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