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 #include <linux/slab.h>
  39
  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                 skb_frag_size_set(frag, 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, 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         union ib_gid *dgid;
 227
 228         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
 229                        wr_id, wc->status);
 230
 231         if (unlikely(wr_id >= ipoib_recvq_size)) {
 232                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
 233                            wr_id, ipoib_recvq_size);
 234                 return;
 235         }
 236
 237         skb  = priv->rx_ring[wr_id].skb;
 238
 239         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 240                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 241                         ipoib_warn(priv, "failed recv event "
 242                                    "(status=%d, wrid=%d vend_err %x)\n",
 243                                    wc->status, wr_id, wc->vendor_err);
 244                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
 245                 dev_kfree_skb_any(skb);
 246                 priv->rx_ring[wr_id].skb = NULL;
 247                 return;
 248         }
 249
 250         /*
 251          * Drop packets that this interface sent, ie multicast packets
 252          * that the HCA has replicated.
 253          */
 254         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
 255                 goto repost;
 256
 257         memcpy(mapping, priv->rx_ring[wr_id].mapping,
 258                IPOIB_UD_RX_SG * sizeof *mapping);
 259
 260         /*
 261          * If we can't allocate a new RX buffer, dump
 262          * this packet and reuse the old buffer.
 263          */
 264         if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
 265                 ++dev->stats.rx_dropped;
 266                 goto repost;
 267         }
 268
 269         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 270                        wc->byte_len, wc->slid);
 271
 272         ipoib_ud_dma_unmap_rx(priv, mapping);
 273         ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
 274
 275         /* First byte of dgid signals multicast when 0xff */
 276         dgid = &((struct ib_grh *)skb->data)->dgid;
 277
 278         if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
 279                 skb->pkt_type = PACKET_HOST;
 280         else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
 281                 skb->pkt_type = PACKET_BROADCAST;
 282         else
 283                 skb->pkt_type = PACKET_MULTICAST;
 284
 285         skb_pull(skb, IB_GRH_BYTES);
 286
 287         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 288         skb_reset_mac_header(skb);
 289         skb_pull(skb, IPOIB_ENCAP_LEN);
 290
 291         ++dev->stats.rx_packets;
 292         dev->stats.rx_bytes += skb->len;
 293
 294         skb->dev = dev;
 295         if ((dev->features & NETIF_F_RXCSUM) && likely(wc->csum_ok))
 296                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 297
 298         napi_gro_receive(&priv->napi, skb);
 299
 300 repost:
 301         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 302                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 303                            "for buf %d\n", wr_id);
 304 }
 305
 306 static int ipoib_dma_map_tx(struct ib_device *ca,
 307                             struct ipoib_tx_buf *tx_req)
 308 {
 309         struct sk_buff *skb = tx_req->skb;
 310         u64 *mapping = tx_req->mapping;
 311         int i;
 312         int off;
 313
 314         if (skb_headlen(skb)) {
 315                 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
 316                                                DMA_TO_DEVICE);
 317                 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
 318                         return -EIO;
 319
 320                 off = 1;
 321         } else
 322                 off = 0;
 323
 324         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 325                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 326                 mapping[i + off] = ib_dma_map_page(ca,
 327                                                  skb_frag_page(frag),
 328                                                  frag->page_offset, skb_frag_size(frag),
 329                                                  DMA_TO_DEVICE);
 330                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
 331                         goto partial_error;
 332         }
 333         return 0;
 334
 335 partial_error:
 336         for (; i > 0; --i) {
 337                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 338
 339                 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
 340         }
 341
 342         if (off)
 343                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 344
 345         return -EIO;
 346 }
 347
 348 static void ipoib_dma_unmap_tx(struct ib_device *ca,
 349                                struct ipoib_tx_buf *tx_req)
 350 {
 351         struct sk_buff *skb = tx_req->skb;
 352         u64 *mapping = tx_req->mapping;
 353         int i;
 354         int off;
 355
 356         if (skb_headlen(skb)) {
 357                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 358                 off = 1;
 359         } else
 360                 off = 0;
 361
 362         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 363                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 364
 365                 ib_dma_unmap_page(ca, mapping[i + off], skb_frag_size(frag),
 366                                   DMA_TO_DEVICE);
 367         }
 368 }
 369
 370 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 371 {
 372         struct ipoib_dev_priv *priv = netdev_priv(dev);
 373         unsigned int wr_id = wc->wr_id;
 374         struct ipoib_tx_buf *tx_req;
 375
 376         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 377                        wr_id, wc->status);
 378
 379         if (unlikely(wr_id >= ipoib_sendq_size)) {
 380                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 381                            wr_id, ipoib_sendq_size);
 382                 return;
 383         }
 384
 385         tx_req = &priv->tx_ring[wr_id];
 386
 387         ipoib_dma_unmap_tx(priv->ca, tx_req);
 388
 389         ++dev->stats.tx_packets;
 390         dev->stats.tx_bytes += tx_req->skb->len;
 391
 392         dev_kfree_skb_any(tx_req->skb);
 393
 394         ++priv->tx_tail;
 395         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 396             netif_queue_stopped(dev) &&
 397             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 398                 netif_wake_queue(dev);
 399
 400         if (wc->status != IB_WC_SUCCESS &&
 401             wc->status != IB_WC_WR_FLUSH_ERR)
 402                 ipoib_warn(priv, "failed send event "
 403                            "(status=%d, wrid=%d vend_err %x)\n",
 404                            wc->status, wr_id, wc->vendor_err);
 405 }
 406
 407 static int poll_tx(struct ipoib_dev_priv *priv)
 408 {
 409         int n, i;
 410
 411         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
 412         for (i = 0; i < n; ++i)
 413                 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
 414
 415         return n == MAX_SEND_CQE;
 416 }
 417
 418 int ipoib_poll(struct napi_struct *napi, int budget)
 419 {
 420         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
 421         struct net_device *dev = priv->dev;
 422         int done;
 423         int t;
 424         int n, i;
 425
 426         done  = 0;
 427
 428 poll_more:
 429         while (done < budget) {
 430                 int max = (budget - done);
 431
 432                 t = min(IPOIB_NUM_WC, max);
 433                 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
 434
 435                 for (i = 0; i < n; i++) {
 436                         struct ib_wc *wc = priv->ibwc + i;
 437
 438                         if (wc->wr_id & IPOIB_OP_RECV) {
 439                                 ++done;
 440                                 if (wc->wr_id & IPOIB_OP_CM)
 441                                         ipoib_cm_handle_rx_wc(dev, wc);
 442                                 else
 443                                         ipoib_ib_handle_rx_wc(dev, wc);
 444                         } else
 445                                 ipoib_cm_handle_tx_wc(priv->dev, wc);
 446                 }
 447
 448                 if (n != t)
 449                         break;
 450         }
 451
 452         if (done < budget) {
 453                 napi_complete(napi);
 454                 if (unlikely(ib_req_notify_cq(priv->recv_cq,
 455                                               IB_CQ_NEXT_COMP |
 456                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 457                     napi_reschedule(napi))
 458                         goto poll_more;
 459         }
 460
 461         return done;
 462 }
 463
 464 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 465 {
 466         struct net_device *dev = dev_ptr;
 467         struct ipoib_dev_priv *priv = netdev_priv(dev);
 468
 469         napi_schedule(&priv->napi);
 470 }
 471
 472 static void drain_tx_cq(struct net_device *dev)
 473 {
 474         struct ipoib_dev_priv *priv = netdev_priv(dev);
 475
 476         netif_tx_lock(dev);
 477         while (poll_tx(priv))
 478                 ; /* nothing */
 479
 480         if (netif_queue_stopped(dev))
 481                 mod_timer(&priv->poll_timer, jiffies + 1);
 482
 483         netif_tx_unlock(dev);
 484 }
 485
 486 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
 487 {
 488         struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
 489
 490         mod_timer(&priv->poll_timer, jiffies);
 491 }
 492
 493 static inline int post_send(struct ipoib_dev_priv *priv,
 494                             unsigned int wr_id,
 495                             struct ib_ah *address, u32 qpn,
 496                             struct ipoib_tx_buf *tx_req,
 497                             void *head, int hlen)
 498 {
 499         struct ib_send_wr *bad_wr;
 500         int i, off;
 501         struct sk_buff *skb = tx_req->skb;
 502         skb_frag_t *frags = skb_shinfo(skb)->frags;
 503         int nr_frags = skb_shinfo(skb)->nr_frags;
 504         u64 *mapping = tx_req->mapping;
 505
 506         if (skb_headlen(skb)) {
 507                 priv->tx_sge[0].addr         = mapping[0];
 508                 priv->tx_sge[0].length       = skb_headlen(skb);
 509                 off = 1;
 510         } else
 511                 off = 0;
 512
 513         for (i = 0; i < nr_frags; ++i) {
 514                 priv->tx_sge[i + off].addr = mapping[i + off];
 515                 priv->tx_sge[i + off].length = skb_frag_size(&frags[i]);
 516         }
 517         priv->tx_wr.num_sge          = nr_frags + off;
 518         priv->tx_wr.wr_id            = wr_id;
 519         priv->tx_wr.wr.ud.remote_qpn = qpn;
 520         priv->tx_wr.wr.ud.ah         = address;
 521
 522         if (head) {
 523                 priv->tx_wr.wr.ud.mss    = skb_shinfo(skb)->gso_size;
 524                 priv->tx_wr.wr.ud.header = head;
 525                 priv->tx_wr.wr.ud.hlen   = hlen;
 526                 priv->tx_wr.opcode       = IB_WR_LSO;
 527         } else
 528                 priv->tx_wr.opcode       = IB_WR_SEND;
 529
 530         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 531 }
 532
 533 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 534                 struct ipoib_ah *address, u32 qpn)
 535 {
 536         struct ipoib_dev_priv *priv = netdev_priv(dev);
 537         struct ipoib_tx_buf *tx_req;
 538         int hlen, rc;
 539         void *phead;
 540
 541         if (skb_is_gso(skb)) {
 542                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
 543                 phead = skb->data;
 544                 if (unlikely(!skb_pull(skb, hlen))) {
 545                         ipoib_warn(priv, "linear data too small\n");
 546                         ++dev->stats.tx_dropped;
 547                         ++dev->stats.tx_errors;
 548                         dev_kfree_skb_any(skb);
 549                         return;
 550                 }
 551         } else {
 552                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 553                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 554                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 555                         ++dev->stats.tx_dropped;
 556                         ++dev->stats.tx_errors;
 557                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 558                         return;
 559                 }
 560                 phead = NULL;
 561                 hlen  = 0;
 562         }
 563
 564         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 565                        skb->len, address, qpn);
 566
 567         /*
 568          * We put the skb into the tx_ring _before_ we call post_send()
 569          * because it's entirely possible that the completion handler will
 570          * run before we execute anything after the post_send().  That
 571          * means we have to make sure everything is properly recorded and
 572          * our state is consistent before we call post_send().
 573          */
 574         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 575         tx_req->skb = skb;
 576         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 577                 ++dev->stats.tx_errors;
 578                 dev_kfree_skb_any(skb);
 579                 return;
 580         }
 581
 582         if (skb->ip_summed == CHECKSUM_PARTIAL)
 583                 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
 584         else
 585                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
 586
 587         if (++priv->tx_outstanding == ipoib_sendq_size) {
 588                 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 589                 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
 590                         ipoib_warn(priv, "request notify on send CQ failed\n");
 591                 netif_stop_queue(dev);
 592         }
 593
 594         rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 595                        address->ah, qpn, tx_req, phead, hlen);
 596         if (unlikely(rc)) {
 597                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 598                 ++dev->stats.tx_errors;
 599                 --priv->tx_outstanding;
 600                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 601                 dev_kfree_skb_any(skb);
 602                 if (netif_queue_stopped(dev))
 603                         netif_wake_queue(dev);
 604         } else {
 605                 dev->trans_start = jiffies;
 606
 607                 address->last_send = priv->tx_head;
 608                 ++priv->tx_head;
 609                 skb_orphan(skb);
 610
 611         }
 612
 613         if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
 614                 while (poll_tx(priv))
 615                         ; /* nothing */
 616 }
 617
 618 static void __ipoib_reap_ah(struct net_device *dev)
 619 {
 620         struct ipoib_dev_priv *priv = netdev_priv(dev);
 621         struct ipoib_ah *ah, *tah;
 622         LIST_HEAD(remove_list);
 623         unsigned long flags;
 624
 625         netif_tx_lock_bh(dev);
 626         spin_lock_irqsave(&priv->lock, flags);
 627
 628         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 629                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 630                         list_del(&ah->list);
 631                         ib_destroy_ah(ah->ah);
 632                         kfree(ah);
 633                 }
 634
 635         spin_unlock_irqrestore(&priv->lock, flags);
 636         netif_tx_unlock_bh(dev);
 637 }
 638
 639 void ipoib_reap_ah(struct work_struct *work)
 640 {
 641         struct ipoib_dev_priv *priv =
 642                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 643         struct net_device *dev = priv->dev;
 644
 645         __ipoib_reap_ah(dev);
 646
 647         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 648                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 649                                    round_jiffies_relative(HZ));
 650 }
 651
 652 static void ipoib_ib_tx_timer_func(unsigned long ctx)
 653 {
 654         drain_tx_cq((struct net_device *)ctx);
 655 }
 656
 657 int ipoib_ib_dev_open(struct net_device *dev)
 658 {
 659         struct ipoib_dev_priv *priv = netdev_priv(dev);
 660         int ret;
 661
 662         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
 663                 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
 664                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 665                 return -1;
 666         }
 667         set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 668
 669         ret = ipoib_init_qp(dev);
 670         if (ret) {
 671                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 672                 return -1;
 673         }
 674
 675         ret = ipoib_ib_post_receives(dev);
 676         if (ret) {
 677                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 678                 ipoib_ib_dev_stop(dev, 1);
 679                 return -1;
 680         }
 681
 682         ret = ipoib_cm_dev_open(dev);
 683         if (ret) {
 684                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 685                 ipoib_ib_dev_stop(dev, 1);
 686                 return -1;
 687         }
 688
 689         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 690         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 691                            round_jiffies_relative(HZ));
 692
 693         if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 694                 napi_enable(&priv->napi);
 695
 696         return 0;
 697 }
 698
 699 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
 700 {
 701         struct ipoib_dev_priv *priv = netdev_priv(dev);
 702         u16 pkey_index = 0;
 703
 704         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
 705                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 706         else
 707                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 708 }
 709
 710 int ipoib_ib_dev_up(struct net_device *dev)
 711 {
 712         struct ipoib_dev_priv *priv = netdev_priv(dev);
 713
 714         ipoib_pkey_dev_check_presence(dev);
 715
 716         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 717                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 718                 return 0;
 719         }
 720
 721         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 722
 723         return ipoib_mcast_start_thread(dev);
 724 }
 725
 726 int ipoib_ib_dev_down(struct net_device *dev, int flush)
 727 {
 728         struct ipoib_dev_priv *priv = netdev_priv(dev);
 729
 730         ipoib_dbg(priv, "downing ib_dev\n");
 731
 732         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 733         netif_carrier_off(dev);
 734
 735         /* Shutdown the P_Key thread if still active */
 736         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 737                 mutex_lock(&pkey_mutex);
 738                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
 739                 cancel_delayed_work(&priv->pkey_poll_task);
 740                 mutex_unlock(&pkey_mutex);
 741                 if (flush)
 742                         flush_workqueue(ipoib_workqueue);
 743         }
 744
 745         ipoib_mcast_stop_thread(dev, flush);
 746         ipoib_mcast_dev_flush(dev);
 747
 748         ipoib_flush_paths(dev);
 749
 750         return 0;
 751 }
 752
 753 static int recvs_pending(struct net_device *dev)
 754 {
 755         struct ipoib_dev_priv *priv = netdev_priv(dev);
 756         int pending = 0;
 757         int i;
 758
 759         for (i = 0; i < ipoib_recvq_size; ++i)
 760                 if (priv->rx_ring[i].skb)
 761                         ++pending;
 762
 763         return pending;
 764 }
 765
 766 void ipoib_drain_cq(struct net_device *dev)
 767 {
 768         struct ipoib_dev_priv *priv = netdev_priv(dev);
 769         int i, n;
 770
 771         /*
 772          * We call completion handling routines that expect to be
 773          * called from the BH-disabled NAPI poll context, so disable
 774          * BHs here too.
 775          */
 776         local_bh_disable();
 777
 778         do {
 779                 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
 780                 for (i = 0; i < n; ++i) {
 781                         /*
 782                          * Convert any successful completions to flush
 783                          * errors to avoid passing packets up the
 784                          * stack after bringing the device down.
 785                          */
 786                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 787                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 788
 789                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 790                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 791                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 792                                 else
 793                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 794                         } else
 795                                 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
 796                 }
 797         } while (n == IPOIB_NUM_WC);
 798
 799         while (poll_tx(priv))
 800                 ; /* nothing */
 801
 802         local_bh_enable();
 803 }
 804
 805 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
 806 {
 807         struct ipoib_dev_priv *priv = netdev_priv(dev);
 808         struct ib_qp_attr qp_attr;
 809         unsigned long begin;
 810         struct ipoib_tx_buf *tx_req;
 811         int i;
 812
 813         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 814                 napi_disable(&priv->napi);
 815
 816         ipoib_cm_dev_stop(dev);
 817
 818         /*
 819          * Move our QP to the error state and then reinitialize in
 820          * when all work requests have completed or have been flushed.
 821          */
 822         qp_attr.qp_state = IB_QPS_ERR;
 823         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 824                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 825
 826         /* Wait for all sends and receives to complete */
 827         begin = jiffies;
 828
 829         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 830                 if (time_after(jiffies, begin + 5 * HZ)) {
 831                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 832                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 833
 834                         /*
 835                          * assume the HW is wedged and just free up
 836                          * all our pending work requests.
 837                          */
 838                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 839                                 tx_req = &priv->tx_ring[priv->tx_tail &
 840                                                         (ipoib_sendq_size - 1)];
 841                                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 842                                 dev_kfree_skb_any(tx_req->skb);
 843                                 ++priv->tx_tail;
 844                                 --priv->tx_outstanding;
 845                         }
 846
 847                         for (i = 0; i < ipoib_recvq_size; ++i) {
 848                                 struct ipoib_rx_buf *rx_req;
 849
 850                                 rx_req = &priv->rx_ring[i];
 851                                 if (!rx_req->skb)
 852                                         continue;
 853                                 ipoib_ud_dma_unmap_rx(priv,
 854                                                       priv->rx_ring[i].mapping);
 855                                 dev_kfree_skb_any(rx_req->skb);
 856                                 rx_req->skb = NULL;
 857                         }
 858
 859                         goto timeout;
 860                 }
 861
 862                 ipoib_drain_cq(dev);
 863
 864                 msleep(1);
 865         }
 866
 867         ipoib_dbg(priv, "All sends and receives done.\n");
 868
 869 timeout:
 870         del_timer_sync(&priv->poll_timer);
 871         qp_attr.qp_state = IB_QPS_RESET;
 872         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 873                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 874
 875         /* Wait for all AHs to be reaped */
 876         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 877         cancel_delayed_work(&priv->ah_reap_task);
 878         if (flush)
 879                 flush_workqueue(ipoib_workqueue);
 880
 881         begin = jiffies;
 882
 883         while (!list_empty(&priv->dead_ahs)) {
 884                 __ipoib_reap_ah(dev);
 885
 886                 if (time_after(jiffies, begin + HZ)) {
 887                         ipoib_warn(priv, "timing out; will leak address handles\n");
 888                         break;
 889                 }
 890
 891                 msleep(1);
 892         }
 893
 894         ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
 895
 896         return 0;
 897 }
 898
 899 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 900 {
 901         struct ipoib_dev_priv *priv = netdev_priv(dev);
 902
 903         priv->ca = ca;
 904         priv->port = port;
 905         priv->qp = NULL;
 906
 907         if (ipoib_transport_dev_init(dev, ca)) {
 908                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 909                 return -ENODEV;
 910         }
 911
 912         setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
 913                     (unsigned long) dev);
 914
 915         if (dev->flags & IFF_UP) {
 916                 if (ipoib_ib_dev_open(dev)) {
 917                         ipoib_transport_dev_cleanup(dev);
 918                         return -ENODEV;
 919                 }
 920         }
 921
 922         return 0;
 923 }
 924
 925 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
 926                                 enum ipoib_flush_level level)
 927 {
 928         struct ipoib_dev_priv *cpriv;
 929         struct net_device *dev = priv->dev;
 930         u16 new_index;
 931
 932         mutex_lock(&priv->vlan_mutex);
 933
 934         /*
 935          * Flush any child interfaces too -- they might be up even if
 936          * the parent is down.
 937          */
 938         list_for_each_entry(cpriv, &priv->child_intfs, list)
 939                 __ipoib_ib_dev_flush(cpriv, level);
 940
 941         mutex_unlock(&priv->vlan_mutex);
 942
 943         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
 944                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
 945                 return;
 946         }
 947
 948         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 949                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
 950                 return;
 951         }
 952
 953         if (level == IPOIB_FLUSH_HEAVY) {
 954                 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
 955                         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 956                         ipoib_ib_dev_down(dev, 0);
 957                         ipoib_ib_dev_stop(dev, 0);
 958                         if (ipoib_pkey_dev_delay_open(dev))
 959                                 return;
 960                 }
 961
 962                 /* restart QP only if P_Key index is changed */
 963                 if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
 964                     new_index == priv->pkey_index) {
 965                         ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
 966                         return;
 967                 }
 968                 priv->pkey_index = new_index;
 969         }
 970
 971         if (level == IPOIB_FLUSH_LIGHT) {
 972                 ipoib_mark_paths_invalid(dev);
 973                 ipoib_mcast_dev_flush(dev);
 974         }
 975
 976         if (level >= IPOIB_FLUSH_NORMAL)
 977                 ipoib_ib_dev_down(dev, 0);
 978
 979         if (level == IPOIB_FLUSH_HEAVY) {
 980                 ipoib_ib_dev_stop(dev, 0);
 981                 ipoib_ib_dev_open(dev);
 982         }
 983
 984         /*
 985          * The device could have been brought down between the start and when
 986          * we get here, don't bring it back up if it's not configured up
 987          */
 988         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 989                 if (level >= IPOIB_FLUSH_NORMAL)
 990                         ipoib_ib_dev_up(dev);
 991                 ipoib_mcast_restart_task(&priv->restart_task);
 992         }
 993 }
 994
 995 void ipoib_ib_dev_flush_light(struct work_struct *work)
 996 {
 997         struct ipoib_dev_priv *priv =
 998                 container_of(work, struct ipoib_dev_priv, flush_light);
 999
1000         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
1001 }
1002
1003 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1004 {
1005         struct ipoib_dev_priv *priv =
1006                 container_of(work, struct ipoib_dev_priv, flush_normal);
1007
1008         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1009 }
1010
1011 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1012 {
1013         struct ipoib_dev_priv *priv =
1014                 container_of(work, struct ipoib_dev_priv, flush_heavy);
1015
1016         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1017 }
1018
1019 void ipoib_ib_dev_cleanup(struct net_device *dev)
1020 {
1021         struct ipoib_dev_priv *priv = netdev_priv(dev);
1022
1023         ipoib_dbg(priv, "cleaning up ib_dev\n");
1024
1025         ipoib_mcast_stop_thread(dev, 1);
1026         ipoib_mcast_dev_flush(dev);
1027
1028         ipoib_transport_dev_cleanup(dev);
1029 }
1030
1031 /*
1032  * Delayed P_Key Assigment Interim Support
1033  *
1034  * The following is initial implementation of delayed P_Key assigment
1035  * mechanism. It is using the same approach implemented for the multicast
1036  * group join. The single goal of this implementation is to quickly address
1037  * Bug #2507. This implementation will probably be removed when the P_Key
1038  * change async notification is available.
1039  */
1040
1041 void ipoib_pkey_poll(struct work_struct *work)
1042 {
1043         struct ipoib_dev_priv *priv =
1044                 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1045         struct net_device *dev = priv->dev;
1046
1047         ipoib_pkey_dev_check_presence(dev);
1048
1049         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1050                 ipoib_open(dev);
1051         else {
1052                 mutex_lock(&pkey_mutex);
1053                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1054                         queue_delayed_work(ipoib_workqueue,
1055                                            &priv->pkey_poll_task,
1056                                            HZ);
1057                 mutex_unlock(&pkey_mutex);
1058         }
1059 }
1060
1061 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1062 {
1063         struct ipoib_dev_priv *priv = netdev_priv(dev);
1064
1065         /* Look for the interface pkey value in the IB Port P_Key table and */
1066         /* set the interface pkey assigment flag                            */
1067         ipoib_pkey_dev_check_presence(dev);
1068
1069         /* P_Key value not assigned yet - start polling */
1070         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1071                 mutex_lock(&pkey_mutex);
1072                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1073                 queue_delayed_work(ipoib_workqueue,
1074                                    &priv->pkey_poll_task,
1075                                    HZ);
1076                 mutex_unlock(&pkey_mutex);
1077                 return 1;
1078         }
1079
1080         return 0;
1081 }