SAFE public projects git trees. - safe/jmp/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  * $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
 235         if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
 236                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 237
 238         netif_receive_skb(skb);
 239
 240 repost:
 241         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 242                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 243                            "for buf %d\n", wr_id);
 244 }
 245
 246 static int ipoib_dma_map_tx(struct ib_device *ca,
 247                             struct ipoib_tx_buf *tx_req)
 248 {
 249         struct sk_buff *skb = tx_req->skb;
 250         u64 *mapping = tx_req->mapping;
 251         int i;
 252
 253         mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
 254                                        DMA_TO_DEVICE);
 255         if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
 256                 return -EIO;
 257
 258         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 259                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 260                 mapping[i + 1] = ib_dma_map_page(ca, frag->page,
 261                                                  frag->page_offset, frag->size,
 262                                                  DMA_TO_DEVICE);
 263                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + 1])))
 264                         goto partial_error;
 265         }
 266         return 0;
 267
 268 partial_error:
 269         ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 270
 271         for (; i > 0; --i) {
 272                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 273                 ib_dma_unmap_page(ca, mapping[i], frag->size, DMA_TO_DEVICE);
 274         }
 275         return -EIO;
 276 }
 277
 278 static void ipoib_dma_unmap_tx(struct ib_device *ca,
 279                                struct ipoib_tx_buf *tx_req)
 280 {
 281         struct sk_buff *skb = tx_req->skb;
 282         u64 *mapping = tx_req->mapping;
 283         int i;
 284
 285         ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 286
 287         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 288                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 289                 ib_dma_unmap_page(ca, mapping[i + 1], frag->size,
 290                                   DMA_TO_DEVICE);
 291         }
 292 }
 293
 294 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 295 {
 296         struct ipoib_dev_priv *priv = netdev_priv(dev);
 297         unsigned int wr_id = wc->wr_id;
 298         struct ipoib_tx_buf *tx_req;
 299         unsigned long flags;
 300
 301         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 302                        wr_id, wc->status);
 303
 304         if (unlikely(wr_id >= ipoib_sendq_size)) {
 305                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 306                            wr_id, ipoib_sendq_size);
 307                 return;
 308         }
 309
 310         tx_req = &priv->tx_ring[wr_id];
 311
 312         ipoib_dma_unmap_tx(priv->ca, tx_req);
 313
 314         ++dev->stats.tx_packets;
 315         dev->stats.tx_bytes += tx_req->skb->len;
 316
 317         dev_kfree_skb_any(tx_req->skb);
 318
 319         spin_lock_irqsave(&priv->tx_lock, flags);
 320         ++priv->tx_tail;
 321         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 322             netif_queue_stopped(dev) &&
 323             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 324                 netif_wake_queue(dev);
 325         spin_unlock_irqrestore(&priv->tx_lock, flags);
 326
 327         if (wc->status != IB_WC_SUCCESS &&
 328             wc->status != IB_WC_WR_FLUSH_ERR)
 329                 ipoib_warn(priv, "failed send event "
 330                            "(status=%d, wrid=%d vend_err %x)\n",
 331                            wc->status, wr_id, wc->vendor_err);
 332 }
 333
 334 int ipoib_poll(struct napi_struct *napi, int budget)
 335 {
 336         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
 337         struct net_device *dev = priv->dev;
 338         int done;
 339         int t;
 340         int n, i;
 341
 342         done  = 0;
 343
 344 poll_more:
 345         while (done < budget) {
 346                 int max = (budget - done);
 347
 348                 t = min(IPOIB_NUM_WC, max);
 349                 n = ib_poll_cq(priv->cq, t, priv->ibwc);
 350
 351                 for (i = 0; i < n; i++) {
 352                         struct ib_wc *wc = priv->ibwc + i;
 353
 354                         if (wc->wr_id & IPOIB_OP_RECV) {
 355                                 ++done;
 356                                 if (wc->wr_id & IPOIB_OP_CM)
 357                                         ipoib_cm_handle_rx_wc(dev, wc);
 358                                 else
 359                                         ipoib_ib_handle_rx_wc(dev, wc);
 360                         } else {
 361                                 if (wc->wr_id & IPOIB_OP_CM)
 362                                         ipoib_cm_handle_tx_wc(dev, wc);
 363                                 else
 364                                         ipoib_ib_handle_tx_wc(dev, wc);
 365                         }
 366                 }
 367
 368                 if (n != t)
 369                         break;
 370         }
 371
 372         if (done < budget) {
 373                 netif_rx_complete(dev, napi);
 374                 if (unlikely(ib_req_notify_cq(priv->cq,
 375                                               IB_CQ_NEXT_COMP |
 376                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 377                     netif_rx_reschedule(dev, napi))
 378                         goto poll_more;
 379         }
 380
 381         return done;
 382 }
 383
 384 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 385 {
 386         struct net_device *dev = dev_ptr;
 387         struct ipoib_dev_priv *priv = netdev_priv(dev);
 388
 389         netif_rx_schedule(dev, &priv->napi);
 390 }
 391
 392 static inline int post_send(struct ipoib_dev_priv *priv,
 393                             unsigned int wr_id,
 394                             struct ib_ah *address, u32 qpn,
 395                             u64 *mapping, int headlen,
 396                             skb_frag_t *frags,
 397                             int nr_frags)
 398 {
 399         struct ib_send_wr *bad_wr;
 400         int i;
 401
 402         priv->tx_sge[0].addr         = mapping[0];
 403         priv->tx_sge[0].length       = headlen;
 404         for (i = 0; i < nr_frags; ++i) {
 405                 priv->tx_sge[i + 1].addr = mapping[i + 1];
 406                 priv->tx_sge[i + 1].length = frags[i].size;
 407         }
 408         priv->tx_wr.num_sge          = nr_frags + 1;
 409         priv->tx_wr.wr_id            = wr_id;
 410         priv->tx_wr.wr.ud.remote_qpn = qpn;
 411         priv->tx_wr.wr.ud.ah         = address;
 412
 413         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 414 }
 415
 416 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 417                 struct ipoib_ah *address, u32 qpn)
 418 {
 419         struct ipoib_dev_priv *priv = netdev_priv(dev);
 420         struct ipoib_tx_buf *tx_req;
 421
 422         if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 423                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 424                            skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 425                 ++dev->stats.tx_dropped;
 426                 ++dev->stats.tx_errors;
 427                 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 428                 return;
 429         }
 430
 431         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 432                        skb->len, address, qpn);
 433
 434         /*
 435          * We put the skb into the tx_ring _before_ we call post_send()
 436          * because it's entirely possible that the completion handler will
 437          * run before we execute anything after the post_send().  That
 438          * means we have to make sure everything is properly recorded and
 439          * our state is consistent before we call post_send().
 440          */
 441         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 442         tx_req->skb = skb;
 443         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 444                 ++dev->stats.tx_errors;
 445                 dev_kfree_skb_any(skb);
 446                 return;
 447         }
 448
 449         if (skb->ip_summed == CHECKSUM_PARTIAL)
 450                 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
 451         else
 452                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
 453
 454         if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 455                                address->ah, qpn,
 456                                tx_req->mapping, skb_headlen(skb),
 457                                skb_shinfo(skb)->frags, skb_shinfo(skb)->nr_frags))) {
 458                 ipoib_warn(priv, "post_send failed\n");
 459                 ++dev->stats.tx_errors;
 460                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 461                 dev_kfree_skb_any(skb);
 462         } else {
 463                 dev->trans_start = jiffies;
 464
 465                 address->last_send = priv->tx_head;
 466                 ++priv->tx_head;
 467
 468                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 469                         ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 470                         netif_stop_queue(dev);
 471                 }
 472         }
 473 }
 474
 475 static void __ipoib_reap_ah(struct net_device *dev)
 476 {
 477         struct ipoib_dev_priv *priv = netdev_priv(dev);
 478         struct ipoib_ah *ah, *tah;
 479         LIST_HEAD(remove_list);
 480
 481         spin_lock_irq(&priv->tx_lock);
 482         spin_lock(&priv->lock);
 483         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 484                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 485                         list_del(&ah->list);
 486                         ib_destroy_ah(ah->ah);
 487                         kfree(ah);
 488                 }
 489         spin_unlock(&priv->lock);
 490         spin_unlock_irq(&priv->tx_lock);
 491 }
 492
 493 void ipoib_reap_ah(struct work_struct *work)
 494 {
 495         struct ipoib_dev_priv *priv =
 496                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 497         struct net_device *dev = priv->dev;
 498
 499         __ipoib_reap_ah(dev);
 500
 501         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 502                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 503                                    round_jiffies_relative(HZ));
 504 }
 505
 506 int ipoib_ib_dev_open(struct net_device *dev)
 507 {
 508         struct ipoib_dev_priv *priv = netdev_priv(dev);
 509         int ret;
 510
 511         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
 512                 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
 513                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 514                 return -1;
 515         }
 516         set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 517
 518         ret = ipoib_init_qp(dev);
 519         if (ret) {
 520                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 521                 return -1;
 522         }
 523
 524         ret = ipoib_ib_post_receives(dev);
 525         if (ret) {
 526                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 527                 ipoib_ib_dev_stop(dev, 1);
 528                 return -1;
 529         }
 530
 531         ret = ipoib_cm_dev_open(dev);
 532         if (ret) {
 533                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 534                 ipoib_ib_dev_stop(dev, 1);
 535                 return -1;
 536         }
 537
 538         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 539         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 540                            round_jiffies_relative(HZ));
 541
 542         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 543
 544         return 0;
 545 }
 546
 547 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
 548 {
 549         struct ipoib_dev_priv *priv = netdev_priv(dev);
 550         u16 pkey_index = 0;
 551
 552         if (ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
 553                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 554         else
 555                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 556 }
 557
 558 int ipoib_ib_dev_up(struct net_device *dev)
 559 {
 560         struct ipoib_dev_priv *priv = netdev_priv(dev);
 561
 562         ipoib_pkey_dev_check_presence(dev);
 563
 564         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 565                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 566                 return 0;
 567         }
 568
 569         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 570
 571         return ipoib_mcast_start_thread(dev);
 572 }
 573
 574 int ipoib_ib_dev_down(struct net_device *dev, int flush)
 575 {
 576         struct ipoib_dev_priv *priv = netdev_priv(dev);
 577
 578         ipoib_dbg(priv, "downing ib_dev\n");
 579
 580         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 581         netif_carrier_off(dev);
 582
 583         /* Shutdown the P_Key thread if still active */
 584         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 585                 mutex_lock(&pkey_mutex);
 586                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
 587                 cancel_delayed_work(&priv->pkey_poll_task);
 588                 mutex_unlock(&pkey_mutex);
 589                 if (flush)
 590                         flush_workqueue(ipoib_workqueue);
 591         }
 592
 593         ipoib_mcast_stop_thread(dev, flush);
 594         ipoib_mcast_dev_flush(dev);
 595
 596         ipoib_flush_paths(dev);
 597
 598         return 0;
 599 }
 600
 601 static int recvs_pending(struct net_device *dev)
 602 {
 603         struct ipoib_dev_priv *priv = netdev_priv(dev);
 604         int pending = 0;
 605         int i;
 606
 607         for (i = 0; i < ipoib_recvq_size; ++i)
 608                 if (priv->rx_ring[i].skb)
 609                         ++pending;
 610
 611         return pending;
 612 }
 613
 614 void ipoib_drain_cq(struct net_device *dev)
 615 {
 616         struct ipoib_dev_priv *priv = netdev_priv(dev);
 617         int i, n;
 618         do {
 619                 n = ib_poll_cq(priv->cq, IPOIB_NUM_WC, priv->ibwc);
 620                 for (i = 0; i < n; ++i) {
 621                         /*
 622                          * Convert any successful completions to flush
 623                          * errors to avoid passing packets up the
 624                          * stack after bringing the device down.
 625                          */
 626                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 627                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 628
 629                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 630                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 631                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 632                                 else
 633                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 634                         } else {
 635                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 636                                         ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
 637                                 else
 638                                         ipoib_ib_handle_tx_wc(dev, priv->ibwc + i);
 639                         }
 640                 }
 641         } while (n == IPOIB_NUM_WC);
 642 }
 643
 644 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
 645 {
 646         struct ipoib_dev_priv *priv = netdev_priv(dev);
 647         struct ib_qp_attr qp_attr;
 648         unsigned long begin;
 649         struct ipoib_tx_buf *tx_req;
 650         int i;
 651
 652         clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 653
 654         ipoib_cm_dev_stop(dev);
 655
 656         /*
 657          * Move our QP to the error state and then reinitialize in
 658          * when all work requests have completed or have been flushed.
 659          */
 660         qp_attr.qp_state = IB_QPS_ERR;
 661         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 662                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 663
 664         /* Wait for all sends and receives to complete */
 665         begin = jiffies;
 666
 667         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 668                 if (time_after(jiffies, begin + 5 * HZ)) {
 669                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 670                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 671
 672                         /*
 673                          * assume the HW is wedged and just free up
 674                          * all our pending work requests.
 675                          */
 676                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 677                                 tx_req = &priv->tx_ring[priv->tx_tail &
 678                                                         (ipoib_sendq_size - 1)];
 679                                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 680                                 dev_kfree_skb_any(tx_req->skb);
 681                                 ++priv->tx_tail;
 682                                 --priv->tx_outstanding;
 683                         }
 684
 685                         for (i = 0; i < ipoib_recvq_size; ++i) {
 686                                 struct ipoib_rx_buf *rx_req;
 687
 688                                 rx_req = &priv->rx_ring[i];
 689                                 if (!rx_req->skb)
 690                                         continue;
 691                                 ib_dma_unmap_single(priv->ca,
 692                                                     rx_req->mapping,
 693                                                     IPOIB_BUF_SIZE,
 694                                                     DMA_FROM_DEVICE);
 695                                 dev_kfree_skb_any(rx_req->skb);
 696                                 rx_req->skb = NULL;
 697                         }
 698
 699                         goto timeout;
 700                 }
 701
 702                 ipoib_drain_cq(dev);
 703
 704                 msleep(1);
 705         }
 706
 707         ipoib_dbg(priv, "All sends and receives done.\n");
 708
 709 timeout:
 710         qp_attr.qp_state = IB_QPS_RESET;
 711         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 712                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 713
 714         /* Wait for all AHs to be reaped */
 715         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 716         cancel_delayed_work(&priv->ah_reap_task);
 717         if (flush)
 718                 flush_workqueue(ipoib_workqueue);
 719
 720         begin = jiffies;
 721
 722         while (!list_empty(&priv->dead_ahs)) {
 723                 __ipoib_reap_ah(dev);
 724
 725                 if (time_after(jiffies, begin + HZ)) {
 726                         ipoib_warn(priv, "timing out; will leak address handles\n");
 727                         break;
 728                 }
 729
 730                 msleep(1);
 731         }
 732
 733         ib_req_notify_cq(priv->cq, IB_CQ_NEXT_COMP);
 734
 735         return 0;
 736 }
 737
 738 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 739 {
 740         struct ipoib_dev_priv *priv = netdev_priv(dev);
 741
 742         priv->ca = ca;
 743         priv->port = port;
 744         priv->qp = NULL;
 745
 746         if (ipoib_transport_dev_init(dev, ca)) {
 747                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 748                 return -ENODEV;
 749         }
 750
 751         if (dev->flags & IFF_UP) {
 752                 if (ipoib_ib_dev_open(dev)) {
 753                         ipoib_transport_dev_cleanup(dev);
 754                         return -ENODEV;
 755                 }
 756         }
 757
 758         return 0;
 759 }
 760
 761 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv, int pkey_event)
 762 {
 763         struct ipoib_dev_priv *cpriv;
 764         struct net_device *dev = priv->dev;
 765         u16 new_index;
 766
 767         mutex_lock(&priv->vlan_mutex);
 768
 769         /*
 770          * Flush any child interfaces too -- they might be up even if
 771          * the parent is down.
 772          */
 773         list_for_each_entry(cpriv, &priv->child_intfs, list)
 774                 __ipoib_ib_dev_flush(cpriv, pkey_event);
 775
 776         mutex_unlock(&priv->vlan_mutex);
 777
 778         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
 779                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
 780                 return;
 781         }
 782
 783         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 784                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
 785                 return;
 786         }
 787
 788         if (pkey_event) {
 789                 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
 790                         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 791                         ipoib_ib_dev_down(dev, 0);
 792                         ipoib_ib_dev_stop(dev, 0);
 793                         ipoib_pkey_dev_delay_open(dev);
 794                         return;
 795                 }
 796                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 797
 798                 /* restart QP only if P_Key index is changed */
 799                 if (new_index == priv->pkey_index) {
 800                         ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
 801                         return;
 802                 }
 803                 priv->pkey_index = new_index;
 804         }
 805
 806         ipoib_dbg(priv, "flushing\n");
 807
 808         ipoib_ib_dev_down(dev, 0);
 809
 810         if (pkey_event) {
 811                 ipoib_ib_dev_stop(dev, 0);
 812                 ipoib_ib_dev_open(dev);
 813         }
 814
 815         /*
 816          * The device could have been brought down between the start and when
 817          * we get here, don't bring it back up if it's not configured up
 818          */
 819         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 820                 ipoib_ib_dev_up(dev);
 821                 ipoib_mcast_restart_task(&priv->restart_task);
 822         }
 823 }
 824
 825 void ipoib_ib_dev_flush(struct work_struct *work)
 826 {
 827         struct ipoib_dev_priv *priv =
 828                 container_of(work, struct ipoib_dev_priv, flush_task);
 829
 830         ipoib_dbg(priv, "Flushing %s\n", priv->dev->name);
 831         __ipoib_ib_dev_flush(priv, 0);
 832 }
 833
 834 void ipoib_pkey_event(struct work_struct *work)
 835 {
 836         struct ipoib_dev_priv *priv =
 837                 container_of(work, struct ipoib_dev_priv, pkey_event_task);
 838
 839         ipoib_dbg(priv, "Flushing %s and restarting its QP\n", priv->dev->name);
 840         __ipoib_ib_dev_flush(priv, 1);
 841 }
 842
 843 void ipoib_ib_dev_cleanup(struct net_device *dev)
 844 {
 845         struct ipoib_dev_priv *priv = netdev_priv(dev);
 846
 847         ipoib_dbg(priv, "cleaning up ib_dev\n");
 848
 849         ipoib_mcast_stop_thread(dev, 1);
 850         ipoib_mcast_dev_flush(dev);
 851
 852         ipoib_transport_dev_cleanup(dev);
 853 }
 854
 855 /*
 856  * Delayed P_Key Assigment Interim Support
 857  *
 858  * The following is initial implementation of delayed P_Key assigment
 859  * mechanism. It is using the same approach implemented for the multicast
 860  * group join. The single goal of this implementation is to quickly address
 861  * Bug #2507. This implementation will probably be removed when the P_Key
 862  * change async notification is available.
 863  */
 864
 865 void ipoib_pkey_poll(struct work_struct *work)
 866 {
 867         struct ipoib_dev_priv *priv =
 868                 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
 869         struct net_device *dev = priv->dev;
 870
 871         ipoib_pkey_dev_check_presence(dev);
 872
 873         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 874                 ipoib_open(dev);
 875         else {
 876                 mutex_lock(&pkey_mutex);
 877                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
 878                         queue_delayed_work(ipoib_workqueue,
 879                                            &priv->pkey_poll_task,
 880                                            HZ);
 881                 mutex_unlock(&pkey_mutex);
 882         }
 883 }
 884
 885 int ipoib_pkey_dev_delay_open(struct net_device *dev)
 886 {
 887         struct ipoib_dev_priv *priv = netdev_priv(dev);
 888
 889         /* Look for the interface pkey value in the IB Port P_Key table and */
 890         /* set the interface pkey assigment flag                            */
 891         ipoib_pkey_dev_check_presence(dev);
 892
 893         /* P_Key value not assigned yet - start polling */
 894         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 895                 mutex_lock(&pkey_mutex);
 896                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
 897                 queue_delayed_work(ipoib_workqueue,
 898                                    &priv->pkey_poll_task,
 899                                    HZ);
 900                 mutex_unlock(&pkey_mutex);
 901                 return 1;
 902         }
 903
 904         return 0;
 905 }