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