git.oblomov.eu Git - linux-2.6/blob - 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/dma-mapping.h>
  38
  39 #include <rdma/ib_cache.h>
  40 #include <linux/ip.h>
  41 #include <linux/tcp.h>
  42
  43 #include "ipoib.h"
  44
  45 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  46 static int data_debug_level;
  47
  48 module_param(data_debug_level, int, 0644);
  49 MODULE_PARM_DESC(data_debug_level,
  50                  "Enable data path debug tracing if > 0");
  51 #endif
  52
  53 static DEFINE_MUTEX(pkey_mutex);
  54
  55 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
  56                                  struct ib_pd *pd, struct ib_ah_attr *attr)
  57 {
  58         struct ipoib_ah *ah;
  59
  60         ah = kmalloc(sizeof *ah, GFP_KERNEL);
  61         if (!ah)
  62                 return NULL;
  63
  64         ah->dev       = dev;
  65         ah->last_send = 0;
  66         kref_init(&ah->ref);
  67
  68         ah->ah = ib_create_ah(pd, attr);
  69         if (IS_ERR(ah->ah)) {
  70                 kfree(ah);
  71                 ah = NULL;
  72         } else
  73                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
  74
  75         return ah;
  76 }
  77
  78 void ipoib_free_ah(struct kref *kref)
  79 {
  80         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
  81         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
  82
  83         unsigned long flags;
  84
  85         spin_lock_irqsave(&priv->lock, flags);
  86         list_add_tail(&ah->list, &priv->dead_ahs);
  87         spin_unlock_irqrestore(&priv->lock, flags);
  88 }
  89
  90 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
  91                                   u64 mapping[IPOIB_UD_RX_SG])
  92 {
  93         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
  94                 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
  95                                     DMA_FROM_DEVICE);
  96                 ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
  97                                   DMA_FROM_DEVICE);
  98         } else
  99                 ib_dma_unmap_single(priv->ca, mapping[0],
 100                                     IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
 101                                     DMA_FROM_DEVICE);
 102 }
 103
 104 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
 105                                    struct sk_buff *skb,
 106                                    unsigned int length)
 107 {
 108         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
 109                 skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
 110                 unsigned int size;
 111                 /*
 112                  * There is only two buffers needed for max_payload = 4K,
 113                  * first buf size is IPOIB_UD_HEAD_SIZE
 114                  */
 115                 skb->tail += IPOIB_UD_HEAD_SIZE;
 116                 skb->len  += length;
 117
 118                 size = length - IPOIB_UD_HEAD_SIZE;
 119
 120                 frag->size     = size;
 121                 skb->data_len += size;
 122                 skb->truesize += size;
 123         } else
 124                 skb_put(skb, length);
 125
 126 }
 127
 128 static int ipoib_ib_post_receive(struct net_device *dev, int id)
 129 {
 130         struct ipoib_dev_priv *priv = netdev_priv(dev);
 131         struct ib_recv_wr *bad_wr;
 132         int ret;
 133
 134         priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
 135         priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
 136         priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
 137
 138
 139         ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
 140         if (unlikely(ret)) {
 141                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
 142                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
 143                 dev_kfree_skb_any(priv->rx_ring[id].skb);
 144                 priv->rx_ring[id].skb = NULL;
 145         }
 146
 147         return ret;
 148 }
 149
 150 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
 151 {
 152         struct ipoib_dev_priv *priv = netdev_priv(dev);
 153         struct sk_buff *skb;
 154         int buf_size;
 155         u64 *mapping;
 156
 157         if (ipoib_ud_need_sg(priv->max_ib_mtu))
 158                 buf_size = IPOIB_UD_HEAD_SIZE;
 159         else
 160                 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
 161
 162         skb = dev_alloc_skb(buf_size + 4);
 163         if (unlikely(!skb))
 164                 return NULL;
 165
 166         /*
 167          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
 168          * header.  So we need 4 more bytes to get to 48 and align the
 169          * IP header to a multiple of 16.
 170          */
 171         skb_reserve(skb, 4);
 172
 173         mapping = priv->rx_ring[id].mapping;
 174         mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
 175                                        DMA_FROM_DEVICE);
 176         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
 177                 goto error;
 178
 179         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
 180                 struct page *page = alloc_page(GFP_ATOMIC);
 181                 if (!page)
 182                         goto partial_error;
 183                 skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
 184                 mapping[1] =
 185                         ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[0].page,
 186                                         0, PAGE_SIZE, DMA_FROM_DEVICE);
 187                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
 188                         goto partial_error;
 189         }
 190
 191         priv->rx_ring[id].skb = skb;
 192         return skb;
 193
 194 partial_error:
 195         ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
 196 error:
 197         dev_kfree_skb_any(skb);
 198         return NULL;
 199 }
 200
 201 static int ipoib_ib_post_receives(struct net_device *dev)
 202 {
 203         struct ipoib_dev_priv *priv = netdev_priv(dev);
 204         int i;
 205
 206         for (i = 0; i < ipoib_recvq_size; ++i) {
 207                 if (!ipoib_alloc_rx_skb(dev, i)) {
 208                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 209                         return -ENOMEM;
 210                 }
 211                 if (ipoib_ib_post_receive(dev, i)) {
 212                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 213                         return -EIO;
 214                 }
 215         }
 216
 217         return 0;
 218 }
 219
 220 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 221 {
 222         struct ipoib_dev_priv *priv = netdev_priv(dev);
 223         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
 224         struct sk_buff *skb;
 225         u64 mapping[IPOIB_UD_RX_SG];
 226
 227         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
 228                        wr_id, wc->status);
 229
 230         if (unlikely(wr_id >= ipoib_recvq_size)) {
 231                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
 232                            wr_id, ipoib_recvq_size);
 233                 return;
 234         }
 235
 236         skb  = priv->rx_ring[wr_id].skb;
 237
 238         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 239                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 240                         ipoib_warn(priv, "failed recv event "
 241                                    "(status=%d, wrid=%d vend_err %x)\n",
 242                                    wc->status, wr_id, wc->vendor_err);
 243                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
 244                 dev_kfree_skb_any(skb);
 245                 priv->rx_ring[wr_id].skb = NULL;
 246                 return;
 247         }
 248
 249         /*
 250          * Drop packets that this interface sent, ie multicast packets
 251          * that the HCA has replicated.
 252          */
 253         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
 254                 goto repost;
 255
 256         memcpy(mapping, priv->rx_ring[wr_id].mapping,
 257                IPOIB_UD_RX_SG * sizeof *mapping);
 258
 259         /*
 260          * If we can't allocate a new RX buffer, dump
 261          * this packet and reuse the old buffer.
 262          */
 263         if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
 264                 ++dev->stats.rx_dropped;
 265                 goto repost;
 266         }
 267
 268         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 269                        wc->byte_len, wc->slid);
 270
 271         ipoib_ud_dma_unmap_rx(priv, mapping);
 272         ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
 273
 274         skb_pull(skb, IB_GRH_BYTES);
 275
 276         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 277         skb_reset_mac_header(skb);
 278         skb_pull(skb, IPOIB_ENCAP_LEN);
 279
 280         dev->last_rx = jiffies;
 281         ++dev->stats.rx_packets;
 282         dev->stats.rx_bytes += skb->len;
 283
 284         skb->dev = dev;
 285         /* XXX get correct PACKET_ type here */
 286         skb->pkt_type = PACKET_HOST;
 287
 288         if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
 289                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 290
 291         if (dev->features & NETIF_F_LRO)
 292                 lro_receive_skb(&priv->lro.lro_mgr, skb, NULL);
 293         else
 294                 netif_receive_skb(skb);
 295
 296 repost:
 297         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 298                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 299                            "for buf %d\n", wr_id);
 300 }
 301
 302 static int ipoib_dma_map_tx(struct ib_device *ca,
 303                             struct ipoib_tx_buf *tx_req)
 304 {
 305         struct sk_buff *skb = tx_req->skb;
 306         u64 *mapping = tx_req->mapping;
 307         int i;
 308         int off;
 309
 310         if (skb_headlen(skb)) {
 311                 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
 312                                                DMA_TO_DEVICE);
 313                 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
 314                         return -EIO;
 315
 316                 off = 1;
 317         } else
 318                 off = 0;
 319
 320         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 321                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 322                 mapping[i + off] = ib_dma_map_page(ca, frag->page,
 323                                                  frag->page_offset, frag->size,
 324                                                  DMA_TO_DEVICE);
 325                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
 326                         goto partial_error;
 327         }
 328         return 0;
 329
 330 partial_error:
 331         for (; i > 0; --i) {
 332                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 333                 ib_dma_unmap_page(ca, mapping[i - !off], frag->size, DMA_TO_DEVICE);
 334         }
 335
 336         if (off)
 337                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 338
 339         return -EIO;
 340 }
 341
 342 static void ipoib_dma_unmap_tx(struct ib_device *ca,
 343                                struct ipoib_tx_buf *tx_req)
 344 {
 345         struct sk_buff *skb = tx_req->skb;
 346         u64 *mapping = tx_req->mapping;
 347         int i;
 348         int off;
 349
 350         if (skb_headlen(skb)) {
 351                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 352                 off = 1;
 353         } else
 354                 off = 0;
 355
 356         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 357                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 358                 ib_dma_unmap_page(ca, mapping[i + off], frag->size,
 359                                   DMA_TO_DEVICE);
 360         }
 361 }
 362
 363 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 364 {
 365         struct ipoib_dev_priv *priv = netdev_priv(dev);
 366         unsigned int wr_id = wc->wr_id;
 367         struct ipoib_tx_buf *tx_req;
 368
 369         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 370                        wr_id, wc->status);
 371
 372         if (unlikely(wr_id >= ipoib_sendq_size)) {
 373                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 374                            wr_id, ipoib_sendq_size);
 375                 return;
 376         }
 377
 378         tx_req = &priv->tx_ring[wr_id];
 379
 380         ipoib_dma_unmap_tx(priv->ca, tx_req);
 381
 382         ++dev->stats.tx_packets;
 383         dev->stats.tx_bytes += tx_req->skb->len;
 384
 385         dev_kfree_skb_any(tx_req->skb);
 386
 387         ++priv->tx_tail;
 388         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 389             netif_queue_stopped(dev) &&
 390             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 391                 netif_wake_queue(dev);
 392
 393         if (wc->status != IB_WC_SUCCESS &&
 394             wc->status != IB_WC_WR_FLUSH_ERR)
 395                 ipoib_warn(priv, "failed send event "
 396                            "(status=%d, wrid=%d vend_err %x)\n",
 397                            wc->status, wr_id, wc->vendor_err);
 398 }
 399
 400 static int poll_tx(struct ipoib_dev_priv *priv)
 401 {
 402         int n, i;
 403
 404         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
 405         for (i = 0; i < n; ++i)
 406                 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
 407
 408         return n == MAX_SEND_CQE;
 409 }
 410
 411 int ipoib_poll(struct napi_struct *napi, int budget)
 412 {
 413         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
 414         struct net_device *dev = priv->dev;
 415         int done;
 416         int t;
 417         int n, i;
 418
 419         done  = 0;
 420
 421 poll_more:
 422         while (done < budget) {
 423                 int max = (budget - done);
 424
 425                 t = min(IPOIB_NUM_WC, max);
 426                 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
 427
 428                 for (i = 0; i < n; i++) {
 429                         struct ib_wc *wc = priv->ibwc + i;
 430
 431                         if (wc->wr_id & IPOIB_OP_RECV) {
 432                                 ++done;
 433                                 if (wc->wr_id & IPOIB_OP_CM)
 434                                         ipoib_cm_handle_rx_wc(dev, wc);
 435                                 else
 436                                         ipoib_ib_handle_rx_wc(dev, wc);
 437                         } else
 438                                 ipoib_cm_handle_tx_wc(priv->dev, wc);
 439                 }
 440
 441                 if (n != t)
 442                         break;
 443         }
 444
 445         if (done < budget) {
 446                 if (dev->features & NETIF_F_LRO)
 447                         lro_flush_all(&priv->lro.lro_mgr);
 448
 449                 netif_rx_complete(dev, napi);
 450                 if (unlikely(ib_req_notify_cq(priv->recv_cq,
 451                                               IB_CQ_NEXT_COMP |
 452                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 453                     netif_rx_reschedule(dev, napi))
 454                         goto poll_more;
 455         }
 456
 457         return done;
 458 }
 459
 460 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 461 {
 462         struct net_device *dev = dev_ptr;
 463         struct ipoib_dev_priv *priv = netdev_priv(dev);
 464
 465         netif_rx_schedule(dev, &priv->napi);
 466 }
 467
 468 static void drain_tx_cq(struct net_device *dev)
 469 {
 470         struct ipoib_dev_priv *priv = netdev_priv(dev);
 471
 472         netif_tx_lock(dev);
 473         while (poll_tx(priv))
 474                 ; /* nothing */
 475
 476         if (netif_queue_stopped(dev))
 477                 mod_timer(&priv->poll_timer, jiffies + 1);
 478
 479         netif_tx_unlock(dev);
 480 }
 481
 482 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
 483 {
 484         struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
 485
 486         mod_timer(&priv->poll_timer, jiffies);
 487 }
 488
 489 static inline int post_send(struct ipoib_dev_priv *priv,
 490                             unsigned int wr_id,
 491                             struct ib_ah *address, u32 qpn,
 492                             struct ipoib_tx_buf *tx_req,
 493                             void *head, int hlen)
 494 {
 495         struct ib_send_wr *bad_wr;
 496         int i, off;
 497         struct sk_buff *skb = tx_req->skb;
 498         skb_frag_t *frags = skb_shinfo(skb)->frags;
 499         int nr_frags = skb_shinfo(skb)->nr_frags;
 500         u64 *mapping = tx_req->mapping;
 501
 502         if (skb_headlen(skb)) {
 503                 priv->tx_sge[0].addr         = mapping[0];
 504                 priv->tx_sge[0].length       = skb_headlen(skb);
 505                 off = 1;
 506         } else
 507                 off = 0;
 508
 509         for (i = 0; i < nr_frags; ++i) {
 510                 priv->tx_sge[i + off].addr = mapping[i + off];
 511                 priv->tx_sge[i + off].length = frags[i].size;
 512         }
 513         priv->tx_wr.num_sge          = nr_frags + off;
 514         priv->tx_wr.wr_id            = wr_id;
 515         priv->tx_wr.wr.ud.remote_qpn = qpn;
 516         priv->tx_wr.wr.ud.ah         = address;
 517
 518         if (head) {
 519                 priv->tx_wr.wr.ud.mss    = skb_shinfo(skb)->gso_size;
 520                 priv->tx_wr.wr.ud.header = head;
 521                 priv->tx_wr.wr.ud.hlen   = hlen;
 522                 priv->tx_wr.opcode       = IB_WR_LSO;
 523         } else
 524                 priv->tx_wr.opcode       = IB_WR_SEND;
 525
 526         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 527 }
 528
 529 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 530                 struct ipoib_ah *address, u32 qpn)
 531 {
 532         struct ipoib_dev_priv *priv = netdev_priv(dev);
 533         struct ipoib_tx_buf *tx_req;
 534         int hlen;
 535         void *phead;
 536
 537         if (skb_is_gso(skb)) {
 538                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
 539                 phead = skb->data;
 540                 if (unlikely(!skb_pull(skb, hlen))) {
 541                         ipoib_warn(priv, "linear data too small\n");
 542                         ++dev->stats.tx_dropped;
 543                         ++dev->stats.tx_errors;
 544                         dev_kfree_skb_any(skb);
 545                         return;
 546                 }
 547         } else {
 548                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 549                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 550                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 551                         ++dev->stats.tx_dropped;
 552                         ++dev->stats.tx_errors;
 553                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 554                         return;
 555                 }
 556                 phead = NULL;
 557                 hlen  = 0;
 558         }
 559
 560         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 561                        skb->len, address, qpn);
 562
 563         /*
 564          * We put the skb into the tx_ring _before_ we call post_send()
 565          * because it's entirely possible that the completion handler will
 566          * run before we execute anything after the post_send().  That
 567          * means we have to make sure everything is properly recorded and
 568          * our state is consistent before we call post_send().
 569          */
 570         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 571         tx_req->skb = skb;
 572         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 573                 ++dev->stats.tx_errors;
 574                 dev_kfree_skb_any(skb);
 575                 return;
 576         }
 577
 578         if (skb->ip_summed == CHECKSUM_PARTIAL)
 579                 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
 580         else
 581                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
 582
 583         if (++priv->tx_outstanding == ipoib_sendq_size) {
 584                 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 585                 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
 586                         ipoib_warn(priv, "request notify on send CQ failed\n");
 587                 netif_stop_queue(dev);
 588         }
 589
 590         if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 591                                address->ah, qpn, tx_req, phead, hlen))) {
 592                 ipoib_warn(priv, "post_send failed\n");
 593                 ++dev->stats.tx_errors;
 594                 --priv->tx_outstanding;
 595                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 596                 dev_kfree_skb_any(skb);
 597                 if (netif_queue_stopped(dev))
 598                         netif_wake_queue(dev);
 599         } else {
 600                 dev->trans_start = jiffies;
 601
 602                 address->last_send = priv->tx_head;
 603                 ++priv->tx_head;
 604                 skb_orphan(skb);
 605
 606         }
 607
 608         if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
 609                 while (poll_tx(priv))
 610                         ; /* nothing */
 611 }
 612
 613 static void __ipoib_reap_ah(struct net_device *dev)
 614 {
 615         struct ipoib_dev_priv *priv = netdev_priv(dev);
 616         struct ipoib_ah *ah, *tah;
 617         LIST_HEAD(remove_list);
 618         unsigned long flags;
 619
 620         netif_tx_lock_bh(dev);
 621         spin_lock_irqsave(&priv->lock, flags);
 622
 623         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 624                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 625                         list_del(&ah->list);
 626                         ib_destroy_ah(ah->ah);
 627                         kfree(ah);
 628                 }
 629
 630         spin_unlock_irqrestore(&priv->lock, flags);
 631         netif_tx_unlock_bh(dev);
 632 }
 633
 634 void ipoib_reap_ah(struct work_struct *work)
 635 {
 636         struct ipoib_dev_priv *priv =
 637                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 638         struct net_device *dev = priv->dev;
 639
 640         __ipoib_reap_ah(dev);
 641
 642         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 643                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 644                                    round_jiffies_relative(HZ));
 645 }
 646
 647 static void ipoib_ib_tx_timer_func(unsigned long ctx)
 648 {
 649         drain_tx_cq((struct net_device *)ctx);
 650 }
 651
 652 int ipoib_ib_dev_open(struct net_device *dev)
 653 {
 654         struct ipoib_dev_priv *priv = netdev_priv(dev);
 655         int ret;
 656
 657         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
 658                 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
 659                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 660                 return -1;
 661         }
 662         set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 663
 664         ret = ipoib_init_qp(dev);
 665         if (ret) {
 666                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 667                 return -1;
 668         }
 669
 670         ret = ipoib_ib_post_receives(dev);
 671         if (ret) {
 672                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 673                 ipoib_ib_dev_stop(dev, 1);
 674                 return -1;
 675         }
 676
 677         ret = ipoib_cm_dev_open(dev);
 678         if (ret) {
 679                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 680                 ipoib_ib_dev_stop(dev, 1);
 681                 return -1;
 682         }
 683
 684         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 685         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 686                            round_jiffies_relative(HZ));
 687
 688         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 689
 690         return 0;
 691 }
 692
 693 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
 694 {
 695         struct ipoib_dev_priv *priv = netdev_priv(dev);
 696         u16 pkey_index = 0;
 697
 698         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
 699                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 700         else
 701                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 702 }
 703
 704 int ipoib_ib_dev_up(struct net_device *dev)
 705 {
 706         struct ipoib_dev_priv *priv = netdev_priv(dev);
 707
 708         ipoib_pkey_dev_check_presence(dev);
 709
 710         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 711                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 712                 return 0;
 713         }
 714
 715         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 716
 717         return ipoib_mcast_start_thread(dev);
 718 }
 719
 720 int ipoib_ib_dev_down(struct net_device *dev, int flush)
 721 {
 722         struct ipoib_dev_priv *priv = netdev_priv(dev);
 723
 724         ipoib_dbg(priv, "downing ib_dev\n");
 725
 726         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 727         netif_carrier_off(dev);
 728
 729         /* Shutdown the P_Key thread if still active */
 730         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 731                 mutex_lock(&pkey_mutex);
 732                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
 733                 cancel_delayed_work(&priv->pkey_poll_task);
 734                 mutex_unlock(&pkey_mutex);
 735                 if (flush)
 736                         flush_workqueue(ipoib_workqueue);
 737         }
 738
 739         ipoib_mcast_stop_thread(dev, flush);
 740         ipoib_mcast_dev_flush(dev);
 741
 742         ipoib_flush_paths(dev);
 743
 744         return 0;
 745 }
 746
 747 static int recvs_pending(struct net_device *dev)
 748 {
 749         struct ipoib_dev_priv *priv = netdev_priv(dev);
 750         int pending = 0;
 751         int i;
 752
 753         for (i = 0; i < ipoib_recvq_size; ++i)
 754                 if (priv->rx_ring[i].skb)
 755                         ++pending;
 756
 757         return pending;
 758 }
 759
 760 void ipoib_drain_cq(struct net_device *dev)
 761 {
 762         struct ipoib_dev_priv *priv = netdev_priv(dev);
 763         int i, n;
 764
 765         /*
 766          * We call completion handling routines that expect to be
 767          * called from the BH-disabled NAPI poll context, so disable
 768          * BHs here too.
 769          */
 770         local_bh_disable();
 771
 772         do {
 773                 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
 774                 for (i = 0; i < n; ++i) {
 775                         /*
 776                          * Convert any successful completions to flush
 777                          * errors to avoid passing packets up the
 778                          * stack after bringing the device down.
 779                          */
 780                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 781                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 782
 783                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 784                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 785                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 786                                 else
 787                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 788                         } else
 789                                 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
 790                 }
 791         } while (n == IPOIB_NUM_WC);
 792
 793         while (poll_tx(priv))
 794                 ; /* nothing */
 795
 796         local_bh_enable();
 797 }
 798
 799 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
 800 {
 801         struct ipoib_dev_priv *priv = netdev_priv(dev);
 802         struct ib_qp_attr qp_attr;
 803         unsigned long begin;
 804         struct ipoib_tx_buf *tx_req;
 805         int i;
 806
 807         clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 808
 809         ipoib_cm_dev_stop(dev);
 810
 811         /*
 812          * Move our QP to the error state and then reinitialize in
 813          * when all work requests have completed or have been flushed.
 814          */
 815         qp_attr.qp_state = IB_QPS_ERR;
 816         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 817                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 818
 819         /* Wait for all sends and receives to complete */
 820         begin = jiffies;
 821
 822         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 823                 if (time_after(jiffies, begin + 5 * HZ)) {
 824                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 825                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 826
 827                         /*
 828                          * assume the HW is wedged and just free up
 829                          * all our pending work requests.
 830                          */
 831                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 832                                 tx_req = &priv->tx_ring[priv->tx_tail &
 833                                                         (ipoib_sendq_size - 1)];
 834                                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 835                                 dev_kfree_skb_any(tx_req->skb);
 836                                 ++priv->tx_tail;
 837                                 --priv->tx_outstanding;
 838                         }
 839
 840                         for (i = 0; i < ipoib_recvq_size; ++i) {
 841                                 struct ipoib_rx_buf *rx_req;
 842
 843                                 rx_req = &priv->rx_ring[i];
 844                                 if (!rx_req->skb)
 845                                         continue;
 846                                 ipoib_ud_dma_unmap_rx(priv,
 847                                                       priv->rx_ring[i].mapping);
 848                                 dev_kfree_skb_any(rx_req->skb);
 849                                 rx_req->skb = NULL;
 850                         }
 851
 852                         goto timeout;
 853                 }
 854
 855                 ipoib_drain_cq(dev);
 856
 857                 msleep(1);
 858         }
 859
 860         ipoib_dbg(priv, "All sends and receives done.\n");
 861
 862 timeout:
 863         del_timer_sync(&priv->poll_timer);
 864         qp_attr.qp_state = IB_QPS_RESET;
 865         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 866                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 867
 868         /* Wait for all AHs to be reaped */
 869         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 870         cancel_delayed_work(&priv->ah_reap_task);
 871         if (flush)
 872                 flush_workqueue(ipoib_workqueue);
 873
 874         begin = jiffies;
 875
 876         while (!list_empty(&priv->dead_ahs)) {
 877                 __ipoib_reap_ah(dev);
 878
 879                 if (time_after(jiffies, begin + HZ)) {
 880                         ipoib_warn(priv, "timing out; will leak address handles\n");
 881                         break;
 882                 }
 883
 884                 msleep(1);
 885         }
 886
 887         ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
 888
 889         return 0;
 890 }
 891
 892 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 893 {
 894         struct ipoib_dev_priv *priv = netdev_priv(dev);
 895
 896         priv->ca = ca;
 897         priv->port = port;
 898         priv->qp = NULL;
 899
 900         if (ipoib_transport_dev_init(dev, ca)) {
 901                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 902                 return -ENODEV;
 903         }
 904
 905         setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
 906                     (unsigned long) dev);
 907
 908         if (dev->flags & IFF_UP) {
 909                 if (ipoib_ib_dev_open(dev)) {
 910                         ipoib_transport_dev_cleanup(dev);
 911                         return -ENODEV;
 912                 }
 913         }
 914
 915         return 0;
 916 }
 917
 918 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
 919                                 enum ipoib_flush_level level)
 920 {
 921         struct ipoib_dev_priv *cpriv;
 922         struct net_device *dev = priv->dev;
 923         u16 new_index;
 924
 925         mutex_lock(&priv->vlan_mutex);
 926
 927         /*
 928          * Flush any child interfaces too -- they might be up even if
 929          * the parent is down.
 930          */
 931         list_for_each_entry(cpriv, &priv->child_intfs, list)
 932                 __ipoib_ib_dev_flush(cpriv, level);
 933
 934         mutex_unlock(&priv->vlan_mutex);
 935
 936         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
 937                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
 938                 return;
 939         }
 940
 941         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 942                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
 943                 return;
 944         }
 945
 946         if (level == IPOIB_FLUSH_HEAVY) {
 947                 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
 948                         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 949                         ipoib_ib_dev_down(dev, 0);
 950                         ipoib_ib_dev_stop(dev, 0);
 951                         if (ipoib_pkey_dev_delay_open(dev))
 952                                 return;
 953                 }
 954
 955                 /* restart QP only if P_Key index is changed */
 956                 if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
 957                     new_index == priv->pkey_index) {
 958                         ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
 959                         return;
 960                 }
 961                 priv->pkey_index = new_index;
 962         }
 963
 964         if (level == IPOIB_FLUSH_LIGHT) {
 965                 ipoib_mark_paths_invalid(dev);
 966                 ipoib_mcast_dev_flush(dev);
 967         }
 968
 969         if (level >= IPOIB_FLUSH_NORMAL)
 970                 ipoib_ib_dev_down(dev, 0);
 971
 972         if (level == IPOIB_FLUSH_HEAVY) {
 973                 ipoib_ib_dev_stop(dev, 0);
 974                 ipoib_ib_dev_open(dev);
 975         }
 976
 977         /*
 978          * The device could have been brought down between the start and when
 979          * we get here, don't bring it back up if it's not configured up
 980          */
 981         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 982                 if (level >= IPOIB_FLUSH_NORMAL)
 983                         ipoib_ib_dev_up(dev);
 984                 ipoib_mcast_restart_task(&priv->restart_task);
 985         }
 986 }
 987
 988 void ipoib_ib_dev_flush_light(struct work_struct *work)
 989 {
 990         struct ipoib_dev_priv *priv =
 991                 container_of(work, struct ipoib_dev_priv, flush_light);
 992
 993         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
 994 }
 995
 996 void ipoib_ib_dev_flush_normal(struct work_struct *work)
 997 {
 998         struct ipoib_dev_priv *priv =
 999                 container_of(work, struct ipoib_dev_priv, flush_normal);
1000
1001         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1002 }
1003
1004 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1005 {
1006         struct ipoib_dev_priv *priv =
1007                 container_of(work, struct ipoib_dev_priv, flush_heavy);
1008
1009         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1010 }
1011
1012 void ipoib_ib_dev_cleanup(struct net_device *dev)
1013 {
1014         struct ipoib_dev_priv *priv = netdev_priv(dev);
1015
1016         ipoib_dbg(priv, "cleaning up ib_dev\n");
1017
1018         ipoib_mcast_stop_thread(dev, 1);
1019         ipoib_mcast_dev_flush(dev);
1020
1021         ipoib_transport_dev_cleanup(dev);
1022 }
1023
1024 /*
1025  * Delayed P_Key Assigment Interim Support
1026  *
1027  * The following is initial implementation of delayed P_Key assigment
1028  * mechanism. It is using the same approach implemented for the multicast
1029  * group join. The single goal of this implementation is to quickly address
1030  * Bug #2507. This implementation will probably be removed when the P_Key
1031  * change async notification is available.
1032  */
1033
1034 void ipoib_pkey_poll(struct work_struct *work)
1035 {
1036         struct ipoib_dev_priv *priv =
1037                 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1038         struct net_device *dev = priv->dev;
1039
1040         ipoib_pkey_dev_check_presence(dev);
1041
1042         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1043                 ipoib_open(dev);
1044         else {
1045                 mutex_lock(&pkey_mutex);
1046                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1047                         queue_delayed_work(ipoib_workqueue,
1048                                            &priv->pkey_poll_task,
1049                                            HZ);
1050                 mutex_unlock(&pkey_mutex);
1051         }
1052 }
1053
1054 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1055 {
1056         struct ipoib_dev_priv *priv = netdev_priv(dev);
1057
1058         /* Look for the interface pkey value in the IB Port P_Key table and */
1059         /* set the interface pkey assigment flag                            */
1060         ipoib_pkey_dev_check_presence(dev);
1061
1062         /* P_Key value not assigned yet - start polling */
1063         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1064                 mutex_lock(&pkey_mutex);
1065                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1066                 queue_delayed_work(ipoib_workqueue,
1067                                    &priv->pkey_poll_task,
1068                                    HZ);
1069                 mutex_unlock(&pkey_mutex);
1070                 return 1;
1071         }
1072
1073         return 0;
1074 }