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
  44 #include "ipoib.h"
  45
  46 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  47 static int data_debug_level;
  48
  49 module_param(data_debug_level, int, 0644);
  50 MODULE_PARM_DESC(data_debug_level,
  51                  "Enable data path debug tracing if > 0");
  52 #endif
  53
  54 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
  55                                  struct ib_pd *pd, struct rdma_ah_attr *attr)
  56 {
  57         struct ipoib_ah *ah;
  58         struct ib_ah *vah;
  59
  60         ah = kmalloc(sizeof *ah, GFP_KERNEL);
  61         if (!ah)
  62                 return ERR_PTR(-ENOMEM);
  63
  64         ah->dev       = dev;
  65         ah->last_send = 0;
  66         kref_init(&ah->ref);
  67
  68         vah = rdma_create_ah(pd, attr);
  69         if (IS_ERR(vah)) {
  70                 kfree(ah);
  71                 ah = (struct ipoib_ah *)vah;
  72         } else {
  73                 ah->ah = vah;
  74                 ipoib_dbg(ipoib_priv(dev), "Created ah %p\n", ah->ah);
  75         }
  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 = ipoib_priv(ah->dev);
  84
  85         unsigned long flags;
  86
  87         spin_lock_irqsave(&priv->lock, flags);
  88         list_add_tail(&ah->list, &priv->dead_ahs);
  89         spin_unlock_irqrestore(&priv->lock, flags);
  90 }
  91
  92 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
  93                                   u64 mapping[IPOIB_UD_RX_SG])
  94 {
  95         ib_dma_unmap_single(priv->ca, mapping[0],
  96                             IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
  97                             DMA_FROM_DEVICE);
  98 }
  99
 100 static int ipoib_ib_post_receive(struct net_device *dev, int id)
 101 {
 102         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 103         struct ib_recv_wr *bad_wr;
 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, &bad_wr);
 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
 410         if (unlikely(netif_queue_stopped(dev) &&
 411                      ((priv->tx_head - priv->tx_tail) <= ipoib_sendq_size >> 1) &&
 412                      test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
 413                 netif_wake_queue(dev);
 414
 415         if (wc->status != IB_WC_SUCCESS &&
 416             wc->status != IB_WC_WR_FLUSH_ERR) {
 417                 struct ipoib_qp_state_validate *qp_work;
 418                 ipoib_warn(priv,
 419                            "failed send event (status=%d, wrid=%d vend_err %#x)\n",
 420                            wc->status, wr_id, wc->vendor_err);
 421                 qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC);
 422                 if (!qp_work)
 423                         return;
 424
 425                 INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work);
 426                 qp_work->priv = priv;
 427                 queue_work(priv->wq, &qp_work->work);
 428         }
 429 }
 430
 431 static int poll_tx(struct ipoib_dev_priv *priv)
 432 {
 433         int n, i;
 434         struct ib_wc *wc;
 435
 436         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
 437         for (i = 0; i < n; ++i) {
 438                 wc = priv->send_wc + i;
 439                 if (wc->wr_id & IPOIB_OP_CM)
 440                         ipoib_cm_handle_tx_wc(priv->dev, priv->send_wc + i);
 441                 else
 442                         ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
 443         }
 444         return n == MAX_SEND_CQE;
 445 }
 446
 447 int ipoib_rx_poll(struct napi_struct *napi, int budget)
 448 {
 449         struct ipoib_dev_priv *priv =
 450                 container_of(napi, struct ipoib_dev_priv, recv_napi);
 451         struct net_device *dev = priv->dev;
 452         int done;
 453         int t;
 454         int n, i;
 455
 456         done  = 0;
 457
 458 poll_more:
 459         while (done < budget) {
 460                 int max = (budget - done);
 461
 462                 t = min(IPOIB_NUM_WC, max);
 463                 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
 464
 465                 for (i = 0; i < n; i++) {
 466                         struct ib_wc *wc = priv->ibwc + i;
 467
 468                         if (wc->wr_id & IPOIB_OP_RECV) {
 469                                 ++done;
 470                                 if (wc->wr_id & IPOIB_OP_CM)
 471                                         ipoib_cm_handle_rx_wc(dev, wc);
 472                                 else
 473                                         ipoib_ib_handle_rx_wc(dev, wc);
 474                         } else {
 475                                 pr_warn("%s: Got unexpected wqe id\n", __func__);
 476                         }
 477                 }
 478
 479                 if (n != t)
 480                         break;
 481         }
 482
 483         if (done < budget) {
 484                 napi_complete(napi);
 485                 if (unlikely(ib_req_notify_cq(priv->recv_cq,
 486                                               IB_CQ_NEXT_COMP |
 487                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 488                     napi_reschedule(napi))
 489                         goto poll_more;
 490         }
 491
 492         return done;
 493 }
 494
 495 int ipoib_tx_poll(struct napi_struct *napi, int budget)
 496 {
 497         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv,
 498                                                    send_napi);
 499         struct net_device *dev = priv->dev;
 500         int n, i;
 501         struct ib_wc *wc;
 502
 503 poll_more:
 504         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
 505
 506         for (i = 0; i < n; i++) {
 507                 wc = priv->send_wc + i;
 508                 if (wc->wr_id & IPOIB_OP_CM)
 509                         ipoib_cm_handle_tx_wc(dev, wc);
 510                 else
 511                         ipoib_ib_handle_tx_wc(dev, wc);
 512         }
 513
 514         if (n < budget) {
 515                 napi_complete(napi);
 516                 if (unlikely(ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
 517                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 518                     napi_reschedule(napi))
 519                         goto poll_more;
 520         }
 521         return n < 0 ? 0 : n;
 522 }
 523
 524 void ipoib_ib_rx_completion(struct ib_cq *cq, void *ctx_ptr)
 525 {
 526         struct ipoib_dev_priv *priv = ctx_ptr;
 527
 528         napi_schedule(&priv->recv_napi);
 529 }
 530
 531 void ipoib_ib_tx_completion(struct ib_cq *cq, void *ctx_ptr)
 532 {
 533         struct ipoib_dev_priv *priv = ctx_ptr;
 534
 535         napi_schedule(&priv->send_napi);
 536 }
 537
 538 static inline int post_send(struct ipoib_dev_priv *priv,
 539                             unsigned int wr_id,
 540                             struct ib_ah *address, u32 dqpn,
 541                             struct ipoib_tx_buf *tx_req,
 542                             void *head, int hlen)
 543 {
 544         struct ib_send_wr *bad_wr;
 545         struct sk_buff *skb = tx_req->skb;
 546
 547         ipoib_build_sge(priv, tx_req);
 548
 549         priv->tx_wr.wr.wr_id    = wr_id;
 550         priv->tx_wr.remote_qpn  = dqpn;
 551         priv->tx_wr.ah          = address;
 552
 553         if (head) {
 554                 priv->tx_wr.mss         = skb_shinfo(skb)->gso_size;
 555                 priv->tx_wr.header      = head;
 556                 priv->tx_wr.hlen        = hlen;
 557                 priv->tx_wr.wr.opcode   = IB_WR_LSO;
 558         } else
 559                 priv->tx_wr.wr.opcode   = IB_WR_SEND;
 560
 561         return ib_post_send(priv->qp, &priv->tx_wr.wr, &bad_wr);
 562 }
 563
 564 int ipoib_send(struct net_device *dev, struct sk_buff *skb,
 565                struct ib_ah *address, u32 dqpn)
 566 {
 567         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 568         struct ipoib_tx_buf *tx_req;
 569         int hlen, rc;
 570         void *phead;
 571         unsigned usable_sge = priv->max_send_sge - !!skb_headlen(skb);
 572
 573         if (skb_is_gso(skb)) {
 574                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
 575                 phead = skb->data;
 576                 if (unlikely(!skb_pull(skb, hlen))) {
 577                         ipoib_warn(priv, "linear data too small\n");
 578                         ++dev->stats.tx_dropped;
 579                         ++dev->stats.tx_errors;
 580                         dev_kfree_skb_any(skb);
 581                         return -1;
 582                 }
 583         } else {
 584                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 585                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 586                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 587                         ++dev->stats.tx_dropped;
 588                         ++dev->stats.tx_errors;
 589                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 590                         return -1;
 591                 }
 592                 phead = NULL;
 593                 hlen  = 0;
 594         }
 595         if (skb_shinfo(skb)->nr_frags > usable_sge) {
 596                 if (skb_linearize(skb) < 0) {
 597                         ipoib_warn(priv, "skb could not be linearized\n");
 598                         ++dev->stats.tx_dropped;
 599                         ++dev->stats.tx_errors;
 600                         dev_kfree_skb_any(skb);
 601                         return -1;
 602                 }
 603                 /* Does skb_linearize return ok without reducing nr_frags? */
 604                 if (skb_shinfo(skb)->nr_frags > usable_sge) {
 605                         ipoib_warn(priv, "too many frags after skb linearize\n");
 606                         ++dev->stats.tx_dropped;
 607                         ++dev->stats.tx_errors;
 608                         dev_kfree_skb_any(skb);
 609                         return -1;
 610                 }
 611         }
 612
 613         ipoib_dbg_data(priv,
 614                        "sending packet, length=%d address=%p dqpn=0x%06x\n",
 615                        skb->len, address, dqpn);
 616
 617         /*
 618          * We put the skb into the tx_ring _before_ we call post_send()
 619          * because it's entirely possible that the completion handler will
 620          * run before we execute anything after the post_send().  That
 621          * means we have to make sure everything is properly recorded and
 622          * our state is consistent before we call post_send().
 623          */
 624         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 625         tx_req->skb = skb;
 626         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 627                 ++dev->stats.tx_errors;
 628                 dev_kfree_skb_any(skb);
 629                 return -1;
 630         }
 631
 632         if (skb->ip_summed == CHECKSUM_PARTIAL)
 633                 priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
 634         else
 635                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 636         /* increase the tx_head after send success, but use it for queue state */
 637         if (priv->tx_head - priv->tx_tail == ipoib_sendq_size - 1) {
 638                 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 639                 netif_stop_queue(dev);
 640         }
 641
 642         skb_orphan(skb);
 643         skb_dst_drop(skb);
 644
 645         if (netif_queue_stopped(dev))
 646                 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
 647                                      IB_CQ_REPORT_MISSED_EVENTS))
 648                         ipoib_warn(priv, "request notify on send CQ failed\n");
 649
 650         rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 651                        address, dqpn, tx_req, phead, hlen);
 652         if (unlikely(rc)) {
 653                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 654                 ++dev->stats.tx_errors;
 655                 ipoib_dma_unmap_tx(priv, tx_req);
 656                 dev_kfree_skb_any(skb);
 657                 if (netif_queue_stopped(dev))
 658                         netif_wake_queue(dev);
 659                 rc = 0;
 660         } else {
 661                 netif_trans_update(dev);
 662
 663                 rc = priv->tx_head;
 664                 ++priv->tx_head;
 665         }
 666         return rc;
 667 }
 668
 669 static void __ipoib_reap_ah(struct net_device *dev)
 670 {
 671         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 672         struct ipoib_ah *ah, *tah;
 673         LIST_HEAD(remove_list);
 674         unsigned long flags;
 675
 676         netif_tx_lock_bh(dev);
 677         spin_lock_irqsave(&priv->lock, flags);
 678
 679         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 680                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 681                         list_del(&ah->list);
 682                         rdma_destroy_ah(ah->ah);
 683                         kfree(ah);
 684                 }
 685
 686         spin_unlock_irqrestore(&priv->lock, flags);
 687         netif_tx_unlock_bh(dev);
 688 }
 689
 690 void ipoib_reap_ah(struct work_struct *work)
 691 {
 692         struct ipoib_dev_priv *priv =
 693                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 694         struct net_device *dev = priv->dev;
 695
 696         __ipoib_reap_ah(dev);
 697
 698         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 699                 queue_delayed_work(priv->wq, &priv->ah_reap_task,
 700                                    round_jiffies_relative(HZ));
 701 }
 702
 703 static void ipoib_flush_ah(struct net_device *dev)
 704 {
 705         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 706
 707         cancel_delayed_work(&priv->ah_reap_task);
 708         flush_workqueue(priv->wq);
 709         ipoib_reap_ah(&priv->ah_reap_task.work);
 710 }
 711
 712 static void ipoib_stop_ah(struct net_device *dev)
 713 {
 714         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 715
 716         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 717         ipoib_flush_ah(dev);
 718 }
 719
 720 static int recvs_pending(struct net_device *dev)
 721 {
 722         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 723         int pending = 0;
 724         int i;
 725
 726         for (i = 0; i < ipoib_recvq_size; ++i)
 727                 if (priv->rx_ring[i].skb)
 728                         ++pending;
 729
 730         return pending;
 731 }
 732
 733 static void check_qp_movement_and_print(struct ipoib_dev_priv *priv,
 734                                         struct ib_qp *qp,
 735                                         enum ib_qp_state new_state)
 736 {
 737         struct ib_qp_attr qp_attr;
 738         struct ib_qp_init_attr query_init_attr;
 739         int ret;
 740
 741         ret = ib_query_qp(qp, &qp_attr, IB_QP_STATE, &query_init_attr);
 742         if (ret) {
 743                 ipoib_warn(priv, "%s: Failed to query QP\n", __func__);
 744                 return;
 745         }
 746         /* print according to the new-state and the previous state.*/
 747         if (new_state == IB_QPS_ERR && qp_attr.qp_state == IB_QPS_RESET)
 748                 ipoib_dbg(priv, "Failed modify QP, IB_QPS_RESET to IB_QPS_ERR, acceptable\n");
 749         else
 750                 ipoib_warn(priv, "Failed to modify QP to state: %d from state: %d\n",
 751                            new_state, qp_attr.qp_state);
 752 }
 753
 754 static void ipoib_napi_enable(struct net_device *dev)
 755 {
 756         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 757
 758         napi_enable(&priv->recv_napi);
 759         napi_enable(&priv->send_napi);
 760 }
 761
 762 static void ipoib_napi_disable(struct net_device *dev)
 763 {
 764         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 765
 766         napi_disable(&priv->recv_napi);
 767         napi_disable(&priv->send_napi);
 768 }
 769
 770 int ipoib_ib_dev_stop_default(struct net_device *dev)
 771 {
 772         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 773         struct ib_qp_attr qp_attr;
 774         unsigned long begin;
 775         struct ipoib_tx_buf *tx_req;
 776         int i;
 777
 778         if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 779                 ipoib_napi_disable(dev);
 780
 781         ipoib_cm_dev_stop(dev);
 782
 783         /*
 784          * Move our QP to the error state and then reinitialize in
 785          * when all work requests have completed or have been flushed.
 786          */
 787         qp_attr.qp_state = IB_QPS_ERR;
 788         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 789                 check_qp_movement_and_print(priv, priv->qp, IB_QPS_ERR);
 790
 791         /* Wait for all sends and receives to complete */
 792         begin = jiffies;
 793
 794         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 795                 if (time_after(jiffies, begin + 5 * HZ)) {
 796                         ipoib_warn(priv,
 797                                    "timing out; %d sends %d receives not completed\n",
 798                                    priv->tx_head - priv->tx_tail,
 799                                    recvs_pending(dev));
 800
 801                         /*
 802                          * assume the HW is wedged and just free up
 803                          * all our pending work requests.
 804                          */
 805                         while ((int)priv->tx_tail - (int)priv->tx_head < 0) {
 806                                 tx_req = &priv->tx_ring[priv->tx_tail &
 807                                                         (ipoib_sendq_size - 1)];
 808                                 ipoib_dma_unmap_tx(priv, tx_req);
 809                                 dev_kfree_skb_any(tx_req->skb);
 810                                 ++priv->tx_tail;
 811                         }
 812
 813                         for (i = 0; i < ipoib_recvq_size; ++i) {
 814                                 struct ipoib_rx_buf *rx_req;
 815
 816                                 rx_req = &priv->rx_ring[i];
 817                                 if (!rx_req->skb)
 818                                         continue;
 819                                 ipoib_ud_dma_unmap_rx(priv,
 820                                                       priv->rx_ring[i].mapping);
 821                                 dev_kfree_skb_any(rx_req->skb);
 822                                 rx_req->skb = NULL;
 823                         }
 824
 825                         goto timeout;
 826                 }
 827
 828                 ipoib_drain_cq(dev);
 829
 830                 usleep_range(1000, 2000);
 831         }
 832
 833         ipoib_dbg(priv, "All sends and receives done.\n");
 834
 835 timeout:
 836         qp_attr.qp_state = IB_QPS_RESET;
 837         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 838                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 839
 840         ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
 841
 842         return 0;
 843 }
 844
 845 int ipoib_ib_dev_stop(struct net_device *dev)
 846 {
 847         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 848
 849         priv->rn_ops->ndo_stop(dev);
 850
 851         clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 852         ipoib_flush_ah(dev);
 853
 854         return 0;
 855 }
 856
 857 int ipoib_ib_dev_open_default(struct net_device *dev)
 858 {
 859         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 860         int ret;
 861
 862         ret = ipoib_init_qp(dev);
 863         if (ret) {
 864                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 865                 return -1;
 866         }
 867
 868         ret = ipoib_ib_post_receives(dev);
 869         if (ret) {
 870                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 871                 goto out;
 872         }
 873
 874         ret = ipoib_cm_dev_open(dev);
 875         if (ret) {
 876                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 877                 goto out;
 878         }
 879
 880         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 881                 ipoib_napi_enable(dev);
 882
 883         return 0;
 884 out:
 885         return -1;
 886 }
 887
 888 int ipoib_ib_dev_open(struct net_device *dev)
 889 {
 890         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 891
 892         ipoib_pkey_dev_check_presence(dev);
 893
 894         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 895                 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
 896                            (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
 897                 return -1;
 898         }
 899
 900         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 901         queue_delayed_work(priv->wq, &priv->ah_reap_task,
 902                            round_jiffies_relative(HZ));
 903
 904         if (priv->rn_ops->ndo_open(dev)) {
 905                 pr_warn("%s: Failed to open dev\n", dev->name);
 906                 goto dev_stop;
 907         }
 908
 909         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 910
 911         return 0;
 912
 913 dev_stop:
 914         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 915         cancel_delayed_work(&priv->ah_reap_task);
 916         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 917         ipoib_ib_dev_stop(dev);
 918         return -1;
 919 }
 920
 921 void ipoib_pkey_dev_check_presence(struct net_device *dev)
 922 {
 923         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 924         struct rdma_netdev *rn = netdev_priv(dev);
 925
 926         if (!(priv->pkey & 0x7fff) ||
 927             ib_find_pkey(priv->ca, priv->port, priv->pkey,
 928                          &priv->pkey_index)) {
 929                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 930         } else {
 931                 if (rn->set_id)
 932                         rn->set_id(dev, priv->pkey_index);
 933                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 934         }
 935 }
 936
 937 void ipoib_ib_dev_up(struct net_device *dev)
 938 {
 939         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 940
 941         ipoib_pkey_dev_check_presence(dev);
 942
 943         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 944                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 945                 return;
 946         }
 947
 948         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 949
 950         ipoib_mcast_start_thread(dev);
 951 }
 952
 953 void ipoib_ib_dev_down(struct net_device *dev)
 954 {
 955         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 956
 957         ipoib_dbg(priv, "downing ib_dev\n");
 958
 959         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 960         netif_carrier_off(dev);
 961
 962         ipoib_mcast_stop_thread(dev);
 963         ipoib_mcast_dev_flush(dev);
 964
 965         ipoib_flush_paths(dev);
 966 }
 967
 968 void ipoib_drain_cq(struct net_device *dev)
 969 {
 970         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 971         int i, n;
 972
 973         /*
 974          * We call completion handling routines that expect to be
 975          * called from the BH-disabled NAPI poll context, so disable
 976          * BHs here too.
 977          */
 978         local_bh_disable();
 979
 980         do {
 981                 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
 982                 for (i = 0; i < n; ++i) {
 983                         /*
 984                          * Convert any successful completions to flush
 985                          * errors to avoid passing packets up the
 986                          * stack after bringing the device down.
 987                          */
 988                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 989                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 990
 991                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 992                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 993                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 994                                 else
 995                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 996                         } else {
 997                                 pr_warn("%s: Got unexpected wqe id\n", __func__);
 998                         }
 999                 }
1000         } while (n == IPOIB_NUM_WC);
1001
1002         while (poll_tx(priv))
1003                 ; /* nothing */
1004
1005         local_bh_enable();
1006 }
1007
1008 /*
1009  * Takes whatever value which is in pkey index 0 and updates priv->pkey
1010  * returns 0 if the pkey value was changed.
1011  */
1012 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
1013 {
1014         int result;
1015         u16 prev_pkey;
1016
1017         prev_pkey = priv->pkey;
1018         result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1019         if (result) {
1020                 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
1021                            priv->port, result);
1022                 return result;
1023         }
1024
1025         priv->pkey |= 0x8000;
1026
1027         if (prev_pkey != priv->pkey) {
1028                 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
1029                           prev_pkey, priv->pkey);
1030                 /*
1031                  * Update the pkey in the broadcast address, while making sure to set
1032                  * the full membership bit, so that we join the right broadcast group.
1033                  */
1034                 priv->dev->broadcast[8] = priv->pkey >> 8;
1035                 priv->dev->broadcast[9] = priv->pkey & 0xff;
1036                 return 0;
1037         }
1038
1039         return 1;
1040 }
1041 /*
1042  * returns 0 if pkey value was found in a different slot.
1043  */
1044 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
1045 {
1046         u16 old_index = priv->pkey_index;
1047
1048         priv->pkey_index = 0;
1049         ipoib_pkey_dev_check_presence(priv->dev);
1050
1051         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
1052             (old_index == priv->pkey_index))
1053                 return 1;
1054         return 0;
1055 }
1056
1057 /*
1058  * returns true if the device address of the ipoib interface has changed and the
1059  * new address is a valid one (i.e in the gid table), return false otherwise.
1060  */
1061 static bool ipoib_dev_addr_changed_valid(struct ipoib_dev_priv *priv)
1062 {
1063         union ib_gid search_gid;
1064         union ib_gid gid0;
1065         union ib_gid *netdev_gid;
1066         int err;
1067         u16 index;
1068         u8 port;
1069         bool ret = false;
1070
1071         netdev_gid = (union ib_gid *)(priv->dev->dev_addr + 4);
1072         if (ib_query_gid(priv->ca, priv->port, 0, &gid0, NULL))
1073                 return false;
1074
1075         netif_addr_lock_bh(priv->dev);
1076
1077         /* The subnet prefix may have changed, update it now so we won't have
1078          * to do it later
1079          */
1080         priv->local_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1081         netdev_gid->global.subnet_prefix = gid0.global.subnet_prefix;
1082         search_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1083
1084         search_gid.global.interface_id = priv->local_gid.global.interface_id;
1085
1086         netif_addr_unlock_bh(priv->dev);
1087
1088         err = ib_find_gid(priv->ca, &search_gid, IB_GID_TYPE_IB,
1089                           priv->dev, &port, &index);
1090
1091         netif_addr_lock_bh(priv->dev);
1092
1093         if (search_gid.global.interface_id !=
1094             priv->local_gid.global.interface_id)
1095                 /* There was a change while we were looking up the gid, bail
1096                  * here and let the next work sort this out
1097                  */
1098                 goto out;
1099
1100         /* The next section of code needs some background:
1101          * Per IB spec the port GUID can't change if the HCA is powered on.
1102          * port GUID is the basis for GID at index 0 which is the basis for
1103          * the default device address of a ipoib interface.
1104          *
1105          * so it seems the flow should be:
1106          * if user_changed_dev_addr && gid in gid tbl
1107          *      set bit dev_addr_set
1108          *      return true
1109          * else
1110          *      return false
1111          *
1112          * The issue is that there are devices that don't follow the spec,
1113          * they change the port GUID when the HCA is powered, so in order
1114          * not to break userspace applications, We need to check if the
1115          * user wanted to control the device address and we assume that
1116          * if he sets the device address back to be based on GID index 0,
1117          * he no longer wishs to control it.
1118          *
1119          * If the user doesn't control the the device address,
1120          * IPOIB_FLAG_DEV_ADDR_SET is set and ib_find_gid failed it means
1121          * the port GUID has changed and GID at index 0 has changed
1122          * so we need to change priv->local_gid and priv->dev->dev_addr
1123          * to reflect the new GID.
1124          */
1125         if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
1126                 if (!err && port == priv->port) {
1127                         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1128                         if (index == 0)
1129                                 clear_bit(IPOIB_FLAG_DEV_ADDR_CTRL,
1130                                           &priv->flags);
1131                         else
1132                                 set_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags);
1133                         ret = true;
1134                 } else {
1135                         ret = false;
1136                 }
1137         } else {
1138                 if (!err && port == priv->port) {
1139                         ret = true;
1140                 } else {
1141                         if (!test_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags)) {
1142                                 memcpy(&priv->local_gid, &gid0,
1143                                        sizeof(priv->local_gid));
1144                                 memcpy(priv->dev->dev_addr + 4, &gid0,
1145                                        sizeof(priv->local_gid));
1146                                 ret = true;
1147                         }
1148                 }
1149         }
1150
1151 out:
1152         netif_addr_unlock_bh(priv->dev);
1153
1154         return ret;
1155 }
1156
1157 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
1158                                 enum ipoib_flush_level level,
1159                                 int nesting)
1160 {
1161         struct ipoib_dev_priv *cpriv;
1162         struct net_device *dev = priv->dev;
1163         int result;
1164
1165         down_read_nested(&priv->vlan_rwsem, nesting);
1166
1167         /*
1168          * Flush any child interfaces too -- they might be up even if
1169          * the parent is down.
1170          */
1171         list_for_each_entry(cpriv, &priv->child_intfs, list)
1172                 __ipoib_ib_dev_flush(cpriv, level, nesting + 1);
1173
1174         up_read(&priv->vlan_rwsem);
1175
1176         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
1177             level != IPOIB_FLUSH_HEAVY) {
1178                 /* Make sure the dev_addr is set even if not flushing */
1179                 if (level == IPOIB_FLUSH_LIGHT)
1180                         ipoib_dev_addr_changed_valid(priv);
1181                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
1182                 return;
1183         }
1184
1185         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1186                 /* interface is down. update pkey and leave. */
1187                 if (level == IPOIB_FLUSH_HEAVY) {
1188                         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1189                                 update_parent_pkey(priv);
1190                         else
1191                                 update_child_pkey(priv);
1192                 } else if (level == IPOIB_FLUSH_LIGHT)
1193                         ipoib_dev_addr_changed_valid(priv);
1194                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1195                 return;
1196         }
1197
1198         if (level == IPOIB_FLUSH_HEAVY) {
1199                 /* child devices chase their origin pkey value, while non-child
1200                  * (parent) devices should always takes what present in pkey index 0
1201                  */
1202                 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1203                         result = update_child_pkey(priv);
1204                         if (result) {
1205                                 /* restart QP only if P_Key index is changed */
1206                                 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
1207                                 return;
1208                         }
1209
1210                 } else {
1211                         result = update_parent_pkey(priv);
1212                         /* restart QP only if P_Key value changed */
1213                         if (result) {
1214                                 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
1215                                 return;
1216                         }
1217                 }
1218         }
1219
1220         if (level == IPOIB_FLUSH_LIGHT) {
1221                 int oper_up;
1222                 ipoib_mark_paths_invalid(dev);
1223                 /* Set IPoIB operation as down to prevent races between:
1224                  * the flush flow which leaves MCG and on the fly joins
1225                  * which can happen during that time. mcast restart task
1226                  * should deal with join requests we missed.
1227                  */
1228                 oper_up = test_and_clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1229                 ipoib_mcast_dev_flush(dev);
1230                 if (oper_up)
1231                         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1232                 ipoib_flush_ah(dev);
1233         }
1234
1235         if (level >= IPOIB_FLUSH_NORMAL)
1236                 ipoib_ib_dev_down(dev);
1237
1238         if (level == IPOIB_FLUSH_HEAVY) {
1239                 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
1240                         ipoib_ib_dev_stop(dev);
1241
1242                 if (ipoib_ib_dev_open(dev))
1243                         return;
1244
1245                 if (netif_queue_stopped(dev))
1246                         netif_start_queue(dev);
1247         }
1248
1249         /*
1250          * The device could have been brought down between the start and when
1251          * we get here, don't bring it back up if it's not configured up
1252          */
1253         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1254                 if (level >= IPOIB_FLUSH_NORMAL)
1255                         ipoib_ib_dev_up(dev);
1256                 if (ipoib_dev_addr_changed_valid(priv))
1257                         ipoib_mcast_restart_task(&priv->restart_task);
1258         }
1259 }
1260
1261 void ipoib_ib_dev_flush_light(struct work_struct *work)
1262 {
1263         struct ipoib_dev_priv *priv =
1264                 container_of(work, struct ipoib_dev_priv, flush_light);
1265
1266         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0);
1267 }
1268
1269 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1270 {
1271         struct ipoib_dev_priv *priv =
1272                 container_of(work, struct ipoib_dev_priv, flush_normal);
1273
1274         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0);
1275 }
1276
1277 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1278 {
1279         struct ipoib_dev_priv *priv =
1280                 container_of(work, struct ipoib_dev_priv, flush_heavy);
1281
1282         rtnl_lock();
1283         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0);
1284         rtnl_unlock();
1285 }
1286
1287 void ipoib_ib_dev_cleanup(struct net_device *dev)
1288 {
1289         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1290
1291         ipoib_dbg(priv, "cleaning up ib_dev\n");
1292         /*
1293          * We must make sure there are no more (path) completions
1294          * that may wish to touch priv fields that are no longer valid
1295          */
1296         ipoib_flush_paths(dev);
1297
1298         ipoib_mcast_stop_thread(dev);
1299         ipoib_mcast_dev_flush(dev);
1300
1301         /*
1302          * All of our ah references aren't free until after
1303          * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1304          * the neighbor garbage collection is stopped and reaped.
1305          * That should all be done now, so make a final ah flush.
1306          */
1307         ipoib_stop_ah(dev);
1308
1309         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
1310
1311         priv->rn_ops->ndo_uninit(dev);
1312
1313         if (priv->pd) {
1314                 ib_dealloc_pd(priv->pd);
1315                 priv->pd = NULL;
1316         }
1317 }
1318