err.no Git - linux-2.6/blob - drivers/infiniband/ulp/ipoib/ipoib_ib.c

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  *
  35  * $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 #include <linux/ip.h>
  43 #include <linux/tcp.h>
  44
  45 #include "ipoib.h"
  46
  47 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  48 static int data_debug_level;
  49
  50 module_param(data_debug_level, int, 0644);
  51 MODULE_PARM_DESC(data_debug_level,
  52                  "Enable data path debug tracing if > 0");
  53 #endif
  54
  55 static DEFINE_MUTEX(pkey_mutex);
  56
  57 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
  58                                  struct ib_pd *pd, struct ib_ah_attr *attr)
  59 {
  60         struct ipoib_ah *ah;
  61
  62         ah = kmalloc(sizeof *ah, GFP_KERNEL);
  63         if (!ah)
  64                 return NULL;
  65
  66         ah->dev       = dev;
  67         ah->last_send = 0;
  68         kref_init(&ah->ref);
  69
  70         ah->ah = ib_create_ah(pd, attr);
  71         if (IS_ERR(ah->ah)) {
  72                 kfree(ah);
  73                 ah = NULL;
  74         } else
  75                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
  76
  77         return ah;
  78 }
  79
  80 void ipoib_free_ah(struct kref *kref)
  81 {
  82         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
  83         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
  84
  85         unsigned long flags;
  86
  87         spin_lock_irqsave(&priv->lock, flags);
  88         list_add_tail(&ah->list, &priv->dead_ahs);
  89         spin_unlock_irqrestore(&priv->lock, flags);
  90 }
  91
  92 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
  93                                   u64 mapping[IPOIB_UD_RX_SG])
  94 {
  95         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
  96                 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
  97                                     DMA_FROM_DEVICE);
  98                 ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
  99                                   DMA_FROM_DEVICE);
 100         } else
 101                 ib_dma_unmap_single(priv->ca, mapping[0],
 102                                     IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
 103                                     DMA_FROM_DEVICE);
 104 }
 105
 106 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
 107                                    struct sk_buff *skb,
 108                                    unsigned int length)
 109 {
 110         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
 111                 skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
 112                 unsigned int size;
 113                 /*
 114                  * There is only two buffers needed for max_payload = 4K,
 115                  * first buf size is IPOIB_UD_HEAD_SIZE
 116                  */
 117                 skb->tail += IPOIB_UD_HEAD_SIZE;
 118                 skb->len  += length;
 119
 120                 size = length - IPOIB_UD_HEAD_SIZE;
 121
 122                 frag->size     = size;
 123                 skb->data_len += size;
 124                 skb->truesize += size;
 125         } else
 126                 skb_put(skb, length);
 127
 128 }
 129
 130 static int ipoib_ib_post_receive(struct net_device *dev, int id)
 131 {
 132         struct ipoib_dev_priv *priv = netdev_priv(dev);
 133         struct ib_recv_wr *bad_wr;
 134         int ret;
 135
 136         priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
 137         priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
 138         priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
 139
 140
 141         ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
 142         if (unlikely(ret)) {
 143                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
 144                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
 145                 dev_kfree_skb_any(priv->rx_ring[id].skb);
 146                 priv->rx_ring[id].skb = NULL;
 147         }
 148
 149         return ret;
 150 }
 151
 152 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
 153 {
 154         struct ipoib_dev_priv *priv = netdev_priv(dev);
 155         struct sk_buff *skb;
 156         int buf_size;
 157         u64 *mapping;
 158
 159         if (ipoib_ud_need_sg(priv->max_ib_mtu))
 160                 buf_size = IPOIB_UD_HEAD_SIZE;
 161         else
 162                 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
 163
 164         skb = dev_alloc_skb(buf_size + 4);
 165         if (unlikely(!skb))
 166                 return NULL;
 167
 168         /*
 169          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
 170          * header.  So we need 4 more bytes to get to 48 and align the
 171          * IP header to a multiple of 16.
 172          */
 173         skb_reserve(skb, 4);
 174
 175         mapping = priv->rx_ring[id].mapping;
 176         mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
 177                                        DMA_FROM_DEVICE);
 178         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
 179                 goto error;
 180
 181         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
 182                 struct page *page = alloc_page(GFP_ATOMIC);
 183                 if (!page)
 184                         goto partial_error;
 185                 skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
 186                 mapping[1] =
 187                         ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[0].page,
 188                                         0, PAGE_SIZE, DMA_FROM_DEVICE);
 189                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
 190                         goto partial_error;
 191         }
 192
 193         priv->rx_ring[id].skb = skb;
 194         return skb;
 195
 196 partial_error:
 197         ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
 198 error:
 199         dev_kfree_skb_any(skb);
 200         return NULL;
 201 }
 202
 203 static int ipoib_ib_post_receives(struct net_device *dev)
 204 {
 205         struct ipoib_dev_priv *priv = netdev_priv(dev);
 206         int i;
 207
 208         for (i = 0; i < ipoib_recvq_size; ++i) {
 209                 if (!ipoib_alloc_rx_skb(dev, i)) {
 210                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 211                         return -ENOMEM;
 212                 }
 213                 if (ipoib_ib_post_receive(dev, i)) {
 214                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 215                         return -EIO;
 216                 }
 217         }
 218
 219         return 0;
 220 }
 221
 222 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 223 {
 224         struct ipoib_dev_priv *priv = netdev_priv(dev);
 225         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
 226         struct sk_buff *skb;
 227         u64 mapping[IPOIB_UD_RX_SG];
 228
 229         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
 230                        wr_id, wc->status);
 231
 232         if (unlikely(wr_id >= ipoib_recvq_size)) {
 233                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
 234                            wr_id, ipoib_recvq_size);
 235                 return;
 236         }
 237
 238         skb  = priv->rx_ring[wr_id].skb;
 239
 240         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 241                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 242                         ipoib_warn(priv, "failed recv event "
 243                                    "(status=%d, wrid=%d vend_err %x)\n",
 244                                    wc->status, wr_id, wc->vendor_err);
 245                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
 246                 dev_kfree_skb_any(skb);
 247                 priv->rx_ring[wr_id].skb = NULL;
 248                 return;
 249         }
 250
 251         /*
 252          * Drop packets that this interface sent, ie multicast packets
 253          * that the HCA has replicated.
 254          */
 255         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
 256                 goto repost;
 257
 258         memcpy(mapping, priv->rx_ring[wr_id].mapping,
 259                IPOIB_UD_RX_SG * sizeof *mapping);
 260
 261         /*
 262          * If we can't allocate a new RX buffer, dump
 263          * this packet and reuse the old buffer.
 264          */
 265         if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
 266                 ++dev->stats.rx_dropped;
 267                 goto repost;
 268         }
 269
 270         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 271                        wc->byte_len, wc->slid);
 272
 273         ipoib_ud_dma_unmap_rx(priv, mapping);
 274         ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
 275
 276         skb_pull(skb, IB_GRH_BYTES);
 277
 278         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 279         skb_reset_mac_header(skb);
 280         skb_pull(skb, IPOIB_ENCAP_LEN);
 281
 282         dev->last_rx = jiffies;
 283         ++dev->stats.rx_packets;
 284         dev->stats.rx_bytes += skb->len;
 285
 286         skb->dev = dev;
 287         /* XXX get correct PACKET_ type here */
 288         skb->pkt_type = PACKET_HOST;
 289
 290         if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
 291                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 292
 293         netif_receive_skb(skb);
 294
 295 repost:
 296         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 297                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 298                            "for buf %d\n", wr_id);
 299 }
 300
 301 static int ipoib_dma_map_tx(struct ib_device *ca,
 302                             struct ipoib_tx_buf *tx_req)
 303 {
 304         struct sk_buff *skb = tx_req->skb;
 305         u64 *mapping = tx_req->mapping;
 306         int i;
 307         int off;
 308
 309         if (skb_headlen(skb)) {
 310                 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
 311                                                DMA_TO_DEVICE);
 312                 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
 313                         return -EIO;
 314
 315                 off = 1;
 316         } else
 317                 off = 0;
 318
 319         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 320                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 321                 mapping[i + off] = ib_dma_map_page(ca, frag->page,
 322                                                  frag->page_offset, frag->size,
 323                                                  DMA_TO_DEVICE);
 324                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
 325                         goto partial_error;
 326         }
 327         return 0;
 328
 329 partial_error:
 330         for (; i > 0; --i) {
 331                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 332                 ib_dma_unmap_page(ca, mapping[i - !off], frag->size, DMA_TO_DEVICE);
 333         }
 334
 335         if (off)
 336                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 337
 338         return -EIO;
 339 }
 340
 341 static void ipoib_dma_unmap_tx(struct ib_device *ca,
 342                                struct ipoib_tx_buf *tx_req)
 343 {
 344         struct sk_buff *skb = tx_req->skb;
 345         u64 *mapping = tx_req->mapping;
 346         int i;
 347         int off;
 348
 349         if (skb_headlen(skb)) {
 350                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 351                 off = 1;
 352         } else
 353                 off = 0;
 354
 355         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 356                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 357                 ib_dma_unmap_page(ca, mapping[i + off], frag->size,
 358                                   DMA_TO_DEVICE);
 359         }
 360 }
 361
 362 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 363 {
 364         struct ipoib_dev_priv *priv = netdev_priv(dev);
 365         unsigned int wr_id = wc->wr_id;
 366         struct ipoib_tx_buf *tx_req;
 367         unsigned long flags;
 368
 369         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 370                        wr_id, wc->status);
 371
 372         if (unlikely(wr_id >= ipoib_sendq_size)) {
 373                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 374                            wr_id, ipoib_sendq_size);
 375                 return;
 376         }
 377
 378         tx_req = &priv->tx_ring[wr_id];
 379
 380         ipoib_dma_unmap_tx(priv->ca, tx_req);
 381
 382         ++dev->stats.tx_packets;
 383         dev->stats.tx_bytes += tx_req->skb->len;
 384
 385         dev_kfree_skb_any(tx_req->skb);
 386
 387         spin_lock_irqsave(&priv->tx_lock, flags);
 388         ++priv->tx_tail;
 389         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 390             netif_queue_stopped(dev) &&
 391             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 392                 netif_wake_queue(dev);
 393         spin_unlock_irqrestore(&priv->tx_lock, flags);
 394
 395         if (wc->status != IB_WC_SUCCESS &&
 396             wc->status != IB_WC_WR_FLUSH_ERR)
 397                 ipoib_warn(priv, "failed send event "
 398                            "(status=%d, wrid=%d vend_err %x)\n",
 399                            wc->status, wr_id, wc->vendor_err);
 400 }
 401
 402 int ipoib_poll(struct napi_struct *napi, int budget)
 403 {
 404         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
 405         struct net_device *dev = priv->dev;
 406         int done;
 407         int t;
 408         int n, i;
 409
 410         done  = 0;
 411
 412 poll_more:
 413         while (done < budget) {
 414                 int max = (budget - done);
 415
 416                 t = min(IPOIB_NUM_WC, max);
 417                 n = ib_poll_cq(priv->cq, t, priv->ibwc);
 418
 419                 for (i = 0; i < n; i++) {
 420                         struct ib_wc *wc = priv->ibwc + i;
 421
 422                         if (wc->wr_id & IPOIB_OP_RECV) {
 423                                 ++done;
 424                                 if (wc->wr_id & IPOIB_OP_CM)
 425                                         ipoib_cm_handle_rx_wc(dev, wc);
 426                                 else
 427                                         ipoib_ib_handle_rx_wc(dev, wc);
 428                         } else {
 429                                 if (wc->wr_id & IPOIB_OP_CM)
 430                                         ipoib_cm_handle_tx_wc(dev, wc);
 431                                 else
 432                                         ipoib_ib_handle_tx_wc(dev, wc);
 433                         }
 434                 }
 435
 436                 if (n != t)
 437                         break;
 438         }
 439
 440         if (done < budget) {
 441                 netif_rx_complete(dev, napi);
 442                 if (unlikely(ib_req_notify_cq(priv->cq,
 443                                               IB_CQ_NEXT_COMP |
 444                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 445                     netif_rx_reschedule(dev, napi))
 446                         goto poll_more;
 447         }
 448
 449         return done;
 450 }
 451
 452 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 453 {
 454         struct net_device *dev = dev_ptr;
 455         struct ipoib_dev_priv *priv = netdev_priv(dev);
 456
 457         netif_rx_schedule(dev, &priv->napi);
 458 }
 459
 460 static inline int post_send(struct ipoib_dev_priv *priv,
 461                             unsigned int wr_id,
 462                             struct ib_ah *address, u32 qpn,
 463                             struct ipoib_tx_buf *tx_req,
 464                             void *head, int hlen)
 465 {
 466         struct ib_send_wr *bad_wr;
 467         int i, off;
 468         struct sk_buff *skb = tx_req->skb;
 469         skb_frag_t *frags = skb_shinfo(skb)->frags;
 470         int nr_frags = skb_shinfo(skb)->nr_frags;
 471         u64 *mapping = tx_req->mapping;
 472
 473         if (skb_headlen(skb)) {
 474                 priv->tx_sge[0].addr         = mapping[0];
 475                 priv->tx_sge[0].length       = skb_headlen(skb);
 476                 off = 1;
 477         } else
 478                 off = 0;
 479
 480         for (i = 0; i < nr_frags; ++i) {
 481                 priv->tx_sge[i + off].addr = mapping[i + off];
 482                 priv->tx_sge[i + off].length = frags[i].size;
 483         }
 484         priv->tx_wr.num_sge          = nr_frags + off;
 485         priv->tx_wr.wr_id            = wr_id;
 486         priv->tx_wr.wr.ud.remote_qpn = qpn;
 487         priv->tx_wr.wr.ud.ah         = address;
 488
 489         if (head) {
 490                 priv->tx_wr.wr.ud.mss    = skb_shinfo(skb)->gso_size;
 491                 priv->tx_wr.wr.ud.header = head;
 492                 priv->tx_wr.wr.ud.hlen   = hlen;
 493                 priv->tx_wr.opcode       = IB_WR_LSO;
 494         } else
 495                 priv->tx_wr.opcode       = IB_WR_SEND;
 496
 497         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 498 }
 499
 500 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 501                 struct ipoib_ah *address, u32 qpn)
 502 {
 503         struct ipoib_dev_priv *priv = netdev_priv(dev);
 504         struct ipoib_tx_buf *tx_req;
 505         int hlen;
 506         void *phead;
 507
 508         if (skb_is_gso(skb)) {
 509                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
 510                 phead = skb->data;
 511                 if (unlikely(!skb_pull(skb, hlen))) {
 512                         ipoib_warn(priv, "linear data too small\n");
 513                         ++dev->stats.tx_dropped;
 514                         ++dev->stats.tx_errors;
 515                         dev_kfree_skb_any(skb);
 516                         return;
 517                 }
 518         } else {
 519                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 520                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 521                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 522                         ++dev->stats.tx_dropped;
 523                         ++dev->stats.tx_errors;
 524                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 525                         return;
 526                 }
 527                 phead = NULL;
 528                 hlen  = 0;
 529         }
 530
 531         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 532                        skb->len, address, qpn);
 533
 534         /*
 535          * We put the skb into the tx_ring _before_ we call post_send()
 536          * because it's entirely possible that the completion handler will
 537          * run before we execute anything after the post_send().  That
 538          * means we have to make sure everything is properly recorded and
 539          * our state is consistent before we call post_send().
 540          */
 541         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 542         tx_req->skb = skb;
 543         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 544                 ++dev->stats.tx_errors;
 545                 dev_kfree_skb_any(skb);
 546                 return;
 547         }
 548
 549         if (skb->ip_summed == CHECKSUM_PARTIAL)
 550                 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
 551         else
 552                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
 553
 554         if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 555                                address->ah, qpn, tx_req, phead, hlen))) {
 556                 ipoib_warn(priv, "post_send failed\n");
 557                 ++dev->stats.tx_errors;
 558                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 559                 dev_kfree_skb_any(skb);
 560         } else {
 561                 dev->trans_start = jiffies;
 562
 563                 address->last_send = priv->tx_head;
 564                 ++priv->tx_head;
 565
 566                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 567                         ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 568                         netif_stop_queue(dev);
 569                 }
 570         }
 571 }
 572
 573 static void __ipoib_reap_ah(struct net_device *dev)
 574 {
 575         struct ipoib_dev_priv *priv = netdev_priv(dev);
 576         struct ipoib_ah *ah, *tah;
 577         LIST_HEAD(remove_list);
 578
 579         spin_lock_irq(&priv->tx_lock);
 580         spin_lock(&priv->lock);
 581         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 582                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 583                         list_del(&ah->list);
 584                         ib_destroy_ah(ah->ah);
 585                         kfree(ah);
 586                 }
 587         spin_unlock(&priv->lock);
 588         spin_unlock_irq(&priv->tx_lock);
 589 }
 590
 591 void ipoib_reap_ah(struct work_struct *work)
 592 {
 593         struct ipoib_dev_priv *priv =
 594                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 595         struct net_device *dev = priv->dev;
 596
 597         __ipoib_reap_ah(dev);
 598
 599         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 600                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 601                                    round_jiffies_relative(HZ));
 602 }
 603
 604 int ipoib_ib_dev_open(struct net_device *dev)
 605 {
 606         struct ipoib_dev_priv *priv = netdev_priv(dev);
 607         int ret;
 608
 609         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
 610                 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
 611                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 612                 return -1;
 613         }
 614         set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 615
 616         ret = ipoib_init_qp(dev);
 617         if (ret) {
 618                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 619                 return -1;
 620         }
 621
 622         ret = ipoib_ib_post_receives(dev);
 623         if (ret) {
 624                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 625                 ipoib_ib_dev_stop(dev, 1);
 626                 return -1;
 627         }
 628
 629         ret = ipoib_cm_dev_open(dev);
 630         if (ret) {
 631                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 632                 ipoib_ib_dev_stop(dev, 1);
 633                 return -1;
 634         }
 635
 636         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 637         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 638                            round_jiffies_relative(HZ));
 639
 640         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 641
 642         return 0;
 643 }
 644
 645 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
 646 {
 647         struct ipoib_dev_priv *priv = netdev_priv(dev);
 648         u16 pkey_index = 0;
 649
 650         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
 651                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 652         else
 653                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 654 }
 655
 656 int ipoib_ib_dev_up(struct net_device *dev)
 657 {
 658         struct ipoib_dev_priv *priv = netdev_priv(dev);
 659
 660         ipoib_pkey_dev_check_presence(dev);
 661
 662         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 663                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 664                 return 0;
 665         }
 666
 667         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 668
 669         return ipoib_mcast_start_thread(dev);
 670 }
 671
 672 int ipoib_ib_dev_down(struct net_device *dev, int flush)
 673 {
 674         struct ipoib_dev_priv *priv = netdev_priv(dev);
 675
 676         ipoib_dbg(priv, "downing ib_dev\n");
 677
 678         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 679         netif_carrier_off(dev);
 680
 681         /* Shutdown the P_Key thread if still active */
 682         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 683                 mutex_lock(&pkey_mutex);
 684                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
 685                 cancel_delayed_work(&priv->pkey_poll_task);
 686                 mutex_unlock(&pkey_mutex);
 687                 if (flush)
 688                         flush_workqueue(ipoib_workqueue);
 689         }
 690
 691         ipoib_mcast_stop_thread(dev, flush);
 692         ipoib_mcast_dev_flush(dev);
 693
 694         ipoib_flush_paths(dev);
 695
 696         return 0;
 697 }
 698
 699 static int recvs_pending(struct net_device *dev)
 700 {
 701         struct ipoib_dev_priv *priv = netdev_priv(dev);
 702         int pending = 0;
 703         int i;
 704
 705         for (i = 0; i < ipoib_recvq_size; ++i)
 706                 if (priv->rx_ring[i].skb)
 707                         ++pending;
 708
 709         return pending;
 710 }
 711
 712 void ipoib_drain_cq(struct net_device *dev)
 713 {
 714         struct ipoib_dev_priv *priv = netdev_priv(dev);
 715         int i, n;
 716         do {
 717                 n = ib_poll_cq(priv->cq, IPOIB_NUM_WC, priv->ibwc);
 718                 for (i = 0; i < n; ++i) {
 719                         /*
 720                          * Convert any successful completions to flush
 721                          * errors to avoid passing packets up the
 722                          * stack after bringing the device down.
 723                          */
 724                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 725                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 726
 727                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 728                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 729                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 730                                 else
 731                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 732                         } else {
 733                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 734                                         ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
 735                                 else
 736                                         ipoib_ib_handle_tx_wc(dev, priv->ibwc + i);
 737                         }
 738                 }
 739         } while (n == IPOIB_NUM_WC);
 740 }
 741
 742 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
 743 {
 744         struct ipoib_dev_priv *priv = netdev_priv(dev);
 745         struct ib_qp_attr qp_attr;
 746         unsigned long begin;
 747         struct ipoib_tx_buf *tx_req;
 748         int i;
 749
 750         clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 751
 752         ipoib_cm_dev_stop(dev);
 753
 754         /*
 755          * Move our QP to the error state and then reinitialize in
 756          * when all work requests have completed or have been flushed.
 757          */
 758         qp_attr.qp_state = IB_QPS_ERR;
 759         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 760                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 761
 762         /* Wait for all sends and receives to complete */
 763         begin = jiffies;
 764
 765         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 766                 if (time_after(jiffies, begin + 5 * HZ)) {
 767                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 768                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 769
 770                         /*
 771                          * assume the HW is wedged and just free up
 772                          * all our pending work requests.
 773                          */
 774                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 775                                 tx_req = &priv->tx_ring[priv->tx_tail &
 776                                                         (ipoib_sendq_size - 1)];
 777                                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 778                                 dev_kfree_skb_any(tx_req->skb);
 779                                 ++priv->tx_tail;
 780                                 --priv->tx_outstanding;
 781                         }
 782
 783                         for (i = 0; i < ipoib_recvq_size; ++i) {
 784                                 struct ipoib_rx_buf *rx_req;
 785
 786                                 rx_req = &priv->rx_ring[i];
 787                                 if (!rx_req->skb)
 788                                         continue;
 789                                 ipoib_ud_dma_unmap_rx(priv,
 790                                                       priv->rx_ring[i].mapping);
 791                                 dev_kfree_skb_any(rx_req->skb);
 792                                 rx_req->skb = NULL;
 793                         }
 794
 795                         goto timeout;
 796                 }
 797
 798                 ipoib_drain_cq(dev);
 799
 800                 msleep(1);
 801         }
 802
 803         ipoib_dbg(priv, "All sends and receives done.\n");
 804
 805 timeout:
 806         qp_attr.qp_state = IB_QPS_RESET;
 807         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 808                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 809
 810         /* Wait for all AHs to be reaped */
 811         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 812         cancel_delayed_work(&priv->ah_reap_task);
 813         if (flush)
 814                 flush_workqueue(ipoib_workqueue);
 815
 816         begin = jiffies;
 817
 818         while (!list_empty(&priv->dead_ahs)) {
 819                 __ipoib_reap_ah(dev);
 820
 821                 if (time_after(jiffies, begin + HZ)) {
 822                         ipoib_warn(priv, "timing out; will leak address handles\n");
 823                         break;
 824                 }
 825
 826                 msleep(1);
 827         }
 828
 829         ib_req_notify_cq(priv->cq, IB_CQ_NEXT_COMP);
 830
 831         return 0;
 832 }
 833
 834 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 835 {
 836         struct ipoib_dev_priv *priv = netdev_priv(dev);
 837
 838         priv->ca = ca;
 839         priv->port = port;
 840         priv->qp = NULL;
 841
 842         if (ipoib_transport_dev_init(dev, ca)) {
 843                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 844                 return -ENODEV;
 845         }
 846
 847         if (dev->flags & IFF_UP) {
 848                 if (ipoib_ib_dev_open(dev)) {
 849                         ipoib_transport_dev_cleanup(dev);
 850                         return -ENODEV;
 851                 }
 852         }
 853
 854         return 0;
 855 }
 856
 857 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv, int pkey_event)
 858 {
 859         struct ipoib_dev_priv *cpriv;
 860         struct net_device *dev = priv->dev;
 861         u16 new_index;
 862
 863         mutex_lock(&priv->vlan_mutex);
 864
 865         /*
 866          * Flush any child interfaces too -- they might be up even if
 867          * the parent is down.
 868          */
 869         list_for_each_entry(cpriv, &priv->child_intfs, list)
 870                 __ipoib_ib_dev_flush(cpriv, pkey_event);
 871
 872         mutex_unlock(&priv->vlan_mutex);
 873
 874         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
 875                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
 876                 return;
 877         }
 878
 879         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 880                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
 881                 return;
 882         }
 883
 884         if (pkey_event) {
 885                 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
 886                         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 887                         ipoib_ib_dev_down(dev, 0);
 888                         ipoib_ib_dev_stop(dev, 0);
 889                         if (ipoib_pkey_dev_delay_open(dev))
 890                                 return;
 891                 }
 892
 893                 /* restart QP only if P_Key index is changed */
 894                 if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
 895                     new_index == priv->pkey_index) {
 896                         ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
 897                         return;
 898                 }
 899                 priv->pkey_index = new_index;
 900         }
 901
 902         ipoib_dbg(priv, "flushing\n");
 903
 904         ipoib_ib_dev_down(dev, 0);
 905
 906         if (pkey_event) {
 907                 ipoib_ib_dev_stop(dev, 0);
 908                 ipoib_ib_dev_open(dev);
 909         }
 910
 911         /*
 912          * The device could have been brought down between the start and when
 913          * we get here, don't bring it back up if it's not configured up
 914          */
 915         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 916                 ipoib_ib_dev_up(dev);
 917                 ipoib_mcast_restart_task(&priv->restart_task);
 918         }
 919 }
 920
 921 void ipoib_ib_dev_flush(struct work_struct *work)
 922 {
 923         struct ipoib_dev_priv *priv =
 924                 container_of(work, struct ipoib_dev_priv, flush_task);
 925
 926         ipoib_dbg(priv, "Flushing %s\n", priv->dev->name);
 927         __ipoib_ib_dev_flush(priv, 0);
 928 }
 929
 930 void ipoib_pkey_event(struct work_struct *work)
 931 {
 932         struct ipoib_dev_priv *priv =
 933                 container_of(work, struct ipoib_dev_priv, pkey_event_task);
 934
 935         ipoib_dbg(priv, "Flushing %s and restarting its QP\n", priv->dev->name);
 936         __ipoib_ib_dev_flush(priv, 1);
 937 }
 938
 939 void ipoib_ib_dev_cleanup(struct net_device *dev)
 940 {
 941         struct ipoib_dev_priv *priv = netdev_priv(dev);
 942
 943         ipoib_dbg(priv, "cleaning up ib_dev\n");
 944
 945         ipoib_mcast_stop_thread(dev, 1);
 946         ipoib_mcast_dev_flush(dev);
 947
 948         ipoib_transport_dev_cleanup(dev);
 949 }
 950
 951 /*
 952  * Delayed P_Key Assigment Interim Support
 953  *
 954  * The following is initial implementation of delayed P_Key assigment
 955  * mechanism. It is using the same approach implemented for the multicast
 956  * group join. The single goal of this implementation is to quickly address
 957  * Bug #2507. This implementation will probably be removed when the P_Key
 958  * change async notification is available.
 959  */
 960
 961 void ipoib_pkey_poll(struct work_struct *work)
 962 {
 963         struct ipoib_dev_priv *priv =
 964                 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
 965         struct net_device *dev = priv->dev;
 966
 967         ipoib_pkey_dev_check_presence(dev);
 968
 969         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
 970                 ipoib_open(dev);
 971         else {
 972                 mutex_lock(&pkey_mutex);
 973                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
 974                         queue_delayed_work(ipoib_workqueue,
 975                                            &priv->pkey_poll_task,
 976                                            HZ);
 977                 mutex_unlock(&pkey_mutex);
 978         }
 979 }
 980
 981 int ipoib_pkey_dev_delay_open(struct net_device *dev)
 982 {
 983         struct ipoib_dev_priv *priv = netdev_priv(dev);
 984
 985         /* Look for the interface pkey value in the IB Port P_Key table and */
 986         /* set the interface pkey assigment flag                            */
 987         ipoib_pkey_dev_check_presence(dev);
 988
 989         /* P_Key value not assigned yet - start polling */
 990         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 991                 mutex_lock(&pkey_mutex);
 992                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
 993                 queue_delayed_work(ipoib_workqueue,
 994                                    &priv->pkey_poll_task,
 995                                    HZ);
 996                 mutex_unlock(&pkey_mutex);
 997                 return 1;
 998         }
 999
1000         return 0;
1001 }