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  * $Id: ipoib_ib.c 1386 2004-12-27 16:23:17Z roland $
  36  */
  37
  38 #include <linux/delay.h>
  39 #include <linux/dma-mapping.h>
  40
  41 #include <rdma/ib_cache.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 int ipoib_ib_post_receive(struct net_device *dev, int id)
  91 {
  92         struct ipoib_dev_priv *priv = netdev_priv(dev);
  93         struct ib_sge list;
  94         struct ib_recv_wr param;
  95         struct ib_recv_wr *bad_wr;
  96         int ret;
  97
  98         list.addr     = priv->rx_ring[id].mapping;
  99         list.length   = IPOIB_BUF_SIZE;
 100         list.lkey     = priv->mr->lkey;
 101
 102         param.next    = NULL;
 103         param.wr_id   = id | IPOIB_OP_RECV;
 104         param.sg_list = &list;
 105         param.num_sge = 1;
 106
 107         ret = ib_post_recv(priv->qp, &param, &bad_wr);
 108         if (unlikely(ret)) {
 109                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
 110                 ib_dma_unmap_single(priv->ca, priv->rx_ring[id].mapping,
 111                                     IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
 112                 dev_kfree_skb_any(priv->rx_ring[id].skb);
 113                 priv->rx_ring[id].skb = NULL;
 114         }
 115
 116         return ret;
 117 }
 118
 119 static int ipoib_alloc_rx_skb(struct net_device *dev, int id)
 120 {
 121         struct ipoib_dev_priv *priv = netdev_priv(dev);
 122         struct sk_buff *skb;
 123         u64 addr;
 124
 125         skb = dev_alloc_skb(IPOIB_BUF_SIZE + 4);
 126         if (!skb)
 127                 return -ENOMEM;
 128
 129         /*
 130          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
 131          * header.  So we need 4 more bytes to get to 48 and align the
 132          * IP header to a multiple of 16.
 133          */
 134         skb_reserve(skb, 4);
 135
 136         addr = ib_dma_map_single(priv->ca, skb->data, IPOIB_BUF_SIZE,
 137                                  DMA_FROM_DEVICE);
 138         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 139                 dev_kfree_skb_any(skb);
 140                 return -EIO;
 141         }
 142
 143         priv->rx_ring[id].skb     = skb;
 144         priv->rx_ring[id].mapping = addr;
 145
 146         return 0;
 147 }
 148
 149 static int ipoib_ib_post_receives(struct net_device *dev)
 150 {
 151         struct ipoib_dev_priv *priv = netdev_priv(dev);
 152         int i;
 153
 154         for (i = 0; i < ipoib_recvq_size; ++i) {
 155                 if (ipoib_alloc_rx_skb(dev, i)) {
 156                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 157                         return -ENOMEM;
 158                 }
 159                 if (ipoib_ib_post_receive(dev, i)) {
 160                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 161                         return -EIO;
 162                 }
 163         }
 164
 165         return 0;
 166 }
 167
 168 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 169 {
 170         struct ipoib_dev_priv *priv = netdev_priv(dev);
 171         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
 172         struct sk_buff *skb;
 173         u64 addr;
 174
 175         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
 176                        wr_id, wc->status);
 177
 178         if (unlikely(wr_id >= ipoib_recvq_size)) {
 179                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
 180                            wr_id, ipoib_recvq_size);
 181                 return;
 182         }
 183
 184         skb  = priv->rx_ring[wr_id].skb;
 185         addr = priv->rx_ring[wr_id].mapping;
 186
 187         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 188                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 189                         ipoib_warn(priv, "failed recv event "
 190                                    "(status=%d, wrid=%d vend_err %x)\n",
 191                                    wc->status, wr_id, wc->vendor_err);
 192                 ib_dma_unmap_single(priv->ca, addr,
 193                                     IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
 194                 dev_kfree_skb_any(skb);
 195                 priv->rx_ring[wr_id].skb = NULL;
 196                 return;
 197         }
 198
 199         /*
 200          * Drop packets that this interface sent, ie multicast packets
 201          * that the HCA has replicated.
 202          */
 203         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
 204                 goto repost;
 205
 206         /*
 207          * If we can't allocate a new RX buffer, dump
 208          * this packet and reuse the old buffer.
 209          */
 210         if (unlikely(ipoib_alloc_rx_skb(dev, wr_id))) {
 211                 ++dev->stats.rx_dropped;
 212                 goto repost;
 213         }
 214
 215         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 216                        wc->byte_len, wc->slid);
 217
 218         ib_dma_unmap_single(priv->ca, addr, IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
 219
 220         skb_put(skb, wc->byte_len);
 221         skb_pull(skb, IB_GRH_BYTES);
 222
 223         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 224         skb_reset_mac_header(skb);
 225         skb_pull(skb, IPOIB_ENCAP_LEN);
 226
 227         dev->last_rx = jiffies;
 228         ++dev->stats.rx_packets;
 229         dev->stats.rx_bytes += skb->len;
 230
 231         skb->dev = dev;
 232         /* XXX get correct PACKET_ type here */
 233         skb->pkt_type = PACKET_HOST;
 234         netif_receive_skb(skb);
 235
 236 repost:
 237         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 238                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 239                            "for buf %d\n", wr_id);
 240 }
 241
 242 static int ipoib_dma_map_tx(struct ib_device *ca,
 243                             struct ipoib_tx_buf *tx_req)
 244 {
 245         struct sk_buff *skb = tx_req->skb;
 246         u64 *mapping = tx_req->mapping;
 247         int i;
 248
 249         mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
 250                                        DMA_TO_DEVICE);
 251         if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
 252                 return -EIO;
 253
 254         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 255                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 256                 mapping[i + 1] = ib_dma_map_page(ca, frag->page,
 257                                                  frag->page_offset, frag->size,
 258                                                  DMA_TO_DEVICE);
 259                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + 1])))
 260                         goto partial_error;
 261         }
 262         return 0;
 263
 264 partial_error:
 265         ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 266
 267         for (; i > 0; --i) {
 268                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 269                 ib_dma_unmap_page(ca, mapping[i], frag->size, DMA_TO_DEVICE);
 270         }
 271         return -EIO;
 272 }
 273
 274 static void ipoib_dma_unmap_tx(struct ib_device *ca,
 275                                struct ipoib_tx_buf *tx_req)
 276 {
 277         struct sk_buff *skb = tx_req->skb;
 278         u64 *mapping = tx_req->mapping;
 279         int i;
 280
 281         ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 282
 283         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 284                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 285                 ib_dma_unmap_page(ca, mapping[i + 1], frag->size,
 286                                   DMA_TO_DEVICE);
 287         }
 288 }
 289
 290 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 291 {
 292         struct ipoib_dev_priv *priv = netdev_priv(dev);
 293         unsigned int wr_id = wc->wr_id;
 294         struct ipoib_tx_buf *tx_req;
 295         unsigned long flags;
 296
 297         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 298                        wr_id, wc->status);
 299
 300         if (unlikely(wr_id >= ipoib_sendq_size)) {
 301                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 302                            wr_id, ipoib_sendq_size);
 303                 return;
 304         }
 305
 306         tx_req = &priv->tx_ring[wr_id];
 307
 308         ipoib_dma_unmap_tx(priv->ca, tx_req);
 309
 310         ++dev->stats.tx_packets;
 311         dev->stats.tx_bytes += tx_req->skb->len;
 312
 313         dev_kfree_skb_any(tx_req->skb);
 314
 315         spin_lock_irqsave(&priv->tx_lock, flags);
 316         ++priv->tx_tail;
 317         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 318             netif_queue_stopped(dev) &&
 319             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 320                 netif_wake_queue(dev);
 321         spin_unlock_irqrestore(&priv->tx_lock, flags);
 322
 323         if (wc->status != IB_WC_SUCCESS &&
 324             wc->status != IB_WC_WR_FLUSH_ERR)
 325                 ipoib_warn(priv, "failed send event "
 326                            "(status=%d, wrid=%d vend_err %x)\n",
 327                            wc->status, wr_id, wc->vendor_err);
 328 }
 329
 330 int ipoib_poll(struct napi_struct *napi, int budget)
 331 {
 332         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
 333         struct net_device *dev = priv->dev;
 334         int done;
 335         int t;
 336         int n, i;
 337
 338         done  = 0;
 339
 340 poll_more:
 341         while (done < budget) {
 342                 int max = (budget - done);
 343
 344                 t = min(IPOIB_NUM_WC, max);
 345                 n = ib_poll_cq(priv->cq, t, priv->ibwc);
 346
 347                 for (i = 0; i < n; i++) {
 348                         struct ib_wc *wc = priv->ibwc + i;
 349
 350                         if (wc->wr_id & IPOIB_OP_RECV) {
 351                                 ++done;
 352                                 if (wc->wr_id & IPOIB_OP_CM)
 353                                         ipoib_cm_handle_rx_wc(dev, wc);
 354                                 else
 355                                         ipoib_ib_handle_rx_wc(dev, wc);
 356                         } else {
 357                                 if (wc->wr_id & IPOIB_OP_CM)
 358                                         ipoib_cm_handle_tx_wc(dev, wc);
 359                                 else
 360                                         ipoib_ib_handle_tx_wc(dev, wc);
 361                         }
 362                 }
 363
 364                 if (n != t)
 365                         break;
 366         }
 367
 368         if (done < budget) {
 369                 netif_rx_complete(dev, napi);
 370                 if (unlikely(ib_req_notify_cq(priv->cq,
 371                                               IB_CQ_NEXT_COMP |
 372                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 373                     netif_rx_reschedule(dev, napi))
 374                         goto poll_more;
 375         }
 376
 377         return done;
 378 }
 379
 380 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 381 {
 382         struct net_device *dev = dev_ptr;
 383         struct ipoib_dev_priv *priv = netdev_priv(dev);
 384
 385         netif_rx_schedule(dev, &priv->napi);
 386 }
 387
 388 static inline int post_send(struct ipoib_dev_priv *priv,
 389                             unsigned int wr_id,
 390                             struct ib_ah *address, u32 qpn,
 391                             u64 *mapping, int headlen,
 392                             skb_frag_t *frags,
 393                             int nr_frags)
 394 {
 395         struct ib_send_wr *bad_wr;
 396         int i;
 397
 398         priv->tx_sge[0].addr         = mapping[0];
 399         priv->tx_sge[0].length       = headlen;
 400         for (i = 0; i < nr_frags; ++i) {
 401                 priv->tx_sge[i + 1].addr = mapping[i + 1];
 402                 priv->tx_sge[i + 1].length = frags[i].size;
 403         }
 404         priv->tx_wr.num_sge          = nr_frags + 1;
 405         priv->tx_wr.wr_id            = wr_id;
 406         priv->tx_wr.wr.ud.remote_qpn = qpn;
 407         priv->tx_wr.wr.ud.ah         = address;
 408
 409         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 410 }
 411
 412 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 413                 struct ipoib_ah *address, u32 qpn)
 414 {
 415         struct ipoib_dev_priv *priv = netdev_priv(dev);
 416         struct ipoib_tx_buf *tx_req;
 417
 418         if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 419                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 420                            skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 421                 ++dev->stats.tx_dropped;
 422                 ++dev->stats.tx_errors;
 423                 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 424                 return;
 425         }
 426
 427         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 428                        skb->len, address, qpn);
 429
 430         /*
 431          * We put the skb into the tx_ring _before_ we call post_send()
 432          * because it's entirely possible that the completion handler will
 433          * run before we execute anything after the post_send().  That
 434          * means we have to make sure everything is properly recorded and
 435          * our state is consistent before we call post_send().
 436          */
 437         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 438         tx_req->skb = skb;
 439         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 440                 ++dev->stats.tx_errors;
 441                 dev_kfree_skb_any(skb);
 442                 return;
 443         }
 444
 445         if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 446                                address->ah, qpn,
 447                                tx_req->mapping, skb_headlen(skb),
 448                                skb_shinfo(skb)->frags, skb_shinfo(skb)->nr_frags))) {
 449                 ipoib_warn(priv, "post_send failed\n");
 450                 ++dev->stats.tx_errors;
 451                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 452                 dev_kfree_skb_any(skb);
 453         } else {
 454                 dev->trans_start = jiffies;
 455
 456                 address->last_send = priv->tx_head;
 457                 ++priv->tx_head;
 458
 459                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 460                         ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 461                         netif_stop_queue(dev);
 462                 }
 463         }
 464 }
 465
 466 static void __ipoib_reap_ah(struct net_device *dev)
 467 {
 468         struct ipoib_dev_priv *priv = netdev_priv(dev);
 469         struct ipoib_ah *ah, *tah;
 470         LIST_HEAD(remove_list);
 471
 472         spin_lock_irq(&priv->tx_lock);
 473         spin_lock(&priv->lock);
 474         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 475                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 476                         list_del(&ah->list);
 477                         ib_destroy_ah(ah->ah);
 478                         kfree(ah);
 479                 }
 480         spin_unlock(&priv->lock);
 481         spin_unlock_irq(&priv->tx_lock);
 482 }
 483
 484 void ipoib_reap_ah(struct work_struct *work)
 485 {
 486         struct ipoib_dev_priv *priv =
 487                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 488         struct net_device *dev = priv->dev;
 489
 490         __ipoib_reap_ah(dev);
 491
 492         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 493                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 494                                    round_jiffies_relative(HZ));
 495 }
 496
 497 int ipoib_ib_dev_open(struct net_device *dev)
 498 {
 499         struct ipoib_dev_priv *priv = netdev_priv(dev);
 500         int ret;
 501
 502         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
 503                 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
 504                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 505                 return -1;
 506         }
 507         set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 508
 509         ret = ipoib_init_qp(dev);
 510         if (ret) {
 511                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 512                 return -1;
 513         }
 514
 515         ret = ipoib_ib_post_receives(dev);
 516         if (ret) {
 517                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 518                 ipoib_ib_dev_stop(dev, 1);
 519                 return -1;
 520         }
 521
 522         ret = ipoib_cm_dev_open(dev);
 523         if (ret) {
 524                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 525                 ipoib_ib_dev_stop(dev, 1);
 526                 return -1;
 527         }
 528
 529         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 530         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 531                            round_jiffies_relative(HZ));
 532
 533         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 534
 535         return 0;
 536 }
 537
 538 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
 539 {
 540         struct ipoib_dev_priv *priv = netdev_priv(dev);
 541         u16 pkey_index = 0;
 542
 543         if (ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
 544                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 545         else
 546                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 547 }
 548
 549 int ipoib_ib_dev_up(struct net_device *dev)
 550 {
 551         struct ipoib_dev_priv *priv = netdev_priv(dev);
 552
 553         ipoib_pkey_dev_check_presence(dev);
 554
 555         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 556                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 557                 return 0;
 558         }
 559
 560         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 561
 562         return ipoib_mcast_start_thread(dev);
 563 }
 564
 565 int ipoib_ib_dev_down(struct net_device *dev, int flush)
 566 {
 567         struct ipoib_dev_priv *priv = netdev_priv(dev);
 568
 569         ipoib_dbg(priv, "downing ib_dev\n");
 570
 571         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 572         netif_carrier_off(dev);
 573
 574         /* Shutdown the P_Key thread if still active */
 575         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 576                 mutex_lock(&pkey_mutex);
 577                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
 578                 cancel_delayed_work(&priv->pkey_poll_task);
 579                 mutex_unlock(&pkey_mutex);
 580                 if (flush)
 581                         flush_workqueue(ipoib_workqueue);
 582         }
 583
 584         ipoib_mcast_stop_thread(dev, flush);
 585         ipoib_mcast_dev_flush(dev);
 586
 587         ipoib_flush_paths(dev);
 588
 589         return 0;
 590 }
 591
 592 static int recvs_pending(struct net_device *dev)
 593 {
 594         struct ipoib_dev_priv *priv = netdev_priv(dev);
 595         int pending = 0;
 596         int i;
 597
 598         for (i = 0; i < ipoib_recvq_size; ++i)
 599                 if (priv->rx_ring[i].skb)
 600                         ++pending;
 601
 602         return pending;
 603 }
 604
 605 void ipoib_drain_cq(struct net_device *dev)
 606 {
 607         struct ipoib_dev_priv *priv = netdev_priv(dev);
 608         int i, n;
 609         do {
 610                 n = ib_poll_cq(priv->cq, IPOIB_NUM_WC, priv->ibwc);
 611                 for (i = 0; i < n; ++i) {
 612                         /*
 613                          * Convert any successful completions to flush
 614                          * errors to avoid passing packets up the
 615                          * stack after bringing the device down.
 616                          */
 617                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 618                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 619
 620                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 621                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 622                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 623                                 else
 624                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 625                         } else {
 626                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 627                                         ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
 628                                 else
 629                                         ipoib_ib_handle_tx_wc(dev, priv->ibwc + i);
 630                         }
 631                 }
 632         } while (n == IPOIB_NUM_WC);
 633 }
 634
 635 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
 636 {
 637         struct ipoib_dev_priv *priv = netdev_priv(dev);
 638         struct ib_qp_attr qp_attr;
 639         unsigned long begin;
 640         struct ipoib_tx_buf *tx_req;
 641         int i;
 642
 643         clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 644
 645         ipoib_cm_dev_stop(dev);
 646
 647         /*
 648          * Move our QP to the error state and then reinitialize in
 649          * when all work requests have completed or have been flushed.
 650          */
 651         qp_attr.qp_state = IB_QPS_ERR;
 652         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 653                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 654
 655         /* Wait for all sends and receives to complete */
 656         begin = jiffies;
 657
 658         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 659                 if (time_after(jiffies, begin + 5 * HZ)) {
 660                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 661                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 662
 663                         /*
 664                          * assume the HW is wedged and just free up
 665                          * all our pending work requests.
 666                          */
 667                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 668                                 tx_req = &priv->tx_ring[priv->tx_tail &
 669                                                         (ipoib_sendq_size - 1)];
 670                                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 671                                 dev_kfree_skb_any(tx_req->skb);
 672                                 ++priv->tx_tail;
 673                                 --priv->tx_outstanding;
 674                         }
 675
 676                         for (i = 0; i < ipoib_recvq_size; ++i) {
 677                                 struct ipoib_rx_buf *rx_req;
 678
 679                                 rx_req = &priv->rx_ring[i];
 680                                 if (!rx_req->skb)
 681                                         continue;
 682                                 ib_dma_unmap_single(priv->ca,
 683                                                     rx_req->mapping,
 684                                                     IPOIB_BUF_SIZE,
 685                                                     DMA_FROM_DEVICE);
 686                                 dev_kfree_skb_any(rx_req->skb);
 687                                 rx_req->skb = NULL;
 688                         }
 689
 690                         goto timeout;
 691                 }
 692
 693                 ipoib_drain_cq(dev);
 694
 695                 msleep(1);
 696         }
 697
 698         ipoib_dbg(priv, "All sends and receives done.\n");
 699
 700 timeout:
 701         qp_attr.qp_state = IB_QPS_RESET;
 702         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 703                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 704
 705         /* Wait for all AHs to be reaped */
 706         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 707         cancel_delayed_work(&priv->ah_reap_task);
 708         if (flush)
 709                 flush_workqueue(ipoib_workqueue);
 710
 711         begin = jiffies;
 712
 713         while (!list_empty(&priv->dead_ahs)) {
 714                 __ipoib_reap_ah(dev);
 715
 716                 if (time_after(jiffies, begin + HZ)) {
 717                         ipoib_warn(priv, "timing out; will leak address handles\n");
 718                         break;
 719                 }
 720
 721                 msleep(1);
 722         }
 723
 724         ib_req_notify_cq(priv->cq, IB_CQ_NEXT_COMP);
 725
 726         return 0;
 727 }
 728
 729 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 730 {
 731         struct ipoib_dev_priv *priv = netdev_priv(dev);
 732
 733         priv->ca = ca;
 734         priv->port = port;
 735         priv->qp = NULL;
 736
 737         if (ipoib_transport_dev_init(dev, ca)) {
 738                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 739                 return -ENODEV;
 740         }
 741
 742         if (dev->flags & IFF_UP) {
 743                 if (ipoib_ib_dev_open(dev)) {
 744                         ipoib_transport_dev_cleanup(dev);
 745                         return -ENODEV;
 746                 }
 747         }
 748
 749         return 0;
 750 }
 751
 752 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv, int pkey_event)
 753 {
 754         struct ipoib_dev_priv *cpriv;
 755         struct net_device *dev = priv->dev;
 756         u16 new_index;
 757
 758         mutex_lock(&priv->vlan_mutex);
 759
 760         /*
 761          * Flush any child interfaces too -- they might be up even if
 762          * the parent is down.
 763          */
 764         list_for_each_entry(cpriv, &priv->child_intfs, list)
 765                 __ipoib_ib_dev_flush(cpriv, pkey_event);
 766
 767         mutex_unlock(&priv->vlan_mutex);
 768
 769         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
 770                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
 771                 return;
 772         }
 773
 774         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 775                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
 776                 return;
 777         }
 778
 779         if (pkey_event) {
 780                 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
 781                         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 782                         ipoib_ib_dev_down(dev, 0);
 783                         ipoib_ib_dev_stop(dev, 0);
 784                         ipoib_pkey_dev_delay_open(dev);
 785                         return;
 786                 }
 787                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 788
 789                 /* restart QP only if P_Key index is changed */
 790                 if (new_index == priv->pkey_index) {
 791                         ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
 792                         return;
 793                 }
 794                 priv->pkey_index = new_index;
 795         }
 796
 797         ipoib_dbg(priv, "flushing\n");
 798
 799         ipoib_ib_dev_down(dev, 0);
 800
 801         if (pkey_event) {
 802                 ipoib_ib_dev_stop(dev, 0);
 803                 ipoib_ib_dev_open(dev);
 804         }
 805
 806         /*
 807          * The device could have been brought down between the start and when
 808          * we get here, don't bring it back up if it's not configured up
 809          */
 810         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 811                 ipoib_ib_dev_up(dev);
 812                 ipoib_mcast_restart_task(&priv->restart_task);
 813         }
 814 }
 815
 816 void ipoib_ib_dev_flush(struct work_struct *work)
 817 {
 818         struct ipoib_dev_priv *priv =
 819                 container_of(work, struct ipoib_dev_priv, flush_task);
 820
 821         ipoib_dbg(priv, "Flushing %s\n", priv->dev->name);
 822         __ipoib_ib_dev_flush(priv, 0);
 823 }
 824
 825 void ipoib_pkey_event(struct work_struct *work)
 826 {
 827         struct ipoib_dev_priv *priv =
 828                 container_of(work, struct ipoib_dev_priv, pkey_event_task);
 829
 830         ipoib_dbg(priv, "Flushing %s and restarting its QP\n", priv->dev->name);
 831         __ipoib_ib_dev_flush(priv, 1);
 832 }
 833
 834 void ipoib_ib_dev_cleanup(struct net_device *dev)
 835 {
 836         struct ipoib_dev_priv *priv = netdev_priv(dev);
 837
 838         ipoib_dbg(priv, "cleaning up ib_dev\n");
 839
 840         ipoib_mcast_stop_thread(dev, 1);
 841         ipoib_mcast_dev_flush(dev);
 842
 843         ipoib_transport_dev_cleanup(dev);
 844 }
 845
 846 /*
 847  * Delayed P_Key Assigment Interim Support
 848  *
 849  * The following is initial implementation of delayed P_Key assigment
 850  * mechanism. It is using the same approach implemented for the multicast
 851  * group join. The single goal of this implementation is to quickly address
 852  * Bug #2507. This implementation will probably be removed when the P_Key
 853  * change async notification is available.
 854  */
 855
 856 void ipoib_pkey_poll(struct work_struct *work)
 857 {
 858         struct ipoib_dev_priv *priv =
 859                 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
 860         struct net_device *dev = priv->dev;
 861
 862         ipoib_pkey_dev_check_presence(dev);
 863
 864         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 865                 ipoib_open(dev);
 866         else {
 867                 mutex_lock(&pkey_mutex);
 868                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
 869                         queue_delayed_work(ipoib_workqueue,
 870                                            &priv->pkey_poll_task,
 871                                            HZ);
 872                 mutex_unlock(&pkey_mutex);
 873         }
 874 }
 875
 876 int ipoib_pkey_dev_delay_open(struct net_device *dev)
 877 {
 878         struct ipoib_dev_priv *priv = netdev_priv(dev);
 879
 880         /* Look for the interface pkey value in the IB Port P_Key table and */
 881         /* set the interface pkey assigment flag                            */
 882         ipoib_pkey_dev_check_presence(dev);
 883
 884         /* P_Key value not assigned yet - start polling */
 885         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 886                 mutex_lock(&pkey_mutex);
 887                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
 888                 queue_delayed_work(ipoib_workqueue,
 889                                    &priv->pkey_poll_task,
 890                                    HZ);
 891                 mutex_unlock(&pkey_mutex);
 892                 return 1;
 893         }
 894
 895         return 0;
 896 }