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 static DEFINE_MUTEX(pkey_mutex);
  54
  55 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
  56                                  struct ib_pd *pd, struct ib_ah_attr *attr)
  57 {
  58         struct ipoib_ah *ah;
  59
  60         ah = kmalloc(sizeof *ah, GFP_KERNEL);
  61         if (!ah)
  62                 return NULL;
  63
  64         ah->dev       = dev;
  65         ah->last_send = 0;
  66         kref_init(&ah->ref);
  67
  68         ah->ah = ib_create_ah(pd, attr);
  69         if (IS_ERR(ah->ah)) {
  70                 kfree(ah);
  71                 ah = NULL;
  72         } else
  73                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
  74
  75         return ah;
  76 }
  77
  78 void ipoib_free_ah(struct kref *kref)
  79 {
  80         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
  81         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
  82
  83         unsigned long flags;
  84
  85         spin_lock_irqsave(&priv->lock, flags);
  86         list_add_tail(&ah->list, &priv->dead_ahs);
  87         spin_unlock_irqrestore(&priv->lock, flags);
  88 }
  89
  90 static int ipoib_ib_post_receive(struct net_device *dev, int id)
  91 {
  92         struct ipoib_dev_priv *priv = netdev_priv(dev);
  93         struct ib_sge list;
  94         struct ib_recv_wr param;
  95         struct ib_recv_wr *bad_wr;
  96         int ret;
  97
  98         list.addr     = priv->rx_ring[id].mapping;
  99         list.length   = IPOIB_BUF_SIZE;
 100         list.lkey     = priv->mr->lkey;
 101
 102         param.next    = NULL;
 103         param.wr_id   = id | IPOIB_OP_RECV;
 104         param.sg_list = &list;
 105         param.num_sge = 1;
 106
 107         ret = ib_post_recv(priv->qp, &param, &bad_wr);
 108         if (unlikely(ret)) {
 109                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
 110                 ib_dma_unmap_single(priv->ca, priv->rx_ring[id].mapping,
 111                                     IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
 112                 dev_kfree_skb_any(priv->rx_ring[id].skb);
 113                 priv->rx_ring[id].skb = NULL;
 114         }
 115
 116         return ret;
 117 }
 118
 119 static int ipoib_alloc_rx_skb(struct net_device *dev, int id)
 120 {
 121         struct ipoib_dev_priv *priv = netdev_priv(dev);
 122         struct sk_buff *skb;
 123         u64 addr;
 124
 125         skb = dev_alloc_skb(IPOIB_BUF_SIZE + 4);
 126         if (!skb)
 127                 return -ENOMEM;
 128
 129         /*
 130          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
 131          * header.  So we need 4 more bytes to get to 48 and align the
 132          * IP header to a multiple of 16.
 133          */
 134         skb_reserve(skb, 4);
 135
 136         addr = ib_dma_map_single(priv->ca, skb->data, IPOIB_BUF_SIZE,
 137                                  DMA_FROM_DEVICE);
 138         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 139                 dev_kfree_skb_any(skb);
 140                 return -EIO;
 141         }
 142
 143         priv->rx_ring[id].skb     = skb;
 144         priv->rx_ring[id].mapping = addr;
 145
 146         return 0;
 147 }
 148
 149 static int ipoib_ib_post_receives(struct net_device *dev)
 150 {
 151         struct ipoib_dev_priv *priv = netdev_priv(dev);
 152         int i;
 153
 154         for (i = 0; i < ipoib_recvq_size; ++i) {
 155                 if (ipoib_alloc_rx_skb(dev, i)) {
 156                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 157                         return -ENOMEM;
 158                 }
 159                 if (ipoib_ib_post_receive(dev, i)) {
 160                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 161                         return -EIO;
 162                 }
 163         }
 164
 165         return 0;
 166 }
 167
 168 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 169 {
 170         struct ipoib_dev_priv *priv = netdev_priv(dev);
 171         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
 172         struct sk_buff *skb;
 173         u64 addr;
 174
 175         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
 176                        wr_id, wc->status);
 177
 178         if (unlikely(wr_id >= ipoib_recvq_size)) {
 179                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
 180                            wr_id, ipoib_recvq_size);
 181                 return;
 182         }
 183
 184         skb  = priv->rx_ring[wr_id].skb;
 185         addr = priv->rx_ring[wr_id].mapping;
 186
 187         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 188                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 189                         ipoib_warn(priv, "failed recv event "
 190                                    "(status=%d, wrid=%d vend_err %x)\n",
 191                                    wc->status, wr_id, wc->vendor_err);
 192                 ib_dma_unmap_single(priv->ca, addr,
 193                                     IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
 194                 dev_kfree_skb_any(skb);
 195                 priv->rx_ring[wr_id].skb = NULL;
 196                 return;
 197         }
 198
 199         /*
 200          * If we can't allocate a new RX buffer, dump
 201          * this packet and reuse the old buffer.
 202          */
 203         if (unlikely(ipoib_alloc_rx_skb(dev, wr_id))) {
 204                 ++priv->stats.rx_dropped;
 205                 goto repost;
 206         }
 207
 208         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 209                        wc->byte_len, wc->slid);
 210
 211         ib_dma_unmap_single(priv->ca, addr, IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
 212
 213         skb_put(skb, wc->byte_len);
 214         skb_pull(skb, IB_GRH_BYTES);
 215
 216         if (wc->slid != priv->local_lid ||
 217             wc->src_qp != priv->qp->qp_num) {
 218                 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 219                 skb_reset_mac_header(skb);
 220                 skb_pull(skb, IPOIB_ENCAP_LEN);
 221
 222                 dev->last_rx = jiffies;
 223                 ++priv->stats.rx_packets;
 224                 priv->stats.rx_bytes += skb->len;
 225
 226                 skb->dev = dev;
 227                 /* XXX get correct PACKET_ type here */
 228                 skb->pkt_type = PACKET_HOST;
 229                 netif_rx_ni(skb);
 230         } else {
 231                 ipoib_dbg_data(priv, "dropping loopback packet\n");
 232                 dev_kfree_skb_any(skb);
 233         }
 234
 235 repost:
 236         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 237                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 238                            "for buf %d\n", wr_id);
 239 }
 240
 241 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 242 {
 243         struct ipoib_dev_priv *priv = netdev_priv(dev);
 244         unsigned int wr_id = wc->wr_id;
 245         struct ipoib_tx_buf *tx_req;
 246         unsigned long flags;
 247
 248         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 249                        wr_id, wc->status);
 250
 251         if (unlikely(wr_id >= ipoib_sendq_size)) {
 252                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 253                            wr_id, ipoib_sendq_size);
 254                 return;
 255         }
 256
 257         tx_req = &priv->tx_ring[wr_id];
 258
 259         ib_dma_unmap_single(priv->ca, tx_req->mapping,
 260                             tx_req->skb->len, DMA_TO_DEVICE);
 261
 262         ++priv->stats.tx_packets;
 263         priv->stats.tx_bytes += tx_req->skb->len;
 264
 265         dev_kfree_skb_any(tx_req->skb);
 266
 267         spin_lock_irqsave(&priv->tx_lock, flags);
 268         ++priv->tx_tail;
 269         if (unlikely(test_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags)) &&
 270             priv->tx_head - priv->tx_tail <= ipoib_sendq_size >> 1) {
 271                 clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
 272                 netif_wake_queue(dev);
 273         }
 274         spin_unlock_irqrestore(&priv->tx_lock, flags);
 275
 276         if (wc->status != IB_WC_SUCCESS &&
 277             wc->status != IB_WC_WR_FLUSH_ERR)
 278                 ipoib_warn(priv, "failed send event "
 279                            "(status=%d, wrid=%d vend_err %x)\n",
 280                            wc->status, wr_id, wc->vendor_err);
 281 }
 282
 283 static void ipoib_ib_handle_wc(struct net_device *dev, struct ib_wc *wc)
 284 {
 285         if (wc->wr_id & IPOIB_CM_OP_SRQ)
 286                 ipoib_cm_handle_rx_wc(dev, wc);
 287         else if (wc->wr_id & IPOIB_OP_RECV)
 288                 ipoib_ib_handle_rx_wc(dev, wc);
 289         else
 290                 ipoib_ib_handle_tx_wc(dev, wc);
 291 }
 292
 293 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 294 {
 295         struct net_device *dev = (struct net_device *) dev_ptr;
 296         struct ipoib_dev_priv *priv = netdev_priv(dev);
 297         int n, i;
 298
 299         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 300         do {
 301                 n = ib_poll_cq(cq, IPOIB_NUM_WC, priv->ibwc);
 302                 for (i = 0; i < n; ++i)
 303                         ipoib_ib_handle_wc(dev, priv->ibwc + i);
 304         } while (n == IPOIB_NUM_WC);
 305 }
 306
 307 static inline int post_send(struct ipoib_dev_priv *priv,
 308                             unsigned int wr_id,
 309                             struct ib_ah *address, u32 qpn,
 310                             u64 addr, int len)
 311 {
 312         struct ib_send_wr *bad_wr;
 313
 314         priv->tx_sge.addr             = addr;
 315         priv->tx_sge.length           = len;
 316
 317         priv->tx_wr.wr_id             = wr_id;
 318         priv->tx_wr.wr.ud.remote_qpn  = qpn;
 319         priv->tx_wr.wr.ud.ah          = address;
 320
 321         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 322 }
 323
 324 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 325                 struct ipoib_ah *address, u32 qpn)
 326 {
 327         struct ipoib_dev_priv *priv = netdev_priv(dev);
 328         struct ipoib_tx_buf *tx_req;
 329         u64 addr;
 330
 331         if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 332                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 333                            skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 334                 ++priv->stats.tx_dropped;
 335                 ++priv->stats.tx_errors;
 336                 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 337                 return;
 338         }
 339
 340         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 341                        skb->len, address, qpn);
 342
 343         /*
 344          * We put the skb into the tx_ring _before_ we call post_send()
 345          * because it's entirely possible that the completion handler will
 346          * run before we execute anything after the post_send().  That
 347          * means we have to make sure everything is properly recorded and
 348          * our state is consistent before we call post_send().
 349          */
 350         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 351         tx_req->skb = skb;
 352         addr = ib_dma_map_single(priv->ca, skb->data, skb->len,
 353                                  DMA_TO_DEVICE);
 354         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 355                 ++priv->stats.tx_errors;
 356                 dev_kfree_skb_any(skb);
 357                 return;
 358         }
 359         tx_req->mapping = addr;
 360
 361         if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 362                                address->ah, qpn, addr, skb->len))) {
 363                 ipoib_warn(priv, "post_send failed\n");
 364                 ++priv->stats.tx_errors;
 365                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
 366                 dev_kfree_skb_any(skb);
 367         } else {
 368                 dev->trans_start = jiffies;
 369
 370                 address->last_send = priv->tx_head;
 371                 ++priv->tx_head;
 372
 373                 if (priv->tx_head - priv->tx_tail == ipoib_sendq_size) {
 374                         ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 375                         netif_stop_queue(dev);
 376                         set_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
 377                 }
 378         }
 379 }
 380
 381 static void __ipoib_reap_ah(struct net_device *dev)
 382 {
 383         struct ipoib_dev_priv *priv = netdev_priv(dev);
 384         struct ipoib_ah *ah, *tah;
 385         LIST_HEAD(remove_list);
 386
 387         spin_lock_irq(&priv->tx_lock);
 388         spin_lock(&priv->lock);
 389         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 390                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 391                         list_del(&ah->list);
 392                         ib_destroy_ah(ah->ah);
 393                         kfree(ah);
 394                 }
 395         spin_unlock(&priv->lock);
 396         spin_unlock_irq(&priv->tx_lock);
 397 }
 398
 399 void ipoib_reap_ah(struct work_struct *work)
 400 {
 401         struct ipoib_dev_priv *priv =
 402                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 403         struct net_device *dev = priv->dev;
 404
 405         __ipoib_reap_ah(dev);
 406
 407         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 408                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
 409 }
 410
 411 int ipoib_ib_dev_open(struct net_device *dev)
 412 {
 413         struct ipoib_dev_priv *priv = netdev_priv(dev);
 414         int ret;
 415
 416         ret = ipoib_init_qp(dev);
 417         if (ret) {
 418                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 419                 return -1;
 420         }
 421
 422         ret = ipoib_ib_post_receives(dev);
 423         if (ret) {
 424                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 425                 ipoib_ib_dev_stop(dev);
 426                 return -1;
 427         }
 428
 429         ret = ipoib_cm_dev_open(dev);
 430         if (ret) {
 431                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 432                 ipoib_ib_dev_stop(dev);
 433                 return -1;
 434         }
 435
 436         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 437         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
 438
 439         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 440
 441         return 0;
 442 }
 443
 444 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
 445 {
 446         struct ipoib_dev_priv *priv = netdev_priv(dev);
 447         u16 pkey_index = 0;
 448
 449         if (ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
 450                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 451         else
 452                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 453 }
 454
 455 int ipoib_ib_dev_up(struct net_device *dev)
 456 {
 457         struct ipoib_dev_priv *priv = netdev_priv(dev);
 458
 459         ipoib_pkey_dev_check_presence(dev);
 460
 461         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 462                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 463                 return 0;
 464         }
 465
 466         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 467
 468         return ipoib_mcast_start_thread(dev);
 469 }
 470
 471 int ipoib_ib_dev_down(struct net_device *dev, int flush)
 472 {
 473         struct ipoib_dev_priv *priv = netdev_priv(dev);
 474
 475         ipoib_dbg(priv, "downing ib_dev\n");
 476
 477         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 478         netif_carrier_off(dev);
 479
 480         /* Shutdown the P_Key thread if still active */
 481         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 482                 mutex_lock(&pkey_mutex);
 483                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
 484                 cancel_delayed_work(&priv->pkey_task);
 485                 mutex_unlock(&pkey_mutex);
 486                 if (flush)
 487                         flush_workqueue(ipoib_workqueue);
 488         }
 489
 490         ipoib_mcast_stop_thread(dev, flush);
 491         ipoib_mcast_dev_flush(dev);
 492
 493         ipoib_flush_paths(dev);
 494
 495         return 0;
 496 }
 497
 498 static int recvs_pending(struct net_device *dev)
 499 {
 500         struct ipoib_dev_priv *priv = netdev_priv(dev);
 501         int pending = 0;
 502         int i;
 503
 504         for (i = 0; i < ipoib_recvq_size; ++i)
 505                 if (priv->rx_ring[i].skb)
 506                         ++pending;
 507
 508         return pending;
 509 }
 510
 511 int ipoib_ib_dev_stop(struct net_device *dev)
 512 {
 513         struct ipoib_dev_priv *priv = netdev_priv(dev);
 514         struct ib_qp_attr qp_attr;
 515         unsigned long begin;
 516         struct ipoib_tx_buf *tx_req;
 517         int i;
 518
 519         clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 520
 521         ipoib_cm_dev_stop(dev);
 522
 523         /*
 524          * Move our QP to the error state and then reinitialize in
 525          * when all work requests have completed or have been flushed.
 526          */
 527         qp_attr.qp_state = IB_QPS_ERR;
 528         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 529                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 530
 531         /* Wait for all sends and receives to complete */
 532         begin = jiffies;
 533
 534         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 535                 if (time_after(jiffies, begin + 5 * HZ)) {
 536                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 537                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 538
 539                         /*
 540                          * assume the HW is wedged and just free up
 541                          * all our pending work requests.
 542                          */
 543                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 544                                 tx_req = &priv->tx_ring[priv->tx_tail &
 545                                                         (ipoib_sendq_size - 1)];
 546                                 ib_dma_unmap_single(priv->ca,
 547                                                     tx_req->mapping,
 548                                                     tx_req->skb->len,
 549                                                     DMA_TO_DEVICE);
 550                                 dev_kfree_skb_any(tx_req->skb);
 551                                 ++priv->tx_tail;
 552                         }
 553
 554                         for (i = 0; i < ipoib_recvq_size; ++i) {
 555                                 struct ipoib_rx_buf *rx_req;
 556
 557                                 rx_req = &priv->rx_ring[i];
 558                                 if (!rx_req->skb)
 559                                         continue;
 560                                 ib_dma_unmap_single(priv->ca,
 561                                                     rx_req->mapping,
 562                                                     IPOIB_BUF_SIZE,
 563                                                     DMA_FROM_DEVICE);
 564                                 dev_kfree_skb_any(rx_req->skb);
 565                                 rx_req->skb = NULL;
 566                         }
 567
 568                         goto timeout;
 569                 }
 570
 571                 msleep(1);
 572         }
 573
 574         ipoib_dbg(priv, "All sends and receives done.\n");
 575
 576 timeout:
 577         qp_attr.qp_state = IB_QPS_RESET;
 578         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 579                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 580
 581         /* Wait for all AHs to be reaped */
 582         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 583         cancel_delayed_work(&priv->ah_reap_task);
 584         flush_workqueue(ipoib_workqueue);
 585
 586         begin = jiffies;
 587
 588         while (!list_empty(&priv->dead_ahs)) {
 589                 __ipoib_reap_ah(dev);
 590
 591                 if (time_after(jiffies, begin + HZ)) {
 592                         ipoib_warn(priv, "timing out; will leak address handles\n");
 593                         break;
 594                 }
 595
 596                 msleep(1);
 597         }
 598
 599         return 0;
 600 }
 601
 602 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 603 {
 604         struct ipoib_dev_priv *priv = netdev_priv(dev);
 605
 606         priv->ca = ca;
 607         priv->port = port;
 608         priv->qp = NULL;
 609
 610         if (ipoib_transport_dev_init(dev, ca)) {
 611                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 612                 return -ENODEV;
 613         }
 614
 615         if (dev->flags & IFF_UP) {
 616                 if (ipoib_ib_dev_open(dev)) {
 617                         ipoib_transport_dev_cleanup(dev);
 618                         return -ENODEV;
 619                 }
 620         }
 621
 622         return 0;
 623 }
 624
 625 void ipoib_ib_dev_flush(struct work_struct *work)
 626 {
 627         struct ipoib_dev_priv *cpriv, *priv =
 628                 container_of(work, struct ipoib_dev_priv, flush_task);
 629         struct net_device *dev = priv->dev;
 630
 631         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) ) {
 632                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
 633                 return;
 634         }
 635
 636         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 637                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
 638                 return;
 639         }
 640
 641         ipoib_dbg(priv, "flushing\n");
 642
 643         ipoib_ib_dev_down(dev, 0);
 644
 645         /*
 646          * The device could have been brought down between the start and when
 647          * we get here, don't bring it back up if it's not configured up
 648          */
 649         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 650                 ipoib_ib_dev_up(dev);
 651                 ipoib_mcast_restart_task(&priv->restart_task);
 652         }
 653
 654         mutex_lock(&priv->vlan_mutex);
 655
 656         /* Flush any child interfaces too */
 657         list_for_each_entry(cpriv, &priv->child_intfs, list)
 658                 ipoib_ib_dev_flush(&cpriv->flush_task);
 659
 660         mutex_unlock(&priv->vlan_mutex);
 661 }
 662
 663 void ipoib_ib_dev_cleanup(struct net_device *dev)
 664 {
 665         struct ipoib_dev_priv *priv = netdev_priv(dev);
 666
 667         ipoib_dbg(priv, "cleaning up ib_dev\n");
 668
 669         ipoib_mcast_stop_thread(dev, 1);
 670         ipoib_mcast_dev_flush(dev);
 671
 672         ipoib_transport_dev_cleanup(dev);
 673 }
 674
 675 /*
 676  * Delayed P_Key Assigment Interim Support
 677  *
 678  * The following is initial implementation of delayed P_Key assigment
 679  * mechanism. It is using the same approach implemented for the multicast
 680  * group join. The single goal of this implementation is to quickly address
 681  * Bug #2507. This implementation will probably be removed when the P_Key
 682  * change async notification is available.
 683  */
 684
 685 void ipoib_pkey_poll(struct work_struct *work)
 686 {
 687         struct ipoib_dev_priv *priv =
 688                 container_of(work, struct ipoib_dev_priv, pkey_task.work);
 689         struct net_device *dev = priv->dev;
 690
 691         ipoib_pkey_dev_check_presence(dev);
 692
 693         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 694                 ipoib_open(dev);
 695         else {
 696                 mutex_lock(&pkey_mutex);
 697                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
 698                         queue_delayed_work(ipoib_workqueue,
 699                                            &priv->pkey_task,
 700                                            HZ);
 701                 mutex_unlock(&pkey_mutex);
 702         }
 703 }
 704
 705 int ipoib_pkey_dev_delay_open(struct net_device *dev)
 706 {
 707         struct ipoib_dev_priv *priv = netdev_priv(dev);
 708
 709         /* Look for the interface pkey value in the IB Port P_Key table and */
 710         /* set the interface pkey assigment flag                            */
 711         ipoib_pkey_dev_check_presence(dev);
 712
 713         /* P_Key value not assigned yet - start polling */
 714         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 715                 mutex_lock(&pkey_mutex);
 716                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
 717                 queue_delayed_work(ipoib_workqueue,
 718                                    &priv->pkey_task,
 719                                    HZ);
 720                 mutex_unlock(&pkey_mutex);
 721                 return 1;
 722         }
 723
 724         return 0;
 725 }