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 DEFINE_MUTEX(pkey_mutex);
  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 int ipoib_ib_post_receive(struct net_device *dev, int id)
  99 {
 100         struct ipoib_dev_priv *priv = netdev_priv(dev);
 101         struct ib_sge list;
 102         struct ib_recv_wr param;
 103         struct ib_recv_wr *bad_wr;
 104         int ret;
 105
 106         list.addr     = priv->rx_ring[id].mapping;
 107         list.length   = IPOIB_BUF_SIZE;
 108         list.lkey     = priv->mr->lkey;
 109
 110         param.next    = NULL;
 111         param.wr_id   = id | IPOIB_OP_RECV;
 112         param.sg_list = &list;
 113         param.num_sge = 1;
 114
 115         ret = ib_post_recv(priv->qp, &param, &bad_wr);
 116         if (unlikely(ret)) {
 117                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
 118                 dma_unmap_single(priv->ca->dma_device,
 119                                  priv->rx_ring[id].mapping,
 120                                  IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
 121                 dev_kfree_skb_any(priv->rx_ring[id].skb);
 122                 priv->rx_ring[id].skb = NULL;
 123         }
 124
 125         return ret;
 126 }
 127
 128 static int ipoib_alloc_rx_skb(struct net_device *dev, int id)
 129 {
 130         struct ipoib_dev_priv *priv = netdev_priv(dev);
 131         struct sk_buff *skb;
 132         dma_addr_t addr;
 133
 134         skb = dev_alloc_skb(IPOIB_BUF_SIZE + 4);
 135         if (!skb)
 136                 return -ENOMEM;
 137
 138         /*
 139          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
 140          * header.  So we need 4 more bytes to get to 48 and align the
 141          * IP header to a multiple of 16.
 142          */
 143         skb_reserve(skb, 4);
 144
 145         addr = dma_map_single(priv->ca->dma_device,
 146                               skb->data, IPOIB_BUF_SIZE,
 147                               DMA_FROM_DEVICE);
 148         if (unlikely(dma_mapping_error(addr))) {
 149                 dev_kfree_skb_any(skb);
 150                 return -EIO;
 151         }
 152
 153         priv->rx_ring[id].skb     = skb;
 154         priv->rx_ring[id].mapping = addr;
 155
 156         return 0;
 157 }
 158
 159 static int ipoib_ib_post_receives(struct net_device *dev)
 160 {
 161         struct ipoib_dev_priv *priv = netdev_priv(dev);
 162         int i;
 163
 164         for (i = 0; i < IPOIB_RX_RING_SIZE; ++i) {
 165                 if (ipoib_alloc_rx_skb(dev, i)) {
 166                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 167                         return -ENOMEM;
 168                 }
 169                 if (ipoib_ib_post_receive(dev, i)) {
 170                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 171                         return -EIO;
 172                 }
 173         }
 174
 175         return 0;
 176 }
 177
 178 static void ipoib_ib_handle_wc(struct net_device *dev,
 179                                struct ib_wc *wc)
 180 {
 181         struct ipoib_dev_priv *priv = netdev_priv(dev);
 182         unsigned int wr_id = wc->wr_id;
 183
 184         ipoib_dbg_data(priv, "called: id %d, op %d, status: %d\n",
 185                        wr_id, wc->opcode, wc->status);
 186
 187         if (wr_id & IPOIB_OP_RECV) {
 188                 wr_id &= ~IPOIB_OP_RECV;
 189
 190                 if (wr_id < IPOIB_RX_RING_SIZE) {
 191                         struct sk_buff *skb  = priv->rx_ring[wr_id].skb;
 192                         dma_addr_t      addr = priv->rx_ring[wr_id].mapping;
 193
 194                         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 195                                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 196                                         ipoib_warn(priv, "failed recv event "
 197                                                    "(status=%d, wrid=%d vend_err %x)\n",
 198                                                    wc->status, wr_id, wc->vendor_err);
 199                                 dma_unmap_single(priv->ca->dma_device, addr,
 200                                                  IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
 201                                 dev_kfree_skb_any(skb);
 202                                 priv->rx_ring[wr_id].skb = NULL;
 203                                 return;
 204                         }
 205
 206                         /*
 207                          * If we can't allocate a new RX buffer, dump
 208                          * this packet and reuse the old buffer.
 209                          */
 210                         if (unlikely(ipoib_alloc_rx_skb(dev, wr_id))) {
 211                                 ++priv->stats.rx_dropped;
 212                                 goto repost;
 213                         }
 214
 215                         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 216                                        wc->byte_len, wc->slid);
 217
 218                         dma_unmap_single(priv->ca->dma_device, addr,
 219                                          IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
 220
 221                         skb_put(skb, wc->byte_len);
 222                         skb_pull(skb, IB_GRH_BYTES);
 223
 224                         if (wc->slid != priv->local_lid ||
 225                             wc->src_qp != priv->qp->qp_num) {
 226                                 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 227                                 skb->mac.raw = skb->data;
 228                                 skb_pull(skb, IPOIB_ENCAP_LEN);
 229
 230                                 dev->last_rx = jiffies;
 231                                 ++priv->stats.rx_packets;
 232                                 priv->stats.rx_bytes += skb->len;
 233
 234                                 skb->dev = dev;
 235                                 /* XXX get correct PACKET_ type here */
 236                                 skb->pkt_type = PACKET_HOST;
 237                                 netif_rx_ni(skb);
 238                         } else {
 239                                 ipoib_dbg_data(priv, "dropping loopback packet\n");
 240                                 dev_kfree_skb_any(skb);
 241                         }
 242
 243                 repost:
 244                         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 245                                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 246                                            "for buf %d\n", wr_id);
 247                 } else
 248                         ipoib_warn(priv, "completion event with wrid %d\n",
 249                                    wr_id);
 250
 251         } else {
 252                 struct ipoib_tx_buf *tx_req;
 253                 unsigned long flags;
 254
 255                 if (wr_id >= IPOIB_TX_RING_SIZE) {
 256                         ipoib_warn(priv, "completion event with wrid %d (> %d)\n",
 257                                    wr_id, IPOIB_TX_RING_SIZE);
 258                         return;
 259                 }
 260
 261                 ipoib_dbg_data(priv, "send complete, wrid %d\n", wr_id);
 262
 263                 tx_req = &priv->tx_ring[wr_id];
 264
 265                 dma_unmap_single(priv->ca->dma_device,
 266                                  pci_unmap_addr(tx_req, mapping),
 267                                  tx_req->skb->len,
 268                                  DMA_TO_DEVICE);
 269
 270                 ++priv->stats.tx_packets;
 271                 priv->stats.tx_bytes += tx_req->skb->len;
 272
 273                 dev_kfree_skb_any(tx_req->skb);
 274
 275                 spin_lock_irqsave(&priv->tx_lock, flags);
 276                 ++priv->tx_tail;
 277                 if (netif_queue_stopped(dev) &&
 278                     priv->tx_head - priv->tx_tail <= IPOIB_TX_RING_SIZE / 2)
 279                         netif_wake_queue(dev);
 280                 spin_unlock_irqrestore(&priv->tx_lock, flags);
 281
 282                 if (wc->status != IB_WC_SUCCESS &&
 283                     wc->status != IB_WC_WR_FLUSH_ERR)
 284                         ipoib_warn(priv, "failed send event "
 285                                    "(status=%d, wrid=%d vend_err %x)\n",
 286                                    wc->status, wr_id, wc->vendor_err);
 287         }
 288 }
 289
 290 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 291 {
 292         struct net_device *dev = (struct net_device *) dev_ptr;
 293         struct ipoib_dev_priv *priv = netdev_priv(dev);
 294         int n, i;
 295
 296         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 297         do {
 298                 n = ib_poll_cq(cq, IPOIB_NUM_WC, priv->ibwc);
 299                 for (i = 0; i < n; ++i)
 300                         ipoib_ib_handle_wc(dev, priv->ibwc + i);
 301         } while (n == IPOIB_NUM_WC);
 302 }
 303
 304 static inline int post_send(struct ipoib_dev_priv *priv,
 305                             unsigned int wr_id,
 306                             struct ib_ah *address, u32 qpn,
 307                             dma_addr_t addr, int len)
 308 {
 309         struct ib_send_wr *bad_wr;
 310
 311         priv->tx_sge.addr             = addr;
 312         priv->tx_sge.length           = len;
 313
 314         priv->tx_wr.wr_id             = wr_id;
 315         priv->tx_wr.wr.ud.remote_qpn  = qpn;
 316         priv->tx_wr.wr.ud.ah          = address;
 317
 318         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 319 }
 320
 321 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 322                 struct ipoib_ah *address, u32 qpn)
 323 {
 324         struct ipoib_dev_priv *priv = netdev_priv(dev);
 325         struct ipoib_tx_buf *tx_req;
 326         dma_addr_t addr;
 327
 328         if (skb->len > dev->mtu + INFINIBAND_ALEN) {
 329                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 330                            skb->len, dev->mtu + INFINIBAND_ALEN);
 331                 ++priv->stats.tx_dropped;
 332                 ++priv->stats.tx_errors;
 333                 dev_kfree_skb_any(skb);
 334                 return;
 335         }
 336
 337         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 338                        skb->len, address, qpn);
 339
 340         /*
 341          * We put the skb into the tx_ring _before_ we call post_send()
 342          * because it's entirely possible that the completion handler will
 343          * run before we execute anything after the post_send().  That
 344          * means we have to make sure everything is properly recorded and
 345          * our state is consistent before we call post_send().
 346          */
 347         tx_req = &priv->tx_ring[priv->tx_head & (IPOIB_TX_RING_SIZE - 1)];
 348         tx_req->skb = skb;
 349         addr = dma_map_single(priv->ca->dma_device, skb->data, skb->len,
 350                               DMA_TO_DEVICE);
 351         pci_unmap_addr_set(tx_req, mapping, addr);
 352
 353         if (unlikely(post_send(priv, priv->tx_head & (IPOIB_TX_RING_SIZE - 1),
 354                                address->ah, qpn, addr, skb->len))) {
 355                 ipoib_warn(priv, "post_send failed\n");
 356                 ++priv->stats.tx_errors;
 357                 dma_unmap_single(priv->ca->dma_device, addr, skb->len,
 358                                  DMA_TO_DEVICE);
 359                 dev_kfree_skb_any(skb);
 360         } else {
 361                 dev->trans_start = jiffies;
 362
 363                 address->last_send = priv->tx_head;
 364                 ++priv->tx_head;
 365
 366                 if (priv->tx_head - priv->tx_tail == IPOIB_TX_RING_SIZE) {
 367                         ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 368                         netif_stop_queue(dev);
 369                 }
 370         }
 371 }
 372
 373 static void __ipoib_reap_ah(struct net_device *dev)
 374 {
 375         struct ipoib_dev_priv *priv = netdev_priv(dev);
 376         struct ipoib_ah *ah, *tah;
 377         LIST_HEAD(remove_list);
 378
 379         spin_lock_irq(&priv->lock);
 380         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 381                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 382                         list_del(&ah->list);
 383                         list_add_tail(&ah->list, &remove_list);
 384                 }
 385         spin_unlock_irq(&priv->lock);
 386
 387         list_for_each_entry_safe(ah, tah, &remove_list, list) {
 388                 ipoib_dbg(priv, "Reaping ah %p\n", ah->ah);
 389                 ib_destroy_ah(ah->ah);
 390                 kfree(ah);
 391         }
 392 }
 393
 394 void ipoib_reap_ah(void *dev_ptr)
 395 {
 396         struct net_device *dev = dev_ptr;
 397         struct ipoib_dev_priv *priv = netdev_priv(dev);
 398
 399         __ipoib_reap_ah(dev);
 400
 401         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 402                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
 403 }
 404
 405 int ipoib_ib_dev_open(struct net_device *dev)
 406 {
 407         struct ipoib_dev_priv *priv = netdev_priv(dev);
 408         int ret;
 409
 410         ret = ipoib_init_qp(dev);
 411         if (ret) {
 412                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 413                 return -1;
 414         }
 415
 416         ret = ipoib_ib_post_receives(dev);
 417         if (ret) {
 418                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 419                 ipoib_ib_dev_stop(dev);
 420                 return -1;
 421         }
 422
 423         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 424         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
 425
 426         return 0;
 427 }
 428
 429 int ipoib_ib_dev_up(struct net_device *dev)
 430 {
 431         struct ipoib_dev_priv *priv = netdev_priv(dev);
 432
 433         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 434
 435         return ipoib_mcast_start_thread(dev);
 436 }
 437
 438 int ipoib_ib_dev_down(struct net_device *dev, int flush)
 439 {
 440         struct ipoib_dev_priv *priv = netdev_priv(dev);
 441
 442         ipoib_dbg(priv, "downing ib_dev\n");
 443
 444         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 445         netif_carrier_off(dev);
 446
 447         /* Shutdown the P_Key thread if still active */
 448         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 449                 mutex_lock(&pkey_mutex);
 450                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
 451                 cancel_delayed_work(&priv->pkey_task);
 452                 mutex_unlock(&pkey_mutex);
 453                 if (flush)
 454                         flush_workqueue(ipoib_workqueue);
 455         }
 456
 457         ipoib_mcast_stop_thread(dev, flush);
 458         ipoib_mcast_dev_flush(dev);
 459
 460         ipoib_flush_paths(dev);
 461
 462         return 0;
 463 }
 464
 465 static int recvs_pending(struct net_device *dev)
 466 {
 467         struct ipoib_dev_priv *priv = netdev_priv(dev);
 468         int pending = 0;
 469         int i;
 470
 471         for (i = 0; i < IPOIB_RX_RING_SIZE; ++i)
 472                 if (priv->rx_ring[i].skb)
 473                         ++pending;
 474
 475         return pending;
 476 }
 477
 478 int ipoib_ib_dev_stop(struct net_device *dev)
 479 {
 480         struct ipoib_dev_priv *priv = netdev_priv(dev);
 481         struct ib_qp_attr qp_attr;
 482         unsigned long begin;
 483         struct ipoib_tx_buf *tx_req;
 484         int i;
 485
 486         /*
 487          * Move our QP to the error state and then reinitialize in
 488          * when all work requests have completed or have been flushed.
 489          */
 490         qp_attr.qp_state = IB_QPS_ERR;
 491         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 492                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 493
 494         /* Wait for all sends and receives to complete */
 495         begin = jiffies;
 496
 497         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 498                 if (time_after(jiffies, begin + 5 * HZ)) {
 499                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 500                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 501
 502                         /*
 503                          * assume the HW is wedged and just free up
 504                          * all our pending work requests.
 505                          */
 506                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 507                                 tx_req = &priv->tx_ring[priv->tx_tail &
 508                                                         (IPOIB_TX_RING_SIZE - 1)];
 509                                 dma_unmap_single(priv->ca->dma_device,
 510                                                  pci_unmap_addr(tx_req, mapping),
 511                                                  tx_req->skb->len,
 512                                                  DMA_TO_DEVICE);
 513                                 dev_kfree_skb_any(tx_req->skb);
 514                                 ++priv->tx_tail;
 515                         }
 516
 517                         for (i = 0; i < IPOIB_RX_RING_SIZE; ++i)
 518                                 if (priv->rx_ring[i].skb) {
 519                                         dma_unmap_single(priv->ca->dma_device,
 520                                                          pci_unmap_addr(&priv->rx_ring[i],
 521                                                                         mapping),
 522                                                          IPOIB_BUF_SIZE,
 523                                                          DMA_FROM_DEVICE);
 524                                         dev_kfree_skb_any(priv->rx_ring[i].skb);
 525                                         priv->rx_ring[i].skb = NULL;
 526                                 }
 527
 528                         goto timeout;
 529                 }
 530
 531                 msleep(1);
 532         }
 533
 534         ipoib_dbg(priv, "All sends and receives done.\n");
 535
 536 timeout:
 537         qp_attr.qp_state = IB_QPS_RESET;
 538         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 539                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 540
 541         /* Wait for all AHs to be reaped */
 542         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 543         cancel_delayed_work(&priv->ah_reap_task);
 544         flush_workqueue(ipoib_workqueue);
 545
 546         begin = jiffies;
 547
 548         while (!list_empty(&priv->dead_ahs)) {
 549                 __ipoib_reap_ah(dev);
 550
 551                 if (time_after(jiffies, begin + HZ)) {
 552                         ipoib_warn(priv, "timing out; will leak address handles\n");
 553                         break;
 554                 }
 555
 556                 msleep(1);
 557         }
 558
 559         return 0;
 560 }
 561
 562 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 563 {
 564         struct ipoib_dev_priv *priv = netdev_priv(dev);
 565
 566         priv->ca = ca;
 567         priv->port = port;
 568         priv->qp = NULL;
 569
 570         if (ipoib_transport_dev_init(dev, ca)) {
 571                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 572                 return -ENODEV;
 573         }
 574
 575         if (dev->flags & IFF_UP) {
 576                 if (ipoib_ib_dev_open(dev)) {
 577                         ipoib_transport_dev_cleanup(dev);
 578                         return -ENODEV;
 579                 }
 580         }
 581
 582         return 0;
 583 }
 584
 585 void ipoib_ib_dev_flush(void *_dev)
 586 {
 587         struct net_device *dev = (struct net_device *)_dev;
 588         struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv;
 589
 590         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 591                 return;
 592
 593         ipoib_dbg(priv, "flushing\n");
 594
 595         ipoib_ib_dev_down(dev, 0);
 596
 597         /*
 598          * The device could have been brought down between the start and when
 599          * we get here, don't bring it back up if it's not configured up
 600          */
 601         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 602                 ipoib_ib_dev_up(dev);
 603
 604         mutex_lock(&priv->vlan_mutex);
 605
 606         /* Flush any child interfaces too */
 607         list_for_each_entry(cpriv, &priv->child_intfs, list)
 608                 ipoib_ib_dev_flush(&cpriv->dev);
 609
 610         mutex_unlock(&priv->vlan_mutex);
 611 }
 612
 613 void ipoib_ib_dev_cleanup(struct net_device *dev)
 614 {
 615         struct ipoib_dev_priv *priv = netdev_priv(dev);
 616
 617         ipoib_dbg(priv, "cleaning up ib_dev\n");
 618
 619         ipoib_mcast_stop_thread(dev, 1);
 620         ipoib_mcast_dev_flush(dev);
 621
 622         ipoib_transport_dev_cleanup(dev);
 623 }
 624
 625 /*
 626  * Delayed P_Key Assigment Interim Support
 627  *
 628  * The following is initial implementation of delayed P_Key assigment
 629  * mechanism. It is using the same approach implemented for the multicast
 630  * group join. The single goal of this implementation is to quickly address
 631  * Bug #2507. This implementation will probably be removed when the P_Key
 632  * change async notification is available.
 633  */
 634
 635 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
 636 {
 637         struct ipoib_dev_priv *priv = netdev_priv(dev);
 638         u16 pkey_index = 0;
 639
 640         if (ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
 641                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 642         else
 643                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 644 }
 645
 646 void ipoib_pkey_poll(void *dev_ptr)
 647 {
 648         struct net_device *dev = dev_ptr;
 649         struct ipoib_dev_priv *priv = netdev_priv(dev);
 650
 651         ipoib_pkey_dev_check_presence(dev);
 652
 653         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 654                 ipoib_open(dev);
 655         else {
 656                 mutex_lock(&pkey_mutex);
 657                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
 658                         queue_delayed_work(ipoib_workqueue,
 659                                            &priv->pkey_task,
 660                                            HZ);
 661                 mutex_unlock(&pkey_mutex);
 662         }
 663 }
 664
 665 int ipoib_pkey_dev_delay_open(struct net_device *dev)
 666 {
 667         struct ipoib_dev_priv *priv = netdev_priv(dev);
 668
 669         /* Look for the interface pkey value in the IB Port P_Key table and */
 670         /* set the interface pkey assigment flag                            */
 671         ipoib_pkey_dev_check_presence(dev);
 672
 673         /* P_Key value not assigned yet - start polling */
 674         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 675                 mutex_lock(&pkey_mutex);
 676                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
 677                 queue_delayed_work(ipoib_workqueue,
 678                                    &priv->pkey_task,
 679                                    HZ);
 680                 mutex_unlock(&pkey_mutex);
 681                 return 1;
 682         }
 683
 684         return 0;
 685 }