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