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
  36 #include <linux/delay.h>
  37 #include <linux/moduleparam.h>
  38 #include <linux/dma-mapping.h>
  39 #include <linux/slab.h>
  40
  41 #include <linux/ip.h>
  42 #include <linux/tcp.h>
  43 #include <rdma/ib_cache.h>
  44
  45 #include "ipoib.h"
  46
  47 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  48 static int data_debug_level;
  49
  50 module_param(data_debug_level, int, 0644);
  51 MODULE_PARM_DESC(data_debug_level,
  52                  "Enable data path debug tracing if > 0");
  53 #endif
  54
  55 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
  56                                  struct ib_pd *pd, struct rdma_ah_attr *attr)
  57 {
  58         struct ipoib_ah *ah;
  59         struct ib_ah *vah;
  60
  61         ah = kmalloc(sizeof(*ah), GFP_KERNEL);
  62         if (!ah)
  63                 return ERR_PTR(-ENOMEM);
  64
  65         ah->dev       = dev;
  66         ah->last_send = 0;
  67         kref_init(&ah->ref);
  68
  69         vah = rdma_create_ah(pd, attr);
  70         if (IS_ERR(vah)) {
  71                 kfree(ah);
  72                 ah = (struct ipoib_ah *)vah;
  73         } else {
  74                 ah->ah = vah;
  75                 ipoib_dbg(ipoib_priv(dev), "Created ah %p\n", ah->ah);
  76         }
  77
  78         return ah;
  79 }
  80
  81 void ipoib_free_ah(struct kref *kref)
  82 {
  83         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
  84         struct ipoib_dev_priv *priv = ipoib_priv(ah->dev);
  85
  86         unsigned long flags;
  87
  88         spin_lock_irqsave(&priv->lock, flags);
  89         list_add_tail(&ah->list, &priv->dead_ahs);
  90         spin_unlock_irqrestore(&priv->lock, flags);
  91 }
  92
  93 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
  94                                   u64 mapping[IPOIB_UD_RX_SG])
  95 {
  96         ib_dma_unmap_single(priv->ca, mapping[0],
  97                             IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
  98                             DMA_FROM_DEVICE);
  99 }
 100
 101 static int ipoib_ib_post_receive(struct net_device *dev, int id)
 102 {
 103         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 104         int ret;
 105
 106         priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
 107         priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
 108         priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
 109
 110
 111         ret = ib_post_recv(priv->qp, &priv->rx_wr, NULL);
 112         if (unlikely(ret)) {
 113                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
 114                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
 115                 dev_kfree_skb_any(priv->rx_ring[id].skb);
 116                 priv->rx_ring[id].skb = NULL;
 117         }
 118
 119         return ret;
 120 }
 121
 122 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
 123 {
 124         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 125         struct sk_buff *skb;
 126         int buf_size;
 127         u64 *mapping;
 128
 129         buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
 130
 131         skb = dev_alloc_skb(buf_size + IPOIB_HARD_LEN);
 132         if (unlikely(!skb))
 133                 return NULL;
 134
 135         /*
 136          * the IP header will be at IPOIP_HARD_LEN + IB_GRH_BYTES, that is
 137          * 64 bytes aligned
 138          */
 139         skb_reserve(skb, sizeof(struct ipoib_pseudo_header));
 140
 141         mapping = priv->rx_ring[id].mapping;
 142         mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
 143                                        DMA_FROM_DEVICE);
 144         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
 145                 goto error;
 146
 147         priv->rx_ring[id].skb = skb;
 148         return skb;
 149 error:
 150         dev_kfree_skb_any(skb);
 151         return NULL;
 152 }
 153
 154 static int ipoib_ib_post_receives(struct net_device *dev)
 155 {
 156         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 157         int i;
 158
 159         for (i = 0; i < ipoib_recvq_size; ++i) {
 160                 if (!ipoib_alloc_rx_skb(dev, i)) {
 161                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 162                         return -ENOMEM;
 163                 }
 164                 if (ipoib_ib_post_receive(dev, i)) {
 165                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 166                         return -EIO;
 167                 }
 168         }
 169
 170         return 0;
 171 }
 172
 173 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 174 {
 175         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 176         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
 177         struct sk_buff *skb;
 178         u64 mapping[IPOIB_UD_RX_SG];
 179         union ib_gid *dgid;
 180         union ib_gid *sgid;
 181
 182         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
 183                        wr_id, wc->status);
 184
 185         if (unlikely(wr_id >= ipoib_recvq_size)) {
 186                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
 187                            wr_id, ipoib_recvq_size);
 188                 return;
 189         }
 190
 191         skb  = priv->rx_ring[wr_id].skb;
 192
 193         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 194                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 195                         ipoib_warn(priv,
 196                                    "failed recv event (status=%d, wrid=%d vend_err %#x)\n",
 197                                    wc->status, wr_id, wc->vendor_err);
 198                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
 199                 dev_kfree_skb_any(skb);
 200                 priv->rx_ring[wr_id].skb = NULL;
 201                 return;
 202         }
 203
 204         memcpy(mapping, priv->rx_ring[wr_id].mapping,
 205                IPOIB_UD_RX_SG * sizeof(*mapping));
 206
 207         /*
 208          * If we can't allocate a new RX buffer, dump
 209          * this packet and reuse the old buffer.
 210          */
 211         if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
 212                 ++dev->stats.rx_dropped;
 213                 goto repost;
 214         }
 215
 216         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 217                        wc->byte_len, wc->slid);
 218
 219         ipoib_ud_dma_unmap_rx(priv, mapping);
 220
 221         skb_put(skb, wc->byte_len);
 222
 223         /* First byte of dgid signals multicast when 0xff */
 224         dgid = &((struct ib_grh *)skb->data)->dgid;
 225
 226         if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
 227                 skb->pkt_type = PACKET_HOST;
 228         else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
 229                 skb->pkt_type = PACKET_BROADCAST;
 230         else
 231                 skb->pkt_type = PACKET_MULTICAST;
 232
 233         sgid = &((struct ib_grh *)skb->data)->sgid;
 234
 235         /*
 236          * Drop packets that this interface sent, ie multicast packets
 237          * that the HCA has replicated.
 238          */
 239         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num) {
 240                 int need_repost = 1;
 241
 242                 if ((wc->wc_flags & IB_WC_GRH) &&
 243                     sgid->global.interface_id != priv->local_gid.global.interface_id)
 244                         need_repost = 0;
 245
 246                 if (need_repost) {
 247                         dev_kfree_skb_any(skb);
 248                         goto repost;
 249                 }
 250         }
 251
 252         skb_pull(skb, IB_GRH_BYTES);
 253
 254         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 255         skb_add_pseudo_hdr(skb);
 256
 257         ++dev->stats.rx_packets;
 258         dev->stats.rx_bytes += skb->len;
 259         if (skb->pkt_type == PACKET_MULTICAST)
 260                 dev->stats.multicast++;
 261
 262         skb->dev = dev;
 263         if ((dev->features & NETIF_F_RXCSUM) &&
 264                         likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
 265                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 266
 267         napi_gro_receive(&priv->recv_napi, skb);
 268
 269 repost:
 270         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 271                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 272                            "for buf %d\n", wr_id);
 273 }
 274
 275 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
 276 {
 277         struct sk_buff *skb = tx_req->skb;
 278         u64 *mapping = tx_req->mapping;
 279         int i;
 280         int off;
 281
 282         if (skb_headlen(skb)) {
 283                 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
 284                                                DMA_TO_DEVICE);
 285                 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
 286                         return -EIO;
 287
 288                 off = 1;
 289         } else
 290                 off = 0;
 291
 292         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 293                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 294                 mapping[i + off] = ib_dma_map_page(ca,
 295                                                  skb_frag_page(frag),
 296                                                  frag->page_offset, skb_frag_size(frag),
 297                                                  DMA_TO_DEVICE);
 298                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
 299                         goto partial_error;
 300         }
 301         return 0;
 302
 303 partial_error:
 304         for (; i > 0; --i) {
 305                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 306
 307                 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
 308         }
 309
 310         if (off)
 311                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 312
 313         return -EIO;
 314 }
 315
 316 void ipoib_dma_unmap_tx(struct ipoib_dev_priv *priv,
 317                         struct ipoib_tx_buf *tx_req)
 318 {
 319         struct sk_buff *skb = tx_req->skb;
 320         u64 *mapping = tx_req->mapping;
 321         int i;
 322         int off;
 323
 324         if (skb_headlen(skb)) {
 325                 ib_dma_unmap_single(priv->ca, mapping[0], skb_headlen(skb),
 326                                     DMA_TO_DEVICE);
 327                 off = 1;
 328         } else
 329                 off = 0;
 330
 331         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 332                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 333
 334                 ib_dma_unmap_page(priv->ca, mapping[i + off],
 335                                   skb_frag_size(frag), DMA_TO_DEVICE);
 336         }
 337 }
 338
 339 /*
 340  * As the result of a completion error the QP Can be transferred to SQE states.
 341  * The function checks if the (send)QP is in SQE state and
 342  * moves it back to RTS state, that in order to have it functional again.
 343  */
 344 static void ipoib_qp_state_validate_work(struct work_struct *work)
 345 {
 346         struct ipoib_qp_state_validate *qp_work =
 347                 container_of(work, struct ipoib_qp_state_validate, work);
 348
 349         struct ipoib_dev_priv *priv = qp_work->priv;
 350         struct ib_qp_attr qp_attr;
 351         struct ib_qp_init_attr query_init_attr;
 352         int ret;
 353
 354         ret = ib_query_qp(priv->qp, &qp_attr, IB_QP_STATE, &query_init_attr);
 355         if (ret) {
 356                 ipoib_warn(priv, "%s: Failed to query QP ret: %d\n",
 357                            __func__, ret);
 358                 goto free_res;
 359         }
 360         pr_info("%s: QP: 0x%x is in state: %d\n",
 361                 __func__, priv->qp->qp_num, qp_attr.qp_state);
 362
 363         /* currently support only in SQE->RTS transition*/
 364         if (qp_attr.qp_state == IB_QPS_SQE) {
 365                 qp_attr.qp_state = IB_QPS_RTS;
 366
 367                 ret = ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE);
 368                 if (ret) {
 369                         pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n",
 370                                 ret, priv->qp->qp_num);
 371                         goto free_res;
 372                 }
 373                 pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n",
 374                         __func__, priv->qp->qp_num);
 375         } else {
 376                 pr_warn("QP (%d) will stay in state: %d\n",
 377                         priv->qp->qp_num, qp_attr.qp_state);
 378         }
 379
 380 free_res:
 381         kfree(qp_work);
 382 }
 383
 384 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 385 {
 386         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 387         unsigned int wr_id = wc->wr_id;
 388         struct ipoib_tx_buf *tx_req;
 389
 390         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 391                        wr_id, wc->status);
 392
 393         if (unlikely(wr_id >= ipoib_sendq_size)) {
 394                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 395                            wr_id, ipoib_sendq_size);
 396                 return;
 397         }
 398
 399         tx_req = &priv->tx_ring[wr_id];
 400
 401         ipoib_dma_unmap_tx(priv, tx_req);
 402
 403         ++dev->stats.tx_packets;
 404         dev->stats.tx_bytes += tx_req->skb->len;
 405
 406         dev_kfree_skb_any(tx_req->skb);
 407
 408         ++priv->tx_tail;
 409         ++priv->global_tx_tail;
 410
 411         if (unlikely(netif_queue_stopped(dev) &&
 412                      ((priv->global_tx_head - priv->global_tx_tail) <=
 413                       ipoib_sendq_size >> 1) &&
 414                      test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
 415                 netif_wake_queue(dev);
 416
 417         if (wc->status != IB_WC_SUCCESS &&
 418             wc->status != IB_WC_WR_FLUSH_ERR) {
 419                 struct ipoib_qp_state_validate *qp_work;
 420                 ipoib_warn(priv,
 421                            "failed send event (status=%d, wrid=%d vend_err %#x)\n",
 422                            wc->status, wr_id, wc->vendor_err);
 423                 qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC);
 424                 if (!qp_work)
 425                         return;
 426
 427                 INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work);
 428                 qp_work->priv = priv;
 429                 queue_work(priv->wq, &qp_work->work);
 430         }
 431 }
 432
 433 static int poll_tx(struct ipoib_dev_priv *priv)
 434 {
 435         int n, i;
 436         struct ib_wc *wc;
 437
 438         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
 439         for (i = 0; i < n; ++i) {
 440                 wc = priv->send_wc + i;
 441                 if (wc->wr_id & IPOIB_OP_CM)
 442                         ipoib_cm_handle_tx_wc(priv->dev, priv->send_wc + i);
 443                 else
 444                         ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
 445         }
 446         return n == MAX_SEND_CQE;
 447 }
 448
 449 int ipoib_rx_poll(struct napi_struct *napi, int budget)
 450 {
 451         struct ipoib_dev_priv *priv =
 452                 container_of(napi, struct ipoib_dev_priv, recv_napi);
 453         struct net_device *dev = priv->dev;
 454         int done;
 455         int t;
 456         int n, i;
 457
 458         done  = 0;
 459
 460 poll_more:
 461         while (done < budget) {
 462                 int max = (budget - done);
 463
 464                 t = min(IPOIB_NUM_WC, max);
 465                 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
 466
 467                 for (i = 0; i < n; i++) {
 468                         struct ib_wc *wc = priv->ibwc + i;
 469
 470                         if (wc->wr_id & IPOIB_OP_RECV) {
 471                                 ++done;
 472                                 if (wc->wr_id & IPOIB_OP_CM)
 473                                         ipoib_cm_handle_rx_wc(dev, wc);
 474                                 else
 475                                         ipoib_ib_handle_rx_wc(dev, wc);
 476                         } else {
 477                                 pr_warn("%s: Got unexpected wqe id\n", __func__);
 478                         }
 479                 }
 480
 481                 if (n != t)
 482                         break;
 483         }
 484
 485         if (done < budget) {
 486                 napi_complete(napi);
 487                 if (unlikely(ib_req_notify_cq(priv->recv_cq,
 488                                               IB_CQ_NEXT_COMP |
 489                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 490                     napi_reschedule(napi))
 491                         goto poll_more;
 492         }
 493
 494         return done;
 495 }
 496
 497 int ipoib_tx_poll(struct napi_struct *napi, int budget)
 498 {
 499         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv,
 500                                                    send_napi);
 501         struct net_device *dev = priv->dev;
 502         int n, i;
 503         struct ib_wc *wc;
 504
 505 poll_more:
 506         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
 507
 508         for (i = 0; i < n; i++) {
 509                 wc = priv->send_wc + i;
 510                 if (wc->wr_id & IPOIB_OP_CM)
 511                         ipoib_cm_handle_tx_wc(dev, wc);
 512                 else
 513                         ipoib_ib_handle_tx_wc(dev, wc);
 514         }
 515
 516         if (n < budget) {
 517                 napi_complete(napi);
 518                 if (unlikely(ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
 519                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 520                     napi_reschedule(napi))
 521                         goto poll_more;
 522         }
 523         return n < 0 ? 0 : n;
 524 }
 525
 526 void ipoib_ib_rx_completion(struct ib_cq *cq, void *ctx_ptr)
 527 {
 528         struct ipoib_dev_priv *priv = ctx_ptr;
 529
 530         napi_schedule(&priv->recv_napi);
 531 }
 532
 533 void ipoib_ib_tx_completion(struct ib_cq *cq, void *ctx_ptr)
 534 {
 535         struct ipoib_dev_priv *priv = ctx_ptr;
 536
 537         napi_schedule(&priv->send_napi);
 538 }
 539
 540 static inline int post_send(struct ipoib_dev_priv *priv,
 541                             unsigned int wr_id,
 542                             struct ib_ah *address, u32 dqpn,
 543                             struct ipoib_tx_buf *tx_req,
 544                             void *head, int hlen)
 545 {
 546         struct sk_buff *skb = tx_req->skb;
 547
 548         ipoib_build_sge(priv, tx_req);
 549
 550         priv->tx_wr.wr.wr_id    = wr_id;
 551         priv->tx_wr.remote_qpn  = dqpn;
 552         priv->tx_wr.ah          = address;
 553
 554         if (head) {
 555                 priv->tx_wr.mss         = skb_shinfo(skb)->gso_size;
 556                 priv->tx_wr.header      = head;
 557                 priv->tx_wr.hlen        = hlen;
 558                 priv->tx_wr.wr.opcode   = IB_WR_LSO;
 559         } else
 560                 priv->tx_wr.wr.opcode   = IB_WR_SEND;
 561
 562         return ib_post_send(priv->qp, &priv->tx_wr.wr, NULL);
 563 }
 564
 565 int ipoib_send(struct net_device *dev, struct sk_buff *skb,
 566                struct ib_ah *address, u32 dqpn)
 567 {
 568         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 569         struct ipoib_tx_buf *tx_req;
 570         int hlen, rc;
 571         void *phead;
 572         unsigned int usable_sge = priv->max_send_sge - !!skb_headlen(skb);
 573
 574         if (skb_is_gso(skb)) {
 575                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
 576                 phead = skb->data;
 577                 if (unlikely(!skb_pull(skb, hlen))) {
 578                         ipoib_warn(priv, "linear data too small\n");
 579                         ++dev->stats.tx_dropped;
 580                         ++dev->stats.tx_errors;
 581                         dev_kfree_skb_any(skb);
 582                         return -1;
 583                 }
 584         } else {
 585                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 586                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 587                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 588                         ++dev->stats.tx_dropped;
 589                         ++dev->stats.tx_errors;
 590                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 591                         return -1;
 592                 }
 593                 phead = NULL;
 594                 hlen  = 0;
 595         }
 596         if (skb_shinfo(skb)->nr_frags > usable_sge) {
 597                 if (skb_linearize(skb) < 0) {
 598                         ipoib_warn(priv, "skb could not be linearized\n");
 599                         ++dev->stats.tx_dropped;
 600                         ++dev->stats.tx_errors;
 601                         dev_kfree_skb_any(skb);
 602                         return -1;
 603                 }
 604                 /* Does skb_linearize return ok without reducing nr_frags? */
 605                 if (skb_shinfo(skb)->nr_frags > usable_sge) {
 606                         ipoib_warn(priv, "too many frags after skb linearize\n");
 607                         ++dev->stats.tx_dropped;
 608                         ++dev->stats.tx_errors;
 609                         dev_kfree_skb_any(skb);
 610                         return -1;
 611                 }
 612         }
 613
 614         ipoib_dbg_data(priv,
 615                        "sending packet, length=%d address=%p dqpn=0x%06x\n",
 616                        skb->len, address, dqpn);
 617
 618         /*
 619          * We put the skb into the tx_ring _before_ we call post_send()
 620          * because it's entirely possible that the completion handler will
 621          * run before we execute anything after the post_send().  That
 622          * means we have to make sure everything is properly recorded and
 623          * our state is consistent before we call post_send().
 624          */
 625         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 626         tx_req->skb = skb;
 627         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 628                 ++dev->stats.tx_errors;
 629                 dev_kfree_skb_any(skb);
 630                 return -1;
 631         }
 632
 633         if (skb->ip_summed == CHECKSUM_PARTIAL)
 634                 priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
 635         else
 636                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 637         /* increase the tx_head after send success, but use it for queue state */
 638         if ((priv->global_tx_head - priv->global_tx_tail) ==
 639             ipoib_sendq_size - 1) {
 640                 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 641                 netif_stop_queue(dev);
 642         }
 643
 644         skb_orphan(skb);
 645         skb_dst_drop(skb);
 646
 647         if (netif_queue_stopped(dev))
 648                 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
 649                                      IB_CQ_REPORT_MISSED_EVENTS) < 0)
 650                         ipoib_warn(priv, "request notify on send CQ failed\n");
 651
 652         rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 653                        address, dqpn, tx_req, phead, hlen);
 654         if (unlikely(rc)) {
 655                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 656                 ++dev->stats.tx_errors;
 657                 ipoib_dma_unmap_tx(priv, tx_req);
 658                 dev_kfree_skb_any(skb);
 659                 if (netif_queue_stopped(dev))
 660                         netif_wake_queue(dev);
 661                 rc = 0;
 662         } else {
 663                 netif_trans_update(dev);
 664
 665                 rc = priv->tx_head;
 666                 ++priv->tx_head;
 667                 ++priv->global_tx_head;
 668         }
 669         return rc;
 670 }
 671
 672 static void ipoib_reap_dead_ahs(struct ipoib_dev_priv *priv)
 673 {
 674         struct ipoib_ah *ah, *tah;
 675         LIST_HEAD(remove_list);
 676         unsigned long flags;
 677
 678         netif_tx_lock_bh(priv->dev);
 679         spin_lock_irqsave(&priv->lock, flags);
 680
 681         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 682                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 683                         list_del(&ah->list);
 684                         rdma_destroy_ah(ah->ah);
 685                         kfree(ah);
 686                 }
 687
 688         spin_unlock_irqrestore(&priv->lock, flags);
 689         netif_tx_unlock_bh(priv->dev);
 690 }
 691
 692 void ipoib_reap_ah(struct work_struct *work)
 693 {
 694         struct ipoib_dev_priv *priv =
 695                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 696
 697         ipoib_reap_dead_ahs(priv);
 698
 699         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 700                 queue_delayed_work(priv->wq, &priv->ah_reap_task,
 701                                    round_jiffies_relative(HZ));
 702 }
 703
 704 static void ipoib_start_ah_reaper(struct ipoib_dev_priv *priv)
 705 {
 706         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 707         queue_delayed_work(priv->wq, &priv->ah_reap_task,
 708                            round_jiffies_relative(HZ));
 709 }
 710
 711 static void ipoib_stop_ah_reaper(struct ipoib_dev_priv *priv)
 712 {
 713         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 714         cancel_delayed_work(&priv->ah_reap_task);
 715         /*
 716          * After ipoib_stop_ah_reaper() we always go through
 717          * ipoib_reap_dead_ahs() which ensures the work is really stopped and
 718          * does a final flush out of the dead_ah's list
 719          */
 720 }
 721
 722 static int recvs_pending(struct net_device *dev)
 723 {
 724         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 725         int pending = 0;
 726         int i;
 727
 728         for (i = 0; i < ipoib_recvq_size; ++i)
 729                 if (priv->rx_ring[i].skb)
 730                         ++pending;
 731
 732         return pending;
 733 }
 734
 735 static void check_qp_movement_and_print(struct ipoib_dev_priv *priv,
 736                                         struct ib_qp *qp,
 737                                         enum ib_qp_state new_state)
 738 {
 739         struct ib_qp_attr qp_attr;
 740         struct ib_qp_init_attr query_init_attr;
 741         int ret;
 742
 743         ret = ib_query_qp(qp, &qp_attr, IB_QP_STATE, &query_init_attr);
 744         if (ret) {
 745                 ipoib_warn(priv, "%s: Failed to query QP\n", __func__);
 746                 return;
 747         }
 748         /* print according to the new-state and the previous state.*/
 749         if (new_state == IB_QPS_ERR && qp_attr.qp_state == IB_QPS_RESET)
 750                 ipoib_dbg(priv, "Failed modify QP, IB_QPS_RESET to IB_QPS_ERR, acceptable\n");
 751         else
 752                 ipoib_warn(priv, "Failed to modify QP to state: %d from state: %d\n",
 753                            new_state, qp_attr.qp_state);
 754 }
 755
 756 static void ipoib_napi_enable(struct net_device *dev)
 757 {
 758         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 759
 760         napi_enable(&priv->recv_napi);
 761         napi_enable(&priv->send_napi);
 762 }
 763
 764 static void ipoib_napi_disable(struct net_device *dev)
 765 {
 766         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 767
 768         napi_disable(&priv->recv_napi);
 769         napi_disable(&priv->send_napi);
 770 }
 771
 772 int ipoib_ib_dev_stop_default(struct net_device *dev)
 773 {
 774         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 775         struct ib_qp_attr qp_attr;
 776         unsigned long begin;
 777         struct ipoib_tx_buf *tx_req;
 778         int i;
 779
 780         if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 781                 ipoib_napi_disable(dev);
 782
 783         ipoib_cm_dev_stop(dev);
 784
 785         /*
 786          * Move our QP to the error state and then reinitialize in
 787          * when all work requests have completed or have been flushed.
 788          */
 789         qp_attr.qp_state = IB_QPS_ERR;
 790         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 791                 check_qp_movement_and_print(priv, priv->qp, IB_QPS_ERR);
 792
 793         /* Wait for all sends and receives to complete */
 794         begin = jiffies;
 795
 796         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 797                 if (time_after(jiffies, begin + 5 * HZ)) {
 798                         ipoib_warn(priv,
 799                                    "timing out; %d sends %d receives not completed\n",
 800                                    priv->tx_head - priv->tx_tail,
 801                                    recvs_pending(dev));
 802
 803                         /*
 804                          * assume the HW is wedged and just free up
 805                          * all our pending work requests.
 806                          */
 807                         while ((int)priv->tx_tail - (int)priv->tx_head < 0) {
 808                                 tx_req = &priv->tx_ring[priv->tx_tail &
 809                                                         (ipoib_sendq_size - 1)];
 810                                 ipoib_dma_unmap_tx(priv, tx_req);
 811                                 dev_kfree_skb_any(tx_req->skb);
 812                                 ++priv->tx_tail;
 813                                 ++priv->global_tx_tail;
 814                         }
 815
 816                         for (i = 0; i < ipoib_recvq_size; ++i) {
 817                                 struct ipoib_rx_buf *rx_req;
 818
 819                                 rx_req = &priv->rx_ring[i];
 820                                 if (!rx_req->skb)
 821                                         continue;
 822                                 ipoib_ud_dma_unmap_rx(priv,
 823                                                       priv->rx_ring[i].mapping);
 824                                 dev_kfree_skb_any(rx_req->skb);
 825                                 rx_req->skb = NULL;
 826                         }
 827
 828                         goto timeout;
 829                 }
 830
 831                 ipoib_drain_cq(dev);
 832
 833                 usleep_range(1000, 2000);
 834         }
 835
 836         ipoib_dbg(priv, "All sends and receives done.\n");
 837
 838 timeout:
 839         qp_attr.qp_state = IB_QPS_RESET;
 840         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 841                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 842
 843         ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
 844
 845         return 0;
 846 }
 847
 848 int ipoib_ib_dev_open_default(struct net_device *dev)
 849 {
 850         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 851         int ret;
 852
 853         ret = ipoib_init_qp(dev);
 854         if (ret) {
 855                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 856                 return -1;
 857         }
 858
 859         ret = ipoib_ib_post_receives(dev);
 860         if (ret) {
 861                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 862                 goto out;
 863         }
 864
 865         ret = ipoib_cm_dev_open(dev);
 866         if (ret) {
 867                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 868                 goto out;
 869         }
 870
 871         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 872                 ipoib_napi_enable(dev);
 873
 874         return 0;
 875 out:
 876         return -1;
 877 }
 878
 879 int ipoib_ib_dev_open(struct net_device *dev)
 880 {
 881         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 882
 883         ipoib_pkey_dev_check_presence(dev);
 884
 885         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 886                 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
 887                            (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
 888                 return -1;
 889         }
 890
 891         ipoib_start_ah_reaper(priv);
 892         if (priv->rn_ops->ndo_open(dev)) {
 893                 pr_warn("%s: Failed to open dev\n", dev->name);
 894                 goto dev_stop;
 895         }
 896
 897         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 898
 899         return 0;
 900
 901 dev_stop:
 902         ipoib_stop_ah_reaper(priv);
 903         return -1;
 904 }
 905
 906 void ipoib_ib_dev_stop(struct net_device *dev)
 907 {
 908         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 909
 910         priv->rn_ops->ndo_stop(dev);
 911
 912         clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 913         ipoib_stop_ah_reaper(priv);
 914 }
 915
 916 void ipoib_pkey_dev_check_presence(struct net_device *dev)
 917 {
 918         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 919         struct rdma_netdev *rn = netdev_priv(dev);
 920
 921         if (!(priv->pkey & 0x7fff) ||
 922             ib_find_pkey(priv->ca, priv->port, priv->pkey,
 923                          &priv->pkey_index)) {
 924                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 925         } else {
 926                 if (rn->set_id)
 927                         rn->set_id(dev, priv->pkey_index);
 928                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 929         }
 930 }
 931
 932 void ipoib_ib_dev_up(struct net_device *dev)
 933 {
 934         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 935
 936         ipoib_pkey_dev_check_presence(dev);
 937
 938         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 939                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 940                 return;
 941         }
 942
 943         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 944
 945         ipoib_mcast_start_thread(dev);
 946 }
 947
 948 void ipoib_ib_dev_down(struct net_device *dev)
 949 {
 950         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 951
 952         ipoib_dbg(priv, "downing ib_dev\n");
 953
 954         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 955         netif_carrier_off(dev);
 956
 957         ipoib_mcast_stop_thread(dev);
 958         ipoib_mcast_dev_flush(dev);
 959
 960         ipoib_flush_paths(dev);
 961 }
 962
 963 void ipoib_drain_cq(struct net_device *dev)
 964 {
 965         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 966         int i, n;
 967
 968         /*
 969          * We call completion handling routines that expect to be
 970          * called from the BH-disabled NAPI poll context, so disable
 971          * BHs here too.
 972          */
 973         local_bh_disable();
 974
 975         do {
 976                 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
 977                 for (i = 0; i < n; ++i) {
 978                         /*
 979                          * Convert any successful completions to flush
 980                          * errors to avoid passing packets up the
 981                          * stack after bringing the device down.
 982                          */
 983                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 984                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 985
 986                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 987                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 988                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 989                                 else
 990                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 991                         } else {
 992                                 pr_warn("%s: Got unexpected wqe id\n", __func__);
 993                         }
 994                 }
 995         } while (n == IPOIB_NUM_WC);
 996
 997         while (poll_tx(priv))
 998                 ; /* nothing */
 999
1000         local_bh_enable();
1001 }
1002
1003 /*
1004  * Takes whatever value which is in pkey index 0 and updates priv->pkey
1005  * returns 0 if the pkey value was changed.
1006  */
1007 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
1008 {
1009         int result;
1010         u16 prev_pkey;
1011
1012         prev_pkey = priv->pkey;
1013         result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1014         if (result) {
1015                 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
1016                            priv->port, result);
1017                 return result;
1018         }
1019
1020         priv->pkey |= 0x8000;
1021
1022         if (prev_pkey != priv->pkey) {
1023                 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
1024                           prev_pkey, priv->pkey);
1025                 /*
1026                  * Update the pkey in the broadcast address, while making sure to set
1027                  * the full membership bit, so that we join the right broadcast group.
1028                  */
1029                 priv->dev->broadcast[8] = priv->pkey >> 8;
1030                 priv->dev->broadcast[9] = priv->pkey & 0xff;
1031                 return 0;
1032         }
1033
1034         return 1;
1035 }
1036 /*
1037  * returns 0 if pkey value was found in a different slot.
1038  */
1039 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
1040 {
1041         u16 old_index = priv->pkey_index;
1042
1043         priv->pkey_index = 0;
1044         ipoib_pkey_dev_check_presence(priv->dev);
1045
1046         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
1047             (old_index == priv->pkey_index))
1048                 return 1;
1049         return 0;
1050 }
1051
1052 /*
1053  * returns true if the device address of the ipoib interface has changed and the
1054  * new address is a valid one (i.e in the gid table), return false otherwise.
1055  */
1056 static bool ipoib_dev_addr_changed_valid(struct ipoib_dev_priv *priv)
1057 {
1058         union ib_gid search_gid;
1059         union ib_gid gid0;
1060         union ib_gid *netdev_gid;
1061         int err;
1062         u16 index;
1063         u8 port;
1064         bool ret = false;
1065
1066         netdev_gid = (union ib_gid *)(priv->dev->dev_addr + 4);
1067         if (rdma_query_gid(priv->ca, priv->port, 0, &gid0))
1068                 return false;
1069
1070         netif_addr_lock_bh(priv->dev);
1071
1072         /* The subnet prefix may have changed, update it now so we won't have
1073          * to do it later
1074          */
1075         priv->local_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1076         netdev_gid->global.subnet_prefix = gid0.global.subnet_prefix;
1077         search_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1078
1079         search_gid.global.interface_id = priv->local_gid.global.interface_id;
1080
1081         netif_addr_unlock_bh(priv->dev);
1082
1083         err = ib_find_gid(priv->ca, &search_gid, &port, &index);
1084
1085         netif_addr_lock_bh(priv->dev);
1086
1087         if (search_gid.global.interface_id !=
1088             priv->local_gid.global.interface_id)
1089                 /* There was a change while we were looking up the gid, bail
1090                  * here and let the next work sort this out
1091                  */
1092                 goto out;
1093
1094         /* The next section of code needs some background:
1095          * Per IB spec the port GUID can't change if the HCA is powered on.
1096          * port GUID is the basis for GID at index 0 which is the basis for
1097          * the default device address of a ipoib interface.
1098          *
1099          * so it seems the flow should be:
1100          * if user_changed_dev_addr && gid in gid tbl
1101          *      set bit dev_addr_set
1102          *      return true
1103          * else
1104          *      return false
1105          *
1106          * The issue is that there are devices that don't follow the spec,
1107          * they change the port GUID when the HCA is powered, so in order
1108          * not to break userspace applications, We need to check if the
1109          * user wanted to control the device address and we assume that
1110          * if he sets the device address back to be based on GID index 0,
1111          * he no longer wishs to control it.
1112          *
1113          * If the user doesn't control the the device address,
1114          * IPOIB_FLAG_DEV_ADDR_SET is set and ib_find_gid failed it means
1115          * the port GUID has changed and GID at index 0 has changed
1116          * so we need to change priv->local_gid and priv->dev->dev_addr
1117          * to reflect the new GID.
1118          */
1119         if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
1120                 if (!err && port == priv->port) {
1121                         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1122                         if (index == 0)
1123                                 clear_bit(IPOIB_FLAG_DEV_ADDR_CTRL,
1124                                           &priv->flags);
1125                         else
1126                                 set_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags);
1127                         ret = true;
1128                 } else {
1129                         ret = false;
1130                 }
1131         } else {
1132                 if (!err && port == priv->port) {
1133                         ret = true;
1134                 } else {
1135                         if (!test_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags)) {
1136                                 memcpy(&priv->local_gid, &gid0,
1137                                        sizeof(priv->local_gid));
1138                                 memcpy(priv->dev->dev_addr + 4, &gid0,
1139                                        sizeof(priv->local_gid));
1140                                 ret = true;
1141                         }
1142                 }
1143         }
1144
1145 out:
1146         netif_addr_unlock_bh(priv->dev);
1147
1148         return ret;
1149 }
1150
1151 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
1152                                 enum ipoib_flush_level level,
1153                                 int nesting)
1154 {
1155         struct ipoib_dev_priv *cpriv;
1156         struct net_device *dev = priv->dev;
1157         int result;
1158
1159         down_read_nested(&priv->vlan_rwsem, nesting);
1160
1161         /*
1162          * Flush any child interfaces too -- they might be up even if
1163          * the parent is down.
1164          */
1165         list_for_each_entry(cpriv, &priv->child_intfs, list)
1166                 __ipoib_ib_dev_flush(cpriv, level, nesting + 1);
1167
1168         up_read(&priv->vlan_rwsem);
1169
1170         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
1171             level != IPOIB_FLUSH_HEAVY) {
1172                 /* Make sure the dev_addr is set even if not flushing */
1173                 if (level == IPOIB_FLUSH_LIGHT)
1174                         ipoib_dev_addr_changed_valid(priv);
1175                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
1176                 return;
1177         }
1178
1179         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1180                 /* interface is down. update pkey and leave. */
1181                 if (level == IPOIB_FLUSH_HEAVY) {
1182                         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1183                                 update_parent_pkey(priv);
1184                         else
1185                                 update_child_pkey(priv);
1186                 } else if (level == IPOIB_FLUSH_LIGHT)
1187                         ipoib_dev_addr_changed_valid(priv);
1188                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1189                 return;
1190         }
1191
1192         if (level == IPOIB_FLUSH_HEAVY) {
1193                 /* child devices chase their origin pkey value, while non-child
1194                  * (parent) devices should always takes what present in pkey index 0
1195                  */
1196                 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1197                         result = update_child_pkey(priv);
1198                         if (result) {
1199                                 /* restart QP only if P_Key index is changed */
1200                                 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
1201                                 return;
1202                         }
1203
1204                 } else {
1205                         result = update_parent_pkey(priv);
1206                         /* restart QP only if P_Key value changed */
1207                         if (result) {
1208                                 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
1209                                 return;
1210                         }
1211                 }
1212         }
1213
1214         if (level == IPOIB_FLUSH_LIGHT) {
1215                 int oper_up;
1216                 ipoib_mark_paths_invalid(dev);
1217                 /* Set IPoIB operation as down to prevent races between:
1218                  * the flush flow which leaves MCG and on the fly joins
1219                  * which can happen during that time. mcast restart task
1220                  * should deal with join requests we missed.
1221                  */
1222                 oper_up = test_and_clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1223                 ipoib_mcast_dev_flush(dev);
1224                 if (oper_up)
1225                         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1226                 ipoib_reap_dead_ahs(priv);
1227         }
1228
1229         if (level >= IPOIB_FLUSH_NORMAL)
1230                 ipoib_ib_dev_down(dev);
1231
1232         if (level == IPOIB_FLUSH_HEAVY) {
1233                 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
1234                         ipoib_ib_dev_stop(dev);
1235
1236                 if (ipoib_ib_dev_open(dev))
1237                         return;
1238
1239                 if (netif_queue_stopped(dev))
1240                         netif_start_queue(dev);
1241         }
1242
1243         /*
1244          * The device could have been brought down between the start and when
1245          * we get here, don't bring it back up if it's not configured up
1246          */
1247         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1248                 if (level >= IPOIB_FLUSH_NORMAL)
1249                         ipoib_ib_dev_up(dev);
1250                 if (ipoib_dev_addr_changed_valid(priv))
1251                         ipoib_mcast_restart_task(&priv->restart_task);
1252         }
1253 }
1254
1255 void ipoib_ib_dev_flush_light(struct work_struct *work)
1256 {
1257         struct ipoib_dev_priv *priv =
1258                 container_of(work, struct ipoib_dev_priv, flush_light);
1259
1260         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0);
1261 }
1262
1263 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1264 {
1265         struct ipoib_dev_priv *priv =
1266                 container_of(work, struct ipoib_dev_priv, flush_normal);
1267
1268         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0);
1269 }
1270
1271 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1272 {
1273         struct ipoib_dev_priv *priv =
1274                 container_of(work, struct ipoib_dev_priv, flush_heavy);
1275
1276         rtnl_lock();
1277         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0);
1278         rtnl_unlock();
1279 }
1280
1281 void ipoib_ib_dev_cleanup(struct net_device *dev)
1282 {
1283         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1284
1285         ipoib_dbg(priv, "cleaning up ib_dev\n");
1286         /*
1287          * We must make sure there are no more (path) completions
1288          * that may wish to touch priv fields that are no longer valid
1289          */
1290         ipoib_flush_paths(dev);
1291
1292         ipoib_mcast_stop_thread(dev);
1293         ipoib_mcast_dev_flush(dev);
1294
1295         /*
1296          * All of our ah references aren't free until after
1297          * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1298          * the neighbor garbage collection is stopped and reaped.
1299          * That should all be done now, so make a final ah flush.
1300          */
1301         ipoib_reap_dead_ahs(priv);
1302
1303         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
1304
1305         priv->rn_ops->ndo_uninit(dev);
1306
1307         if (priv->pd) {
1308                 ib_dealloc_pd(priv->pd);
1309                 priv->pd = NULL;
1310         }
1311 }
1312