Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / net / ethernet / chelsio / cxgb3 / cxgb3_offload.c
diff --git a/kernel/drivers/net/ethernet/chelsio/cxgb3/cxgb3_offload.c b/kernel/drivers/net/ethernet/chelsio/cxgb3/cxgb3_offload.c
new file mode 100644 (file)
index 0000000..b0cbb2b
--- /dev/null
@@ -0,0 +1,1427 @@
+/*
+ * Copyright (c) 2006-2008 Chelsio, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+
+#include <linux/list.h>
+#include <linux/slab.h>
+#include <net/neighbour.h>
+#include <linux/notifier.h>
+#include <linux/atomic.h>
+#include <linux/proc_fs.h>
+#include <linux/if_vlan.h>
+#include <net/netevent.h>
+#include <linux/highmem.h>
+#include <linux/vmalloc.h>
+#include <linux/export.h>
+
+#include "common.h"
+#include "regs.h"
+#include "cxgb3_ioctl.h"
+#include "cxgb3_ctl_defs.h"
+#include "cxgb3_defs.h"
+#include "l2t.h"
+#include "firmware_exports.h"
+#include "cxgb3_offload.h"
+
+static LIST_HEAD(client_list);
+static LIST_HEAD(ofld_dev_list);
+static DEFINE_MUTEX(cxgb3_db_lock);
+
+static DEFINE_RWLOCK(adapter_list_lock);
+static LIST_HEAD(adapter_list);
+
+static const unsigned int MAX_ATIDS = 64 * 1024;
+static const unsigned int ATID_BASE = 0x10000;
+
+static void cxgb_neigh_update(struct neighbour *neigh);
+static void cxgb_redirect(struct dst_entry *old, struct dst_entry *new,
+                         struct neighbour *neigh, const void *daddr);
+
+static inline int offload_activated(struct t3cdev *tdev)
+{
+       const struct adapter *adapter = tdev2adap(tdev);
+
+       return test_bit(OFFLOAD_DEVMAP_BIT, &adapter->open_device_map);
+}
+
+/**
+ *     cxgb3_register_client - register an offload client
+ *     @client: the client
+ *
+ *     Add the client to the client list,
+ *     and call backs the client for each activated offload device
+ */
+void cxgb3_register_client(struct cxgb3_client *client)
+{
+       struct t3cdev *tdev;
+
+       mutex_lock(&cxgb3_db_lock);
+       list_add_tail(&client->client_list, &client_list);
+
+       if (client->add) {
+               list_for_each_entry(tdev, &ofld_dev_list, ofld_dev_list) {
+                       if (offload_activated(tdev))
+                               client->add(tdev);
+               }
+       }
+       mutex_unlock(&cxgb3_db_lock);
+}
+
+EXPORT_SYMBOL(cxgb3_register_client);
+
+/**
+ *     cxgb3_unregister_client - unregister an offload client
+ *     @client: the client
+ *
+ *     Remove the client to the client list,
+ *     and call backs the client for each activated offload device.
+ */
+void cxgb3_unregister_client(struct cxgb3_client *client)
+{
+       struct t3cdev *tdev;
+
+       mutex_lock(&cxgb3_db_lock);
+       list_del(&client->client_list);
+
+       if (client->remove) {
+               list_for_each_entry(tdev, &ofld_dev_list, ofld_dev_list) {
+                       if (offload_activated(tdev))
+                               client->remove(tdev);
+               }
+       }
+       mutex_unlock(&cxgb3_db_lock);
+}
+
+EXPORT_SYMBOL(cxgb3_unregister_client);
+
+/**
+ *     cxgb3_add_clients - activate registered clients for an offload device
+ *     @tdev: the offload device
+ *
+ *     Call backs all registered clients once a offload device is activated
+ */
+void cxgb3_add_clients(struct t3cdev *tdev)
+{
+       struct cxgb3_client *client;
+
+       mutex_lock(&cxgb3_db_lock);
+       list_for_each_entry(client, &client_list, client_list) {
+               if (client->add)
+                       client->add(tdev);
+       }
+       mutex_unlock(&cxgb3_db_lock);
+}
+
+/**
+ *     cxgb3_remove_clients - deactivates registered clients
+ *                            for an offload device
+ *     @tdev: the offload device
+ *
+ *     Call backs all registered clients once a offload device is deactivated
+ */
+void cxgb3_remove_clients(struct t3cdev *tdev)
+{
+       struct cxgb3_client *client;
+
+       mutex_lock(&cxgb3_db_lock);
+       list_for_each_entry(client, &client_list, client_list) {
+               if (client->remove)
+                       client->remove(tdev);
+       }
+       mutex_unlock(&cxgb3_db_lock);
+}
+
+void cxgb3_event_notify(struct t3cdev *tdev, u32 event, u32 port)
+{
+       struct cxgb3_client *client;
+
+       mutex_lock(&cxgb3_db_lock);
+       list_for_each_entry(client, &client_list, client_list) {
+               if (client->event_handler)
+                       client->event_handler(tdev, event, port);
+       }
+       mutex_unlock(&cxgb3_db_lock);
+}
+
+static struct net_device *get_iff_from_mac(struct adapter *adapter,
+                                          const unsigned char *mac,
+                                          unsigned int vlan)
+{
+       int i;
+
+       for_each_port(adapter, i) {
+               struct net_device *dev = adapter->port[i];
+
+               if (ether_addr_equal(dev->dev_addr, mac)) {
+                       rcu_read_lock();
+                       if (vlan && vlan != VLAN_VID_MASK) {
+                               dev = __vlan_find_dev_deep_rcu(dev, htons(ETH_P_8021Q), vlan);
+                       } else if (netif_is_bond_slave(dev)) {
+                               struct net_device *upper_dev;
+
+                               while ((upper_dev =
+                                       netdev_master_upper_dev_get_rcu(dev)))
+                                       dev = upper_dev;
+                       }
+                       rcu_read_unlock();
+                       return dev;
+               }
+       }
+       return NULL;
+}
+
+static int cxgb_ulp_iscsi_ctl(struct adapter *adapter, unsigned int req,
+                             void *data)
+{
+       int i;
+       int ret = 0;
+       unsigned int val = 0;
+       struct ulp_iscsi_info *uiip = data;
+
+       switch (req) {
+       case ULP_ISCSI_GET_PARAMS:
+               uiip->pdev = adapter->pdev;
+               uiip->llimit = t3_read_reg(adapter, A_ULPRX_ISCSI_LLIMIT);
+               uiip->ulimit = t3_read_reg(adapter, A_ULPRX_ISCSI_ULIMIT);
+               uiip->tagmask = t3_read_reg(adapter, A_ULPRX_ISCSI_TAGMASK);
+
+               val = t3_read_reg(adapter, A_ULPRX_ISCSI_PSZ);
+               for (i = 0; i < 4; i++, val >>= 8)
+                       uiip->pgsz_factor[i] = val & 0xFF;
+
+               val = t3_read_reg(adapter, A_TP_PARA_REG7);
+               uiip->max_txsz =
+               uiip->max_rxsz = min((val >> S_PMMAXXFERLEN0)&M_PMMAXXFERLEN0,
+                                    (val >> S_PMMAXXFERLEN1)&M_PMMAXXFERLEN1);
+               /*
+                * On tx, the iscsi pdu has to be <= tx page size and has to
+                * fit into the Tx PM FIFO.
+                */
+               val = min(adapter->params.tp.tx_pg_size,
+                         t3_read_reg(adapter, A_PM1_TX_CFG) >> 17);
+               uiip->max_txsz = min(val, uiip->max_txsz);
+
+               /* set MaxRxData to 16224 */
+               val = t3_read_reg(adapter, A_TP_PARA_REG2);
+               if ((val >> S_MAXRXDATA) != 0x3f60) {
+                       val &= (M_RXCOALESCESIZE << S_RXCOALESCESIZE);
+                       val |= V_MAXRXDATA(0x3f60);
+                       pr_info("%s, iscsi set MaxRxData to 16224 (0x%x)\n",
+                               adapter->name, val);
+                       t3_write_reg(adapter, A_TP_PARA_REG2, val);
+               }
+
+               /*
+                * on rx, the iscsi pdu has to be < rx page size and the
+                * the max rx data length programmed in TP
+                */
+               val = min(adapter->params.tp.rx_pg_size,
+                         ((t3_read_reg(adapter, A_TP_PARA_REG2)) >>
+                               S_MAXRXDATA) & M_MAXRXDATA);
+               uiip->max_rxsz = min(val, uiip->max_rxsz);
+               break;
+       case ULP_ISCSI_SET_PARAMS:
+               t3_write_reg(adapter, A_ULPRX_ISCSI_TAGMASK, uiip->tagmask);
+               /* program the ddp page sizes */
+               for (i = 0; i < 4; i++)
+                       val |= (uiip->pgsz_factor[i] & 0xF) << (8 * i);
+               if (val && (val != t3_read_reg(adapter, A_ULPRX_ISCSI_PSZ))) {
+                       pr_info("%s, setting iscsi pgsz 0x%x, %u,%u,%u,%u\n",
+                               adapter->name, val, uiip->pgsz_factor[0],
+                               uiip->pgsz_factor[1], uiip->pgsz_factor[2],
+                               uiip->pgsz_factor[3]);
+                       t3_write_reg(adapter, A_ULPRX_ISCSI_PSZ, val);
+               }
+               break;
+       default:
+               ret = -EOPNOTSUPP;
+       }
+       return ret;
+}
+
+/* Response queue used for RDMA events. */
+#define ASYNC_NOTIF_RSPQ 0
+
+static int cxgb_rdma_ctl(struct adapter *adapter, unsigned int req, void *data)
+{
+       int ret = 0;
+
+       switch (req) {
+       case RDMA_GET_PARAMS: {
+               struct rdma_info *rdma = data;
+               struct pci_dev *pdev = adapter->pdev;
+
+               rdma->udbell_physbase = pci_resource_start(pdev, 2);
+               rdma->udbell_len = pci_resource_len(pdev, 2);
+               rdma->tpt_base =
+                       t3_read_reg(adapter, A_ULPTX_TPT_LLIMIT);
+               rdma->tpt_top = t3_read_reg(adapter, A_ULPTX_TPT_ULIMIT);
+               rdma->pbl_base =
+                       t3_read_reg(adapter, A_ULPTX_PBL_LLIMIT);
+               rdma->pbl_top = t3_read_reg(adapter, A_ULPTX_PBL_ULIMIT);
+               rdma->rqt_base = t3_read_reg(adapter, A_ULPRX_RQ_LLIMIT);
+               rdma->rqt_top = t3_read_reg(adapter, A_ULPRX_RQ_ULIMIT);
+               rdma->kdb_addr = adapter->regs + A_SG_KDOORBELL;
+               rdma->pdev = pdev;
+               break;
+       }
+       case RDMA_CQ_OP:{
+               unsigned long flags;
+               struct rdma_cq_op *rdma = data;
+
+               /* may be called in any context */
+               spin_lock_irqsave(&adapter->sge.reg_lock, flags);
+               ret = t3_sge_cqcntxt_op(adapter, rdma->id, rdma->op,
+                                       rdma->credits);
+               spin_unlock_irqrestore(&adapter->sge.reg_lock, flags);
+               break;
+       }
+       case RDMA_GET_MEM:{
+               struct ch_mem_range *t = data;
+               struct mc7 *mem;
+
+               if ((t->addr & 7) || (t->len & 7))
+                       return -EINVAL;
+               if (t->mem_id == MEM_CM)
+                       mem = &adapter->cm;
+               else if (t->mem_id == MEM_PMRX)
+                       mem = &adapter->pmrx;
+               else if (t->mem_id == MEM_PMTX)
+                       mem = &adapter->pmtx;
+               else
+                       return -EINVAL;
+
+               ret =
+                       t3_mc7_bd_read(mem, t->addr / 8, t->len / 8,
+                                       (u64 *) t->buf);
+               if (ret)
+                       return ret;
+               break;
+       }
+       case RDMA_CQ_SETUP:{
+               struct rdma_cq_setup *rdma = data;
+
+               spin_lock_irq(&adapter->sge.reg_lock);
+               ret =
+                       t3_sge_init_cqcntxt(adapter, rdma->id,
+                                       rdma->base_addr, rdma->size,
+                                       ASYNC_NOTIF_RSPQ,
+                                       rdma->ovfl_mode, rdma->credits,
+                                       rdma->credit_thres);
+               spin_unlock_irq(&adapter->sge.reg_lock);
+               break;
+       }
+       case RDMA_CQ_DISABLE:
+               spin_lock_irq(&adapter->sge.reg_lock);
+               ret = t3_sge_disable_cqcntxt(adapter, *(unsigned int *)data);
+               spin_unlock_irq(&adapter->sge.reg_lock);
+               break;
+       case RDMA_CTRL_QP_SETUP:{
+               struct rdma_ctrlqp_setup *rdma = data;
+
+               spin_lock_irq(&adapter->sge.reg_lock);
+               ret = t3_sge_init_ecntxt(adapter, FW_RI_SGEEC_START, 0,
+                                               SGE_CNTXT_RDMA,
+                                               ASYNC_NOTIF_RSPQ,
+                                               rdma->base_addr, rdma->size,
+                                               FW_RI_TID_START, 1, 0);
+               spin_unlock_irq(&adapter->sge.reg_lock);
+               break;
+       }
+       case RDMA_GET_MIB: {
+               spin_lock(&adapter->stats_lock);
+               t3_tp_get_mib_stats(adapter, (struct tp_mib_stats *)data);
+               spin_unlock(&adapter->stats_lock);
+               break;
+       }
+       default:
+               ret = -EOPNOTSUPP;
+       }
+       return ret;
+}
+
+static int cxgb_offload_ctl(struct t3cdev *tdev, unsigned int req, void *data)
+{
+       struct adapter *adapter = tdev2adap(tdev);
+       struct tid_range *tid;
+       struct mtutab *mtup;
+       struct iff_mac *iffmacp;
+       struct ddp_params *ddpp;
+       struct adap_ports *ports;
+       struct ofld_page_info *rx_page_info;
+       struct tp_params *tp = &adapter->params.tp;
+       int i;
+
+       switch (req) {
+       case GET_MAX_OUTSTANDING_WR:
+               *(unsigned int *)data = FW_WR_NUM;
+               break;
+       case GET_WR_LEN:
+               *(unsigned int *)data = WR_FLITS;
+               break;
+       case GET_TX_MAX_CHUNK:
+               *(unsigned int *)data = 1 << 20;        /* 1MB */
+               break;
+       case GET_TID_RANGE:
+               tid = data;
+               tid->num = t3_mc5_size(&adapter->mc5) -
+                   adapter->params.mc5.nroutes -
+                   adapter->params.mc5.nfilters - adapter->params.mc5.nservers;
+               tid->base = 0;
+               break;
+       case GET_STID_RANGE:
+               tid = data;
+               tid->num = adapter->params.mc5.nservers;
+               tid->base = t3_mc5_size(&adapter->mc5) - tid->num -
+                   adapter->params.mc5.nfilters - adapter->params.mc5.nroutes;
+               break;
+       case GET_L2T_CAPACITY:
+               *(unsigned int *)data = 2048;
+               break;
+       case GET_MTUS:
+               mtup = data;
+               mtup->size = NMTUS;
+               mtup->mtus = adapter->params.mtus;
+               break;
+       case GET_IFF_FROM_MAC:
+               iffmacp = data;
+               iffmacp->dev = get_iff_from_mac(adapter, iffmacp->mac_addr,
+                                               iffmacp->vlan_tag &
+                                               VLAN_VID_MASK);
+               break;
+       case GET_DDP_PARAMS:
+               ddpp = data;
+               ddpp->llimit = t3_read_reg(adapter, A_ULPRX_TDDP_LLIMIT);
+               ddpp->ulimit = t3_read_reg(adapter, A_ULPRX_TDDP_ULIMIT);
+               ddpp->tag_mask = t3_read_reg(adapter, A_ULPRX_TDDP_TAGMASK);
+               break;
+       case GET_PORTS:
+               ports = data;
+               ports->nports = adapter->params.nports;
+               for_each_port(adapter, i)
+                       ports->lldevs[i] = adapter->port[i];
+               break;
+       case ULP_ISCSI_GET_PARAMS:
+       case ULP_ISCSI_SET_PARAMS:
+               if (!offload_running(adapter))
+                       return -EAGAIN;
+               return cxgb_ulp_iscsi_ctl(adapter, req, data);
+       case RDMA_GET_PARAMS:
+       case RDMA_CQ_OP:
+       case RDMA_CQ_SETUP:
+       case RDMA_CQ_DISABLE:
+       case RDMA_CTRL_QP_SETUP:
+       case RDMA_GET_MEM:
+       case RDMA_GET_MIB:
+               if (!offload_running(adapter))
+                       return -EAGAIN;
+               return cxgb_rdma_ctl(adapter, req, data);
+       case GET_RX_PAGE_INFO:
+               rx_page_info = data;
+               rx_page_info->page_size = tp->rx_pg_size;
+               rx_page_info->num = tp->rx_num_pgs;
+               break;
+       case GET_ISCSI_IPV4ADDR: {
+               struct iscsi_ipv4addr *p = data;
+               struct port_info *pi = netdev_priv(p->dev);
+               p->ipv4addr = pi->iscsi_ipv4addr;
+               break;
+       }
+       case GET_EMBEDDED_INFO: {
+               struct ch_embedded_info *e = data;
+
+               spin_lock(&adapter->stats_lock);
+               t3_get_fw_version(adapter, &e->fw_vers);
+               t3_get_tp_version(adapter, &e->tp_vers);
+               spin_unlock(&adapter->stats_lock);
+               break;
+       }
+       default:
+               return -EOPNOTSUPP;
+       }
+       return 0;
+}
+
+/*
+ * Dummy handler for Rx offload packets in case we get an offload packet before
+ * proper processing is setup.  This complains and drops the packet as it isn't
+ * normal to get offload packets at this stage.
+ */
+static int rx_offload_blackhole(struct t3cdev *dev, struct sk_buff **skbs,
+                               int n)
+{
+       while (n--)
+               dev_kfree_skb_any(skbs[n]);
+       return 0;
+}
+
+static void dummy_neigh_update(struct t3cdev *dev, struct neighbour *neigh)
+{
+}
+
+void cxgb3_set_dummy_ops(struct t3cdev *dev)
+{
+       dev->recv = rx_offload_blackhole;
+       dev->neigh_update = dummy_neigh_update;
+}
+
+/*
+ * Free an active-open TID.
+ */
+void *cxgb3_free_atid(struct t3cdev *tdev, int atid)
+{
+       struct tid_info *t = &(T3C_DATA(tdev))->tid_maps;
+       union active_open_entry *p = atid2entry(t, atid);
+       void *ctx = p->t3c_tid.ctx;
+
+       spin_lock_bh(&t->atid_lock);
+       p->next = t->afree;
+       t->afree = p;
+       t->atids_in_use--;
+       spin_unlock_bh(&t->atid_lock);
+
+       return ctx;
+}
+
+EXPORT_SYMBOL(cxgb3_free_atid);
+
+/*
+ * Free a server TID and return it to the free pool.
+ */
+void cxgb3_free_stid(struct t3cdev *tdev, int stid)
+{
+       struct tid_info *t = &(T3C_DATA(tdev))->tid_maps;
+       union listen_entry *p = stid2entry(t, stid);
+
+       spin_lock_bh(&t->stid_lock);
+       p->next = t->sfree;
+       t->sfree = p;
+       t->stids_in_use--;
+       spin_unlock_bh(&t->stid_lock);
+}
+
+EXPORT_SYMBOL(cxgb3_free_stid);
+
+void cxgb3_insert_tid(struct t3cdev *tdev, struct cxgb3_client *client,
+                     void *ctx, unsigned int tid)
+{
+       struct tid_info *t = &(T3C_DATA(tdev))->tid_maps;
+
+       t->tid_tab[tid].client = client;
+       t->tid_tab[tid].ctx = ctx;
+       atomic_inc(&t->tids_in_use);
+}
+
+EXPORT_SYMBOL(cxgb3_insert_tid);
+
+/*
+ * Populate a TID_RELEASE WR.  The skb must be already propely sized.
+ */
+static inline void mk_tid_release(struct sk_buff *skb, unsigned int tid)
+{
+       struct cpl_tid_release *req;
+
+       skb->priority = CPL_PRIORITY_SETUP;
+       req = (struct cpl_tid_release *)__skb_put(skb, sizeof(*req));
+       req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_FORWARD));
+       OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_TID_RELEASE, tid));
+}
+
+static void t3_process_tid_release_list(struct work_struct *work)
+{
+       struct t3c_data *td = container_of(work, struct t3c_data,
+                                          tid_release_task);
+       struct sk_buff *skb;
+       struct t3cdev *tdev = td->dev;
+
+
+       spin_lock_bh(&td->tid_release_lock);
+       while (td->tid_release_list) {
+               struct t3c_tid_entry *p = td->tid_release_list;
+
+               td->tid_release_list = p->ctx;
+               spin_unlock_bh(&td->tid_release_lock);
+
+               skb = alloc_skb(sizeof(struct cpl_tid_release),
+                               GFP_KERNEL);
+               if (!skb)
+                       skb = td->nofail_skb;
+               if (!skb) {
+                       spin_lock_bh(&td->tid_release_lock);
+                       p->ctx = (void *)td->tid_release_list;
+                       td->tid_release_list = p;
+                       break;
+               }
+               mk_tid_release(skb, p - td->tid_maps.tid_tab);
+               cxgb3_ofld_send(tdev, skb);
+               p->ctx = NULL;
+               if (skb == td->nofail_skb)
+                       td->nofail_skb =
+                               alloc_skb(sizeof(struct cpl_tid_release),
+                                       GFP_KERNEL);
+               spin_lock_bh(&td->tid_release_lock);
+       }
+       td->release_list_incomplete = (td->tid_release_list == NULL) ? 0 : 1;
+       spin_unlock_bh(&td->tid_release_lock);
+
+       if (!td->nofail_skb)
+               td->nofail_skb =
+                       alloc_skb(sizeof(struct cpl_tid_release),
+                               GFP_KERNEL);
+}
+
+/* use ctx as a next pointer in the tid release list */
+void cxgb3_queue_tid_release(struct t3cdev *tdev, unsigned int tid)
+{
+       struct t3c_data *td = T3C_DATA(tdev);
+       struct t3c_tid_entry *p = &td->tid_maps.tid_tab[tid];
+
+       spin_lock_bh(&td->tid_release_lock);
+       p->ctx = (void *)td->tid_release_list;
+       p->client = NULL;
+       td->tid_release_list = p;
+       if (!p->ctx || td->release_list_incomplete)
+               schedule_work(&td->tid_release_task);
+       spin_unlock_bh(&td->tid_release_lock);
+}
+
+EXPORT_SYMBOL(cxgb3_queue_tid_release);
+
+/*
+ * Remove a tid from the TID table.  A client may defer processing its last
+ * CPL message if it is locked at the time it arrives, and while the message
+ * sits in the client's backlog the TID may be reused for another connection.
+ * To handle this we atomically switch the TID association if it still points
+ * to the original client context.
+ */
+void cxgb3_remove_tid(struct t3cdev *tdev, void *ctx, unsigned int tid)
+{
+       struct tid_info *t = &(T3C_DATA(tdev))->tid_maps;
+
+       BUG_ON(tid >= t->ntids);
+       if (tdev->type == T3A)
+               (void)cmpxchg(&t->tid_tab[tid].ctx, ctx, NULL);
+       else {
+               struct sk_buff *skb;
+
+               skb = alloc_skb(sizeof(struct cpl_tid_release), GFP_ATOMIC);
+               if (likely(skb)) {
+                       mk_tid_release(skb, tid);
+                       cxgb3_ofld_send(tdev, skb);
+                       t->tid_tab[tid].ctx = NULL;
+               } else
+                       cxgb3_queue_tid_release(tdev, tid);
+       }
+       atomic_dec(&t->tids_in_use);
+}
+
+EXPORT_SYMBOL(cxgb3_remove_tid);
+
+int cxgb3_alloc_atid(struct t3cdev *tdev, struct cxgb3_client *client,
+                    void *ctx)
+{
+       int atid = -1;
+       struct tid_info *t = &(T3C_DATA(tdev))->tid_maps;
+
+       spin_lock_bh(&t->atid_lock);
+       if (t->afree &&
+           t->atids_in_use + atomic_read(&t->tids_in_use) + MC5_MIN_TIDS <=
+           t->ntids) {
+               union active_open_entry *p = t->afree;
+
+               atid = (p - t->atid_tab) + t->atid_base;
+               t->afree = p->next;
+               p->t3c_tid.ctx = ctx;
+               p->t3c_tid.client = client;
+               t->atids_in_use++;
+       }
+       spin_unlock_bh(&t->atid_lock);
+       return atid;
+}
+
+EXPORT_SYMBOL(cxgb3_alloc_atid);
+
+int cxgb3_alloc_stid(struct t3cdev *tdev, struct cxgb3_client *client,
+                    void *ctx)
+{
+       int stid = -1;
+       struct tid_info *t = &(T3C_DATA(tdev))->tid_maps;
+
+       spin_lock_bh(&t->stid_lock);
+       if (t->sfree) {
+               union listen_entry *p = t->sfree;
+
+               stid = (p - t->stid_tab) + t->stid_base;
+               t->sfree = p->next;
+               p->t3c_tid.ctx = ctx;
+               p->t3c_tid.client = client;
+               t->stids_in_use++;
+       }
+       spin_unlock_bh(&t->stid_lock);
+       return stid;
+}
+
+EXPORT_SYMBOL(cxgb3_alloc_stid);
+
+/* Get the t3cdev associated with a net_device */
+struct t3cdev *dev2t3cdev(struct net_device *dev)
+{
+       const struct port_info *pi = netdev_priv(dev);
+
+       return (struct t3cdev *)pi->adapter;
+}
+
+EXPORT_SYMBOL(dev2t3cdev);
+
+static int do_smt_write_rpl(struct t3cdev *dev, struct sk_buff *skb)
+{
+       struct cpl_smt_write_rpl *rpl = cplhdr(skb);
+
+       if (rpl->status != CPL_ERR_NONE)
+               pr_err("Unexpected SMT_WRITE_RPL status %u for entry %u\n",
+                      rpl->status, GET_TID(rpl));
+
+       return CPL_RET_BUF_DONE;
+}
+
+static int do_l2t_write_rpl(struct t3cdev *dev, struct sk_buff *skb)
+{
+       struct cpl_l2t_write_rpl *rpl = cplhdr(skb);
+
+       if (rpl->status != CPL_ERR_NONE)
+               pr_err("Unexpected L2T_WRITE_RPL status %u for entry %u\n",
+                      rpl->status, GET_TID(rpl));
+
+       return CPL_RET_BUF_DONE;
+}
+
+static int do_rte_write_rpl(struct t3cdev *dev, struct sk_buff *skb)
+{
+       struct cpl_rte_write_rpl *rpl = cplhdr(skb);
+
+       if (rpl->status != CPL_ERR_NONE)
+               pr_err("Unexpected RTE_WRITE_RPL status %u for entry %u\n",
+                      rpl->status, GET_TID(rpl));
+
+       return CPL_RET_BUF_DONE;
+}
+
+static int do_act_open_rpl(struct t3cdev *dev, struct sk_buff *skb)
+{
+       struct cpl_act_open_rpl *rpl = cplhdr(skb);
+       unsigned int atid = G_TID(ntohl(rpl->atid));
+       struct t3c_tid_entry *t3c_tid;
+
+       t3c_tid = lookup_atid(&(T3C_DATA(dev))->tid_maps, atid);
+       if (t3c_tid && t3c_tid->ctx && t3c_tid->client &&
+           t3c_tid->client->handlers &&
+           t3c_tid->client->handlers[CPL_ACT_OPEN_RPL]) {
+               return t3c_tid->client->handlers[CPL_ACT_OPEN_RPL] (dev, skb,
+                                                                   t3c_tid->
+                                                                   ctx);
+       } else {
+               pr_err("%s: received clientless CPL command 0x%x\n",
+                      dev->name, CPL_ACT_OPEN_RPL);
+               return CPL_RET_BUF_DONE | CPL_RET_BAD_MSG;
+       }
+}
+
+static int do_stid_rpl(struct t3cdev *dev, struct sk_buff *skb)
+{
+       union opcode_tid *p = cplhdr(skb);
+       unsigned int stid = G_TID(ntohl(p->opcode_tid));
+       struct t3c_tid_entry *t3c_tid;
+
+       t3c_tid = lookup_stid(&(T3C_DATA(dev))->tid_maps, stid);
+       if (t3c_tid && t3c_tid->ctx && t3c_tid->client->handlers &&
+           t3c_tid->client->handlers[p->opcode]) {
+               return t3c_tid->client->handlers[p->opcode] (dev, skb,
+                                                            t3c_tid->ctx);
+       } else {
+               pr_err("%s: received clientless CPL command 0x%x\n",
+                      dev->name, p->opcode);
+               return CPL_RET_BUF_DONE | CPL_RET_BAD_MSG;
+       }
+}
+
+static int do_hwtid_rpl(struct t3cdev *dev, struct sk_buff *skb)
+{
+       union opcode_tid *p = cplhdr(skb);
+       unsigned int hwtid = G_TID(ntohl(p->opcode_tid));
+       struct t3c_tid_entry *t3c_tid;
+
+       t3c_tid = lookup_tid(&(T3C_DATA(dev))->tid_maps, hwtid);
+       if (t3c_tid && t3c_tid->ctx && t3c_tid->client->handlers &&
+           t3c_tid->client->handlers[p->opcode]) {
+               return t3c_tid->client->handlers[p->opcode]
+                   (dev, skb, t3c_tid->ctx);
+       } else {
+               pr_err("%s: received clientless CPL command 0x%x\n",
+                      dev->name, p->opcode);
+               return CPL_RET_BUF_DONE | CPL_RET_BAD_MSG;
+       }
+}
+
+static int do_cr(struct t3cdev *dev, struct sk_buff *skb)
+{
+       struct cpl_pass_accept_req *req = cplhdr(skb);
+       unsigned int stid = G_PASS_OPEN_TID(ntohl(req->tos_tid));
+       struct tid_info *t = &(T3C_DATA(dev))->tid_maps;
+       struct t3c_tid_entry *t3c_tid;
+       unsigned int tid = GET_TID(req);
+
+       if (unlikely(tid >= t->ntids)) {
+               printk("%s: passive open TID %u too large\n",
+                      dev->name, tid);
+               t3_fatal_err(tdev2adap(dev));
+               return CPL_RET_BUF_DONE;
+       }
+
+       t3c_tid = lookup_stid(t, stid);
+       if (t3c_tid && t3c_tid->ctx && t3c_tid->client->handlers &&
+           t3c_tid->client->handlers[CPL_PASS_ACCEPT_REQ]) {
+               return t3c_tid->client->handlers[CPL_PASS_ACCEPT_REQ]
+                   (dev, skb, t3c_tid->ctx);
+       } else {
+               pr_err("%s: received clientless CPL command 0x%x\n",
+                      dev->name, CPL_PASS_ACCEPT_REQ);
+               return CPL_RET_BUF_DONE | CPL_RET_BAD_MSG;
+       }
+}
+
+/*
+ * Returns an sk_buff for a reply CPL message of size len.  If the input
+ * sk_buff has no other users it is trimmed and reused, otherwise a new buffer
+ * is allocated.  The input skb must be of size at least len.  Note that this
+ * operation does not destroy the original skb data even if it decides to reuse
+ * the buffer.
+ */
+static struct sk_buff *cxgb3_get_cpl_reply_skb(struct sk_buff *skb, size_t len,
+                                              gfp_t gfp)
+{
+       if (likely(!skb_cloned(skb))) {
+               BUG_ON(skb->len < len);
+               __skb_trim(skb, len);
+               skb_get(skb);
+       } else {
+               skb = alloc_skb(len, gfp);
+               if (skb)
+                       __skb_put(skb, len);
+       }
+       return skb;
+}
+
+static int do_abort_req_rss(struct t3cdev *dev, struct sk_buff *skb)
+{
+       union opcode_tid *p = cplhdr(skb);
+       unsigned int hwtid = G_TID(ntohl(p->opcode_tid));
+       struct t3c_tid_entry *t3c_tid;
+
+       t3c_tid = lookup_tid(&(T3C_DATA(dev))->tid_maps, hwtid);
+       if (t3c_tid && t3c_tid->ctx && t3c_tid->client->handlers &&
+           t3c_tid->client->handlers[p->opcode]) {
+               return t3c_tid->client->handlers[p->opcode]
+                   (dev, skb, t3c_tid->ctx);
+       } else {
+               struct cpl_abort_req_rss *req = cplhdr(skb);
+               struct cpl_abort_rpl *rpl;
+               struct sk_buff *reply_skb;
+               unsigned int tid = GET_TID(req);
+               u8 cmd = req->status;
+
+               if (req->status == CPL_ERR_RTX_NEG_ADVICE ||
+                   req->status == CPL_ERR_PERSIST_NEG_ADVICE)
+                       goto out;
+
+               reply_skb = cxgb3_get_cpl_reply_skb(skb,
+                                                   sizeof(struct
+                                                          cpl_abort_rpl),
+                                                   GFP_ATOMIC);
+
+               if (!reply_skb) {
+                       printk("do_abort_req_rss: couldn't get skb!\n");
+                       goto out;
+               }
+               reply_skb->priority = CPL_PRIORITY_DATA;
+               __skb_put(reply_skb, sizeof(struct cpl_abort_rpl));
+               rpl = cplhdr(reply_skb);
+               rpl->wr.wr_hi =
+                   htonl(V_WR_OP(FW_WROPCODE_OFLD_HOST_ABORT_CON_RPL));
+               rpl->wr.wr_lo = htonl(V_WR_TID(tid));
+               OPCODE_TID(rpl) = htonl(MK_OPCODE_TID(CPL_ABORT_RPL, tid));
+               rpl->cmd = cmd;
+               cxgb3_ofld_send(dev, reply_skb);
+out:
+               return CPL_RET_BUF_DONE;
+       }
+}
+
+static int do_act_establish(struct t3cdev *dev, struct sk_buff *skb)
+{
+       struct cpl_act_establish *req = cplhdr(skb);
+       unsigned int atid = G_PASS_OPEN_TID(ntohl(req->tos_tid));
+       struct tid_info *t = &(T3C_DATA(dev))->tid_maps;
+       struct t3c_tid_entry *t3c_tid;
+       unsigned int tid = GET_TID(req);
+
+       if (unlikely(tid >= t->ntids)) {
+               printk("%s: active establish TID %u too large\n",
+                      dev->name, tid);
+               t3_fatal_err(tdev2adap(dev));
+               return CPL_RET_BUF_DONE;
+       }
+
+       t3c_tid = lookup_atid(t, atid);
+       if (t3c_tid && t3c_tid->ctx && t3c_tid->client->handlers &&
+           t3c_tid->client->handlers[CPL_ACT_ESTABLISH]) {
+               return t3c_tid->client->handlers[CPL_ACT_ESTABLISH]
+                   (dev, skb, t3c_tid->ctx);
+       } else {
+               pr_err("%s: received clientless CPL command 0x%x\n",
+                      dev->name, CPL_ACT_ESTABLISH);
+               return CPL_RET_BUF_DONE | CPL_RET_BAD_MSG;
+       }
+}
+
+static int do_trace(struct t3cdev *dev, struct sk_buff *skb)
+{
+       struct cpl_trace_pkt *p = cplhdr(skb);
+
+       skb->protocol = htons(0xffff);
+       skb->dev = dev->lldev;
+       skb_pull(skb, sizeof(*p));
+       skb_reset_mac_header(skb);
+       netif_receive_skb(skb);
+       return 0;
+}
+
+/*
+ * That skb would better have come from process_responses() where we abuse
+ * ->priority and ->csum to carry our data.  NB: if we get to per-arch
+ * ->csum, the things might get really interesting here.
+ */
+
+static inline u32 get_hwtid(struct sk_buff *skb)
+{
+       return ntohl((__force __be32)skb->priority) >> 8 & 0xfffff;
+}
+
+static inline u32 get_opcode(struct sk_buff *skb)
+{
+       return G_OPCODE(ntohl((__force __be32)skb->csum));
+}
+
+static int do_term(struct t3cdev *dev, struct sk_buff *skb)
+{
+       unsigned int hwtid = get_hwtid(skb);
+       unsigned int opcode = get_opcode(skb);
+       struct t3c_tid_entry *t3c_tid;
+
+       t3c_tid = lookup_tid(&(T3C_DATA(dev))->tid_maps, hwtid);
+       if (t3c_tid && t3c_tid->ctx && t3c_tid->client->handlers &&
+           t3c_tid->client->handlers[opcode]) {
+               return t3c_tid->client->handlers[opcode] (dev, skb,
+                                                         t3c_tid->ctx);
+       } else {
+               pr_err("%s: received clientless CPL command 0x%x\n",
+                      dev->name, opcode);
+               return CPL_RET_BUF_DONE | CPL_RET_BAD_MSG;
+       }
+}
+
+static int nb_callback(struct notifier_block *self, unsigned long event,
+                      void *ctx)
+{
+       switch (event) {
+       case (NETEVENT_NEIGH_UPDATE):{
+               cxgb_neigh_update((struct neighbour *)ctx);
+               break;
+       }
+       case (NETEVENT_REDIRECT):{
+               struct netevent_redirect *nr = ctx;
+               cxgb_redirect(nr->old, nr->new, nr->neigh,
+                             nr->daddr);
+               cxgb_neigh_update(nr->neigh);
+               break;
+       }
+       default:
+               break;
+       }
+       return 0;
+}
+
+static struct notifier_block nb = {
+       .notifier_call = nb_callback
+};
+
+/*
+ * Process a received packet with an unknown/unexpected CPL opcode.
+ */
+static int do_bad_cpl(struct t3cdev *dev, struct sk_buff *skb)
+{
+       pr_err("%s: received bad CPL command 0x%x\n", dev->name, *skb->data);
+       return CPL_RET_BUF_DONE | CPL_RET_BAD_MSG;
+}
+
+/*
+ * Handlers for each CPL opcode
+ */
+static cpl_handler_func cpl_handlers[NUM_CPL_CMDS];
+
+/*
+ * Add a new handler to the CPL dispatch table.  A NULL handler may be supplied
+ * to unregister an existing handler.
+ */
+void t3_register_cpl_handler(unsigned int opcode, cpl_handler_func h)
+{
+       if (opcode < NUM_CPL_CMDS)
+               cpl_handlers[opcode] = h ? h : do_bad_cpl;
+       else
+               pr_err("T3C: handler registration for opcode %x failed\n",
+                      opcode);
+}
+
+EXPORT_SYMBOL(t3_register_cpl_handler);
+
+/*
+ * T3CDEV's receive method.
+ */
+static int process_rx(struct t3cdev *dev, struct sk_buff **skbs, int n)
+{
+       while (n--) {
+               struct sk_buff *skb = *skbs++;
+               unsigned int opcode = get_opcode(skb);
+               int ret = cpl_handlers[opcode] (dev, skb);
+
+#if VALIDATE_TID
+               if (ret & CPL_RET_UNKNOWN_TID) {
+                       union opcode_tid *p = cplhdr(skb);
+
+                       pr_err("%s: CPL message (opcode %u) had unknown TID %u\n",
+                              dev->name, opcode, G_TID(ntohl(p->opcode_tid)));
+               }
+#endif
+               if (ret & CPL_RET_BUF_DONE)
+                       kfree_skb(skb);
+       }
+       return 0;
+}
+
+/*
+ * Sends an sk_buff to a T3C driver after dealing with any active network taps.
+ */
+int cxgb3_ofld_send(struct t3cdev *dev, struct sk_buff *skb)
+{
+       int r;
+
+       local_bh_disable();
+       r = dev->send(dev, skb);
+       local_bh_enable();
+       return r;
+}
+
+EXPORT_SYMBOL(cxgb3_ofld_send);
+
+static int is_offloading(struct net_device *dev)
+{
+       struct adapter *adapter;
+       int i;
+
+       read_lock_bh(&adapter_list_lock);
+       list_for_each_entry(adapter, &adapter_list, adapter_list) {
+               for_each_port(adapter, i) {
+                       if (dev == adapter->port[i]) {
+                               read_unlock_bh(&adapter_list_lock);
+                               return 1;
+                       }
+               }
+       }
+       read_unlock_bh(&adapter_list_lock);
+       return 0;
+}
+
+static void cxgb_neigh_update(struct neighbour *neigh)
+{
+       struct net_device *dev;
+
+       if (!neigh)
+               return;
+       dev = neigh->dev;
+       if (dev && (is_offloading(dev))) {
+               struct t3cdev *tdev = dev2t3cdev(dev);
+
+               BUG_ON(!tdev);
+               t3_l2t_update(tdev, neigh);
+       }
+}
+
+static void set_l2t_ix(struct t3cdev *tdev, u32 tid, struct l2t_entry *e)
+{
+       struct sk_buff *skb;
+       struct cpl_set_tcb_field *req;
+
+       skb = alloc_skb(sizeof(*req), GFP_ATOMIC);
+       if (!skb) {
+               pr_err("%s: cannot allocate skb!\n", __func__);
+               return;
+       }
+       skb->priority = CPL_PRIORITY_CONTROL;
+       req = (struct cpl_set_tcb_field *)skb_put(skb, sizeof(*req));
+       req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_FORWARD));
+       OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_SET_TCB_FIELD, tid));
+       req->reply = 0;
+       req->cpu_idx = 0;
+       req->word = htons(W_TCB_L2T_IX);
+       req->mask = cpu_to_be64(V_TCB_L2T_IX(M_TCB_L2T_IX));
+       req->val = cpu_to_be64(V_TCB_L2T_IX(e->idx));
+       tdev->send(tdev, skb);
+}
+
+static void cxgb_redirect(struct dst_entry *old, struct dst_entry *new,
+                         struct neighbour *neigh,
+                         const void *daddr)
+{
+       struct net_device *dev;
+       struct tid_info *ti;
+       struct t3cdev *tdev;
+       u32 tid;
+       int update_tcb;
+       struct l2t_entry *e;
+       struct t3c_tid_entry *te;
+
+       dev = neigh->dev;
+
+       if (!is_offloading(dev))
+               return;
+       tdev = dev2t3cdev(dev);
+       BUG_ON(!tdev);
+
+       /* Add new L2T entry */
+       e = t3_l2t_get(tdev, new, dev, daddr);
+       if (!e) {
+               pr_err("%s: couldn't allocate new l2t entry!\n", __func__);
+               return;
+       }
+
+       /* Walk tid table and notify clients of dst change. */
+       ti = &(T3C_DATA(tdev))->tid_maps;
+       for (tid = 0; tid < ti->ntids; tid++) {
+               te = lookup_tid(ti, tid);
+               BUG_ON(!te);
+               if (te && te->ctx && te->client && te->client->redirect) {
+                       update_tcb = te->client->redirect(te->ctx, old, new, e);
+                       if (update_tcb) {
+                               rcu_read_lock();
+                               l2t_hold(L2DATA(tdev), e);
+                               rcu_read_unlock();
+                               set_l2t_ix(tdev, tid, e);
+                       }
+               }
+       }
+       l2t_release(tdev, e);
+}
+
+/*
+ * Allocate a chunk of memory using kmalloc or, if that fails, vmalloc.
+ * The allocated memory is cleared.
+ */
+void *cxgb_alloc_mem(unsigned long size)
+{
+       void *p = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
+
+       if (!p)
+               p = vzalloc(size);
+       return p;
+}
+
+/*
+ * Free memory allocated through t3_alloc_mem().
+ */
+void cxgb_free_mem(void *addr)
+{
+       if (is_vmalloc_addr(addr))
+               vfree(addr);
+       else
+               kfree(addr);
+}
+
+/*
+ * Allocate and initialize the TID tables.  Returns 0 on success.
+ */
+static int init_tid_tabs(struct tid_info *t, unsigned int ntids,
+                        unsigned int natids, unsigned int nstids,
+                        unsigned int atid_base, unsigned int stid_base)
+{
+       unsigned long size = ntids * sizeof(*t->tid_tab) +
+           natids * sizeof(*t->atid_tab) + nstids * sizeof(*t->stid_tab);
+
+       t->tid_tab = cxgb_alloc_mem(size);
+       if (!t->tid_tab)
+               return -ENOMEM;
+
+       t->stid_tab = (union listen_entry *)&t->tid_tab[ntids];
+       t->atid_tab = (union active_open_entry *)&t->stid_tab[nstids];
+       t->ntids = ntids;
+       t->nstids = nstids;
+       t->stid_base = stid_base;
+       t->sfree = NULL;
+       t->natids = natids;
+       t->atid_base = atid_base;
+       t->afree = NULL;
+       t->stids_in_use = t->atids_in_use = 0;
+       atomic_set(&t->tids_in_use, 0);
+       spin_lock_init(&t->stid_lock);
+       spin_lock_init(&t->atid_lock);
+
+       /*
+        * Setup the free lists for stid_tab and atid_tab.
+        */
+       if (nstids) {
+               while (--nstids)
+                       t->stid_tab[nstids - 1].next = &t->stid_tab[nstids];
+               t->sfree = t->stid_tab;
+       }
+       if (natids) {
+               while (--natids)
+                       t->atid_tab[natids - 1].next = &t->atid_tab[natids];
+               t->afree = t->atid_tab;
+       }
+       return 0;
+}
+
+static void free_tid_maps(struct tid_info *t)
+{
+       cxgb_free_mem(t->tid_tab);
+}
+
+static inline void add_adapter(struct adapter *adap)
+{
+       write_lock_bh(&adapter_list_lock);
+       list_add_tail(&adap->adapter_list, &adapter_list);
+       write_unlock_bh(&adapter_list_lock);
+}
+
+static inline void remove_adapter(struct adapter *adap)
+{
+       write_lock_bh(&adapter_list_lock);
+       list_del(&adap->adapter_list);
+       write_unlock_bh(&adapter_list_lock);
+}
+
+int cxgb3_offload_activate(struct adapter *adapter)
+{
+       struct t3cdev *dev = &adapter->tdev;
+       int natids, err;
+       struct t3c_data *t;
+       struct tid_range stid_range, tid_range;
+       struct mtutab mtutab;
+       unsigned int l2t_capacity;
+       struct l2t_data *l2td;
+
+       t = kzalloc(sizeof(*t), GFP_KERNEL);
+       if (!t)
+               return -ENOMEM;
+
+       err = -EOPNOTSUPP;
+       if (dev->ctl(dev, GET_TX_MAX_CHUNK, &t->tx_max_chunk) < 0 ||
+           dev->ctl(dev, GET_MAX_OUTSTANDING_WR, &t->max_wrs) < 0 ||
+           dev->ctl(dev, GET_L2T_CAPACITY, &l2t_capacity) < 0 ||
+           dev->ctl(dev, GET_MTUS, &mtutab) < 0 ||
+           dev->ctl(dev, GET_TID_RANGE, &tid_range) < 0 ||
+           dev->ctl(dev, GET_STID_RANGE, &stid_range) < 0)
+               goto out_free;
+
+       err = -ENOMEM;
+       l2td = t3_init_l2t(l2t_capacity);
+       if (!l2td)
+               goto out_free;
+
+       natids = min(tid_range.num / 2, MAX_ATIDS);
+       err = init_tid_tabs(&t->tid_maps, tid_range.num, natids,
+                           stid_range.num, ATID_BASE, stid_range.base);
+       if (err)
+               goto out_free_l2t;
+
+       t->mtus = mtutab.mtus;
+       t->nmtus = mtutab.size;
+
+       INIT_WORK(&t->tid_release_task, t3_process_tid_release_list);
+       spin_lock_init(&t->tid_release_lock);
+       INIT_LIST_HEAD(&t->list_node);
+       t->dev = dev;
+
+       RCU_INIT_POINTER(dev->l2opt, l2td);
+       T3C_DATA(dev) = t;
+       dev->recv = process_rx;
+       dev->neigh_update = t3_l2t_update;
+
+       /* Register netevent handler once */
+       if (list_empty(&adapter_list))
+               register_netevent_notifier(&nb);
+
+       t->nofail_skb = alloc_skb(sizeof(struct cpl_tid_release), GFP_KERNEL);
+       t->release_list_incomplete = 0;
+
+       add_adapter(adapter);
+       return 0;
+
+out_free_l2t:
+       t3_free_l2t(l2td);
+out_free:
+       kfree(t);
+       return err;
+}
+
+static void clean_l2_data(struct rcu_head *head)
+{
+       struct l2t_data *d = container_of(head, struct l2t_data, rcu_head);
+       t3_free_l2t(d);
+}
+
+
+void cxgb3_offload_deactivate(struct adapter *adapter)
+{
+       struct t3cdev *tdev = &adapter->tdev;
+       struct t3c_data *t = T3C_DATA(tdev);
+       struct l2t_data *d;
+
+       remove_adapter(adapter);
+       if (list_empty(&adapter_list))
+               unregister_netevent_notifier(&nb);
+
+       free_tid_maps(&t->tid_maps);
+       T3C_DATA(tdev) = NULL;
+       rcu_read_lock();
+       d = L2DATA(tdev);
+       rcu_read_unlock();
+       RCU_INIT_POINTER(tdev->l2opt, NULL);
+       call_rcu(&d->rcu_head, clean_l2_data);
+       if (t->nofail_skb)
+               kfree_skb(t->nofail_skb);
+       kfree(t);
+}
+
+static inline void register_tdev(struct t3cdev *tdev)
+{
+       static int unit;
+
+       mutex_lock(&cxgb3_db_lock);
+       snprintf(tdev->name, sizeof(tdev->name), "ofld_dev%d", unit++);
+       list_add_tail(&tdev->ofld_dev_list, &ofld_dev_list);
+       mutex_unlock(&cxgb3_db_lock);
+}
+
+static inline void unregister_tdev(struct t3cdev *tdev)
+{
+       mutex_lock(&cxgb3_db_lock);
+       list_del(&tdev->ofld_dev_list);
+       mutex_unlock(&cxgb3_db_lock);
+}
+
+static inline int adap2type(struct adapter *adapter)
+{
+       int type = 0;
+
+       switch (adapter->params.rev) {
+       case T3_REV_A:
+               type = T3A;
+               break;
+       case T3_REV_B:
+       case T3_REV_B2:
+               type = T3B;
+               break;
+       case T3_REV_C:
+               type = T3C;
+               break;
+       }
+       return type;
+}
+
+void cxgb3_adapter_ofld(struct adapter *adapter)
+{
+       struct t3cdev *tdev = &adapter->tdev;
+
+       INIT_LIST_HEAD(&tdev->ofld_dev_list);
+
+       cxgb3_set_dummy_ops(tdev);
+       tdev->send = t3_offload_tx;
+       tdev->ctl = cxgb_offload_ctl;
+       tdev->type = adap2type(adapter);
+
+       register_tdev(tdev);
+}
+
+void cxgb3_adapter_unofld(struct adapter *adapter)
+{
+       struct t3cdev *tdev = &adapter->tdev;
+
+       tdev->recv = NULL;
+       tdev->neigh_update = NULL;
+
+       unregister_tdev(tdev);
+}
+
+void __init cxgb3_offload_init(void)
+{
+       int i;
+
+       for (i = 0; i < NUM_CPL_CMDS; ++i)
+               cpl_handlers[i] = do_bad_cpl;
+
+       t3_register_cpl_handler(CPL_SMT_WRITE_RPL, do_smt_write_rpl);
+       t3_register_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl);
+       t3_register_cpl_handler(CPL_RTE_WRITE_RPL, do_rte_write_rpl);
+       t3_register_cpl_handler(CPL_PASS_OPEN_RPL, do_stid_rpl);
+       t3_register_cpl_handler(CPL_CLOSE_LISTSRV_RPL, do_stid_rpl);
+       t3_register_cpl_handler(CPL_PASS_ACCEPT_REQ, do_cr);
+       t3_register_cpl_handler(CPL_PASS_ESTABLISH, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_ABORT_RPL_RSS, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_ABORT_RPL, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_RX_URG_NOTIFY, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_RX_DATA, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_TX_DATA_ACK, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_TX_DMA_ACK, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_ACT_OPEN_RPL, do_act_open_rpl);
+       t3_register_cpl_handler(CPL_PEER_CLOSE, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_CLOSE_CON_RPL, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req_rss);
+       t3_register_cpl_handler(CPL_ACT_ESTABLISH, do_act_establish);
+       t3_register_cpl_handler(CPL_SET_TCB_RPL, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_GET_TCB_RPL, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_RDMA_TERMINATE, do_term);
+       t3_register_cpl_handler(CPL_RDMA_EC_STATUS, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_TRACE_PKT, do_trace);
+       t3_register_cpl_handler(CPL_RX_DATA_DDP, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_RX_DDP_COMPLETE, do_hwtid_rpl);
+       t3_register_cpl_handler(CPL_ISCSI_HDR, do_hwtid_rpl);
+}