SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 #include <linux/delay.h>
  41
  42 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  43 static int data_debug_level;
  44
  45 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  46 MODULE_PARM_DESC(cm_data_debug_level,
  47                  "Enable data path debug tracing for connected mode if > 0");
  48 #endif
  49
  50 #include "ipoib.h"
  51
  52 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  53
  54 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  55 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  56 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  57 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  58
  59 struct ipoib_cm_id {
  60         struct ib_cm_id *id;
  61         int flags;
  62         u32 remote_qpn;
  63         u32 remote_mtu;
  64 };
  65
  66 static struct ib_qp_attr ipoib_cm_err_attr = {
  67         .qp_state = IB_QPS_ERR
  68 };
  69
  70 #define IPOIB_CM_RX_DRAIN_WRID 0x7fffffff
  71
  72 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  73         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  74         .opcode = IB_WR_SEND,
  75 };
  76
  77 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  78                                struct ib_cm_event *event);
  79
  80 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  81                                   u64 mapping[IPOIB_CM_RX_SG])
  82 {
  83         int i;
  84
  85         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  86
  87         for (i = 0; i < frags; ++i)
  88                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  89 }
  90
  91 static int ipoib_cm_post_receive(struct net_device *dev, int id)
  92 {
  93         struct ipoib_dev_priv *priv = netdev_priv(dev);
  94         struct ib_recv_wr *bad_wr;
  95         int i, ret;
  96
  97         priv->cm.rx_wr.wr_id = id | IPOIB_CM_OP_SRQ;
  98
  99         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 100                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 101
 102         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 103         if (unlikely(ret)) {
 104                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 105                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 106                                       priv->cm.srq_ring[id].mapping);
 107                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 108                 priv->cm.srq_ring[id].skb = NULL;
 109         }
 110
 111         return ret;
 112 }
 113
 114 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev, int id, int frags,
 115                                              u64 mapping[IPOIB_CM_RX_SG])
 116 {
 117         struct ipoib_dev_priv *priv = netdev_priv(dev);
 118         struct sk_buff *skb;
 119         int i;
 120
 121         skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
 122         if (unlikely(!skb))
 123                 return NULL;
 124
 125         /*
 126          * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
 127          * IP header to a multiple of 16.
 128          */
 129         skb_reserve(skb, 12);
 130
 131         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 132                                        DMA_FROM_DEVICE);
 133         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 134                 dev_kfree_skb_any(skb);
 135                 return NULL;
 136         }
 137
 138         for (i = 0; i < frags; i++) {
 139                 struct page *page = alloc_page(GFP_ATOMIC);
 140
 141                 if (!page)
 142                         goto partial_error;
 143                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 144
 145                 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
 146                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 147                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 148                         goto partial_error;
 149         }
 150
 151         priv->cm.srq_ring[id].skb = skb;
 152         return skb;
 153
 154 partial_error:
 155
 156         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 157
 158         for (; i >= 0; --i)
 159                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
 160
 161         dev_kfree_skb_any(skb);
 162         return NULL;
 163 }
 164
 165 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv* priv)
 166 {
 167         struct ib_send_wr *bad_wr;
 168         struct ipoib_cm_rx *p;
 169
 170         /* We only reserved 1 extra slot in CQ for drain WRs, so
 171          * make sure we have at most 1 outstanding WR. */
 172         if (list_empty(&priv->cm.rx_flush_list) ||
 173             !list_empty(&priv->cm.rx_drain_list))
 174                 return;
 175
 176         /*
 177          * QPs on flush list are error state.  This way, a "flush
 178          * error" WC will be immediately generated for each WR we post.
 179          */
 180         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 181         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 182                 ipoib_warn(priv, "failed to post drain wr\n");
 183
 184         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 185 }
 186
 187 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 188 {
 189         struct ipoib_cm_rx *p = ctx;
 190         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 191         unsigned long flags;
 192
 193         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 194                 return;
 195
 196         spin_lock_irqsave(&priv->lock, flags);
 197         list_move(&p->list, &priv->cm.rx_flush_list);
 198         p->state = IPOIB_CM_RX_FLUSH;
 199         ipoib_cm_start_rx_drain(priv);
 200         spin_unlock_irqrestore(&priv->lock, flags);
 201 }
 202
 203 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 204                                            struct ipoib_cm_rx *p)
 205 {
 206         struct ipoib_dev_priv *priv = netdev_priv(dev);
 207         struct ib_qp_init_attr attr = {
 208                 .event_handler = ipoib_cm_rx_event_handler,
 209                 .send_cq = priv->cq, /* For drain WR */
 210                 .recv_cq = priv->cq,
 211                 .srq = priv->cm.srq,
 212                 .cap.max_send_wr = 1, /* For drain WR */
 213                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 214                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 215                 .qp_type = IB_QPT_RC,
 216                 .qp_context = p,
 217         };
 218         return ib_create_qp(priv->pd, &attr);
 219 }
 220
 221 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 222                                   struct ib_cm_id *cm_id, struct ib_qp *qp,
 223                                   unsigned psn)
 224 {
 225         struct ipoib_dev_priv *priv = netdev_priv(dev);
 226         struct ib_qp_attr qp_attr;
 227         int qp_attr_mask, ret;
 228
 229         qp_attr.qp_state = IB_QPS_INIT;
 230         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 231         if (ret) {
 232                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 233                 return ret;
 234         }
 235         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 236         if (ret) {
 237                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 238                 return ret;
 239         }
 240         qp_attr.qp_state = IB_QPS_RTR;
 241         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 242         if (ret) {
 243                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 244                 return ret;
 245         }
 246         qp_attr.rq_psn = psn;
 247         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 248         if (ret) {
 249                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 250                 return ret;
 251         }
 252
 253         /*
 254          * Current Mellanox HCA firmware won't generate completions
 255          * with error for drain WRs unless the QP has been moved to
 256          * RTS first. This work-around leaves a window where a QP has
 257          * moved to error asynchronously, but this will eventually get
 258          * fixed in firmware, so let's not error out if modify QP
 259          * fails.
 260          */
 261         qp_attr.qp_state = IB_QPS_RTS;
 262         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 263         if (ret) {
 264                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 265                 return 0;
 266         }
 267         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 268         if (ret) {
 269                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 270                 return 0;
 271         }
 272
 273         return 0;
 274 }
 275
 276 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 277                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 278                              unsigned psn)
 279 {
 280         struct ipoib_dev_priv *priv = netdev_priv(dev);
 281         struct ipoib_cm_data data = {};
 282         struct ib_cm_rep_param rep = {};
 283
 284         data.qpn = cpu_to_be32(priv->qp->qp_num);
 285         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 286
 287         rep.private_data = &data;
 288         rep.private_data_len = sizeof data;
 289         rep.flow_control = 0;
 290         rep.rnr_retry_count = req->rnr_retry_count;
 291         rep.target_ack_delay = 20; /* FIXME */
 292         rep.srq = 1;
 293         rep.qp_num = qp->qp_num;
 294         rep.starting_psn = psn;
 295         return ib_send_cm_rep(cm_id, &rep);
 296 }
 297
 298 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 299 {
 300         struct net_device *dev = cm_id->context;
 301         struct ipoib_dev_priv *priv = netdev_priv(dev);
 302         struct ipoib_cm_rx *p;
 303         unsigned psn;
 304         int ret;
 305
 306         ipoib_dbg(priv, "REQ arrived\n");
 307         p = kzalloc(sizeof *p, GFP_KERNEL);
 308         if (!p)
 309                 return -ENOMEM;
 310         p->dev = dev;
 311         p->id = cm_id;
 312         cm_id->context = p;
 313         p->state = IPOIB_CM_RX_LIVE;
 314         p->jiffies = jiffies;
 315         INIT_LIST_HEAD(&p->list);
 316
 317         p->qp = ipoib_cm_create_rx_qp(dev, p);
 318         if (IS_ERR(p->qp)) {
 319                 ret = PTR_ERR(p->qp);
 320                 goto err_qp;
 321         }
 322
 323         psn = random32() & 0xffffff;
 324         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 325         if (ret)
 326                 goto err_modify;
 327
 328         spin_lock_irq(&priv->lock);
 329         queue_delayed_work(ipoib_workqueue,
 330                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 331         /* Add this entry to passive ids list head, but do not re-add it
 332          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 333         p->jiffies = jiffies;
 334         if (p->state == IPOIB_CM_RX_LIVE)
 335                 list_move(&p->list, &priv->cm.passive_ids);
 336         spin_unlock_irq(&priv->lock);
 337
 338         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 339         if (ret) {
 340                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 341                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 342                         ipoib_warn(priv, "unable to move qp to error state\n");
 343         }
 344         return 0;
 345
 346 err_modify:
 347         ib_destroy_qp(p->qp);
 348 err_qp:
 349         kfree(p);
 350         return ret;
 351 }
 352
 353 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 354                                struct ib_cm_event *event)
 355 {
 356         struct ipoib_cm_rx *p;
 357         struct ipoib_dev_priv *priv;
 358
 359         switch (event->event) {
 360         case IB_CM_REQ_RECEIVED:
 361                 return ipoib_cm_req_handler(cm_id, event);
 362         case IB_CM_DREQ_RECEIVED:
 363                 p = cm_id->context;
 364                 ib_send_cm_drep(cm_id, NULL, 0);
 365                 /* Fall through */
 366         case IB_CM_REJ_RECEIVED:
 367                 p = cm_id->context;
 368                 priv = netdev_priv(p->dev);
 369                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 370                         ipoib_warn(priv, "unable to move qp to error state\n");
 371                 /* Fall through */
 372         default:
 373                 return 0;
 374         }
 375 }
 376 /* Adjust length of skb with fragments to match received data */
 377 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 378                           unsigned int length, struct sk_buff *toskb)
 379 {
 380         int i, num_frags;
 381         unsigned int size;
 382
 383         /* put header into skb */
 384         size = min(length, hdr_space);
 385         skb->tail += size;
 386         skb->len += size;
 387         length -= size;
 388
 389         num_frags = skb_shinfo(skb)->nr_frags;
 390         for (i = 0; i < num_frags; i++) {
 391                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 392
 393                 if (length == 0) {
 394                         /* don't need this page */
 395                         skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
 396                         --skb_shinfo(skb)->nr_frags;
 397                 } else {
 398                         size = min(length, (unsigned) PAGE_SIZE);
 399
 400                         frag->size = size;
 401                         skb->data_len += size;
 402                         skb->truesize += size;
 403                         skb->len += size;
 404                         length -= size;
 405                 }
 406         }
 407 }
 408
 409 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 410 {
 411         struct ipoib_dev_priv *priv = netdev_priv(dev);
 412         unsigned int wr_id = wc->wr_id & ~IPOIB_CM_OP_SRQ;
 413         struct sk_buff *skb, *newskb;
 414         struct ipoib_cm_rx *p;
 415         unsigned long flags;
 416         u64 mapping[IPOIB_CM_RX_SG];
 417         int frags;
 418
 419         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 420                        wr_id, wc->status);
 421
 422         if (unlikely(wr_id >= ipoib_recvq_size)) {
 423                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~IPOIB_CM_OP_SRQ)) {
 424                         spin_lock_irqsave(&priv->lock, flags);
 425                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 426                         ipoib_cm_start_rx_drain(priv);
 427                         queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 428                         spin_unlock_irqrestore(&priv->lock, flags);
 429                 } else
 430                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 431                                    wr_id, ipoib_recvq_size);
 432                 return;
 433         }
 434
 435         skb  = priv->cm.srq_ring[wr_id].skb;
 436
 437         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 438                 ipoib_dbg(priv, "cm recv error "
 439                            "(status=%d, wrid=%d vend_err %x)\n",
 440                            wc->status, wr_id, wc->vendor_err);
 441                 ++priv->stats.rx_dropped;
 442                 goto repost;
 443         }
 444
 445         if (!likely(wr_id & IPOIB_CM_RX_UPDATE_MASK)) {
 446                 p = wc->qp->qp_context;
 447                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 448                         spin_lock_irqsave(&priv->lock, flags);
 449                         p->jiffies = jiffies;
 450                         /* Move this entry to list head, but do not re-add it
 451                          * if it has been moved out of list. */
 452                         if (p->state == IPOIB_CM_RX_LIVE)
 453                                 list_move(&p->list, &priv->cm.passive_ids);
 454                         spin_unlock_irqrestore(&priv->lock, flags);
 455                 }
 456         }
 457
 458         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 459                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 460
 461         newskb = ipoib_cm_alloc_rx_skb(dev, wr_id, frags, mapping);
 462         if (unlikely(!newskb)) {
 463                 /*
 464                  * If we can't allocate a new RX buffer, dump
 465                  * this packet and reuse the old buffer.
 466                  */
 467                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 468                 ++priv->stats.rx_dropped;
 469                 goto repost;
 470         }
 471
 472         ipoib_cm_dma_unmap_rx(priv, frags, priv->cm.srq_ring[wr_id].mapping);
 473         memcpy(priv->cm.srq_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 474
 475         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 476                        wc->byte_len, wc->slid);
 477
 478         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 479
 480         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 481         skb_reset_mac_header(skb);
 482         skb_pull(skb, IPOIB_ENCAP_LEN);
 483
 484         dev->last_rx = jiffies;
 485         ++priv->stats.rx_packets;
 486         priv->stats.rx_bytes += skb->len;
 487
 488         skb->dev = dev;
 489         /* XXX get correct PACKET_ type here */
 490         skb->pkt_type = PACKET_HOST;
 491         netif_receive_skb(skb);
 492
 493 repost:
 494         if (unlikely(ipoib_cm_post_receive(dev, wr_id)))
 495                 ipoib_warn(priv, "ipoib_cm_post_receive failed "
 496                            "for buf %d\n", wr_id);
 497 }
 498
 499 static inline int post_send(struct ipoib_dev_priv *priv,
 500                             struct ipoib_cm_tx *tx,
 501                             unsigned int wr_id,
 502                             u64 addr, int len)
 503 {
 504         struct ib_send_wr *bad_wr;
 505
 506         priv->tx_sge.addr             = addr;
 507         priv->tx_sge.length           = len;
 508
 509         priv->tx_wr.wr_id             = wr_id;
 510
 511         return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
 512 }
 513
 514 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 515 {
 516         struct ipoib_dev_priv *priv = netdev_priv(dev);
 517         struct ipoib_tx_buf *tx_req;
 518         u64 addr;
 519
 520         if (unlikely(skb->len > tx->mtu)) {
 521                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 522                            skb->len, tx->mtu);
 523                 ++priv->stats.tx_dropped;
 524                 ++priv->stats.tx_errors;
 525                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 526                 return;
 527         }
 528
 529         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 530                        tx->tx_head, skb->len, tx->qp->qp_num);
 531
 532         /*
 533          * We put the skb into the tx_ring _before_ we call post_send()
 534          * because it's entirely possible that the completion handler will
 535          * run before we execute anything after the post_send().  That
 536          * means we have to make sure everything is properly recorded and
 537          * our state is consistent before we call post_send().
 538          */
 539         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 540         tx_req->skb = skb;
 541         addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
 542         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 543                 ++priv->stats.tx_errors;
 544                 dev_kfree_skb_any(skb);
 545                 return;
 546         }
 547
 548         tx_req->mapping = addr;
 549
 550         if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
 551                                 addr, skb->len))) {
 552                 ipoib_warn(priv, "post_send failed\n");
 553                 ++priv->stats.tx_errors;
 554                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
 555                 dev_kfree_skb_any(skb);
 556         } else {
 557                 dev->trans_start = jiffies;
 558                 ++tx->tx_head;
 559
 560                 if (tx->tx_head - tx->tx_tail == ipoib_sendq_size) {
 561                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 562                                   tx->qp->qp_num);
 563                         netif_stop_queue(dev);
 564                         set_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags);
 565                 }
 566         }
 567 }
 568
 569 static void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ipoib_cm_tx *tx,
 570                                   struct ib_wc *wc)
 571 {
 572         struct ipoib_dev_priv *priv = netdev_priv(dev);
 573         unsigned int wr_id = wc->wr_id;
 574         struct ipoib_tx_buf *tx_req;
 575         unsigned long flags;
 576
 577         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 578                        wr_id, wc->status);
 579
 580         if (unlikely(wr_id >= ipoib_sendq_size)) {
 581                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 582                            wr_id, ipoib_sendq_size);
 583                 return;
 584         }
 585
 586         tx_req = &tx->tx_ring[wr_id];
 587
 588         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
 589
 590         /* FIXME: is this right? Shouldn't we only increment on success? */
 591         ++priv->stats.tx_packets;
 592         priv->stats.tx_bytes += tx_req->skb->len;
 593
 594         dev_kfree_skb_any(tx_req->skb);
 595
 596         spin_lock_irqsave(&priv->tx_lock, flags);
 597         ++tx->tx_tail;
 598         if (unlikely(test_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags)) &&
 599             tx->tx_head - tx->tx_tail <= ipoib_sendq_size >> 1) {
 600                 clear_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags);
 601                 netif_wake_queue(dev);
 602         }
 603
 604         if (wc->status != IB_WC_SUCCESS &&
 605             wc->status != IB_WC_WR_FLUSH_ERR) {
 606                 struct ipoib_neigh *neigh;
 607
 608                 ipoib_dbg(priv, "failed cm send event "
 609                            "(status=%d, wrid=%d vend_err %x)\n",
 610                            wc->status, wr_id, wc->vendor_err);
 611
 612                 spin_lock(&priv->lock);
 613                 neigh = tx->neigh;
 614
 615                 if (neigh) {
 616                         neigh->cm = NULL;
 617                         list_del(&neigh->list);
 618                         if (neigh->ah)
 619                                 ipoib_put_ah(neigh->ah);
 620                         ipoib_neigh_free(dev, neigh);
 621
 622                         tx->neigh = NULL;
 623                 }
 624
 625                 /* queue would be re-started anyway when TX is destroyed,
 626                  * but it makes sense to do it ASAP here. */
 627                 if (test_and_clear_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags))
 628                         netif_wake_queue(dev);
 629
 630                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 631                         list_move(&tx->list, &priv->cm.reap_list);
 632                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 633                 }
 634
 635                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 636
 637                 spin_unlock(&priv->lock);
 638         }
 639
 640         spin_unlock_irqrestore(&priv->tx_lock, flags);
 641 }
 642
 643 static void ipoib_cm_tx_completion(struct ib_cq *cq, void *tx_ptr)
 644 {
 645         struct ipoib_cm_tx *tx = tx_ptr;
 646         int n, i;
 647
 648         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 649         do {
 650                 n = ib_poll_cq(cq, IPOIB_NUM_WC, tx->ibwc);
 651                 for (i = 0; i < n; ++i)
 652                         ipoib_cm_handle_tx_wc(tx->dev, tx, tx->ibwc + i);
 653         } while (n == IPOIB_NUM_WC);
 654 }
 655
 656 int ipoib_cm_dev_open(struct net_device *dev)
 657 {
 658         struct ipoib_dev_priv *priv = netdev_priv(dev);
 659         int ret;
 660
 661         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 662                 return 0;
 663
 664         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 665         if (IS_ERR(priv->cm.id)) {
 666                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 667                 ret = PTR_ERR(priv->cm.id);
 668                 goto err_cm;
 669         }
 670
 671         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 672                            0, NULL);
 673         if (ret) {
 674                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 675                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 676                 goto err_listen;
 677         }
 678
 679         return 0;
 680
 681 err_listen:
 682         ib_destroy_cm_id(priv->cm.id);
 683 err_cm:
 684         priv->cm.id = NULL;
 685         return ret;
 686 }
 687
 688 void ipoib_cm_dev_stop(struct net_device *dev)
 689 {
 690         struct ipoib_dev_priv *priv = netdev_priv(dev);
 691         struct ipoib_cm_rx *p, *n;
 692         unsigned long begin;
 693         LIST_HEAD(list);
 694         int ret;
 695
 696         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 697                 return;
 698
 699         ib_destroy_cm_id(priv->cm.id);
 700         priv->cm.id = NULL;
 701
 702         spin_lock_irq(&priv->lock);
 703         while (!list_empty(&priv->cm.passive_ids)) {
 704                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 705                 list_move(&p->list, &priv->cm.rx_error_list);
 706                 p->state = IPOIB_CM_RX_ERROR;
 707                 spin_unlock_irq(&priv->lock);
 708                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 709                 if (ret)
 710                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 711                 spin_lock_irq(&priv->lock);
 712         }
 713
 714         /* Wait for all RX to be drained */
 715         begin = jiffies;
 716
 717         while (!list_empty(&priv->cm.rx_error_list) ||
 718                !list_empty(&priv->cm.rx_flush_list) ||
 719                !list_empty(&priv->cm.rx_drain_list)) {
 720                 if (time_after(jiffies, begin + 5 * HZ)) {
 721                         ipoib_warn(priv, "RX drain timing out\n");
 722
 723                         /*
 724                          * assume the HW is wedged and just free up everything.
 725                          */
 726                         list_splice_init(&priv->cm.rx_flush_list, &list);
 727                         list_splice_init(&priv->cm.rx_error_list, &list);
 728                         list_splice_init(&priv->cm.rx_drain_list, &list);
 729                         break;
 730                 }
 731                 spin_unlock_irq(&priv->lock);
 732                 msleep(1);
 733                 ipoib_drain_cq(dev);
 734                 spin_lock_irq(&priv->lock);
 735         }
 736
 737         list_splice_init(&priv->cm.rx_reap_list, &list);
 738
 739         spin_unlock_irq(&priv->lock);
 740
 741         list_for_each_entry_safe(p, n, &list, list) {
 742                 ib_destroy_cm_id(p->id);
 743                 ib_destroy_qp(p->qp);
 744                 kfree(p);
 745         }
 746
 747         cancel_delayed_work(&priv->cm.stale_task);
 748 }
 749
 750 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 751 {
 752         struct ipoib_cm_tx *p = cm_id->context;
 753         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 754         struct ipoib_cm_data *data = event->private_data;
 755         struct sk_buff_head skqueue;
 756         struct ib_qp_attr qp_attr;
 757         int qp_attr_mask, ret;
 758         struct sk_buff *skb;
 759
 760         p->mtu = be32_to_cpu(data->mtu);
 761
 762         if (p->mtu < priv->dev->mtu + IPOIB_ENCAP_LEN) {
 763                 ipoib_warn(priv, "Rejecting connection: mtu %d < device mtu %d + 4\n",
 764                            p->mtu, priv->dev->mtu);
 765                 return -EINVAL;
 766         }
 767
 768         qp_attr.qp_state = IB_QPS_RTR;
 769         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 770         if (ret) {
 771                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 772                 return ret;
 773         }
 774
 775         qp_attr.rq_psn = 0 /* FIXME */;
 776         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 777         if (ret) {
 778                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 779                 return ret;
 780         }
 781
 782         qp_attr.qp_state = IB_QPS_RTS;
 783         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 784         if (ret) {
 785                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 786                 return ret;
 787         }
 788         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 789         if (ret) {
 790                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 791                 return ret;
 792         }
 793
 794         skb_queue_head_init(&skqueue);
 795
 796         spin_lock_irq(&priv->lock);
 797         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 798         if (p->neigh)
 799                 while ((skb = __skb_dequeue(&p->neigh->queue)))
 800                         __skb_queue_tail(&skqueue, skb);
 801         spin_unlock_irq(&priv->lock);
 802
 803         while ((skb = __skb_dequeue(&skqueue))) {
 804                 skb->dev = p->dev;
 805                 if (dev_queue_xmit(skb))
 806                         ipoib_warn(priv, "dev_queue_xmit failed "
 807                                    "to requeue packet\n");
 808         }
 809
 810         ret = ib_send_cm_rtu(cm_id, NULL, 0);
 811         if (ret) {
 812                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
 813                 return ret;
 814         }
 815         return 0;
 816 }
 817
 818 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ib_cq *cq)
 819 {
 820         struct ipoib_dev_priv *priv = netdev_priv(dev);
 821         struct ib_qp_init_attr attr = {};
 822         attr.recv_cq = priv->cq;
 823         attr.srq = priv->cm.srq;
 824         attr.cap.max_send_wr = ipoib_sendq_size;
 825         attr.cap.max_send_sge = 1;
 826         attr.sq_sig_type = IB_SIGNAL_ALL_WR;
 827         attr.qp_type = IB_QPT_RC;
 828         attr.send_cq = cq;
 829         return ib_create_qp(priv->pd, &attr);
 830 }
 831
 832 static int ipoib_cm_send_req(struct net_device *dev,
 833                              struct ib_cm_id *id, struct ib_qp *qp,
 834                              u32 qpn,
 835                              struct ib_sa_path_rec *pathrec)
 836 {
 837         struct ipoib_dev_priv *priv = netdev_priv(dev);
 838         struct ipoib_cm_data data = {};
 839         struct ib_cm_req_param req = {};
 840
 841         data.qpn = cpu_to_be32(priv->qp->qp_num);
 842         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 843
 844         req.primary_path              = pathrec;
 845         req.alternate_path            = NULL;
 846         req.service_id                = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
 847         req.qp_num                    = qp->qp_num;
 848         req.qp_type                   = qp->qp_type;
 849         req.private_data              = &data;
 850         req.private_data_len          = sizeof data;
 851         req.flow_control              = 0;
 852
 853         req.starting_psn              = 0; /* FIXME */
 854
 855         /*
 856          * Pick some arbitrary defaults here; we could make these
 857          * module parameters if anyone cared about setting them.
 858          */
 859         req.responder_resources       = 4;
 860         req.remote_cm_response_timeout = 20;
 861         req.local_cm_response_timeout  = 20;
 862         req.retry_count               = 0; /* RFC draft warns against retries */
 863         req.rnr_retry_count           = 0; /* RFC draft warns against retries */
 864         req.max_cm_retries            = 15;
 865         req.srq                       = 1;
 866         return ib_send_cm_req(id, &req);
 867 }
 868
 869 static int ipoib_cm_modify_tx_init(struct net_device *dev,
 870                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
 871 {
 872         struct ipoib_dev_priv *priv = netdev_priv(dev);
 873         struct ib_qp_attr qp_attr;
 874         int qp_attr_mask, ret;
 875         ret = ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
 876         if (ret) {
 877                 ipoib_warn(priv, "pkey 0x%x not in cache: %d\n", priv->pkey, ret);
 878                 return ret;
 879         }
 880
 881         qp_attr.qp_state = IB_QPS_INIT;
 882         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
 883         qp_attr.port_num = priv->port;
 884         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
 885
 886         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 887         if (ret) {
 888                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
 889                 return ret;
 890         }
 891         return 0;
 892 }
 893
 894 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
 895                             struct ib_sa_path_rec *pathrec)
 896 {
 897         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 898         int ret;
 899
 900         p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring,
 901                                 GFP_KERNEL);
 902         if (!p->tx_ring) {
 903                 ipoib_warn(priv, "failed to allocate tx ring\n");
 904                 ret = -ENOMEM;
 905                 goto err_tx;
 906         }
 907
 908         p->cq = ib_create_cq(priv->ca, ipoib_cm_tx_completion, NULL, p,
 909                              ipoib_sendq_size + 1, 0);
 910         if (IS_ERR(p->cq)) {
 911                 ret = PTR_ERR(p->cq);
 912                 ipoib_warn(priv, "failed to allocate tx cq: %d\n", ret);
 913                 goto err_cq;
 914         }
 915
 916         ret = ib_req_notify_cq(p->cq, IB_CQ_NEXT_COMP);
 917         if (ret) {
 918                 ipoib_warn(priv, "failed to request completion notification: %d\n", ret);
 919                 goto err_req_notify;
 920         }
 921
 922         p->qp = ipoib_cm_create_tx_qp(p->dev, p->cq);
 923         if (IS_ERR(p->qp)) {
 924                 ret = PTR_ERR(p->qp);
 925                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
 926                 goto err_qp;
 927         }
 928
 929         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
 930         if (IS_ERR(p->id)) {
 931                 ret = PTR_ERR(p->id);
 932                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
 933                 goto err_id;
 934         }
 935
 936         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
 937         if (ret) {
 938                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
 939                 goto err_modify;
 940         }
 941
 942         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
 943         if (ret) {
 944                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
 945                 goto err_send_cm;
 946         }
 947
 948         ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
 949                   p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
 950
 951         return 0;
 952
 953 err_send_cm:
 954 err_modify:
 955         ib_destroy_cm_id(p->id);
 956 err_id:
 957         p->id = NULL;
 958         ib_destroy_qp(p->qp);
 959 err_req_notify:
 960 err_qp:
 961         p->qp = NULL;
 962         ib_destroy_cq(p->cq);
 963 err_cq:
 964         p->cq = NULL;
 965 err_tx:
 966         return ret;
 967 }
 968
 969 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
 970 {
 971         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 972         struct ipoib_tx_buf *tx_req;
 973
 974         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
 975                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
 976
 977         if (p->id)
 978                 ib_destroy_cm_id(p->id);
 979
 980         if (p->qp)
 981                 ib_destroy_qp(p->qp);
 982
 983         if (p->cq)
 984                 ib_destroy_cq(p->cq);
 985
 986         if (test_bit(IPOIB_FLAG_NETIF_STOPPED, &p->flags))
 987                 netif_wake_queue(p->dev);
 988
 989         if (p->tx_ring) {
 990                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
 991                         tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
 992                         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
 993                                          DMA_TO_DEVICE);
 994                         dev_kfree_skb_any(tx_req->skb);
 995                         ++p->tx_tail;
 996                 }
 997
 998                 kfree(p->tx_ring);
 999         }
1000
1001         kfree(p);
1002 }
1003
1004 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1005                                struct ib_cm_event *event)
1006 {
1007         struct ipoib_cm_tx *tx = cm_id->context;
1008         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1009         struct net_device *dev = priv->dev;
1010         struct ipoib_neigh *neigh;
1011         int ret;
1012
1013         switch (event->event) {
1014         case IB_CM_DREQ_RECEIVED:
1015                 ipoib_dbg(priv, "DREQ received.\n");
1016                 ib_send_cm_drep(cm_id, NULL, 0);
1017                 break;
1018         case IB_CM_REP_RECEIVED:
1019                 ipoib_dbg(priv, "REP received.\n");
1020                 ret = ipoib_cm_rep_handler(cm_id, event);
1021                 if (ret)
1022                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1023                                        NULL, 0, NULL, 0);
1024                 break;
1025         case IB_CM_REQ_ERROR:
1026         case IB_CM_REJ_RECEIVED:
1027         case IB_CM_TIMEWAIT_EXIT:
1028                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1029                 spin_lock_irq(&priv->tx_lock);
1030                 spin_lock(&priv->lock);
1031                 neigh = tx->neigh;
1032
1033                 if (neigh) {
1034                         neigh->cm = NULL;
1035                         list_del(&neigh->list);
1036                         if (neigh->ah)
1037                                 ipoib_put_ah(neigh->ah);
1038                         ipoib_neigh_free(dev, neigh);
1039
1040                         tx->neigh = NULL;
1041                 }
1042
1043                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1044                         list_move(&tx->list, &priv->cm.reap_list);
1045                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1046                 }
1047
1048                 spin_unlock(&priv->lock);
1049                 spin_unlock_irq(&priv->tx_lock);
1050                 break;
1051         default:
1052                 break;
1053         }
1054
1055         return 0;
1056 }
1057
1058 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1059                                        struct ipoib_neigh *neigh)
1060 {
1061         struct ipoib_dev_priv *priv = netdev_priv(dev);
1062         struct ipoib_cm_tx *tx;
1063
1064         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1065         if (!tx)
1066                 return NULL;
1067
1068         neigh->cm = tx;
1069         tx->neigh = neigh;
1070         tx->path = path;
1071         tx->dev = dev;
1072         list_add(&tx->list, &priv->cm.start_list);
1073         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1074         queue_work(ipoib_workqueue, &priv->cm.start_task);
1075         return tx;
1076 }
1077
1078 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1079 {
1080         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1081         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1082                 list_move(&tx->list, &priv->cm.reap_list);
1083                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1084                 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1085                           IPOIB_GID_ARG(tx->neigh->dgid));
1086                 tx->neigh = NULL;
1087         }
1088 }
1089
1090 static void ipoib_cm_tx_start(struct work_struct *work)
1091 {
1092         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1093                                                    cm.start_task);
1094         struct net_device *dev = priv->dev;
1095         struct ipoib_neigh *neigh;
1096         struct ipoib_cm_tx *p;
1097         unsigned long flags;
1098         int ret;
1099
1100         struct ib_sa_path_rec pathrec;
1101         u32 qpn;
1102
1103         spin_lock_irqsave(&priv->tx_lock, flags);
1104         spin_lock(&priv->lock);
1105         while (!list_empty(&priv->cm.start_list)) {
1106                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1107                 list_del_init(&p->list);
1108                 neigh = p->neigh;
1109                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1110                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1111                 spin_unlock(&priv->lock);
1112                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1113                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1114                 spin_lock_irqsave(&priv->tx_lock, flags);
1115                 spin_lock(&priv->lock);
1116                 if (ret) {
1117                         neigh = p->neigh;
1118                         if (neigh) {
1119                                 neigh->cm = NULL;
1120                                 list_del(&neigh->list);
1121                                 if (neigh->ah)
1122                                         ipoib_put_ah(neigh->ah);
1123                                 ipoib_neigh_free(dev, neigh);
1124                         }
1125                         list_del(&p->list);
1126                         kfree(p);
1127                 }
1128         }
1129         spin_unlock(&priv->lock);
1130         spin_unlock_irqrestore(&priv->tx_lock, flags);
1131 }
1132
1133 static void ipoib_cm_tx_reap(struct work_struct *work)
1134 {
1135         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1136                                                    cm.reap_task);
1137         struct ipoib_cm_tx *p;
1138
1139         spin_lock_irq(&priv->tx_lock);
1140         spin_lock(&priv->lock);
1141         while (!list_empty(&priv->cm.reap_list)) {
1142                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1143                 list_del(&p->list);
1144                 spin_unlock(&priv->lock);
1145                 spin_unlock_irq(&priv->tx_lock);
1146                 ipoib_cm_tx_destroy(p);
1147                 spin_lock_irq(&priv->tx_lock);
1148                 spin_lock(&priv->lock);
1149         }
1150         spin_unlock(&priv->lock);
1151         spin_unlock_irq(&priv->tx_lock);
1152 }
1153
1154 static void ipoib_cm_skb_reap(struct work_struct *work)
1155 {
1156         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1157                                                    cm.skb_task);
1158         struct net_device *dev = priv->dev;
1159         struct sk_buff *skb;
1160
1161         unsigned mtu = priv->mcast_mtu;
1162
1163         spin_lock_irq(&priv->tx_lock);
1164         spin_lock(&priv->lock);
1165         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1166                 spin_unlock(&priv->lock);
1167                 spin_unlock_irq(&priv->tx_lock);
1168                 if (skb->protocol == htons(ETH_P_IP))
1169                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1170 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1171                 else if (skb->protocol == htons(ETH_P_IPV6))
1172                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, dev);
1173 #endif
1174                 dev_kfree_skb_any(skb);
1175                 spin_lock_irq(&priv->tx_lock);
1176                 spin_lock(&priv->lock);
1177         }
1178         spin_unlock(&priv->lock);
1179         spin_unlock_irq(&priv->tx_lock);
1180 }
1181
1182 void ipoib_cm_skb_too_long(struct net_device* dev, struct sk_buff *skb,
1183                            unsigned int mtu)
1184 {
1185         struct ipoib_dev_priv *priv = netdev_priv(dev);
1186         int e = skb_queue_empty(&priv->cm.skb_queue);
1187
1188         if (skb->dst)
1189                 skb->dst->ops->update_pmtu(skb->dst, mtu);
1190
1191         skb_queue_tail(&priv->cm.skb_queue, skb);
1192         if (e)
1193                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1194 }
1195
1196 static void ipoib_cm_rx_reap(struct work_struct *work)
1197 {
1198         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1199                                                    cm.rx_reap_task);
1200         struct ipoib_cm_rx *p, *n;
1201         LIST_HEAD(list);
1202
1203         spin_lock_irq(&priv->lock);
1204         list_splice_init(&priv->cm.rx_reap_list, &list);
1205         spin_unlock_irq(&priv->lock);
1206
1207         list_for_each_entry_safe(p, n, &list, list) {
1208                 ib_destroy_cm_id(p->id);
1209                 ib_destroy_qp(p->qp);
1210                 kfree(p);
1211         }
1212 }
1213
1214 static void ipoib_cm_stale_task(struct work_struct *work)
1215 {
1216         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1217                                                    cm.stale_task.work);
1218         struct ipoib_cm_rx *p;
1219         int ret;
1220
1221         spin_lock_irq(&priv->lock);
1222         while (!list_empty(&priv->cm.passive_ids)) {
1223                 /* List is sorted by LRU, start from tail,
1224                  * stop when we see a recently used entry */
1225                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1226                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1227                         break;
1228                 list_move(&p->list, &priv->cm.rx_error_list);
1229                 p->state = IPOIB_CM_RX_ERROR;
1230                 spin_unlock_irq(&priv->lock);
1231                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1232                 if (ret)
1233                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1234                 spin_lock_irq(&priv->lock);
1235         }
1236
1237         if (!list_empty(&priv->cm.passive_ids))
1238                 queue_delayed_work(ipoib_workqueue,
1239                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1240         spin_unlock_irq(&priv->lock);
1241 }
1242
1243
1244 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1245                          char *buf)
1246 {
1247         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1248
1249         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1250                 return sprintf(buf, "connected\n");
1251         else
1252                 return sprintf(buf, "datagram\n");
1253 }
1254
1255 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1256                         const char *buf, size_t count)
1257 {
1258         struct net_device *dev = to_net_dev(d);
1259         struct ipoib_dev_priv *priv = netdev_priv(dev);
1260
1261         /* flush paths if we switch modes so that connections are restarted */
1262         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1263                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1264                 ipoib_warn(priv, "enabling connected mode "
1265                            "will cause multicast packet drops\n");
1266                 ipoib_flush_paths(dev);
1267                 return count;
1268         }
1269
1270         if (!strcmp(buf, "datagram\n")) {
1271                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1272                 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1273                 ipoib_flush_paths(dev);
1274                 return count;
1275         }
1276
1277         return -EINVAL;
1278 }
1279
1280 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1281
1282 int ipoib_cm_add_mode_attr(struct net_device *dev)
1283 {
1284         return device_create_file(&dev->dev, &dev_attr_mode);
1285 }
1286
1287 int ipoib_cm_dev_init(struct net_device *dev)
1288 {
1289         struct ipoib_dev_priv *priv = netdev_priv(dev);
1290         struct ib_srq_init_attr srq_init_attr = {
1291                 .attr = {
1292                         .max_wr  = ipoib_recvq_size,
1293                         .max_sge = IPOIB_CM_RX_SG
1294                 }
1295         };
1296         int ret, i;
1297
1298         INIT_LIST_HEAD(&priv->cm.passive_ids);
1299         INIT_LIST_HEAD(&priv->cm.reap_list);
1300         INIT_LIST_HEAD(&priv->cm.start_list);
1301         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1302         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1303         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1304         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1305         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1306         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1307         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1308         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1309         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1310
1311         skb_queue_head_init(&priv->cm.skb_queue);
1312
1313         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1314         if (IS_ERR(priv->cm.srq)) {
1315                 ret = PTR_ERR(priv->cm.srq);
1316                 priv->cm.srq = NULL;
1317                 return ret;
1318         }
1319
1320         priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1321                                     GFP_KERNEL);
1322         if (!priv->cm.srq_ring) {
1323                 printk(KERN_WARNING "%s: failed to allocate CM ring (%d entries)\n",
1324                        priv->ca->name, ipoib_recvq_size);
1325                 ipoib_cm_dev_cleanup(dev);
1326                 return -ENOMEM;
1327         }
1328
1329         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
1330                 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1331
1332         priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1333         for (i = 1; i < IPOIB_CM_RX_SG; ++i)
1334                 priv->cm.rx_sge[i].length = PAGE_SIZE;
1335         priv->cm.rx_wr.next = NULL;
1336         priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1337         priv->cm.rx_wr.num_sge = IPOIB_CM_RX_SG;
1338
1339         for (i = 0; i < ipoib_recvq_size; ++i) {
1340                 if (!ipoib_cm_alloc_rx_skb(dev, i, IPOIB_CM_RX_SG - 1,
1341                                            priv->cm.srq_ring[i].mapping)) {
1342                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
1343                         ipoib_cm_dev_cleanup(dev);
1344                         return -ENOMEM;
1345                 }
1346                 if (ipoib_cm_post_receive(dev, i)) {
1347                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
1348                         ipoib_cm_dev_cleanup(dev);
1349                         return -EIO;
1350                 }
1351         }
1352
1353         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1354         return 0;
1355 }
1356
1357 void ipoib_cm_dev_cleanup(struct net_device *dev)
1358 {
1359         struct ipoib_dev_priv *priv = netdev_priv(dev);
1360         int i, ret;
1361
1362         if (!priv->cm.srq)
1363                 return;
1364
1365         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1366
1367         ret = ib_destroy_srq(priv->cm.srq);
1368         if (ret)
1369                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1370
1371         priv->cm.srq = NULL;
1372         if (!priv->cm.srq_ring)
1373                 return;
1374         for (i = 0; i < ipoib_recvq_size; ++i)
1375                 if (priv->cm.srq_ring[i].skb) {
1376                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
1377                                               priv->cm.srq_ring[i].mapping);
1378                         dev_kfree_skb_any(priv->cm.srq_ring[i].skb);
1379                         priv->cm.srq_ring[i].skb = NULL;
1380                 }
1381         kfree(priv->cm.srq_ring);
1382         priv->cm.srq_ring = NULL;
1383 }