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