drivers/infiniband/ulp/ipoib/ipoib_ib.c

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