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 static DEFINE_MUTEX(pkey_mutex);
  55
  56 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
  57                                  struct ib_pd *pd, struct ib_ah_attr *attr)
  58 {
  59         struct ipoib_ah *ah;
  60         struct ib_ah *vah;
  61
  62         ah = kmalloc(sizeof *ah, GFP_KERNEL);
  63         if (!ah)
  64                 return ERR_PTR(-ENOMEM);
  65
  66         ah->dev       = dev;
  67         ah->last_send = 0;
  68         kref_init(&ah->ref);
  69
  70         vah = ib_create_ah(pd, attr);
  71         if (IS_ERR(vah)) {
  72                 kfree(ah);
  73                 ah = (struct ipoib_ah *)vah;
  74         } else {
  75                 ah->ah = vah;
  76                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
  77         }
  78
  79         return ah;
  80 }
  81
  82 void ipoib_free_ah(struct kref *kref)
  83 {
  84         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
  85         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
  86
  87         unsigned long flags;
  88
  89         spin_lock_irqsave(&priv->lock, flags);
  90         list_add_tail(&ah->list, &priv->dead_ahs);
  91         spin_unlock_irqrestore(&priv->lock, flags);
  92 }
  93
  94 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
  95                                   u64 mapping[IPOIB_UD_RX_SG])
  96 {
  97         ib_dma_unmap_single(priv->ca, mapping[0],
  98                             IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
  99                             DMA_FROM_DEVICE);
 100 }
 101
 102 static int ipoib_ib_post_receive(struct net_device *dev, int id)
 103 {
 104         struct ipoib_dev_priv *priv = netdev_priv(dev);
 105         struct ib_recv_wr *bad_wr;
 106         int ret;
 107
 108         priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
 109         priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
 110         priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
 111
 112
 113         ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
 114         if (unlikely(ret)) {
 115                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
 116                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
 117                 dev_kfree_skb_any(priv->rx_ring[id].skb);
 118                 priv->rx_ring[id].skb = NULL;
 119         }
 120
 121         return ret;
 122 }
 123
 124 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
 125 {
 126         struct ipoib_dev_priv *priv = netdev_priv(dev);
 127         struct sk_buff *skb;
 128         int buf_size;
 129         u64 *mapping;
 130
 131         buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
 132
 133         skb = dev_alloc_skb(buf_size + IPOIB_ENCAP_LEN);
 134         if (unlikely(!skb))
 135                 return NULL;
 136
 137         /*
 138          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
 139          * header.  So we need 4 more bytes to get to 48 and align the
 140          * IP header to a multiple of 16.
 141          */
 142         skb_reserve(skb, 4);
 143
 144         mapping = priv->rx_ring[id].mapping;
 145         mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
 146                                        DMA_FROM_DEVICE);
 147         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
 148                 goto error;
 149
 150         priv->rx_ring[id].skb = skb;
 151         return skb;
 152 error:
 153         dev_kfree_skb_any(skb);
 154         return NULL;
 155 }
 156
 157 static int ipoib_ib_post_receives(struct net_device *dev)
 158 {
 159         struct ipoib_dev_priv *priv = netdev_priv(dev);
 160         int i;
 161
 162         for (i = 0; i < ipoib_recvq_size; ++i) {
 163                 if (!ipoib_alloc_rx_skb(dev, i)) {
 164                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 165                         return -ENOMEM;
 166                 }
 167                 if (ipoib_ib_post_receive(dev, i)) {
 168                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 169                         return -EIO;
 170                 }
 171         }
 172
 173         return 0;
 174 }
 175
 176 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 177 {
 178         struct ipoib_dev_priv *priv = netdev_priv(dev);
 179         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
 180         struct sk_buff *skb;
 181         u64 mapping[IPOIB_UD_RX_SG];
 182         union ib_gid *dgid;
 183         union ib_gid *sgid;
 184
 185         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
 186                        wr_id, wc->status);
 187
 188         if (unlikely(wr_id >= ipoib_recvq_size)) {
 189                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
 190                            wr_id, ipoib_recvq_size);
 191                 return;
 192         }
 193
 194         skb  = priv->rx_ring[wr_id].skb;
 195
 196         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 197                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 198                         ipoib_warn(priv, "failed recv event "
 199                                    "(status=%d, wrid=%d vend_err %x)\n",
 200                                    wc->status, wr_id, wc->vendor_err);
 201                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
 202                 dev_kfree_skb_any(skb);
 203                 priv->rx_ring[wr_id].skb = NULL;
 204                 return;
 205         }
 206
 207         memcpy(mapping, priv->rx_ring[wr_id].mapping,
 208                IPOIB_UD_RX_SG * sizeof *mapping);
 209
 210         /*
 211          * If we can't allocate a new RX buffer, dump
 212          * this packet and reuse the old buffer.
 213          */
 214         if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
 215                 ++dev->stats.rx_dropped;
 216                 goto repost;
 217         }
 218
 219         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 220                        wc->byte_len, wc->slid);
 221
 222         ipoib_ud_dma_unmap_rx(priv, mapping);
 223
 224         skb_put(skb, wc->byte_len);
 225
 226         /* First byte of dgid signals multicast when 0xff */
 227         dgid = &((struct ib_grh *)skb->data)->dgid;
 228
 229         if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
 230                 skb->pkt_type = PACKET_HOST;
 231         else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
 232                 skb->pkt_type = PACKET_BROADCAST;
 233         else
 234                 skb->pkt_type = PACKET_MULTICAST;
 235
 236         sgid = &((struct ib_grh *)skb->data)->sgid;
 237
 238         /*
 239          * Drop packets that this interface sent, ie multicast packets
 240          * that the HCA has replicated.
 241          */
 242         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num) {
 243                 int need_repost = 1;
 244
 245                 if ((wc->wc_flags & IB_WC_GRH) &&
 246                     sgid->global.interface_id != priv->local_gid.global.interface_id)
 247                         need_repost = 0;
 248
 249                 if (need_repost) {
 250                         dev_kfree_skb_any(skb);
 251                         goto repost;
 252                 }
 253         }
 254
 255         skb_pull(skb, IB_GRH_BYTES);
 256
 257         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 258         skb_reset_mac_header(skb);
 259         skb_pull(skb, IPOIB_ENCAP_LEN);
 260
 261         ++dev->stats.rx_packets;
 262         dev->stats.rx_bytes += skb->len;
 263
 264         skb->dev = dev;
 265         if ((dev->features & NETIF_F_RXCSUM) &&
 266                         likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
 267                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 268
 269         napi_gro_receive(&priv->napi, skb);
 270
 271 repost:
 272         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 273                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 274                            "for buf %d\n", wr_id);
 275 }
 276
 277 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
 278 {
 279         struct sk_buff *skb = tx_req->skb;
 280         u64 *mapping = tx_req->mapping;
 281         int i;
 282         int off;
 283
 284         if (skb_headlen(skb)) {
 285                 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
 286                                                DMA_TO_DEVICE);
 287                 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
 288                         return -EIO;
 289
 290                 off = 1;
 291         } else
 292                 off = 0;
 293
 294         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 295                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 296                 mapping[i + off] = ib_dma_map_page(ca,
 297                                                  skb_frag_page(frag),
 298                                                  frag->page_offset, skb_frag_size(frag),
 299                                                  DMA_TO_DEVICE);
 300                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
 301                         goto partial_error;
 302         }
 303         return 0;
 304
 305 partial_error:
 306         for (; i > 0; --i) {
 307                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 308
 309                 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
 310         }
 311
 312         if (off)
 313                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 314
 315         return -EIO;
 316 }
 317
 318 void ipoib_dma_unmap_tx(struct ipoib_dev_priv *priv,
 319                         struct ipoib_tx_buf *tx_req)
 320 {
 321         struct sk_buff *skb = tx_req->skb;
 322         u64 *mapping = tx_req->mapping;
 323         int i;
 324         int off;
 325
 326         if (skb_headlen(skb)) {
 327                 ib_dma_unmap_single(priv->ca, mapping[0], skb_headlen(skb),
 328                                     DMA_TO_DEVICE);
 329                 off = 1;
 330         } else
 331                 off = 0;
 332
 333         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 334                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 335
 336                 ib_dma_unmap_page(priv->ca, mapping[i + off],
 337                                   skb_frag_size(frag), DMA_TO_DEVICE);
 338         }
 339 }
 340
 341 /*
 342  * As the result of a completion error the QP Can be transferred to SQE states.
 343  * The function checks if the (send)QP is in SQE state and
 344  * moves it back to RTS state, that in order to have it functional again.
 345  */
 346 static void ipoib_qp_state_validate_work(struct work_struct *work)
 347 {
 348         struct ipoib_qp_state_validate *qp_work =
 349                 container_of(work, struct ipoib_qp_state_validate, work);
 350
 351         struct ipoib_dev_priv *priv = qp_work->priv;
 352         struct ib_qp_attr qp_attr;
 353         struct ib_qp_init_attr query_init_attr;
 354         int ret;
 355
 356         ret = ib_query_qp(priv->qp, &qp_attr, IB_QP_STATE, &query_init_attr);
 357         if (ret) {
 358                 ipoib_warn(priv, "%s: Failed to query QP ret: %d\n",
 359                            __func__, ret);
 360                 goto free_res;
 361         }
 362         pr_info("%s: QP: 0x%x is in state: %d\n",
 363                 __func__, priv->qp->qp_num, qp_attr.qp_state);
 364
 365         /* currently support only in SQE->RTS transition*/
 366         if (qp_attr.qp_state == IB_QPS_SQE) {
 367                 qp_attr.qp_state = IB_QPS_RTS;
 368
 369                 ret = ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE);
 370                 if (ret) {
 371                         pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n",
 372                                 ret, priv->qp->qp_num);
 373                         goto free_res;
 374                 }
 375                 pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n",
 376                         __func__, priv->qp->qp_num);
 377         } else {
 378                 pr_warn("QP (%d) will stay in state: %d\n",
 379                         priv->qp->qp_num, qp_attr.qp_state);
 380         }
 381
 382 free_res:
 383         kfree(qp_work);
 384 }
 385
 386 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 387 {
 388         struct ipoib_dev_priv *priv = netdev_priv(dev);
 389         unsigned int wr_id = wc->wr_id;
 390         struct ipoib_tx_buf *tx_req;
 391
 392         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 393                        wr_id, wc->status);
 394
 395         if (unlikely(wr_id >= ipoib_sendq_size)) {
 396                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 397                            wr_id, ipoib_sendq_size);
 398                 return;
 399         }
 400
 401         tx_req = &priv->tx_ring[wr_id];
 402
 403         ipoib_dma_unmap_tx(priv, tx_req);
 404
 405         ++dev->stats.tx_packets;
 406         dev->stats.tx_bytes += tx_req->skb->len;
 407
 408         dev_kfree_skb_any(tx_req->skb);
 409
 410         ++priv->tx_tail;
 411         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 412             netif_queue_stopped(dev) &&
 413             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 414                 netif_wake_queue(dev);
 415
 416         if (wc->status != IB_WC_SUCCESS &&
 417             wc->status != IB_WC_WR_FLUSH_ERR) {
 418                 struct ipoib_qp_state_validate *qp_work;
 419                 ipoib_warn(priv, "failed send event "
 420                            "(status=%d, wrid=%d vend_err %x)\n",
 421                            wc->status, wr_id, wc->vendor_err);
 422                 qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC);
 423                 if (!qp_work) {
 424                         ipoib_warn(priv, "%s Failed alloc ipoib_qp_state_validate for qp: 0x%x\n",
 425                                    __func__, priv->qp->qp_num);
 426                         return;
 427                 }
 428
 429                 INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work);
 430                 qp_work->priv = priv;
 431                 queue_work(priv->wq, &qp_work->work);
 432         }
 433 }
 434
 435 static int poll_tx(struct ipoib_dev_priv *priv)
 436 {
 437         int n, i;
 438
 439         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
 440         for (i = 0; i < n; ++i)
 441                 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
 442
 443         return n == MAX_SEND_CQE;
 444 }
 445
 446 int ipoib_poll(struct napi_struct *napi, int budget)
 447 {
 448         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
 449         struct net_device *dev = priv->dev;
 450         int done;
 451         int t;
 452         int n, i;
 453
 454         done  = 0;
 455
 456 poll_more:
 457         while (done < budget) {
 458                 int max = (budget - done);
 459
 460                 t = min(IPOIB_NUM_WC, max);
 461                 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
 462
 463                 for (i = 0; i < n; i++) {
 464                         struct ib_wc *wc = priv->ibwc + i;
 465
 466                         if (wc->wr_id & IPOIB_OP_RECV) {
 467                                 ++done;
 468                                 if (wc->wr_id & IPOIB_OP_CM)
 469                                         ipoib_cm_handle_rx_wc(dev, wc);
 470                                 else
 471                                         ipoib_ib_handle_rx_wc(dev, wc);
 472                         } else
 473                                 ipoib_cm_handle_tx_wc(priv->dev, wc);
 474                 }
 475
 476                 if (n != t)
 477                         break;
 478         }
 479
 480         if (done < budget) {
 481                 napi_complete(napi);
 482                 if (unlikely(ib_req_notify_cq(priv->recv_cq,
 483                                               IB_CQ_NEXT_COMP |
 484                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 485                     napi_reschedule(napi))
 486                         goto poll_more;
 487         }
 488
 489         return done;
 490 }
 491
 492 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 493 {
 494         struct net_device *dev = dev_ptr;
 495         struct ipoib_dev_priv *priv = netdev_priv(dev);
 496
 497         napi_schedule(&priv->napi);
 498 }
 499
 500 static void drain_tx_cq(struct net_device *dev)
 501 {
 502         struct ipoib_dev_priv *priv = netdev_priv(dev);
 503
 504         netif_tx_lock(dev);
 505         while (poll_tx(priv))
 506                 ; /* nothing */
 507
 508         if (netif_queue_stopped(dev))
 509                 mod_timer(&priv->poll_timer, jiffies + 1);
 510
 511         netif_tx_unlock(dev);
 512 }
 513
 514 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
 515 {
 516         struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
 517
 518         mod_timer(&priv->poll_timer, jiffies);
 519 }
 520
 521 static inline int post_send(struct ipoib_dev_priv *priv,
 522                             unsigned int wr_id,
 523                             struct ib_ah *address, u32 qpn,
 524                             struct ipoib_tx_buf *tx_req,
 525                             void *head, int hlen)
 526 {
 527         struct ib_send_wr *bad_wr;
 528         struct sk_buff *skb = tx_req->skb;
 529
 530         ipoib_build_sge(priv, tx_req);
 531
 532         priv->tx_wr.wr.wr_id    = wr_id;
 533         priv->tx_wr.remote_qpn  = qpn;
 534         priv->tx_wr.ah          = address;
 535
 536         if (head) {
 537                 priv->tx_wr.mss         = skb_shinfo(skb)->gso_size;
 538                 priv->tx_wr.header      = head;
 539                 priv->tx_wr.hlen        = hlen;
 540                 priv->tx_wr.wr.opcode   = IB_WR_LSO;
 541         } else
 542                 priv->tx_wr.wr.opcode   = IB_WR_SEND;
 543
 544         return ib_post_send(priv->qp, &priv->tx_wr.wr, &bad_wr);
 545 }
 546
 547 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 548                 struct ipoib_ah *address, u32 qpn)
 549 {
 550         struct ipoib_dev_priv *priv = netdev_priv(dev);
 551         struct ipoib_tx_buf *tx_req;
 552         int hlen, rc;
 553         void *phead;
 554         unsigned usable_sge = priv->max_send_sge - !!skb_headlen(skb);
 555
 556         if (skb_is_gso(skb)) {
 557                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
 558                 phead = skb->data;
 559                 if (unlikely(!skb_pull(skb, hlen))) {
 560                         ipoib_warn(priv, "linear data too small\n");
 561                         ++dev->stats.tx_dropped;
 562                         ++dev->stats.tx_errors;
 563                         dev_kfree_skb_any(skb);
 564                         return;
 565                 }
 566         } else {
 567                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 568                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 569                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 570                         ++dev->stats.tx_dropped;
 571                         ++dev->stats.tx_errors;
 572                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 573                         return;
 574                 }
 575                 phead = NULL;
 576                 hlen  = 0;
 577         }
 578         if (skb_shinfo(skb)->nr_frags > usable_sge) {
 579                 if (skb_linearize(skb) < 0) {
 580                         ipoib_warn(priv, "skb could not be linearized\n");
 581                         ++dev->stats.tx_dropped;
 582                         ++dev->stats.tx_errors;
 583                         dev_kfree_skb_any(skb);
 584                         return;
 585                 }
 586                 /* Does skb_linearize return ok without reducing nr_frags? */
 587                 if (skb_shinfo(skb)->nr_frags > usable_sge) {
 588                         ipoib_warn(priv, "too many frags after skb linearize\n");
 589                         ++dev->stats.tx_dropped;
 590                         ++dev->stats.tx_errors;
 591                         dev_kfree_skb_any(skb);
 592                         return;
 593                 }
 594         }
 595
 596         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 597                        skb->len, address, qpn);
 598
 599         /*
 600          * We put the skb into the tx_ring _before_ we call post_send()
 601          * because it's entirely possible that the completion handler will
 602          * run before we execute anything after the post_send().  That
 603          * means we have to make sure everything is properly recorded and
 604          * our state is consistent before we call post_send().
 605          */
 606         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 607         tx_req->skb = skb;
 608         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 609                 ++dev->stats.tx_errors;
 610                 dev_kfree_skb_any(skb);
 611                 return;
 612         }
 613
 614         if (skb->ip_summed == CHECKSUM_PARTIAL)
 615                 priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
 616         else
 617                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 618
 619         if (++priv->tx_outstanding == ipoib_sendq_size) {
 620                 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 621                 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
 622                         ipoib_warn(priv, "request notify on send CQ failed\n");
 623                 netif_stop_queue(dev);
 624         }
 625
 626         skb_orphan(skb);
 627         skb_dst_drop(skb);
 628
 629         rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 630                        address->ah, qpn, tx_req, phead, hlen);
 631         if (unlikely(rc)) {
 632                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 633                 ++dev->stats.tx_errors;
 634                 --priv->tx_outstanding;
 635                 ipoib_dma_unmap_tx(priv, tx_req);
 636                 dev_kfree_skb_any(skb);
 637                 if (netif_queue_stopped(dev))
 638                         netif_wake_queue(dev);
 639         } else {
 640                 dev->trans_start = jiffies;
 641
 642                 address->last_send = priv->tx_head;
 643                 ++priv->tx_head;
 644         }
 645
 646         if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
 647                 while (poll_tx(priv))
 648                         ; /* nothing */
 649 }
 650
 651 static void __ipoib_reap_ah(struct net_device *dev)
 652 {
 653         struct ipoib_dev_priv *priv = netdev_priv(dev);
 654         struct ipoib_ah *ah, *tah;
 655         LIST_HEAD(remove_list);
 656         unsigned long flags;
 657
 658         netif_tx_lock_bh(dev);
 659         spin_lock_irqsave(&priv->lock, flags);
 660
 661         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 662                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 663                         list_del(&ah->list);
 664                         ib_destroy_ah(ah->ah);
 665                         kfree(ah);
 666                 }
 667
 668         spin_unlock_irqrestore(&priv->lock, flags);
 669         netif_tx_unlock_bh(dev);
 670 }
 671
 672 void ipoib_reap_ah(struct work_struct *work)
 673 {
 674         struct ipoib_dev_priv *priv =
 675                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 676         struct net_device *dev = priv->dev;
 677
 678         __ipoib_reap_ah(dev);
 679
 680         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 681                 queue_delayed_work(priv->wq, &priv->ah_reap_task,
 682                                    round_jiffies_relative(HZ));
 683 }
 684
 685 static void ipoib_flush_ah(struct net_device *dev)
 686 {
 687         struct ipoib_dev_priv *priv = netdev_priv(dev);
 688
 689         cancel_delayed_work(&priv->ah_reap_task);
 690         flush_workqueue(priv->wq);
 691         ipoib_reap_ah(&priv->ah_reap_task.work);
 692 }
 693
 694 static void ipoib_stop_ah(struct net_device *dev)
 695 {
 696         struct ipoib_dev_priv *priv = netdev_priv(dev);
 697
 698         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 699         ipoib_flush_ah(dev);
 700 }
 701
 702 static void ipoib_ib_tx_timer_func(unsigned long ctx)
 703 {
 704         drain_tx_cq((struct net_device *)ctx);
 705 }
 706
 707 int ipoib_ib_dev_open(struct net_device *dev)
 708 {
 709         struct ipoib_dev_priv *priv = netdev_priv(dev);
 710         int ret;
 711
 712         ipoib_pkey_dev_check_presence(dev);
 713
 714         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 715                 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
 716                            (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
 717                 return -1;
 718         }
 719
 720         ret = ipoib_init_qp(dev);
 721         if (ret) {
 722                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 723                 return -1;
 724         }
 725
 726         ret = ipoib_ib_post_receives(dev);
 727         if (ret) {
 728                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 729                 goto dev_stop;
 730         }
 731
 732         ret = ipoib_cm_dev_open(dev);
 733         if (ret) {
 734                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 735                 goto dev_stop;
 736         }
 737
 738         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 739         queue_delayed_work(priv->wq, &priv->ah_reap_task,
 740                            round_jiffies_relative(HZ));
 741
 742         if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 743                 napi_enable(&priv->napi);
 744
 745         return 0;
 746 dev_stop:
 747         if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 748                 napi_enable(&priv->napi);
 749         ipoib_ib_dev_stop(dev);
 750         return -1;
 751 }
 752
 753 void ipoib_pkey_dev_check_presence(struct net_device *dev)
 754 {
 755         struct ipoib_dev_priv *priv = netdev_priv(dev);
 756
 757         if (!(priv->pkey & 0x7fff) ||
 758             ib_find_pkey(priv->ca, priv->port, priv->pkey,
 759                          &priv->pkey_index))
 760                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 761         else
 762                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 763 }
 764
 765 int ipoib_ib_dev_up(struct net_device *dev)
 766 {
 767         struct ipoib_dev_priv *priv = netdev_priv(dev);
 768
 769         ipoib_pkey_dev_check_presence(dev);
 770
 771         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 772                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 773                 return 0;
 774         }
 775
 776         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 777
 778         return ipoib_mcast_start_thread(dev);
 779 }
 780
 781 int ipoib_ib_dev_down(struct net_device *dev)
 782 {
 783         struct ipoib_dev_priv *priv = netdev_priv(dev);
 784
 785         ipoib_dbg(priv, "downing ib_dev\n");
 786
 787         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 788         netif_carrier_off(dev);
 789
 790         ipoib_mcast_stop_thread(dev);
 791         ipoib_mcast_dev_flush(dev);
 792
 793         ipoib_flush_paths(dev);
 794
 795         return 0;
 796 }
 797
 798 static int recvs_pending(struct net_device *dev)
 799 {
 800         struct ipoib_dev_priv *priv = netdev_priv(dev);
 801         int pending = 0;
 802         int i;
 803
 804         for (i = 0; i < ipoib_recvq_size; ++i)
 805                 if (priv->rx_ring[i].skb)
 806                         ++pending;
 807
 808         return pending;
 809 }
 810
 811 void ipoib_drain_cq(struct net_device *dev)
 812 {
 813         struct ipoib_dev_priv *priv = netdev_priv(dev);
 814         int i, n;
 815
 816         /*
 817          * We call completion handling routines that expect to be
 818          * called from the BH-disabled NAPI poll context, so disable
 819          * BHs here too.
 820          */
 821         local_bh_disable();
 822
 823         do {
 824                 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
 825                 for (i = 0; i < n; ++i) {
 826                         /*
 827                          * Convert any successful completions to flush
 828                          * errors to avoid passing packets up the
 829                          * stack after bringing the device down.
 830                          */
 831                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 832                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 833
 834                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 835                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 836                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 837                                 else
 838                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 839                         } else
 840                                 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
 841                 }
 842         } while (n == IPOIB_NUM_WC);
 843
 844         while (poll_tx(priv))
 845                 ; /* nothing */
 846
 847         local_bh_enable();
 848 }
 849
 850 int ipoib_ib_dev_stop(struct net_device *dev)
 851 {
 852         struct ipoib_dev_priv *priv = netdev_priv(dev);
 853         struct ib_qp_attr qp_attr;
 854         unsigned long begin;
 855         struct ipoib_tx_buf *tx_req;
 856         int i;
 857
 858         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 859                 napi_disable(&priv->napi);
 860
 861         ipoib_cm_dev_stop(dev);
 862
 863         /*
 864          * Move our QP to the error state and then reinitialize in
 865          * when all work requests have completed or have been flushed.
 866          */
 867         qp_attr.qp_state = IB_QPS_ERR;
 868         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 869                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 870
 871         /* Wait for all sends and receives to complete */
 872         begin = jiffies;
 873
 874         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 875                 if (time_after(jiffies, begin + 5 * HZ)) {
 876                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 877                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 878
 879                         /*
 880                          * assume the HW is wedged and just free up
 881                          * all our pending work requests.
 882                          */
 883                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 884                                 tx_req = &priv->tx_ring[priv->tx_tail &
 885                                                         (ipoib_sendq_size - 1)];
 886                                 ipoib_dma_unmap_tx(priv, tx_req);
 887                                 dev_kfree_skb_any(tx_req->skb);
 888                                 ++priv->tx_tail;
 889                                 --priv->tx_outstanding;
 890                         }
 891
 892                         for (i = 0; i < ipoib_recvq_size; ++i) {
 893                                 struct ipoib_rx_buf *rx_req;
 894
 895                                 rx_req = &priv->rx_ring[i];
 896                                 if (!rx_req->skb)
 897                                         continue;
 898                                 ipoib_ud_dma_unmap_rx(priv,
 899                                                       priv->rx_ring[i].mapping);
 900                                 dev_kfree_skb_any(rx_req->skb);
 901                                 rx_req->skb = NULL;
 902                         }
 903
 904                         goto timeout;
 905                 }
 906
 907                 ipoib_drain_cq(dev);
 908
 909                 msleep(1);
 910         }
 911
 912         ipoib_dbg(priv, "All sends and receives done.\n");
 913
 914 timeout:
 915         del_timer_sync(&priv->poll_timer);
 916         qp_attr.qp_state = IB_QPS_RESET;
 917         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 918                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 919
 920         ipoib_flush_ah(dev);
 921
 922         ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
 923
 924         return 0;
 925 }
 926
 927 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 928 {
 929         struct ipoib_dev_priv *priv = netdev_priv(dev);
 930
 931         priv->ca = ca;
 932         priv->port = port;
 933         priv->qp = NULL;
 934
 935         if (ipoib_transport_dev_init(dev, ca)) {
 936                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 937                 return -ENODEV;
 938         }
 939
 940         setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
 941                     (unsigned long) dev);
 942
 943         if (dev->flags & IFF_UP) {
 944                 if (ipoib_ib_dev_open(dev)) {
 945                         ipoib_transport_dev_cleanup(dev);
 946                         return -ENODEV;
 947                 }
 948         }
 949
 950         return 0;
 951 }
 952
 953 /*
 954  * Takes whatever value which is in pkey index 0 and updates priv->pkey
 955  * returns 0 if the pkey value was changed.
 956  */
 957 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
 958 {
 959         int result;
 960         u16 prev_pkey;
 961
 962         prev_pkey = priv->pkey;
 963         result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
 964         if (result) {
 965                 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
 966                            priv->port, result);
 967                 return result;
 968         }
 969
 970         priv->pkey |= 0x8000;
 971
 972         if (prev_pkey != priv->pkey) {
 973                 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
 974                           prev_pkey, priv->pkey);
 975                 /*
 976                  * Update the pkey in the broadcast address, while making sure to set
 977                  * the full membership bit, so that we join the right broadcast group.
 978                  */
 979                 priv->dev->broadcast[8] = priv->pkey >> 8;
 980                 priv->dev->broadcast[9] = priv->pkey & 0xff;
 981                 return 0;
 982         }
 983
 984         return 1;
 985 }
 986 /*
 987  * returns 0 if pkey value was found in a different slot.
 988  */
 989 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
 990 {
 991         u16 old_index = priv->pkey_index;
 992
 993         priv->pkey_index = 0;
 994         ipoib_pkey_dev_check_presence(priv->dev);
 995
 996         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
 997             (old_index == priv->pkey_index))
 998                 return 1;
 999         return 0;
1000 }
1001
1002 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
1003                                 enum ipoib_flush_level level,
1004                                 int nesting)
1005 {
1006         struct ipoib_dev_priv *cpriv;
1007         struct net_device *dev = priv->dev;
1008         int result;
1009
1010         down_read_nested(&priv->vlan_rwsem, nesting);
1011
1012         /*
1013          * Flush any child interfaces too -- they might be up even if
1014          * the parent is down.
1015          */
1016         list_for_each_entry(cpriv, &priv->child_intfs, list)
1017                 __ipoib_ib_dev_flush(cpriv, level, nesting + 1);
1018
1019         up_read(&priv->vlan_rwsem);
1020
1021         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
1022             level != IPOIB_FLUSH_HEAVY) {
1023                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
1024                 return;
1025         }
1026
1027         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1028                 /* interface is down. update pkey and leave. */
1029                 if (level == IPOIB_FLUSH_HEAVY) {
1030                         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1031                                 update_parent_pkey(priv);
1032                         else
1033                                 update_child_pkey(priv);
1034                 }
1035                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1036                 return;
1037         }
1038
1039         if (level == IPOIB_FLUSH_HEAVY) {
1040                 /* child devices chase their origin pkey value, while non-child
1041                  * (parent) devices should always takes what present in pkey index 0
1042                  */
1043                 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1044                         result = update_child_pkey(priv);
1045                         if (result) {
1046                                 /* restart QP only if P_Key index is changed */
1047                                 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
1048                                 return;
1049                         }
1050
1051                 } else {
1052                         result = update_parent_pkey(priv);
1053                         /* restart QP only if P_Key value changed */
1054                         if (result) {
1055                                 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
1056                                 return;
1057                         }
1058                 }
1059         }
1060
1061         if (level == IPOIB_FLUSH_LIGHT) {
1062                 ipoib_mark_paths_invalid(dev);
1063                 ipoib_mcast_dev_flush(dev);
1064                 ipoib_flush_ah(dev);
1065         }
1066
1067         if (level >= IPOIB_FLUSH_NORMAL)
1068                 ipoib_ib_dev_down(dev);
1069
1070         if (level == IPOIB_FLUSH_HEAVY) {
1071                 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
1072                         ipoib_ib_dev_stop(dev);
1073                 if (ipoib_ib_dev_open(dev) != 0)
1074                         return;
1075                 if (netif_queue_stopped(dev))
1076                         netif_start_queue(dev);
1077         }
1078
1079         /*
1080          * The device could have been brought down between the start and when
1081          * we get here, don't bring it back up if it's not configured up
1082          */
1083         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1084                 if (level >= IPOIB_FLUSH_NORMAL)
1085                         ipoib_ib_dev_up(dev);
1086                 ipoib_mcast_restart_task(&priv->restart_task);
1087         }
1088 }
1089
1090 void ipoib_ib_dev_flush_light(struct work_struct *work)
1091 {
1092         struct ipoib_dev_priv *priv =
1093                 container_of(work, struct ipoib_dev_priv, flush_light);
1094
1095         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0);
1096 }
1097
1098 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1099 {
1100         struct ipoib_dev_priv *priv =
1101                 container_of(work, struct ipoib_dev_priv, flush_normal);
1102
1103         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0);
1104 }
1105
1106 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1107 {
1108         struct ipoib_dev_priv *priv =
1109                 container_of(work, struct ipoib_dev_priv, flush_heavy);
1110
1111         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0);
1112 }
1113
1114 void ipoib_ib_dev_cleanup(struct net_device *dev)
1115 {
1116         struct ipoib_dev_priv *priv = netdev_priv(dev);
1117
1118         ipoib_dbg(priv, "cleaning up ib_dev\n");
1119         /*
1120          * We must make sure there are no more (path) completions
1121          * that may wish to touch priv fields that are no longer valid
1122          */
1123         ipoib_flush_paths(dev);
1124
1125         ipoib_mcast_stop_thread(dev);
1126         ipoib_mcast_dev_flush(dev);
1127
1128         /*
1129          * All of our ah references aren't free until after
1130          * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1131          * the neighbor garbage collection is stopped and reaped.
1132          * That should all be done now, so make a final ah flush.
1133          */
1134         ipoib_stop_ah(dev);
1135
1136         ipoib_transport_dev_cleanup(dev);
1137 }
1138
1139