drivers/infiniband/ulp/ipoib/ipoib_cm.c

   1 /*
   2  * Copyright (c) 2006 Mellanox Technologies. All rights reserved
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  * $Id$
  33  */
  34
  35 #include <rdma/ib_cm.h>
  36 #include <rdma/ib_cache.h>
  37 #include <net/dst.h>
  38 #include <net/icmp.h>
  39 #include <linux/icmpv6.h>
  40
  41 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  42 static int data_debug_level;
  43
  44 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  45 MODULE_PARM_DESC(cm_data_debug_level,
  46                  "Enable data path debug tracing for connected mode if > 0");
  47 #endif
  48
  49 #include "ipoib.h"
  50
  51 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  52
  53 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  54 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  55 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  56 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  57
  58 struct ipoib_cm_id {
  59         struct ib_cm_id *id;
  60         int flags;
  61         u32 remote_qpn;
  62         u32 remote_mtu;
  63 };
  64
  65 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  66                                struct ib_cm_event *event);
  67
  68 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  69                                   u64 mapping[IPOIB_CM_RX_SG])
  70 {
  71         int i;
  72
  73         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  74
  75         for (i = 0; i < frags; ++i)
  76                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  77 }
  78
  79 static int ipoib_cm_post_receive(struct net_device *dev, int id)
  80 {
  81         struct ipoib_dev_priv *priv = netdev_priv(dev);
  82         struct ib_recv_wr *bad_wr;
  83         int i, ret;
  84
  85         priv->cm.rx_wr.wr_id = id | IPOIB_CM_OP_SRQ;
  86
  87         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
  88                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
  89
  90         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
  91         if (unlikely(ret)) {
  92                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
  93                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
  94                                       priv->cm.srq_ring[id].mapping);
  95                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
  96                 priv->cm.srq_ring[id].skb = NULL;
  97         }
  98
  99         return ret;
 100 }
 101
 102 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev, int id, int frags,
 103                                              u64 mapping[IPOIB_CM_RX_SG])
 104 {
 105         struct ipoib_dev_priv *priv = netdev_priv(dev);
 106         struct sk_buff *skb;
 107         int i;
 108
 109         skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
 110         if (unlikely(!skb))
 111                 return NULL;
 112
 113         /*
 114          * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
 115          * IP header to a multiple of 16.
 116          */
 117         skb_reserve(skb, 12);
 118
 119         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 120                                        DMA_FROM_DEVICE);
 121         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 122                 dev_kfree_skb_any(skb);
 123                 return NULL;
 124         }
 125
 126         for (i = 0; i < frags; i++) {
 127                 struct page *page = alloc_page(GFP_ATOMIC);
 128
 129                 if (!page)
 130                         goto partial_error;
 131                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 132
 133                 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
 134                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 135                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 136                         goto partial_error;
 137         }
 138
 139         priv->cm.srq_ring[id].skb = skb;
 140         return skb;
 141
 142 partial_error:
 143
 144         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 145
 146         for (; i >= 0; --i)
 147                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
 148
 149         dev_kfree_skb_any(skb);
 150         return NULL;
 151 }
 152
 153 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 154                                            struct ipoib_cm_rx *p)
 155 {
 156         struct ipoib_dev_priv *priv = netdev_priv(dev);
 157         struct ib_qp_init_attr attr = {
 158                 .send_cq = priv->cq, /* does not matter, we never send anything */
 159                 .recv_cq = priv->cq,
 160                 .srq = priv->cm.srq,
 161                 .cap.max_send_wr = 1, /* FIXME: 0 Seems not to work */
 162                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 163                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 164                 .qp_type = IB_QPT_RC,
 165                 .qp_context = p,
 166         };
 167         return ib_create_qp(priv->pd, &attr);
 168 }
 169
 170 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 171                                   struct ib_cm_id *cm_id, struct ib_qp *qp,
 172                                   unsigned psn)
 173 {
 174         struct ipoib_dev_priv *priv = netdev_priv(dev);
 175         struct ib_qp_attr qp_attr;
 176         int qp_attr_mask, ret;
 177
 178         qp_attr.qp_state = IB_QPS_INIT;
 179         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 180         if (ret) {
 181                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 182                 return ret;
 183         }
 184         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 185         if (ret) {
 186                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 187                 return ret;
 188         }
 189         qp_attr.qp_state = IB_QPS_RTR;
 190         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 191         if (ret) {
 192                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 193                 return ret;
 194         }
 195         qp_attr.rq_psn = psn;
 196         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 197         if (ret) {
 198                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 199                 return ret;
 200         }
 201         return 0;
 202 }
 203
 204 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 205                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 206                              unsigned psn)
 207 {
 208         struct ipoib_dev_priv *priv = netdev_priv(dev);
 209         struct ipoib_cm_data data = {};
 210         struct ib_cm_rep_param rep = {};
 211
 212         data.qpn = cpu_to_be32(priv->qp->qp_num);
 213         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 214
 215         rep.private_data = &data;
 216         rep.private_data_len = sizeof data;
 217         rep.flow_control = 0;
 218         rep.rnr_retry_count = req->rnr_retry_count;
 219         rep.target_ack_delay = 20; /* FIXME */
 220         rep.srq = 1;
 221         rep.qp_num = qp->qp_num;
 222         rep.starting_psn = psn;
 223         return ib_send_cm_rep(cm_id, &rep);
 224 }
 225
 226 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 227 {
 228         struct net_device *dev = cm_id->context;
 229         struct ipoib_dev_priv *priv = netdev_priv(dev);
 230         struct ipoib_cm_rx *p;
 231         unsigned psn;
 232         int ret;
 233
 234         ipoib_dbg(priv, "REQ arrived\n");
 235         p = kzalloc(sizeof *p, GFP_KERNEL);
 236         if (!p)
 237                 return -ENOMEM;
 238         p->dev = dev;
 239         p->id = cm_id;
 240         p->qp = ipoib_cm_create_rx_qp(dev, p);
 241         if (IS_ERR(p->qp)) {
 242                 ret = PTR_ERR(p->qp);
 243                 goto err_qp;
 244         }
 245
 246         psn = random32() & 0xffffff;
 247         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 248         if (ret)
 249                 goto err_modify;
 250
 251         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 252         if (ret) {
 253                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 254                 goto err_rep;
 255         }
 256
 257         cm_id->context = p;
 258         p->jiffies = jiffies;
 259         spin_lock_irq(&priv->lock);
 260         list_add(&p->list, &priv->cm.passive_ids);
 261         spin_unlock_irq(&priv->lock);
 262         queue_delayed_work(ipoib_workqueue,
 263                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 264         return 0;
 265
 266 err_rep:
 267 err_modify:
 268         ib_destroy_qp(p->qp);
 269 err_qp:
 270         kfree(p);
 271         return ret;
 272 }
 273
 274 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 275                                struct ib_cm_event *event)
 276 {
 277         struct ipoib_cm_rx *p;
 278         struct ipoib_dev_priv *priv;
 279         int ret;
 280
 281         switch (event->event) {
 282         case IB_CM_REQ_RECEIVED:
 283                 return ipoib_cm_req_handler(cm_id, event);
 284         case IB_CM_DREQ_RECEIVED:
 285                 p = cm_id->context;
 286                 ib_send_cm_drep(cm_id, NULL, 0);
 287                 /* Fall through */
 288         case IB_CM_REJ_RECEIVED:
 289                 p = cm_id->context;
 290                 priv = netdev_priv(p->dev);
 291                 spin_lock_irq(&priv->lock);
 292                 if (list_empty(&p->list))
 293                         ret = 0; /* Connection is going away already. */
 294                 else {
 295                         list_del_init(&p->list);
 296                         ret = -ECONNRESET;
 297                 }
 298                 spin_unlock_irq(&priv->lock);
 299                 if (ret) {
 300                         ib_destroy_qp(p->qp);
 301                         kfree(p);
 302                         return ret;
 303                 }
 304                 return 0;
 305         default:
 306                 return 0;
 307         }
 308 }
 309 /* Adjust length of skb with fragments to match received data */
 310 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 311                           unsigned int length, struct sk_buff *toskb)
 312 {
 313         int i, num_frags;
 314         unsigned int size;
 315
 316         /* put header into skb */
 317         size = min(length, hdr_space);
 318         skb->tail += size;
 319         skb->len += size;
 320         length -= size;
 321
 322         num_frags = skb_shinfo(skb)->nr_frags;
 323         for (i = 0; i < num_frags; i++) {
 324                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 325
 326                 if (length == 0) {
 327                         /* don't need this page */
 328                         skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
 329                         --skb_shinfo(skb)->nr_frags;
 330                 } else {
 331                         size = min(length, (unsigned) PAGE_SIZE);
 332
 333                         frag->size = size;
 334                         skb->data_len += size;
 335                         skb->truesize += size;
 336                         skb->len += size;
 337                         length -= size;
 338                 }
 339         }
 340 }
 341
 342 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 343 {
 344         struct ipoib_dev_priv *priv = netdev_priv(dev);
 345         unsigned int wr_id = wc->wr_id & ~IPOIB_CM_OP_SRQ;
 346         struct sk_buff *skb, *newskb;
 347         struct ipoib_cm_rx *p;
 348         unsigned long flags;
 349         u64 mapping[IPOIB_CM_RX_SG];
 350         int frags;
 351
 352         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 353                        wr_id, wc->status);
 354
 355         if (unlikely(wr_id >= ipoib_recvq_size)) {
 356                 ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 357                            wr_id, ipoib_recvq_size);
 358                 return;
 359         }
 360
 361         skb  = priv->cm.srq_ring[wr_id].skb;
 362
 363         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 364                 ipoib_dbg(priv, "cm recv error "
 365                            "(status=%d, wrid=%d vend_err %x)\n",
 366                            wc->status, wr_id, wc->vendor_err);
 367                 ++priv->stats.rx_dropped;
 368                 goto repost;
 369         }
 370
 371         if (!likely(wr_id & IPOIB_CM_RX_UPDATE_MASK)) {
 372                 p = wc->qp->qp_context;
 373                 if (time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 374                         spin_lock_irqsave(&priv->lock, flags);
 375                         p->jiffies = jiffies;
 376                         /* Move this entry to list head, but do
 377                          * not re-add it if it has been removed. */
 378                         if (!list_empty(&p->list))
 379                                 list_move(&p->list, &priv->cm.passive_ids);
 380                         spin_unlock_irqrestore(&priv->lock, flags);
 381                         queue_delayed_work(ipoib_workqueue,
 382                                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 383                 }
 384         }
 385
 386         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 387                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 388
 389         newskb = ipoib_cm_alloc_rx_skb(dev, wr_id, frags, mapping);
 390         if (unlikely(!newskb)) {
 391                 /*
 392                  * If we can't allocate a new RX buffer, dump
 393                  * this packet and reuse the old buffer.
 394                  */
 395                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 396                 ++priv->stats.rx_dropped;
 397                 goto repost;
 398         }
 399
 400         ipoib_cm_dma_unmap_rx(priv, frags, priv->cm.srq_ring[wr_id].mapping);
 401         memcpy(priv->cm.srq_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 402
 403         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 404                        wc->byte_len, wc->slid);
 405
 406         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 407
 408         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 409         skb_reset_mac_header(skb);
 410         skb_pull(skb, IPOIB_ENCAP_LEN);
 411
 412         dev->last_rx = jiffies;
 413         ++priv->stats.rx_packets;
 414         priv->stats.rx_bytes += skb->len;
 415
 416         skb->dev = dev;
 417         /* XXX get correct PACKET_ type here */
 418         skb->pkt_type = PACKET_HOST;
 419         netif_rx_ni(skb);
 420
 421 repost:
 422         if (unlikely(ipoib_cm_post_receive(dev, wr_id)))
 423                 ipoib_warn(priv, "ipoib_cm_post_receive failed "
 424                            "for buf %d\n", wr_id);
 425 }
 426
 427 static inline int post_send(struct ipoib_dev_priv *priv,
 428                             struct ipoib_cm_tx *tx,
 429                             unsigned int wr_id,
 430                             u64 addr, int len)
 431 {
 432         struct ib_send_wr *bad_wr;
 433
 434         priv->tx_sge.addr             = addr;
 435         priv->tx_sge.length           = len;
 436
 437         priv->tx_wr.wr_id             = wr_id;
 438
 439         return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
 440 }
 441
 442 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 443 {
 444         struct ipoib_dev_priv *priv = netdev_priv(dev);
 445         struct ipoib_tx_buf *tx_req;
 446         u64 addr;
 447
 448         if (unlikely(skb->len > tx->mtu)) {
 449                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 450                            skb->len, tx->mtu);
 451                 ++priv->stats.tx_dropped;
 452                 ++priv->stats.tx_errors;
 453                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 454                 return;
 455         }
 456
 457         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 458                        tx->tx_head, skb->len, tx->qp->qp_num);
 459
 460         /*
 461          * We put the skb into the tx_ring _before_ we call post_send()
 462          * because it's entirely possible that the completion handler will
 463          * run before we execute anything after the post_send().  That
 464          * means we have to make sure everything is properly recorded and
 465          * our state is consistent before we call post_send().
 466          */
 467         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 468         tx_req->skb = skb;
 469         addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
 470         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 471                 ++priv->stats.tx_errors;
 472                 dev_kfree_skb_any(skb);
 473                 return;
 474         }
 475
 476         tx_req->mapping = addr;
 477
 478         if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
 479                                 addr, skb->len))) {
 480                 ipoib_warn(priv, "post_send failed\n");
 481                 ++priv->stats.tx_errors;
 482                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
 483                 dev_kfree_skb_any(skb);
 484         } else {
 485                 dev->trans_start = jiffies;
 486                 ++tx->tx_head;
 487
 488                 if (tx->tx_head - tx->tx_tail == ipoib_sendq_size) {
 489                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 490                                   tx->qp->qp_num);
 491                         netif_stop_queue(dev);
 492                         set_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags);
 493                 }
 494         }
 495 }
 496
 497 static void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ipoib_cm_tx *tx,
 498                                   struct ib_wc *wc)
 499 {
 500         struct ipoib_dev_priv *priv = netdev_priv(dev);
 501         unsigned int wr_id = wc->wr_id;
 502         struct ipoib_tx_buf *tx_req;
 503         unsigned long flags;
 504
 505         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 506                        wr_id, wc->status);
 507
 508         if (unlikely(wr_id >= ipoib_sendq_size)) {
 509                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 510                            wr_id, ipoib_sendq_size);
 511                 return;
 512         }
 513
 514         tx_req = &tx->tx_ring[wr_id];
 515
 516         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
 517
 518         /* FIXME: is this right? Shouldn't we only increment on success? */
 519         ++priv->stats.tx_packets;
 520         priv->stats.tx_bytes += tx_req->skb->len;
 521
 522         dev_kfree_skb_any(tx_req->skb);
 523
 524         spin_lock_irqsave(&priv->tx_lock, flags);
 525         ++tx->tx_tail;
 526         if (unlikely(test_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags)) &&
 527             tx->tx_head - tx->tx_tail <= ipoib_sendq_size >> 1) {
 528                 clear_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags);
 529                 netif_wake_queue(dev);
 530         }
 531
 532         if (wc->status != IB_WC_SUCCESS &&
 533             wc->status != IB_WC_WR_FLUSH_ERR) {
 534                 struct ipoib_neigh *neigh;
 535
 536                 ipoib_dbg(priv, "failed cm send event "
 537                            "(status=%d, wrid=%d vend_err %x)\n",
 538                            wc->status, wr_id, wc->vendor_err);
 539
 540                 spin_lock(&priv->lock);
 541                 neigh = tx->neigh;
 542
 543                 if (neigh) {
 544                         neigh->cm = NULL;
 545                         list_del(&neigh->list);
 546                         if (neigh->ah)
 547                                 ipoib_put_ah(neigh->ah);
 548                         ipoib_neigh_free(dev, neigh);
 549
 550                         tx->neigh = NULL;
 551                 }
 552
 553                 /* queue would be re-started anyway when TX is destroyed,
 554                  * but it makes sense to do it ASAP here. */
 555                 if (test_and_clear_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags))
 556                         netif_wake_queue(dev);
 557
 558                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 559                         list_move(&tx->list, &priv->cm.reap_list);
 560                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 561                 }
 562
 563                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 564
 565                 spin_unlock(&priv->lock);
 566         }
 567
 568         spin_unlock_irqrestore(&priv->tx_lock, flags);
 569 }
 570
 571 static void ipoib_cm_tx_completion(struct ib_cq *cq, void *tx_ptr)
 572 {
 573         struct ipoib_cm_tx *tx = tx_ptr;
 574         int n, i;
 575
 576         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 577         do {
 578                 n = ib_poll_cq(cq, IPOIB_NUM_WC, tx->ibwc);
 579                 for (i = 0; i < n; ++i)
 580                         ipoib_cm_handle_tx_wc(tx->dev, tx, tx->ibwc + i);
 581         } while (n == IPOIB_NUM_WC);
 582 }
 583
 584 int ipoib_cm_dev_open(struct net_device *dev)
 585 {
 586         struct ipoib_dev_priv *priv = netdev_priv(dev);
 587         int ret;
 588
 589         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 590                 return 0;
 591
 592         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 593         if (IS_ERR(priv->cm.id)) {
 594                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 595                 return IS_ERR(priv->cm.id);
 596         }
 597
 598         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 599                            0, NULL);
 600         if (ret) {
 601                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 602                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 603                 ib_destroy_cm_id(priv->cm.id);
 604                 return ret;
 605         }
 606         return 0;
 607 }
 608
 609 void ipoib_cm_dev_stop(struct net_device *dev)
 610 {
 611         struct ipoib_dev_priv *priv = netdev_priv(dev);
 612         struct ipoib_cm_rx *p;
 613
 614         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 615                 return;
 616
 617         ib_destroy_cm_id(priv->cm.id);
 618         spin_lock_irq(&priv->lock);
 619         while (!list_empty(&priv->cm.passive_ids)) {
 620                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 621                 list_del_init(&p->list);
 622                 spin_unlock_irq(&priv->lock);
 623                 ib_destroy_cm_id(p->id);
 624                 ib_destroy_qp(p->qp);
 625                 kfree(p);
 626                 spin_lock_irq(&priv->lock);
 627         }
 628         spin_unlock_irq(&priv->lock);
 629
 630         cancel_delayed_work(&priv->cm.stale_task);
 631 }
 632
 633 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 634 {
 635         struct ipoib_cm_tx *p = cm_id->context;
 636         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 637         struct ipoib_cm_data *data = event->private_data;
 638         struct sk_buff_head skqueue;
 639         struct ib_qp_attr qp_attr;
 640         int qp_attr_mask, ret;
 641         struct sk_buff *skb;
 642
 643         p->mtu = be32_to_cpu(data->mtu);
 644
 645         if (p->mtu < priv->dev->mtu + IPOIB_ENCAP_LEN) {
 646                 ipoib_warn(priv, "Rejecting connection: mtu %d < device mtu %d + 4\n",
 647                            p->mtu, priv->dev->mtu);
 648                 return -EINVAL;
 649         }
 650
 651         qp_attr.qp_state = IB_QPS_RTR;
 652         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 653         if (ret) {
 654                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 655                 return ret;
 656         }
 657
 658         qp_attr.rq_psn = 0 /* FIXME */;
 659         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 660         if (ret) {
 661                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 662                 return ret;
 663         }
 664
 665         qp_attr.qp_state = IB_QPS_RTS;
 666         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 667         if (ret) {
 668                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 669                 return ret;
 670         }
 671         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 672         if (ret) {
 673                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 674                 return ret;
 675         }
 676
 677         skb_queue_head_init(&skqueue);
 678
 679         spin_lock_irq(&priv->lock);
 680         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 681         if (p->neigh)
 682                 while ((skb = __skb_dequeue(&p->neigh->queue)))
 683                         __skb_queue_tail(&skqueue, skb);
 684         spin_unlock_irq(&priv->lock);
 685
 686         while ((skb = __skb_dequeue(&skqueue))) {
 687                 skb->dev = p->dev;
 688                 if (dev_queue_xmit(skb))
 689                         ipoib_warn(priv, "dev_queue_xmit failed "
 690                                    "to requeue packet\n");
 691         }
 692
 693         ret = ib_send_cm_rtu(cm_id, NULL, 0);
 694         if (ret) {
 695                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
 696                 return ret;
 697         }
 698         return 0;
 699 }
 700
 701 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ib_cq *cq)
 702 {
 703         struct ipoib_dev_priv *priv = netdev_priv(dev);
 704         struct ib_qp_init_attr attr = {};
 705         attr.recv_cq = priv->cq;
 706         attr.srq = priv->cm.srq;
 707         attr.cap.max_send_wr = ipoib_sendq_size;
 708         attr.cap.max_send_sge = 1;
 709         attr.sq_sig_type = IB_SIGNAL_ALL_WR;
 710         attr.qp_type = IB_QPT_RC;
 711         attr.send_cq = cq;
 712         return ib_create_qp(priv->pd, &attr);
 713 }
 714
 715 static int ipoib_cm_send_req(struct net_device *dev,
 716                              struct ib_cm_id *id, struct ib_qp *qp,
 717                              u32 qpn,
 718                              struct ib_sa_path_rec *pathrec)
 719 {
 720         struct ipoib_dev_priv *priv = netdev_priv(dev);
 721         struct ipoib_cm_data data = {};
 722         struct ib_cm_req_param req = {};
 723
 724         data.qpn = cpu_to_be32(priv->qp->qp_num);
 725         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 726
 727         req.primary_path              = pathrec;
 728         req.alternate_path            = NULL;
 729         req.service_id                = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
 730         req.qp_num                    = qp->qp_num;
 731         req.qp_type                   = qp->qp_type;
 732         req.private_data              = &data;
 733         req.private_data_len          = sizeof data;
 734         req.flow_control              = 0;
 735
 736         req.starting_psn              = 0; /* FIXME */
 737
 738         /*
 739          * Pick some arbitrary defaults here; we could make these
 740          * module parameters if anyone cared about setting them.
 741          */
 742         req.responder_resources       = 4;
 743         req.remote_cm_response_timeout = 20;
 744         req.local_cm_response_timeout  = 20;
 745         req.retry_count               = 0; /* RFC draft warns against retries */
 746         req.rnr_retry_count           = 0; /* RFC draft warns against retries */
 747         req.max_cm_retries            = 15;
 748         req.srq                       = 1;
 749         return ib_send_cm_req(id, &req);
 750 }
 751
 752 static int ipoib_cm_modify_tx_init(struct net_device *dev,
 753                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
 754 {
 755         struct ipoib_dev_priv *priv = netdev_priv(dev);
 756         struct ib_qp_attr qp_attr;
 757         int qp_attr_mask, ret;
 758         ret = ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
 759         if (ret) {
 760                 ipoib_warn(priv, "pkey 0x%x not in cache: %d\n", priv->pkey, ret);
 761                 return ret;
 762         }
 763
 764         qp_attr.qp_state = IB_QPS_INIT;
 765         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
 766         qp_attr.port_num = priv->port;
 767         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
 768
 769         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 770         if (ret) {
 771                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
 772                 return ret;
 773         }
 774         return 0;
 775 }
 776
 777 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
 778                             struct ib_sa_path_rec *pathrec)
 779 {
 780         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 781         int ret;
 782
 783         p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring,
 784                                 GFP_KERNEL);
 785         if (!p->tx_ring) {
 786                 ipoib_warn(priv, "failed to allocate tx ring\n");
 787                 ret = -ENOMEM;
 788                 goto err_tx;
 789         }
 790
 791         p->cq = ib_create_cq(priv->ca, ipoib_cm_tx_completion, NULL, p,
 792                              ipoib_sendq_size + 1);
 793         if (IS_ERR(p->cq)) {
 794                 ret = PTR_ERR(p->cq);
 795                 ipoib_warn(priv, "failed to allocate tx cq: %d\n", ret);
 796                 goto err_cq;
 797         }
 798
 799         ret = ib_req_notify_cq(p->cq, IB_CQ_NEXT_COMP);
 800         if (ret) {
 801                 ipoib_warn(priv, "failed to request completion notification: %d\n", ret);
 802                 goto err_req_notify;
 803         }
 804
 805         p->qp = ipoib_cm_create_tx_qp(p->dev, p->cq);
 806         if (IS_ERR(p->qp)) {
 807                 ret = PTR_ERR(p->qp);
 808                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
 809                 goto err_qp;
 810         }
 811
 812         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
 813         if (IS_ERR(p->id)) {
 814                 ret = PTR_ERR(p->id);
 815                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
 816                 goto err_id;
 817         }
 818
 819         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
 820         if (ret) {
 821                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
 822                 goto err_modify;
 823         }
 824
 825         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
 826         if (ret) {
 827                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
 828                 goto err_send_cm;
 829         }
 830
 831         ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
 832                   p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
 833
 834         return 0;
 835
 836 err_send_cm:
 837 err_modify:
 838         ib_destroy_cm_id(p->id);
 839 err_id:
 840         p->id = NULL;
 841         ib_destroy_qp(p->qp);
 842 err_req_notify:
 843 err_qp:
 844         p->qp = NULL;
 845         ib_destroy_cq(p->cq);
 846 err_cq:
 847         p->cq = NULL;
 848 err_tx:
 849         return ret;
 850 }
 851
 852 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
 853 {
 854         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 855         struct ipoib_tx_buf *tx_req;
 856
 857         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
 858                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
 859
 860         if (p->id)
 861                 ib_destroy_cm_id(p->id);
 862
 863         if (p->qp)
 864                 ib_destroy_qp(p->qp);
 865
 866         if (p->cq)
 867                 ib_destroy_cq(p->cq);
 868
 869         if (test_bit(IPOIB_FLAG_NETIF_STOPPED, &p->flags))
 870                 netif_wake_queue(p->dev);
 871
 872         if (p->tx_ring) {
 873                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
 874                         tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
 875                         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
 876                                          DMA_TO_DEVICE);
 877                         dev_kfree_skb_any(tx_req->skb);
 878                         ++p->tx_tail;
 879                 }
 880
 881                 kfree(p->tx_ring);
 882         }
 883
 884         kfree(p);
 885 }
 886
 887 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
 888                                struct ib_cm_event *event)
 889 {
 890         struct ipoib_cm_tx *tx = cm_id->context;
 891         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
 892         struct net_device *dev = priv->dev;
 893         struct ipoib_neigh *neigh;
 894         int ret;
 895
 896         switch (event->event) {
 897         case IB_CM_DREQ_RECEIVED:
 898                 ipoib_dbg(priv, "DREQ received.\n");
 899                 ib_send_cm_drep(cm_id, NULL, 0);
 900                 break;
 901         case IB_CM_REP_RECEIVED:
 902                 ipoib_dbg(priv, "REP received.\n");
 903                 ret = ipoib_cm_rep_handler(cm_id, event);
 904                 if (ret)
 905                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 906                                        NULL, 0, NULL, 0);
 907                 break;
 908         case IB_CM_REQ_ERROR:
 909         case IB_CM_REJ_RECEIVED:
 910         case IB_CM_TIMEWAIT_EXIT:
 911                 ipoib_dbg(priv, "CM error %d.\n", event->event);
 912                 spin_lock_irq(&priv->tx_lock);
 913                 spin_lock(&priv->lock);
 914                 neigh = tx->neigh;
 915
 916                 if (neigh) {
 917                         neigh->cm = NULL;
 918                         list_del(&neigh->list);
 919                         if (neigh->ah)
 920                                 ipoib_put_ah(neigh->ah);
 921                         ipoib_neigh_free(dev, neigh);
 922
 923                         tx->neigh = NULL;
 924                 }
 925
 926                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 927                         list_move(&tx->list, &priv->cm.reap_list);
 928                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 929                 }
 930
 931                 spin_unlock(&priv->lock);
 932                 spin_unlock_irq(&priv->tx_lock);
 933                 break;
 934         default:
 935                 break;
 936         }
 937
 938         return 0;
 939 }
 940
 941 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
 942                                        struct ipoib_neigh *neigh)
 943 {
 944         struct ipoib_dev_priv *priv = netdev_priv(dev);
 945         struct ipoib_cm_tx *tx;
 946
 947         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
 948         if (!tx)
 949                 return NULL;
 950
 951         neigh->cm = tx;
 952         tx->neigh = neigh;
 953         tx->path = path;
 954         tx->dev = dev;
 955         list_add(&tx->list, &priv->cm.start_list);
 956         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
 957         queue_work(ipoib_workqueue, &priv->cm.start_task);
 958         return tx;
 959 }
 960
 961 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
 962 {
 963         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
 964         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 965                 list_move(&tx->list, &priv->cm.reap_list);
 966                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
 967                 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
 968                           IPOIB_GID_ARG(tx->neigh->dgid));
 969                 tx->neigh = NULL;
 970         }
 971 }
 972
 973 static void ipoib_cm_tx_start(struct work_struct *work)
 974 {
 975         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 976                                                    cm.start_task);
 977         struct net_device *dev = priv->dev;
 978         struct ipoib_neigh *neigh;
 979         struct ipoib_cm_tx *p;
 980         unsigned long flags;
 981         int ret;
 982
 983         struct ib_sa_path_rec pathrec;
 984         u32 qpn;
 985
 986         spin_lock_irqsave(&priv->tx_lock, flags);
 987         spin_lock(&priv->lock);
 988         while (!list_empty(&priv->cm.start_list)) {
 989                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
 990                 list_del_init(&p->list);
 991                 neigh = p->neigh;
 992                 qpn = IPOIB_QPN(neigh->neighbour->ha);
 993                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
 994                 spin_unlock(&priv->lock);
 995                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 996                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
 997                 spin_lock_irqsave(&priv->tx_lock, flags);
 998                 spin_lock(&priv->lock);
 999                 if (ret) {
1000                         neigh = p->neigh;
1001                         if (neigh) {
1002                                 neigh->cm = NULL;
1003                                 list_del(&neigh->list);
1004                                 if (neigh->ah)
1005                                         ipoib_put_ah(neigh->ah);
1006                                 ipoib_neigh_free(dev, neigh);
1007                         }
1008                         list_del(&p->list);
1009                         kfree(p);
1010                 }
1011         }
1012         spin_unlock(&priv->lock);
1013         spin_unlock_irqrestore(&priv->tx_lock, flags);
1014 }
1015
1016 static void ipoib_cm_tx_reap(struct work_struct *work)
1017 {
1018         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1019                                                    cm.reap_task);
1020         struct ipoib_cm_tx *p;
1021
1022         spin_lock_irq(&priv->tx_lock);
1023         spin_lock(&priv->lock);
1024         while (!list_empty(&priv->cm.reap_list)) {
1025                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1026                 list_del(&p->list);
1027                 spin_unlock(&priv->lock);
1028                 spin_unlock_irq(&priv->tx_lock);
1029                 ipoib_cm_tx_destroy(p);
1030                 spin_lock_irq(&priv->tx_lock);
1031                 spin_lock(&priv->lock);
1032         }
1033         spin_unlock(&priv->lock);
1034         spin_unlock_irq(&priv->tx_lock);
1035 }
1036
1037 static void ipoib_cm_skb_reap(struct work_struct *work)
1038 {
1039         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1040                                                    cm.skb_task);
1041         struct net_device *dev = priv->dev;
1042         struct sk_buff *skb;
1043
1044         unsigned mtu = priv->mcast_mtu;
1045
1046         spin_lock_irq(&priv->tx_lock);
1047         spin_lock(&priv->lock);
1048         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1049                 spin_unlock(&priv->lock);
1050                 spin_unlock_irq(&priv->tx_lock);
1051                 if (skb->protocol == htons(ETH_P_IP))
1052                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1053 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1054                 else if (skb->protocol == htons(ETH_P_IPV6))
1055                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, dev);
1056 #endif
1057                 dev_kfree_skb_any(skb);
1058                 spin_lock_irq(&priv->tx_lock);
1059                 spin_lock(&priv->lock);
1060         }
1061         spin_unlock(&priv->lock);
1062         spin_unlock_irq(&priv->tx_lock);
1063 }
1064
1065 void ipoib_cm_skb_too_long(struct net_device* dev, struct sk_buff *skb,
1066                            unsigned int mtu)
1067 {
1068         struct ipoib_dev_priv *priv = netdev_priv(dev);
1069         int e = skb_queue_empty(&priv->cm.skb_queue);
1070
1071         if (skb->dst)
1072                 skb->dst->ops->update_pmtu(skb->dst, mtu);
1073
1074         skb_queue_tail(&priv->cm.skb_queue, skb);
1075         if (e)
1076                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1077 }
1078
1079 static void ipoib_cm_stale_task(struct work_struct *work)
1080 {
1081         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1082                                                    cm.stale_task.work);
1083         struct ipoib_cm_rx *p;
1084
1085         spin_lock_irq(&priv->lock);
1086         while (!list_empty(&priv->cm.passive_ids)) {
1087                 /* List if sorted by LRU, start from tail,
1088                  * stop when we see a recently used entry */
1089                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1090                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1091                         break;
1092                 list_del_init(&p->list);
1093                 spin_unlock_irq(&priv->lock);
1094                 ib_destroy_cm_id(p->id);
1095                 ib_destroy_qp(p->qp);
1096                 kfree(p);
1097                 spin_lock_irq(&priv->lock);
1098         }
1099         spin_unlock_irq(&priv->lock);
1100 }
1101
1102
1103 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1104                          char *buf)
1105 {
1106         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1107
1108         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1109                 return sprintf(buf, "connected\n");
1110         else
1111                 return sprintf(buf, "datagram\n");
1112 }
1113
1114 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1115                         const char *buf, size_t count)
1116 {
1117         struct net_device *dev = to_net_dev(d);
1118         struct ipoib_dev_priv *priv = netdev_priv(dev);
1119
1120         /* flush paths if we switch modes so that connections are restarted */
1121         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1122                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1123                 ipoib_warn(priv, "enabling connected mode "
1124                            "will cause multicast packet drops\n");
1125                 ipoib_flush_paths(dev);
1126                 return count;
1127         }
1128
1129         if (!strcmp(buf, "datagram\n")) {
1130                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1131                 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1132                 ipoib_flush_paths(dev);
1133                 return count;
1134         }
1135
1136         return -EINVAL;
1137 }
1138
1139 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1140
1141 int ipoib_cm_add_mode_attr(struct net_device *dev)
1142 {
1143         return device_create_file(&dev->dev, &dev_attr_mode);
1144 }
1145
1146 int ipoib_cm_dev_init(struct net_device *dev)
1147 {
1148         struct ipoib_dev_priv *priv = netdev_priv(dev);
1149         struct ib_srq_init_attr srq_init_attr = {
1150                 .attr = {
1151                         .max_wr  = ipoib_recvq_size,
1152                         .max_sge = IPOIB_CM_RX_SG
1153                 }
1154         };
1155         int ret, i;
1156
1157         INIT_LIST_HEAD(&priv->cm.passive_ids);
1158         INIT_LIST_HEAD(&priv->cm.reap_list);
1159         INIT_LIST_HEAD(&priv->cm.start_list);
1160         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1161         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1162         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1163         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1164
1165         skb_queue_head_init(&priv->cm.skb_queue);
1166
1167         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1168         if (IS_ERR(priv->cm.srq)) {
1169                 ret = PTR_ERR(priv->cm.srq);
1170                 priv->cm.srq = NULL;
1171                 return ret;
1172         }
1173
1174         priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1175                                     GFP_KERNEL);
1176         if (!priv->cm.srq_ring) {
1177                 printk(KERN_WARNING "%s: failed to allocate CM ring (%d entries)\n",
1178                        priv->ca->name, ipoib_recvq_size);
1179                 ipoib_cm_dev_cleanup(dev);
1180                 return -ENOMEM;
1181         }
1182
1183         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
1184                 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1185
1186         priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1187         for (i = 1; i < IPOIB_CM_RX_SG; ++i)
1188                 priv->cm.rx_sge[i].length = PAGE_SIZE;
1189         priv->cm.rx_wr.next = NULL;
1190         priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1191         priv->cm.rx_wr.num_sge = IPOIB_CM_RX_SG;
1192
1193         for (i = 0; i < ipoib_recvq_size; ++i) {
1194                 if (!ipoib_cm_alloc_rx_skb(dev, i, IPOIB_CM_RX_SG - 1,
1195                                            priv->cm.srq_ring[i].mapping)) {
1196                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
1197                         ipoib_cm_dev_cleanup(dev);
1198                         return -ENOMEM;
1199                 }
1200                 if (ipoib_cm_post_receive(dev, i)) {
1201                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
1202                         ipoib_cm_dev_cleanup(dev);
1203                         return -EIO;
1204                 }
1205         }
1206
1207         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1208         return 0;
1209 }
1210
1211 void ipoib_cm_dev_cleanup(struct net_device *dev)
1212 {
1213         struct ipoib_dev_priv *priv = netdev_priv(dev);
1214         int i, ret;
1215
1216         if (!priv->cm.srq)
1217                 return;
1218
1219         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1220
1221         ret = ib_destroy_srq(priv->cm.srq);
1222         if (ret)
1223                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1224
1225         priv->cm.srq = NULL;
1226         if (!priv->cm.srq_ring)
1227                 return;
1228         for (i = 0; i < ipoib_recvq_size; ++i)
1229                 if (priv->cm.srq_ring[i].skb) {
1230                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
1231                                               priv->cm.srq_ring[i].mapping);
1232                         dev_kfree_skb_any(priv->cm.srq_ring[i].skb);
1233                         priv->cm.srq_ring[i].skb = NULL;
1234                 }
1235         kfree(priv->cm.srq_ring);
1236         priv->cm.srq_ring = NULL;
1237 }