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/dma-mapping.h>
  38
  39 #include <linux/ip.h>
  40 #include <linux/tcp.h>
  41
  42 #include "ipoib.h"
  43
  44 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  45 static int data_debug_level;
  46
  47 module_param(data_debug_level, int, 0644);
  48 MODULE_PARM_DESC(data_debug_level,
  49                  "Enable data path debug tracing if > 0");
  50 #endif
  51
  52 static DEFINE_MUTEX(pkey_mutex);
  53
  54 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
  55                                  struct ib_pd *pd, struct ib_ah_attr *attr)
  56 {
  57         struct ipoib_ah *ah;
  58
  59         ah = kmalloc(sizeof *ah, GFP_KERNEL);
  60         if (!ah)
  61                 return NULL;
  62
  63         ah->dev       = dev;
  64         ah->last_send = 0;
  65         kref_init(&ah->ref);
  66
  67         ah->ah = ib_create_ah(pd, attr);
  68         if (IS_ERR(ah->ah)) {
  69                 kfree(ah);
  70                 ah = NULL;
  71         } else
  72                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
  73
  74         return ah;
  75 }
  76
  77 void ipoib_free_ah(struct kref *kref)
  78 {
  79         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
  80         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
  81
  82         unsigned long flags;
  83
  84         spin_lock_irqsave(&priv->lock, flags);
  85         list_add_tail(&ah->list, &priv->dead_ahs);
  86         spin_unlock_irqrestore(&priv->lock, flags);
  87 }
  88
  89 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
  90                                   u64 mapping[IPOIB_UD_RX_SG])
  91 {
  92         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
  93                 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
  94                                     DMA_FROM_DEVICE);
  95                 ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
  96                                   DMA_FROM_DEVICE);
  97         } else
  98                 ib_dma_unmap_single(priv->ca, mapping[0],
  99                                     IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
 100                                     DMA_FROM_DEVICE);
 101 }
 102
 103 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
 104                                    struct sk_buff *skb,
 105                                    unsigned int length)
 106 {
 107         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
 108                 skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
 109                 unsigned int size;
 110                 /*
 111                  * There is only two buffers needed for max_payload = 4K,
 112                  * first buf size is IPOIB_UD_HEAD_SIZE
 113                  */
 114                 skb->tail += IPOIB_UD_HEAD_SIZE;
 115                 skb->len  += length;
 116
 117                 size = length - IPOIB_UD_HEAD_SIZE;
 118
 119                 frag->size     = size;
 120                 skb->data_len += size;
 121                 skb->truesize += size;
 122         } else
 123                 skb_put(skb, length);
 124
 125 }
 126
 127 static int ipoib_ib_post_receive(struct net_device *dev, int id)
 128 {
 129         struct ipoib_dev_priv *priv = netdev_priv(dev);
 130         struct ib_recv_wr *bad_wr;
 131         int ret;
 132
 133         priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
 134         priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
 135         priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
 136
 137
 138         ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
 139         if (unlikely(ret)) {
 140                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
 141                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
 142                 dev_kfree_skb_any(priv->rx_ring[id].skb);
 143                 priv->rx_ring[id].skb = NULL;
 144         }
 145
 146         return ret;
 147 }
 148
 149 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
 150 {
 151         struct ipoib_dev_priv *priv = netdev_priv(dev);
 152         struct sk_buff *skb;
 153         int buf_size;
 154         u64 *mapping;
 155
 156         if (ipoib_ud_need_sg(priv->max_ib_mtu))
 157                 buf_size = IPOIB_UD_HEAD_SIZE;
 158         else
 159                 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
 160
 161         skb = dev_alloc_skb(buf_size + 4);
 162         if (unlikely(!skb))
 163                 return NULL;
 164
 165         /*
 166          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
 167          * header.  So we need 4 more bytes to get to 48 and align the
 168          * IP header to a multiple of 16.
 169          */
 170         skb_reserve(skb, 4);
 171
 172         mapping = priv->rx_ring[id].mapping;
 173         mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
 174                                        DMA_FROM_DEVICE);
 175         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
 176                 goto error;
 177
 178         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
 179                 struct page *page = alloc_page(GFP_ATOMIC);
 180                 if (!page)
 181                         goto partial_error;
 182                 skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
 183                 mapping[1] =
 184                         ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[0].page,
 185                                         0, PAGE_SIZE, DMA_FROM_DEVICE);
 186                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
 187                         goto partial_error;
 188         }
 189
 190         priv->rx_ring[id].skb = skb;
 191         return skb;
 192
 193 partial_error:
 194         ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
 195 error:
 196         dev_kfree_skb_any(skb);
 197         return NULL;
 198 }
 199
 200 static int ipoib_ib_post_receives(struct net_device *dev)
 201 {
 202         struct ipoib_dev_priv *priv = netdev_priv(dev);
 203         int i;
 204
 205         for (i = 0; i < ipoib_recvq_size; ++i) {
 206                 if (!ipoib_alloc_rx_skb(dev, i)) {
 207                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 208                         return -ENOMEM;
 209                 }
 210                 if (ipoib_ib_post_receive(dev, i)) {
 211                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
 212                         return -EIO;
 213                 }
 214         }
 215
 216         return 0;
 217 }
 218
 219 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 220 {
 221         struct ipoib_dev_priv *priv = netdev_priv(dev);
 222         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
 223         struct sk_buff *skb;
 224         u64 mapping[IPOIB_UD_RX_SG];
 225
 226         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
 227                        wr_id, wc->status);
 228
 229         if (unlikely(wr_id >= ipoib_recvq_size)) {
 230                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
 231                            wr_id, ipoib_recvq_size);
 232                 return;
 233         }
 234
 235         skb  = priv->rx_ring[wr_id].skb;
 236
 237         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 238                 if (wc->status != IB_WC_WR_FLUSH_ERR)
 239                         ipoib_warn(priv, "failed recv event "
 240                                    "(status=%d, wrid=%d vend_err %x)\n",
 241                                    wc->status, wr_id, wc->vendor_err);
 242                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
 243                 dev_kfree_skb_any(skb);
 244                 priv->rx_ring[wr_id].skb = NULL;
 245                 return;
 246         }
 247
 248         /*
 249          * Drop packets that this interface sent, ie multicast packets
 250          * that the HCA has replicated.
 251          */
 252         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
 253                 goto repost;
 254
 255         memcpy(mapping, priv->rx_ring[wr_id].mapping,
 256                IPOIB_UD_RX_SG * sizeof *mapping);
 257
 258         /*
 259          * If we can't allocate a new RX buffer, dump
 260          * this packet and reuse the old buffer.
 261          */
 262         if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
 263                 ++dev->stats.rx_dropped;
 264                 goto repost;
 265         }
 266
 267         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 268                        wc->byte_len, wc->slid);
 269
 270         ipoib_ud_dma_unmap_rx(priv, mapping);
 271         ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
 272
 273         skb_pull(skb, IB_GRH_BYTES);
 274
 275         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 276         skb_reset_mac_header(skb);
 277         skb_pull(skb, IPOIB_ENCAP_LEN);
 278
 279         dev->last_rx = jiffies;
 280         ++dev->stats.rx_packets;
 281         dev->stats.rx_bytes += skb->len;
 282
 283         skb->dev = dev;
 284         /* XXX get correct PACKET_ type here */
 285         skb->pkt_type = PACKET_HOST;
 286
 287         if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
 288                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 289
 290         if (dev->features & NETIF_F_LRO)
 291                 lro_receive_skb(&priv->lro.lro_mgr, skb, NULL);
 292         else
 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
 368         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 369                        wr_id, wc->status);
 370
 371         if (unlikely(wr_id >= ipoib_sendq_size)) {
 372                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
 373                            wr_id, ipoib_sendq_size);
 374                 return;
 375         }
 376
 377         tx_req = &priv->tx_ring[wr_id];
 378
 379         ipoib_dma_unmap_tx(priv->ca, tx_req);
 380
 381         ++dev->stats.tx_packets;
 382         dev->stats.tx_bytes += tx_req->skb->len;
 383
 384         dev_kfree_skb_any(tx_req->skb);
 385
 386         ++priv->tx_tail;
 387         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 388             netif_queue_stopped(dev) &&
 389             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 390                 netif_wake_queue(dev);
 391
 392         if (wc->status != IB_WC_SUCCESS &&
 393             wc->status != IB_WC_WR_FLUSH_ERR)
 394                 ipoib_warn(priv, "failed send event "
 395                            "(status=%d, wrid=%d vend_err %x)\n",
 396                            wc->status, wr_id, wc->vendor_err);
 397 }
 398
 399 static int poll_tx(struct ipoib_dev_priv *priv)
 400 {
 401         int n, i;
 402
 403         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
 404         for (i = 0; i < n; ++i)
 405                 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
 406
 407         return n == MAX_SEND_CQE;
 408 }
 409
 410 int ipoib_poll(struct napi_struct *napi, int budget)
 411 {
 412         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
 413         struct net_device *dev = priv->dev;
 414         int done;
 415         int t;
 416         int n, i;
 417
 418         done  = 0;
 419
 420 poll_more:
 421         while (done < budget) {
 422                 int max = (budget - done);
 423
 424                 t = min(IPOIB_NUM_WC, max);
 425                 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
 426
 427                 for (i = 0; i < n; i++) {
 428                         struct ib_wc *wc = priv->ibwc + i;
 429
 430                         if (wc->wr_id & IPOIB_OP_RECV) {
 431                                 ++done;
 432                                 if (wc->wr_id & IPOIB_OP_CM)
 433                                         ipoib_cm_handle_rx_wc(dev, wc);
 434                                 else
 435                                         ipoib_ib_handle_rx_wc(dev, wc);
 436                         } else
 437                                 ipoib_cm_handle_tx_wc(priv->dev, wc);
 438                 }
 439
 440                 if (n != t)
 441                         break;
 442         }
 443
 444         if (done < budget) {
 445                 if (dev->features & NETIF_F_LRO)
 446                         lro_flush_all(&priv->lro.lro_mgr);
 447
 448                 napi_complete(napi);
 449                 if (unlikely(ib_req_notify_cq(priv->recv_cq,
 450                                               IB_CQ_NEXT_COMP |
 451                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
 452                     napi_reschedule(napi))
 453                         goto poll_more;
 454         }
 455
 456         return done;
 457 }
 458
 459 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
 460 {
 461         struct net_device *dev = dev_ptr;
 462         struct ipoib_dev_priv *priv = netdev_priv(dev);
 463
 464         napi_schedule(&priv->napi);
 465 }
 466
 467 static void drain_tx_cq(struct net_device *dev)
 468 {
 469         struct ipoib_dev_priv *priv = netdev_priv(dev);
 470
 471         netif_tx_lock(dev);
 472         while (poll_tx(priv))
 473                 ; /* nothing */
 474
 475         if (netif_queue_stopped(dev))
 476                 mod_timer(&priv->poll_timer, jiffies + 1);
 477
 478         netif_tx_unlock(dev);
 479 }
 480
 481 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
 482 {
 483         struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
 484
 485         mod_timer(&priv->poll_timer, jiffies);
 486 }
 487
 488 static inline int post_send(struct ipoib_dev_priv *priv,
 489                             unsigned int wr_id,
 490                             struct ib_ah *address, u32 qpn,
 491                             struct ipoib_tx_buf *tx_req,
 492                             void *head, int hlen)
 493 {
 494         struct ib_send_wr *bad_wr;
 495         int i, off;
 496         struct sk_buff *skb = tx_req->skb;
 497         skb_frag_t *frags = skb_shinfo(skb)->frags;
 498         int nr_frags = skb_shinfo(skb)->nr_frags;
 499         u64 *mapping = tx_req->mapping;
 500
 501         if (skb_headlen(skb)) {
 502                 priv->tx_sge[0].addr         = mapping[0];
 503                 priv->tx_sge[0].length       = skb_headlen(skb);
 504                 off = 1;
 505         } else
 506                 off = 0;
 507
 508         for (i = 0; i < nr_frags; ++i) {
 509                 priv->tx_sge[i + off].addr = mapping[i + off];
 510                 priv->tx_sge[i + off].length = frags[i].size;
 511         }
 512         priv->tx_wr.num_sge          = nr_frags + off;
 513         priv->tx_wr.wr_id            = wr_id;
 514         priv->tx_wr.wr.ud.remote_qpn = qpn;
 515         priv->tx_wr.wr.ud.ah         = address;
 516
 517         if (head) {
 518                 priv->tx_wr.wr.ud.mss    = skb_shinfo(skb)->gso_size;
 519                 priv->tx_wr.wr.ud.header = head;
 520                 priv->tx_wr.wr.ud.hlen   = hlen;
 521                 priv->tx_wr.opcode       = IB_WR_LSO;
 522         } else
 523                 priv->tx_wr.opcode       = IB_WR_SEND;
 524
 525         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
 526 }
 527
 528 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 529                 struct ipoib_ah *address, u32 qpn)
 530 {
 531         struct ipoib_dev_priv *priv = netdev_priv(dev);
 532         struct ipoib_tx_buf *tx_req;
 533         int hlen;
 534         void *phead;
 535
 536         if (skb_is_gso(skb)) {
 537                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
 538                 phead = skb->data;
 539                 if (unlikely(!skb_pull(skb, hlen))) {
 540                         ipoib_warn(priv, "linear data too small\n");
 541                         ++dev->stats.tx_dropped;
 542                         ++dev->stats.tx_errors;
 543                         dev_kfree_skb_any(skb);
 544                         return;
 545                 }
 546         } else {
 547                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
 548                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 549                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
 550                         ++dev->stats.tx_dropped;
 551                         ++dev->stats.tx_errors;
 552                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
 553                         return;
 554                 }
 555                 phead = NULL;
 556                 hlen  = 0;
 557         }
 558
 559         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
 560                        skb->len, address, qpn);
 561
 562         /*
 563          * We put the skb into the tx_ring _before_ we call post_send()
 564          * because it's entirely possible that the completion handler will
 565          * run before we execute anything after the post_send().  That
 566          * means we have to make sure everything is properly recorded and
 567          * our state is consistent before we call post_send().
 568          */
 569         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
 570         tx_req->skb = skb;
 571         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 572                 ++dev->stats.tx_errors;
 573                 dev_kfree_skb_any(skb);
 574                 return;
 575         }
 576
 577         if (skb->ip_summed == CHECKSUM_PARTIAL)
 578                 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
 579         else
 580                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
 581
 582         if (++priv->tx_outstanding == ipoib_sendq_size) {
 583                 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 584                 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
 585                         ipoib_warn(priv, "request notify on send CQ failed\n");
 586                 netif_stop_queue(dev);
 587         }
 588
 589         if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
 590                                address->ah, qpn, tx_req, phead, hlen))) {
 591                 ipoib_warn(priv, "post_send failed\n");
 592                 ++dev->stats.tx_errors;
 593                 --priv->tx_outstanding;
 594                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 595                 dev_kfree_skb_any(skb);
 596                 if (netif_queue_stopped(dev))
 597                         netif_wake_queue(dev);
 598         } else {
 599                 dev->trans_start = jiffies;
 600
 601                 address->last_send = priv->tx_head;
 602                 ++priv->tx_head;
 603                 skb_orphan(skb);
 604
 605         }
 606
 607         if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
 608                 while (poll_tx(priv))
 609                         ; /* nothing */
 610 }
 611
 612 static void __ipoib_reap_ah(struct net_device *dev)
 613 {
 614         struct ipoib_dev_priv *priv = netdev_priv(dev);
 615         struct ipoib_ah *ah, *tah;
 616         LIST_HEAD(remove_list);
 617         unsigned long flags;
 618
 619         netif_tx_lock_bh(dev);
 620         spin_lock_irqsave(&priv->lock, flags);
 621
 622         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
 623                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 624                         list_del(&ah->list);
 625                         ib_destroy_ah(ah->ah);
 626                         kfree(ah);
 627                 }
 628
 629         spin_unlock_irqrestore(&priv->lock, flags);
 630         netif_tx_unlock_bh(dev);
 631 }
 632
 633 void ipoib_reap_ah(struct work_struct *work)
 634 {
 635         struct ipoib_dev_priv *priv =
 636                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
 637         struct net_device *dev = priv->dev;
 638
 639         __ipoib_reap_ah(dev);
 640
 641         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
 642                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 643                                    round_jiffies_relative(HZ));
 644 }
 645
 646 static void ipoib_ib_tx_timer_func(unsigned long ctx)
 647 {
 648         drain_tx_cq((struct net_device *)ctx);
 649 }
 650
 651 int ipoib_ib_dev_open(struct net_device *dev)
 652 {
 653         struct ipoib_dev_priv *priv = netdev_priv(dev);
 654         int ret;
 655
 656         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
 657                 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
 658                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 659                 return -1;
 660         }
 661         set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 662
 663         ret = ipoib_init_qp(dev);
 664         if (ret) {
 665                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
 666                 return -1;
 667         }
 668
 669         ret = ipoib_ib_post_receives(dev);
 670         if (ret) {
 671                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
 672                 ipoib_ib_dev_stop(dev, 1);
 673                 return -1;
 674         }
 675
 676         ret = ipoib_cm_dev_open(dev);
 677         if (ret) {
 678                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
 679                 ipoib_ib_dev_stop(dev, 1);
 680                 return -1;
 681         }
 682
 683         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
 684         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
 685                            round_jiffies_relative(HZ));
 686
 687         if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 688                 napi_enable(&priv->napi);
 689
 690         return 0;
 691 }
 692
 693 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
 694 {
 695         struct ipoib_dev_priv *priv = netdev_priv(dev);
 696         u16 pkey_index = 0;
 697
 698         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
 699                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 700         else
 701                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 702 }
 703
 704 int ipoib_ib_dev_up(struct net_device *dev)
 705 {
 706         struct ipoib_dev_priv *priv = netdev_priv(dev);
 707
 708         ipoib_pkey_dev_check_presence(dev);
 709
 710         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 711                 ipoib_dbg(priv, "PKEY is not assigned.\n");
 712                 return 0;
 713         }
 714
 715         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 716
 717         return ipoib_mcast_start_thread(dev);
 718 }
 719
 720 int ipoib_ib_dev_down(struct net_device *dev, int flush)
 721 {
 722         struct ipoib_dev_priv *priv = netdev_priv(dev);
 723
 724         ipoib_dbg(priv, "downing ib_dev\n");
 725
 726         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
 727         netif_carrier_off(dev);
 728
 729         /* Shutdown the P_Key thread if still active */
 730         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
 731                 mutex_lock(&pkey_mutex);
 732                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
 733                 cancel_delayed_work(&priv->pkey_poll_task);
 734                 mutex_unlock(&pkey_mutex);
 735                 if (flush)
 736                         flush_workqueue(ipoib_workqueue);
 737         }
 738
 739         ipoib_mcast_stop_thread(dev, flush);
 740         ipoib_mcast_dev_flush(dev);
 741
 742         ipoib_flush_paths(dev);
 743
 744         return 0;
 745 }
 746
 747 static int recvs_pending(struct net_device *dev)
 748 {
 749         struct ipoib_dev_priv *priv = netdev_priv(dev);
 750         int pending = 0;
 751         int i;
 752
 753         for (i = 0; i < ipoib_recvq_size; ++i)
 754                 if (priv->rx_ring[i].skb)
 755                         ++pending;
 756
 757         return pending;
 758 }
 759
 760 void ipoib_drain_cq(struct net_device *dev)
 761 {
 762         struct ipoib_dev_priv *priv = netdev_priv(dev);
 763         int i, n;
 764
 765         /*
 766          * We call completion handling routines that expect to be
 767          * called from the BH-disabled NAPI poll context, so disable
 768          * BHs here too.
 769          */
 770         local_bh_disable();
 771
 772         do {
 773                 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
 774                 for (i = 0; i < n; ++i) {
 775                         /*
 776                          * Convert any successful completions to flush
 777                          * errors to avoid passing packets up the
 778                          * stack after bringing the device down.
 779                          */
 780                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
 781                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
 782
 783                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
 784                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
 785                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 786                                 else
 787                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
 788                         } else
 789                                 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
 790                 }
 791         } while (n == IPOIB_NUM_WC);
 792
 793         while (poll_tx(priv))
 794                 ; /* nothing */
 795
 796         local_bh_enable();
 797 }
 798
 799 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
 800 {
 801         struct ipoib_dev_priv *priv = netdev_priv(dev);
 802         struct ib_qp_attr qp_attr;
 803         unsigned long begin;
 804         struct ipoib_tx_buf *tx_req;
 805         int i;
 806
 807         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
 808                 napi_disable(&priv->napi);
 809
 810         ipoib_cm_dev_stop(dev);
 811
 812         /*
 813          * Move our QP to the error state and then reinitialize in
 814          * when all work requests have completed or have been flushed.
 815          */
 816         qp_attr.qp_state = IB_QPS_ERR;
 817         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 818                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
 819
 820         /* Wait for all sends and receives to complete */
 821         begin = jiffies;
 822
 823         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
 824                 if (time_after(jiffies, begin + 5 * HZ)) {
 825                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
 826                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
 827
 828                         /*
 829                          * assume the HW is wedged and just free up
 830                          * all our pending work requests.
 831                          */
 832                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 833                                 tx_req = &priv->tx_ring[priv->tx_tail &
 834                                                         (ipoib_sendq_size - 1)];
 835                                 ipoib_dma_unmap_tx(priv->ca, tx_req);
 836                                 dev_kfree_skb_any(tx_req->skb);
 837                                 ++priv->tx_tail;
 838                                 --priv->tx_outstanding;
 839                         }
 840
 841                         for (i = 0; i < ipoib_recvq_size; ++i) {
 842                                 struct ipoib_rx_buf *rx_req;
 843
 844                                 rx_req = &priv->rx_ring[i];
 845                                 if (!rx_req->skb)
 846                                         continue;
 847                                 ipoib_ud_dma_unmap_rx(priv,
 848                                                       priv->rx_ring[i].mapping);
 849                                 dev_kfree_skb_any(rx_req->skb);
 850                                 rx_req->skb = NULL;
 851                         }
 852
 853                         goto timeout;
 854                 }
 855
 856                 ipoib_drain_cq(dev);
 857
 858                 msleep(1);
 859         }
 860
 861         ipoib_dbg(priv, "All sends and receives done.\n");
 862
 863 timeout:
 864         del_timer_sync(&priv->poll_timer);
 865         qp_attr.qp_state = IB_QPS_RESET;
 866         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
 867                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
 868
 869         /* Wait for all AHs to be reaped */
 870         set_bit(IPOIB_STOP_REAPER, &priv->flags);
 871         cancel_delayed_work(&priv->ah_reap_task);
 872         if (flush)
 873                 flush_workqueue(ipoib_workqueue);
 874
 875         begin = jiffies;
 876
 877         while (!list_empty(&priv->dead_ahs)) {
 878                 __ipoib_reap_ah(dev);
 879
 880                 if (time_after(jiffies, begin + HZ)) {
 881                         ipoib_warn(priv, "timing out; will leak address handles\n");
 882                         break;
 883                 }
 884
 885                 msleep(1);
 886         }
 887
 888         ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
 889
 890         return 0;
 891 }
 892
 893 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
 894 {
 895         struct ipoib_dev_priv *priv = netdev_priv(dev);
 896
 897         priv->ca = ca;
 898         priv->port = port;
 899         priv->qp = NULL;
 900
 901         if (ipoib_transport_dev_init(dev, ca)) {
 902                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
 903                 return -ENODEV;
 904         }
 905
 906         setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
 907                     (unsigned long) dev);
 908
 909         if (dev->flags & IFF_UP) {
 910                 if (ipoib_ib_dev_open(dev)) {
 911                         ipoib_transport_dev_cleanup(dev);
 912                         return -ENODEV;
 913                 }
 914         }
 915
 916         return 0;
 917 }
 918
 919 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
 920                                 enum ipoib_flush_level level)
 921 {
 922         struct ipoib_dev_priv *cpriv;
 923         struct net_device *dev = priv->dev;
 924         u16 new_index;
 925
 926         mutex_lock(&priv->vlan_mutex);
 927
 928         /*
 929          * Flush any child interfaces too -- they might be up even if
 930          * the parent is down.
 931          */
 932         list_for_each_entry(cpriv, &priv->child_intfs, list)
 933                 __ipoib_ib_dev_flush(cpriv, level);
 934
 935         mutex_unlock(&priv->vlan_mutex);
 936
 937         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
 938                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
 939                 return;
 940         }
 941
 942         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 943                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
 944                 return;
 945         }
 946
 947         if (level == IPOIB_FLUSH_HEAVY) {
 948                 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
 949                         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 950                         ipoib_ib_dev_down(dev, 0);
 951                         ipoib_ib_dev_stop(dev, 0);
 952                         if (ipoib_pkey_dev_delay_open(dev))
 953                                 return;
 954                 }
 955
 956                 /* restart QP only if P_Key index is changed */
 957                 if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
 958                     new_index == priv->pkey_index) {
 959                         ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
 960                         return;
 961                 }
 962                 priv->pkey_index = new_index;
 963         }
 964
 965         if (level == IPOIB_FLUSH_LIGHT) {
 966                 ipoib_mark_paths_invalid(dev);
 967                 ipoib_mcast_dev_flush(dev);
 968         }
 969
 970         if (level >= IPOIB_FLUSH_NORMAL)
 971                 ipoib_ib_dev_down(dev, 0);
 972
 973         if (level == IPOIB_FLUSH_HEAVY) {
 974                 ipoib_ib_dev_stop(dev, 0);
 975                 ipoib_ib_dev_open(dev);
 976         }
 977
 978         /*
 979          * The device could have been brought down between the start and when
 980          * we get here, don't bring it back up if it's not configured up
 981          */
 982         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
 983                 if (level >= IPOIB_FLUSH_NORMAL)
 984                         ipoib_ib_dev_up(dev);
 985                 ipoib_mcast_restart_task(&priv->restart_task);
 986         }
 987 }
 988
 989 void ipoib_ib_dev_flush_light(struct work_struct *work)
 990 {
 991         struct ipoib_dev_priv *priv =
 992                 container_of(work, struct ipoib_dev_priv, flush_light);
 993
 994         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
 995 }
 996
 997 void ipoib_ib_dev_flush_normal(struct work_struct *work)
 998 {
 999         struct ipoib_dev_priv *priv =
1000                 container_of(work, struct ipoib_dev_priv, flush_normal);
1001
1002         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1003 }
1004
1005 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1006 {
1007         struct ipoib_dev_priv *priv =
1008                 container_of(work, struct ipoib_dev_priv, flush_heavy);
1009
1010         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1011 }
1012
1013 void ipoib_ib_dev_cleanup(struct net_device *dev)
1014 {
1015         struct ipoib_dev_priv *priv = netdev_priv(dev);
1016
1017         ipoib_dbg(priv, "cleaning up ib_dev\n");
1018
1019         ipoib_mcast_stop_thread(dev, 1);
1020         ipoib_mcast_dev_flush(dev);
1021
1022         ipoib_transport_dev_cleanup(dev);
1023 }
1024
1025 /*
1026  * Delayed P_Key Assigment Interim Support
1027  *
1028  * The following is initial implementation of delayed P_Key assigment
1029  * mechanism. It is using the same approach implemented for the multicast
1030  * group join. The single goal of this implementation is to quickly address
1031  * Bug #2507. This implementation will probably be removed when the P_Key
1032  * change async notification is available.
1033  */
1034
1035 void ipoib_pkey_poll(struct work_struct *work)
1036 {
1037         struct ipoib_dev_priv *priv =
1038                 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1039         struct net_device *dev = priv->dev;
1040
1041         ipoib_pkey_dev_check_presence(dev);
1042
1043         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1044                 ipoib_open(dev);
1045         else {
1046                 mutex_lock(&pkey_mutex);
1047                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1048                         queue_delayed_work(ipoib_workqueue,
1049                                            &priv->pkey_poll_task,
1050                                            HZ);
1051                 mutex_unlock(&pkey_mutex);
1052         }
1053 }
1054
1055 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1056 {
1057         struct ipoib_dev_priv *priv = netdev_priv(dev);
1058
1059         /* Look for the interface pkey value in the IB Port P_Key table and */
1060         /* set the interface pkey assigment flag                            */
1061         ipoib_pkey_dev_check_presence(dev);
1062
1063         /* P_Key value not assigned yet - start polling */
1064         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1065                 mutex_lock(&pkey_mutex);
1066                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1067                 queue_delayed_work(ipoib_workqueue,
1068                                    &priv->pkey_poll_task,
1069                                    HZ);
1070                 mutex_unlock(&pkey_mutex);
1071                 return 1;
1072         }
1073
1074         return 0;
1075 }