcp_window = tcph->window;
+
+       lro_desc->pkt_aggr_cnt = 1;
+       lro_desc->ip_tot_len = ntohs(iph->tot_len);
+
+       if (tcph->doff == 8) {
+               ptr = (u32 *)(tcph+1);
+               lro_desc->tcp_saw_tstamp = 1;
+               lro_desc->tcp_rcv_tsval = *(ptr+1);
+               lro_desc->tcp_rcv_tsecr = *(ptr+2);
+       }
+
+       lro_desc->mss = tcp_data_len;
+       lro_desc->vgrp = vgrp;
+       lro_desc->vlan_tag = vlan_tag;
+       lro_desc->active = 1;
+
+       lro_desc->data_csum = lro_tcp_data_csum(iph, tcph,
+                                               tcp_data_len);
+}
+
+static inline void lro_clear_desc(struct net_lro_desc *lro_desc)
+{
+       memset(lro_desc, 0, sizeof(struct net_lro_desc));
+}
+
+static void lro_add_common(struct net_lro_desc *lro_desc, struct iphdr *iph,
+                          struct tcphdr *tcph, int tcp_data_len)
+{
+       struct sk_buff *parent = lro_desc->parent;
+       u32 *topt;
+
+       lro_desc->pkt_aggr_cnt++;
+       lro_desc->ip_tot_len += tcp_data_len;
+       lro_desc->tcp_next_seq += tcp_data_len;
+       lro_desc->tcp_window = tcph->window;
+       lro_desc->tcp_ack = tcph->ack_seq;
+
+       /* don't update tcp_rcv_tsval, would not work with PAWS */
+       if (lro_desc->tcp_saw_tstamp) {
+               topt = (u32 *) (tcph + 1);
+               lro_desc->tcp_rcv_tsecr = *(topt + 2);
+       }
+
+       lro_desc->data_csum = csum_block_add(lro_desc->data_csum,
+                                            lro_tcp_data_csum(iph, tcph,
+                                                              tcp_data_len),
+                                            parent->len);
+
+       parent->len += tcp_data_len;
+       parent->data_len += tcp_data_len;
+       if (tcp_data_len > lro_desc->mss)
+               lro_desc->mss = tcp_data_len;
+}
+
+static void lro_add_packet(struct net_lro_desc *lro_desc, struct sk_buff *skb,
+                          struct iphdr *iph, struct tcphdr *tcph)
+{
+       struct sk_buff *parent = lro_desc->parent;
+       int tcp_data_len = TCP_PAYLOAD_LENGTH(iph, tcph);
+
+       lro_add_common(lro_desc, iph, tcph, tcp_data_len);
+
+       skb_pull(skb, (skb->len - tcp_data_len));
+       parent->truesize += skb->truesize;
+
+       if (lro_desc->last_skb)
+               lro_desc->last_skb->next = skb;
+       else
+               skb_shinfo(parent)->frag_list = skb;
+
+       lro_desc->last_skb = skb;
+}
+
+static void lro_add_frags(struct net_lro_desc *lro_desc,
+                         int len, int hlen, int truesize,
+                         struct skb_frag_struct *skb_frags,
+                         struct iphdr *iph, struct tcphdr *tcph)
+{
+       struct sk_buff *skb = lro_desc->parent;
+       int tcp_data_len = TCP_PAYLOAD_LENGTH(iph, tcph);
+
+       lro_add_common(lro_desc, iph, tcph, tcp_data_len);
+
+       skb->truesize += truesize;
+
+       skb_frags[0].page_offset += hlen;
+       skb_frags[0].size -= hlen;
+
+       while (tcp_data_len > 0) {
+               *(lro_desc->next_frag) = *skb_frags;
+               tcp_data_len -= skb_frags->size;
+               lro_desc->next_frag++;
+               skb_frags++;
+               skb_shinfo(skb)->nr_frags++;
+       }
+}
+
+static int lro_check_tcp_conn(struct net_lro_desc *lro_desc,
+                             struct iphdr *iph,
+                             struct tcphdr *tcph)
+{
+       if ((lro_desc->iph->saddr != iph->saddr)
+           || (lro_desc->iph->daddr != iph->daddr)
+           || (lro_desc->tcph->source != tcph->source)
+           || (lro_desc->tcph->dest != tcph->dest))
+               return -1;
+       return 0;
+}
+
+static struct net_lro_desc *lro_get_desc(struct net_lro_mgr *lro_mgr,
+                                        struct net_lro_desc *lro_arr,
+                                        struct iphdr *iph,
+                                        struct tcphdr *tcph)
+{
+       struct net_lro_desc *lro_desc = NULL;
+       struct net_lro_desc *tmp;
+       int max_desc = lro_mgr->max_desc;
+       int i;
+
+       for (i = 0; i < max_desc; i++) {
+               tmp = &lro_arr[i];
+               if (tmp->active)
+                       if (!lro_check_tcp_conn(tmp, iph, tcph)) {
+                               lro_desc = tmp;
+                               goto out;
+                       }
+       }
+
+       for (i = 0; i < max_desc; i++) {
+               if (!lro_arr[i].active) {
+                       lro_desc = &lro_arr[i];
+                       goto out;
+               }
+       }
+
+       LRO_INC_STATS(lro_mgr, no_desc);
+out:
+       return lro_desc;
+}
+
+static void lro_flush(struct net_lro_mgr *lro_mgr,
+                     struct net_lro_desc *lro_desc)
+{
+       if (lro_desc->pkt_aggr_cnt > 1)
+               lro_update_tcp_ip_header(lro_desc);
+
+       skb_shinfo(lro_desc->parent)->gso_size = lro_desc->mss;
+
+       if (lro_desc->vgrp) {
+               if (test_bit(LRO_F_NAPI, &lro_mgr->features))
+                       vlan_hwaccel_receive_skb(lro_desc->parent,
+                                                lro_desc->vgrp,
+                                                lro_desc->vlan_tag);
+               else
+                       vlan_hwaccel_rx(lro_desc->parent,
+                                       lro_desc->vgrp,
+                                       lro_desc->vlan_tag);
+
+       } else {
+               if (test_bit(LRO_F_NAPI, &lro_mgr->features))
+                       netif_receive_skb(lro_desc->parent);
+               else
+                       netif_rx(lro_desc->parent);
+       }
+
+       LRO_INC_STATS(lro_mgr, flushed);
+       lro_clear_desc(lro_desc);
+}
+
+static int __lro_proc_skb(struct net_lro_mgr *lro_mgr, struct sk_buff *skb,
+                         struct vlan_group *vgrp, u16 vlan_tag, void *priv)
+{
+       struct net_lro_desc *lro_desc;
+       struct iphdr *iph;
+       struct tcphdr *tcph;
+       u64 flags;
+       int vlan_hdr_len = 0;
+
+       if (!lro_mgr->get_skb_header
+           || lro_mgr->get_skb_header(skb, (void *)&iph, (void *)&tcph,
+                                      &flags, priv))
+               goto out;
+
+       if (!(flags & LRO_IPV4) || !(flags & LRO_TCP))
+               goto out;
+
+       lro_desc = lro_get_desc(lro_mgr, lro_mgr->lro_arr, iph, tcph);
+       if (!lro_desc)
+               goto out;
+
+       if ((skb->protocol == htons(ETH_P_8021Q))
+           && !test_bit(LRO_F_EXTRACT_VLAN_ID, &lro_mgr->features))
+               vlan_hdr_len = VLAN_HLEN;
+
+       if (!lro_desc->active) { /* start new lro session */
+               if (lro_tcp_ip_check(iph, tcph, skb->len - vlan_hdr_len, NULL))
+                       goto out;
+
+               skb->ip_summed = lro_mgr->ip_summed_aggr;
+               lro_init_desc(lro_desc, skb, iph, tcph, vlan_tag, vgrp);
+               LRO_INC_STATS(lro_mgr, aggregated);
+               return 0;
+       }
+
+       if (lro_desc->tcp_next_seq != ntohl(tcph->seq))
+               goto out2;
+
+       if (lro_tcp_ip_check(iph, tcph, skb->len, lro_desc))
+               goto out2;
+
+       lro_add_packet(lro_desc, skb, iph, tcph);
+       LRO_INC_STATS(lro_mgr, aggregated);
+
+       if ((lro_desc->pkt_aggr_cnt >= lro_mgr->max_aggr) ||
+           lro_desc->parent->len > (0xFFFF - lro_mgr->dev->mtu))
+               lro_flush(lro_mgr, lro_desc);
+
+       return 0;
+
+out2: /* send aggregated SKBs to stack */
+       lro_flush(lro_mgr, lro_desc);
+
+out:  /* Original SKB has to be posted to stack */
+       skb->ip_summed = lro_mgr->ip_summed;
+       return 1;
+}
+
+
+static struct sk_buff *lro_gen_skb(struct net_lro_mgr *lro_mgr,
+                                  struct skb_frag_struct *frags,
+                                  int len, int true_size,
+                                  void *mac_hdr,
+                                  int hlen, __wsum sum,
+                                  u32 ip_summed)
+{
+       struct sk_buff *skb;
+       struct skb_frag_struct *skb_frags;
+       int data_len = len;
+       int hdr_len = min(len, hlen);
+
+       skb = netdev_alloc_skb(lro_mgr->dev, hlen);
+       if (!skb)
+               return NULL;
+
+       skb->len = len;
+       skb->data_len = len - hdr_len;
+       skb->truesize += true_size;
+       skb->tail += hdr_len;
+
+       memcpy(skb->data, mac_hdr, hdr_len);
+
+       skb_frags = skb_shinfo(skb)->frags;
+       while (data_len > 0) {
+               *skb_frags = *frags;
+               data_len -= frags->size;
+               skb_frags++;
+               frags++;
+               skb_shinfo(skb)->nr_frags++;
+       }
+
+       skb_shinfo(skb)->frags[0].page_offset += hdr_len;
+       skb_shinfo(skb)->frags[0].size -= hdr_len;
+
+       skb->ip_summed = ip_summed;
+       skb->csum = sum;
+       skb->protocol = eth_type_trans(skb, lro_mgr->dev);
+       return skb;
+}
+
+static struct sk_buff *__lro_proc_segment(struct net_lro_mgr *lro_mgr,
+                                         struct skb_frag_struct *frags,
+                                         int len, int true_size,
+                                         struct vlan_group *vgrp,
+                                         u16 vlan_tag, void *priv, __wsum sum)
+{
+       struct net_lro_desc *lro_desc;
+       struct iphdr *iph;
+       struct tcphdr *tcph;
+       struct sk_buff *skb;
+       u64 flags;
+       void *mac_hdr;
+       int mac_hdr_len;
+       int hdr_len = LRO_MAX_PG_HLEN;
+       int vlan_hdr_len = 0;
+
+       if (!lro_mgr->get_frag_header
+           || lro_mgr->get_frag_header(frags, (void *)&mac_hdr, (void *)&iph,
+                                       (void *)&tcph, &flags, priv)) {
+               mac_hdr = page_address(frags->page) + frags->page_offset;
+               goto out1;
+       }
+
+       if (!(flags & LRO_IPV4) || !(flags & LRO_TCP))
+               goto out1;
+
+       hdr_len = (int)((void *)(tcph) + TCP_HDR_LEN(tcph) - mac_hdr);
+       mac_hdr_len = (int)((void *)(iph) - mac_hdr);
+
+       lro_desc = lro_get_desc(lro_mgr, lro_mgr->lro_arr, iph, tcph);
+       if (!lro_desc)
+               goto out1;
+
+       if (!lro_desc->active) { /* start new lro session */
+               if (lro_tcp_ip_check(iph, tcph, len - mac_hdr_len, NULL))
+                       goto out1;
+
+               skb = lro_gen_skb(lro_mgr, frags, len, true_size, mac_hdr,
+                                 hdr_len, 0, lro_mgr->ip_summed_aggr);
+               if (!skb)
+                       goto out;
+
+               if ((skb->protocol == htons(ETH_P_8021Q))
+                   && !test_bit(LRO_F_EXTRACT_VLAN_ID, &lro_mgr->features))
+                       vlan_hdr_len = VLAN_HLEN;
+
+               iph = (void *)(skb->data + vlan_hdr_len);
+               tcph = (void *)((u8 *)skb->data + vlan_hdr_len
+                               + IP_HDR_LEN(iph));
+
+               lro_init_desc(lro_desc, skb, iph, tcph, 0, NULL);
+               LRO_INC_STATS(lro_mgr, aggregated);
+               return 0;
+       }
+
+       if (lro_desc->tcp_next_seq != ntohl(tcph->seq))
+               goto out2;
+
+       if (lro_tcp_ip_check(iph, tcph, len - mac_hdr_len, lro_desc))
+               goto out2;
+
+       lro_add_frags(lro_desc, len, hdr_len, true_size, frags, iph, tcph);
+       LRO_INC_STATS(lro_mgr, aggregated);
+
+       if ((skb_shinfo(lro_desc->parent)->nr_frags >= lro_mgr->max_aggr) ||
+           lro_desc->parent->len > (0xFFFF - lro_mgr->dev->mtu))
+               lro_flush(lro_mgr, lro_desc);
+
+       return NULL;
+
+out2: /* send aggregated packets to the stack */
+       lro_flush(lro_mgr, lro_desc);
+
+out1:  /* Original packet has to be posted to the stack */
+       skb = lro_gen_skb(lro_mgr, frags, len, true_size, mac_hdr,
+                         hdr_len, sum, lro_mgr->ip_summed);
+out:
+       return skb;
+}
+
+void lro_receive_skb(struct net_lro_mgr *lro_mgr,
+                    struct sk_buff *skb,
+                    void *priv)
+{
+       if (__lro_proc_skb(lro_mgr, skb, NULL, 0, priv)) {
+               if (test_bit(LRO_F_NAPI, &lro_mgr->features))
+                       netif_receive_skb(skb);
+               else
+                       netif_rx(skb);
+       }
+}
+EXPORT_SYMBOL(lro_receive_skb);
+
+void lro_vlan_hwaccel_receive_skb(struct net_lro_mgr *lro_mgr,
+                                 struct sk_buff *skb,
+                                 struct vlan_group *vgrp,
+                                 u16 vlan_tag,
+                                 void *priv)
+{
+       if (__lro_proc_skb(lro_mgr, skb, vgrp, vlan_tag, priv)) {
+               if (test_bit(LRO_F_NAPI, &lro_mgr->features))
+                       vlan_hwaccel_receive_skb(skb, vgrp, vlan_tag);
+               else
+                       vlan_hwaccel_rx(skb, vgrp, vlan_tag);
+       }
+}
+EXPORT_SYMBOL(lro_vlan_hwaccel_receive_skb);
+
+void lro_receive_frags(struct net_lro_mgr *lro_mgr,
+                      struct skb_frag_struct *frags,
+                      int len, int true_size, void *priv, __wsum sum)
+{
+       struct sk_buff *skb;
+
+       skb = __lro_proc_segment(lro_mgr, frags, len, true_size, NULL, 0,
+                                priv, sum);
+       if (!skb)
+               return;
+
+       if (test_bit(LRO_F_NAPI, &lro_mgr->features))
+               netif_receive_skb(skb);
+       else
+               netif_rx(skb);
+}
+EXPORT_SYMBOL(lro_receive_frags);
+
+void lro_vlan_hwaccel_receive_frags(struct net_lro_mgr *lro_mgr,
+                                   struct skb_frag_struct *frags,
+                                   int len, int true_size,
+                                   struct vlan_group *vgrp,
+                                   u16 vlan_tag, void *priv, __wsum sum)
+{
+       struct sk_buff *skb;
+
+       skb = __lro_proc_segment(lro_mgr, frags, len, true_size, vgrp,
+                                vlan_tag, priv, sum);
+       if (!skb)
+               return;
+
+       if (test_bit(LRO_F_NAPI, &lro_mgr->features))
+               vlan_hwaccel_receive_skb(skb, vgrp, vlan_tag);
+       else
+               vlan_hwaccel_rx(skb, vgrp, vlan_tag);
+}
+EXPORT_SYMBOL(lro_vlan_hwaccel_receive_frags);
+
+void lro_flush_all(struct net_lro_mgr *lro_mgr)
+{
+       int i;
+       struct net_lro_desc *lro_desc = lro_mgr->lro_arr;
+
+       for (i = 0; i < lro_mgr->max_desc; i++) {
+               if (lro_desc[i].active)
+                       lro_flush(lro_mgr, &lro_desc[i]);
+       }
+}
+EXPORT_SYMBOL(lro_flush_all);
+
+void lro_flush_pkt(struct net_lro_mgr *lro_mgr,
+                 struct iphdr *iph, struct tcphdr *tcph)
+{
+       struct net_lro_desc *lro_desc;
+
+       lro_desc = lro_get_desc(lro_mgr, lro_mgr->lro_arr, iph, tcph);
+       if (lro_desc->active)
+               lro_flush(lro_mgr, lro_desc);
+}
+EXPORT_SYMBOL(lro_flush_pkt);