drivers/infiniband/ulp/ipoib/ipoib_cm.c

   1 /*
   2  * Copyright (c) 2006 Mellanox Technologies. All rights reserved
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  * $Id$
  33  */
  34
  35 #include <rdma/ib_cm.h>
  36 #include <rdma/ib_cache.h>
  37 #include <net/dst.h>
  38 #include <net/icmp.h>
  39 #include <linux/icmpv6.h>
  40 #include <linux/delay.h>
  41
  42 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  43 static int data_debug_level;
  44
  45 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  46 MODULE_PARM_DESC(cm_data_debug_level,
  47                  "Enable data path debug tracing for connected mode if > 0");
  48 #endif
  49
  50 #include "ipoib.h"
  51
  52 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  53
  54 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  55 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  56 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  57 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  58
  59 struct ipoib_cm_id {
  60         struct ib_cm_id *id;
  61         int flags;
  62         u32 remote_qpn;
  63         u32 remote_mtu;
  64 };
  65
  66 static struct ib_qp_attr ipoib_cm_err_attr = {
  67         .qp_state = IB_QPS_ERR
  68 };
  69
  70 #define IPOIB_CM_RX_DRAIN_WRID 0x7fffffff
  71
  72 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  73         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  74         .opcode = IB_WR_SEND,
  75 };
  76
  77 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  78                                struct ib_cm_event *event);
  79
  80 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  81                                   u64 mapping[IPOIB_CM_RX_SG])
  82 {
  83         int i;
  84
  85         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  86
  87         for (i = 0; i < frags; ++i)
  88                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  89 }
  90
  91 static int ipoib_cm_post_receive(struct net_device *dev, int id)
  92 {
  93         struct ipoib_dev_priv *priv = netdev_priv(dev);
  94         struct ib_recv_wr *bad_wr;
  95         int i, ret;
  96
  97         priv->cm.rx_wr.wr_id = id | IPOIB_CM_OP_SRQ;
  98
  99         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 100                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 101
 102         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 103         if (unlikely(ret)) {
 104                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 105                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 106                                       priv->cm.srq_ring[id].mapping);
 107                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 108                 priv->cm.srq_ring[id].skb = NULL;
 109         }
 110
 111         return ret;
 112 }
 113
 114 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev, int id, int frags,
 115                                              u64 mapping[IPOIB_CM_RX_SG])
 116 {
 117         struct ipoib_dev_priv *priv = netdev_priv(dev);
 118         struct sk_buff *skb;
 119         int i;
 120
 121         skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
 122         if (unlikely(!skb))
 123                 return NULL;
 124
 125         /*
 126          * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
 127          * IP header to a multiple of 16.
 128          */
 129         skb_reserve(skb, 12);
 130
 131         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 132                                        DMA_FROM_DEVICE);
 133         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 134                 dev_kfree_skb_any(skb);
 135                 return NULL;
 136         }
 137
 138         for (i = 0; i < frags; i++) {
 139                 struct page *page = alloc_page(GFP_ATOMIC);
 140
 141                 if (!page)
 142                         goto partial_error;
 143                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 144
 145                 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
 146                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 147                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 148                         goto partial_error;
 149         }
 150
 151         priv->cm.srq_ring[id].skb = skb;
 152         return skb;
 153
 154 partial_error:
 155
 156         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 157
 158         for (; i >= 0; --i)
 159                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
 160
 161         dev_kfree_skb_any(skb);
 162         return NULL;
 163 }
 164
 165 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv* priv)
 166 {
 167         struct ib_send_wr *bad_wr;
 168         struct ipoib_cm_rx *p;
 169
 170         /* We only reserved 1 extra slot in CQ for drain WRs, so
 171          * make sure we have at most 1 outstanding WR. */
 172         if (list_empty(&priv->cm.rx_flush_list) ||
 173             !list_empty(&priv->cm.rx_drain_list))
 174                 return;
 175
 176         /*
 177          * QPs on flush list are error state.  This way, a "flush
 178          * error" WC will be immediately generated for each WR we post.
 179          */
 180         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 181         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 182                 ipoib_warn(priv, "failed to post drain wr\n");
 183
 184         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 185 }
 186
 187 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 188 {
 189         struct ipoib_cm_rx *p = ctx;
 190         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 191         unsigned long flags;
 192
 193         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 194                 return;
 195
 196         spin_lock_irqsave(&priv->lock, flags);
 197         list_move(&p->list, &priv->cm.rx_flush_list);
 198         p->state = IPOIB_CM_RX_FLUSH;
 199         ipoib_cm_start_rx_drain(priv);
 200         spin_unlock_irqrestore(&priv->lock, flags);
 201 }
 202
 203 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 204                                            struct ipoib_cm_rx *p)
 205 {
 206         struct ipoib_dev_priv *priv = netdev_priv(dev);
 207         struct ib_qp_init_attr attr = {
 208                 .event_handler = ipoib_cm_rx_event_handler,
 209                 .send_cq = priv->cq, /* For drain WR */
 210                 .recv_cq = priv->cq,
 211                 .srq = priv->cm.srq,
 212                 .cap.max_send_wr = 1, /* For drain WR */
 213                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 214                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 215                 .qp_type = IB_QPT_RC,
 216                 .qp_context = p,
 217         };
 218         return ib_create_qp(priv->pd, &attr);
 219 }
 220
 221 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 222                                   struct ib_cm_id *cm_id, struct ib_qp *qp,
 223                                   unsigned psn)
 224 {
 225         struct ipoib_dev_priv *priv = netdev_priv(dev);
 226         struct ib_qp_attr qp_attr;
 227         int qp_attr_mask, ret;
 228
 229         qp_attr.qp_state = IB_QPS_INIT;
 230         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 231         if (ret) {
 232                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 233                 return ret;
 234         }
 235         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 236         if (ret) {
 237                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 238                 return ret;
 239         }
 240         qp_attr.qp_state = IB_QPS_RTR;
 241         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 242         if (ret) {
 243                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 244                 return ret;
 245         }
 246         qp_attr.rq_psn = psn;
 247         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 248         if (ret) {
 249                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 250                 return ret;
 251         }
 252
 253         /*
 254          * Current Mellanox HCA firmware won't generate completions
 255          * with error for drain WRs unless the QP has been moved to
 256          * RTS first. This work-around leaves a window where a QP has
 257          * moved to error asynchronously, but this will eventually get
 258          * fixed in firmware, so let's not error out if modify QP
 259          * fails.
 260          */
 261         qp_attr.qp_state = IB_QPS_RTS;
 262         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 263         if (ret) {
 264                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 265                 return 0;
 266         }
 267         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 268         if (ret) {
 269                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 270                 return 0;
 271         }
 272
 273         return 0;
 274 }
 275
 276 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 277                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 278                              unsigned psn)
 279 {
 280         struct ipoib_dev_priv *priv = netdev_priv(dev);
 281         struct ipoib_cm_data data = {};
 282         struct ib_cm_rep_param rep = {};
 283
 284         data.qpn = cpu_to_be32(priv->qp->qp_num);
 285         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 286
 287         rep.private_data = &data;
 288         rep.private_data_len = sizeof data;
 289         rep.flow_control = 0;
 290         rep.rnr_retry_count = req->rnr_retry_count;
 291         rep.target_ack_delay = 20; /* FIXME */
 292         rep.srq = 1;
 293         rep.qp_num = qp->qp_num;
 294         rep.starting_psn = psn;
 295         return ib_send_cm_rep(cm_id, &rep);
 296 }
 297
 298 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 299 {
 300         struct net_device *dev = cm_id->context;
 301         struct ipoib_dev_priv *priv = netdev_priv(dev);
 302         struct ipoib_cm_rx *p;
 303         unsigned psn;
 304         int ret;
 305
 306         ipoib_dbg(priv, "REQ arrived\n");
 307         p = kzalloc(sizeof *p, GFP_KERNEL);
 308         if (!p)
 309                 return -ENOMEM;
 310         p->dev = dev;
 311         p->id = cm_id;
 312         p->qp = ipoib_cm_create_rx_qp(dev, p);
 313         if (IS_ERR(p->qp)) {
 314                 ret = PTR_ERR(p->qp);
 315                 goto err_qp;
 316         }
 317
 318         psn = random32() & 0xffffff;
 319         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 320         if (ret)
 321                 goto err_modify;
 322
 323         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 324         if (ret) {
 325                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 326                 goto err_rep;
 327         }
 328
 329         cm_id->context = p;
 330         p->jiffies = jiffies;
 331         p->state = IPOIB_CM_RX_LIVE;
 332         spin_lock_irq(&priv->lock);
 333         if (list_empty(&priv->cm.passive_ids))
 334                 queue_delayed_work(ipoib_workqueue,
 335                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 336         list_add(&p->list, &priv->cm.passive_ids);
 337         spin_unlock_irq(&priv->lock);
 338         return 0;
 339
 340 err_rep:
 341 err_modify:
 342         ib_destroy_qp(p->qp);
 343 err_qp:
 344         kfree(p);
 345         return ret;
 346 }
 347
 348 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 349                                struct ib_cm_event *event)
 350 {
 351         struct ipoib_cm_rx *p;
 352         struct ipoib_dev_priv *priv;
 353
 354         switch (event->event) {
 355         case IB_CM_REQ_RECEIVED:
 356                 return ipoib_cm_req_handler(cm_id, event);
 357         case IB_CM_DREQ_RECEIVED:
 358                 p = cm_id->context;
 359                 ib_send_cm_drep(cm_id, NULL, 0);
 360                 /* Fall through */
 361         case IB_CM_REJ_RECEIVED:
 362                 p = cm_id->context;
 363                 priv = netdev_priv(p->dev);
 364                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 365                         ipoib_warn(priv, "unable to move qp to error state\n");
 366                 /* Fall through */
 367         default:
 368                 return 0;
 369         }
 370 }
 371 /* Adjust length of skb with fragments to match received data */
 372 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 373                           unsigned int length, struct sk_buff *toskb)
 374 {
 375         int i, num_frags;
 376         unsigned int size;
 377
 378         /* put header into skb */
 379         size = min(length, hdr_space);
 380         skb->tail += size;
 381         skb->len += size;
 382         length -= size;
 383
 384         num_frags = skb_shinfo(skb)->nr_frags;
 385         for (i = 0; i < num_frags; i++) {
 386                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 387
 388                 if (length == 0) {
 389                         /* don't need this page */
 390                         skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
 391                         --skb_shinfo(skb)->nr_frags;
 392                 } else {
 393                         size = min(length, (unsigned) PAGE_SIZE);
 394
 395                         frag->size = size;
 396                         skb->data_len += size;
 397                         skb->truesize += size;
 398                         skb->len += size;
 399                         length -= size;
 400                 }
 401         }
 402 }
 403
 404 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 405 {
 406         struct ipoib_dev_priv *priv = netdev_priv(dev);
 407         unsigned int wr_id = wc->wr_id & ~IPOIB_CM_OP_SRQ;
 408         struct sk_buff *skb, *newskb;
 409         struct ipoib_cm_rx *p;
 410         unsigned long flags;
 411         u64 mapping[IPOIB_CM_RX_SG];
 412         int frags;
 413
 414         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 415                        wr_id, wc->status);
 416
 417         if (unlikely(wr_id >= ipoib_recvq_size)) {
 418                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~IPOIB_CM_OP_SRQ)) {
 419                         spin_lock_irqsave(&priv->lock, flags);
 420                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 421                         ipoib_cm_start_rx_drain(priv);
 422                         queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 423                         spin_unlock_irqrestore(&priv->lock, flags);
 424                 } else
 425                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 426                                    wr_id, ipoib_recvq_size);
 427                 return;
 428         }
 429
 430         skb  = priv->cm.srq_ring[wr_id].skb;
 431
 432         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 433                 ipoib_dbg(priv, "cm recv error "
 434                            "(status=%d, wrid=%d vend_err %x)\n",
 435                            wc->status, wr_id, wc->vendor_err);
 436                 ++priv->stats.rx_dropped;
 437                 goto repost;
 438         }
 439
 440         if (!likely(wr_id & IPOIB_CM_RX_UPDATE_MASK)) {
 441                 p = wc->qp->qp_context;
 442                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 443                         spin_lock_irqsave(&priv->lock, flags);
 444                         p->jiffies = jiffies;
 445                         /* Move this entry to list head, but do not re-add it
 446                          * if it has been moved out of list. */
 447                         if (p->state == IPOIB_CM_RX_LIVE)
 448                                 list_move(&p->list, &priv->cm.passive_ids);
 449                         spin_unlock_irqrestore(&priv->lock, flags);
 450                 }
 451         }
 452
 453         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 454                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 455
 456         newskb = ipoib_cm_alloc_rx_skb(dev, wr_id, frags, mapping);
 457         if (unlikely(!newskb)) {
 458                 /*
 459                  * If we can't allocate a new RX buffer, dump
 460                  * this packet and reuse the old buffer.
 461                  */
 462                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 463                 ++priv->stats.rx_dropped;
 464                 goto repost;
 465         }
 466
 467         ipoib_cm_dma_unmap_rx(priv, frags, priv->cm.srq_ring[wr_id].mapping);
 468         memcpy(priv->cm.srq_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 469
 470         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 471                        wc->byte_len, wc->slid);
 472
 473         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 474
 475         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 476         skb_reset_mac_header(skb);
 477         skb_pull(skb, IPOIB_ENCAP_LEN);
 478
 479         dev->last_rx = jiffies;
 480         ++priv->stats.rx_packets;
 481         priv->stats.rx_bytes += skb->len;
 482
 483         skb->dev = dev;
 484         /* XXX get correct PACKET_ type here */
 485         skb->pkt_type = PACKET_HOST;
 486         netif_receive_skb(skb);
 487
 488 repost:
 489         if (unlikely(ipoib_cm_post_receive(dev, wr_id)))
 490                 ipoib_warn(priv, "ipoib_cm_post_receive failed "
 491                            "for buf %d\n", wr_id);
 492 }
 493
 494 static inline int post_send(struct ipoib_dev_priv *priv,
 495                             struct ipoib_cm_tx *tx,
 496                             unsigned int wr_id,
 497                             u64 addr, int len)
 498 {
 499         struct ib_send_wr *bad_wr;
 500
 501         priv->tx_sge.addr             = addr;
 502         priv->tx_sge.length           = len;
 503
 504         priv->tx_wr.wr_id             = wr_id;
 505
 506         return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
 507 }
 508
 509 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 510 {
 511         struct ipoib_dev_priv *priv = netdev_priv(dev);
 512         struct ipoib_tx_buf *tx_req;
 513         u64 addr;
 514
 515         if (unlikely(skb->len > tx->mtu)) {
 516                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 517                            skb->len, tx->mtu);
 518                 ++priv->stats.tx_dropped;
 519                 ++priv->stats.tx_errors;
 520                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 521                 return;
 522         }
 523
 524         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 525                        tx->tx_head, skb->len, tx->qp->qp_num);
 526
 527         /*
 528          * We put the skb into the tx_ring _before_ we call post_send()
 529          * because it's entirely possible that the completion handler will
 530          * run before we execute anything after the post_send().  That
 531          * means we have to make sure everything is properly recorded and
 532          * our state is consistent before we call post_send().
 533          */
 534         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 535         tx_req->skb = skb;
 536         addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
 537         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 538                 ++priv->stats.tx_errors;
 539                 dev_kfree_skb_any(skb);
 540                 return;
 541         }
 542
 543         tx_req->mapping = addr;
 544
 545         if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
 546                                 addr, skb->len))) {
 547                 ipoib_warn(priv, "post_send failed\n");
 548                 ++priv->stats.tx_errors;
 549                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
 550                 dev_kfree_skb_any(skb);
 551         } else {
 552                 dev->trans_start = jiffies;
 553                 ++tx->tx_head;
 554
 555                 if (tx->tx_head - tx->tx_tail == ipoib_sendq_size) {
 556                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 557                                   tx->qp->qp_num);
 558                         netif_stop_queue(dev);
 559                         set_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags);
 560                 }
 561         }
 562 }
 563
 564 static void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ipoib_cm_tx *tx,
 565                                   struct ib_wc *wc)
 566 {
 567         struct ipoib_dev_priv *priv = netdev_priv(dev);
 568         unsigned int wr_id = wc->wr_id;
 569         struct ipoib_tx_buf *tx_req;
 570         unsigned long flags;
 571
 572         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 573                        wr_id, wc->status);
 574
 575         if (unlikely(wr_id >= ipoib_sendq_size)) {
 576                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 577                            wr_id, ipoib_sendq_size);
 578                 return;
 579         }
 580
 581         tx_req = &tx->tx_ring[wr_id];
 582
 583         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
 584
 585         /* FIXME: is this right? Shouldn't we only increment on success? */
 586         ++priv->stats.tx_packets;
 587         priv->stats.tx_bytes += tx_req->skb->len;
 588
 589         dev_kfree_skb_any(tx_req->skb);
 590
 591         spin_lock_irqsave(&priv->tx_lock, flags);
 592         ++tx->tx_tail;
 593         if (unlikely(test_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags)) &&
 594             tx->tx_head - tx->tx_tail <= ipoib_sendq_size >> 1) {
 595                 clear_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags);
 596                 netif_wake_queue(dev);
 597         }
 598
 599         if (wc->status != IB_WC_SUCCESS &&
 600             wc->status != IB_WC_WR_FLUSH_ERR) {
 601                 struct ipoib_neigh *neigh;
 602
 603                 ipoib_dbg(priv, "failed cm send event "
 604                            "(status=%d, wrid=%d vend_err %x)\n",
 605                            wc->status, wr_id, wc->vendor_err);
 606
 607                 spin_lock(&priv->lock);
 608                 neigh = tx->neigh;
 609
 610                 if (neigh) {
 611                         neigh->cm = NULL;
 612                         list_del(&neigh->list);
 613                         if (neigh->ah)
 614                                 ipoib_put_ah(neigh->ah);
 615                         ipoib_neigh_free(dev, neigh);
 616
 617                         tx->neigh = NULL;
 618                 }
 619
 620                 /* queue would be re-started anyway when TX is destroyed,
 621                  * but it makes sense to do it ASAP here. */
 622                 if (test_and_clear_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags))
 623                         netif_wake_queue(dev);
 624
 625                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 626                         list_move(&tx->list, &priv->cm.reap_list);
 627                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 628                 }
 629
 630                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 631
 632                 spin_unlock(&priv->lock);
 633         }
 634
 635         spin_unlock_irqrestore(&priv->tx_lock, flags);
 636 }
 637
 638 static void ipoib_cm_tx_completion(struct ib_cq *cq, void *tx_ptr)
 639 {
 640         struct ipoib_cm_tx *tx = tx_ptr;
 641         int n, i;
 642
 643         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 644         do {
 645                 n = ib_poll_cq(cq, IPOIB_NUM_WC, tx->ibwc);
 646                 for (i = 0; i < n; ++i)
 647                         ipoib_cm_handle_tx_wc(tx->dev, tx, tx->ibwc + i);
 648         } while (n == IPOIB_NUM_WC);
 649 }
 650
 651 int ipoib_cm_dev_open(struct net_device *dev)
 652 {
 653         struct ipoib_dev_priv *priv = netdev_priv(dev);
 654         int ret;
 655
 656         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 657                 return 0;
 658
 659         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 660         if (IS_ERR(priv->cm.id)) {
 661                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 662                 ret = PTR_ERR(priv->cm.id);
 663                 goto err_cm;
 664         }
 665
 666         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 667                            0, NULL);
 668         if (ret) {
 669                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 670                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 671                 goto err_listen;
 672         }
 673
 674         return 0;
 675
 676 err_listen:
 677         ib_destroy_cm_id(priv->cm.id);
 678 err_cm:
 679         priv->cm.id = NULL;
 680         return ret;
 681 }
 682
 683 void ipoib_cm_dev_stop(struct net_device *dev)
 684 {
 685         struct ipoib_dev_priv *priv = netdev_priv(dev);
 686         struct ipoib_cm_rx *p, *n;
 687         unsigned long begin;
 688         LIST_HEAD(list);
 689         int ret;
 690
 691         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 692                 return;
 693
 694         ib_destroy_cm_id(priv->cm.id);
 695         priv->cm.id = NULL;
 696
 697         spin_lock_irq(&priv->lock);
 698         while (!list_empty(&priv->cm.passive_ids)) {
 699                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 700                 list_move(&p->list, &priv->cm.rx_error_list);
 701                 p->state = IPOIB_CM_RX_ERROR;
 702                 spin_unlock_irq(&priv->lock);
 703                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 704                 if (ret)
 705                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 706                 spin_lock_irq(&priv->lock);
 707         }
 708
 709         /* Wait for all RX to be drained */
 710         begin = jiffies;
 711
 712         while (!list_empty(&priv->cm.rx_error_list) ||
 713                !list_empty(&priv->cm.rx_flush_list) ||
 714                !list_empty(&priv->cm.rx_drain_list)) {
 715                 if (time_after(jiffies, begin + 5 * HZ)) {
 716                         ipoib_warn(priv, "RX drain timing out\n");
 717
 718                         /*
 719                          * assume the HW is wedged and just free up everything.
 720                          */
 721                         list_splice_init(&priv->cm.rx_flush_list, &list);
 722                         list_splice_init(&priv->cm.rx_error_list, &list);
 723                         list_splice_init(&priv->cm.rx_drain_list, &list);
 724                         break;
 725                 }
 726                 spin_unlock_irq(&priv->lock);
 727                 msleep(1);
 728                 ipoib_drain_cq(dev);
 729                 spin_lock_irq(&priv->lock);
 730         }
 731
 732         list_splice_init(&priv->cm.rx_reap_list, &list);
 733
 734         spin_unlock_irq(&priv->lock);
 735
 736         list_for_each_entry_safe(p, n, &list, list) {
 737                 ib_destroy_cm_id(p->id);
 738                 ib_destroy_qp(p->qp);
 739                 kfree(p);
 740         }
 741
 742         cancel_delayed_work(&priv->cm.stale_task);
 743 }
 744
 745 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 746 {
 747         struct ipoib_cm_tx *p = cm_id->context;
 748         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 749         struct ipoib_cm_data *data = event->private_data;
 750         struct sk_buff_head skqueue;
 751         struct ib_qp_attr qp_attr;
 752         int qp_attr_mask, ret;
 753         struct sk_buff *skb;
 754
 755         p->mtu = be32_to_cpu(data->mtu);
 756
 757         if (p->mtu < priv->dev->mtu + IPOIB_ENCAP_LEN) {
 758                 ipoib_warn(priv, "Rejecting connection: mtu %d < device mtu %d + 4\n",
 759                            p->mtu, priv->dev->mtu);
 760                 return -EINVAL;
 761         }
 762
 763         qp_attr.qp_state = IB_QPS_RTR;
 764         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 765         if (ret) {
 766                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 767                 return ret;
 768         }
 769
 770         qp_attr.rq_psn = 0 /* FIXME */;
 771         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 772         if (ret) {
 773                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 774                 return ret;
 775         }
 776
 777         qp_attr.qp_state = IB_QPS_RTS;
 778         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 779         if (ret) {
 780                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 781                 return ret;
 782         }
 783         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 784         if (ret) {
 785                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 786                 return ret;
 787         }
 788
 789         skb_queue_head_init(&skqueue);
 790
 791         spin_lock_irq(&priv->lock);
 792         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 793         if (p->neigh)
 794                 while ((skb = __skb_dequeue(&p->neigh->queue)))
 795                         __skb_queue_tail(&skqueue, skb);
 796         spin_unlock_irq(&priv->lock);
 797
 798         while ((skb = __skb_dequeue(&skqueue))) {
 799                 skb->dev = p->dev;
 800                 if (dev_queue_xmit(skb))
 801                         ipoib_warn(priv, "dev_queue_xmit failed "
 802                                    "to requeue packet\n");
 803         }
 804
 805         ret = ib_send_cm_rtu(cm_id, NULL, 0);
 806         if (ret) {
 807                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
 808                 return ret;
 809         }
 810         return 0;
 811 }
 812
 813 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ib_cq *cq)
 814 {
 815         struct ipoib_dev_priv *priv = netdev_priv(dev);
 816         struct ib_qp_init_attr attr = {};
 817         attr.recv_cq = priv->cq;
 818         attr.srq = priv->cm.srq;
 819         attr.cap.max_send_wr = ipoib_sendq_size;
 820         attr.cap.max_send_sge = 1;
 821         attr.sq_sig_type = IB_SIGNAL_ALL_WR;
 822         attr.qp_type = IB_QPT_RC;
 823         attr.send_cq = cq;
 824         return ib_create_qp(priv->pd, &attr);
 825 }
 826
 827 static int ipoib_cm_send_req(struct net_device *dev,
 828                              struct ib_cm_id *id, struct ib_qp *qp,
 829                              u32 qpn,
 830                              struct ib_sa_path_rec *pathrec)
 831 {
 832         struct ipoib_dev_priv *priv = netdev_priv(dev);
 833         struct ipoib_cm_data data = {};
 834         struct ib_cm_req_param req = {};
 835
 836         data.qpn = cpu_to_be32(priv->qp->qp_num);
 837         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 838
 839         req.primary_path              = pathrec;
 840         req.alternate_path            = NULL;
 841         req.service_id                = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
 842         req.qp_num                    = qp->qp_num;
 843         req.qp_type                   = qp->qp_type;
 844         req.private_data              = &data;
 845         req.private_data_len          = sizeof data;
 846         req.flow_control              = 0;
 847
 848         req.starting_psn              = 0; /* FIXME */
 849
 850         /*
 851          * Pick some arbitrary defaults here; we could make these
 852          * module parameters if anyone cared about setting them.
 853          */
 854         req.responder_resources       = 4;
 855         req.remote_cm_response_timeout = 20;
 856         req.local_cm_response_timeout  = 20;
 857         req.retry_count               = 0; /* RFC draft warns against retries */
 858         req.rnr_retry_count           = 0; /* RFC draft warns against retries */
 859         req.max_cm_retries            = 15;
 860         req.srq                       = 1;
 861         return ib_send_cm_req(id, &req);
 862 }
 863
 864 static int ipoib_cm_modify_tx_init(struct net_device *dev,
 865                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
 866 {
 867         struct ipoib_dev_priv *priv = netdev_priv(dev);
 868         struct ib_qp_attr qp_attr;
 869         int qp_attr_mask, ret;
 870         ret = ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
 871         if (ret) {
 872                 ipoib_warn(priv, "pkey 0x%x not in cache: %d\n", priv->pkey, ret);
 873                 return ret;
 874         }
 875
 876         qp_attr.qp_state = IB_QPS_INIT;
 877         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
 878         qp_attr.port_num = priv->port;
 879         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
 880
 881         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 882         if (ret) {
 883                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
 884                 return ret;
 885         }
 886         return 0;
 887 }
 888
 889 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
 890                             struct ib_sa_path_rec *pathrec)
 891 {
 892         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 893         int ret;
 894
 895         p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring,
 896                                 GFP_KERNEL);
 897         if (!p->tx_ring) {
 898                 ipoib_warn(priv, "failed to allocate tx ring\n");
 899                 ret = -ENOMEM;
 900                 goto err_tx;
 901         }
 902
 903         p->cq = ib_create_cq(priv->ca, ipoib_cm_tx_completion, NULL, p,
 904                              ipoib_sendq_size + 1, 0);
 905         if (IS_ERR(p->cq)) {
 906                 ret = PTR_ERR(p->cq);
 907                 ipoib_warn(priv, "failed to allocate tx cq: %d\n", ret);
 908                 goto err_cq;
 909         }
 910
 911         ret = ib_req_notify_cq(p->cq, IB_CQ_NEXT_COMP);
 912         if (ret) {
 913                 ipoib_warn(priv, "failed to request completion notification: %d\n", ret);
 914                 goto err_req_notify;
 915         }
 916
 917         p->qp = ipoib_cm_create_tx_qp(p->dev, p->cq);
 918         if (IS_ERR(p->qp)) {
 919                 ret = PTR_ERR(p->qp);
 920                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
 921                 goto err_qp;
 922         }
 923
 924         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
 925         if (IS_ERR(p->id)) {
 926                 ret = PTR_ERR(p->id);
 927                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
 928                 goto err_id;
 929         }
 930
 931         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
 932         if (ret) {
 933                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
 934                 goto err_modify;
 935         }
 936
 937         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
 938         if (ret) {
 939                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
 940                 goto err_send_cm;
 941         }
 942
 943         ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
 944                   p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
 945
 946         return 0;
 947
 948 err_send_cm:
 949 err_modify:
 950         ib_destroy_cm_id(p->id);
 951 err_id:
 952         p->id = NULL;
 953         ib_destroy_qp(p->qp);
 954 err_req_notify:
 955 err_qp:
 956         p->qp = NULL;
 957         ib_destroy_cq(p->cq);
 958 err_cq:
 959         p->cq = NULL;
 960 err_tx:
 961         return ret;
 962 }
 963
 964 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
 965 {
 966         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 967         struct ipoib_tx_buf *tx_req;
 968
 969         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
 970                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
 971
 972         if (p->id)
 973                 ib_destroy_cm_id(p->id);
 974
 975         if (p->qp)
 976                 ib_destroy_qp(p->qp);
 977
 978         if (p->cq)
 979                 ib_destroy_cq(p->cq);
 980
 981         if (test_bit(IPOIB_FLAG_NETIF_STOPPED, &p->flags))
 982                 netif_wake_queue(p->dev);
 983
 984         if (p->tx_ring) {
 985                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
 986                         tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
 987                         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
 988                                          DMA_TO_DEVICE);
 989                         dev_kfree_skb_any(tx_req->skb);
 990                         ++p->tx_tail;
 991                 }
 992
 993                 kfree(p->tx_ring);
 994         }
 995
 996         kfree(p);
 997 }
 998
 999 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1000                                struct ib_cm_event *event)
1001 {
1002         struct ipoib_cm_tx *tx = cm_id->context;
1003         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1004         struct net_device *dev = priv->dev;
1005         struct ipoib_neigh *neigh;
1006         int ret;
1007
1008         switch (event->event) {
1009         case IB_CM_DREQ_RECEIVED:
1010                 ipoib_dbg(priv, "DREQ received.\n");
1011                 ib_send_cm_drep(cm_id, NULL, 0);
1012                 break;
1013         case IB_CM_REP_RECEIVED:
1014                 ipoib_dbg(priv, "REP received.\n");
1015                 ret = ipoib_cm_rep_handler(cm_id, event);
1016                 if (ret)
1017                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1018                                        NULL, 0, NULL, 0);
1019                 break;
1020         case IB_CM_REQ_ERROR:
1021         case IB_CM_REJ_RECEIVED:
1022         case IB_CM_TIMEWAIT_EXIT:
1023                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1024                 spin_lock_irq(&priv->tx_lock);
1025                 spin_lock(&priv->lock);
1026                 neigh = tx->neigh;
1027
1028                 if (neigh) {
1029                         neigh->cm = NULL;
1030                         list_del(&neigh->list);
1031                         if (neigh->ah)
1032                                 ipoib_put_ah(neigh->ah);
1033                         ipoib_neigh_free(dev, neigh);
1034
1035                         tx->neigh = NULL;
1036                 }
1037
1038                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1039                         list_move(&tx->list, &priv->cm.reap_list);
1040                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1041                 }
1042
1043                 spin_unlock(&priv->lock);
1044                 spin_unlock_irq(&priv->tx_lock);
1045                 break;
1046         default:
1047                 break;
1048         }
1049
1050         return 0;
1051 }
1052
1053 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1054                                        struct ipoib_neigh *neigh)
1055 {
1056         struct ipoib_dev_priv *priv = netdev_priv(dev);
1057         struct ipoib_cm_tx *tx;
1058
1059         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1060         if (!tx)
1061                 return NULL;
1062
1063         neigh->cm = tx;
1064         tx->neigh = neigh;
1065         tx->path = path;
1066         tx->dev = dev;
1067         list_add(&tx->list, &priv->cm.start_list);
1068         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1069         queue_work(ipoib_workqueue, &priv->cm.start_task);
1070         return tx;
1071 }
1072
1073 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1074 {
1075         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1076         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1077                 list_move(&tx->list, &priv->cm.reap_list);
1078                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1079                 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1080                           IPOIB_GID_ARG(tx->neigh->dgid));
1081                 tx->neigh = NULL;
1082         }
1083 }
1084
1085 static void ipoib_cm_tx_start(struct work_struct *work)
1086 {
1087         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1088                                                    cm.start_task);
1089         struct net_device *dev = priv->dev;
1090         struct ipoib_neigh *neigh;
1091         struct ipoib_cm_tx *p;
1092         unsigned long flags;
1093         int ret;
1094
1095         struct ib_sa_path_rec pathrec;
1096         u32 qpn;
1097
1098         spin_lock_irqsave(&priv->tx_lock, flags);
1099         spin_lock(&priv->lock);
1100         while (!list_empty(&priv->cm.start_list)) {
1101                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1102                 list_del_init(&p->list);
1103                 neigh = p->neigh;
1104                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1105                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1106                 spin_unlock(&priv->lock);
1107                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1108                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1109                 spin_lock_irqsave(&priv->tx_lock, flags);
1110                 spin_lock(&priv->lock);
1111                 if (ret) {
1112                         neigh = p->neigh;
1113                         if (neigh) {
1114                                 neigh->cm = NULL;
1115                                 list_del(&neigh->list);
1116                                 if (neigh->ah)
1117                                         ipoib_put_ah(neigh->ah);
1118                                 ipoib_neigh_free(dev, neigh);
1119                         }
1120                         list_del(&p->list);
1121                         kfree(p);
1122                 }
1123         }
1124         spin_unlock(&priv->lock);
1125         spin_unlock_irqrestore(&priv->tx_lock, flags);
1126 }
1127
1128 static void ipoib_cm_tx_reap(struct work_struct *work)
1129 {
1130         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1131                                                    cm.reap_task);
1132         struct ipoib_cm_tx *p;
1133
1134         spin_lock_irq(&priv->tx_lock);
1135         spin_lock(&priv->lock);
1136         while (!list_empty(&priv->cm.reap_list)) {
1137                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1138                 list_del(&p->list);
1139                 spin_unlock(&priv->lock);
1140                 spin_unlock_irq(&priv->tx_lock);
1141                 ipoib_cm_tx_destroy(p);
1142                 spin_lock_irq(&priv->tx_lock);
1143                 spin_lock(&priv->lock);
1144         }
1145         spin_unlock(&priv->lock);
1146         spin_unlock_irq(&priv->tx_lock);
1147 }
1148
1149 static void ipoib_cm_skb_reap(struct work_struct *work)
1150 {
1151         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1152                                                    cm.skb_task);
1153         struct net_device *dev = priv->dev;
1154         struct sk_buff *skb;
1155
1156         unsigned mtu = priv->mcast_mtu;
1157
1158         spin_lock_irq(&priv->tx_lock);
1159         spin_lock(&priv->lock);
1160         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1161                 spin_unlock(&priv->lock);
1162                 spin_unlock_irq(&priv->tx_lock);
1163                 if (skb->protocol == htons(ETH_P_IP))
1164                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1165 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1166                 else if (skb->protocol == htons(ETH_P_IPV6))
1167                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, dev);
1168 #endif
1169                 dev_kfree_skb_any(skb);
1170                 spin_lock_irq(&priv->tx_lock);
1171                 spin_lock(&priv->lock);
1172         }
1173         spin_unlock(&priv->lock);
1174         spin_unlock_irq(&priv->tx_lock);
1175 }
1176
1177 void ipoib_cm_skb_too_long(struct net_device* dev, struct sk_buff *skb,
1178                            unsigned int mtu)
1179 {
1180         struct ipoib_dev_priv *priv = netdev_priv(dev);
1181         int e = skb_queue_empty(&priv->cm.skb_queue);
1182
1183         if (skb->dst)
1184                 skb->dst->ops->update_pmtu(skb->dst, mtu);
1185
1186         skb_queue_tail(&priv->cm.skb_queue, skb);
1187         if (e)
1188                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1189 }
1190
1191 static void ipoib_cm_rx_reap(struct work_struct *work)
1192 {
1193         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1194                                                    cm.rx_reap_task);
1195         struct ipoib_cm_rx *p, *n;
1196         LIST_HEAD(list);
1197
1198         spin_lock_irq(&priv->lock);
1199         list_splice_init(&priv->cm.rx_reap_list, &list);
1200         spin_unlock_irq(&priv->lock);
1201
1202         list_for_each_entry_safe(p, n, &list, list) {
1203                 ib_destroy_cm_id(p->id);
1204                 ib_destroy_qp(p->qp);
1205                 kfree(p);
1206         }
1207 }
1208
1209 static void ipoib_cm_stale_task(struct work_struct *work)
1210 {
1211         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1212                                                    cm.stale_task.work);
1213         struct ipoib_cm_rx *p;
1214         int ret;
1215
1216         spin_lock_irq(&priv->lock);
1217         while (!list_empty(&priv->cm.passive_ids)) {
1218                 /* List is sorted by LRU, start from tail,
1219                  * stop when we see a recently used entry */
1220                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1221                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1222                         break;
1223                 list_move(&p->list, &priv->cm.rx_error_list);
1224                 p->state = IPOIB_CM_RX_ERROR;
1225                 spin_unlock_irq(&priv->lock);
1226                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1227                 if (ret)
1228                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1229                 spin_lock_irq(&priv->lock);
1230         }
1231
1232         if (!list_empty(&priv->cm.passive_ids))
1233                 queue_delayed_work(ipoib_workqueue,
1234                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1235         spin_unlock_irq(&priv->lock);
1236 }
1237
1238
1239 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1240                          char *buf)
1241 {
1242         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1243
1244         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1245                 return sprintf(buf, "connected\n");
1246         else
1247                 return sprintf(buf, "datagram\n");
1248 }
1249
1250 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1251                         const char *buf, size_t count)
1252 {
1253         struct net_device *dev = to_net_dev(d);
1254         struct ipoib_dev_priv *priv = netdev_priv(dev);
1255
1256         /* flush paths if we switch modes so that connections are restarted */
1257         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1258                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1259                 ipoib_warn(priv, "enabling connected mode "
1260                            "will cause multicast packet drops\n");
1261                 ipoib_flush_paths(dev);
1262                 return count;
1263         }
1264
1265         if (!strcmp(buf, "datagram\n")) {
1266                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1267                 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1268                 ipoib_flush_paths(dev);
1269                 return count;
1270         }
1271
1272         return -EINVAL;
1273 }
1274
1275 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1276
1277 int ipoib_cm_add_mode_attr(struct net_device *dev)
1278 {
1279         return device_create_file(&dev->dev, &dev_attr_mode);
1280 }
1281
1282 int ipoib_cm_dev_init(struct net_device *dev)
1283 {
1284         struct ipoib_dev_priv *priv = netdev_priv(dev);
1285         struct ib_srq_init_attr srq_init_attr = {
1286                 .attr = {
1287                         .max_wr  = ipoib_recvq_size,
1288                         .max_sge = IPOIB_CM_RX_SG
1289                 }
1290         };
1291         int ret, i;
1292
1293         INIT_LIST_HEAD(&priv->cm.passive_ids);
1294         INIT_LIST_HEAD(&priv->cm.reap_list);
1295         INIT_LIST_HEAD(&priv->cm.start_list);
1296         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1297         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1298         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1299         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1300         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1301         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1302         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1303         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1304         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1305
1306         skb_queue_head_init(&priv->cm.skb_queue);
1307
1308         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1309         if (IS_ERR(priv->cm.srq)) {
1310                 ret = PTR_ERR(priv->cm.srq);
1311                 priv->cm.srq = NULL;
1312                 return ret;
1313         }
1314
1315         priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1316                                     GFP_KERNEL);
1317         if (!priv->cm.srq_ring) {
1318                 printk(KERN_WARNING "%s: failed to allocate CM ring (%d entries)\n",
1319                        priv->ca->name, ipoib_recvq_size);
1320                 ipoib_cm_dev_cleanup(dev);
1321                 return -ENOMEM;
1322         }
1323
1324         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
1325                 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1326
1327         priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1328         for (i = 1; i < IPOIB_CM_RX_SG; ++i)
1329                 priv->cm.rx_sge[i].length = PAGE_SIZE;
1330         priv->cm.rx_wr.next = NULL;
1331         priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1332         priv->cm.rx_wr.num_sge = IPOIB_CM_RX_SG;
1333
1334         for (i = 0; i < ipoib_recvq_size; ++i) {
1335                 if (!ipoib_cm_alloc_rx_skb(dev, i, IPOIB_CM_RX_SG - 1,
1336                                            priv->cm.srq_ring[i].mapping)) {
1337                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
1338                         ipoib_cm_dev_cleanup(dev);
1339                         return -ENOMEM;
1340                 }
1341                 if (ipoib_cm_post_receive(dev, i)) {
1342                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
1343                         ipoib_cm_dev_cleanup(dev);
1344                         return -EIO;
1345                 }
1346         }
1347
1348         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1349         return 0;
1350 }
1351
1352 void ipoib_cm_dev_cleanup(struct net_device *dev)
1353 {
1354         struct ipoib_dev_priv *priv = netdev_priv(dev);
1355         int i, ret;
1356
1357         if (!priv->cm.srq)
1358                 return;
1359
1360         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1361
1362         ret = ib_destroy_srq(priv->cm.srq);
1363         if (ret)
1364                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1365
1366         priv->cm.srq = NULL;
1367         if (!priv->cm.srq_ring)
1368                 return;
1369         for (i = 0; i < ipoib_recvq_size; ++i)
1370                 if (priv->cm.srq_ring[i].skb) {
1371                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
1372                                               priv->cm.srq_ring[i].mapping);
1373                         dev_kfree_skb_any(priv->cm.srq_ring[i].skb);
1374                         priv->cm.srq_ring[i].skb = NULL;
1375                 }
1376         kfree(priv->cm.srq_ring);
1377         priv->cm.srq_ring = NULL;
1378 }