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
  61         ah = kmalloc(sizeof *ah, GFP_KERNEL);
  62         if (!ah)
  63                 return NULL;
  64
  65         ah->dev       = dev;
  66         ah->last_send = 0;
  67         kref_init(&ah->ref);
  68
  69         ah->ah = ib_create_ah(pd, attr);
  70         if (IS_ERR(ah->ah)) {
  71                 kfree(ah);
  72                 ah = NULL;
  73         } else
  74                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
  75
  76         return ah;
  77 }
  78
  79 void ipoib_free_ah(struct kref *kref)
  80 {
  81         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
  82         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
  83
  84         unsigned long flags;
  85
  86         spin_lock_irqsave(&priv->lock, flags);
  87         list_add_tail(&ah->list, &priv->dead_ahs);
  88         spin_unlock_irqrestore(&priv->lock, flags);
  89 }
  90
  91 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
  92                                   u64 mapping[IPOIB_UD_RX_SG])
  93 {
  94         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
  95                 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
  96                                     DMA_FROM_DEVICE);
  97                 ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
  98                                   DMA_FROM_DEVICE);
  99         } else
 100                 ib_dma_unmap_single(priv->ca, mapping[0],
 101                                     IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
 102                                     DMA_FROM_DEVICE);
 103 }
 104
 105 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
 106                                    struct sk_buff *skb,
 107                                    unsigned int length)
 108 {
 109         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
 110                 skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
 111                 unsigned int size;
 112                 /*
 113                  * There is only two buffers needed for max_payload = 4K,
 114                  * first buf size is IPOIB_UD_HEAD_SIZE
 115                  */
 116                 skb->tail += IPOIB_UD_HEAD_SIZE;
 117                 skb->len  += length;
 118
 119                 size = length - IPOIB_UD_HEAD_SIZE;
 120
 121                 skb_frag_size_set(frag, size);
 122                 skb->data_len += size;
 123                 skb->truesize += size;
 124         } else
 125                 skb_put(skb, length);
 126
 127 }
 128
 129 static int ipoib_ib_post_receive(struct net_device *dev, int id)
 130 {
 131         struct ipoib_dev_priv *priv = netdev_priv(dev);
 132         struct ib_recv_wr *bad_wr;
 133         int ret;
 134
 135         priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
 136         priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
 137         priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
 138
 139
 140         ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
 141         if (unlikely(ret)) {
 142                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
 143                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
 144                 dev_kfree_skb_any(priv->rx_ring[id].skb);
 145                 priv->rx_ring[id].skb = NULL;
 146         }
 147
 148         return ret;
 149 }
 150
 151 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
 152 {
 153         struct ipoib_dev_priv *priv = netdev_priv(dev);
 154         struct sk_buff *skb;
 155         int buf_size;
 156         u64 *mapping;
 157
 158         if (ipoib_ud_need_sg(priv->max_ib_mtu))
 159                 buf_size = IPOIB_UD_HEAD_SIZE;
 160         else
 161                 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
 162
 163         skb = dev_alloc_skb(buf_size + 4);
 164         if (unlikely(!skb))
 165                 return NULL;
 166
 167         /*
 168          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
 169          * header.  So we need 4 more bytes to get to 48 and align the
 170          * IP header to a multiple of 16.
 171          */
 172         skb_reserve(skb, 4);
 173
 174         mapping = priv->rx_ring[id].mapping;
 175         mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
 176                                        DMA_FROM_DEVICE);
 177         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
 178                 goto error;
 179
 180         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
 181                 struct page *page = alloc_page(GFP_ATOMIC);
 182                 if (!page)
 183                         goto partial_error;
 184                 skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
 185                 mapping[1] =
 186                         ib_dma_map_page(priv->ca, page,
 187                                         0, PAGE_SIZE, DMA_FROM_DEVICE);
 188                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
 189                         goto partial_error;
 190         }
 191
 192         priv->rx_ring[id].skb = skb;
 193         return skb;
 194
 195 partial_error:
 196         ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
 197 error:
 198         dev_kfree_skb_any(skb);
 199         return NULL;
 200 }
 201
 202 static int ipoib_ib_post_receives(struct net_device *dev)
 203 {
 204         struct ipoib_dev_priv *priv = netdev_priv(dev);
 205         int i;
 206
 207         for (i = 0; i < ipoib_recvq_size; ++i) {
 208                 if (!ipoib_alloc_rx_skb(dev, i)) {
 209                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 210                         return -ENOMEM;
 211                 }
 212                 if (ipoib_ib_post_receive(dev, i)) {
 213                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 214                         return -EIO;
 215                 }
 216         }
 217
 218         return 0;
 219 }
 220
 221 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 222 {
 223         struct ipoib_dev_priv *priv = netdev_priv(dev);
 224         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
 225         struct sk_buff *skb;
 226         u64 mapping[IPOIB_UD_RX_SG];
 227         union ib_gid *dgid;
 228
 229         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
 230                        wr_id, wc->status);
 231
 232         if (unlikely(wr_id >= ipoib_recvq_size)) {
 233                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
 234                            wr_id, ipoib_recvq_size);
 235                 return;
 236         }
 237
 238         skb  = priv->rx_ring[wr_id].skb;
 239
 240         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 241                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 242                         ipoib_warn(priv, "failed recv event "
 243                                    "(status=%d, wrid=%d vend_err %x)\n",
 244                                    wc->status, wr_id, wc->vendor_err);
 245                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
 246                 dev_kfree_skb_any(skb);
 247                 priv->rx_ring[wr_id].skb = NULL;
 248                 return;
 249         }
 250
 251         /*
 252          * Drop packets that this interface sent, ie multicast packets
 253          * that the HCA has replicated.
 254          */
 255         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
 256                 goto repost;
 257
 258         memcpy(mapping, priv->rx_ring[wr_id].mapping,
 259                IPOIB_UD_RX_SG * sizeof *mapping);
 260
 261         /*
 262          * If we can't allocate a new RX buffer, dump
 263          * this packet and reuse the old buffer.
 264          */
 265         if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
 266                 ++dev->stats.rx_dropped;
 267                 goto repost;
 268         }
 269
 270         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 271                        wc->byte_len, wc->slid);
 272
 273         ipoib_ud_dma_unmap_rx(priv, mapping);
 274         ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
 275
 276         /* First byte of dgid signals multicast when 0xff */
 277         dgid = &((struct ib_grh *)skb->data)->dgid;
 278
 279         if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
 280                 skb->pkt_type = PACKET_HOST;
 281         else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
 282                 skb->pkt_type = PACKET_BROADCAST;
 283         else
 284                 skb->pkt_type = PACKET_MULTICAST;
 285
 286         skb_pull(skb, IB_GRH_BYTES);
 287
 288         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 289         skb_reset_mac_header(skb);
 290         skb_pull(skb, IPOIB_ENCAP_LEN);
 291
 292         ++dev->stats.rx_packets;
 293         dev->stats.rx_bytes += skb->len;
 294
 295         skb->dev = dev;
 296         if ((dev->features & NETIF_F_RXCSUM) && likely(wc->csum_ok))
 297                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 298
 299         napi_gro_receive(&priv->napi, skb);
 300
 301 repost:
 302         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 303                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 304                            "for buf %d\n", wr_id);
 305 }
 306
 307 static int ipoib_dma_map_tx(struct ib_device *ca,
 308                             struct ipoib_tx_buf *tx_req)
 309 {
 310         struct sk_buff *skb = tx_req->skb;
 311         u64 *mapping = tx_req->mapping;
 312         int i;
 313         int off;
 314
 315         if (skb_headlen(skb)) {
 316                 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
 317                                                DMA_TO_DEVICE);
 318                 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
 319                         return -EIO;
 320
 321                 off = 1;
 322         } else
 323                 off = 0;
 324
 325         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 326                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 327                 mapping[i + off] = ib_dma_map_page(ca,
 328                                                  skb_frag_page(frag),
 329                                                  frag->page_offset, skb_frag_size(frag),
 330                                                  DMA_TO_DEVICE);
 331                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
 332                         goto partial_error;
 333         }
 334         return 0;
 335
 336 partial_error:
 337         for (; i > 0; --i) {
 338                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 339
 340                 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
 341         }
 342
 343         if (off)
 344                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 345
 346         return -EIO;
 347 }
 348
 349 static void ipoib_dma_unmap_tx(struct ib_device *ca,
 350                                struct ipoib_tx_buf *tx_req)
 351 {
 352         struct sk_buff *skb = tx_req->skb;
 353         u64 *mapping = tx_req->mapping;
 354         int i;
 355         int off;
 356
 357         if (skb_headlen(skb)) {
 358                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 359                 off = 1;
 360         } else
 361                 off = 0;
 362
 363         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 364                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 365
 366                 ib_dma_unmap_page(ca, mapping[i + off], skb_frag_size(frag),
 367                                   DMA_TO_DEVICE);
 368         }
 369 }
 370
 371 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 372 {
 373         struct ipoib_dev_priv *priv = netdev_priv(dev);
 374         unsigned int wr_id = wc->wr_id;
 375         struct ipoib_tx_buf *tx_req;
 376
 377         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 378                        wr_id, wc->status);
 379
 380         if (unlikely(wr_id >= ipoib_sendq_size)) {
 381                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 382                            wr_id, ipoib_sendq_size);
 383                 return;
 384         }
 385
 386         tx_req = &priv->tx_ring[wr_id];
 387
 388         ipoib_dma_unmap_tx(priv->ca, tx_req);
 389
 390         ++dev->stats.tx_packets;
 391         dev->stats.tx_bytes += tx_req->skb->len;
 392
 393         dev_kfree_skb_any(tx_req->skb);
 394
 395         ++priv->tx_tail;
 396         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 397             netif_queue_stopped(dev) &&
 398             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 399                 netif_wake_queue(dev);
 400
 401         if (wc->status != IB_WC_SUCCESS &&
 402             wc->status != IB_WC_WR_FLUSH_ERR)
 403                 ipoib_warn(priv, "failed send event "
 404                            "(status=%d, wrid=%d vend_err %x)\n",
 405                            wc->status, wr_id, wc->vendor_err);
 406 }
 407
 408 static int poll_tx(struct ipoib_dev_priv *priv)
 409 {
 410         int n, i;
 411
 412         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
 413         for (i = 0; i < n; ++i)
 414                 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
 415
 416         return n == MAX_SEND_CQE;
 417 }
 418
 419 int ipoib_poll(struct napi_struct *napi, int budget)
 420 {
 421         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
 422         struct net_device *dev = priv->dev;
 423         int done;
 424         int t;
 425         int n, i;
 426
 427         done  = 0;
 428
 429 poll_more:
 430         while (done < budget) {
 431                 int max = (budget - done);
 432
 433                 t = min(IPOIB_NUM_WC, max);
 434                 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
 435
 436                 for (i = 0; i < n; i++) {
 437                         struct ib_wc *wc = priv->ibwc + i;
 438
 439                         if (wc->wr_id & IPOIB_OP_RECV) {
 440                                 ++done;
 441                                 if (wc->wr_id & IPOIB_OP_CM)
 442                                         ipoib_cm_handle_rx_wc(dev, wc);
 443                                 else
 444                                         ipoib_ib_handle_rx_wc(dev, wc);
 445                         } else
 446                                 ipoib_cm_handle_tx_wc(priv->dev, wc);
 447                 }
 448
 449                 if (n != t)
 450                         break;
 451         }
 452
 453         if (done < budget) {
 454                 napi_complete(napi);
 455                 if (unlikely(ib_req_notify_cq(priv->recv_cq,
 456                                               IB_CQ_NEXT_COMP |
 457                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 458                     napi_reschedule(napi))
 459                         goto poll_more;
 460         }
 461
 462         return done;
 463 }
 464
 465 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 466 {
 467         struct net_device *dev = dev_ptr;
 468         struct ipoib_dev_priv *priv = netdev_priv(dev);
 469
 470         napi_schedule(&priv->napi);
 471 }
 472
 473 static void drain_tx_cq(struct net_device *dev)
 474 {
 475         struct ipoib_dev_priv *priv = netdev_priv(dev);
 476
 477         netif_tx_lock(dev);
 478         while (poll_tx(priv))
 479                 ; /* nothing */
 480
 481         if (netif_queue_stopped(dev))
 482                 mod_timer(&priv->poll_timer, jiffies + 1);
 483
 484         netif_tx_unlock(dev);
 485 }
 486
 487 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
 488 {
 489         struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
 490
 491         mod_timer(&priv->poll_timer, jiffies);
 492 }
 493
 494 static inline int post_send(struct ipoib_dev_priv *priv,
 495                             unsigned int wr_id,
 496                             struct ib_ah *address, u32 qpn,
 497                             struct ipoib_tx_buf *tx_req,
 498                             void *head, int hlen)
 499 {
 500         struct ib_send_wr *bad_wr;
 501         int i, off;
 502         struct sk_buff *skb = tx_req->skb;
 503         skb_frag_t *frags = skb_shinfo(skb)->frags;
 504         int nr_frags = skb_shinfo(skb)->nr_frags;
 505         u64 *mapping = tx_req->mapping;
 506
 507         if (skb_headlen(skb)) {
 508                 priv->tx_sge[0].addr         = mapping[0];
 509                 priv->tx_sge[0].length       = skb_headlen(skb);
 510                 off = 1;
 511         } else
 512                 off = 0;
 513
 514         for (i = 0; i < nr_frags; ++i) {
 515                 priv->tx_sge[i + off].addr = mapping[i + off];
 516                 priv->tx_sge[i + off].length = skb_frag_size(&frags[i]);
 517         }
 518         priv->tx_wr.num_sge          = nr_frags + off;
 519         priv->tx_wr.wr_id            = wr_id;
 520         priv->tx_wr.wr.ud.remote_qpn = qpn;
 521         priv->tx_wr.wr.ud.ah         = address;
 522
 523         if (head) {
 524                 priv->tx_wr.wr.ud.mss    = skb_shinfo(skb)->gso_size;
 525                 priv->tx_wr.wr.ud.header = head;
 526                 priv->tx_wr.wr.ud.hlen   = hlen;
 527                 priv->tx_wr.opcode       = IB_WR_LSO;
 528         } else
 529                 priv->tx_wr.opcode       = IB_WR_SEND;
 530
 531         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 532 }
 533
 534 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 535                 struct ipoib_ah *address, u32 qpn)
 536 {
 537         struct ipoib_dev_priv *priv = netdev_priv(dev);
 538         struct ipoib_tx_buf *tx_req;
 539         int hlen, rc;
 540         void *phead;
 541
 542         if (skb_is_gso(skb)) {
 543                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
 544                 phead = skb->data;
 545                 if (unlikely(!skb_pull(skb, hlen))) {
 546                         ipoib_warn(priv, "linear data too small\n");
 547                         ++dev->stats.tx_dropped;
 548                         ++dev->stats.tx_errors;
 549                         dev_kfree_skb_any(skb);
 550                         return;
 551                 }
 552         } else {
 553                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 554                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 555                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 556                         ++dev->stats.tx_dropped;
 557                         ++dev->stats.tx_errors;
 558                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 559                         return;
 560                 }
 561                 phead = NULL;
 562                 hlen  = 0;
 563         }
 564
 565         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 566                        skb->len, address, qpn);
 567
 568         /*
 569          * We put the skb into the tx_ring _before_ we call post_send()
 570          * because it's entirely possible that the completion handler will
 571          * run before we execute anything after the post_send().  That
 572          * means we have to make sure everything is properly recorded and
 573          * our state is consistent before we call post_send().
 574          */
 575         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 576         tx_req->skb = skb;
 577         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 578                 ++dev->stats.tx_errors;
 579                 dev_kfree_skb_any(skb);
 580                 return;
 581         }
 582
 583         if (skb->ip_summed == CHECKSUM_PARTIAL)
 584                 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
 585         else
 586                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
 587
 588         if (++priv->tx_outstanding == ipoib_sendq_size) {
 589                 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 590                 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
 591                         ipoib_warn(priv, "request notify on send CQ failed\n");
 592                 netif_stop_queue(dev);
 593         }
 594
 595         rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 596                        address->ah, qpn, tx_req, phead, hlen);
 597         if (unlikely(rc)) {
 598                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 599                 ++dev->stats.tx_errors;
 600                 --priv->tx_outstanding;
 601                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 602                 dev_kfree_skb_any(skb);
 603                 if (netif_queue_stopped(dev))
 604                         netif_wake_queue(dev);
 605         } else {
 606                 dev->trans_start = jiffies;
 607
 608                 address->last_send = priv->tx_head;
 609                 ++priv->tx_head;
 610                 skb_orphan(skb);
 611
 612         }
 613
 614         if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
 615                 while (poll_tx(priv))
 616                         ; /* nothing */
 617 }
 618
 619 static void __ipoib_reap_ah(struct net_device *dev)
 620 {
 621         struct ipoib_dev_priv *priv = netdev_priv(dev);
 622         struct ipoib_ah *ah, *tah;
 623         LIST_HEAD(remove_list);
 624         unsigned long flags;
 625
 626         netif_tx_lock_bh(dev);
 627         spin_lock_irqsave(&priv->lock, flags);
 628
 629         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 630                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 631                         list_del(&ah->list);
 632                         ib_destroy_ah(ah->ah);
 633                         kfree(ah);
 634                 }
 635
 636         spin_unlock_irqrestore(&priv->lock, flags);
 637         netif_tx_unlock_bh(dev);
 638 }
 639
 640 void ipoib_reap_ah(struct work_struct *work)
 641 {
 642         struct ipoib_dev_priv *priv =
 643                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 644         struct net_device *dev = priv->dev;
 645
 646         __ipoib_reap_ah(dev);
 647
 648         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 649                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 650                                    round_jiffies_relative(HZ));
 651 }
 652
 653 static void ipoib_ib_tx_timer_func(unsigned long ctx)
 654 {
 655         drain_tx_cq((struct net_device *)ctx);
 656 }
 657
 658 int ipoib_ib_dev_open(struct net_device *dev)
 659 {
 660         struct ipoib_dev_priv *priv = netdev_priv(dev);
 661         int ret;
 662
 663         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
 664                 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
 665                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 666                 return -1;
 667         }
 668         set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 669
 670         ret = ipoib_init_qp(dev);
 671         if (ret) {
 672                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 673                 return -1;
 674         }
 675
 676         ret = ipoib_ib_post_receives(dev);
 677         if (ret) {
 678                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 679                 ipoib_ib_dev_stop(dev, 1);
 680                 return -1;
 681         }
 682
 683         ret = ipoib_cm_dev_open(dev);
 684         if (ret) {
 685                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 686                 ipoib_ib_dev_stop(dev, 1);
 687                 return -1;
 688         }
 689
 690         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 691         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 692                            round_jiffies_relative(HZ));
 693
 694         if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 695                 napi_enable(&priv->napi);
 696
 697         return 0;
 698 }
 699
 700 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
 701 {
 702         struct ipoib_dev_priv *priv = netdev_priv(dev);
 703         u16 pkey_index = 0;
 704
 705         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
 706                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 707         else
 708                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 709 }
 710
 711 int ipoib_ib_dev_up(struct net_device *dev)
 712 {
 713         struct ipoib_dev_priv *priv = netdev_priv(dev);
 714
 715         ipoib_pkey_dev_check_presence(dev);
 716
 717         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 718                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 719                 return 0;
 720         }
 721
 722         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 723
 724         return ipoib_mcast_start_thread(dev);
 725 }
 726
 727 int ipoib_ib_dev_down(struct net_device *dev, int flush)
 728 {
 729         struct ipoib_dev_priv *priv = netdev_priv(dev);
 730
 731         ipoib_dbg(priv, "downing ib_dev\n");
 732
 733         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 734         netif_carrier_off(dev);
 735
 736         /* Shutdown the P_Key thread if still active */
 737         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 738                 mutex_lock(&pkey_mutex);
 739                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
 740                 cancel_delayed_work(&priv->pkey_poll_task);
 741                 mutex_unlock(&pkey_mutex);
 742                 if (flush)
 743                         flush_workqueue(ipoib_workqueue);
 744         }
 745
 746         ipoib_mcast_stop_thread(dev, flush);
 747         ipoib_mcast_dev_flush(dev);
 748
 749         ipoib_flush_paths(dev);
 750
 751         return 0;
 752 }
 753
 754 static int recvs_pending(struct net_device *dev)
 755 {
 756         struct ipoib_dev_priv *priv = netdev_priv(dev);
 757         int pending = 0;
 758         int i;
 759
 760         for (i = 0; i < ipoib_recvq_size; ++i)
 761                 if (priv->rx_ring[i].skb)
 762                         ++pending;
 763
 764         return pending;
 765 }
 766
 767 void ipoib_drain_cq(struct net_device *dev)
 768 {
 769         struct ipoib_dev_priv *priv = netdev_priv(dev);
 770         int i, n;
 771
 772         /*
 773          * We call completion handling routines that expect to be
 774          * called from the BH-disabled NAPI poll context, so disable
 775          * BHs here too.
 776          */
 777         local_bh_disable();
 778
 779         do {
 780                 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
 781                 for (i = 0; i < n; ++i) {
 782                         /*
 783                          * Convert any successful completions to flush
 784                          * errors to avoid passing packets up the
 785                          * stack after bringing the device down.
 786                          */
 787                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 788                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 789
 790                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 791                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 792                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 793                                 else
 794                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 795                         } else
 796                                 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
 797                 }
 798         } while (n == IPOIB_NUM_WC);
 799
 800         while (poll_tx(priv))
 801                 ; /* nothing */
 802
 803         local_bh_enable();
 804 }
 805
 806 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
 807 {
 808         struct ipoib_dev_priv *priv = netdev_priv(dev);
 809         struct ib_qp_attr qp_attr;
 810         unsigned long begin;
 811         struct ipoib_tx_buf *tx_req;
 812         int i;
 813
 814         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 815                 napi_disable(&priv->napi);
 816
 817         ipoib_cm_dev_stop(dev);
 818
 819         /*
 820          * Move our QP to the error state and then reinitialize in
 821          * when all work requests have completed or have been flushed.
 822          */
 823         qp_attr.qp_state = IB_QPS_ERR;
 824         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 825                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 826
 827         /* Wait for all sends and receives to complete */
 828         begin = jiffies;
 829
 830         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 831                 if (time_after(jiffies, begin + 5 * HZ)) {
 832                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 833                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 834
 835                         /*
 836                          * assume the HW is wedged and just free up
 837                          * all our pending work requests.
 838                          */
 839                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 840                                 tx_req = &priv->tx_ring[priv->tx_tail &
 841                                                         (ipoib_sendq_size - 1)];
 842                                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 843                                 dev_kfree_skb_any(tx_req->skb);
 844                                 ++priv->tx_tail;
 845                                 --priv->tx_outstanding;
 846                         }
 847
 848                         for (i = 0; i < ipoib_recvq_size; ++i) {
 849                                 struct ipoib_rx_buf *rx_req;
 850
 851                                 rx_req = &priv->rx_ring[i];
 852                                 if (!rx_req->skb)
 853                                         continue;
 854                                 ipoib_ud_dma_unmap_rx(priv,
 855                                                       priv->rx_ring[i].mapping);
 856                                 dev_kfree_skb_any(rx_req->skb);
 857                                 rx_req->skb = NULL;
 858                         }
 859
 860                         goto timeout;
 861                 }
 862
 863                 ipoib_drain_cq(dev);
 864
 865                 msleep(1);
 866         }
 867
 868         ipoib_dbg(priv, "All sends and receives done.\n");
 869
 870 timeout:
 871         del_timer_sync(&priv->poll_timer);
 872         qp_attr.qp_state = IB_QPS_RESET;
 873         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 874                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 875
 876         /* Wait for all AHs to be reaped */
 877         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 878         cancel_delayed_work(&priv->ah_reap_task);
 879         if (flush)
 880                 flush_workqueue(ipoib_workqueue);
 881
 882         begin = jiffies;
 883
 884         while (!list_empty(&priv->dead_ahs)) {
 885                 __ipoib_reap_ah(dev);
 886
 887                 if (time_after(jiffies, begin + HZ)) {
 888                         ipoib_warn(priv, "timing out; will leak address handles\n");
 889                         break;
 890                 }
 891
 892                 msleep(1);
 893         }
 894
 895         ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
 896
 897         return 0;
 898 }
 899
 900 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 901 {
 902         struct ipoib_dev_priv *priv = netdev_priv(dev);
 903
 904         priv->ca = ca;
 905         priv->port = port;
 906         priv->qp = NULL;
 907
 908         if (ipoib_transport_dev_init(dev, ca)) {
 909                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 910                 return -ENODEV;
 911         }
 912
 913         setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
 914                     (unsigned long) dev);
 915
 916         if (dev->flags & IFF_UP) {
 917                 if (ipoib_ib_dev_open(dev)) {
 918                         ipoib_transport_dev_cleanup(dev);
 919                         return -ENODEV;
 920                 }
 921         }
 922
 923         return 0;
 924 }
 925
 926 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
 927                                 enum ipoib_flush_level level)
 928 {
 929         struct ipoib_dev_priv *cpriv;
 930         struct net_device *dev = priv->dev;
 931         u16 new_index;
 932
 933         mutex_lock(&priv->vlan_mutex);
 934
 935         /*
 936          * Flush any child interfaces too -- they might be up even if
 937          * the parent is down.
 938          */
 939         list_for_each_entry(cpriv, &priv->child_intfs, list)
 940                 __ipoib_ib_dev_flush(cpriv, level);
 941
 942         mutex_unlock(&priv->vlan_mutex);
 943
 944         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
 945                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
 946                 return;
 947         }
 948
 949         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 950                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
 951                 return;
 952         }
 953
 954         if (level == IPOIB_FLUSH_HEAVY) {
 955                 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
 956                         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 957                         ipoib_ib_dev_down(dev, 0);
 958                         ipoib_ib_dev_stop(dev, 0);
 959                         if (ipoib_pkey_dev_delay_open(dev))
 960                                 return;
 961                 }
 962
 963                 /* restart QP only if P_Key index is changed */
 964                 if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
 965                     new_index == priv->pkey_index) {
 966                         ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
 967                         return;
 968                 }
 969                 priv->pkey_index = new_index;
 970         }
 971
 972         if (level == IPOIB_FLUSH_LIGHT) {
 973                 ipoib_mark_paths_invalid(dev);
 974                 ipoib_mcast_dev_flush(dev);
 975         }
 976
 977         if (level >= IPOIB_FLUSH_NORMAL)
 978                 ipoib_ib_dev_down(dev, 0);
 979
 980         if (level == IPOIB_FLUSH_HEAVY) {
 981                 ipoib_ib_dev_stop(dev, 0);
 982                 ipoib_ib_dev_open(dev);
 983         }
 984
 985         /*
 986          * The device could have been brought down between the start and when
 987          * we get here, don't bring it back up if it's not configured up
 988          */
 989         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 990                 if (level >= IPOIB_FLUSH_NORMAL)
 991                         ipoib_ib_dev_up(dev);
 992                 ipoib_mcast_restart_task(&priv->restart_task);
 993         }
 994 }
 995
 996 void ipoib_ib_dev_flush_light(struct work_struct *work)
 997 {
 998         struct ipoib_dev_priv *priv =
 999                 container_of(work, struct ipoib_dev_priv, flush_light);
1000
1001         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
1002 }
1003
1004 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1005 {
1006         struct ipoib_dev_priv *priv =
1007                 container_of(work, struct ipoib_dev_priv, flush_normal);
1008
1009         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1010 }
1011
1012 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1013 {
1014         struct ipoib_dev_priv *priv =
1015                 container_of(work, struct ipoib_dev_priv, flush_heavy);
1016
1017         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1018 }
1019
1020 void ipoib_ib_dev_cleanup(struct net_device *dev)
1021 {
1022         struct ipoib_dev_priv *priv = netdev_priv(dev);
1023
1024         ipoib_dbg(priv, "cleaning up ib_dev\n");
1025
1026         ipoib_mcast_stop_thread(dev, 1);
1027         ipoib_mcast_dev_flush(dev);
1028
1029         ipoib_transport_dev_cleanup(dev);
1030 }
1031
1032 /*
1033  * Delayed P_Key Assigment Interim Support
1034  *
1035  * The following is initial implementation of delayed P_Key assigment
1036  * mechanism. It is using the same approach implemented for the multicast
1037  * group join. The single goal of this implementation is to quickly address
1038  * Bug #2507. This implementation will probably be removed when the P_Key
1039  * change async notification is available.
1040  */
1041
1042 void ipoib_pkey_poll(struct work_struct *work)
1043 {
1044         struct ipoib_dev_priv *priv =
1045                 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1046         struct net_device *dev = priv->dev;
1047
1048         ipoib_pkey_dev_check_presence(dev);
1049
1050         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1051                 ipoib_open(dev);
1052         else {
1053                 mutex_lock(&pkey_mutex);
1054                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1055                         queue_delayed_work(ipoib_workqueue,
1056                                            &priv->pkey_poll_task,
1057                                            HZ);
1058                 mutex_unlock(&pkey_mutex);
1059         }
1060 }
1061
1062 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1063 {
1064         struct ipoib_dev_priv *priv = netdev_priv(dev);
1065
1066         /* Look for the interface pkey value in the IB Port P_Key table and */
1067         /* set the interface pkey assigment flag                            */
1068         ipoib_pkey_dev_check_presence(dev);
1069
1070         /* P_Key value not assigned yet - start polling */
1071         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1072                 mutex_lock(&pkey_mutex);
1073                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1074                 queue_delayed_work(ipoib_workqueue,
1075                                    &priv->pkey_poll_task,
1076                                    HZ);
1077                 mutex_unlock(&pkey_mutex);
1078                 return 1;
1079         }
1080
1081         return 0;
1082 }