git.oblomov.eu Git - 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         LIST_HEAD(list);
 828         int ret;
 829
 830         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 831                 return;
 832
 833         ib_destroy_cm_id(priv->cm.id);
 834         priv->cm.id = NULL;
 835
 836         spin_lock_irq(&priv->lock);
 837         while (!list_empty(&priv->cm.passive_ids)) {
 838                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 839                 list_move(&p->list, &priv->cm.rx_error_list);
 840                 p->state = IPOIB_CM_RX_ERROR;
 841                 spin_unlock_irq(&priv->lock);
 842                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 843                 if (ret)
 844                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 845                 spin_lock_irq(&priv->lock);
 846         }
 847
 848         /* Wait for all RX to be drained */
 849         begin = jiffies;
 850
 851         while (!list_empty(&priv->cm.rx_error_list) ||
 852                !list_empty(&priv->cm.rx_flush_list) ||
 853                !list_empty(&priv->cm.rx_drain_list)) {
 854                 if (time_after(jiffies, begin + 5 * HZ)) {
 855                         ipoib_warn(priv, "RX drain timing out\n");
 856
 857                         /*
 858                          * assume the HW is wedged and just free up everything.
 859                          */
 860                         list_splice_init(&priv->cm.rx_flush_list, &list);
 861                         list_splice_init(&priv->cm.rx_error_list, &list);
 862                         list_splice_init(&priv->cm.rx_drain_list, &list);
 863                         break;
 864                 }
 865                 spin_unlock_irq(&priv->lock);
 866                 msleep(1);
 867                 ipoib_drain_cq(dev);
 868                 spin_lock_irq(&priv->lock);
 869         }
 870
 871         spin_unlock_irq(&priv->lock);
 872
 873         ipoib_cm_free_rx_reap_list(dev);
 874
 875         cancel_delayed_work(&priv->cm.stale_task);
 876 }
 877
 878 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 879 {
 880         struct ipoib_cm_tx *p = cm_id->context;
 881         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 882         struct ipoib_cm_data *data = event->private_data;
 883         struct sk_buff_head skqueue;
 884         struct ib_qp_attr qp_attr;
 885         int qp_attr_mask, ret;
 886         struct sk_buff *skb;
 887
 888         p->mtu = be32_to_cpu(data->mtu);
 889
 890         if (p->mtu <= IPOIB_ENCAP_LEN) {
 891                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 892                            p->mtu, IPOIB_ENCAP_LEN);
 893                 return -EINVAL;
 894         }
 895
 896         qp_attr.qp_state = IB_QPS_RTR;
 897         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 898         if (ret) {
 899                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 900                 return ret;
 901         }
 902
 903         qp_attr.rq_psn = 0 /* FIXME */;
 904         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 905         if (ret) {
 906                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 907                 return ret;
 908         }
 909
 910         qp_attr.qp_state = IB_QPS_RTS;
 911         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 912         if (ret) {
 913                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 914                 return ret;
 915         }
 916         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 917         if (ret) {
 918                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 919                 return ret;
 920         }
 921
 922         skb_queue_head_init(&skqueue);
 923
 924         spin_lock_irq(&priv->lock);
 925         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 926         if (p->neigh)
 927                 while ((skb = __skb_dequeue(&p->neigh->queue)))
 928                         __skb_queue_tail(&skqueue, skb);
 929         spin_unlock_irq(&priv->lock);
 930
 931         while ((skb = __skb_dequeue(&skqueue))) {
 932                 skb->dev = p->dev;
 933                 if (dev_queue_xmit(skb))
 934                         ipoib_warn(priv, "dev_queue_xmit failed "
 935                                    "to requeue packet\n");
 936         }
 937
 938         ret = ib_send_cm_rtu(cm_id, NULL, 0);
 939         if (ret) {
 940                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
 941                 return ret;
 942         }
 943         return 0;
 944 }
 945
 946 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
 947 {
 948         struct ipoib_dev_priv *priv = netdev_priv(dev);
 949         struct ib_qp_init_attr attr = {
 950                 .send_cq                = priv->cq,
 951                 .recv_cq                = priv->cq,
 952                 .srq                    = priv->cm.srq,
 953                 .cap.max_send_wr        = ipoib_sendq_size,
 954                 .cap.max_send_sge       = 1,
 955                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
 956                 .qp_type                = IB_QPT_RC,
 957                 .qp_context             = tx
 958         };
 959
 960         return ib_create_qp(priv->pd, &attr);
 961 }
 962
 963 static int ipoib_cm_send_req(struct net_device *dev,
 964                              struct ib_cm_id *id, struct ib_qp *qp,
 965                              u32 qpn,
 966                              struct ib_sa_path_rec *pathrec)
 967 {
 968         struct ipoib_dev_priv *priv = netdev_priv(dev);
 969         struct ipoib_cm_data data = {};
 970         struct ib_cm_req_param req = {};
 971
 972         data.qpn = cpu_to_be32(priv->qp->qp_num);
 973         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 974
 975         req.primary_path                = pathrec;
 976         req.alternate_path              = NULL;
 977         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
 978         req.qp_num                      = qp->qp_num;
 979         req.qp_type                     = qp->qp_type;
 980         req.private_data                = &data;
 981         req.private_data_len            = sizeof data;
 982         req.flow_control                = 0;
 983
 984         req.starting_psn                = 0; /* FIXME */
 985
 986         /*
 987          * Pick some arbitrary defaults here; we could make these
 988          * module parameters if anyone cared about setting them.
 989          */
 990         req.responder_resources         = 4;
 991         req.remote_cm_response_timeout  = 20;
 992         req.local_cm_response_timeout   = 20;
 993         req.retry_count                 = 0; /* RFC draft warns against retries */
 994         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
 995         req.max_cm_retries              = 15;
 996         req.srq                         = ipoib_cm_has_srq(dev);
 997         return ib_send_cm_req(id, &req);
 998 }
 999
1000 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1001                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1002 {
1003         struct ipoib_dev_priv *priv = netdev_priv(dev);
1004         struct ib_qp_attr qp_attr;
1005         int qp_attr_mask, ret;
1006         ret = ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1007         if (ret) {
1008                 ipoib_warn(priv, "pkey 0x%x not in cache: %d\n", priv->pkey, ret);
1009                 return ret;
1010         }
1011
1012         qp_attr.qp_state = IB_QPS_INIT;
1013         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1014         qp_attr.port_num = priv->port;
1015         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1016
1017         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1018         if (ret) {
1019                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1020                 return ret;
1021         }
1022         return 0;
1023 }
1024
1025 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1026                             struct ib_sa_path_rec *pathrec)
1027 {
1028         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1029         int ret;
1030
1031         p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring,
1032                                 GFP_KERNEL);
1033         if (!p->tx_ring) {
1034                 ipoib_warn(priv, "failed to allocate tx ring\n");
1035                 ret = -ENOMEM;
1036                 goto err_tx;
1037         }
1038
1039         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1040         if (IS_ERR(p->qp)) {
1041                 ret = PTR_ERR(p->qp);
1042                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1043                 goto err_qp;
1044         }
1045
1046         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1047         if (IS_ERR(p->id)) {
1048                 ret = PTR_ERR(p->id);
1049                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1050                 goto err_id;
1051         }
1052
1053         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1054         if (ret) {
1055                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1056                 goto err_modify;
1057         }
1058
1059         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1060         if (ret) {
1061                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1062                 goto err_send_cm;
1063         }
1064
1065         ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
1066                   p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
1067
1068         return 0;
1069
1070 err_send_cm:
1071 err_modify:
1072         ib_destroy_cm_id(p->id);
1073 err_id:
1074         p->id = NULL;
1075         ib_destroy_qp(p->qp);
1076 err_qp:
1077         p->qp = NULL;
1078 err_tx:
1079         return ret;
1080 }
1081
1082 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1083 {
1084         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1085         struct ipoib_tx_buf *tx_req;
1086         unsigned long flags;
1087         unsigned long begin;
1088
1089         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1090                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1091
1092         if (p->id)
1093                 ib_destroy_cm_id(p->id);
1094
1095         if (p->tx_ring) {
1096                 /* Wait for all sends to complete */
1097                 begin = jiffies;
1098                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1099                         if (time_after(jiffies, begin + 5 * HZ)) {
1100                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1101                                            p->tx_head - p->tx_tail);
1102                                 goto timeout;
1103                         }
1104
1105                         msleep(1);
1106                 }
1107         }
1108
1109 timeout:
1110
1111         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1112                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1113                 ib_dma_unmap_single(priv->ca, tx_req->mapping[0], tx_req->skb->len,
1114                                     DMA_TO_DEVICE);
1115                 dev_kfree_skb_any(tx_req->skb);
1116                 ++p->tx_tail;
1117                 spin_lock_irqsave(&priv->tx_lock, flags);
1118                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1119                     netif_queue_stopped(p->dev) &&
1120                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1121                         netif_wake_queue(p->dev);
1122                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1123         }
1124
1125         if (p->qp)
1126                 ib_destroy_qp(p->qp);
1127
1128         kfree(p->tx_ring);
1129         kfree(p);
1130 }
1131
1132 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1133                                struct ib_cm_event *event)
1134 {
1135         struct ipoib_cm_tx *tx = cm_id->context;
1136         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1137         struct net_device *dev = priv->dev;
1138         struct ipoib_neigh *neigh;
1139         int ret;
1140
1141         switch (event->event) {
1142         case IB_CM_DREQ_RECEIVED:
1143                 ipoib_dbg(priv, "DREQ received.\n");
1144                 ib_send_cm_drep(cm_id, NULL, 0);
1145                 break;
1146         case IB_CM_REP_RECEIVED:
1147                 ipoib_dbg(priv, "REP received.\n");
1148                 ret = ipoib_cm_rep_handler(cm_id, event);
1149                 if (ret)
1150                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1151                                        NULL, 0, NULL, 0);
1152                 break;
1153         case IB_CM_REQ_ERROR:
1154         case IB_CM_REJ_RECEIVED:
1155         case IB_CM_TIMEWAIT_EXIT:
1156                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1157                 spin_lock_irq(&priv->tx_lock);
1158                 spin_lock(&priv->lock);
1159                 neigh = tx->neigh;
1160
1161                 if (neigh) {
1162                         neigh->cm = NULL;
1163                         list_del(&neigh->list);
1164                         if (neigh->ah)
1165                                 ipoib_put_ah(neigh->ah);
1166                         ipoib_neigh_free(dev, neigh);
1167
1168                         tx->neigh = NULL;
1169                 }
1170
1171                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1172                         list_move(&tx->list, &priv->cm.reap_list);
1173                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1174                 }
1175
1176                 spin_unlock(&priv->lock);
1177                 spin_unlock_irq(&priv->tx_lock);
1178                 break;
1179         default:
1180                 break;
1181         }
1182
1183         return 0;
1184 }
1185
1186 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1187                                        struct ipoib_neigh *neigh)
1188 {
1189         struct ipoib_dev_priv *priv = netdev_priv(dev);
1190         struct ipoib_cm_tx *tx;
1191
1192         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1193         if (!tx)
1194                 return NULL;
1195
1196         neigh->cm = tx;
1197         tx->neigh = neigh;
1198         tx->path = path;
1199         tx->dev = dev;
1200         list_add(&tx->list, &priv->cm.start_list);
1201         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1202         queue_work(ipoib_workqueue, &priv->cm.start_task);
1203         return tx;
1204 }
1205
1206 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1207 {
1208         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1209         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1210                 list_move(&tx->list, &priv->cm.reap_list);
1211                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1212                 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1213                           IPOIB_GID_ARG(tx->neigh->dgid));
1214                 tx->neigh = NULL;
1215         }
1216 }
1217
1218 static void ipoib_cm_tx_start(struct work_struct *work)
1219 {
1220         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1221                                                    cm.start_task);
1222         struct net_device *dev = priv->dev;
1223         struct ipoib_neigh *neigh;
1224         struct ipoib_cm_tx *p;
1225         unsigned long flags;
1226         int ret;
1227
1228         struct ib_sa_path_rec pathrec;
1229         u32 qpn;
1230
1231         spin_lock_irqsave(&priv->tx_lock, flags);
1232         spin_lock(&priv->lock);
1233         while (!list_empty(&priv->cm.start_list)) {
1234                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1235                 list_del_init(&p->list);
1236                 neigh = p->neigh;
1237                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1238                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1239                 spin_unlock(&priv->lock);
1240                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1241                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1242                 spin_lock_irqsave(&priv->tx_lock, flags);
1243                 spin_lock(&priv->lock);
1244                 if (ret) {
1245                         neigh = p->neigh;
1246                         if (neigh) {
1247                                 neigh->cm = NULL;
1248                                 list_del(&neigh->list);
1249                                 if (neigh->ah)
1250                                         ipoib_put_ah(neigh->ah);
1251                                 ipoib_neigh_free(dev, neigh);
1252                         }
1253                         list_del(&p->list);
1254                         kfree(p);
1255                 }
1256         }
1257         spin_unlock(&priv->lock);
1258         spin_unlock_irqrestore(&priv->tx_lock, flags);
1259 }
1260
1261 static void ipoib_cm_tx_reap(struct work_struct *work)
1262 {
1263         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1264                                                    cm.reap_task);
1265         struct ipoib_cm_tx *p;
1266
1267         spin_lock_irq(&priv->tx_lock);
1268         spin_lock(&priv->lock);
1269         while (!list_empty(&priv->cm.reap_list)) {
1270                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1271                 list_del(&p->list);
1272                 spin_unlock(&priv->lock);
1273                 spin_unlock_irq(&priv->tx_lock);
1274                 ipoib_cm_tx_destroy(p);
1275                 spin_lock_irq(&priv->tx_lock);
1276                 spin_lock(&priv->lock);
1277         }
1278         spin_unlock(&priv->lock);
1279         spin_unlock_irq(&priv->tx_lock);
1280 }
1281
1282 static void ipoib_cm_skb_reap(struct work_struct *work)
1283 {
1284         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1285                                                    cm.skb_task);
1286         struct sk_buff *skb;
1287
1288         unsigned mtu = priv->mcast_mtu;
1289
1290         spin_lock_irq(&priv->tx_lock);
1291         spin_lock(&priv->lock);
1292         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1293                 spin_unlock(&priv->lock);
1294                 spin_unlock_irq(&priv->tx_lock);
1295                 if (skb->protocol == htons(ETH_P_IP))
1296                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1297 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1298                 else if (skb->protocol == htons(ETH_P_IPV6))
1299                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, priv->dev);
1300 #endif
1301                 dev_kfree_skb_any(skb);
1302                 spin_lock_irq(&priv->tx_lock);
1303                 spin_lock(&priv->lock);
1304         }
1305         spin_unlock(&priv->lock);
1306         spin_unlock_irq(&priv->tx_lock);
1307 }
1308
1309 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1310                            unsigned int mtu)
1311 {
1312         struct ipoib_dev_priv *priv = netdev_priv(dev);
1313         int e = skb_queue_empty(&priv->cm.skb_queue);
1314
1315         if (skb->dst)
1316                 skb->dst->ops->update_pmtu(skb->dst, mtu);
1317
1318         skb_queue_tail(&priv->cm.skb_queue, skb);
1319         if (e)
1320                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1321 }
1322
1323 static void ipoib_cm_rx_reap(struct work_struct *work)
1324 {
1325         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1326                                                 cm.rx_reap_task)->dev);
1327 }
1328
1329 static void ipoib_cm_stale_task(struct work_struct *work)
1330 {
1331         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1332                                                    cm.stale_task.work);
1333         struct ipoib_cm_rx *p;
1334         int ret;
1335
1336         spin_lock_irq(&priv->lock);
1337         while (!list_empty(&priv->cm.passive_ids)) {
1338                 /* List is sorted by LRU, start from tail,
1339                  * stop when we see a recently used entry */
1340                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1341                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1342                         break;
1343                 list_move(&p->list, &priv->cm.rx_error_list);
1344                 p->state = IPOIB_CM_RX_ERROR;
1345                 spin_unlock_irq(&priv->lock);
1346                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1347                 if (ret)
1348                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1349                 spin_lock_irq(&priv->lock);
1350         }
1351
1352         if (!list_empty(&priv->cm.passive_ids))
1353                 queue_delayed_work(ipoib_workqueue,
1354                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1355         spin_unlock_irq(&priv->lock);
1356 }
1357
1358
1359 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1360                          char *buf)
1361 {
1362         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1363
1364         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1365                 return sprintf(buf, "connected\n");
1366         else
1367                 return sprintf(buf, "datagram\n");
1368 }
1369
1370 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1371                         const char *buf, size_t count)
1372 {
1373         struct net_device *dev = to_net_dev(d);
1374         struct ipoib_dev_priv *priv = netdev_priv(dev);
1375
1376         /* flush paths if we switch modes so that connections are restarted */
1377         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1378                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1379                 ipoib_warn(priv, "enabling connected mode "
1380                            "will cause multicast packet drops\n");
1381                 ipoib_flush_paths(dev);
1382                 return count;
1383         }
1384
1385         if (!strcmp(buf, "datagram\n")) {
1386                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1387                 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1388                 ipoib_flush_paths(dev);
1389                 return count;
1390         }
1391
1392         return -EINVAL;
1393 }
1394
1395 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1396
1397 int ipoib_cm_add_mode_attr(struct net_device *dev)
1398 {
1399         return device_create_file(&dev->dev, &dev_attr_mode);
1400 }
1401
1402 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1403 {
1404         struct ipoib_dev_priv *priv = netdev_priv(dev);
1405         struct ib_srq_init_attr srq_init_attr = {
1406                 .attr = {
1407                         .max_wr  = ipoib_recvq_size,
1408                         .max_sge = max_sge
1409                 }
1410         };
1411
1412         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1413         if (IS_ERR(priv->cm.srq)) {
1414                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1415                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1416                                priv->ca->name, PTR_ERR(priv->cm.srq));
1417                 priv->cm.srq = NULL;
1418                 return;
1419         }
1420
1421         priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1422                                     GFP_KERNEL);
1423         if (!priv->cm.srq_ring) {
1424                 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1425                        priv->ca->name, ipoib_recvq_size);
1426                 ib_destroy_srq(priv->cm.srq);
1427                 priv->cm.srq = NULL;
1428         }
1429 }
1430
1431 int ipoib_cm_dev_init(struct net_device *dev)
1432 {
1433         struct ipoib_dev_priv *priv = netdev_priv(dev);
1434         int i, ret;
1435         struct ib_device_attr attr;
1436
1437         INIT_LIST_HEAD(&priv->cm.passive_ids);
1438         INIT_LIST_HEAD(&priv->cm.reap_list);
1439         INIT_LIST_HEAD(&priv->cm.start_list);
1440         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1441         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1442         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1443         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1444         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1445         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1446         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1447         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1448         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1449
1450         skb_queue_head_init(&priv->cm.skb_queue);
1451
1452         ret = ib_query_device(priv->ca, &attr);
1453         if (ret) {
1454                 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1455                 return ret;
1456         }
1457
1458         ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1459
1460         attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1461         ipoib_cm_create_srq(dev, attr.max_srq_sge);
1462         if (ipoib_cm_has_srq(dev)) {
1463                 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1464                 priv->cm.num_frags  = attr.max_srq_sge;
1465                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1466                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1467         } else {
1468                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1469                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1470         }
1471
1472         for (i = 0; i < priv->cm.num_frags; ++i)
1473                 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1474
1475         priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1476         for (i = 1; i < priv->cm.num_frags; ++i)
1477                 priv->cm.rx_sge[i].length = PAGE_SIZE;
1478         priv->cm.rx_wr.next = NULL;
1479         priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1480         priv->cm.rx_wr.num_sge = priv->cm.num_frags;
1481
1482         if (ipoib_cm_has_srq(dev)) {
1483                 for (i = 0; i < ipoib_recvq_size; ++i) {
1484                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1485                                                    priv->cm.num_frags - 1,
1486                                                    priv->cm.srq_ring[i].mapping)) {
1487                                 ipoib_warn(priv, "failed to allocate "
1488                                            "receive buffer %d\n", i);
1489                                 ipoib_cm_dev_cleanup(dev);
1490                                 return -ENOMEM;
1491                         }
1492
1493                         if (ipoib_cm_post_receive_srq(dev, i)) {
1494                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1495                                            "failed for buf %d\n", i);
1496                                 ipoib_cm_dev_cleanup(dev);
1497                                 return -EIO;
1498                         }
1499                 }
1500         }
1501
1502         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1503         return 0;
1504 }
1505
1506 void ipoib_cm_dev_cleanup(struct net_device *dev)
1507 {
1508         struct ipoib_dev_priv *priv = netdev_priv(dev);
1509         int ret;
1510
1511         if (!priv->cm.srq)
1512                 return;
1513
1514         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1515
1516         ret = ib_destroy_srq(priv->cm.srq);
1517         if (ret)
1518                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1519
1520         priv->cm.srq = NULL;
1521         if (!priv->cm.srq_ring)
1522                 return;
1523
1524         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1525         priv->cm.srq_ring = NULL;
1526 }