drivers/infiniband/ulp/ipoib/ipoib_ib.c

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/delay.h>
  37 #include <linux/moduleparam.h>
  38 #include <linux/dma-mapping.h>
  39 #include <linux/slab.h>
  40
  41 #include <linux/ip.h>
  42 #include <linux/tcp.h>
  43
  44 #include "ipoib.h"
  45
  46 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  47 static int data_debug_level;
  48
  49 module_param(data_debug_level, int, 0644);
  50 MODULE_PARM_DESC(data_debug_level,
  51                  "Enable data path debug tracing if > 0");
  52 #endif
  53
  54 static DEFINE_MUTEX(pkey_mutex);
  55
  56 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
  57                                  struct ib_pd *pd, struct ib_ah_attr *attr)
  58 {
  59         struct ipoib_ah *ah;
  60         struct ib_ah *vah;
  61
  62         ah = kmalloc(sizeof *ah, GFP_KERNEL);
  63         if (!ah)
  64                 return ERR_PTR(-ENOMEM);
  65
  66         ah->dev       = dev;
  67         ah->last_send = 0;
  68         kref_init(&ah->ref);
  69
  70         vah = ib_create_ah(pd, attr);
  71         if (IS_ERR(vah)) {
  72                 kfree(ah);
  73                 ah = (struct ipoib_ah *)vah;
  74         } else {
  75                 ah->ah = vah;
  76                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
  77         }
  78
  79         return ah;
  80 }
  81
  82 void ipoib_free_ah(struct kref *kref)
  83 {
  84         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
  85         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
  86
  87         unsigned long flags;
  88
  89         spin_lock_irqsave(&priv->lock, flags);
  90         list_add_tail(&ah->list, &priv->dead_ahs);
  91         spin_unlock_irqrestore(&priv->lock, flags);
  92 }
  93
  94 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
  95                                   u64 mapping[IPOIB_UD_RX_SG])
  96 {
  97         ib_dma_unmap_single(priv->ca, mapping[0],
  98                             IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
  99                             DMA_FROM_DEVICE);
 100 }
 101
 102 static int ipoib_ib_post_receive(struct net_device *dev, int id)
 103 {
 104         struct ipoib_dev_priv *priv = netdev_priv(dev);
 105         struct ib_recv_wr *bad_wr;
 106         int ret;
 107
 108         priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
 109         priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
 110         priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
 111
 112
 113         ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
 114         if (unlikely(ret)) {
 115                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
 116                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
 117                 dev_kfree_skb_any(priv->rx_ring[id].skb);
 118                 priv->rx_ring[id].skb = NULL;
 119         }
 120
 121         return ret;
 122 }
 123
 124 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
 125 {
 126         struct ipoib_dev_priv *priv = netdev_priv(dev);
 127         struct sk_buff *skb;
 128         int buf_size;
 129         u64 *mapping;
 130
 131         buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
 132
 133         skb = dev_alloc_skb(buf_size + IPOIB_ENCAP_LEN);
 134         if (unlikely(!skb))
 135                 return NULL;
 136
 137         /*
 138          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
 139          * header.  So we need 4 more bytes to get to 48 and align the
 140          * IP header to a multiple of 16.
 141          */
 142         skb_reserve(skb, 4);
 143
 144         mapping = priv->rx_ring[id].mapping;
 145         mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
 146                                        DMA_FROM_DEVICE);
 147         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
 148                 goto error;
 149
 150         priv->rx_ring[id].skb = skb;
 151         return skb;
 152 error:
 153         dev_kfree_skb_any(skb);
 154         return NULL;
 155 }
 156
 157 static int ipoib_ib_post_receives(struct net_device *dev)
 158 {
 159         struct ipoib_dev_priv *priv = netdev_priv(dev);
 160         int i;
 161
 162         for (i = 0; i < ipoib_recvq_size; ++i) {
 163                 if (!ipoib_alloc_rx_skb(dev, i)) {
 164                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 165                         return -ENOMEM;
 166                 }
 167                 if (ipoib_ib_post_receive(dev, i)) {
 168                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 169                         return -EIO;
 170                 }
 171         }
 172
 173         return 0;
 174 }
 175
 176 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 177 {
 178         struct ipoib_dev_priv *priv = netdev_priv(dev);
 179         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
 180         struct sk_buff *skb;
 181         u64 mapping[IPOIB_UD_RX_SG];
 182         union ib_gid *dgid;
 183
 184         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
 185                        wr_id, wc->status);
 186
 187         if (unlikely(wr_id >= ipoib_recvq_size)) {
 188                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
 189                            wr_id, ipoib_recvq_size);
 190                 return;
 191         }
 192
 193         skb  = priv->rx_ring[wr_id].skb;
 194
 195         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 196                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 197                         ipoib_warn(priv, "failed recv event "
 198                                    "(status=%d, wrid=%d vend_err %x)\n",
 199                                    wc->status, wr_id, wc->vendor_err);
 200                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
 201                 dev_kfree_skb_any(skb);
 202                 priv->rx_ring[wr_id].skb = NULL;
 203                 return;
 204         }
 205
 206         /*
 207          * Drop packets that this interface sent, ie multicast packets
 208          * that the HCA has replicated.
 209          */
 210         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
 211                 goto repost;
 212
 213         memcpy(mapping, priv->rx_ring[wr_id].mapping,
 214                IPOIB_UD_RX_SG * sizeof *mapping);
 215
 216         /*
 217          * If we can't allocate a new RX buffer, dump
 218          * this packet and reuse the old buffer.
 219          */
 220         if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
 221                 ++dev->stats.rx_dropped;
 222                 goto repost;
 223         }
 224
 225         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 226                        wc->byte_len, wc->slid);
 227
 228         ipoib_ud_dma_unmap_rx(priv, mapping);
 229
 230         skb_put(skb, wc->byte_len);
 231
 232         /* First byte of dgid signals multicast when 0xff */
 233         dgid = &((struct ib_grh *)skb->data)->dgid;
 234
 235         if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
 236                 skb->pkt_type = PACKET_HOST;
 237         else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
 238                 skb->pkt_type = PACKET_BROADCAST;
 239         else
 240                 skb->pkt_type = PACKET_MULTICAST;
 241
 242         skb_pull(skb, IB_GRH_BYTES);
 243
 244         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 245         skb_reset_mac_header(skb);
 246         skb_pull(skb, IPOIB_ENCAP_LEN);
 247
 248         skb->truesize = SKB_TRUESIZE(skb->len);
 249
 250         ++dev->stats.rx_packets;
 251         dev->stats.rx_bytes += skb->len;
 252
 253         skb->dev = dev;
 254         if ((dev->features & NETIF_F_RXCSUM) &&
 255                         likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
 256                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 257
 258         napi_gro_receive(&priv->napi, skb);
 259
 260 repost:
 261         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
 262                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
 263                            "for buf %d\n", wr_id);
 264 }
 265
 266 static int ipoib_dma_map_tx(struct ib_device *ca,
 267                             struct ipoib_tx_buf *tx_req)
 268 {
 269         struct sk_buff *skb = tx_req->skb;
 270         u64 *mapping = tx_req->mapping;
 271         int i;
 272         int off;
 273
 274         if (skb_headlen(skb)) {
 275                 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
 276                                                DMA_TO_DEVICE);
 277                 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
 278                         return -EIO;
 279
 280                 off = 1;
 281         } else
 282                 off = 0;
 283
 284         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 285                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 286                 mapping[i + off] = ib_dma_map_page(ca,
 287                                                  skb_frag_page(frag),
 288                                                  frag->page_offset, skb_frag_size(frag),
 289                                                  DMA_TO_DEVICE);
 290                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
 291                         goto partial_error;
 292         }
 293         return 0;
 294
 295 partial_error:
 296         for (; i > 0; --i) {
 297                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 298
 299                 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
 300         }
 301
 302         if (off)
 303                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 304
 305         return -EIO;
 306 }
 307
 308 static void ipoib_dma_unmap_tx(struct ib_device *ca,
 309                                struct ipoib_tx_buf *tx_req)
 310 {
 311         struct sk_buff *skb = tx_req->skb;
 312         u64 *mapping = tx_req->mapping;
 313         int i;
 314         int off;
 315
 316         if (skb_headlen(skb)) {
 317                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
 318                 off = 1;
 319         } else
 320                 off = 0;
 321
 322         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
 323                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 324
 325                 ib_dma_unmap_page(ca, mapping[i + off], skb_frag_size(frag),
 326                                   DMA_TO_DEVICE);
 327         }
 328 }
 329
 330 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 331 {
 332         struct ipoib_dev_priv *priv = netdev_priv(dev);
 333         unsigned int wr_id = wc->wr_id;
 334         struct ipoib_tx_buf *tx_req;
 335
 336         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 337                        wr_id, wc->status);
 338
 339         if (unlikely(wr_id >= ipoib_sendq_size)) {
 340                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 341                            wr_id, ipoib_sendq_size);
 342                 return;
 343         }
 344
 345         tx_req = &priv->tx_ring[wr_id];
 346
 347         ipoib_dma_unmap_tx(priv->ca, tx_req);
 348
 349         ++dev->stats.tx_packets;
 350         dev->stats.tx_bytes += tx_req->skb->len;
 351
 352         dev_kfree_skb_any(tx_req->skb);
 353
 354         ++priv->tx_tail;
 355         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 356             netif_queue_stopped(dev) &&
 357             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 358                 netif_wake_queue(dev);
 359
 360         if (wc->status != IB_WC_SUCCESS &&
 361             wc->status != IB_WC_WR_FLUSH_ERR)
 362                 ipoib_warn(priv, "failed send event "
 363                            "(status=%d, wrid=%d vend_err %x)\n",
 364                            wc->status, wr_id, wc->vendor_err);
 365 }
 366
 367 static int poll_tx(struct ipoib_dev_priv *priv)
 368 {
 369         int n, i;
 370
 371         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
 372         for (i = 0; i < n; ++i)
 373                 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
 374
 375         return n == MAX_SEND_CQE;
 376 }
 377
 378 int ipoib_poll(struct napi_struct *napi, int budget)
 379 {
 380         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
 381         struct net_device *dev = priv->dev;
 382         int done;
 383         int t;
 384         int n, i;
 385
 386         done  = 0;
 387
 388 poll_more:
 389         while (done < budget) {
 390                 int max = (budget - done);
 391
 392                 t = min(IPOIB_NUM_WC, max);
 393                 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
 394
 395                 for (i = 0; i < n; i++) {
 396                         struct ib_wc *wc = priv->ibwc + i;
 397
 398                         if (wc->wr_id & IPOIB_OP_RECV) {
 399                                 ++done;
 400                                 if (wc->wr_id & IPOIB_OP_CM)
 401                                         ipoib_cm_handle_rx_wc(dev, wc);
 402                                 else
 403                                         ipoib_ib_handle_rx_wc(dev, wc);
 404                         } else
 405                                 ipoib_cm_handle_tx_wc(priv->dev, wc);
 406                 }
 407
 408                 if (n != t)
 409                         break;
 410         }
 411
 412         if (done < budget) {
 413                 napi_complete(napi);
 414                 if (unlikely(ib_req_notify_cq(priv->recv_cq,
 415                                               IB_CQ_NEXT_COMP |
 416                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 417                     napi_reschedule(napi))
 418                         goto poll_more;
 419         }
 420
 421         return done;
 422 }
 423
 424 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 425 {
 426         struct net_device *dev = dev_ptr;
 427         struct ipoib_dev_priv *priv = netdev_priv(dev);
 428
 429         napi_schedule(&priv->napi);
 430 }
 431
 432 static void drain_tx_cq(struct net_device *dev)
 433 {
 434         struct ipoib_dev_priv *priv = netdev_priv(dev);
 435
 436         netif_tx_lock(dev);
 437         while (poll_tx(priv))
 438                 ; /* nothing */
 439
 440         if (netif_queue_stopped(dev))
 441                 mod_timer(&priv->poll_timer, jiffies + 1);
 442
 443         netif_tx_unlock(dev);
 444 }
 445
 446 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
 447 {
 448         struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
 449
 450         mod_timer(&priv->poll_timer, jiffies);
 451 }
 452
 453 static inline int post_send(struct ipoib_dev_priv *priv,
 454                             unsigned int wr_id,
 455                             struct ib_ah *address, u32 qpn,
 456                             struct ipoib_tx_buf *tx_req,
 457                             void *head, int hlen)
 458 {
 459         struct ib_send_wr *bad_wr;
 460         int i, off;
 461         struct sk_buff *skb = tx_req->skb;
 462         skb_frag_t *frags = skb_shinfo(skb)->frags;
 463         int nr_frags = skb_shinfo(skb)->nr_frags;
 464         u64 *mapping = tx_req->mapping;
 465
 466         if (skb_headlen(skb)) {
 467                 priv->tx_sge[0].addr         = mapping[0];
 468                 priv->tx_sge[0].length       = skb_headlen(skb);
 469                 off = 1;
 470         } else
 471                 off = 0;
 472
 473         for (i = 0; i < nr_frags; ++i) {
 474                 priv->tx_sge[i + off].addr = mapping[i + off];
 475                 priv->tx_sge[i + off].length = skb_frag_size(&frags[i]);
 476         }
 477         priv->tx_wr.num_sge          = nr_frags + off;
 478         priv->tx_wr.wr_id            = wr_id;
 479         priv->tx_wr.wr.ud.remote_qpn = qpn;
 480         priv->tx_wr.wr.ud.ah         = address;
 481
 482         if (head) {
 483                 priv->tx_wr.wr.ud.mss    = skb_shinfo(skb)->gso_size;
 484                 priv->tx_wr.wr.ud.header = head;
 485                 priv->tx_wr.wr.ud.hlen   = hlen;
 486                 priv->tx_wr.opcode       = IB_WR_LSO;
 487         } else
 488                 priv->tx_wr.opcode       = IB_WR_SEND;
 489
 490         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 491 }
 492
 493 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 494                 struct ipoib_ah *address, u32 qpn)
 495 {
 496         struct ipoib_dev_priv *priv = netdev_priv(dev);
 497         struct ipoib_tx_buf *tx_req;
 498         int hlen, rc;
 499         void *phead;
 500
 501         if (skb_is_gso(skb)) {
 502                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
 503                 phead = skb->data;
 504                 if (unlikely(!skb_pull(skb, hlen))) {
 505                         ipoib_warn(priv, "linear data too small\n");
 506                         ++dev->stats.tx_dropped;
 507                         ++dev->stats.tx_errors;
 508                         dev_kfree_skb_any(skb);
 509                         return;
 510                 }
 511         } else {
 512                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 513                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 514                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 515                         ++dev->stats.tx_dropped;
 516                         ++dev->stats.tx_errors;
 517                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 518                         return;
 519                 }
 520                 phead = NULL;
 521                 hlen  = 0;
 522         }
 523
 524         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 525                        skb->len, address, qpn);
 526
 527         /*
 528          * We put the skb into the tx_ring _before_ we call post_send()
 529          * because it's entirely possible that the completion handler will
 530          * run before we execute anything after the post_send().  That
 531          * means we have to make sure everything is properly recorded and
 532          * our state is consistent before we call post_send().
 533          */
 534         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 535         tx_req->skb = skb;
 536         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 537                 ++dev->stats.tx_errors;
 538                 dev_kfree_skb_any(skb);
 539                 return;
 540         }
 541
 542         if (skb->ip_summed == CHECKSUM_PARTIAL)
 543                 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
 544         else
 545                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
 546
 547         if (++priv->tx_outstanding == ipoib_sendq_size) {
 548                 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 549                 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
 550                         ipoib_warn(priv, "request notify on send CQ failed\n");
 551                 netif_stop_queue(dev);
 552         }
 553
 554         skb_orphan(skb);
 555         skb_dst_drop(skb);
 556
 557         rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 558                        address->ah, qpn, tx_req, phead, hlen);
 559         if (unlikely(rc)) {
 560                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 561                 ++dev->stats.tx_errors;
 562                 --priv->tx_outstanding;
 563                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 564                 dev_kfree_skb_any(skb);
 565                 if (netif_queue_stopped(dev))
 566                         netif_wake_queue(dev);
 567         } else {
 568                 dev->trans_start = jiffies;
 569
 570                 address->last_send = priv->tx_head;
 571                 ++priv->tx_head;
 572         }
 573
 574         if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
 575                 while (poll_tx(priv))
 576                         ; /* nothing */
 577 }
 578
 579 static void __ipoib_reap_ah(struct net_device *dev)
 580 {
 581         struct ipoib_dev_priv *priv = netdev_priv(dev);
 582         struct ipoib_ah *ah, *tah;
 583         LIST_HEAD(remove_list);
 584         unsigned long flags;
 585
 586         netif_tx_lock_bh(dev);
 587         spin_lock_irqsave(&priv->lock, flags);
 588
 589         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 590                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 591                         list_del(&ah->list);
 592                         ib_destroy_ah(ah->ah);
 593                         kfree(ah);
 594                 }
 595
 596         spin_unlock_irqrestore(&priv->lock, flags);
 597         netif_tx_unlock_bh(dev);
 598 }
 599
 600 void ipoib_reap_ah(struct work_struct *work)
 601 {
 602         struct ipoib_dev_priv *priv =
 603                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 604         struct net_device *dev = priv->dev;
 605
 606         __ipoib_reap_ah(dev);
 607
 608         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 609                 queue_delayed_work(priv->wq, &priv->ah_reap_task,
 610                                    round_jiffies_relative(HZ));
 611 }
 612
 613 static void ipoib_flush_ah(struct net_device *dev)
 614 {
 615         struct ipoib_dev_priv *priv = netdev_priv(dev);
 616
 617         cancel_delayed_work(&priv->ah_reap_task);
 618         flush_workqueue(priv->wq);
 619         ipoib_reap_ah(&priv->ah_reap_task.work);
 620 }
 621
 622 static void ipoib_stop_ah(struct net_device *dev)
 623 {
 624         struct ipoib_dev_priv *priv = netdev_priv(dev);
 625
 626         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 627         ipoib_flush_ah(dev);
 628 }
 629
 630 static void ipoib_ib_tx_timer_func(unsigned long ctx)
 631 {
 632         drain_tx_cq((struct net_device *)ctx);
 633 }
 634
 635 int ipoib_ib_dev_open(struct net_device *dev)
 636 {
 637         struct ipoib_dev_priv *priv = netdev_priv(dev);
 638         int ret;
 639
 640         ipoib_pkey_dev_check_presence(dev);
 641
 642         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 643                 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
 644                            (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
 645                 return -1;
 646         }
 647
 648         ret = ipoib_init_qp(dev);
 649         if (ret) {
 650                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 651                 return -1;
 652         }
 653
 654         ret = ipoib_ib_post_receives(dev);
 655         if (ret) {
 656                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 657                 goto dev_stop;
 658         }
 659
 660         ret = ipoib_cm_dev_open(dev);
 661         if (ret) {
 662                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 663                 goto dev_stop;
 664         }
 665
 666         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 667         queue_delayed_work(priv->wq, &priv->ah_reap_task,
 668                            round_jiffies_relative(HZ));
 669
 670         if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 671                 napi_enable(&priv->napi);
 672
 673         return 0;
 674 dev_stop:
 675         if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 676                 napi_enable(&priv->napi);
 677         ipoib_ib_dev_stop(dev);
 678         return -1;
 679 }
 680
 681 void ipoib_pkey_dev_check_presence(struct net_device *dev)
 682 {
 683         struct ipoib_dev_priv *priv = netdev_priv(dev);
 684
 685         if (!(priv->pkey & 0x7fff) ||
 686             ib_find_pkey(priv->ca, priv->port, priv->pkey,
 687                          &priv->pkey_index))
 688                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 689         else
 690                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 691 }
 692
 693 int ipoib_ib_dev_up(struct net_device *dev)
 694 {
 695         struct ipoib_dev_priv *priv = netdev_priv(dev);
 696
 697         ipoib_pkey_dev_check_presence(dev);
 698
 699         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 700                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 701                 return 0;
 702         }
 703
 704         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 705
 706         return ipoib_mcast_start_thread(dev);
 707 }
 708
 709 int ipoib_ib_dev_down(struct net_device *dev)
 710 {
 711         struct ipoib_dev_priv *priv = netdev_priv(dev);
 712
 713         ipoib_dbg(priv, "downing ib_dev\n");
 714
 715         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 716         netif_carrier_off(dev);
 717
 718         ipoib_mcast_stop_thread(dev);
 719         ipoib_mcast_dev_flush(dev);
 720
 721         ipoib_flush_paths(dev);
 722
 723         return 0;
 724 }
 725
 726 static int recvs_pending(struct net_device *dev)
 727 {
 728         struct ipoib_dev_priv *priv = netdev_priv(dev);
 729         int pending = 0;
 730         int i;
 731
 732         for (i = 0; i < ipoib_recvq_size; ++i)
 733                 if (priv->rx_ring[i].skb)
 734                         ++pending;
 735
 736         return pending;
 737 }
 738
 739 void ipoib_drain_cq(struct net_device *dev)
 740 {
 741         struct ipoib_dev_priv *priv = netdev_priv(dev);
 742         int i, n;
 743
 744         /*
 745          * We call completion handling routines that expect to be
 746          * called from the BH-disabled NAPI poll context, so disable
 747          * BHs here too.
 748          */
 749         local_bh_disable();
 750
 751         do {
 752                 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
 753                 for (i = 0; i < n; ++i) {
 754                         /*
 755                          * Convert any successful completions to flush
 756                          * errors to avoid passing packets up the
 757                          * stack after bringing the device down.
 758                          */
 759                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 760                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 761
 762                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 763                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 764                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 765                                 else
 766                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 767                         } else
 768                                 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
 769                 }
 770         } while (n == IPOIB_NUM_WC);
 771
 772         while (poll_tx(priv))
 773                 ; /* nothing */
 774
 775         local_bh_enable();
 776 }
 777
 778 int ipoib_ib_dev_stop(struct net_device *dev)
 779 {
 780         struct ipoib_dev_priv *priv = netdev_priv(dev);
 781         struct ib_qp_attr qp_attr;
 782         unsigned long begin;
 783         struct ipoib_tx_buf *tx_req;
 784         int i;
 785
 786         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 787                 napi_disable(&priv->napi);
 788
 789         ipoib_cm_dev_stop(dev);
 790
 791         /*
 792          * Move our QP to the error state and then reinitialize in
 793          * when all work requests have completed or have been flushed.
 794          */
 795         qp_attr.qp_state = IB_QPS_ERR;
 796         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 797                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 798
 799         /* Wait for all sends and receives to complete */
 800         begin = jiffies;
 801
 802         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 803                 if (time_after(jiffies, begin + 5 * HZ)) {
 804                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 805                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 806
 807                         /*
 808                          * assume the HW is wedged and just free up
 809                          * all our pending work requests.
 810                          */
 811                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 812                                 tx_req = &priv->tx_ring[priv->tx_tail &
 813                                                         (ipoib_sendq_size - 1)];
 814                                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 815                                 dev_kfree_skb_any(tx_req->skb);
 816                                 ++priv->tx_tail;
 817                                 --priv->tx_outstanding;
 818                         }
 819
 820                         for (i = 0; i < ipoib_recvq_size; ++i) {
 821                                 struct ipoib_rx_buf *rx_req;
 822
 823                                 rx_req = &priv->rx_ring[i];
 824                                 if (!rx_req->skb)
 825                                         continue;
 826                                 ipoib_ud_dma_unmap_rx(priv,
 827                                                       priv->rx_ring[i].mapping);
 828                                 dev_kfree_skb_any(rx_req->skb);
 829                                 rx_req->skb = NULL;
 830                         }
 831
 832                         goto timeout;
 833                 }
 834
 835                 ipoib_drain_cq(dev);
 836
 837                 msleep(1);
 838         }
 839
 840         ipoib_dbg(priv, "All sends and receives done.\n");
 841
 842 timeout:
 843         del_timer_sync(&priv->poll_timer);
 844         qp_attr.qp_state = IB_QPS_RESET;
 845         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 846                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 847
 848         ipoib_flush_ah(dev);
 849
 850         ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
 851
 852         return 0;
 853 }
 854
 855 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 856 {
 857         struct ipoib_dev_priv *priv = netdev_priv(dev);
 858
 859         priv->ca = ca;
 860         priv->port = port;
 861         priv->qp = NULL;
 862
 863         if (ipoib_transport_dev_init(dev, ca)) {
 864                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 865                 return -ENODEV;
 866         }
 867
 868         setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
 869                     (unsigned long) dev);
 870
 871         if (dev->flags & IFF_UP) {
 872                 if (ipoib_ib_dev_open(dev)) {
 873                         ipoib_transport_dev_cleanup(dev);
 874                         return -ENODEV;
 875                 }
 876         }
 877
 878         return 0;
 879 }
 880
 881 /*
 882  * Takes whatever value which is in pkey index 0 and updates priv->pkey
 883  * returns 0 if the pkey value was changed.
 884  */
 885 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
 886 {
 887         int result;
 888         u16 prev_pkey;
 889
 890         prev_pkey = priv->pkey;
 891         result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
 892         if (result) {
 893                 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
 894                            priv->port, result);
 895                 return result;
 896         }
 897
 898         priv->pkey |= 0x8000;
 899
 900         if (prev_pkey != priv->pkey) {
 901                 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
 902                           prev_pkey, priv->pkey);
 903                 /*
 904                  * Update the pkey in the broadcast address, while making sure to set
 905                  * the full membership bit, so that we join the right broadcast group.
 906                  */
 907                 priv->dev->broadcast[8] = priv->pkey >> 8;
 908                 priv->dev->broadcast[9] = priv->pkey & 0xff;
 909                 return 0;
 910         }
 911
 912         return 1;
 913 }
 914 /*
 915  * returns 0 if pkey value was found in a different slot.
 916  */
 917 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
 918 {
 919         u16 old_index = priv->pkey_index;
 920
 921         priv->pkey_index = 0;
 922         ipoib_pkey_dev_check_presence(priv->dev);
 923
 924         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
 925             (old_index == priv->pkey_index))
 926                 return 1;
 927         return 0;
 928 }
 929
 930 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
 931                                 enum ipoib_flush_level level)
 932 {
 933         struct ipoib_dev_priv *cpriv;
 934         struct net_device *dev = priv->dev;
 935         int result;
 936
 937         down_read(&priv->vlan_rwsem);
 938
 939         /*
 940          * Flush any child interfaces too -- they might be up even if
 941          * the parent is down.
 942          */
 943         list_for_each_entry(cpriv, &priv->child_intfs, list)
 944                 __ipoib_ib_dev_flush(cpriv, level);
 945
 946         up_read(&priv->vlan_rwsem);
 947
 948         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
 949             level != IPOIB_FLUSH_HEAVY) {
 950                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
 951                 return;
 952         }
 953
 954         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 955                 /* interface is down. update pkey and leave. */
 956                 if (level == IPOIB_FLUSH_HEAVY) {
 957                         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
 958                                 update_parent_pkey(priv);
 959                         else
 960                                 update_child_pkey(priv);
 961                 }
 962                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
 963                 return;
 964         }
 965
 966         if (level == IPOIB_FLUSH_HEAVY) {
 967                 /* child devices chase their origin pkey value, while non-child
 968                  * (parent) devices should always takes what present in pkey index 0
 969                  */
 970                 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
 971                         result = update_child_pkey(priv);
 972                         if (result) {
 973                                 /* restart QP only if P_Key index is changed */
 974                                 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
 975                                 return;
 976                         }
 977
 978                 } else {
 979                         result = update_parent_pkey(priv);
 980                         /* restart QP only if P_Key value changed */
 981                         if (result) {
 982                                 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
 983                                 return;
 984                         }
 985                 }
 986         }
 987
 988         if (level == IPOIB_FLUSH_LIGHT) {
 989                 ipoib_mark_paths_invalid(dev);
 990                 ipoib_mcast_dev_flush(dev);
 991                 ipoib_flush_ah(dev);
 992         }
 993
 994         if (level >= IPOIB_FLUSH_NORMAL)
 995                 ipoib_ib_dev_down(dev);
 996
 997         if (level == IPOIB_FLUSH_HEAVY) {
 998                 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 999                         ipoib_ib_dev_stop(dev);
1000                 if (ipoib_ib_dev_open(dev) != 0)
1001                         return;
1002                 if (netif_queue_stopped(dev))
1003                         netif_start_queue(dev);
1004         }
1005
1006         /*
1007          * The device could have been brought down between the start and when
1008          * we get here, don't bring it back up if it's not configured up
1009          */
1010         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1011                 if (level >= IPOIB_FLUSH_NORMAL)
1012                         ipoib_ib_dev_up(dev);
1013                 ipoib_mcast_restart_task(&priv->restart_task);
1014         }
1015 }
1016
1017 void ipoib_ib_dev_flush_light(struct work_struct *work)
1018 {
1019         struct ipoib_dev_priv *priv =
1020                 container_of(work, struct ipoib_dev_priv, flush_light);
1021
1022         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
1023 }
1024
1025 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1026 {
1027         struct ipoib_dev_priv *priv =
1028                 container_of(work, struct ipoib_dev_priv, flush_normal);
1029
1030         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1031 }
1032
1033 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1034 {
1035         struct ipoib_dev_priv *priv =
1036                 container_of(work, struct ipoib_dev_priv, flush_heavy);
1037
1038         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1039 }
1040
1041 void ipoib_ib_dev_cleanup(struct net_device *dev)
1042 {
1043         struct ipoib_dev_priv *priv = netdev_priv(dev);
1044
1045         ipoib_dbg(priv, "cleaning up ib_dev\n");
1046         /*
1047          * We must make sure there are no more (path) completions
1048          * that may wish to touch priv fields that are no longer valid
1049          */
1050         ipoib_flush_paths(dev);
1051
1052         ipoib_mcast_stop_thread(dev);
1053         ipoib_mcast_dev_flush(dev);
1054
1055         /*
1056          * All of our ah references aren't free until after
1057          * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1058          * the neighbor garbage collection is stopped and reaped.
1059          * That should all be done now, so make a final ah flush.
1060          */
1061         ipoib_stop_ah(dev);
1062
1063         ipoib_transport_dev_cleanup(dev);
1064 }
1065
1066