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