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
  33 #include <rdma/ib_cm.h>
  34 #include <rdma/ib_cache.h>
  35 #include <net/dst.h>
  36 #include <net/icmp.h>
  37 #include <linux/icmpv6.h>
  38 #include <linux/delay.h>
  39 #include <linux/vmalloc.h>
  40
  41 #include "ipoib.h"
  42
  43 int ipoib_max_conn_qp = 128;
  44
  45 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
  46 MODULE_PARM_DESC(max_nonsrq_conn_qp,
  47                  "Max number of connected-mode QPs per interface "
  48                  "(applied only if shared receive queue is not available)");
  49
  50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  51 static int data_debug_level;
  52
  53 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  54 MODULE_PARM_DESC(cm_data_debug_level,
  55                  "Enable data path debug tracing for connected mode if > 0");
  56 #endif
  57
  58 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  59
  60 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  61 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  62 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  63 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  64
  65 static struct ib_qp_attr ipoib_cm_err_attr = {
  66         .qp_state = IB_QPS_ERR
  67 };
  68
  69 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
  70
  71 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  72         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  73         .opcode = IB_WR_SEND,
  74 };
  75
  76 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  77                                struct ib_cm_event *event);
  78
  79 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  80                                   u64 mapping[IPOIB_CM_RX_SG])
  81 {
  82         int i;
  83
  84         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  85
  86         for (i = 0; i < frags; ++i)
  87                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  88 }
  89
  90 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
  91 {
  92         struct ipoib_dev_priv *priv = netdev_priv(dev);
  93         struct ib_recv_wr *bad_wr;
  94         int i, ret;
  95
  96         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
  97
  98         for (i = 0; i < priv->cm.num_frags; ++i)
  99                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 100
 101         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 102         if (unlikely(ret)) {
 103                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 104                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 105                                       priv->cm.srq_ring[id].mapping);
 106                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 107                 priv->cm.srq_ring[id].skb = NULL;
 108         }
 109
 110         return ret;
 111 }
 112
 113 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 114                                         struct ipoib_cm_rx *rx, int id)
 115 {
 116         struct ipoib_dev_priv *priv = netdev_priv(dev);
 117         struct ib_recv_wr *bad_wr;
 118         int i, ret;
 119
 120         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 121
 122         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 123                 priv->cm.rx_sge[i].addr = rx->rx_ring[id].mapping[i];
 124
 125         ret = ib_post_recv(rx->qp, &priv->cm.rx_wr, &bad_wr);
 126         if (unlikely(ret)) {
 127                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 128                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 129                                       rx->rx_ring[id].mapping);
 130                 dev_kfree_skb_any(rx->rx_ring[id].skb);
 131                 rx->rx_ring[id].skb = NULL;
 132         }
 133
 134         return ret;
 135 }
 136
 137 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
 138                                              struct ipoib_cm_rx_buf *rx_ring,
 139                                              int id, int frags,
 140                                              u64 mapping[IPOIB_CM_RX_SG])
 141 {
 142         struct ipoib_dev_priv *priv = netdev_priv(dev);
 143         struct sk_buff *skb;
 144         int i;
 145
 146         skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
 147         if (unlikely(!skb))
 148                 return NULL;
 149
 150         /*
 151          * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
 152          * IP header to a multiple of 16.
 153          */
 154         skb_reserve(skb, 12);
 155
 156         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 157                                        DMA_FROM_DEVICE);
 158         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 159                 dev_kfree_skb_any(skb);
 160                 return NULL;
 161         }
 162
 163         for (i = 0; i < frags; i++) {
 164                 struct page *page = alloc_page(GFP_ATOMIC);
 165
 166                 if (!page)
 167                         goto partial_error;
 168                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 169
 170                 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
 171                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 172                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 173                         goto partial_error;
 174         }
 175
 176         rx_ring[id].skb = skb;
 177         return skb;
 178
 179 partial_error:
 180
 181         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 182
 183         for (; i > 0; --i)
 184                 ib_dma_unmap_single(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
 185
 186         dev_kfree_skb_any(skb);
 187         return NULL;
 188 }
 189
 190 static void ipoib_cm_free_rx_ring(struct net_device *dev,
 191                                   struct ipoib_cm_rx_buf *rx_ring)
 192 {
 193         struct ipoib_dev_priv *priv = netdev_priv(dev);
 194         int i;
 195
 196         for (i = 0; i < ipoib_recvq_size; ++i)
 197                 if (rx_ring[i].skb) {
 198                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 199                                               rx_ring[i].mapping);
 200                         dev_kfree_skb_any(rx_ring[i].skb);
 201                 }
 202
 203         kfree(rx_ring);
 204 }
 205
 206 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
 207 {
 208         struct ib_send_wr *bad_wr;
 209         struct ipoib_cm_rx *p;
 210
 211         /* We only reserved 1 extra slot in CQ for drain WRs, so
 212          * make sure we have at most 1 outstanding WR. */
 213         if (list_empty(&priv->cm.rx_flush_list) ||
 214             !list_empty(&priv->cm.rx_drain_list))
 215                 return;
 216
 217         /*
 218          * QPs on flush list are error state.  This way, a "flush
 219          * error" WC will be immediately generated for each WR we post.
 220          */
 221         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 222         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 223                 ipoib_warn(priv, "failed to post drain wr\n");
 224
 225         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 226 }
 227
 228 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 229 {
 230         struct ipoib_cm_rx *p = ctx;
 231         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 232         unsigned long flags;
 233
 234         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 235                 return;
 236
 237         spin_lock_irqsave(&priv->lock, flags);
 238         list_move(&p->list, &priv->cm.rx_flush_list);
 239         p->state = IPOIB_CM_RX_FLUSH;
 240         ipoib_cm_start_rx_drain(priv);
 241         spin_unlock_irqrestore(&priv->lock, flags);
 242 }
 243
 244 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 245                                            struct ipoib_cm_rx *p)
 246 {
 247         struct ipoib_dev_priv *priv = netdev_priv(dev);
 248         struct ib_qp_init_attr attr = {
 249                 .event_handler = ipoib_cm_rx_event_handler,
 250                 .send_cq = priv->recv_cq, /* For drain WR */
 251                 .recv_cq = priv->recv_cq,
 252                 .srq = priv->cm.srq,
 253                 .cap.max_send_wr = 1, /* For drain WR */
 254                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 255                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 256                 .qp_type = IB_QPT_RC,
 257                 .qp_context = p,
 258         };
 259
 260         if (!ipoib_cm_has_srq(dev)) {
 261                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 262                 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
 263         }
 264
 265         return ib_create_qp(priv->pd, &attr);
 266 }
 267
 268 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 269                                  struct ib_cm_id *cm_id, struct ib_qp *qp,
 270                                  unsigned psn)
 271 {
 272         struct ipoib_dev_priv *priv = netdev_priv(dev);
 273         struct ib_qp_attr qp_attr;
 274         int qp_attr_mask, ret;
 275
 276         qp_attr.qp_state = IB_QPS_INIT;
 277         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 278         if (ret) {
 279                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 280                 return ret;
 281         }
 282         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 283         if (ret) {
 284                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 285                 return ret;
 286         }
 287         qp_attr.qp_state = IB_QPS_RTR;
 288         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 289         if (ret) {
 290                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 291                 return ret;
 292         }
 293         qp_attr.rq_psn = psn;
 294         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 295         if (ret) {
 296                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 297                 return ret;
 298         }
 299
 300         /*
 301          * Current Mellanox HCA firmware won't generate completions
 302          * with error for drain WRs unless the QP has been moved to
 303          * RTS first. This work-around leaves a window where a QP has
 304          * moved to error asynchronously, but this will eventually get
 305          * fixed in firmware, so let's not error out if modify QP
 306          * fails.
 307          */
 308         qp_attr.qp_state = IB_QPS_RTS;
 309         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 310         if (ret) {
 311                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 312                 return 0;
 313         }
 314         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 315         if (ret) {
 316                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 317                 return 0;
 318         }
 319
 320         return 0;
 321 }
 322
 323 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
 324                                    struct ipoib_cm_rx *rx)
 325 {
 326         struct ipoib_dev_priv *priv = netdev_priv(dev);
 327         int ret;
 328         int i;
 329
 330         rx->rx_ring = kcalloc(ipoib_recvq_size, sizeof *rx->rx_ring, GFP_KERNEL);
 331         if (!rx->rx_ring)
 332                 return -ENOMEM;
 333
 334         spin_lock_irq(&priv->lock);
 335
 336         if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
 337                 spin_unlock_irq(&priv->lock);
 338                 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
 339                 ret = -EINVAL;
 340                 goto err_free;
 341         } else
 342                 ++priv->cm.nonsrq_conn_qp;
 343
 344         spin_unlock_irq(&priv->lock);
 345
 346         for (i = 0; i < ipoib_recvq_size; ++i) {
 347                 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
 348                                            rx->rx_ring[i].mapping)) {
 349                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 350                                 ret = -ENOMEM;
 351                                 goto err_count;
 352                         }
 353                 ret = ipoib_cm_post_receive_nonsrq(dev, rx, i);
 354                 if (ret) {
 355                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
 356                                    "failed for buf %d\n", i);
 357                         ret = -EIO;
 358                         goto err_count;
 359                 }
 360         }
 361
 362         rx->recv_count = ipoib_recvq_size;
 363
 364         return 0;
 365
 366 err_count:
 367         spin_lock_irq(&priv->lock);
 368         --priv->cm.nonsrq_conn_qp;
 369         spin_unlock_irq(&priv->lock);
 370
 371 err_free:
 372         ipoib_cm_free_rx_ring(dev, rx->rx_ring);
 373
 374         return ret;
 375 }
 376
 377 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 378                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 379                              unsigned psn)
 380 {
 381         struct ipoib_dev_priv *priv = netdev_priv(dev);
 382         struct ipoib_cm_data data = {};
 383         struct ib_cm_rep_param rep = {};
 384
 385         data.qpn = cpu_to_be32(priv->qp->qp_num);
 386         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 387
 388         rep.private_data = &data;
 389         rep.private_data_len = sizeof data;
 390         rep.flow_control = 0;
 391         rep.rnr_retry_count = req->rnr_retry_count;
 392         rep.srq = ipoib_cm_has_srq(dev);
 393         rep.qp_num = qp->qp_num;
 394         rep.starting_psn = psn;
 395         return ib_send_cm_rep(cm_id, &rep);
 396 }
 397
 398 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 399 {
 400         struct net_device *dev = cm_id->context;
 401         struct ipoib_dev_priv *priv = netdev_priv(dev);
 402         struct ipoib_cm_rx *p;
 403         unsigned psn;
 404         int ret;
 405
 406         ipoib_dbg(priv, "REQ arrived\n");
 407         p = kzalloc(sizeof *p, GFP_KERNEL);
 408         if (!p)
 409                 return -ENOMEM;
 410         p->dev = dev;
 411         p->id = cm_id;
 412         cm_id->context = p;
 413         p->state = IPOIB_CM_RX_LIVE;
 414         p->jiffies = jiffies;
 415         INIT_LIST_HEAD(&p->list);
 416
 417         p->qp = ipoib_cm_create_rx_qp(dev, p);
 418         if (IS_ERR(p->qp)) {
 419                 ret = PTR_ERR(p->qp);
 420                 goto err_qp;
 421         }
 422
 423         psn = random32() & 0xffffff;
 424         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 425         if (ret)
 426                 goto err_modify;
 427
 428         if (!ipoib_cm_has_srq(dev)) {
 429                 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
 430                 if (ret)
 431                         goto err_modify;
 432         }
 433
 434         spin_lock_irq(&priv->lock);
 435         queue_delayed_work(ipoib_workqueue,
 436                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 437         /* Add this entry to passive ids list head, but do not re-add it
 438          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 439         p->jiffies = jiffies;
 440         if (p->state == IPOIB_CM_RX_LIVE)
 441                 list_move(&p->list, &priv->cm.passive_ids);
 442         spin_unlock_irq(&priv->lock);
 443
 444         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 445         if (ret) {
 446                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 447                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 448                         ipoib_warn(priv, "unable to move qp to error state\n");
 449         }
 450         return 0;
 451
 452 err_modify:
 453         ib_destroy_qp(p->qp);
 454 err_qp:
 455         kfree(p);
 456         return ret;
 457 }
 458
 459 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 460                                struct ib_cm_event *event)
 461 {
 462         struct ipoib_cm_rx *p;
 463         struct ipoib_dev_priv *priv;
 464
 465         switch (event->event) {
 466         case IB_CM_REQ_RECEIVED:
 467                 return ipoib_cm_req_handler(cm_id, event);
 468         case IB_CM_DREQ_RECEIVED:
 469                 p = cm_id->context;
 470                 ib_send_cm_drep(cm_id, NULL, 0);
 471                 /* Fall through */
 472         case IB_CM_REJ_RECEIVED:
 473                 p = cm_id->context;
 474                 priv = netdev_priv(p->dev);
 475                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 476                         ipoib_warn(priv, "unable to move qp to error state\n");
 477                 /* Fall through */
 478         default:
 479                 return 0;
 480         }
 481 }
 482 /* Adjust length of skb with fragments to match received data */
 483 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 484                           unsigned int length, struct sk_buff *toskb)
 485 {
 486         int i, num_frags;
 487         unsigned int size;
 488
 489         /* put header into skb */
 490         size = min(length, hdr_space);
 491         skb->tail += size;
 492         skb->len += size;
 493         length -= size;
 494
 495         num_frags = skb_shinfo(skb)->nr_frags;
 496         for (i = 0; i < num_frags; i++) {
 497                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 498
 499                 if (length == 0) {
 500                         /* don't need this page */
 501                         skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
 502                         --skb_shinfo(skb)->nr_frags;
 503                 } else {
 504                         size = min(length, (unsigned) PAGE_SIZE);
 505
 506                         frag->size = size;
 507                         skb->data_len += size;
 508                         skb->truesize += size;
 509                         skb->len += size;
 510                         length -= size;
 511                 }
 512         }
 513 }
 514
 515 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 516 {
 517         struct ipoib_dev_priv *priv = netdev_priv(dev);
 518         struct ipoib_cm_rx_buf *rx_ring;
 519         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 520         struct sk_buff *skb, *newskb;
 521         struct ipoib_cm_rx *p;
 522         unsigned long flags;
 523         u64 mapping[IPOIB_CM_RX_SG];
 524         int frags;
 525         int has_srq;
 526         struct sk_buff *small_skb;
 527
 528         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 529                        wr_id, wc->status);
 530
 531         if (unlikely(wr_id >= ipoib_recvq_size)) {
 532                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 533                         spin_lock_irqsave(&priv->lock, flags);
 534                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 535                         ipoib_cm_start_rx_drain(priv);
 536                         queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 537                         spin_unlock_irqrestore(&priv->lock, flags);
 538                 } else
 539                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 540                                    wr_id, ipoib_recvq_size);
 541                 return;
 542         }
 543
 544         p = wc->qp->qp_context;
 545
 546         has_srq = ipoib_cm_has_srq(dev);
 547         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 548
 549         skb = rx_ring[wr_id].skb;
 550
 551         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 552                 ipoib_dbg(priv, "cm recv error "
 553                            "(status=%d, wrid=%d vend_err %x)\n",
 554                            wc->status, wr_id, wc->vendor_err);
 555                 ++dev->stats.rx_dropped;
 556                 if (has_srq)
 557                         goto repost;
 558                 else {
 559                         if (!--p->recv_count) {
 560                                 spin_lock_irqsave(&priv->lock, flags);
 561                                 list_move(&p->list, &priv->cm.rx_reap_list);
 562                                 spin_unlock_irqrestore(&priv->lock, flags);
 563                                 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 564                         }
 565                         return;
 566                 }
 567         }
 568
 569         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 570                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 571                         spin_lock_irqsave(&priv->lock, flags);
 572                         p->jiffies = jiffies;
 573                         /* Move this entry to list head, but do not re-add it
 574                          * if it has been moved out of list. */
 575                         if (p->state == IPOIB_CM_RX_LIVE)
 576                                 list_move(&p->list, &priv->cm.passive_ids);
 577                         spin_unlock_irqrestore(&priv->lock, flags);
 578                 }
 579         }
 580
 581         if (wc->byte_len < IPOIB_CM_COPYBREAK) {
 582                 int dlen = wc->byte_len;
 583
 584                 small_skb = dev_alloc_skb(dlen + 12);
 585                 if (small_skb) {
 586                         skb_reserve(small_skb, 12);
 587                         ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
 588                                                    dlen, DMA_FROM_DEVICE);
 589                         skb_copy_from_linear_data(skb, small_skb->data, dlen);
 590                         ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
 591                                                       dlen, DMA_FROM_DEVICE);
 592                         skb_put(small_skb, dlen);
 593                         skb = small_skb;
 594                         goto copied;
 595                 }
 596         }
 597
 598         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 599                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 600
 601         newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
 602         if (unlikely(!newskb)) {
 603                 /*
 604                  * If we can't allocate a new RX buffer, dump
 605                  * this packet and reuse the old buffer.
 606                  */
 607                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 608                 ++dev->stats.rx_dropped;
 609                 goto repost;
 610         }
 611
 612         ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
 613         memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 614
 615         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 616                        wc->byte_len, wc->slid);
 617
 618         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 619
 620 copied:
 621         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 622         skb_reset_mac_header(skb);
 623         skb_pull(skb, IPOIB_ENCAP_LEN);
 624
 625         dev->last_rx = jiffies;
 626         ++dev->stats.rx_packets;
 627         dev->stats.rx_bytes += skb->len;
 628
 629         skb->dev = dev;
 630         /* XXX get correct PACKET_ type here */
 631         skb->pkt_type = PACKET_HOST;
 632         netif_receive_skb(skb);
 633
 634 repost:
 635         if (has_srq) {
 636                 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
 637                         ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
 638                                    "for buf %d\n", wr_id);
 639         } else {
 640                 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p, wr_id))) {
 641                         --p->recv_count;
 642                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
 643                                    "for buf %d\n", wr_id);
 644                 }
 645         }
 646 }
 647
 648 static inline int post_send(struct ipoib_dev_priv *priv,
 649                             struct ipoib_cm_tx *tx,
 650                             unsigned int wr_id,
 651                             u64 addr, int len)
 652 {
 653         struct ib_send_wr *bad_wr;
 654
 655         priv->tx_sge[0].addr          = addr;
 656         priv->tx_sge[0].length        = len;
 657
 658         priv->tx_wr.num_sge     = 1;
 659         priv->tx_wr.wr_id       = wr_id | IPOIB_OP_CM;
 660
 661         return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
 662 }
 663
 664 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 665 {
 666         struct ipoib_dev_priv *priv = netdev_priv(dev);
 667         struct ipoib_tx_buf *tx_req;
 668         u64 addr;
 669
 670         if (unlikely(skb->len > tx->mtu)) {
 671                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 672                            skb->len, tx->mtu);
 673                 ++dev->stats.tx_dropped;
 674                 ++dev->stats.tx_errors;
 675                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 676                 return;
 677         }
 678
 679         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 680                        tx->tx_head, skb->len, tx->qp->qp_num);
 681
 682         /*
 683          * We put the skb into the tx_ring _before_ we call post_send()
 684          * because it's entirely possible that the completion handler will
 685          * run before we execute anything after the post_send().  That
 686          * means we have to make sure everything is properly recorded and
 687          * our state is consistent before we call post_send().
 688          */
 689         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 690         tx_req->skb = skb;
 691         addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
 692         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 693                 ++dev->stats.tx_errors;
 694                 dev_kfree_skb_any(skb);
 695                 return;
 696         }
 697
 698         tx_req->mapping[0] = addr;
 699
 700         if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
 701                                addr, skb->len))) {
 702                 ipoib_warn(priv, "post_send failed\n");
 703                 ++dev->stats.tx_errors;
 704                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
 705                 dev_kfree_skb_any(skb);
 706         } else {
 707                 dev->trans_start = jiffies;
 708                 ++tx->tx_head;
 709
 710                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 711                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 712                                   tx->qp->qp_num);
 713                         netif_stop_queue(dev);
 714                 }
 715         }
 716 }
 717
 718 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 719 {
 720         struct ipoib_dev_priv *priv = netdev_priv(dev);
 721         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 722         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 723         struct ipoib_tx_buf *tx_req;
 724         unsigned long flags;
 725
 726         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 727                        wr_id, wc->status);
 728
 729         if (unlikely(wr_id >= ipoib_sendq_size)) {
 730                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 731                            wr_id, ipoib_sendq_size);
 732                 return;
 733         }
 734
 735         tx_req = &tx->tx_ring[wr_id];
 736
 737         ib_dma_unmap_single(priv->ca, tx_req->mapping[0], tx_req->skb->len, DMA_TO_DEVICE);
 738
 739         /* FIXME: is this right? Shouldn't we only increment on success? */
 740         ++dev->stats.tx_packets;
 741         dev->stats.tx_bytes += tx_req->skb->len;
 742
 743         dev_kfree_skb_any(tx_req->skb);
 744
 745         spin_lock_irqsave(&priv->tx_lock, flags);
 746         ++tx->tx_tail;
 747         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 748             netif_queue_stopped(dev) &&
 749             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 750                 netif_wake_queue(dev);
 751
 752         if (wc->status != IB_WC_SUCCESS &&
 753             wc->status != IB_WC_WR_FLUSH_ERR) {
 754                 struct ipoib_neigh *neigh;
 755
 756                 ipoib_dbg(priv, "failed cm send event "
 757                            "(status=%d, wrid=%d vend_err %x)\n",
 758                            wc->status, wr_id, wc->vendor_err);
 759
 760                 spin_lock(&priv->lock);
 761                 neigh = tx->neigh;
 762
 763                 if (neigh) {
 764                         neigh->cm = NULL;
 765                         list_del(&neigh->list);
 766                         if (neigh->ah)
 767                                 ipoib_put_ah(neigh->ah);
 768                         ipoib_neigh_free(dev, neigh);
 769
 770                         tx->neigh = NULL;
 771                 }
 772
 773                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 774                         list_move(&tx->list, &priv->cm.reap_list);
 775                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 776                 }
 777
 778                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 779
 780                 spin_unlock(&priv->lock);
 781         }
 782
 783         spin_unlock_irqrestore(&priv->tx_lock, flags);
 784 }
 785
 786 int ipoib_cm_dev_open(struct net_device *dev)
 787 {
 788         struct ipoib_dev_priv *priv = netdev_priv(dev);
 789         int ret;
 790
 791         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 792                 return 0;
 793
 794         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 795         if (IS_ERR(priv->cm.id)) {
 796                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 797                 ret = PTR_ERR(priv->cm.id);
 798                 goto err_cm;
 799         }
 800
 801         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 802                            0, NULL);
 803         if (ret) {
 804                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 805                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 806                 goto err_listen;
 807         }
 808
 809         return 0;
 810
 811 err_listen:
 812         ib_destroy_cm_id(priv->cm.id);
 813 err_cm:
 814         priv->cm.id = NULL;
 815         return ret;
 816 }
 817
 818 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
 819 {
 820         struct ipoib_dev_priv *priv = netdev_priv(dev);
 821         struct ipoib_cm_rx *rx, *n;
 822         LIST_HEAD(list);
 823
 824         spin_lock_irq(&priv->lock);
 825         list_splice_init(&priv->cm.rx_reap_list, &list);
 826         spin_unlock_irq(&priv->lock);
 827
 828         list_for_each_entry_safe(rx, n, &list, list) {
 829                 ib_destroy_cm_id(rx->id);
 830                 ib_destroy_qp(rx->qp);
 831                 if (!ipoib_cm_has_srq(dev)) {
 832                         ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
 833                         spin_lock_irq(&priv->lock);
 834                         --priv->cm.nonsrq_conn_qp;
 835                         spin_unlock_irq(&priv->lock);
 836                 }
 837                 kfree(rx);
 838         }
 839 }
 840
 841 void ipoib_cm_dev_stop(struct net_device *dev)
 842 {
 843         struct ipoib_dev_priv *priv = netdev_priv(dev);
 844         struct ipoib_cm_rx *p;
 845         unsigned long begin;
 846         int ret;
 847
 848         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 849                 return;
 850
 851         ib_destroy_cm_id(priv->cm.id);
 852         priv->cm.id = NULL;
 853
 854         spin_lock_irq(&priv->lock);
 855         while (!list_empty(&priv->cm.passive_ids)) {
 856                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 857                 list_move(&p->list, &priv->cm.rx_error_list);
 858                 p->state = IPOIB_CM_RX_ERROR;
 859                 spin_unlock_irq(&priv->lock);
 860                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 861                 if (ret)
 862                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 863                 spin_lock_irq(&priv->lock);
 864         }
 865
 866         /* Wait for all RX to be drained */
 867         begin = jiffies;
 868
 869         while (!list_empty(&priv->cm.rx_error_list) ||
 870                !list_empty(&priv->cm.rx_flush_list) ||
 871                !list_empty(&priv->cm.rx_drain_list)) {
 872                 if (time_after(jiffies, begin + 5 * HZ)) {
 873                         ipoib_warn(priv, "RX drain timing out\n");
 874
 875                         /*
 876                          * assume the HW is wedged and just free up everything.
 877                          */
 878                         list_splice_init(&priv->cm.rx_flush_list,
 879                                          &priv->cm.rx_reap_list);
 880                         list_splice_init(&priv->cm.rx_error_list,
 881                                          &priv->cm.rx_reap_list);
 882                         list_splice_init(&priv->cm.rx_drain_list,
 883                                          &priv->cm.rx_reap_list);
 884                         break;
 885                 }
 886                 spin_unlock_irq(&priv->lock);
 887                 msleep(1);
 888                 ipoib_drain_cq(dev);
 889                 spin_lock_irq(&priv->lock);
 890         }
 891
 892         spin_unlock_irq(&priv->lock);
 893
 894         ipoib_cm_free_rx_reap_list(dev);
 895
 896         cancel_delayed_work(&priv->cm.stale_task);
 897 }
 898
 899 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 900 {
 901         struct ipoib_cm_tx *p = cm_id->context;
 902         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 903         struct ipoib_cm_data *data = event->private_data;
 904         struct sk_buff_head skqueue;
 905         struct ib_qp_attr qp_attr;
 906         int qp_attr_mask, ret;
 907         struct sk_buff *skb;
 908
 909         p->mtu = be32_to_cpu(data->mtu);
 910
 911         if (p->mtu <= IPOIB_ENCAP_LEN) {
 912                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 913                            p->mtu, IPOIB_ENCAP_LEN);
 914                 return -EINVAL;
 915         }
 916
 917         qp_attr.qp_state = IB_QPS_RTR;
 918         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 919         if (ret) {
 920                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 921                 return ret;
 922         }
 923
 924         qp_attr.rq_psn = 0 /* FIXME */;
 925         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 926         if (ret) {
 927                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 928                 return ret;
 929         }
 930
 931         qp_attr.qp_state = IB_QPS_RTS;
 932         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 933         if (ret) {
 934                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 935                 return ret;
 936         }
 937         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 938         if (ret) {
 939                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 940                 return ret;
 941         }
 942
 943         skb_queue_head_init(&skqueue);
 944
 945         spin_lock_irq(&priv->lock);
 946         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 947         if (p->neigh)
 948                 while ((skb = __skb_dequeue(&p->neigh->queue)))
 949                         __skb_queue_tail(&skqueue, skb);
 950         spin_unlock_irq(&priv->lock);
 951
 952         while ((skb = __skb_dequeue(&skqueue))) {
 953                 skb->dev = p->dev;
 954                 if (dev_queue_xmit(skb))
 955                         ipoib_warn(priv, "dev_queue_xmit failed "
 956                                    "to requeue packet\n");
 957         }
 958
 959         ret = ib_send_cm_rtu(cm_id, NULL, 0);
 960         if (ret) {
 961                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
 962                 return ret;
 963         }
 964         return 0;
 965 }
 966
 967 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
 968 {
 969         struct ipoib_dev_priv *priv = netdev_priv(dev);
 970         struct ib_qp_init_attr attr = {
 971                 .send_cq                = priv->recv_cq,
 972                 .recv_cq                = priv->recv_cq,
 973                 .srq                    = priv->cm.srq,
 974                 .cap.max_send_wr        = ipoib_sendq_size,
 975                 .cap.max_send_sge       = 1,
 976                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
 977                 .qp_type                = IB_QPT_RC,
 978                 .qp_context             = tx
 979         };
 980
 981         return ib_create_qp(priv->pd, &attr);
 982 }
 983
 984 static int ipoib_cm_send_req(struct net_device *dev,
 985                              struct ib_cm_id *id, struct ib_qp *qp,
 986                              u32 qpn,
 987                              struct ib_sa_path_rec *pathrec)
 988 {
 989         struct ipoib_dev_priv *priv = netdev_priv(dev);
 990         struct ipoib_cm_data data = {};
 991         struct ib_cm_req_param req = {};
 992
 993         data.qpn = cpu_to_be32(priv->qp->qp_num);
 994         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 995
 996         req.primary_path                = pathrec;
 997         req.alternate_path              = NULL;
 998         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
 999         req.qp_num                      = qp->qp_num;
1000         req.qp_type                     = qp->qp_type;
1001         req.private_data                = &data;
1002         req.private_data_len            = sizeof data;
1003         req.flow_control                = 0;
1004
1005         req.starting_psn                = 0; /* FIXME */
1006
1007         /*
1008          * Pick some arbitrary defaults here; we could make these
1009          * module parameters if anyone cared about setting them.
1010          */
1011         req.responder_resources         = 4;
1012         req.remote_cm_response_timeout  = 20;
1013         req.local_cm_response_timeout   = 20;
1014         req.retry_count                 = 0; /* RFC draft warns against retries */
1015         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
1016         req.max_cm_retries              = 15;
1017         req.srq                         = ipoib_cm_has_srq(dev);
1018         return ib_send_cm_req(id, &req);
1019 }
1020
1021 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1022                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1023 {
1024         struct ipoib_dev_priv *priv = netdev_priv(dev);
1025         struct ib_qp_attr qp_attr;
1026         int qp_attr_mask, ret;
1027         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1028         if (ret) {
1029                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1030                 return ret;
1031         }
1032
1033         qp_attr.qp_state = IB_QPS_INIT;
1034         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1035         qp_attr.port_num = priv->port;
1036         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1037
1038         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1039         if (ret) {
1040                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1041                 return ret;
1042         }
1043         return 0;
1044 }
1045
1046 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1047                             struct ib_sa_path_rec *pathrec)
1048 {
1049         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1050         int ret;
1051
1052         p->tx_ring = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1053         if (!p->tx_ring) {
1054                 ipoib_warn(priv, "failed to allocate tx ring\n");
1055                 ret = -ENOMEM;
1056                 goto err_tx;
1057         }
1058         memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1059
1060         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1061         if (IS_ERR(p->qp)) {
1062                 ret = PTR_ERR(p->qp);
1063                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1064                 goto err_qp;
1065         }
1066
1067         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1068         if (IS_ERR(p->id)) {
1069                 ret = PTR_ERR(p->id);
1070                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1071                 goto err_id;
1072         }
1073
1074         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1075         if (ret) {
1076                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1077                 goto err_modify;
1078         }
1079
1080         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1081         if (ret) {
1082                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1083                 goto err_send_cm;
1084         }
1085
1086         ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
1087                   p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
1088
1089         return 0;
1090
1091 err_send_cm:
1092 err_modify:
1093         ib_destroy_cm_id(p->id);
1094 err_id:
1095         p->id = NULL;
1096         ib_destroy_qp(p->qp);
1097 err_qp:
1098         p->qp = NULL;
1099         vfree(p->tx_ring);
1100 err_tx:
1101         return ret;
1102 }
1103
1104 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1105 {
1106         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1107         struct ipoib_tx_buf *tx_req;
1108         unsigned long flags;
1109         unsigned long begin;
1110
1111         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1112                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1113
1114         if (p->id)
1115                 ib_destroy_cm_id(p->id);
1116
1117         if (p->tx_ring) {
1118                 /* Wait for all sends to complete */
1119                 begin = jiffies;
1120                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1121                         if (time_after(jiffies, begin + 5 * HZ)) {
1122                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1123                                            p->tx_head - p->tx_tail);
1124                                 goto timeout;
1125                         }
1126
1127                         msleep(1);
1128                 }
1129         }
1130
1131 timeout:
1132
1133         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1134                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1135                 ib_dma_unmap_single(priv->ca, tx_req->mapping[0], tx_req->skb->len,
1136                                     DMA_TO_DEVICE);
1137                 dev_kfree_skb_any(tx_req->skb);
1138                 ++p->tx_tail;
1139                 spin_lock_irqsave(&priv->tx_lock, flags);
1140                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1141                     netif_queue_stopped(p->dev) &&
1142                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1143                         netif_wake_queue(p->dev);
1144                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1145         }
1146
1147         if (p->qp)
1148                 ib_destroy_qp(p->qp);
1149
1150         vfree(p->tx_ring);
1151         kfree(p);
1152 }
1153
1154 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1155                                struct ib_cm_event *event)
1156 {
1157         struct ipoib_cm_tx *tx = cm_id->context;
1158         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1159         struct net_device *dev = priv->dev;
1160         struct ipoib_neigh *neigh;
1161         int ret;
1162
1163         switch (event->event) {
1164         case IB_CM_DREQ_RECEIVED:
1165                 ipoib_dbg(priv, "DREQ received.\n");
1166                 ib_send_cm_drep(cm_id, NULL, 0);
1167                 break;
1168         case IB_CM_REP_RECEIVED:
1169                 ipoib_dbg(priv, "REP received.\n");
1170                 ret = ipoib_cm_rep_handler(cm_id, event);
1171                 if (ret)
1172                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1173                                        NULL, 0, NULL, 0);
1174                 break;
1175         case IB_CM_REQ_ERROR:
1176         case IB_CM_REJ_RECEIVED:
1177         case IB_CM_TIMEWAIT_EXIT:
1178                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1179                 spin_lock_irq(&priv->tx_lock);
1180                 spin_lock(&priv->lock);
1181                 neigh = tx->neigh;
1182
1183                 if (neigh) {
1184                         neigh->cm = NULL;
1185                         list_del(&neigh->list);
1186                         if (neigh->ah)
1187                                 ipoib_put_ah(neigh->ah);
1188                         ipoib_neigh_free(dev, neigh);
1189
1190                         tx->neigh = NULL;
1191                 }
1192
1193                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1194                         list_move(&tx->list, &priv->cm.reap_list);
1195                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1196                 }
1197
1198                 spin_unlock(&priv->lock);
1199                 spin_unlock_irq(&priv->tx_lock);
1200                 break;
1201         default:
1202                 break;
1203         }
1204
1205         return 0;
1206 }
1207
1208 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1209                                        struct ipoib_neigh *neigh)
1210 {
1211         struct ipoib_dev_priv *priv = netdev_priv(dev);
1212         struct ipoib_cm_tx *tx;
1213
1214         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1215         if (!tx)
1216                 return NULL;
1217
1218         neigh->cm = tx;
1219         tx->neigh = neigh;
1220         tx->path = path;
1221         tx->dev = dev;
1222         list_add(&tx->list, &priv->cm.start_list);
1223         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1224         queue_work(ipoib_workqueue, &priv->cm.start_task);
1225         return tx;
1226 }
1227
1228 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1229 {
1230         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1231         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1232                 list_move(&tx->list, &priv->cm.reap_list);
1233                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1234                 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1235                           IPOIB_GID_ARG(tx->neigh->dgid));
1236                 tx->neigh = NULL;
1237         }
1238 }
1239
1240 static void ipoib_cm_tx_start(struct work_struct *work)
1241 {
1242         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1243                                                    cm.start_task);
1244         struct net_device *dev = priv->dev;
1245         struct ipoib_neigh *neigh;
1246         struct ipoib_cm_tx *p;
1247         unsigned long flags;
1248         int ret;
1249
1250         struct ib_sa_path_rec pathrec;
1251         u32 qpn;
1252
1253         spin_lock_irqsave(&priv->tx_lock, flags);
1254         spin_lock(&priv->lock);
1255         while (!list_empty(&priv->cm.start_list)) {
1256                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1257                 list_del_init(&p->list);
1258                 neigh = p->neigh;
1259                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1260                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1261                 spin_unlock(&priv->lock);
1262                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1263                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1264                 spin_lock_irqsave(&priv->tx_lock, flags);
1265                 spin_lock(&priv->lock);
1266                 if (ret) {
1267                         neigh = p->neigh;
1268                         if (neigh) {
1269                                 neigh->cm = NULL;
1270                                 list_del(&neigh->list);
1271                                 if (neigh->ah)
1272                                         ipoib_put_ah(neigh->ah);
1273                                 ipoib_neigh_free(dev, neigh);
1274                         }
1275                         list_del(&p->list);
1276                         kfree(p);
1277                 }
1278         }
1279         spin_unlock(&priv->lock);
1280         spin_unlock_irqrestore(&priv->tx_lock, flags);
1281 }
1282
1283 static void ipoib_cm_tx_reap(struct work_struct *work)
1284 {
1285         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1286                                                    cm.reap_task);
1287         struct ipoib_cm_tx *p;
1288
1289         spin_lock_irq(&priv->tx_lock);
1290         spin_lock(&priv->lock);
1291         while (!list_empty(&priv->cm.reap_list)) {
1292                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1293                 list_del(&p->list);
1294                 spin_unlock(&priv->lock);
1295                 spin_unlock_irq(&priv->tx_lock);
1296                 ipoib_cm_tx_destroy(p);
1297                 spin_lock_irq(&priv->tx_lock);
1298                 spin_lock(&priv->lock);
1299         }
1300         spin_unlock(&priv->lock);
1301         spin_unlock_irq(&priv->tx_lock);
1302 }
1303
1304 static void ipoib_cm_skb_reap(struct work_struct *work)
1305 {
1306         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1307                                                    cm.skb_task);
1308         struct sk_buff *skb;
1309
1310         unsigned mtu = priv->mcast_mtu;
1311
1312         spin_lock_irq(&priv->tx_lock);
1313         spin_lock(&priv->lock);
1314         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1315                 spin_unlock(&priv->lock);
1316                 spin_unlock_irq(&priv->tx_lock);
1317                 if (skb->protocol == htons(ETH_P_IP))
1318                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1319 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1320                 else if (skb->protocol == htons(ETH_P_IPV6))
1321                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, priv->dev);
1322 #endif
1323                 dev_kfree_skb_any(skb);
1324                 spin_lock_irq(&priv->tx_lock);
1325                 spin_lock(&priv->lock);
1326         }
1327         spin_unlock(&priv->lock);
1328         spin_unlock_irq(&priv->tx_lock);
1329 }
1330
1331 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1332                            unsigned int mtu)
1333 {
1334         struct ipoib_dev_priv *priv = netdev_priv(dev);
1335         int e = skb_queue_empty(&priv->cm.skb_queue);
1336
1337         if (skb->dst)
1338                 skb->dst->ops->update_pmtu(skb->dst, mtu);
1339
1340         skb_queue_tail(&priv->cm.skb_queue, skb);
1341         if (e)
1342                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1343 }
1344
1345 static void ipoib_cm_rx_reap(struct work_struct *work)
1346 {
1347         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1348                                                 cm.rx_reap_task)->dev);
1349 }
1350
1351 static void ipoib_cm_stale_task(struct work_struct *work)
1352 {
1353         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1354                                                    cm.stale_task.work);
1355         struct ipoib_cm_rx *p;
1356         int ret;
1357
1358         spin_lock_irq(&priv->lock);
1359         while (!list_empty(&priv->cm.passive_ids)) {
1360                 /* List is sorted by LRU, start from tail,
1361                  * stop when we see a recently used entry */
1362                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1363                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1364                         break;
1365                 list_move(&p->list, &priv->cm.rx_error_list);
1366                 p->state = IPOIB_CM_RX_ERROR;
1367                 spin_unlock_irq(&priv->lock);
1368                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1369                 if (ret)
1370                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1371                 spin_lock_irq(&priv->lock);
1372         }
1373
1374         if (!list_empty(&priv->cm.passive_ids))
1375                 queue_delayed_work(ipoib_workqueue,
1376                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1377         spin_unlock_irq(&priv->lock);
1378 }
1379
1380
1381 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1382                          char *buf)
1383 {
1384         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1385
1386         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1387                 return sprintf(buf, "connected\n");
1388         else
1389                 return sprintf(buf, "datagram\n");
1390 }
1391
1392 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1393                         const char *buf, size_t count)
1394 {
1395         struct net_device *dev = to_net_dev(d);
1396         struct ipoib_dev_priv *priv = netdev_priv(dev);
1397
1398         /* flush paths if we switch modes so that connections are restarted */
1399         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1400                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1401                 ipoib_warn(priv, "enabling connected mode "
1402                            "will cause multicast packet drops\n");
1403
1404                 dev->features &= ~(NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO);
1405                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1406
1407                 ipoib_flush_paths(dev);
1408                 return count;
1409         }
1410
1411         if (!strcmp(buf, "datagram\n")) {
1412                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1413                 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1414                 ipoib_flush_paths(dev);
1415
1416                 if (test_bit(IPOIB_FLAG_CSUM, &priv->flags)) {
1417                         dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG;
1418                         if (priv->hca_caps & IB_DEVICE_UD_TSO)
1419                                 dev->features |= NETIF_F_TSO;
1420                 }
1421
1422                 return count;
1423         }
1424
1425         return -EINVAL;
1426 }
1427
1428 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1429
1430 int ipoib_cm_add_mode_attr(struct net_device *dev)
1431 {
1432         return device_create_file(&dev->dev, &dev_attr_mode);
1433 }
1434
1435 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1436 {
1437         struct ipoib_dev_priv *priv = netdev_priv(dev);
1438         struct ib_srq_init_attr srq_init_attr = {
1439                 .attr = {
1440                         .max_wr  = ipoib_recvq_size,
1441                         .max_sge = max_sge
1442                 }
1443         };
1444
1445         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1446         if (IS_ERR(priv->cm.srq)) {
1447                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1448                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1449                                priv->ca->name, PTR_ERR(priv->cm.srq));
1450                 priv->cm.srq = NULL;
1451                 return;
1452         }
1453
1454         priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1455                                     GFP_KERNEL);
1456         if (!priv->cm.srq_ring) {
1457                 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1458                        priv->ca->name, ipoib_recvq_size);
1459                 ib_destroy_srq(priv->cm.srq);
1460                 priv->cm.srq = NULL;
1461         }
1462 }
1463
1464 int ipoib_cm_dev_init(struct net_device *dev)
1465 {
1466         struct ipoib_dev_priv *priv = netdev_priv(dev);
1467         int i, ret;
1468         struct ib_device_attr attr;
1469
1470         INIT_LIST_HEAD(&priv->cm.passive_ids);
1471         INIT_LIST_HEAD(&priv->cm.reap_list);
1472         INIT_LIST_HEAD(&priv->cm.start_list);
1473         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1474         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1475         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1476         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1477         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1478         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1479         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1480         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1481         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1482
1483         skb_queue_head_init(&priv->cm.skb_queue);
1484
1485         ret = ib_query_device(priv->ca, &attr);
1486         if (ret) {
1487                 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1488                 return ret;
1489         }
1490
1491         ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1492
1493         attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1494         ipoib_cm_create_srq(dev, attr.max_srq_sge);
1495         if (ipoib_cm_has_srq(dev)) {
1496                 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1497                 priv->cm.num_frags  = attr.max_srq_sge;
1498                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1499                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1500         } else {
1501                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1502                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1503         }
1504
1505         for (i = 0; i < priv->cm.num_frags; ++i)
1506                 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1507
1508         priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1509         for (i = 1; i < priv->cm.num_frags; ++i)
1510                 priv->cm.rx_sge[i].length = PAGE_SIZE;
1511         priv->cm.rx_wr.next = NULL;
1512         priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1513         priv->cm.rx_wr.num_sge = priv->cm.num_frags;
1514
1515         if (ipoib_cm_has_srq(dev)) {
1516                 for (i = 0; i < ipoib_recvq_size; ++i) {
1517                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1518                                                    priv->cm.num_frags - 1,
1519                                                    priv->cm.srq_ring[i].mapping)) {
1520                                 ipoib_warn(priv, "failed to allocate "
1521                                            "receive buffer %d\n", i);
1522                                 ipoib_cm_dev_cleanup(dev);
1523                                 return -ENOMEM;
1524                         }
1525
1526                         if (ipoib_cm_post_receive_srq(dev, i)) {
1527                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1528                                            "failed for buf %d\n", i);
1529                                 ipoib_cm_dev_cleanup(dev);
1530                                 return -EIO;
1531                         }
1532                 }
1533         }
1534
1535         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1536         return 0;
1537 }
1538
1539 void ipoib_cm_dev_cleanup(struct net_device *dev)
1540 {
1541         struct ipoib_dev_priv *priv = netdev_priv(dev);
1542         int ret;
1543
1544         if (!priv->cm.srq)
1545                 return;
1546
1547         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1548
1549         ret = ib_destroy_srq(priv->cm.srq);
1550         if (ret)
1551                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1552
1553         priv->cm.srq = NULL;
1554         if (!priv->cm.srq_ring)
1555                 return;
1556
1557         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1558         priv->cm.srq_ring = NULL;
1559 }