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