drivers/infiniband/ulp/ipoib/ipoib_cm.c

   1 /*
   2  * Copyright (c) 2006 Mellanox Technologies. All rights reserved
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <rdma/ib_cm.h>
  34 #include <rdma/ib_cache.h>
  35 #include <net/dst.h>
  36 #include <net/icmp.h>
  37 #include <linux/icmpv6.h>
  38 #include <linux/delay.h>
  39 #include <linux/vmalloc.h>
  40
  41 #include "ipoib.h"
  42
  43 int ipoib_max_conn_qp = 128;
  44
  45 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
  46 MODULE_PARM_DESC(max_nonsrq_conn_qp,
  47                  "Max number of connected-mode QPs per interface "
  48                  "(applied only if shared receive queue is not available)");
  49
  50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  51 static int data_debug_level;
  52
  53 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  54 MODULE_PARM_DESC(cm_data_debug_level,
  55                  "Enable data path debug tracing for connected mode if > 0");
  56 #endif
  57
  58 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  59
  60 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  61 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  62 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  63 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  64
  65 static struct ib_qp_attr ipoib_cm_err_attr = {
  66         .qp_state = IB_QPS_ERR
  67 };
  68
  69 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
  70
  71 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  72         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  73         .opcode = IB_WR_SEND,
  74 };
  75
  76 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  77                                struct ib_cm_event *event);
  78
  79 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  80                                   u64 mapping[IPOIB_CM_RX_SG])
  81 {
  82         int i;
  83
  84         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  85
  86         for (i = 0; i < frags; ++i)
  87                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  88 }
  89
  90 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
  91 {
  92         struct ipoib_dev_priv *priv = netdev_priv(dev);
  93         struct ib_recv_wr *bad_wr;
  94         int i, ret;
  95
  96         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
  97
  98         for (i = 0; i < priv->cm.num_frags; ++i)
  99                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 100
 101         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 102         if (unlikely(ret)) {
 103                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 104                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 105                                       priv->cm.srq_ring[id].mapping);
 106                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 107                 priv->cm.srq_ring[id].skb = NULL;
 108         }
 109
 110         return ret;
 111 }
 112
 113 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 114                                         struct ipoib_cm_rx *rx, int id)
 115 {
 116         struct ipoib_dev_priv *priv = netdev_priv(dev);
 117         struct ib_recv_wr *bad_wr;
 118         int i, ret;
 119
 120         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 121
 122         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 123                 priv->cm.rx_sge[i].addr = rx->rx_ring[id].mapping[i];
 124
 125         ret = ib_post_recv(rx->qp, &priv->cm.rx_wr, &bad_wr);
 126         if (unlikely(ret)) {
 127                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 128                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 129                                       rx->rx_ring[id].mapping);
 130                 dev_kfree_skb_any(rx->rx_ring[id].skb);
 131                 rx->rx_ring[id].skb = NULL;
 132         }
 133
 134         return ret;
 135 }
 136
 137 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
 138                                              struct ipoib_cm_rx_buf *rx_ring,
 139                                              int id, int frags,
 140                                              u64 mapping[IPOIB_CM_RX_SG])
 141 {
 142         struct ipoib_dev_priv *priv = netdev_priv(dev);
 143         struct sk_buff *skb;
 144         int i;
 145
 146         skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
 147         if (unlikely(!skb))
 148                 return NULL;
 149
 150         /*
 151          * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
 152          * IP header to a multiple of 16.
 153          */
 154         skb_reserve(skb, 12);
 155
 156         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 157                                        DMA_FROM_DEVICE);
 158         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 159                 dev_kfree_skb_any(skb);
 160                 return NULL;
 161         }
 162
 163         for (i = 0; i < frags; i++) {
 164                 struct page *page = alloc_page(GFP_ATOMIC);
 165
 166                 if (!page)
 167                         goto partial_error;
 168                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 169
 170                 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
 171                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 172                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 173                         goto partial_error;
 174         }
 175
 176         rx_ring[id].skb = skb;
 177         return skb;
 178
 179 partial_error:
 180
 181         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 182
 183         for (; i > 0; --i)
 184                 ib_dma_unmap_single(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
 185
 186         dev_kfree_skb_any(skb);
 187         return NULL;
 188 }
 189
 190 static void ipoib_cm_free_rx_ring(struct net_device *dev,
 191                                   struct ipoib_cm_rx_buf *rx_ring)
 192 {
 193         struct ipoib_dev_priv *priv = netdev_priv(dev);
 194         int i;
 195
 196         for (i = 0; i < ipoib_recvq_size; ++i)
 197                 if (rx_ring[i].skb) {
 198                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 199                                               rx_ring[i].mapping);
 200                         dev_kfree_skb_any(rx_ring[i].skb);
 201                 }
 202
 203         kfree(rx_ring);
 204 }
 205
 206 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
 207 {
 208         struct ib_send_wr *bad_wr;
 209         struct ipoib_cm_rx *p;
 210
 211         /* We only reserved 1 extra slot in CQ for drain WRs, so
 212          * make sure we have at most 1 outstanding WR. */
 213         if (list_empty(&priv->cm.rx_flush_list) ||
 214             !list_empty(&priv->cm.rx_drain_list))
 215                 return;
 216
 217         /*
 218          * QPs on flush list are error state.  This way, a "flush
 219          * error" WC will be immediately generated for each WR we post.
 220          */
 221         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 222         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 223                 ipoib_warn(priv, "failed to post drain wr\n");
 224
 225         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 226 }
 227
 228 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 229 {
 230         struct ipoib_cm_rx *p = ctx;
 231         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 232         unsigned long flags;
 233
 234         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 235                 return;
 236
 237         spin_lock_irqsave(&priv->lock, flags);
 238         list_move(&p->list, &priv->cm.rx_flush_list);
 239         p->state = IPOIB_CM_RX_FLUSH;
 240         ipoib_cm_start_rx_drain(priv);
 241         spin_unlock_irqrestore(&priv->lock, flags);
 242 }
 243
 244 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 245                                            struct ipoib_cm_rx *p)
 246 {
 247         struct ipoib_dev_priv *priv = netdev_priv(dev);
 248         struct ib_qp_init_attr attr = {
 249                 .event_handler = ipoib_cm_rx_event_handler,
 250                 .send_cq = priv->recv_cq, /* For drain WR */
 251                 .recv_cq = priv->recv_cq,
 252                 .srq = priv->cm.srq,
 253                 .cap.max_send_wr = 1, /* For drain WR */
 254                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 255                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 256                 .qp_type = IB_QPT_RC,
 257                 .qp_context = p,
 258         };
 259
 260         if (!ipoib_cm_has_srq(dev)) {
 261                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 262                 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
 263         }
 264
 265         return ib_create_qp(priv->pd, &attr);
 266 }
 267
 268 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 269                                  struct ib_cm_id *cm_id, struct ib_qp *qp,
 270                                  unsigned psn)
 271 {
 272         struct ipoib_dev_priv *priv = netdev_priv(dev);
 273         struct ib_qp_attr qp_attr;
 274         int qp_attr_mask, ret;
 275
 276         qp_attr.qp_state = IB_QPS_INIT;
 277         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 278         if (ret) {
 279                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 280                 return ret;
 281         }
 282         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 283         if (ret) {
 284                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 285                 return ret;
 286         }
 287         qp_attr.qp_state = IB_QPS_RTR;
 288         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 289         if (ret) {
 290                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 291                 return ret;
 292         }
 293         qp_attr.rq_psn = psn;
 294         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 295         if (ret) {
 296                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 297                 return ret;
 298         }
 299
 300         /*
 301          * Current Mellanox HCA firmware won't generate completions
 302          * with error for drain WRs unless the QP has been moved to
 303          * RTS first. This work-around leaves a window where a QP has
 304          * moved to error asynchronously, but this will eventually get
 305          * fixed in firmware, so let's not error out if modify QP
 306          * fails.
 307          */
 308         qp_attr.qp_state = IB_QPS_RTS;
 309         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 310         if (ret) {
 311                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 312                 return 0;
 313         }
 314         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 315         if (ret) {
 316                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 317                 return 0;
 318         }
 319
 320         return 0;
 321 }
 322
 323 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
 324                                    struct ipoib_cm_rx *rx)
 325 {
 326         struct ipoib_dev_priv *priv = netdev_priv(dev);
 327         int ret;
 328         int i;
 329
 330         rx->rx_ring = kcalloc(ipoib_recvq_size, sizeof *rx->rx_ring, GFP_KERNEL);
 331         if (!rx->rx_ring)
 332                 return -ENOMEM;
 333
 334         spin_lock_irq(&priv->lock);
 335
 336         if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
 337                 spin_unlock_irq(&priv->lock);
 338                 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
 339                 ret = -EINVAL;
 340                 goto err_free;
 341         } else
 342                 ++priv->cm.nonsrq_conn_qp;
 343
 344         spin_unlock_irq(&priv->lock);
 345
 346         for (i = 0; i < ipoib_recvq_size; ++i) {
 347                 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
 348                                            rx->rx_ring[i].mapping)) {
 349                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 350                                 ret = -ENOMEM;
 351                                 goto err_count;
 352                         }
 353                 ret = ipoib_cm_post_receive_nonsrq(dev, rx, i);
 354                 if (ret) {
 355                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
 356                                    "failed for buf %d\n", i);
 357                         ret = -EIO;
 358                         goto err_count;
 359                 }
 360         }
 361
 362         rx->recv_count = ipoib_recvq_size;
 363
 364         return 0;
 365
 366 err_count:
 367         spin_lock_irq(&priv->lock);
 368         --priv->cm.nonsrq_conn_qp;
 369         spin_unlock_irq(&priv->lock);
 370
 371 err_free:
 372         ipoib_cm_free_rx_ring(dev, rx->rx_ring);
 373
 374         return ret;
 375 }
 376
 377 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 378                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 379                              unsigned psn)
 380 {
 381         struct ipoib_dev_priv *priv = netdev_priv(dev);
 382         struct ipoib_cm_data data = {};
 383         struct ib_cm_rep_param rep = {};
 384
 385         data.qpn = cpu_to_be32(priv->qp->qp_num);
 386         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 387
 388         rep.private_data = &data;
 389         rep.private_data_len = sizeof data;
 390         rep.flow_control = 0;
 391         rep.rnr_retry_count = req->rnr_retry_count;
 392         rep.srq = ipoib_cm_has_srq(dev);
 393         rep.qp_num = qp->qp_num;
 394         rep.starting_psn = psn;
 395         return ib_send_cm_rep(cm_id, &rep);
 396 }
 397
 398 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 399 {
 400         struct net_device *dev = cm_id->context;
 401         struct ipoib_dev_priv *priv = netdev_priv(dev);
 402         struct ipoib_cm_rx *p;
 403         unsigned psn;
 404         int ret;
 405
 406         ipoib_dbg(priv, "REQ arrived\n");
 407         p = kzalloc(sizeof *p, GFP_KERNEL);
 408         if (!p)
 409                 return -ENOMEM;
 410         p->dev = dev;
 411         p->id = cm_id;
 412         cm_id->context = p;
 413         p->state = IPOIB_CM_RX_LIVE;
 414         p->jiffies = jiffies;
 415         INIT_LIST_HEAD(&p->list);
 416
 417         p->qp = ipoib_cm_create_rx_qp(dev, p);
 418         if (IS_ERR(p->qp)) {
 419                 ret = PTR_ERR(p->qp);
 420                 goto err_qp;
 421         }
 422
 423         psn = random32() & 0xffffff;
 424         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 425         if (ret)
 426                 goto err_modify;
 427
 428         if (!ipoib_cm_has_srq(dev)) {
 429                 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
 430                 if (ret)
 431                         goto err_modify;
 432         }
 433
 434         spin_lock_irq(&priv->lock);
 435         queue_delayed_work(ipoib_workqueue,
 436                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 437         /* Add this entry to passive ids list head, but do not re-add it
 438          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 439         p->jiffies = jiffies;
 440         if (p->state == IPOIB_CM_RX_LIVE)
 441                 list_move(&p->list, &priv->cm.passive_ids);
 442         spin_unlock_irq(&priv->lock);
 443
 444         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 445         if (ret) {
 446                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 447                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 448                         ipoib_warn(priv, "unable to move qp to error state\n");
 449         }
 450         return 0;
 451
 452 err_modify:
 453         ib_destroy_qp(p->qp);
 454 err_qp:
 455         kfree(p);
 456         return ret;
 457 }
 458
 459 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 460                                struct ib_cm_event *event)
 461 {
 462         struct ipoib_cm_rx *p;
 463         struct ipoib_dev_priv *priv;
 464
 465         switch (event->event) {
 466         case IB_CM_REQ_RECEIVED:
 467                 return ipoib_cm_req_handler(cm_id, event);
 468         case IB_CM_DREQ_RECEIVED:
 469                 p = cm_id->context;
 470                 ib_send_cm_drep(cm_id, NULL, 0);
 471                 /* Fall through */
 472         case IB_CM_REJ_RECEIVED:
 473                 p = cm_id->context;
 474                 priv = netdev_priv(p->dev);
 475                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 476                         ipoib_warn(priv, "unable to move qp to error state\n");
 477                 /* Fall through */
 478         default:
 479                 return 0;
 480         }
 481 }
 482 /* Adjust length of skb with fragments to match received data */
 483 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 484                           unsigned int length, struct sk_buff *toskb)
 485 {
 486         int i, num_frags;
 487         unsigned int size;
 488
 489         /* put header into skb */
 490         size = min(length, hdr_space);
 491         skb->tail += size;
 492         skb->len += size;
 493         length -= size;
 494
 495         num_frags = skb_shinfo(skb)->nr_frags;
 496         for (i = 0; i < num_frags; i++) {
 497                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 498
 499                 if (length == 0) {
 500                         /* don't need this page */
 501                         skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
 502                         --skb_shinfo(skb)->nr_frags;
 503                 } else {
 504                         size = min(length, (unsigned) PAGE_SIZE);
 505
 506                         frag->size = size;
 507                         skb->data_len += size;
 508                         skb->truesize += size;
 509                         skb->len += size;
 510                         length -= size;
 511                 }
 512         }
 513 }
 514
 515 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 516 {
 517         struct ipoib_dev_priv *priv = netdev_priv(dev);
 518         struct ipoib_cm_rx_buf *rx_ring;
 519         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 520         struct sk_buff *skb, *newskb;
 521         struct ipoib_cm_rx *p;
 522         unsigned long flags;
 523         u64 mapping[IPOIB_CM_RX_SG];
 524         int frags;
 525         int has_srq;
 526
 527         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 528                        wr_id, wc->status);
 529
 530         if (unlikely(wr_id >= ipoib_recvq_size)) {
 531                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 532                         spin_lock_irqsave(&priv->lock, flags);
 533                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 534                         ipoib_cm_start_rx_drain(priv);
 535                         queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 536                         spin_unlock_irqrestore(&priv->lock, flags);
 537                 } else
 538                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 539                                    wr_id, ipoib_recvq_size);
 540                 return;
 541         }
 542
 543         p = wc->qp->qp_context;
 544
 545         has_srq = ipoib_cm_has_srq(dev);
 546         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 547
 548         skb = rx_ring[wr_id].skb;
 549
 550         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 551                 ipoib_dbg(priv, "cm recv error "
 552                            "(status=%d, wrid=%d vend_err %x)\n",
 553                            wc->status, wr_id, wc->vendor_err);
 554                 ++dev->stats.rx_dropped;
 555                 if (has_srq)
 556                         goto repost;
 557                 else {
 558                         if (!--p->recv_count) {
 559                                 spin_lock_irqsave(&priv->lock, flags);
 560                                 list_move(&p->list, &priv->cm.rx_reap_list);
 561                                 spin_unlock_irqrestore(&priv->lock, flags);
 562                                 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 563                         }
 564                         return;
 565                 }
 566         }
 567
 568         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 569                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 570                         spin_lock_irqsave(&priv->lock, flags);
 571                         p->jiffies = jiffies;
 572                         /* Move this entry to list head, but do not re-add it
 573                          * if it has been moved out of list. */
 574                         if (p->state == IPOIB_CM_RX_LIVE)
 575                                 list_move(&p->list, &priv->cm.passive_ids);
 576                         spin_unlock_irqrestore(&priv->lock, flags);
 577                 }
 578         }
 579
 580         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 581                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 582
 583         newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
 584         if (unlikely(!newskb)) {
 585                 /*
 586                  * If we can't allocate a new RX buffer, dump
 587                  * this packet and reuse the old buffer.
 588                  */
 589                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 590                 ++dev->stats.rx_dropped;
 591                 goto repost;
 592         }
 593
 594         ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
 595         memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 596
 597         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 598                        wc->byte_len, wc->slid);
 599
 600         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 601
 602         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 603         skb_reset_mac_header(skb);
 604         skb_pull(skb, IPOIB_ENCAP_LEN);
 605
 606         dev->last_rx = jiffies;
 607         ++dev->stats.rx_packets;
 608         dev->stats.rx_bytes += skb->len;
 609
 610         skb->dev = dev;
 611         /* XXX get correct PACKET_ type here */
 612         skb->pkt_type = PACKET_HOST;
 613         netif_receive_skb(skb);
 614
 615 repost:
 616         if (has_srq) {
 617                 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
 618                         ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
 619                                    "for buf %d\n", wr_id);
 620         } else {
 621                 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p, wr_id))) {
 622                         --p->recv_count;
 623                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
 624                                    "for buf %d\n", wr_id);
 625                 }
 626         }
 627 }
 628
 629 static inline int post_send(struct ipoib_dev_priv *priv,
 630                             struct ipoib_cm_tx *tx,
 631                             unsigned int wr_id,
 632                             u64 addr, int len)
 633 {
 634         struct ib_send_wr *bad_wr;
 635
 636         priv->tx_sge[0].addr          = addr;
 637         priv->tx_sge[0].length        = len;
 638
 639         priv->tx_wr.num_sge     = 1;
 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->recv_cq,
 953                 .recv_cq                = priv->recv_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_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1009         if (ret) {
1010                 ipoib_warn(priv, "pkey 0x%x not found: %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 = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1034         if (!p->tx_ring) {
1035                 ipoib_warn(priv, "failed to allocate tx ring\n");
1036                 ret = -ENOMEM;
1037                 goto err_tx;
1038         }
1039         memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
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         vfree(p->tx_ring);
1081 err_tx:
1082         return ret;
1083 }
1084
1085 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1086 {
1087         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1088         struct ipoib_tx_buf *tx_req;
1089         unsigned long flags;
1090         unsigned long begin;
1091
1092         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1093                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1094
1095         if (p->id)
1096                 ib_destroy_cm_id(p->id);
1097
1098         if (p->tx_ring) {
1099                 /* Wait for all sends to complete */
1100                 begin = jiffies;
1101                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1102                         if (time_after(jiffies, begin + 5 * HZ)) {
1103                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1104                                            p->tx_head - p->tx_tail);
1105                                 goto timeout;
1106                         }
1107
1108                         msleep(1);
1109                 }
1110         }
1111
1112 timeout:
1113
1114         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1115                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1116                 ib_dma_unmap_single(priv->ca, tx_req->mapping[0], tx_req->skb->len,
1117                                     DMA_TO_DEVICE);
1118                 dev_kfree_skb_any(tx_req->skb);
1119                 ++p->tx_tail;
1120                 spin_lock_irqsave(&priv->tx_lock, flags);
1121                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1122                     netif_queue_stopped(p->dev) &&
1123                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1124                         netif_wake_queue(p->dev);
1125                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1126         }
1127
1128         if (p->qp)
1129                 ib_destroy_qp(p->qp);
1130
1131         vfree(p->tx_ring);
1132         kfree(p);
1133 }
1134
1135 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1136                                struct ib_cm_event *event)
1137 {
1138         struct ipoib_cm_tx *tx = cm_id->context;
1139         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1140         struct net_device *dev = priv->dev;
1141         struct ipoib_neigh *neigh;
1142         int ret;
1143
1144         switch (event->event) {
1145         case IB_CM_DREQ_RECEIVED:
1146                 ipoib_dbg(priv, "DREQ received.\n");
1147                 ib_send_cm_drep(cm_id, NULL, 0);
1148                 break;
1149         case IB_CM_REP_RECEIVED:
1150                 ipoib_dbg(priv, "REP received.\n");
1151                 ret = ipoib_cm_rep_handler(cm_id, event);
1152                 if (ret)
1153                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1154                                        NULL, 0, NULL, 0);
1155                 break;
1156         case IB_CM_REQ_ERROR:
1157         case IB_CM_REJ_RECEIVED:
1158         case IB_CM_TIMEWAIT_EXIT:
1159                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1160                 spin_lock_irq(&priv->tx_lock);
1161                 spin_lock(&priv->lock);
1162                 neigh = tx->neigh;
1163
1164                 if (neigh) {
1165                         neigh->cm = NULL;
1166                         list_del(&neigh->list);
1167                         if (neigh->ah)
1168                                 ipoib_put_ah(neigh->ah);
1169                         ipoib_neigh_free(dev, neigh);
1170
1171                         tx->neigh = NULL;
1172                 }
1173
1174                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1175                         list_move(&tx->list, &priv->cm.reap_list);
1176                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1177                 }
1178
1179                 spin_unlock(&priv->lock);
1180                 spin_unlock_irq(&priv->tx_lock);
1181                 break;
1182         default:
1183                 break;
1184         }
1185
1186         return 0;
1187 }
1188
1189 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1190                                        struct ipoib_neigh *neigh)
1191 {
1192         struct ipoib_dev_priv *priv = netdev_priv(dev);
1193         struct ipoib_cm_tx *tx;
1194
1195         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1196         if (!tx)
1197                 return NULL;
1198
1199         neigh->cm = tx;
1200         tx->neigh = neigh;
1201         tx->path = path;
1202         tx->dev = dev;
1203         list_add(&tx->list, &priv->cm.start_list);
1204         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1205         queue_work(ipoib_workqueue, &priv->cm.start_task);
1206         return tx;
1207 }
1208
1209 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1210 {
1211         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1212         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1213                 list_move(&tx->list, &priv->cm.reap_list);
1214                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1215                 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1216                           IPOIB_GID_ARG(tx->neigh->dgid));
1217                 tx->neigh = NULL;
1218         }
1219 }
1220
1221 static void ipoib_cm_tx_start(struct work_struct *work)
1222 {
1223         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1224                                                    cm.start_task);
1225         struct net_device *dev = priv->dev;
1226         struct ipoib_neigh *neigh;
1227         struct ipoib_cm_tx *p;
1228         unsigned long flags;
1229         int ret;
1230
1231         struct ib_sa_path_rec pathrec;
1232         u32 qpn;
1233
1234         spin_lock_irqsave(&priv->tx_lock, flags);
1235         spin_lock(&priv->lock);
1236         while (!list_empty(&priv->cm.start_list)) {
1237                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1238                 list_del_init(&p->list);
1239                 neigh = p->neigh;
1240                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1241                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1242                 spin_unlock(&priv->lock);
1243                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1244                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1245                 spin_lock_irqsave(&priv->tx_lock, flags);
1246                 spin_lock(&priv->lock);
1247                 if (ret) {
1248                         neigh = p->neigh;
1249                         if (neigh) {
1250                                 neigh->cm = NULL;
1251                                 list_del(&neigh->list);
1252                                 if (neigh->ah)
1253                                         ipoib_put_ah(neigh->ah);
1254                                 ipoib_neigh_free(dev, neigh);
1255                         }
1256                         list_del(&p->list);
1257                         kfree(p);
1258                 }
1259         }
1260         spin_unlock(&priv->lock);
1261         spin_unlock_irqrestore(&priv->tx_lock, flags);
1262 }
1263
1264 static void ipoib_cm_tx_reap(struct work_struct *work)
1265 {
1266         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1267                                                    cm.reap_task);
1268         struct ipoib_cm_tx *p;
1269
1270         spin_lock_irq(&priv->tx_lock);
1271         spin_lock(&priv->lock);
1272         while (!list_empty(&priv->cm.reap_list)) {
1273                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1274                 list_del(&p->list);
1275                 spin_unlock(&priv->lock);
1276                 spin_unlock_irq(&priv->tx_lock);
1277                 ipoib_cm_tx_destroy(p);
1278                 spin_lock_irq(&priv->tx_lock);
1279                 spin_lock(&priv->lock);
1280         }
1281         spin_unlock(&priv->lock);
1282         spin_unlock_irq(&priv->tx_lock);
1283 }
1284
1285 static void ipoib_cm_skb_reap(struct work_struct *work)
1286 {
1287         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1288                                                    cm.skb_task);
1289         struct sk_buff *skb;
1290
1291         unsigned mtu = priv->mcast_mtu;
1292
1293         spin_lock_irq(&priv->tx_lock);
1294         spin_lock(&priv->lock);
1295         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1296                 spin_unlock(&priv->lock);
1297                 spin_unlock_irq(&priv->tx_lock);
1298                 if (skb->protocol == htons(ETH_P_IP))
1299                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1300 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1301                 else if (skb->protocol == htons(ETH_P_IPV6))
1302                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, priv->dev);
1303 #endif
1304                 dev_kfree_skb_any(skb);
1305                 spin_lock_irq(&priv->tx_lock);
1306                 spin_lock(&priv->lock);
1307         }
1308         spin_unlock(&priv->lock);
1309         spin_unlock_irq(&priv->tx_lock);
1310 }
1311
1312 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1313                            unsigned int mtu)
1314 {
1315         struct ipoib_dev_priv *priv = netdev_priv(dev);
1316         int e = skb_queue_empty(&priv->cm.skb_queue);
1317
1318         if (skb->dst)
1319                 skb->dst->ops->update_pmtu(skb->dst, mtu);
1320
1321         skb_queue_tail(&priv->cm.skb_queue, skb);
1322         if (e)
1323                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1324 }
1325
1326 static void ipoib_cm_rx_reap(struct work_struct *work)
1327 {
1328         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1329                                                 cm.rx_reap_task)->dev);
1330 }
1331
1332 static void ipoib_cm_stale_task(struct work_struct *work)
1333 {
1334         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1335                                                    cm.stale_task.work);
1336         struct ipoib_cm_rx *p;
1337         int ret;
1338
1339         spin_lock_irq(&priv->lock);
1340         while (!list_empty(&priv->cm.passive_ids)) {
1341                 /* List is sorted by LRU, start from tail,
1342                  * stop when we see a recently used entry */
1343                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1344                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1345                         break;
1346                 list_move(&p->list, &priv->cm.rx_error_list);
1347                 p->state = IPOIB_CM_RX_ERROR;
1348                 spin_unlock_irq(&priv->lock);
1349                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1350                 if (ret)
1351                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1352                 spin_lock_irq(&priv->lock);
1353         }
1354
1355         if (!list_empty(&priv->cm.passive_ids))
1356                 queue_delayed_work(ipoib_workqueue,
1357                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1358         spin_unlock_irq(&priv->lock);
1359 }
1360
1361
1362 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1363                          char *buf)
1364 {
1365         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1366
1367         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1368                 return sprintf(buf, "connected\n");
1369         else
1370                 return sprintf(buf, "datagram\n");
1371 }
1372
1373 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1374                         const char *buf, size_t count)
1375 {
1376         struct net_device *dev = to_net_dev(d);
1377         struct ipoib_dev_priv *priv = netdev_priv(dev);
1378
1379         /* flush paths if we switch modes so that connections are restarted */
1380         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1381                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1382                 ipoib_warn(priv, "enabling connected mode "
1383                            "will cause multicast packet drops\n");
1384
1385                 dev->features &= ~(NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO);
1386                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1387
1388                 ipoib_flush_paths(dev);
1389                 return count;
1390         }
1391
1392         if (!strcmp(buf, "datagram\n")) {
1393                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1394                 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1395                 ipoib_flush_paths(dev);
1396
1397                 if (test_bit(IPOIB_FLAG_CSUM, &priv->flags)) {
1398                         dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG;
1399                         if (priv->hca_caps & IB_DEVICE_UD_TSO)
1400                                 dev->features |= NETIF_F_TSO;
1401                 }
1402
1403                 return count;
1404         }
1405
1406         return -EINVAL;
1407 }
1408
1409 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1410
1411 int ipoib_cm_add_mode_attr(struct net_device *dev)
1412 {
1413         return device_create_file(&dev->dev, &dev_attr_mode);
1414 }
1415
1416 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1417 {
1418         struct ipoib_dev_priv *priv = netdev_priv(dev);
1419         struct ib_srq_init_attr srq_init_attr = {
1420                 .attr = {
1421                         .max_wr  = ipoib_recvq_size,
1422                         .max_sge = max_sge
1423                 }
1424         };
1425
1426         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1427         if (IS_ERR(priv->cm.srq)) {
1428                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1429                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1430                                priv->ca->name, PTR_ERR(priv->cm.srq));
1431                 priv->cm.srq = NULL;
1432                 return;
1433         }
1434
1435         priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1436                                     GFP_KERNEL);
1437         if (!priv->cm.srq_ring) {
1438                 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1439                        priv->ca->name, ipoib_recvq_size);
1440                 ib_destroy_srq(priv->cm.srq);
1441                 priv->cm.srq = NULL;
1442         }
1443 }
1444
1445 int ipoib_cm_dev_init(struct net_device *dev)
1446 {
1447         struct ipoib_dev_priv *priv = netdev_priv(dev);
1448         int i, ret;
1449         struct ib_device_attr attr;
1450
1451         INIT_LIST_HEAD(&priv->cm.passive_ids);
1452         INIT_LIST_HEAD(&priv->cm.reap_list);
1453         INIT_LIST_HEAD(&priv->cm.start_list);
1454         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1455         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1456         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1457         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1458         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1459         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1460         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1461         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1462         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1463
1464         skb_queue_head_init(&priv->cm.skb_queue);
1465
1466         ret = ib_query_device(priv->ca, &attr);
1467         if (ret) {
1468                 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1469                 return ret;
1470         }
1471
1472         ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1473
1474         attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1475         ipoib_cm_create_srq(dev, attr.max_srq_sge);
1476         if (ipoib_cm_has_srq(dev)) {
1477                 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1478                 priv->cm.num_frags  = attr.max_srq_sge;
1479                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1480                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1481         } else {
1482                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1483                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1484         }
1485
1486         for (i = 0; i < priv->cm.num_frags; ++i)
1487                 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1488
1489         priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1490         for (i = 1; i < priv->cm.num_frags; ++i)
1491                 priv->cm.rx_sge[i].length = PAGE_SIZE;
1492         priv->cm.rx_wr.next = NULL;
1493         priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1494         priv->cm.rx_wr.num_sge = priv->cm.num_frags;
1495
1496         if (ipoib_cm_has_srq(dev)) {
1497                 for (i = 0; i < ipoib_recvq_size; ++i) {
1498                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1499                                                    priv->cm.num_frags - 1,
1500                                                    priv->cm.srq_ring[i].mapping)) {
1501                                 ipoib_warn(priv, "failed to allocate "
1502                                            "receive buffer %d\n", i);
1503                                 ipoib_cm_dev_cleanup(dev);
1504                                 return -ENOMEM;
1505                         }
1506
1507                         if (ipoib_cm_post_receive_srq(dev, i)) {
1508                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1509                                            "failed for buf %d\n", i);
1510                                 ipoib_cm_dev_cleanup(dev);
1511                                 return -EIO;
1512                         }
1513                 }
1514         }
1515
1516         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1517         return 0;
1518 }
1519
1520 void ipoib_cm_dev_cleanup(struct net_device *dev)
1521 {
1522         struct ipoib_dev_priv *priv = netdev_priv(dev);
1523         int ret;
1524
1525         if (!priv->cm.srq)
1526                 return;
1527
1528         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1529
1530         ret = ib_destroy_srq(priv->cm.srq);
1531         if (ret)
1532                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1533
1534         priv->cm.srq = NULL;
1535         if (!priv->cm.srq_ring)
1536                 return;
1537
1538         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1539         priv->cm.srq_ring = NULL;
1540 }