git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/ulp/ipoib/ipoib_ib.c

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  *
  35  * $Id: ipoib_ib.c 1386 2004-12-27 16:23:17Z roland $
  36  */
  37
  38 #include <linux/delay.h>
  39 #include <linux/dma-mapping.h>
  40
  41 #include <rdma/ib_cache.h>
  42
  43 #include "ipoib.h"
  44
  45 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  46 static int data_debug_level;
  47
  48 module_param(data_debug_level, int, 0644);
  49 MODULE_PARM_DESC(data_debug_level,
  50                  "Enable data path debug tracing if > 0");
  51 #endif
  52
  53 #define IPOIB_OP_RECV   (1ul << 31)
  54
  55 static DECLARE_MUTEX(pkey_sem);
  56
  57 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
  58                                  struct ib_pd *pd, struct ib_ah_attr *attr)
  59 {
  60         struct ipoib_ah *ah;
  61
  62         ah = kmalloc(sizeof *ah, GFP_KERNEL);
  63         if (!ah)
  64                 return NULL;
  65
  66         ah->dev       = dev;
  67         ah->last_send = 0;
  68         kref_init(&ah->ref);
  69
  70         ah->ah = ib_create_ah(pd, attr);
  71         if (IS_ERR(ah->ah)) {
  72                 kfree(ah);
  73                 ah = NULL;
  74         } else
  75                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
  76
  77         return ah;
  78 }
  79
  80 void ipoib_free_ah(struct kref *kref)
  81 {
  82         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
  83         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
  84
  85         unsigned long flags;
  86
  87         if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
  88                 ipoib_dbg(priv, "Freeing ah %p\n", ah->ah);
  89                 ib_destroy_ah(ah->ah);
  90                 kfree(ah);
  91         } else {
  92                 spin_lock_irqsave(&priv->lock, flags);
  93                 list_add_tail(&ah->list, &priv->dead_ahs);
  94                 spin_unlock_irqrestore(&priv->lock, flags);
  95         }
  96 }
  97
  98 static inline int ipoib_ib_receive(struct ipoib_dev_priv *priv,
  99                                    unsigned int wr_id,
 100                                    dma_addr_t addr)
 101 {
 102         struct ib_sge list = {
 103                 .addr    = addr,
 104                 .length  = IPOIB_BUF_SIZE,
 105                 .lkey    = priv->mr->lkey,
 106         };
 107         struct ib_recv_wr param = {
 108                 .wr_id      = wr_id | IPOIB_OP_RECV,
 109                 .sg_list    = &list,
 110                 .num_sge    = 1,
 111         };
 112         struct ib_recv_wr *bad_wr;
 113
 114         return ib_post_recv(priv->qp, &param, &bad_wr);
 115 }
 116
 117 static int ipoib_ib_post_receive(struct net_device *dev, int id)
 118 {
 119         struct ipoib_dev_priv *priv = netdev_priv(dev);
 120         struct sk_buff *skb;
 121         dma_addr_t addr;
 122         int ret;
 123
 124         skb = dev_alloc_skb(IPOIB_BUF_SIZE + 4);
 125         if (!skb) {
 126                 ipoib_warn(priv, "failed to allocate receive buffer\n");
 127
 128                 priv->rx_ring[id].skb = NULL;
 129                 return -ENOMEM;
 130         }
 131         skb_reserve(skb, 4);    /* 16 byte align IP header */
 132         priv->rx_ring[id].skb = skb;
 133         addr = dma_map_single(priv->ca->dma_device,
 134                               skb->data, IPOIB_BUF_SIZE,
 135                               DMA_FROM_DEVICE);
 136         pci_unmap_addr_set(&priv->rx_ring[id], mapping, addr);
 137
 138         ret = ipoib_ib_receive(priv, id, addr);
 139         if (ret) {
 140                 ipoib_warn(priv, "ipoib_ib_receive failed for buf %d (%d)\n",
 141                            id, ret);
 142                 dma_unmap_single(priv->ca->dma_device, addr,
 143                                  IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
 144                 dev_kfree_skb_any(skb);
 145                 priv->rx_ring[id].skb = NULL;
 146         }
 147
 148         return ret;
 149 }
 150
 151 static int ipoib_ib_post_receives(struct net_device *dev)
 152 {
 153         struct ipoib_dev_priv *priv = netdev_priv(dev);
 154         int i;
 155
 156         for (i = 0; i < IPOIB_RX_RING_SIZE; ++i) {
 157                 if (ipoib_ib_post_receive(dev, i)) {
 158                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 159                         return -EIO;
 160                 }
 161         }
 162
 163         return 0;
 164 }
 165
 166 static void ipoib_ib_handle_wc(struct net_device *dev,
 167                                struct ib_wc *wc)
 168 {
 169         struct ipoib_dev_priv *priv = netdev_priv(dev);
 170         unsigned int wr_id = wc->wr_id;
 171
 172         ipoib_dbg_data(priv, "called: id %d, op %d, status: %d\n",
 173                        wr_id, wc->opcode, wc->status);
 174
 175         if (wr_id & IPOIB_OP_RECV) {
 176                 wr_id &= ~IPOIB_OP_RECV;
 177
 178                 if (wr_id < IPOIB_RX_RING_SIZE) {
 179                         struct sk_buff *skb = priv->rx_ring[wr_id].skb;
 180
 181                         priv->rx_ring[wr_id].skb = NULL;
 182
 183                         dma_unmap_single(priv->ca->dma_device,
 184                                          pci_unmap_addr(&priv->rx_ring[wr_id],
 185                                                         mapping),
 186                                          IPOIB_BUF_SIZE,
 187                                          DMA_FROM_DEVICE);
 188
 189                         if (wc->status != IB_WC_SUCCESS) {
 190                                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 191                                         ipoib_warn(priv, "failed recv event "
 192                                                    "(status=%d, wrid=%d vend_err %x)\n",
 193                                                    wc->status, wr_id, wc->vendor_err);
 194                                 dev_kfree_skb_any(skb);
 195                                 return;
 196                         }
 197
 198                         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 199                                        wc->byte_len, wc->slid);
 200
 201                         skb_put(skb, wc->byte_len);
 202                         skb_pull(skb, IB_GRH_BYTES);
 203
 204                         if (wc->slid != priv->local_lid ||
 205                             wc->src_qp != priv->qp->qp_num) {
 206                                 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 207                                 skb->mac.raw = skb->data;
 208                                 skb_pull(skb, IPOIB_ENCAP_LEN);
 209
 210                                 dev->last_rx = jiffies;
 211                                 ++priv->stats.rx_packets;
 212                                 priv->stats.rx_bytes += skb->len;
 213
 214                                 skb->dev = dev;
 215                                 /* XXX get correct PACKET_ type here */
 216                                 skb->pkt_type = PACKET_HOST;
 217                                 netif_rx_ni(skb);
 218                         } else {
 219                                 ipoib_dbg_data(priv, "dropping loopback packet\n");
 220                                 dev_kfree_skb_any(skb);
 221                         }
 222
 223                         /* repost receive */
 224                         if (ipoib_ib_post_receive(dev, wr_id))
 225                                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 226                                            "for buf %d\n", wr_id);
 227                 } else
 228                         ipoib_warn(priv, "completion event with wrid %d\n",
 229                                    wr_id);
 230
 231         } else {
 232                 struct ipoib_buf *tx_req;
 233                 unsigned long flags;
 234
 235                 if (wr_id >= IPOIB_TX_RING_SIZE) {
 236                         ipoib_warn(priv, "completion event with wrid %d (> %d)\n",
 237                                    wr_id, IPOIB_TX_RING_SIZE);
 238                         return;
 239                 }
 240
 241                 ipoib_dbg_data(priv, "send complete, wrid %d\n", wr_id);
 242
 243                 tx_req = &priv->tx_ring[wr_id];
 244
 245                 dma_unmap_single(priv->ca->dma_device,
 246                                  pci_unmap_addr(tx_req, mapping),
 247                                  tx_req->skb->len,
 248                                  DMA_TO_DEVICE);
 249
 250                 ++priv->stats.tx_packets;
 251                 priv->stats.tx_bytes += tx_req->skb->len;
 252
 253                 dev_kfree_skb_any(tx_req->skb);
 254
 255                 spin_lock_irqsave(&priv->tx_lock, flags);
 256                 ++priv->tx_tail;
 257                 if (netif_queue_stopped(dev) &&
 258                     priv->tx_head - priv->tx_tail <= IPOIB_TX_RING_SIZE / 2)
 259                         netif_wake_queue(dev);
 260                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 261
 262                 if (wc->status != IB_WC_SUCCESS &&
 263                     wc->status != IB_WC_WR_FLUSH_ERR)
 264                         ipoib_warn(priv, "failed send event "
 265                                    "(status=%d, wrid=%d vend_err %x)\n",
 266                                    wc->status, wr_id, wc->vendor_err);
 267         }
 268 }
 269
 270 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 271 {
 272         struct net_device *dev = (struct net_device *) dev_ptr;
 273         struct ipoib_dev_priv *priv = netdev_priv(dev);
 274         int n, i;
 275
 276         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 277         do {
 278                 n = ib_poll_cq(cq, IPOIB_NUM_WC, priv->ibwc);
 279                 for (i = 0; i < n; ++i)
 280                         ipoib_ib_handle_wc(dev, priv->ibwc + i);
 281         } while (n == IPOIB_NUM_WC);
 282 }
 283
 284 static inline int post_send(struct ipoib_dev_priv *priv,
 285                             unsigned int wr_id,
 286                             struct ib_ah *address, u32 qpn,
 287                             dma_addr_t addr, int len)
 288 {
 289         struct ib_send_wr *bad_wr;
 290
 291         priv->tx_sge.addr             = addr;
 292         priv->tx_sge.length           = len;
 293
 294         priv->tx_wr.wr_id             = wr_id;
 295         priv->tx_wr.wr.ud.remote_qpn  = qpn;
 296         priv->tx_wr.wr.ud.ah          = address;
 297
 298         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 299 }
 300
 301 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 302                 struct ipoib_ah *address, u32 qpn)
 303 {
 304         struct ipoib_dev_priv *priv = netdev_priv(dev);
 305         struct ipoib_buf *tx_req;
 306         dma_addr_t addr;
 307
 308         if (skb->len > dev->mtu + INFINIBAND_ALEN) {
 309                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 310                            skb->len, dev->mtu + INFINIBAND_ALEN);
 311                 ++priv->stats.tx_dropped;
 312                 ++priv->stats.tx_errors;
 313                 dev_kfree_skb_any(skb);
 314                 return;
 315         }
 316
 317         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 318                        skb->len, address, qpn);
 319
 320         /*
 321          * We put the skb into the tx_ring _before_ we call post_send()
 322          * because it's entirely possible that the completion handler will
 323          * run before we execute anything after the post_send().  That
 324          * means we have to make sure everything is properly recorded and
 325          * our state is consistent before we call post_send().
 326          */
 327         tx_req = &priv->tx_ring[priv->tx_head & (IPOIB_TX_RING_SIZE - 1)];
 328         tx_req->skb = skb;
 329         addr = dma_map_single(priv->ca->dma_device, skb->data, skb->len,
 330                               DMA_TO_DEVICE);
 331         pci_unmap_addr_set(tx_req, mapping, addr);
 332
 333         if (unlikely(post_send(priv, priv->tx_head & (IPOIB_TX_RING_SIZE - 1),
 334                                address->ah, qpn, addr, skb->len))) {
 335                 ipoib_warn(priv, "post_send failed\n");
 336                 ++priv->stats.tx_errors;
 337                 dma_unmap_single(priv->ca->dma_device, addr, skb->len,
 338                                  DMA_TO_DEVICE);
 339                 dev_kfree_skb_any(skb);
 340         } else {
 341                 dev->trans_start = jiffies;
 342
 343                 address->last_send = priv->tx_head;
 344                 ++priv->tx_head;
 345
 346                 if (priv->tx_head - priv->tx_tail == IPOIB_TX_RING_SIZE) {
 347                         ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 348                         netif_stop_queue(dev);
 349                 }
 350         }
 351 }
 352
 353 static void __ipoib_reap_ah(struct net_device *dev)
 354 {
 355         struct ipoib_dev_priv *priv = netdev_priv(dev);
 356         struct ipoib_ah *ah, *tah;
 357         LIST_HEAD(remove_list);
 358
 359         spin_lock_irq(&priv->lock);
 360         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 361                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 362                         list_del(&ah->list);
 363                         list_add_tail(&ah->list, &remove_list);
 364                 }
 365         spin_unlock_irq(&priv->lock);
 366
 367         list_for_each_entry_safe(ah, tah, &remove_list, list) {
 368                 ipoib_dbg(priv, "Reaping ah %p\n", ah->ah);
 369                 ib_destroy_ah(ah->ah);
 370                 kfree(ah);
 371         }
 372 }
 373
 374 void ipoib_reap_ah(void *dev_ptr)
 375 {
 376         struct net_device *dev = dev_ptr;
 377         struct ipoib_dev_priv *priv = netdev_priv(dev);
 378
 379         __ipoib_reap_ah(dev);
 380
 381         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 382                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
 383 }
 384
 385 int ipoib_ib_dev_open(struct net_device *dev)
 386 {
 387         struct ipoib_dev_priv *priv = netdev_priv(dev);
 388         int ret;
 389
 390         ret = ipoib_qp_create(dev);
 391         if (ret) {
 392                 ipoib_warn(priv, "ipoib_qp_create returned %d\n", ret);
 393                 return -1;
 394         }
 395
 396         ret = ipoib_ib_post_receives(dev);
 397         if (ret) {
 398                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 399                 return -1;
 400         }
 401
 402         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 403         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
 404
 405         return 0;
 406 }
 407
 408 int ipoib_ib_dev_up(struct net_device *dev)
 409 {
 410         struct ipoib_dev_priv *priv = netdev_priv(dev);
 411
 412         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 413
 414         return ipoib_mcast_start_thread(dev);
 415 }
 416
 417 int ipoib_ib_dev_down(struct net_device *dev)
 418 {
 419         struct ipoib_dev_priv *priv = netdev_priv(dev);
 420
 421         ipoib_dbg(priv, "downing ib_dev\n");
 422
 423         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 424         netif_carrier_off(dev);
 425
 426         /* Shutdown the P_Key thread if still active */
 427         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 428                 down(&pkey_sem);
 429                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
 430                 cancel_delayed_work(&priv->pkey_task);
 431                 up(&pkey_sem);
 432                 flush_workqueue(ipoib_workqueue);
 433         }
 434
 435         ipoib_mcast_stop_thread(dev);
 436
 437         /*
 438          * Flush the multicast groups first so we stop any multicast joins. The
 439          * completion thread may have already died and we may deadlock waiting
 440          * for the completion thread to finish some multicast joins.
 441          */
 442         ipoib_mcast_dev_flush(dev);
 443
 444         /* Delete broadcast and local addresses since they will be recreated */
 445         ipoib_mcast_dev_down(dev);
 446
 447         ipoib_flush_paths(dev);
 448
 449         return 0;
 450 }
 451
 452 static int recvs_pending(struct net_device *dev)
 453 {
 454         struct ipoib_dev_priv *priv = netdev_priv(dev);
 455         int pending = 0;
 456         int i;
 457
 458         for (i = 0; i < IPOIB_RX_RING_SIZE; ++i)
 459                 if (priv->rx_ring[i].skb)
 460                         ++pending;
 461
 462         return pending;
 463 }
 464
 465 int ipoib_ib_dev_stop(struct net_device *dev)
 466 {
 467         struct ipoib_dev_priv *priv = netdev_priv(dev);
 468         struct ib_qp_attr qp_attr;
 469         int attr_mask;
 470         unsigned long begin;
 471         struct ipoib_buf *tx_req;
 472         int i;
 473
 474         /* Kill the existing QP and allocate a new one */
 475         qp_attr.qp_state = IB_QPS_ERR;
 476         attr_mask        = IB_QP_STATE;
 477         if (ib_modify_qp(priv->qp, &qp_attr, attr_mask))
 478                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 479
 480         /* Wait for all sends and receives to complete */
 481         begin = jiffies;
 482
 483         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 484                 if (time_after(jiffies, begin + 5 * HZ)) {
 485                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 486                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 487
 488                         /*
 489                          * assume the HW is wedged and just free up
 490                          * all our pending work requests.
 491                          */
 492                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 493                                 tx_req = &priv->tx_ring[priv->tx_tail &
 494                                                         (IPOIB_TX_RING_SIZE - 1)];
 495                                 dma_unmap_single(priv->ca->dma_device,
 496                                                  pci_unmap_addr(tx_req, mapping),
 497                                                  tx_req->skb->len,
 498                                                  DMA_TO_DEVICE);
 499                                 dev_kfree_skb_any(tx_req->skb);
 500                                 ++priv->tx_tail;
 501                         }
 502
 503                         for (i = 0; i < IPOIB_RX_RING_SIZE; ++i)
 504                                 if (priv->rx_ring[i].skb) {
 505                                         dma_unmap_single(priv->ca->dma_device,
 506                                                          pci_unmap_addr(&priv->rx_ring[i],
 507                                                                         mapping),
 508                                                          IPOIB_BUF_SIZE,
 509                                                          DMA_FROM_DEVICE);
 510                                         dev_kfree_skb_any(priv->rx_ring[i].skb);
 511                                         priv->rx_ring[i].skb = NULL;
 512                                 }
 513
 514                         goto timeout;
 515                 }
 516
 517                 msleep(1);
 518         }
 519
 520         ipoib_dbg(priv, "All sends and receives done.\n");
 521
 522 timeout:
 523         qp_attr.qp_state = IB_QPS_RESET;
 524         attr_mask        = IB_QP_STATE;
 525         if (ib_modify_qp(priv->qp, &qp_attr, attr_mask))
 526                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 527
 528         /* Wait for all AHs to be reaped */
 529         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 530         cancel_delayed_work(&priv->ah_reap_task);
 531         flush_workqueue(ipoib_workqueue);
 532
 533         begin = jiffies;
 534
 535         while (!list_empty(&priv->dead_ahs)) {
 536                 __ipoib_reap_ah(dev);
 537
 538                 if (time_after(jiffies, begin + HZ)) {
 539                         ipoib_warn(priv, "timing out; will leak address handles\n");
 540                         break;
 541                 }
 542
 543                 msleep(1);
 544         }
 545
 546         return 0;
 547 }
 548
 549 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 550 {
 551         struct ipoib_dev_priv *priv = netdev_priv(dev);
 552
 553         priv->ca = ca;
 554         priv->port = port;
 555         priv->qp = NULL;
 556
 557         if (ipoib_transport_dev_init(dev, ca)) {
 558                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 559                 return -ENODEV;
 560         }
 561
 562         if (dev->flags & IFF_UP) {
 563                 if (ipoib_ib_dev_open(dev)) {
 564                         ipoib_transport_dev_cleanup(dev);
 565                         return -ENODEV;
 566                 }
 567         }
 568
 569         return 0;
 570 }
 571
 572 void ipoib_ib_dev_flush(void *_dev)
 573 {
 574         struct net_device *dev = (struct net_device *)_dev;
 575         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv;
 576
 577         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 578                 return;
 579
 580         ipoib_dbg(priv, "flushing\n");
 581
 582         ipoib_ib_dev_down(dev);
 583
 584         /*
 585          * The device could have been brought down between the start and when
 586          * we get here, don't bring it back up if it's not configured up
 587          */
 588         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 589                 ipoib_ib_dev_up(dev);
 590
 591         /* Flush any child interfaces too */
 592         list_for_each_entry(cpriv, &priv->child_intfs, list)
 593                 ipoib_ib_dev_flush(&cpriv->dev);
 594 }
 595
 596 void ipoib_ib_dev_cleanup(struct net_device *dev)
 597 {
 598         struct ipoib_dev_priv *priv = netdev_priv(dev);
 599
 600         ipoib_dbg(priv, "cleaning up ib_dev\n");
 601
 602         ipoib_mcast_stop_thread(dev);
 603
 604         /* Delete the broadcast address and the local address */
 605         ipoib_mcast_dev_down(dev);
 606
 607         ipoib_transport_dev_cleanup(dev);
 608 }
 609
 610 /*
 611  * Delayed P_Key Assigment Interim Support
 612  *
 613  * The following is initial implementation of delayed P_Key assigment
 614  * mechanism. It is using the same approach implemented for the multicast
 615  * group join. The single goal of this implementation is to quickly address
 616  * Bug #2507. This implementation will probably be removed when the P_Key
 617  * change async notification is available.
 618  */
 619 int ipoib_open(struct net_device *dev);
 620
 621 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
 622 {
 623         struct ipoib_dev_priv *priv = netdev_priv(dev);
 624         u16 pkey_index = 0;
 625
 626         if (ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
 627                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 628         else
 629                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 630 }
 631
 632 void ipoib_pkey_poll(void *dev_ptr)
 633 {
 634         struct net_device *dev = dev_ptr;
 635         struct ipoib_dev_priv *priv = netdev_priv(dev);
 636
 637         ipoib_pkey_dev_check_presence(dev);
 638
 639         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 640                 ipoib_open(dev);
 641         else {
 642                 down(&pkey_sem);
 643                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
 644                         queue_delayed_work(ipoib_workqueue,
 645                                            &priv->pkey_task,
 646                                            HZ);
 647                 up(&pkey_sem);
 648         }
 649 }
 650
 651 int ipoib_pkey_dev_delay_open(struct net_device *dev)
 652 {
 653         struct ipoib_dev_priv *priv = netdev_priv(dev);
 654
 655         /* Look for the interface pkey value in the IB Port P_Key table and */
 656         /* set the interface pkey assigment flag                            */
 657         ipoib_pkey_dev_check_presence(dev);
 658
 659         /* P_Key value not assigned yet - start polling */
 660         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 661                 down(&pkey_sem);
 662                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
 663                 queue_delayed_work(ipoib_workqueue,
 664                                    &priv->pkey_task,
 665                                    HZ);
 666                 up(&pkey_sem);
 667                 return 1;
 668         }
 669
 670         return 0;
 671 }