kernel/drivers/infiniband/hw/mlx4/mad.c

   1 /*
   2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <rdma/ib_mad.h>
  34 #include <rdma/ib_smi.h>
  35 #include <rdma/ib_sa.h>
  36 #include <rdma/ib_cache.h>
  37
  38 #include <linux/random.h>
  39 #include <linux/mlx4/cmd.h>
  40 #include <linux/gfp.h>
  41 #include <rdma/ib_pma.h>
  42
  43 #include "mlx4_ib.h"
  44
  45 enum {
  46         MLX4_IB_VENDOR_CLASS1 = 0x9,
  47         MLX4_IB_VENDOR_CLASS2 = 0xa
  48 };
  49
  50 #define MLX4_TUN_SEND_WRID_SHIFT 34
  51 #define MLX4_TUN_QPN_SHIFT 32
  52 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
  53 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
  54
  55 #define MLX4_TUN_IS_RECV(a)  (((a) >>  MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
  56 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
  57
  58  /* Port mgmt change event handling */
  59
  60 #define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
  61 #define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
  62 #define NUM_IDX_IN_PKEY_TBL_BLK 32
  63 #define GUID_TBL_ENTRY_SIZE 8      /* size in bytes */
  64 #define GUID_TBL_BLK_NUM_ENTRIES 8
  65 #define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
  66
  67 struct mlx4_mad_rcv_buf {
  68         struct ib_grh grh;
  69         u8 payload[256];
  70 } __packed;
  71
  72 struct mlx4_mad_snd_buf {
  73         u8 payload[256];
  74 } __packed;
  75
  76 struct mlx4_tunnel_mad {
  77         struct ib_grh grh;
  78         struct mlx4_ib_tunnel_header hdr;
  79         struct ib_mad mad;
  80 } __packed;
  81
  82 struct mlx4_rcv_tunnel_mad {
  83         struct mlx4_rcv_tunnel_hdr hdr;
  84         struct ib_grh grh;
  85         struct ib_mad mad;
  86 } __packed;
  87
  88 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
  89 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num);
  90 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
  91                                 int block, u32 change_bitmap);
  92
  93 __be64 mlx4_ib_gen_node_guid(void)
  94 {
  95 #define NODE_GUID_HI    ((u64) (((u64)IB_OPENIB_OUI) << 40))
  96         return cpu_to_be64(NODE_GUID_HI | prandom_u32());
  97 }
  98
  99 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
 100 {
 101         return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
 102                 cpu_to_be64(0xff00000000000000LL);
 103 }
 104
 105 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
 106                  int port, const struct ib_wc *in_wc,
 107                  const struct ib_grh *in_grh,
 108                  const void *in_mad, void *response_mad)
 109 {
 110         struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
 111         void *inbox;
 112         int err;
 113         u32 in_modifier = port;
 114         u8 op_modifier = 0;
 115
 116         inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
 117         if (IS_ERR(inmailbox))
 118                 return PTR_ERR(inmailbox);
 119         inbox = inmailbox->buf;
 120
 121         outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
 122         if (IS_ERR(outmailbox)) {
 123                 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
 124                 return PTR_ERR(outmailbox);
 125         }
 126
 127         memcpy(inbox, in_mad, 256);
 128
 129         /*
 130          * Key check traps can't be generated unless we have in_wc to
 131          * tell us where to send the trap.
 132          */
 133         if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
 134                 op_modifier |= 0x1;
 135         if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
 136                 op_modifier |= 0x2;
 137         if (mlx4_is_mfunc(dev->dev) &&
 138             (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
 139                 op_modifier |= 0x8;
 140
 141         if (in_wc) {
 142                 struct {
 143                         __be32          my_qpn;
 144                         u32             reserved1;
 145                         __be32          rqpn;
 146                         u8              sl;
 147                         u8              g_path;
 148                         u16             reserved2[2];
 149                         __be16          pkey;
 150                         u32             reserved3[11];
 151                         u8              grh[40];
 152                 } *ext_info;
 153
 154                 memset(inbox + 256, 0, 256);
 155                 ext_info = inbox + 256;
 156
 157                 ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
 158                 ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
 159                 ext_info->sl     = in_wc->sl << 4;
 160                 ext_info->g_path = in_wc->dlid_path_bits |
 161                         (in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
 162                 ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
 163
 164                 if (in_grh)
 165                         memcpy(ext_info->grh, in_grh, 40);
 166
 167                 op_modifier |= 0x4;
 168
 169                 in_modifier |= in_wc->slid << 16;
 170         }
 171
 172         err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
 173                            mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
 174                            MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
 175                            (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
 176
 177         if (!err)
 178                 memcpy(response_mad, outmailbox->buf, 256);
 179
 180         mlx4_free_cmd_mailbox(dev->dev, inmailbox);
 181         mlx4_free_cmd_mailbox(dev->dev, outmailbox);
 182
 183         return err;
 184 }
 185
 186 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
 187 {
 188         struct ib_ah *new_ah;
 189         struct ib_ah_attr ah_attr;
 190         unsigned long flags;
 191
 192         if (!dev->send_agent[port_num - 1][0])
 193                 return;
 194
 195         memset(&ah_attr, 0, sizeof ah_attr);
 196         ah_attr.dlid     = lid;
 197         ah_attr.sl       = sl;
 198         ah_attr.port_num = port_num;
 199
 200         new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
 201                               &ah_attr);
 202         if (IS_ERR(new_ah))
 203                 return;
 204
 205         spin_lock_irqsave(&dev->sm_lock, flags);
 206         if (dev->sm_ah[port_num - 1])
 207                 ib_destroy_ah(dev->sm_ah[port_num - 1]);
 208         dev->sm_ah[port_num - 1] = new_ah;
 209         spin_unlock_irqrestore(&dev->sm_lock, flags);
 210 }
 211
 212 /*
 213  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
 214  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
 215  */
 216 static void smp_snoop(struct ib_device *ibdev, u8 port_num, const struct ib_mad *mad,
 217                       u16 prev_lid)
 218 {
 219         struct ib_port_info *pinfo;
 220         u16 lid;
 221         __be16 *base;
 222         u32 bn, pkey_change_bitmap;
 223         int i;
 224
 225
 226         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 227         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 228              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 229             mad->mad_hdr.method == IB_MGMT_METHOD_SET)
 230                 switch (mad->mad_hdr.attr_id) {
 231                 case IB_SMP_ATTR_PORT_INFO:
 232                         pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
 233                         lid = be16_to_cpu(pinfo->lid);
 234
 235                         update_sm_ah(dev, port_num,
 236                                      be16_to_cpu(pinfo->sm_lid),
 237                                      pinfo->neighbormtu_mastersmsl & 0xf);
 238
 239                         if (pinfo->clientrereg_resv_subnetto & 0x80)
 240                                 handle_client_rereg_event(dev, port_num);
 241
 242                         if (prev_lid != lid)
 243                                 handle_lid_change_event(dev, port_num);
 244                         break;
 245
 246                 case IB_SMP_ATTR_PKEY_TABLE:
 247                         if (!mlx4_is_mfunc(dev->dev)) {
 248                                 mlx4_ib_dispatch_event(dev, port_num,
 249                                                        IB_EVENT_PKEY_CHANGE);
 250                                 break;
 251                         }
 252
 253                         /* at this point, we are running in the master.
 254                          * Slaves do not receive SMPs.
 255                          */
 256                         bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
 257                         base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
 258                         pkey_change_bitmap = 0;
 259                         for (i = 0; i < 32; i++) {
 260                                 pr_debug("PKEY[%d] = x%x\n",
 261                                          i + bn*32, be16_to_cpu(base[i]));
 262                                 if (be16_to_cpu(base[i]) !=
 263                                     dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
 264                                         pkey_change_bitmap |= (1 << i);
 265                                         dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
 266                                                 be16_to_cpu(base[i]);
 267                                 }
 268                         }
 269                         pr_debug("PKEY Change event: port=%d, "
 270                                  "block=0x%x, change_bitmap=0x%x\n",
 271                                  port_num, bn, pkey_change_bitmap);
 272
 273                         if (pkey_change_bitmap) {
 274                                 mlx4_ib_dispatch_event(dev, port_num,
 275                                                        IB_EVENT_PKEY_CHANGE);
 276                                 if (!dev->sriov.is_going_down)
 277                                         __propagate_pkey_ev(dev, port_num, bn,
 278                                                             pkey_change_bitmap);
 279                         }
 280                         break;
 281
 282                 case IB_SMP_ATTR_GUID_INFO:
 283                         /* paravirtualized master's guid is guid 0 -- does not change */
 284                         if (!mlx4_is_master(dev->dev))
 285                                 mlx4_ib_dispatch_event(dev, port_num,
 286                                                        IB_EVENT_GID_CHANGE);
 287                         /*if master, notify relevant slaves*/
 288                         if (mlx4_is_master(dev->dev) &&
 289                             !dev->sriov.is_going_down) {
 290                                 bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
 291                                 mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
 292                                                                     (u8 *)(&((struct ib_smp *)mad)->data));
 293                                 mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
 294                                                                      (u8 *)(&((struct ib_smp *)mad)->data));
 295                         }
 296                         break;
 297
 298                 default:
 299                         break;
 300                 }
 301 }
 302
 303 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
 304                                 int block, u32 change_bitmap)
 305 {
 306         int i, ix, slave, err;
 307         int have_event = 0;
 308
 309         for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
 310                 if (slave == mlx4_master_func_num(dev->dev))
 311                         continue;
 312                 if (!mlx4_is_slave_active(dev->dev, slave))
 313                         continue;
 314
 315                 have_event = 0;
 316                 for (i = 0; i < 32; i++) {
 317                         if (!(change_bitmap & (1 << i)))
 318                                 continue;
 319                         for (ix = 0;
 320                              ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
 321                                 if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
 322                                     [ix] == i + 32 * block) {
 323                                         err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
 324                                         pr_debug("propagate_pkey_ev: slave %d,"
 325                                                  " port %d, ix %d (%d)\n",
 326                                                  slave, port_num, ix, err);
 327                                         have_event = 1;
 328                                         break;
 329                                 }
 330                         }
 331                         if (have_event)
 332                                 break;
 333                 }
 334         }
 335 }
 336
 337 static void node_desc_override(struct ib_device *dev,
 338                                struct ib_mad *mad)
 339 {
 340         unsigned long flags;
 341
 342         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 343              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 344             mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
 345             mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
 346                 spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
 347                 memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
 348                 spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
 349         }
 350 }
 351
 352 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
 353 {
 354         int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
 355         struct ib_mad_send_buf *send_buf;
 356         struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
 357         int ret;
 358         unsigned long flags;
 359
 360         if (agent) {
 361                 send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
 362                                               IB_MGMT_MAD_DATA, GFP_ATOMIC,
 363                                               IB_MGMT_BASE_VERSION);
 364                 if (IS_ERR(send_buf))
 365                         return;
 366                 /*
 367                  * We rely here on the fact that MLX QPs don't use the
 368                  * address handle after the send is posted (this is
 369                  * wrong following the IB spec strictly, but we know
 370                  * it's OK for our devices).
 371                  */
 372                 spin_lock_irqsave(&dev->sm_lock, flags);
 373                 memcpy(send_buf->mad, mad, sizeof *mad);
 374                 if ((send_buf->ah = dev->sm_ah[port_num - 1]))
 375                         ret = ib_post_send_mad(send_buf, NULL);
 376                 else
 377                         ret = -EINVAL;
 378                 spin_unlock_irqrestore(&dev->sm_lock, flags);
 379
 380                 if (ret)
 381                         ib_free_send_mad(send_buf);
 382         }
 383 }
 384
 385 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
 386                                                              struct ib_sa_mad *sa_mad)
 387 {
 388         int ret = 0;
 389
 390         /* dispatch to different sa handlers */
 391         switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
 392         case IB_SA_ATTR_MC_MEMBER_REC:
 393                 ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
 394                 break;
 395         default:
 396                 break;
 397         }
 398         return ret;
 399 }
 400
 401 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
 402 {
 403         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 404         int i;
 405
 406         for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
 407                 if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
 408                         return i;
 409         }
 410         return -1;
 411 }
 412
 413
 414 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
 415                                    u8 port, u16 pkey, u16 *ix)
 416 {
 417         int i, ret;
 418         u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
 419         u16 slot_pkey;
 420
 421         if (slave == mlx4_master_func_num(dev->dev))
 422                 return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
 423
 424         unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
 425
 426         for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
 427                 if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
 428                         continue;
 429
 430                 pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
 431
 432                 ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
 433                 if (ret)
 434                         continue;
 435                 if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
 436                         if (slot_pkey & 0x8000) {
 437                                 *ix = (u16) pkey_ix;
 438                                 return 0;
 439                         } else {
 440                                 /* take first partial pkey index found */
 441                                 if (partial_ix == 0xFF)
 442                                         partial_ix = pkey_ix;
 443                         }
 444                 }
 445         }
 446
 447         if (partial_ix < 0xFF) {
 448                 *ix = (u16) partial_ix;
 449                 return 0;
 450         }
 451
 452         return -EINVAL;
 453 }
 454
 455 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
 456                           enum ib_qp_type dest_qpt, struct ib_wc *wc,
 457                           struct ib_grh *grh, struct ib_mad *mad)
 458 {
 459         struct ib_sge list;
 460         struct ib_ud_wr wr;
 461         struct ib_send_wr *bad_wr;
 462         struct mlx4_ib_demux_pv_ctx *tun_ctx;
 463         struct mlx4_ib_demux_pv_qp *tun_qp;
 464         struct mlx4_rcv_tunnel_mad *tun_mad;
 465         struct ib_ah_attr attr;
 466         struct ib_ah *ah;
 467         struct ib_qp *src_qp = NULL;
 468         unsigned tun_tx_ix = 0;
 469         int dqpn;
 470         int ret = 0;
 471         u16 tun_pkey_ix;
 472         u16 cached_pkey;
 473         u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
 474
 475         if (dest_qpt > IB_QPT_GSI)
 476                 return -EINVAL;
 477
 478         tun_ctx = dev->sriov.demux[port-1].tun[slave];
 479
 480         /* check if proxy qp created */
 481         if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
 482                 return -EAGAIN;
 483
 484         if (!dest_qpt)
 485                 tun_qp = &tun_ctx->qp[0];
 486         else
 487                 tun_qp = &tun_ctx->qp[1];
 488
 489         /* compute P_Key index to put in tunnel header for slave */
 490         if (dest_qpt) {
 491                 u16 pkey_ix;
 492                 ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
 493                 if (ret)
 494                         return -EINVAL;
 495
 496                 ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
 497                 if (ret)
 498                         return -EINVAL;
 499                 tun_pkey_ix = pkey_ix;
 500         } else
 501                 tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
 502
 503         dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
 504
 505         /* get tunnel tx data buf for slave */
 506         src_qp = tun_qp->qp;
 507
 508         /* create ah. Just need an empty one with the port num for the post send.
 509          * The driver will set the force loopback bit in post_send */
 510         memset(&attr, 0, sizeof attr);
 511         attr.port_num = port;
 512         if (is_eth) {
 513                 memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
 514                 attr.ah_flags = IB_AH_GRH;
 515         }
 516         ah = ib_create_ah(tun_ctx->pd, &attr);
 517         if (IS_ERR(ah))
 518                 return -ENOMEM;
 519
 520         /* allocate tunnel tx buf after pass failure returns */
 521         spin_lock(&tun_qp->tx_lock);
 522         if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
 523             (MLX4_NUM_TUNNEL_BUFS - 1))
 524                 ret = -EAGAIN;
 525         else
 526                 tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
 527         spin_unlock(&tun_qp->tx_lock);
 528         if (ret)
 529                 goto out;
 530
 531         tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
 532         if (tun_qp->tx_ring[tun_tx_ix].ah)
 533                 ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
 534         tun_qp->tx_ring[tun_tx_ix].ah = ah;
 535         ib_dma_sync_single_for_cpu(&dev->ib_dev,
 536                                    tun_qp->tx_ring[tun_tx_ix].buf.map,
 537                                    sizeof (struct mlx4_rcv_tunnel_mad),
 538                                    DMA_TO_DEVICE);
 539
 540         /* copy over to tunnel buffer */
 541         if (grh)
 542                 memcpy(&tun_mad->grh, grh, sizeof *grh);
 543         memcpy(&tun_mad->mad, mad, sizeof *mad);
 544
 545         /* adjust tunnel data */
 546         tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
 547         tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
 548         tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
 549
 550         if (is_eth) {
 551                 u16 vlan = 0;
 552                 if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
 553                                                 NULL)) {
 554                         /* VST mode */
 555                         if (vlan != wc->vlan_id)
 556                                 /* Packet vlan is not the VST-assigned vlan.
 557                                  * Drop the packet.
 558                                  */
 559                                 goto out;
 560                          else
 561                                 /* Remove the vlan tag before forwarding
 562                                  * the packet to the VF.
 563                                  */
 564                                 vlan = 0xffff;
 565                 } else {
 566                         vlan = wc->vlan_id;
 567                 }
 568
 569                 tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
 570                 memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
 571                 memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
 572         } else {
 573                 tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
 574                 tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
 575         }
 576
 577         ib_dma_sync_single_for_device(&dev->ib_dev,
 578                                       tun_qp->tx_ring[tun_tx_ix].buf.map,
 579                                       sizeof (struct mlx4_rcv_tunnel_mad),
 580                                       DMA_TO_DEVICE);
 581
 582         list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
 583         list.length = sizeof (struct mlx4_rcv_tunnel_mad);
 584         list.lkey = tun_ctx->pd->local_dma_lkey;
 585
 586         wr.ah = ah;
 587         wr.port_num = port;
 588         wr.remote_qkey = IB_QP_SET_QKEY;
 589         wr.remote_qpn = dqpn;
 590         wr.wr.next = NULL;
 591         wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
 592         wr.wr.sg_list = &list;
 593         wr.wr.num_sge = 1;
 594         wr.wr.opcode = IB_WR_SEND;
 595         wr.wr.send_flags = IB_SEND_SIGNALED;
 596
 597         ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
 598 out:
 599         if (ret)
 600                 ib_destroy_ah(ah);
 601         return ret;
 602 }
 603
 604 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
 605                         struct ib_wc *wc, struct ib_grh *grh,
 606                         struct ib_mad *mad)
 607 {
 608         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 609         int err;
 610         int slave;
 611         u8 *slave_id;
 612         int is_eth = 0;
 613
 614         if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
 615                 is_eth = 0;
 616         else
 617                 is_eth = 1;
 618
 619         if (is_eth) {
 620                 if (!(wc->wc_flags & IB_WC_GRH)) {
 621                         mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
 622                         return -EINVAL;
 623                 }
 624                 if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
 625                         mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
 626                         return -EINVAL;
 627                 }
 628                 if (mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave)) {
 629                         mlx4_ib_warn(ibdev, "failed matching grh\n");
 630                         return -ENOENT;
 631                 }
 632                 if (slave >= dev->dev->caps.sqp_demux) {
 633                         mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
 634                                      slave, dev->dev->caps.sqp_demux);
 635                         return -ENOENT;
 636                 }
 637
 638                 if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
 639                         return 0;
 640
 641                 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
 642                 if (err)
 643                         pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
 644                                  slave, err);
 645                 return 0;
 646         }
 647
 648         /* Initially assume that this mad is for us */
 649         slave = mlx4_master_func_num(dev->dev);
 650
 651         /* See if the slave id is encoded in a response mad */
 652         if (mad->mad_hdr.method & 0x80) {
 653                 slave_id = (u8 *) &mad->mad_hdr.tid;
 654                 slave = *slave_id;
 655                 if (slave != 255) /*255 indicates the dom0*/
 656                         *slave_id = 0; /* remap tid */
 657         }
 658
 659         /* If a grh is present, we demux according to it */
 660         if (wc->wc_flags & IB_WC_GRH) {
 661                 slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
 662                 if (slave < 0) {
 663                         mlx4_ib_warn(ibdev, "failed matching grh\n");
 664                         return -ENOENT;
 665                 }
 666         }
 667         /* Class-specific handling */
 668         switch (mad->mad_hdr.mgmt_class) {
 669         case IB_MGMT_CLASS_SUBN_LID_ROUTED:
 670         case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
 671                 /* 255 indicates the dom0 */
 672                 if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
 673                         if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
 674                                 return -EPERM;
 675                         /* for a VF. drop unsolicited MADs */
 676                         if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
 677                                 mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
 678                                              slave, mad->mad_hdr.mgmt_class,
 679                                              mad->mad_hdr.method);
 680                                 return -EINVAL;
 681                         }
 682                 }
 683                 break;
 684         case IB_MGMT_CLASS_SUBN_ADM:
 685                 if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
 686                                              (struct ib_sa_mad *) mad))
 687                         return 0;
 688                 break;
 689         case IB_MGMT_CLASS_CM:
 690                 if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
 691                         return 0;
 692                 break;
 693         case IB_MGMT_CLASS_DEVICE_MGMT:
 694                 if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
 695                         return 0;
 696                 break;
 697         default:
 698                 /* Drop unsupported classes for slaves in tunnel mode */
 699                 if (slave != mlx4_master_func_num(dev->dev)) {
 700                         pr_debug("dropping unsupported ingress mad from class:%d "
 701                                  "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
 702                         return 0;
 703                 }
 704         }
 705         /*make sure that no slave==255 was not handled yet.*/
 706         if (slave >= dev->dev->caps.sqp_demux) {
 707                 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
 708                              slave, dev->dev->caps.sqp_demux);
 709                 return -ENOENT;
 710         }
 711
 712         err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
 713         if (err)
 714                 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
 715                          slave, err);
 716         return 0;
 717 }
 718
 719 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 720                         const struct ib_wc *in_wc, const struct ib_grh *in_grh,
 721                         const struct ib_mad *in_mad, struct ib_mad *out_mad)
 722 {
 723         u16 slid, prev_lid = 0;
 724         int err;
 725         struct ib_port_attr pattr;
 726
 727         if (in_wc && in_wc->qp->qp_num) {
 728                 pr_debug("received MAD: slid:%d sqpn:%d "
 729                         "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
 730                         in_wc->slid, in_wc->src_qp,
 731                         in_wc->dlid_path_bits,
 732                         in_wc->qp->qp_num,
 733                         in_wc->wc_flags,
 734                         in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
 735                         be16_to_cpu(in_mad->mad_hdr.attr_id));
 736                 if (in_wc->wc_flags & IB_WC_GRH) {
 737                         pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
 738                                  be64_to_cpu(in_grh->sgid.global.subnet_prefix),
 739                                  be64_to_cpu(in_grh->sgid.global.interface_id));
 740                         pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
 741                                  be64_to_cpu(in_grh->dgid.global.subnet_prefix),
 742                                  be64_to_cpu(in_grh->dgid.global.interface_id));
 743                 }
 744         }
 745
 746         slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
 747
 748         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
 749                 forward_trap(to_mdev(ibdev), port_num, in_mad);
 750                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
 751         }
 752
 753         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 754             in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 755                 if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
 756                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
 757                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
 758                         return IB_MAD_RESULT_SUCCESS;
 759
 760                 /*
 761                  * Don't process SMInfo queries -- the SMA can't handle them.
 762                  */
 763                 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
 764                         return IB_MAD_RESULT_SUCCESS;
 765         } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
 766                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
 767                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
 768                    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
 769                 if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
 770                     in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
 771                         return IB_MAD_RESULT_SUCCESS;
 772         } else
 773                 return IB_MAD_RESULT_SUCCESS;
 774
 775         if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 776              in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 777             in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
 778             in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
 779             !ib_query_port(ibdev, port_num, &pattr))
 780                 prev_lid = pattr.lid;
 781
 782         err = mlx4_MAD_IFC(to_mdev(ibdev),
 783                            (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
 784                            (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
 785                            MLX4_MAD_IFC_NET_VIEW,
 786                            port_num, in_wc, in_grh, in_mad, out_mad);
 787         if (err)
 788                 return IB_MAD_RESULT_FAILURE;
 789
 790         if (!out_mad->mad_hdr.status) {
 791                 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
 792                         smp_snoop(ibdev, port_num, in_mad, prev_lid);
 793                 /* slaves get node desc from FW */
 794                 if (!mlx4_is_slave(to_mdev(ibdev)->dev))
 795                         node_desc_override(ibdev, out_mad);
 796         }
 797
 798         /* set return bit in status of directed route responses */
 799         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 800                 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
 801
 802         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
 803                 /* no response for trap repress */
 804                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
 805
 806         return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
 807 }
 808
 809 static void edit_counter(struct mlx4_counter *cnt,
 810                                         struct ib_pma_portcounters *pma_cnt)
 811 {
 812         ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
 813                              (be64_to_cpu(cnt->tx_bytes) >> 2));
 814         ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
 815                              (be64_to_cpu(cnt->rx_bytes) >> 2));
 816         ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
 817                              be64_to_cpu(cnt->tx_frames));
 818         ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
 819                              be64_to_cpu(cnt->rx_frames));
 820 }
 821
 822 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 823                         const struct ib_wc *in_wc, const struct ib_grh *in_grh,
 824                         const struct ib_mad *in_mad, struct ib_mad *out_mad)
 825 {
 826         struct mlx4_counter counter_stats;
 827         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 828         struct counter_index *tmp_counter;
 829         int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
 830
 831         if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
 832                 return -EINVAL;
 833
 834         memset(&counter_stats, 0, sizeof(counter_stats));
 835         mutex_lock(&dev->counters_table[port_num - 1].mutex);
 836         list_for_each_entry(tmp_counter,
 837                             &dev->counters_table[port_num - 1].counters_list,
 838                             list) {
 839                 err = mlx4_get_counter_stats(dev->dev,
 840                                              tmp_counter->index,
 841                                              &counter_stats, 0);
 842                 if (err) {
 843                         err = IB_MAD_RESULT_FAILURE;
 844                         stats_avail = 0;
 845                         break;
 846                 }
 847                 stats_avail = 1;
 848         }
 849         mutex_unlock(&dev->counters_table[port_num - 1].mutex);
 850         if (stats_avail) {
 851                 memset(out_mad->data, 0, sizeof out_mad->data);
 852                 switch (counter_stats.counter_mode & 0xf) {
 853                 case 0:
 854                         edit_counter(&counter_stats,
 855                                      (void *)(out_mad->data + 40));
 856                         err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
 857                         break;
 858                 default:
 859                         err = IB_MAD_RESULT_FAILURE;
 860                 }
 861         }
 862
 863         return err;
 864 }
 865
 866 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 867                         const struct ib_wc *in_wc, const struct ib_grh *in_grh,
 868                         const struct ib_mad_hdr *in, size_t in_mad_size,
 869                         struct ib_mad_hdr *out, size_t *out_mad_size,
 870                         u16 *out_mad_pkey_index)
 871 {
 872         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 873         const struct ib_mad *in_mad = (const struct ib_mad *)in;
 874         struct ib_mad *out_mad = (struct ib_mad *)out;
 875         enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
 876
 877         if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
 878                          *out_mad_size != sizeof(*out_mad)))
 879                 return IB_MAD_RESULT_FAILURE;
 880
 881         /* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
 882          * queries, should be called only by VFs and for that specific purpose
 883          */
 884         if (link == IB_LINK_LAYER_INFINIBAND) {
 885                 if (mlx4_is_slave(dev->dev) &&
 886                     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
 887                     in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS)
 888                         return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
 889                                                 in_grh, in_mad, out_mad);
 890
 891                 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
 892                                       in_grh, in_mad, out_mad);
 893         }
 894
 895         if (link == IB_LINK_LAYER_ETHERNET)
 896                 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
 897                                         in_grh, in_mad, out_mad);
 898
 899         return -EINVAL;
 900 }
 901
 902 static void send_handler(struct ib_mad_agent *agent,
 903                          struct ib_mad_send_wc *mad_send_wc)
 904 {
 905         if (mad_send_wc->send_buf->context[0])
 906                 ib_destroy_ah(mad_send_wc->send_buf->context[0]);
 907         ib_free_send_mad(mad_send_wc->send_buf);
 908 }
 909
 910 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
 911 {
 912         struct ib_mad_agent *agent;
 913         int p, q;
 914         int ret;
 915         enum rdma_link_layer ll;
 916
 917         for (p = 0; p < dev->num_ports; ++p) {
 918                 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
 919                 for (q = 0; q <= 1; ++q) {
 920                         if (ll == IB_LINK_LAYER_INFINIBAND) {
 921                                 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
 922                                                               q ? IB_QPT_GSI : IB_QPT_SMI,
 923                                                               NULL, 0, send_handler,
 924                                                               NULL, NULL, 0);
 925                                 if (IS_ERR(agent)) {
 926                                         ret = PTR_ERR(agent);
 927                                         goto err;
 928                                 }
 929                                 dev->send_agent[p][q] = agent;
 930                         } else
 931                                 dev->send_agent[p][q] = NULL;
 932                 }
 933         }
 934
 935         return 0;
 936
 937 err:
 938         for (p = 0; p < dev->num_ports; ++p)
 939                 for (q = 0; q <= 1; ++q)
 940                         if (dev->send_agent[p][q])
 941                                 ib_unregister_mad_agent(dev->send_agent[p][q]);
 942
 943         return ret;
 944 }
 945
 946 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
 947 {
 948         struct ib_mad_agent *agent;
 949         int p, q;
 950
 951         for (p = 0; p < dev->num_ports; ++p) {
 952                 for (q = 0; q <= 1; ++q) {
 953                         agent = dev->send_agent[p][q];
 954                         if (agent) {
 955                                 dev->send_agent[p][q] = NULL;
 956                                 ib_unregister_mad_agent(agent);
 957                         }
 958                 }
 959
 960                 if (dev->sm_ah[p])
 961                         ib_destroy_ah(dev->sm_ah[p]);
 962         }
 963 }
 964
 965 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
 966 {
 967         mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
 968
 969         if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
 970                 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
 971                                             MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
 972 }
 973
 974 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
 975 {
 976         /* re-configure the alias-guid and mcg's */
 977         if (mlx4_is_master(dev->dev)) {
 978                 mlx4_ib_invalidate_all_guid_record(dev, port_num);
 979
 980                 if (!dev->sriov.is_going_down) {
 981                         mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
 982                         mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
 983                                                     MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
 984                 }
 985         }
 986         mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
 987 }
 988
 989 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
 990                               struct mlx4_eqe *eqe)
 991 {
 992         __propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
 993                             GET_MASK_FROM_EQE(eqe));
 994 }
 995
 996 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
 997                                       u32 guid_tbl_blk_num, u32 change_bitmap)
 998 {
 999         struct ib_smp *in_mad  = NULL;
1000         struct ib_smp *out_mad  = NULL;
1001         u16 i;
1002
1003         if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1004                 return;
1005
1006         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1007         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1008         if (!in_mad || !out_mad) {
1009                 mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
1010                 goto out;
1011         }
1012
1013         guid_tbl_blk_num  *= 4;
1014
1015         for (i = 0; i < 4; i++) {
1016                 if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1017                         continue;
1018                 memset(in_mad, 0, sizeof *in_mad);
1019                 memset(out_mad, 0, sizeof *out_mad);
1020
1021                 in_mad->base_version  = 1;
1022                 in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1023                 in_mad->class_version = 1;
1024                 in_mad->method        = IB_MGMT_METHOD_GET;
1025                 in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1026                 in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1027
1028                 if (mlx4_MAD_IFC(dev,
1029                                  MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1030                                  port_num, NULL, NULL, in_mad, out_mad)) {
1031                         mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1032                         goto out;
1033                 }
1034
1035                 mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1036                                                     port_num,
1037                                                     (u8 *)(&((struct ib_smp *)out_mad)->data));
1038                 mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1039                                                      port_num,
1040                                                      (u8 *)(&((struct ib_smp *)out_mad)->data));
1041         }
1042
1043 out:
1044         kfree(in_mad);
1045         kfree(out_mad);
1046         return;
1047 }
1048
1049 void handle_port_mgmt_change_event(struct work_struct *work)
1050 {
1051         struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1052         struct mlx4_ib_dev *dev = ew->ib_dev;
1053         struct mlx4_eqe *eqe = &(ew->ib_eqe);
1054         u8 port = eqe->event.port_mgmt_change.port;
1055         u32 changed_attr;
1056         u32 tbl_block;
1057         u32 change_bitmap;
1058
1059         switch (eqe->subtype) {
1060         case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1061                 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1062
1063                 /* Update the SM ah - This should be done before handling
1064                    the other changed attributes so that MADs can be sent to the SM */
1065                 if (changed_attr & MSTR_SM_CHANGE_MASK) {
1066                         u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1067                         u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1068                         update_sm_ah(dev, port, lid, sl);
1069                 }
1070
1071                 /* Check if it is a lid change event */
1072                 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1073                         handle_lid_change_event(dev, port);
1074
1075                 /* Generate GUID changed event */
1076                 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1077                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1078                         /*if master, notify all slaves*/
1079                         if (mlx4_is_master(dev->dev))
1080                                 mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1081                                                             MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1082                 }
1083
1084                 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1085                         handle_client_rereg_event(dev, port);
1086                 break;
1087
1088         case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1089                 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1090                 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1091                         propagate_pkey_ev(dev, port, eqe);
1092                 break;
1093         case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1094                 /* paravirtualized master's guid is guid 0 -- does not change */
1095                 if (!mlx4_is_master(dev->dev))
1096                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1097                 /*if master, notify relevant slaves*/
1098                 else if (!dev->sriov.is_going_down) {
1099                         tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1100                         change_bitmap = GET_MASK_FROM_EQE(eqe);
1101                         handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1102                 }
1103                 break;
1104         default:
1105                 pr_warn("Unsupported subtype 0x%x for "
1106                         "Port Management Change event\n", eqe->subtype);
1107         }
1108
1109         kfree(ew);
1110 }
1111
1112 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1113                             enum ib_event_type type)
1114 {
1115         struct ib_event event;
1116
1117         event.device            = &dev->ib_dev;
1118         event.element.port_num  = port_num;
1119         event.event             = type;
1120
1121         ib_dispatch_event(&event);
1122 }
1123
1124 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1125 {
1126         unsigned long flags;
1127         struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1128         struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1129         spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1130         if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1131                 queue_work(ctx->wq, &ctx->work);
1132         spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1133 }
1134
1135 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1136                                   struct mlx4_ib_demux_pv_qp *tun_qp,
1137                                   int index)
1138 {
1139         struct ib_sge sg_list;
1140         struct ib_recv_wr recv_wr, *bad_recv_wr;
1141         int size;
1142
1143         size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1144                 sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1145
1146         sg_list.addr = tun_qp->ring[index].map;
1147         sg_list.length = size;
1148         sg_list.lkey = ctx->pd->local_dma_lkey;
1149
1150         recv_wr.next = NULL;
1151         recv_wr.sg_list = &sg_list;
1152         recv_wr.num_sge = 1;
1153         recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1154                 MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1155         ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1156                                       size, DMA_FROM_DEVICE);
1157         return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1158 }
1159
1160 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1161                 int slave, struct ib_sa_mad *sa_mad)
1162 {
1163         int ret = 0;
1164
1165         /* dispatch to different sa handlers */
1166         switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1167         case IB_SA_ATTR_MC_MEMBER_REC:
1168                 ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1169                 break;
1170         default:
1171                 break;
1172         }
1173         return ret;
1174 }
1175
1176 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1177 {
1178         int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1179
1180         return (qpn >= proxy_start && qpn <= proxy_start + 1);
1181 }
1182
1183
1184 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1185                          enum ib_qp_type dest_qpt, u16 pkey_index,
1186                          u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1187                          u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1188 {
1189         struct ib_sge list;
1190         struct ib_ud_wr wr;
1191         struct ib_send_wr *bad_wr;
1192         struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1193         struct mlx4_ib_demux_pv_qp *sqp;
1194         struct mlx4_mad_snd_buf *sqp_mad;
1195         struct ib_ah *ah;
1196         struct ib_qp *send_qp = NULL;
1197         unsigned wire_tx_ix = 0;
1198         int ret = 0;
1199         u16 wire_pkey_ix;
1200         int src_qpnum;
1201         u8 sgid_index;
1202
1203
1204         sqp_ctx = dev->sriov.sqps[port-1];
1205
1206         /* check if proxy qp created */
1207         if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1208                 return -EAGAIN;
1209
1210         if (dest_qpt == IB_QPT_SMI) {
1211                 src_qpnum = 0;
1212                 sqp = &sqp_ctx->qp[0];
1213                 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1214         } else {
1215                 src_qpnum = 1;
1216                 sqp = &sqp_ctx->qp[1];
1217                 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1218         }
1219
1220         send_qp = sqp->qp;
1221
1222         /* create ah */
1223         sgid_index = attr->grh.sgid_index;
1224         attr->grh.sgid_index = 0;
1225         ah = ib_create_ah(sqp_ctx->pd, attr);
1226         if (IS_ERR(ah))
1227                 return -ENOMEM;
1228         attr->grh.sgid_index = sgid_index;
1229         to_mah(ah)->av.ib.gid_index = sgid_index;
1230         /* get rid of force-loopback bit */
1231         to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1232         spin_lock(&sqp->tx_lock);
1233         if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1234             (MLX4_NUM_TUNNEL_BUFS - 1))
1235                 ret = -EAGAIN;
1236         else
1237                 wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1238         spin_unlock(&sqp->tx_lock);
1239         if (ret)
1240                 goto out;
1241
1242         sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1243         if (sqp->tx_ring[wire_tx_ix].ah)
1244                 ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1245         sqp->tx_ring[wire_tx_ix].ah = ah;
1246         ib_dma_sync_single_for_cpu(&dev->ib_dev,
1247                                    sqp->tx_ring[wire_tx_ix].buf.map,
1248                                    sizeof (struct mlx4_mad_snd_buf),
1249                                    DMA_TO_DEVICE);
1250
1251         memcpy(&sqp_mad->payload, mad, sizeof *mad);
1252
1253         ib_dma_sync_single_for_device(&dev->ib_dev,
1254                                       sqp->tx_ring[wire_tx_ix].buf.map,
1255                                       sizeof (struct mlx4_mad_snd_buf),
1256                                       DMA_TO_DEVICE);
1257
1258         list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1259         list.length = sizeof (struct mlx4_mad_snd_buf);
1260         list.lkey = sqp_ctx->pd->local_dma_lkey;
1261
1262         wr.ah = ah;
1263         wr.port_num = port;
1264         wr.pkey_index = wire_pkey_ix;
1265         wr.remote_qkey = qkey;
1266         wr.remote_qpn = remote_qpn;
1267         wr.wr.next = NULL;
1268         wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1269         wr.wr.sg_list = &list;
1270         wr.wr.num_sge = 1;
1271         wr.wr.opcode = IB_WR_SEND;
1272         wr.wr.send_flags = IB_SEND_SIGNALED;
1273         if (s_mac)
1274                 memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1275         if (vlan_id < 0x1000)
1276                 vlan_id |= (attr->sl & 7) << 13;
1277         to_mah(ah)->av.eth.vlan = cpu_to_be16(vlan_id);
1278
1279
1280         ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1281 out:
1282         if (ret)
1283                 ib_destroy_ah(ah);
1284         return ret;
1285 }
1286
1287 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1288 {
1289         if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1290                 return slave;
1291         return mlx4_get_base_gid_ix(dev->dev, slave, port);
1292 }
1293
1294 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1295                                     struct ib_ah_attr *ah_attr)
1296 {
1297         if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1298                 ah_attr->grh.sgid_index = slave;
1299         else
1300                 ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1301 }
1302
1303 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1304 {
1305         struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1306         struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1307         int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1308         struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1309         struct mlx4_ib_ah ah;
1310         struct ib_ah_attr ah_attr;
1311         u8 *slave_id;
1312         int slave;
1313         int port;
1314         u16 vlan_id;
1315
1316         /* Get slave that sent this packet */
1317         if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1318             wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1319             (wc->src_qp & 0x1) != ctx->port - 1 ||
1320             wc->src_qp & 0x4) {
1321                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1322                 return;
1323         }
1324         slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1325         if (slave != ctx->slave) {
1326                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1327                              "belongs to another slave\n", wc->src_qp);
1328                 return;
1329         }
1330
1331         /* Map transaction ID */
1332         ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1333                                    sizeof (struct mlx4_tunnel_mad),
1334                                    DMA_FROM_DEVICE);
1335         switch (tunnel->mad.mad_hdr.method) {
1336         case IB_MGMT_METHOD_SET:
1337         case IB_MGMT_METHOD_GET:
1338         case IB_MGMT_METHOD_REPORT:
1339         case IB_SA_METHOD_GET_TABLE:
1340         case IB_SA_METHOD_DELETE:
1341         case IB_SA_METHOD_GET_MULTI:
1342         case IB_SA_METHOD_GET_TRACE_TBL:
1343                 slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1344                 if (*slave_id) {
1345                         mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1346                                      "class:%d slave:%d\n", *slave_id,
1347                                      tunnel->mad.mad_hdr.mgmt_class, slave);
1348                         return;
1349                 } else
1350                         *slave_id = slave;
1351         default:
1352                 /* nothing */;
1353         }
1354
1355         /* Class-specific handling */
1356         switch (tunnel->mad.mad_hdr.mgmt_class) {
1357         case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1358         case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1359                 if (slave != mlx4_master_func_num(dev->dev) &&
1360                     !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1361                         return;
1362                 break;
1363         case IB_MGMT_CLASS_SUBN_ADM:
1364                 if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1365                               (struct ib_sa_mad *) &tunnel->mad))
1366                         return;
1367                 break;
1368         case IB_MGMT_CLASS_CM:
1369                 if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1370                               (struct ib_mad *) &tunnel->mad))
1371                         return;
1372                 break;
1373         case IB_MGMT_CLASS_DEVICE_MGMT:
1374                 if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1375                     tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1376                         return;
1377                 break;
1378         default:
1379                 /* Drop unsupported classes for slaves in tunnel mode */
1380                 if (slave != mlx4_master_func_num(dev->dev)) {
1381                         mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1382                                      "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1383                         return;
1384                 }
1385         }
1386
1387         /* We are using standard ib_core services to send the mad, so generate a
1388          * stadard address handle by decoding the tunnelled mlx4_ah fields */
1389         memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1390         ah.ibah.device = ctx->ib_dev;
1391
1392         port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1393         port = mlx4_slave_convert_port(dev->dev, slave, port);
1394         if (port < 0)
1395                 return;
1396         ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1397
1398         mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1399         if (ah_attr.ah_flags & IB_AH_GRH)
1400                 fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1401
1402         memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1403         vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1404         /* if slave have default vlan use it */
1405         mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1406                                     &vlan_id, &ah_attr.sl);
1407
1408         mlx4_ib_send_to_wire(dev, slave, ctx->port,
1409                              is_proxy_qp0(dev, wc->src_qp, slave) ?
1410                              IB_QPT_SMI : IB_QPT_GSI,
1411                              be16_to_cpu(tunnel->hdr.pkey_index),
1412                              be32_to_cpu(tunnel->hdr.remote_qpn),
1413                              be32_to_cpu(tunnel->hdr.qkey),
1414                              &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1415 }
1416
1417 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1418                                  enum ib_qp_type qp_type, int is_tun)
1419 {
1420         int i;
1421         struct mlx4_ib_demux_pv_qp *tun_qp;
1422         int rx_buf_size, tx_buf_size;
1423
1424         if (qp_type > IB_QPT_GSI)
1425                 return -EINVAL;
1426
1427         tun_qp = &ctx->qp[qp_type];
1428
1429         tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1430                                GFP_KERNEL);
1431         if (!tun_qp->ring)
1432                 return -ENOMEM;
1433
1434         tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1435                                   sizeof (struct mlx4_ib_tun_tx_buf),
1436                                   GFP_KERNEL);
1437         if (!tun_qp->tx_ring) {
1438                 kfree(tun_qp->ring);
1439                 tun_qp->ring = NULL;
1440                 return -ENOMEM;
1441         }
1442
1443         if (is_tun) {
1444                 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1445                 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1446         } else {
1447                 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1448                 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1449         }
1450
1451         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1452                 tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1453                 if (!tun_qp->ring[i].addr)
1454                         goto err;
1455                 tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1456                                                         tun_qp->ring[i].addr,
1457                                                         rx_buf_size,
1458                                                         DMA_FROM_DEVICE);
1459                 if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1460                         kfree(tun_qp->ring[i].addr);
1461                         goto err;
1462                 }
1463         }
1464
1465         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1466                 tun_qp->tx_ring[i].buf.addr =
1467                         kmalloc(tx_buf_size, GFP_KERNEL);
1468                 if (!tun_qp->tx_ring[i].buf.addr)
1469                         goto tx_err;
1470                 tun_qp->tx_ring[i].buf.map =
1471                         ib_dma_map_single(ctx->ib_dev,
1472                                           tun_qp->tx_ring[i].buf.addr,
1473                                           tx_buf_size,
1474                                           DMA_TO_DEVICE);
1475                 if (ib_dma_mapping_error(ctx->ib_dev,
1476                                          tun_qp->tx_ring[i].buf.map)) {
1477                         kfree(tun_qp->tx_ring[i].buf.addr);
1478                         goto tx_err;
1479                 }
1480                 tun_qp->tx_ring[i].ah = NULL;
1481         }
1482         spin_lock_init(&tun_qp->tx_lock);
1483         tun_qp->tx_ix_head = 0;
1484         tun_qp->tx_ix_tail = 0;
1485         tun_qp->proxy_qpt = qp_type;
1486
1487         return 0;
1488
1489 tx_err:
1490         while (i > 0) {
1491                 --i;
1492                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1493                                     tx_buf_size, DMA_TO_DEVICE);
1494                 kfree(tun_qp->tx_ring[i].buf.addr);
1495         }
1496         kfree(tun_qp->tx_ring);
1497         tun_qp->tx_ring = NULL;
1498         i = MLX4_NUM_TUNNEL_BUFS;
1499 err:
1500         while (i > 0) {
1501                 --i;
1502                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1503                                     rx_buf_size, DMA_FROM_DEVICE);
1504                 kfree(tun_qp->ring[i].addr);
1505         }
1506         kfree(tun_qp->ring);
1507         tun_qp->ring = NULL;
1508         return -ENOMEM;
1509 }
1510
1511 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1512                                      enum ib_qp_type qp_type, int is_tun)
1513 {
1514         int i;
1515         struct mlx4_ib_demux_pv_qp *tun_qp;
1516         int rx_buf_size, tx_buf_size;
1517
1518         if (qp_type > IB_QPT_GSI)
1519                 return;
1520
1521         tun_qp = &ctx->qp[qp_type];
1522         if (is_tun) {
1523                 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1524                 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1525         } else {
1526                 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1527                 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1528         }
1529
1530
1531         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1532                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1533                                     rx_buf_size, DMA_FROM_DEVICE);
1534                 kfree(tun_qp->ring[i].addr);
1535         }
1536
1537         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1538                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1539                                     tx_buf_size, DMA_TO_DEVICE);
1540                 kfree(tun_qp->tx_ring[i].buf.addr);
1541                 if (tun_qp->tx_ring[i].ah)
1542                         ib_destroy_ah(tun_qp->tx_ring[i].ah);
1543         }
1544         kfree(tun_qp->tx_ring);
1545         kfree(tun_qp->ring);
1546 }
1547
1548 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1549 {
1550         struct mlx4_ib_demux_pv_ctx *ctx;
1551         struct mlx4_ib_demux_pv_qp *tun_qp;
1552         struct ib_wc wc;
1553         int ret;
1554         ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1555         ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1556
1557         while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1558                 tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1559                 if (wc.status == IB_WC_SUCCESS) {
1560                         switch (wc.opcode) {
1561                         case IB_WC_RECV:
1562                                 mlx4_ib_multiplex_mad(ctx, &wc);
1563                                 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1564                                                              wc.wr_id &
1565                                                              (MLX4_NUM_TUNNEL_BUFS - 1));
1566                                 if (ret)
1567                                         pr_err("Failed reposting tunnel "
1568                                                "buf:%lld\n", wc.wr_id);
1569                                 break;
1570                         case IB_WC_SEND:
1571                                 pr_debug("received tunnel send completion:"
1572                                          "wrid=0x%llx, status=0x%x\n",
1573                                          wc.wr_id, wc.status);
1574                                 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1575                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1576                                 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1577                                         = NULL;
1578                                 spin_lock(&tun_qp->tx_lock);
1579                                 tun_qp->tx_ix_tail++;
1580                                 spin_unlock(&tun_qp->tx_lock);
1581
1582                                 break;
1583                         default:
1584                                 break;
1585                         }
1586                 } else  {
1587                         pr_debug("mlx4_ib: completion error in tunnel: %d."
1588                                  " status = %d, wrid = 0x%llx\n",
1589                                  ctx->slave, wc.status, wc.wr_id);
1590                         if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1591                                 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1592                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1593                                 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1594                                         = NULL;
1595                                 spin_lock(&tun_qp->tx_lock);
1596                                 tun_qp->tx_ix_tail++;
1597                                 spin_unlock(&tun_qp->tx_lock);
1598                         }
1599                 }
1600         }
1601 }
1602
1603 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1604 {
1605         struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1606
1607         /* It's worse than that! He's dead, Jim! */
1608         pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1609                event->event, sqp->port);
1610 }
1611
1612 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1613                             enum ib_qp_type qp_type, int create_tun)
1614 {
1615         int i, ret;
1616         struct mlx4_ib_demux_pv_qp *tun_qp;
1617         struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1618         struct ib_qp_attr attr;
1619         int qp_attr_mask_INIT;
1620
1621         if (qp_type > IB_QPT_GSI)
1622                 return -EINVAL;
1623
1624         tun_qp = &ctx->qp[qp_type];
1625
1626         memset(&qp_init_attr, 0, sizeof qp_init_attr);
1627         qp_init_attr.init_attr.send_cq = ctx->cq;
1628         qp_init_attr.init_attr.recv_cq = ctx->cq;
1629         qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1630         qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1631         qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1632         qp_init_attr.init_attr.cap.max_send_sge = 1;
1633         qp_init_attr.init_attr.cap.max_recv_sge = 1;
1634         if (create_tun) {
1635                 qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1636                 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1637                 qp_init_attr.port = ctx->port;
1638                 qp_init_attr.slave = ctx->slave;
1639                 qp_init_attr.proxy_qp_type = qp_type;
1640                 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1641                            IB_QP_QKEY | IB_QP_PORT;
1642         } else {
1643                 qp_init_attr.init_attr.qp_type = qp_type;
1644                 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1645                 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1646         }
1647         qp_init_attr.init_attr.port_num = ctx->port;
1648         qp_init_attr.init_attr.qp_context = ctx;
1649         qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1650         tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1651         if (IS_ERR(tun_qp->qp)) {
1652                 ret = PTR_ERR(tun_qp->qp);
1653                 tun_qp->qp = NULL;
1654                 pr_err("Couldn't create %s QP (%d)\n",
1655                        create_tun ? "tunnel" : "special", ret);
1656                 return ret;
1657         }
1658
1659         memset(&attr, 0, sizeof attr);
1660         attr.qp_state = IB_QPS_INIT;
1661         ret = 0;
1662         if (create_tun)
1663                 ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1664                                               ctx->port, IB_DEFAULT_PKEY_FULL,
1665                                               &attr.pkey_index);
1666         if (ret || !create_tun)
1667                 attr.pkey_index =
1668                         to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1669         attr.qkey = IB_QP1_QKEY;
1670         attr.port_num = ctx->port;
1671         ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1672         if (ret) {
1673                 pr_err("Couldn't change %s qp state to INIT (%d)\n",
1674                        create_tun ? "tunnel" : "special", ret);
1675                 goto err_qp;
1676         }
1677         attr.qp_state = IB_QPS_RTR;
1678         ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1679         if (ret) {
1680                 pr_err("Couldn't change %s qp state to RTR (%d)\n",
1681                        create_tun ? "tunnel" : "special", ret);
1682                 goto err_qp;
1683         }
1684         attr.qp_state = IB_QPS_RTS;
1685         attr.sq_psn = 0;
1686         ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1687         if (ret) {
1688                 pr_err("Couldn't change %s qp state to RTS (%d)\n",
1689                        create_tun ? "tunnel" : "special", ret);
1690                 goto err_qp;
1691         }
1692
1693         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1694                 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1695                 if (ret) {
1696                         pr_err(" mlx4_ib_post_pv_buf error"
1697                                " (err = %d, i = %d)\n", ret, i);
1698                         goto err_qp;
1699                 }
1700         }
1701         return 0;
1702
1703 err_qp:
1704         ib_destroy_qp(tun_qp->qp);
1705         tun_qp->qp = NULL;
1706         return ret;
1707 }
1708
1709 /*
1710  * IB MAD completion callback for real SQPs
1711  */
1712 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1713 {
1714         struct mlx4_ib_demux_pv_ctx *ctx;
1715         struct mlx4_ib_demux_pv_qp *sqp;
1716         struct ib_wc wc;
1717         struct ib_grh *grh;
1718         struct ib_mad *mad;
1719
1720         ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1721         ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1722
1723         while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1724                 sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1725                 if (wc.status == IB_WC_SUCCESS) {
1726                         switch (wc.opcode) {
1727                         case IB_WC_SEND:
1728                                 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1729                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1730                                 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1731                                         = NULL;
1732                                 spin_lock(&sqp->tx_lock);
1733                                 sqp->tx_ix_tail++;
1734                                 spin_unlock(&sqp->tx_lock);
1735                                 break;
1736                         case IB_WC_RECV:
1737                                 mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1738                                                 (sqp->ring[wc.wr_id &
1739                                                 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1740                                 grh = &(((struct mlx4_mad_rcv_buf *)
1741                                                 (sqp->ring[wc.wr_id &
1742                                                 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1743                                 mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1744                                 if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1745                                                            (MLX4_NUM_TUNNEL_BUFS - 1)))
1746                                         pr_err("Failed reposting SQP "
1747                                                "buf:%lld\n", wc.wr_id);
1748                                 break;
1749                         default:
1750                                 BUG_ON(1);
1751                                 break;
1752                         }
1753                 } else  {
1754                         pr_debug("mlx4_ib: completion error in tunnel: %d."
1755                                  " status = %d, wrid = 0x%llx\n",
1756                                  ctx->slave, wc.status, wc.wr_id);
1757                         if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1758                                 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1759                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1760                                 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1761                                         = NULL;
1762                                 spin_lock(&sqp->tx_lock);
1763                                 sqp->tx_ix_tail++;
1764                                 spin_unlock(&sqp->tx_lock);
1765                         }
1766                 }
1767         }
1768 }
1769
1770 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1771                                struct mlx4_ib_demux_pv_ctx **ret_ctx)
1772 {
1773         struct mlx4_ib_demux_pv_ctx *ctx;
1774
1775         *ret_ctx = NULL;
1776         ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1777         if (!ctx) {
1778                 pr_err("failed allocating pv resource context "
1779                        "for port %d, slave %d\n", port, slave);
1780                 return -ENOMEM;
1781         }
1782
1783         ctx->ib_dev = &dev->ib_dev;
1784         ctx->port = port;
1785         ctx->slave = slave;
1786         *ret_ctx = ctx;
1787         return 0;
1788 }
1789
1790 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1791 {
1792         if (dev->sriov.demux[port - 1].tun[slave]) {
1793                 kfree(dev->sriov.demux[port - 1].tun[slave]);
1794                 dev->sriov.demux[port - 1].tun[slave] = NULL;
1795         }
1796 }
1797
1798 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1799                                int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1800 {
1801         int ret, cq_size;
1802         struct ib_cq_init_attr cq_attr = {};
1803
1804         if (ctx->state != DEMUX_PV_STATE_DOWN)
1805                 return -EEXIST;
1806
1807         ctx->state = DEMUX_PV_STATE_STARTING;
1808         /* have QP0 only if link layer is IB */
1809         if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1810             IB_LINK_LAYER_INFINIBAND)
1811                 ctx->has_smi = 1;
1812
1813         if (ctx->has_smi) {
1814                 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1815                 if (ret) {
1816                         pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1817                         goto err_out;
1818                 }
1819         }
1820
1821         ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1822         if (ret) {
1823                 pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1824                 goto err_out_qp0;
1825         }
1826
1827         cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1828         if (ctx->has_smi)
1829                 cq_size *= 2;
1830
1831         cq_attr.cqe = cq_size;
1832         ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1833                                NULL, ctx, &cq_attr);
1834         if (IS_ERR(ctx->cq)) {
1835                 ret = PTR_ERR(ctx->cq);
1836                 pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1837                 goto err_buf;
1838         }
1839
1840         ctx->pd = ib_alloc_pd(ctx->ib_dev);
1841         if (IS_ERR(ctx->pd)) {
1842                 ret = PTR_ERR(ctx->pd);
1843                 pr_err("Couldn't create tunnel PD (%d)\n", ret);
1844                 goto err_cq;
1845         }
1846
1847         if (ctx->has_smi) {
1848                 ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1849                 if (ret) {
1850                         pr_err("Couldn't create %s QP0 (%d)\n",
1851                                create_tun ? "tunnel for" : "",  ret);
1852                         goto err_pd;
1853                 }
1854         }
1855
1856         ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1857         if (ret) {
1858                 pr_err("Couldn't create %s QP1 (%d)\n",
1859                        create_tun ? "tunnel for" : "",  ret);
1860                 goto err_qp0;
1861         }
1862
1863         if (create_tun)
1864                 INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1865         else
1866                 INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1867
1868         ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1869
1870         ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1871         if (ret) {
1872                 pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1873                 goto err_wq;
1874         }
1875         ctx->state = DEMUX_PV_STATE_ACTIVE;
1876         return 0;
1877
1878 err_wq:
1879         ctx->wq = NULL;
1880         ib_destroy_qp(ctx->qp[1].qp);
1881         ctx->qp[1].qp = NULL;
1882
1883
1884 err_qp0:
1885         if (ctx->has_smi)
1886                 ib_destroy_qp(ctx->qp[0].qp);
1887         ctx->qp[0].qp = NULL;
1888
1889 err_pd:
1890         ib_dealloc_pd(ctx->pd);
1891         ctx->pd = NULL;
1892
1893 err_cq:
1894         ib_destroy_cq(ctx->cq);
1895         ctx->cq = NULL;
1896
1897 err_buf:
1898         mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1899
1900 err_out_qp0:
1901         if (ctx->has_smi)
1902                 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1903 err_out:
1904         ctx->state = DEMUX_PV_STATE_DOWN;
1905         return ret;
1906 }
1907
1908 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1909                                  struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1910 {
1911         if (!ctx)
1912                 return;
1913         if (ctx->state > DEMUX_PV_STATE_DOWN) {
1914                 ctx->state = DEMUX_PV_STATE_DOWNING;
1915                 if (flush)
1916                         flush_workqueue(ctx->wq);
1917                 if (ctx->has_smi) {
1918                         ib_destroy_qp(ctx->qp[0].qp);
1919                         ctx->qp[0].qp = NULL;
1920                         mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1921                 }
1922                 ib_destroy_qp(ctx->qp[1].qp);
1923                 ctx->qp[1].qp = NULL;
1924                 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1925                 ib_dealloc_pd(ctx->pd);
1926                 ctx->pd = NULL;
1927                 ib_destroy_cq(ctx->cq);
1928                 ctx->cq = NULL;
1929                 ctx->state = DEMUX_PV_STATE_DOWN;
1930         }
1931 }
1932
1933 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1934                                   int port, int do_init)
1935 {
1936         int ret = 0;
1937
1938         if (!do_init) {
1939                 clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1940                 /* for master, destroy real sqp resources */
1941                 if (slave == mlx4_master_func_num(dev->dev))
1942                         destroy_pv_resources(dev, slave, port,
1943                                              dev->sriov.sqps[port - 1], 1);
1944                 /* destroy the tunnel qp resources */
1945                 destroy_pv_resources(dev, slave, port,
1946                                      dev->sriov.demux[port - 1].tun[slave], 1);
1947                 return 0;
1948         }
1949
1950         /* create the tunnel qp resources */
1951         ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1952                                   dev->sriov.demux[port - 1].tun[slave]);
1953
1954         /* for master, create the real sqp resources */
1955         if (!ret && slave == mlx4_master_func_num(dev->dev))
1956                 ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1957                                           dev->sriov.sqps[port - 1]);
1958         return ret;
1959 }
1960
1961 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1962 {
1963         struct mlx4_ib_demux_work *dmxw;
1964
1965         dmxw = container_of(work, struct mlx4_ib_demux_work, work);
1966         mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
1967                                dmxw->do_init);
1968         kfree(dmxw);
1969         return;
1970 }
1971
1972 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
1973                                        struct mlx4_ib_demux_ctx *ctx,
1974                                        int port)
1975 {
1976         char name[12];
1977         int ret = 0;
1978         int i;
1979
1980         ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
1981                            sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
1982         if (!ctx->tun)
1983                 return -ENOMEM;
1984
1985         ctx->dev = dev;
1986         ctx->port = port;
1987         ctx->ib_dev = &dev->ib_dev;
1988
1989         for (i = 0;
1990              i < min(dev->dev->caps.sqp_demux,
1991              (u16)(dev->dev->persist->num_vfs + 1));
1992              i++) {
1993                 struct mlx4_active_ports actv_ports =
1994                         mlx4_get_active_ports(dev->dev, i);
1995
1996                 if (!test_bit(port - 1, actv_ports.ports))
1997                         continue;
1998
1999                 ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2000                 if (ret) {
2001                         ret = -ENOMEM;
2002                         goto err_mcg;
2003                 }
2004         }
2005
2006         ret = mlx4_ib_mcg_port_init(ctx);
2007         if (ret) {
2008                 pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2009                 goto err_mcg;
2010         }
2011
2012         snprintf(name, sizeof name, "mlx4_ibt%d", port);
2013         ctx->wq = create_singlethread_workqueue(name);
2014         if (!ctx->wq) {
2015                 pr_err("Failed to create tunnelling WQ for port %d\n", port);
2016                 ret = -ENOMEM;
2017                 goto err_wq;
2018         }
2019
2020         snprintf(name, sizeof name, "mlx4_ibud%d", port);
2021         ctx->ud_wq = create_singlethread_workqueue(name);
2022         if (!ctx->ud_wq) {
2023                 pr_err("Failed to create up/down WQ for port %d\n", port);
2024                 ret = -ENOMEM;
2025                 goto err_udwq;
2026         }
2027
2028         return 0;
2029
2030 err_udwq:
2031         destroy_workqueue(ctx->wq);
2032         ctx->wq = NULL;
2033
2034 err_wq:
2035         mlx4_ib_mcg_port_cleanup(ctx, 1);
2036 err_mcg:
2037         for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2038                 free_pv_object(dev, i, port);
2039         kfree(ctx->tun);
2040         ctx->tun = NULL;
2041         return ret;
2042 }
2043
2044 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2045 {
2046         if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2047                 sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2048                 flush_workqueue(sqp_ctx->wq);
2049                 if (sqp_ctx->has_smi) {
2050                         ib_destroy_qp(sqp_ctx->qp[0].qp);
2051                         sqp_ctx->qp[0].qp = NULL;
2052                         mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2053                 }
2054                 ib_destroy_qp(sqp_ctx->qp[1].qp);
2055                 sqp_ctx->qp[1].qp = NULL;
2056                 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2057                 ib_dealloc_pd(sqp_ctx->pd);
2058                 sqp_ctx->pd = NULL;
2059                 ib_destroy_cq(sqp_ctx->cq);
2060                 sqp_ctx->cq = NULL;
2061                 sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2062         }
2063 }
2064
2065 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2066 {
2067         int i;
2068         if (ctx) {
2069                 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2070                 mlx4_ib_mcg_port_cleanup(ctx, 1);
2071                 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2072                         if (!ctx->tun[i])
2073                                 continue;
2074                         if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2075                                 ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2076                 }
2077                 flush_workqueue(ctx->wq);
2078                 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2079                         destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2080                         free_pv_object(dev, i, ctx->port);
2081                 }
2082                 kfree(ctx->tun);
2083                 destroy_workqueue(ctx->ud_wq);
2084                 destroy_workqueue(ctx->wq);
2085         }
2086 }
2087
2088 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2089 {
2090         int i;
2091
2092         if (!mlx4_is_master(dev->dev))
2093                 return;
2094         /* initialize or tear down tunnel QPs for the master */
2095         for (i = 0; i < dev->dev->caps.num_ports; i++)
2096                 mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2097         return;
2098 }
2099
2100 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2101 {
2102         int i = 0;
2103         int err;
2104
2105         if (!mlx4_is_mfunc(dev->dev))
2106                 return 0;
2107
2108         dev->sriov.is_going_down = 0;
2109         spin_lock_init(&dev->sriov.going_down_lock);
2110         mlx4_ib_cm_paravirt_init(dev);
2111
2112         mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2113
2114         if (mlx4_is_slave(dev->dev)) {
2115                 mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2116                 return 0;
2117         }
2118
2119         for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2120                 if (i == mlx4_master_func_num(dev->dev))
2121                         mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2122                 else
2123                         mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2124         }
2125
2126         err = mlx4_ib_init_alias_guid_service(dev);
2127         if (err) {
2128                 mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2129                 goto paravirt_err;
2130         }
2131         err = mlx4_ib_device_register_sysfs(dev);
2132         if (err) {
2133                 mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2134                 goto sysfs_err;
2135         }
2136
2137         mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2138                      dev->dev->caps.sqp_demux);
2139         for (i = 0; i < dev->num_ports; i++) {
2140                 union ib_gid gid;
2141                 err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2142                 if (err)
2143                         goto demux_err;
2144                 dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2145                 err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2146                                       &dev->sriov.sqps[i]);
2147                 if (err)
2148                         goto demux_err;
2149                 err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2150                 if (err)
2151                         goto free_pv;
2152         }
2153         mlx4_ib_master_tunnels(dev, 1);
2154         return 0;
2155
2156 free_pv:
2157         free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2158 demux_err:
2159         while (--i >= 0) {
2160                 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2161                 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2162         }
2163         mlx4_ib_device_unregister_sysfs(dev);
2164
2165 sysfs_err:
2166         mlx4_ib_destroy_alias_guid_service(dev);
2167
2168 paravirt_err:
2169         mlx4_ib_cm_paravirt_clean(dev, -1);
2170
2171         return err;
2172 }
2173
2174 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2175 {
2176         int i;
2177         unsigned long flags;
2178
2179         if (!mlx4_is_mfunc(dev->dev))
2180                 return;
2181
2182         spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2183         dev->sriov.is_going_down = 1;
2184         spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2185         if (mlx4_is_master(dev->dev)) {
2186                 for (i = 0; i < dev->num_ports; i++) {
2187                         flush_workqueue(dev->sriov.demux[i].ud_wq);
2188                         mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2189                         kfree(dev->sriov.sqps[i]);
2190                         dev->sriov.sqps[i] = NULL;
2191                         mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2192                 }
2193
2194                 mlx4_ib_cm_paravirt_clean(dev, -1);
2195                 mlx4_ib_destroy_alias_guid_service(dev);
2196                 mlx4_ib_device_unregister_sysfs(dev);
2197         }
2198 }