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 /* Counters should be saturate once they reach their maximum value */
  68 #define ASSIGN_32BIT_COUNTER(counter, value) do {\
  69         if ((value) > U32_MAX)                   \
  70                 counter = cpu_to_be32(U32_MAX); \
  71         else                                     \
  72                 counter = cpu_to_be32(value);    \
  73 } while (0)
  74
  75 struct mlx4_mad_rcv_buf {
  76         struct ib_grh grh;
  77         u8 payload[256];
  78 } __packed;
  79
  80 struct mlx4_mad_snd_buf {
  81         u8 payload[256];
  82 } __packed;
  83
  84 struct mlx4_tunnel_mad {
  85         struct ib_grh grh;
  86         struct mlx4_ib_tunnel_header hdr;
  87         struct ib_mad mad;
  88 } __packed;
  89
  90 struct mlx4_rcv_tunnel_mad {
  91         struct mlx4_rcv_tunnel_hdr hdr;
  92         struct ib_grh grh;
  93         struct ib_mad mad;
  94 } __packed;
  95
  96 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
  97 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num);
  98 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
  99                                 int block, u32 change_bitmap);
 100
 101 __be64 mlx4_ib_gen_node_guid(void)
 102 {
 103 #define NODE_GUID_HI    ((u64) (((u64)IB_OPENIB_OUI) << 40))
 104         return cpu_to_be64(NODE_GUID_HI | prandom_u32());
 105 }
 106
 107 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
 108 {
 109         return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
 110                 cpu_to_be64(0xff00000000000000LL);
 111 }
 112
 113 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
 114                  int port, struct ib_wc *in_wc, struct ib_grh *in_grh,
 115                  void *in_mad, void *response_mad)
 116 {
 117         struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
 118         void *inbox;
 119         int err;
 120         u32 in_modifier = port;
 121         u8 op_modifier = 0;
 122
 123         inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
 124         if (IS_ERR(inmailbox))
 125                 return PTR_ERR(inmailbox);
 126         inbox = inmailbox->buf;
 127
 128         outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
 129         if (IS_ERR(outmailbox)) {
 130                 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
 131                 return PTR_ERR(outmailbox);
 132         }
 133
 134         memcpy(inbox, in_mad, 256);
 135
 136         /*
 137          * Key check traps can't be generated unless we have in_wc to
 138          * tell us where to send the trap.
 139          */
 140         if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
 141                 op_modifier |= 0x1;
 142         if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
 143                 op_modifier |= 0x2;
 144         if (mlx4_is_mfunc(dev->dev) &&
 145             (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
 146                 op_modifier |= 0x8;
 147
 148         if (in_wc) {
 149                 struct {
 150                         __be32          my_qpn;
 151                         u32             reserved1;
 152                         __be32          rqpn;
 153                         u8              sl;
 154                         u8              g_path;
 155                         u16             reserved2[2];
 156                         __be16          pkey;
 157                         u32             reserved3[11];
 158                         u8              grh[40];
 159                 } *ext_info;
 160
 161                 memset(inbox + 256, 0, 256);
 162                 ext_info = inbox + 256;
 163
 164                 ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
 165                 ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
 166                 ext_info->sl     = in_wc->sl << 4;
 167                 ext_info->g_path = in_wc->dlid_path_bits |
 168                         (in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
 169                 ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
 170
 171                 if (in_grh)
 172                         memcpy(ext_info->grh, in_grh, 40);
 173
 174                 op_modifier |= 0x4;
 175
 176                 in_modifier |= in_wc->slid << 16;
 177         }
 178
 179         err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
 180                            mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
 181                            MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
 182                            (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
 183
 184         if (!err)
 185                 memcpy(response_mad, outmailbox->buf, 256);
 186
 187         mlx4_free_cmd_mailbox(dev->dev, inmailbox);
 188         mlx4_free_cmd_mailbox(dev->dev, outmailbox);
 189
 190         return err;
 191 }
 192
 193 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
 194 {
 195         struct ib_ah *new_ah;
 196         struct ib_ah_attr ah_attr;
 197         unsigned long flags;
 198
 199         if (!dev->send_agent[port_num - 1][0])
 200                 return;
 201
 202         memset(&ah_attr, 0, sizeof ah_attr);
 203         ah_attr.dlid     = lid;
 204         ah_attr.sl       = sl;
 205         ah_attr.port_num = port_num;
 206
 207         new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
 208                               &ah_attr);
 209         if (IS_ERR(new_ah))
 210                 return;
 211
 212         spin_lock_irqsave(&dev->sm_lock, flags);
 213         if (dev->sm_ah[port_num - 1])
 214                 ib_destroy_ah(dev->sm_ah[port_num - 1]);
 215         dev->sm_ah[port_num - 1] = new_ah;
 216         spin_unlock_irqrestore(&dev->sm_lock, flags);
 217 }
 218
 219 /*
 220  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
 221  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
 222  */
 223 static void smp_snoop(struct ib_device *ibdev, u8 port_num, struct ib_mad *mad,
 224                       u16 prev_lid)
 225 {
 226         struct ib_port_info *pinfo;
 227         u16 lid;
 228         __be16 *base;
 229         u32 bn, pkey_change_bitmap;
 230         int i;
 231
 232
 233         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 234         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 235              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 236             mad->mad_hdr.method == IB_MGMT_METHOD_SET)
 237                 switch (mad->mad_hdr.attr_id) {
 238                 case IB_SMP_ATTR_PORT_INFO:
 239                         pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
 240                         lid = be16_to_cpu(pinfo->lid);
 241
 242                         update_sm_ah(dev, port_num,
 243                                      be16_to_cpu(pinfo->sm_lid),
 244                                      pinfo->neighbormtu_mastersmsl & 0xf);
 245
 246                         if (pinfo->clientrereg_resv_subnetto & 0x80)
 247                                 handle_client_rereg_event(dev, port_num);
 248
 249                         if (prev_lid != lid)
 250                                 handle_lid_change_event(dev, port_num);
 251                         break;
 252
 253                 case IB_SMP_ATTR_PKEY_TABLE:
 254                         if (!mlx4_is_mfunc(dev->dev)) {
 255                                 mlx4_ib_dispatch_event(dev, port_num,
 256                                                        IB_EVENT_PKEY_CHANGE);
 257                                 break;
 258                         }
 259
 260                         /* at this point, we are running in the master.
 261                          * Slaves do not receive SMPs.
 262                          */
 263                         bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
 264                         base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
 265                         pkey_change_bitmap = 0;
 266                         for (i = 0; i < 32; i++) {
 267                                 pr_debug("PKEY[%d] = x%x\n",
 268                                          i + bn*32, be16_to_cpu(base[i]));
 269                                 if (be16_to_cpu(base[i]) !=
 270                                     dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
 271                                         pkey_change_bitmap |= (1 << i);
 272                                         dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
 273                                                 be16_to_cpu(base[i]);
 274                                 }
 275                         }
 276                         pr_debug("PKEY Change event: port=%d, "
 277                                  "block=0x%x, change_bitmap=0x%x\n",
 278                                  port_num, bn, pkey_change_bitmap);
 279
 280                         if (pkey_change_bitmap) {
 281                                 mlx4_ib_dispatch_event(dev, port_num,
 282                                                        IB_EVENT_PKEY_CHANGE);
 283                                 if (!dev->sriov.is_going_down)
 284                                         __propagate_pkey_ev(dev, port_num, bn,
 285                                                             pkey_change_bitmap);
 286                         }
 287                         break;
 288
 289                 case IB_SMP_ATTR_GUID_INFO:
 290                         /* paravirtualized master's guid is guid 0 -- does not change */
 291                         if (!mlx4_is_master(dev->dev))
 292                                 mlx4_ib_dispatch_event(dev, port_num,
 293                                                        IB_EVENT_GID_CHANGE);
 294                         /*if master, notify relevant slaves*/
 295                         if (mlx4_is_master(dev->dev) &&
 296                             !dev->sriov.is_going_down) {
 297                                 bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
 298                                 mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
 299                                                                     (u8 *)(&((struct ib_smp *)mad)->data));
 300                                 mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
 301                                                                      (u8 *)(&((struct ib_smp *)mad)->data));
 302                         }
 303                         break;
 304
 305                 default:
 306                         break;
 307                 }
 308 }
 309
 310 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
 311                                 int block, u32 change_bitmap)
 312 {
 313         int i, ix, slave, err;
 314         int have_event = 0;
 315
 316         for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
 317                 if (slave == mlx4_master_func_num(dev->dev))
 318                         continue;
 319                 if (!mlx4_is_slave_active(dev->dev, slave))
 320                         continue;
 321
 322                 have_event = 0;
 323                 for (i = 0; i < 32; i++) {
 324                         if (!(change_bitmap & (1 << i)))
 325                                 continue;
 326                         for (ix = 0;
 327                              ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
 328                                 if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
 329                                     [ix] == i + 32 * block) {
 330                                         err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
 331                                         pr_debug("propagate_pkey_ev: slave %d,"
 332                                                  " port %d, ix %d (%d)\n",
 333                                                  slave, port_num, ix, err);
 334                                         have_event = 1;
 335                                         break;
 336                                 }
 337                         }
 338                         if (have_event)
 339                                 break;
 340                 }
 341         }
 342 }
 343
 344 static void node_desc_override(struct ib_device *dev,
 345                                struct ib_mad *mad)
 346 {
 347         unsigned long flags;
 348
 349         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 350              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 351             mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
 352             mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
 353                 spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
 354                 memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
 355                 spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
 356         }
 357 }
 358
 359 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, struct ib_mad *mad)
 360 {
 361         int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
 362         struct ib_mad_send_buf *send_buf;
 363         struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
 364         int ret;
 365         unsigned long flags;
 366
 367         if (agent) {
 368                 send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
 369                                               IB_MGMT_MAD_DATA, GFP_ATOMIC);
 370                 if (IS_ERR(send_buf))
 371                         return;
 372                 /*
 373                  * We rely here on the fact that MLX QPs don't use the
 374                  * address handle after the send is posted (this is
 375                  * wrong following the IB spec strictly, but we know
 376                  * it's OK for our devices).
 377                  */
 378                 spin_lock_irqsave(&dev->sm_lock, flags);
 379                 memcpy(send_buf->mad, mad, sizeof *mad);
 380                 if ((send_buf->ah = dev->sm_ah[port_num - 1]))
 381                         ret = ib_post_send_mad(send_buf, NULL);
 382                 else
 383                         ret = -EINVAL;
 384                 spin_unlock_irqrestore(&dev->sm_lock, flags);
 385
 386                 if (ret)
 387                         ib_free_send_mad(send_buf);
 388         }
 389 }
 390
 391 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
 392                                                              struct ib_sa_mad *sa_mad)
 393 {
 394         int ret = 0;
 395
 396         /* dispatch to different sa handlers */
 397         switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
 398         case IB_SA_ATTR_MC_MEMBER_REC:
 399                 ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
 400                 break;
 401         default:
 402                 break;
 403         }
 404         return ret;
 405 }
 406
 407 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
 408 {
 409         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 410         int i;
 411
 412         for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
 413                 if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
 414                         return i;
 415         }
 416         return -1;
 417 }
 418
 419
 420 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
 421                                    u8 port, u16 pkey, u16 *ix)
 422 {
 423         int i, ret;
 424         u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
 425         u16 slot_pkey;
 426
 427         if (slave == mlx4_master_func_num(dev->dev))
 428                 return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
 429
 430         unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
 431
 432         for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
 433                 if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
 434                         continue;
 435
 436                 pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
 437
 438                 ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
 439                 if (ret)
 440                         continue;
 441                 if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
 442                         if (slot_pkey & 0x8000) {
 443                                 *ix = (u16) pkey_ix;
 444                                 return 0;
 445                         } else {
 446                                 /* take first partial pkey index found */
 447                                 if (partial_ix == 0xFF)
 448                                         partial_ix = pkey_ix;
 449                         }
 450                 }
 451         }
 452
 453         if (partial_ix < 0xFF) {
 454                 *ix = (u16) partial_ix;
 455                 return 0;
 456         }
 457
 458         return -EINVAL;
 459 }
 460
 461 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
 462                           enum ib_qp_type dest_qpt, struct ib_wc *wc,
 463                           struct ib_grh *grh, struct ib_mad *mad)
 464 {
 465         struct ib_sge list;
 466         struct ib_send_wr wr, *bad_wr;
 467         struct mlx4_ib_demux_pv_ctx *tun_ctx;
 468         struct mlx4_ib_demux_pv_qp *tun_qp;
 469         struct mlx4_rcv_tunnel_mad *tun_mad;
 470         struct ib_ah_attr attr;
 471         struct ib_ah *ah;
 472         struct ib_qp *src_qp = NULL;
 473         unsigned tun_tx_ix = 0;
 474         int dqpn;
 475         int ret = 0;
 476         u16 tun_pkey_ix;
 477         u16 cached_pkey;
 478         u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
 479
 480         if (dest_qpt > IB_QPT_GSI)
 481                 return -EINVAL;
 482
 483         tun_ctx = dev->sriov.demux[port-1].tun[slave];
 484
 485         /* check if proxy qp created */
 486         if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
 487                 return -EAGAIN;
 488
 489         if (!dest_qpt)
 490                 tun_qp = &tun_ctx->qp[0];
 491         else
 492                 tun_qp = &tun_ctx->qp[1];
 493
 494         /* compute P_Key index to put in tunnel header for slave */
 495         if (dest_qpt) {
 496                 u16 pkey_ix;
 497                 ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
 498                 if (ret)
 499                         return -EINVAL;
 500
 501                 ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
 502                 if (ret)
 503                         return -EINVAL;
 504                 tun_pkey_ix = pkey_ix;
 505         } else
 506                 tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
 507
 508         dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
 509
 510         /* get tunnel tx data buf for slave */
 511         src_qp = tun_qp->qp;
 512
 513         /* create ah. Just need an empty one with the port num for the post send.
 514          * The driver will set the force loopback bit in post_send */
 515         memset(&attr, 0, sizeof attr);
 516         attr.port_num = port;
 517         if (is_eth) {
 518                 memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
 519                 attr.ah_flags = IB_AH_GRH;
 520         }
 521         ah = ib_create_ah(tun_ctx->pd, &attr);
 522         if (IS_ERR(ah))
 523                 return -ENOMEM;
 524
 525         /* allocate tunnel tx buf after pass failure returns */
 526         spin_lock(&tun_qp->tx_lock);
 527         if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
 528             (MLX4_NUM_TUNNEL_BUFS - 1))
 529                 ret = -EAGAIN;
 530         else
 531                 tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
 532         spin_unlock(&tun_qp->tx_lock);
 533         if (ret)
 534                 goto out;
 535
 536         tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
 537         if (tun_qp->tx_ring[tun_tx_ix].ah)
 538                 ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
 539         tun_qp->tx_ring[tun_tx_ix].ah = ah;
 540         ib_dma_sync_single_for_cpu(&dev->ib_dev,
 541                                    tun_qp->tx_ring[tun_tx_ix].buf.map,
 542                                    sizeof (struct mlx4_rcv_tunnel_mad),
 543                                    DMA_TO_DEVICE);
 544
 545         /* copy over to tunnel buffer */
 546         if (grh)
 547                 memcpy(&tun_mad->grh, grh, sizeof *grh);
 548         memcpy(&tun_mad->mad, mad, sizeof *mad);
 549
 550         /* adjust tunnel data */
 551         tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
 552         tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
 553         tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
 554
 555         if (is_eth) {
 556                 u16 vlan = 0;
 557                 if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
 558                                                 NULL)) {
 559                         /* VST mode */
 560                         if (vlan != wc->vlan_id)
 561                                 /* Packet vlan is not the VST-assigned vlan.
 562                                  * Drop the packet.
 563                                  */
 564                                 goto out;
 565                          else
 566                                 /* Remove the vlan tag before forwarding
 567                                  * the packet to the VF.
 568                                  */
 569                                 vlan = 0xffff;
 570                 } else {
 571                         vlan = wc->vlan_id;
 572                 }
 573
 574                 tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
 575                 memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
 576                 memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
 577         } else {
 578                 tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
 579                 tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
 580         }
 581
 582         ib_dma_sync_single_for_device(&dev->ib_dev,
 583                                       tun_qp->tx_ring[tun_tx_ix].buf.map,
 584                                       sizeof (struct mlx4_rcv_tunnel_mad),
 585                                       DMA_TO_DEVICE);
 586
 587         list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
 588         list.length = sizeof (struct mlx4_rcv_tunnel_mad);
 589         list.lkey = tun_ctx->mr->lkey;
 590
 591         wr.wr.ud.ah = ah;
 592         wr.wr.ud.port_num = port;
 593         wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
 594         wr.wr.ud.remote_qpn = dqpn;
 595         wr.next = NULL;
 596         wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
 597         wr.sg_list = &list;
 598         wr.num_sge = 1;
 599         wr.opcode = IB_WR_SEND;
 600         wr.send_flags = IB_SEND_SIGNALED;
 601
 602         ret = ib_post_send(src_qp, &wr, &bad_wr);
 603 out:
 604         if (ret)
 605                 ib_destroy_ah(ah);
 606         return ret;
 607 }
 608
 609 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
 610                         struct ib_wc *wc, struct ib_grh *grh,
 611                         struct ib_mad *mad)
 612 {
 613         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 614         int err;
 615         int slave;
 616         u8 *slave_id;
 617         int is_eth = 0;
 618
 619         if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
 620                 is_eth = 0;
 621         else
 622                 is_eth = 1;
 623
 624         if (is_eth) {
 625                 if (!(wc->wc_flags & IB_WC_GRH)) {
 626                         mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
 627                         return -EINVAL;
 628                 }
 629                 if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
 630                         mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
 631                         return -EINVAL;
 632                 }
 633                 if (mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave)) {
 634                         mlx4_ib_warn(ibdev, "failed matching grh\n");
 635                         return -ENOENT;
 636                 }
 637                 if (slave >= dev->dev->caps.sqp_demux) {
 638                         mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
 639                                      slave, dev->dev->caps.sqp_demux);
 640                         return -ENOENT;
 641                 }
 642
 643                 if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
 644                         return 0;
 645
 646                 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
 647                 if (err)
 648                         pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
 649                                  slave, err);
 650                 return 0;
 651         }
 652
 653         /* Initially assume that this mad is for us */
 654         slave = mlx4_master_func_num(dev->dev);
 655
 656         /* See if the slave id is encoded in a response mad */
 657         if (mad->mad_hdr.method & 0x80) {
 658                 slave_id = (u8 *) &mad->mad_hdr.tid;
 659                 slave = *slave_id;
 660                 if (slave != 255) /*255 indicates the dom0*/
 661                         *slave_id = 0; /* remap tid */
 662         }
 663
 664         /* If a grh is present, we demux according to it */
 665         if (wc->wc_flags & IB_WC_GRH) {
 666                 slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
 667                 if (slave < 0) {
 668                         mlx4_ib_warn(ibdev, "failed matching grh\n");
 669                         return -ENOENT;
 670                 }
 671         }
 672         /* Class-specific handling */
 673         switch (mad->mad_hdr.mgmt_class) {
 674         case IB_MGMT_CLASS_SUBN_LID_ROUTED:
 675         case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
 676                 /* 255 indicates the dom0 */
 677                 if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
 678                         if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
 679                                 return -EPERM;
 680                         /* for a VF. drop unsolicited MADs */
 681                         if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
 682                                 mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
 683                                              slave, mad->mad_hdr.mgmt_class,
 684                                              mad->mad_hdr.method);
 685                                 return -EINVAL;
 686                         }
 687                 }
 688                 break;
 689         case IB_MGMT_CLASS_SUBN_ADM:
 690                 if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
 691                                              (struct ib_sa_mad *) mad))
 692                         return 0;
 693                 break;
 694         case IB_MGMT_CLASS_CM:
 695                 if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
 696                         return 0;
 697                 break;
 698         case IB_MGMT_CLASS_DEVICE_MGMT:
 699                 if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
 700                         return 0;
 701                 break;
 702         default:
 703                 /* Drop unsupported classes for slaves in tunnel mode */
 704                 if (slave != mlx4_master_func_num(dev->dev)) {
 705                         pr_debug("dropping unsupported ingress mad from class:%d "
 706                                  "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
 707                         return 0;
 708                 }
 709         }
 710         /*make sure that no slave==255 was not handled yet.*/
 711         if (slave >= dev->dev->caps.sqp_demux) {
 712                 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
 713                              slave, dev->dev->caps.sqp_demux);
 714                 return -ENOENT;
 715         }
 716
 717         err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
 718         if (err)
 719                 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
 720                          slave, err);
 721         return 0;
 722 }
 723
 724 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 725                         struct ib_wc *in_wc, struct ib_grh *in_grh,
 726                         struct ib_mad *in_mad, struct ib_mad *out_mad)
 727 {
 728         u16 slid, prev_lid = 0;
 729         int err;
 730         struct ib_port_attr pattr;
 731
 732         if (in_wc && in_wc->qp->qp_num) {
 733                 pr_debug("received MAD: slid:%d sqpn:%d "
 734                         "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
 735                         in_wc->slid, in_wc->src_qp,
 736                         in_wc->dlid_path_bits,
 737                         in_wc->qp->qp_num,
 738                         in_wc->wc_flags,
 739                         in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
 740                         be16_to_cpu(in_mad->mad_hdr.attr_id));
 741                 if (in_wc->wc_flags & IB_WC_GRH) {
 742                         pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
 743                                  be64_to_cpu(in_grh->sgid.global.subnet_prefix),
 744                                  be64_to_cpu(in_grh->sgid.global.interface_id));
 745                         pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
 746                                  be64_to_cpu(in_grh->dgid.global.subnet_prefix),
 747                                  be64_to_cpu(in_grh->dgid.global.interface_id));
 748                 }
 749         }
 750
 751         slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
 752
 753         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
 754                 forward_trap(to_mdev(ibdev), port_num, in_mad);
 755                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
 756         }
 757
 758         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 759             in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 760                 if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
 761                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
 762                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
 763                         return IB_MAD_RESULT_SUCCESS;
 764
 765                 /*
 766                  * Don't process SMInfo queries -- the SMA can't handle them.
 767                  */
 768                 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
 769                         return IB_MAD_RESULT_SUCCESS;
 770         } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
 771                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
 772                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
 773                    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
 774                 if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
 775                     in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
 776                         return IB_MAD_RESULT_SUCCESS;
 777         } else
 778                 return IB_MAD_RESULT_SUCCESS;
 779
 780         if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 781              in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 782             in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
 783             in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
 784             !ib_query_port(ibdev, port_num, &pattr))
 785                 prev_lid = pattr.lid;
 786
 787         err = mlx4_MAD_IFC(to_mdev(ibdev),
 788                            (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
 789                            (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
 790                            MLX4_MAD_IFC_NET_VIEW,
 791                            port_num, in_wc, in_grh, in_mad, out_mad);
 792         if (err)
 793                 return IB_MAD_RESULT_FAILURE;
 794
 795         if (!out_mad->mad_hdr.status) {
 796                 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
 797                         smp_snoop(ibdev, port_num, in_mad, prev_lid);
 798                 /* slaves get node desc from FW */
 799                 if (!mlx4_is_slave(to_mdev(ibdev)->dev))
 800                         node_desc_override(ibdev, out_mad);
 801         }
 802
 803         /* set return bit in status of directed route responses */
 804         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 805                 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
 806
 807         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
 808                 /* no response for trap repress */
 809                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
 810
 811         return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
 812 }
 813
 814 static void edit_counter(struct mlx4_counter *cnt,
 815                                         struct ib_pma_portcounters *pma_cnt)
 816 {
 817         ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
 818                              (be64_to_cpu(cnt->tx_bytes) >> 2));
 819         ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
 820                              (be64_to_cpu(cnt->rx_bytes) >> 2));
 821         ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
 822                              be64_to_cpu(cnt->tx_frames));
 823         ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
 824                              be64_to_cpu(cnt->rx_frames));
 825 }
 826
 827 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 828                         struct ib_wc *in_wc, struct ib_grh *in_grh,
 829                         struct ib_mad *in_mad, struct ib_mad *out_mad)
 830 {
 831         struct mlx4_cmd_mailbox *mailbox;
 832         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 833         int err;
 834         u32 inmod = dev->counters[port_num - 1].index & 0xffff;
 835         u8 mode;
 836
 837         if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
 838                 return -EINVAL;
 839
 840         mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
 841         if (IS_ERR(mailbox))
 842                 return IB_MAD_RESULT_FAILURE;
 843
 844         err = mlx4_cmd_box(dev->dev, 0, mailbox->dma, inmod, 0,
 845                            MLX4_CMD_QUERY_IF_STAT, MLX4_CMD_TIME_CLASS_C,
 846                            MLX4_CMD_WRAPPED);
 847         if (err)
 848                 err = IB_MAD_RESULT_FAILURE;
 849         else {
 850                 memset(out_mad->data, 0, sizeof out_mad->data);
 851                 mode = ((struct mlx4_counter *)mailbox->buf)->counter_mode;
 852                 switch (mode & 0xf) {
 853                 case 0:
 854                         edit_counter(mailbox->buf,
 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         mlx4_free_cmd_mailbox(dev->dev, mailbox);
 864
 865         return err;
 866 }
 867
 868 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 869                         struct ib_wc *in_wc, struct ib_grh *in_grh,
 870                         struct ib_mad *in_mad, struct ib_mad *out_mad)
 871 {
 872         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 873         switch (rdma_port_get_link_layer(ibdev, port_num)) {
 874         case IB_LINK_LAYER_INFINIBAND:
 875                 if (!mlx4_is_slave(dev->dev))
 876                         return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
 877                                               in_grh, in_mad, out_mad);
 878         case IB_LINK_LAYER_ETHERNET:
 879                 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
 880                                           in_grh, in_mad, out_mad);
 881         default:
 882                 return -EINVAL;
 883         }
 884 }
 885
 886 static void send_handler(struct ib_mad_agent *agent,
 887                          struct ib_mad_send_wc *mad_send_wc)
 888 {
 889         if (mad_send_wc->send_buf->context[0])
 890                 ib_destroy_ah(mad_send_wc->send_buf->context[0]);
 891         ib_free_send_mad(mad_send_wc->send_buf);
 892 }
 893
 894 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
 895 {
 896         struct ib_mad_agent *agent;
 897         int p, q;
 898         int ret;
 899         enum rdma_link_layer ll;
 900
 901         for (p = 0; p < dev->num_ports; ++p) {
 902                 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
 903                 for (q = 0; q <= 1; ++q) {
 904                         if (ll == IB_LINK_LAYER_INFINIBAND) {
 905                                 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
 906                                                               q ? IB_QPT_GSI : IB_QPT_SMI,
 907                                                               NULL, 0, send_handler,
 908                                                               NULL, NULL, 0);
 909                                 if (IS_ERR(agent)) {
 910                                         ret = PTR_ERR(agent);
 911                                         goto err;
 912                                 }
 913                                 dev->send_agent[p][q] = agent;
 914                         } else
 915                                 dev->send_agent[p][q] = NULL;
 916                 }
 917         }
 918
 919         return 0;
 920
 921 err:
 922         for (p = 0; p < dev->num_ports; ++p)
 923                 for (q = 0; q <= 1; ++q)
 924                         if (dev->send_agent[p][q])
 925                                 ib_unregister_mad_agent(dev->send_agent[p][q]);
 926
 927         return ret;
 928 }
 929
 930 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
 931 {
 932         struct ib_mad_agent *agent;
 933         int p, q;
 934
 935         for (p = 0; p < dev->num_ports; ++p) {
 936                 for (q = 0; q <= 1; ++q) {
 937                         agent = dev->send_agent[p][q];
 938                         if (agent) {
 939                                 dev->send_agent[p][q] = NULL;
 940                                 ib_unregister_mad_agent(agent);
 941                         }
 942                 }
 943
 944                 if (dev->sm_ah[p])
 945                         ib_destroy_ah(dev->sm_ah[p]);
 946         }
 947 }
 948
 949 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
 950 {
 951         mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
 952
 953         if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
 954                 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
 955                                             MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
 956 }
 957
 958 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
 959 {
 960         /* re-configure the alias-guid and mcg's */
 961         if (mlx4_is_master(dev->dev)) {
 962                 mlx4_ib_invalidate_all_guid_record(dev, port_num);
 963
 964                 if (!dev->sriov.is_going_down) {
 965                         mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
 966                         mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
 967                                                     MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
 968                 }
 969         }
 970         mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
 971 }
 972
 973 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
 974                               struct mlx4_eqe *eqe)
 975 {
 976         __propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
 977                             GET_MASK_FROM_EQE(eqe));
 978 }
 979
 980 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
 981                                       u32 guid_tbl_blk_num, u32 change_bitmap)
 982 {
 983         struct ib_smp *in_mad  = NULL;
 984         struct ib_smp *out_mad  = NULL;
 985         u16 i;
 986
 987         if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
 988                 return;
 989
 990         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
 991         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 992         if (!in_mad || !out_mad) {
 993                 mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
 994                 goto out;
 995         }
 996
 997         guid_tbl_blk_num  *= 4;
 998
 999         for (i = 0; i < 4; i++) {
1000                 if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1001                         continue;
1002                 memset(in_mad, 0, sizeof *in_mad);
1003                 memset(out_mad, 0, sizeof *out_mad);
1004
1005                 in_mad->base_version  = 1;
1006                 in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1007                 in_mad->class_version = 1;
1008                 in_mad->method        = IB_MGMT_METHOD_GET;
1009                 in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1010                 in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1011
1012                 if (mlx4_MAD_IFC(dev,
1013                                  MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1014                                  port_num, NULL, NULL, in_mad, out_mad)) {
1015                         mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1016                         goto out;
1017                 }
1018
1019                 mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1020                                                     port_num,
1021                                                     (u8 *)(&((struct ib_smp *)out_mad)->data));
1022                 mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1023                                                      port_num,
1024                                                      (u8 *)(&((struct ib_smp *)out_mad)->data));
1025         }
1026
1027 out:
1028         kfree(in_mad);
1029         kfree(out_mad);
1030         return;
1031 }
1032
1033 void handle_port_mgmt_change_event(struct work_struct *work)
1034 {
1035         struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1036         struct mlx4_ib_dev *dev = ew->ib_dev;
1037         struct mlx4_eqe *eqe = &(ew->ib_eqe);
1038         u8 port = eqe->event.port_mgmt_change.port;
1039         u32 changed_attr;
1040         u32 tbl_block;
1041         u32 change_bitmap;
1042
1043         switch (eqe->subtype) {
1044         case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1045                 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1046
1047                 /* Update the SM ah - This should be done before handling
1048                    the other changed attributes so that MADs can be sent to the SM */
1049                 if (changed_attr & MSTR_SM_CHANGE_MASK) {
1050                         u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1051                         u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1052                         update_sm_ah(dev, port, lid, sl);
1053                 }
1054
1055                 /* Check if it is a lid change event */
1056                 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1057                         handle_lid_change_event(dev, port);
1058
1059                 /* Generate GUID changed event */
1060                 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1061                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1062                         /*if master, notify all slaves*/
1063                         if (mlx4_is_master(dev->dev))
1064                                 mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1065                                                             MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1066                 }
1067
1068                 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1069                         handle_client_rereg_event(dev, port);
1070                 break;
1071
1072         case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1073                 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1074                 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1075                         propagate_pkey_ev(dev, port, eqe);
1076                 break;
1077         case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1078                 /* paravirtualized master's guid is guid 0 -- does not change */
1079                 if (!mlx4_is_master(dev->dev))
1080                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1081                 /*if master, notify relevant slaves*/
1082                 else if (!dev->sriov.is_going_down) {
1083                         tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1084                         change_bitmap = GET_MASK_FROM_EQE(eqe);
1085                         handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1086                 }
1087                 break;
1088         default:
1089                 pr_warn("Unsupported subtype 0x%x for "
1090                         "Port Management Change event\n", eqe->subtype);
1091         }
1092
1093         kfree(ew);
1094 }
1095
1096 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1097                             enum ib_event_type type)
1098 {
1099         struct ib_event event;
1100
1101         event.device            = &dev->ib_dev;
1102         event.element.port_num  = port_num;
1103         event.event             = type;
1104
1105         ib_dispatch_event(&event);
1106 }
1107
1108 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1109 {
1110         unsigned long flags;
1111         struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1112         struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1113         spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1114         if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1115                 queue_work(ctx->wq, &ctx->work);
1116         spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1117 }
1118
1119 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1120                                   struct mlx4_ib_demux_pv_qp *tun_qp,
1121                                   int index)
1122 {
1123         struct ib_sge sg_list;
1124         struct ib_recv_wr recv_wr, *bad_recv_wr;
1125         int size;
1126
1127         size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1128                 sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1129
1130         sg_list.addr = tun_qp->ring[index].map;
1131         sg_list.length = size;
1132         sg_list.lkey = ctx->mr->lkey;
1133
1134         recv_wr.next = NULL;
1135         recv_wr.sg_list = &sg_list;
1136         recv_wr.num_sge = 1;
1137         recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1138                 MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1139         ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1140                                       size, DMA_FROM_DEVICE);
1141         return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1142 }
1143
1144 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1145                 int slave, struct ib_sa_mad *sa_mad)
1146 {
1147         int ret = 0;
1148
1149         /* dispatch to different sa handlers */
1150         switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1151         case IB_SA_ATTR_MC_MEMBER_REC:
1152                 ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1153                 break;
1154         default:
1155                 break;
1156         }
1157         return ret;
1158 }
1159
1160 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1161 {
1162         int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1163
1164         return (qpn >= proxy_start && qpn <= proxy_start + 1);
1165 }
1166
1167
1168 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1169                          enum ib_qp_type dest_qpt, u16 pkey_index,
1170                          u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1171                          u8 *s_mac, struct ib_mad *mad)
1172 {
1173         struct ib_sge list;
1174         struct ib_send_wr wr, *bad_wr;
1175         struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1176         struct mlx4_ib_demux_pv_qp *sqp;
1177         struct mlx4_mad_snd_buf *sqp_mad;
1178         struct ib_ah *ah;
1179         struct ib_qp *send_qp = NULL;
1180         unsigned wire_tx_ix = 0;
1181         int ret = 0;
1182         u16 wire_pkey_ix;
1183         int src_qpnum;
1184         u8 sgid_index;
1185
1186
1187         sqp_ctx = dev->sriov.sqps[port-1];
1188
1189         /* check if proxy qp created */
1190         if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1191                 return -EAGAIN;
1192
1193         if (dest_qpt == IB_QPT_SMI) {
1194                 src_qpnum = 0;
1195                 sqp = &sqp_ctx->qp[0];
1196                 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1197         } else {
1198                 src_qpnum = 1;
1199                 sqp = &sqp_ctx->qp[1];
1200                 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1201         }
1202
1203         send_qp = sqp->qp;
1204
1205         /* create ah */
1206         sgid_index = attr->grh.sgid_index;
1207         attr->grh.sgid_index = 0;
1208         ah = ib_create_ah(sqp_ctx->pd, attr);
1209         if (IS_ERR(ah))
1210                 return -ENOMEM;
1211         attr->grh.sgid_index = sgid_index;
1212         to_mah(ah)->av.ib.gid_index = sgid_index;
1213         /* get rid of force-loopback bit */
1214         to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1215         spin_lock(&sqp->tx_lock);
1216         if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1217             (MLX4_NUM_TUNNEL_BUFS - 1))
1218                 ret = -EAGAIN;
1219         else
1220                 wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1221         spin_unlock(&sqp->tx_lock);
1222         if (ret)
1223                 goto out;
1224
1225         sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1226         if (sqp->tx_ring[wire_tx_ix].ah)
1227                 ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1228         sqp->tx_ring[wire_tx_ix].ah = ah;
1229         ib_dma_sync_single_for_cpu(&dev->ib_dev,
1230                                    sqp->tx_ring[wire_tx_ix].buf.map,
1231                                    sizeof (struct mlx4_mad_snd_buf),
1232                                    DMA_TO_DEVICE);
1233
1234         memcpy(&sqp_mad->payload, mad, sizeof *mad);
1235
1236         ib_dma_sync_single_for_device(&dev->ib_dev,
1237                                       sqp->tx_ring[wire_tx_ix].buf.map,
1238                                       sizeof (struct mlx4_mad_snd_buf),
1239                                       DMA_TO_DEVICE);
1240
1241         list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1242         list.length = sizeof (struct mlx4_mad_snd_buf);
1243         list.lkey = sqp_ctx->mr->lkey;
1244
1245         wr.wr.ud.ah = ah;
1246         wr.wr.ud.port_num = port;
1247         wr.wr.ud.pkey_index = wire_pkey_ix;
1248         wr.wr.ud.remote_qkey = qkey;
1249         wr.wr.ud.remote_qpn = remote_qpn;
1250         wr.next = NULL;
1251         wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1252         wr.sg_list = &list;
1253         wr.num_sge = 1;
1254         wr.opcode = IB_WR_SEND;
1255         wr.send_flags = IB_SEND_SIGNALED;
1256         if (s_mac)
1257                 memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1258
1259
1260         ret = ib_post_send(send_qp, &wr, &bad_wr);
1261 out:
1262         if (ret)
1263                 ib_destroy_ah(ah);
1264         return ret;
1265 }
1266
1267 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1268 {
1269         if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1270                 return slave;
1271         return mlx4_get_base_gid_ix(dev->dev, slave, port);
1272 }
1273
1274 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1275                                     struct ib_ah_attr *ah_attr)
1276 {
1277         if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1278                 ah_attr->grh.sgid_index = slave;
1279         else
1280                 ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1281 }
1282
1283 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1284 {
1285         struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1286         struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1287         int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1288         struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1289         struct mlx4_ib_ah ah;
1290         struct ib_ah_attr ah_attr;
1291         u8 *slave_id;
1292         int slave;
1293         int port;
1294
1295         /* Get slave that sent this packet */
1296         if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1297             wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1298             (wc->src_qp & 0x1) != ctx->port - 1 ||
1299             wc->src_qp & 0x4) {
1300                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1301                 return;
1302         }
1303         slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1304         if (slave != ctx->slave) {
1305                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1306                              "belongs to another slave\n", wc->src_qp);
1307                 return;
1308         }
1309
1310         /* Map transaction ID */
1311         ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1312                                    sizeof (struct mlx4_tunnel_mad),
1313                                    DMA_FROM_DEVICE);
1314         switch (tunnel->mad.mad_hdr.method) {
1315         case IB_MGMT_METHOD_SET:
1316         case IB_MGMT_METHOD_GET:
1317         case IB_MGMT_METHOD_REPORT:
1318         case IB_SA_METHOD_GET_TABLE:
1319         case IB_SA_METHOD_DELETE:
1320         case IB_SA_METHOD_GET_MULTI:
1321         case IB_SA_METHOD_GET_TRACE_TBL:
1322                 slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1323                 if (*slave_id) {
1324                         mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1325                                      "class:%d slave:%d\n", *slave_id,
1326                                      tunnel->mad.mad_hdr.mgmt_class, slave);
1327                         return;
1328                 } else
1329                         *slave_id = slave;
1330         default:
1331                 /* nothing */;
1332         }
1333
1334         /* Class-specific handling */
1335         switch (tunnel->mad.mad_hdr.mgmt_class) {
1336         case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1337         case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1338                 if (slave != mlx4_master_func_num(dev->dev) &&
1339                     !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1340                         return;
1341                 break;
1342         case IB_MGMT_CLASS_SUBN_ADM:
1343                 if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1344                               (struct ib_sa_mad *) &tunnel->mad))
1345                         return;
1346                 break;
1347         case IB_MGMT_CLASS_CM:
1348                 if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1349                               (struct ib_mad *) &tunnel->mad))
1350                         return;
1351                 break;
1352         case IB_MGMT_CLASS_DEVICE_MGMT:
1353                 if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1354                     tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1355                         return;
1356                 break;
1357         default:
1358                 /* Drop unsupported classes for slaves in tunnel mode */
1359                 if (slave != mlx4_master_func_num(dev->dev)) {
1360                         mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1361                                      "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1362                         return;
1363                 }
1364         }
1365
1366         /* We are using standard ib_core services to send the mad, so generate a
1367          * stadard address handle by decoding the tunnelled mlx4_ah fields */
1368         memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1369         ah.ibah.device = ctx->ib_dev;
1370
1371         port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1372         port = mlx4_slave_convert_port(dev->dev, slave, port);
1373         if (port < 0)
1374                 return;
1375         ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1376
1377         mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1378         if (ah_attr.ah_flags & IB_AH_GRH)
1379                 fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1380
1381         memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1382         ah_attr.vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1383         /* if slave have default vlan use it */
1384         mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1385                                     &ah_attr.vlan_id, &ah_attr.sl);
1386
1387         mlx4_ib_send_to_wire(dev, slave, ctx->port,
1388                              is_proxy_qp0(dev, wc->src_qp, slave) ?
1389                              IB_QPT_SMI : IB_QPT_GSI,
1390                              be16_to_cpu(tunnel->hdr.pkey_index),
1391                              be32_to_cpu(tunnel->hdr.remote_qpn),
1392                              be32_to_cpu(tunnel->hdr.qkey),
1393                              &ah_attr, wc->smac, &tunnel->mad);
1394 }
1395
1396 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1397                                  enum ib_qp_type qp_type, int is_tun)
1398 {
1399         int i;
1400         struct mlx4_ib_demux_pv_qp *tun_qp;
1401         int rx_buf_size, tx_buf_size;
1402
1403         if (qp_type > IB_QPT_GSI)
1404                 return -EINVAL;
1405
1406         tun_qp = &ctx->qp[qp_type];
1407
1408         tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1409                                GFP_KERNEL);
1410         if (!tun_qp->ring)
1411                 return -ENOMEM;
1412
1413         tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1414                                   sizeof (struct mlx4_ib_tun_tx_buf),
1415                                   GFP_KERNEL);
1416         if (!tun_qp->tx_ring) {
1417                 kfree(tun_qp->ring);
1418                 tun_qp->ring = NULL;
1419                 return -ENOMEM;
1420         }
1421
1422         if (is_tun) {
1423                 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1424                 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1425         } else {
1426                 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1427                 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1428         }
1429
1430         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1431                 tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1432                 if (!tun_qp->ring[i].addr)
1433                         goto err;
1434                 tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1435                                                         tun_qp->ring[i].addr,
1436                                                         rx_buf_size,
1437                                                         DMA_FROM_DEVICE);
1438                 if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1439                         kfree(tun_qp->ring[i].addr);
1440                         goto err;
1441                 }
1442         }
1443
1444         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1445                 tun_qp->tx_ring[i].buf.addr =
1446                         kmalloc(tx_buf_size, GFP_KERNEL);
1447                 if (!tun_qp->tx_ring[i].buf.addr)
1448                         goto tx_err;
1449                 tun_qp->tx_ring[i].buf.map =
1450                         ib_dma_map_single(ctx->ib_dev,
1451                                           tun_qp->tx_ring[i].buf.addr,
1452                                           tx_buf_size,
1453                                           DMA_TO_DEVICE);
1454                 if (ib_dma_mapping_error(ctx->ib_dev,
1455                                          tun_qp->tx_ring[i].buf.map)) {
1456                         kfree(tun_qp->tx_ring[i].buf.addr);
1457                         goto tx_err;
1458                 }
1459                 tun_qp->tx_ring[i].ah = NULL;
1460         }
1461         spin_lock_init(&tun_qp->tx_lock);
1462         tun_qp->tx_ix_head = 0;
1463         tun_qp->tx_ix_tail = 0;
1464         tun_qp->proxy_qpt = qp_type;
1465
1466         return 0;
1467
1468 tx_err:
1469         while (i > 0) {
1470                 --i;
1471                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1472                                     tx_buf_size, DMA_TO_DEVICE);
1473                 kfree(tun_qp->tx_ring[i].buf.addr);
1474         }
1475         kfree(tun_qp->tx_ring);
1476         tun_qp->tx_ring = NULL;
1477         i = MLX4_NUM_TUNNEL_BUFS;
1478 err:
1479         while (i > 0) {
1480                 --i;
1481                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1482                                     rx_buf_size, DMA_FROM_DEVICE);
1483                 kfree(tun_qp->ring[i].addr);
1484         }
1485         kfree(tun_qp->ring);
1486         tun_qp->ring = NULL;
1487         return -ENOMEM;
1488 }
1489
1490 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1491                                      enum ib_qp_type qp_type, int is_tun)
1492 {
1493         int i;
1494         struct mlx4_ib_demux_pv_qp *tun_qp;
1495         int rx_buf_size, tx_buf_size;
1496
1497         if (qp_type > IB_QPT_GSI)
1498                 return;
1499
1500         tun_qp = &ctx->qp[qp_type];
1501         if (is_tun) {
1502                 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1503                 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1504         } else {
1505                 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1506                 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1507         }
1508
1509
1510         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1511                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1512                                     rx_buf_size, DMA_FROM_DEVICE);
1513                 kfree(tun_qp->ring[i].addr);
1514         }
1515
1516         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1517                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1518                                     tx_buf_size, DMA_TO_DEVICE);
1519                 kfree(tun_qp->tx_ring[i].buf.addr);
1520                 if (tun_qp->tx_ring[i].ah)
1521                         ib_destroy_ah(tun_qp->tx_ring[i].ah);
1522         }
1523         kfree(tun_qp->tx_ring);
1524         kfree(tun_qp->ring);
1525 }
1526
1527 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1528 {
1529         struct mlx4_ib_demux_pv_ctx *ctx;
1530         struct mlx4_ib_demux_pv_qp *tun_qp;
1531         struct ib_wc wc;
1532         int ret;
1533         ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1534         ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1535
1536         while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1537                 tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1538                 if (wc.status == IB_WC_SUCCESS) {
1539                         switch (wc.opcode) {
1540                         case IB_WC_RECV:
1541                                 mlx4_ib_multiplex_mad(ctx, &wc);
1542                                 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1543                                                              wc.wr_id &
1544                                                              (MLX4_NUM_TUNNEL_BUFS - 1));
1545                                 if (ret)
1546                                         pr_err("Failed reposting tunnel "
1547                                                "buf:%lld\n", wc.wr_id);
1548                                 break;
1549                         case IB_WC_SEND:
1550                                 pr_debug("received tunnel send completion:"
1551                                          "wrid=0x%llx, status=0x%x\n",
1552                                          wc.wr_id, wc.status);
1553                                 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1554                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1555                                 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1556                                         = NULL;
1557                                 spin_lock(&tun_qp->tx_lock);
1558                                 tun_qp->tx_ix_tail++;
1559                                 spin_unlock(&tun_qp->tx_lock);
1560
1561                                 break;
1562                         default:
1563                                 break;
1564                         }
1565                 } else  {
1566                         pr_debug("mlx4_ib: completion error in tunnel: %d."
1567                                  " status = %d, wrid = 0x%llx\n",
1568                                  ctx->slave, wc.status, wc.wr_id);
1569                         if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1570                                 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1571                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1572                                 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1573                                         = NULL;
1574                                 spin_lock(&tun_qp->tx_lock);
1575                                 tun_qp->tx_ix_tail++;
1576                                 spin_unlock(&tun_qp->tx_lock);
1577                         }
1578                 }
1579         }
1580 }
1581
1582 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1583 {
1584         struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1585
1586         /* It's worse than that! He's dead, Jim! */
1587         pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1588                event->event, sqp->port);
1589 }
1590
1591 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1592                             enum ib_qp_type qp_type, int create_tun)
1593 {
1594         int i, ret;
1595         struct mlx4_ib_demux_pv_qp *tun_qp;
1596         struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1597         struct ib_qp_attr attr;
1598         int qp_attr_mask_INIT;
1599
1600         if (qp_type > IB_QPT_GSI)
1601                 return -EINVAL;
1602
1603         tun_qp = &ctx->qp[qp_type];
1604
1605         memset(&qp_init_attr, 0, sizeof qp_init_attr);
1606         qp_init_attr.init_attr.send_cq = ctx->cq;
1607         qp_init_attr.init_attr.recv_cq = ctx->cq;
1608         qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1609         qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1610         qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1611         qp_init_attr.init_attr.cap.max_send_sge = 1;
1612         qp_init_attr.init_attr.cap.max_recv_sge = 1;
1613         if (create_tun) {
1614                 qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1615                 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1616                 qp_init_attr.port = ctx->port;
1617                 qp_init_attr.slave = ctx->slave;
1618                 qp_init_attr.proxy_qp_type = qp_type;
1619                 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1620                            IB_QP_QKEY | IB_QP_PORT;
1621         } else {
1622                 qp_init_attr.init_attr.qp_type = qp_type;
1623                 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1624                 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1625         }
1626         qp_init_attr.init_attr.port_num = ctx->port;
1627         qp_init_attr.init_attr.qp_context = ctx;
1628         qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1629         tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1630         if (IS_ERR(tun_qp->qp)) {
1631                 ret = PTR_ERR(tun_qp->qp);
1632                 tun_qp->qp = NULL;
1633                 pr_err("Couldn't create %s QP (%d)\n",
1634                        create_tun ? "tunnel" : "special", ret);
1635                 return ret;
1636         }
1637
1638         memset(&attr, 0, sizeof attr);
1639         attr.qp_state = IB_QPS_INIT;
1640         ret = 0;
1641         if (create_tun)
1642                 ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1643                                               ctx->port, IB_DEFAULT_PKEY_FULL,
1644                                               &attr.pkey_index);
1645         if (ret || !create_tun)
1646                 attr.pkey_index =
1647                         to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1648         attr.qkey = IB_QP1_QKEY;
1649         attr.port_num = ctx->port;
1650         ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1651         if (ret) {
1652                 pr_err("Couldn't change %s qp state to INIT (%d)\n",
1653                        create_tun ? "tunnel" : "special", ret);
1654                 goto err_qp;
1655         }
1656         attr.qp_state = IB_QPS_RTR;
1657         ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1658         if (ret) {
1659                 pr_err("Couldn't change %s qp state to RTR (%d)\n",
1660                        create_tun ? "tunnel" : "special", ret);
1661                 goto err_qp;
1662         }
1663         attr.qp_state = IB_QPS_RTS;
1664         attr.sq_psn = 0;
1665         ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1666         if (ret) {
1667                 pr_err("Couldn't change %s qp state to RTS (%d)\n",
1668                        create_tun ? "tunnel" : "special", ret);
1669                 goto err_qp;
1670         }
1671
1672         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1673                 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1674                 if (ret) {
1675                         pr_err(" mlx4_ib_post_pv_buf error"
1676                                " (err = %d, i = %d)\n", ret, i);
1677                         goto err_qp;
1678                 }
1679         }
1680         return 0;
1681
1682 err_qp:
1683         ib_destroy_qp(tun_qp->qp);
1684         tun_qp->qp = NULL;
1685         return ret;
1686 }
1687
1688 /*
1689  * IB MAD completion callback for real SQPs
1690  */
1691 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1692 {
1693         struct mlx4_ib_demux_pv_ctx *ctx;
1694         struct mlx4_ib_demux_pv_qp *sqp;
1695         struct ib_wc wc;
1696         struct ib_grh *grh;
1697         struct ib_mad *mad;
1698
1699         ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1700         ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1701
1702         while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1703                 sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1704                 if (wc.status == IB_WC_SUCCESS) {
1705                         switch (wc.opcode) {
1706                         case IB_WC_SEND:
1707                                 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1708                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1709                                 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1710                                         = NULL;
1711                                 spin_lock(&sqp->tx_lock);
1712                                 sqp->tx_ix_tail++;
1713                                 spin_unlock(&sqp->tx_lock);
1714                                 break;
1715                         case IB_WC_RECV:
1716                                 mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1717                                                 (sqp->ring[wc.wr_id &
1718                                                 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1719                                 grh = &(((struct mlx4_mad_rcv_buf *)
1720                                                 (sqp->ring[wc.wr_id &
1721                                                 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1722                                 mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1723                                 if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1724                                                            (MLX4_NUM_TUNNEL_BUFS - 1)))
1725                                         pr_err("Failed reposting SQP "
1726                                                "buf:%lld\n", wc.wr_id);
1727                                 break;
1728                         default:
1729                                 BUG_ON(1);
1730                                 break;
1731                         }
1732                 } else  {
1733                         pr_debug("mlx4_ib: completion error in tunnel: %d."
1734                                  " status = %d, wrid = 0x%llx\n",
1735                                  ctx->slave, wc.status, wc.wr_id);
1736                         if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1737                                 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1738                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1739                                 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1740                                         = NULL;
1741                                 spin_lock(&sqp->tx_lock);
1742                                 sqp->tx_ix_tail++;
1743                                 spin_unlock(&sqp->tx_lock);
1744                         }
1745                 }
1746         }
1747 }
1748
1749 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1750                                struct mlx4_ib_demux_pv_ctx **ret_ctx)
1751 {
1752         struct mlx4_ib_demux_pv_ctx *ctx;
1753
1754         *ret_ctx = NULL;
1755         ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1756         if (!ctx) {
1757                 pr_err("failed allocating pv resource context "
1758                        "for port %d, slave %d\n", port, slave);
1759                 return -ENOMEM;
1760         }
1761
1762         ctx->ib_dev = &dev->ib_dev;
1763         ctx->port = port;
1764         ctx->slave = slave;
1765         *ret_ctx = ctx;
1766         return 0;
1767 }
1768
1769 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1770 {
1771         if (dev->sriov.demux[port - 1].tun[slave]) {
1772                 kfree(dev->sriov.demux[port - 1].tun[slave]);
1773                 dev->sriov.demux[port - 1].tun[slave] = NULL;
1774         }
1775 }
1776
1777 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1778                                int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1779 {
1780         int ret, cq_size;
1781
1782         if (ctx->state != DEMUX_PV_STATE_DOWN)
1783                 return -EEXIST;
1784
1785         ctx->state = DEMUX_PV_STATE_STARTING;
1786         /* have QP0 only if link layer is IB */
1787         if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1788             IB_LINK_LAYER_INFINIBAND)
1789                 ctx->has_smi = 1;
1790
1791         if (ctx->has_smi) {
1792                 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1793                 if (ret) {
1794                         pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1795                         goto err_out;
1796                 }
1797         }
1798
1799         ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1800         if (ret) {
1801                 pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1802                 goto err_out_qp0;
1803         }
1804
1805         cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1806         if (ctx->has_smi)
1807                 cq_size *= 2;
1808
1809         ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1810                                NULL, ctx, cq_size, 0);
1811         if (IS_ERR(ctx->cq)) {
1812                 ret = PTR_ERR(ctx->cq);
1813                 pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1814                 goto err_buf;
1815         }
1816
1817         ctx->pd = ib_alloc_pd(ctx->ib_dev);
1818         if (IS_ERR(ctx->pd)) {
1819                 ret = PTR_ERR(ctx->pd);
1820                 pr_err("Couldn't create tunnel PD (%d)\n", ret);
1821                 goto err_cq;
1822         }
1823
1824         ctx->mr = ib_get_dma_mr(ctx->pd, IB_ACCESS_LOCAL_WRITE);
1825         if (IS_ERR(ctx->mr)) {
1826                 ret = PTR_ERR(ctx->mr);
1827                 pr_err("Couldn't get tunnel DMA MR (%d)\n", ret);
1828                 goto err_pd;
1829         }
1830
1831         if (ctx->has_smi) {
1832                 ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1833                 if (ret) {
1834                         pr_err("Couldn't create %s QP0 (%d)\n",
1835                                create_tun ? "tunnel for" : "",  ret);
1836                         goto err_mr;
1837                 }
1838         }
1839
1840         ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1841         if (ret) {
1842                 pr_err("Couldn't create %s QP1 (%d)\n",
1843                        create_tun ? "tunnel for" : "",  ret);
1844                 goto err_qp0;
1845         }
1846
1847         if (create_tun)
1848                 INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1849         else
1850                 INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1851
1852         ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1853
1854         ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1855         if (ret) {
1856                 pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1857                 goto err_wq;
1858         }
1859         ctx->state = DEMUX_PV_STATE_ACTIVE;
1860         return 0;
1861
1862 err_wq:
1863         ctx->wq = NULL;
1864         ib_destroy_qp(ctx->qp[1].qp);
1865         ctx->qp[1].qp = NULL;
1866
1867
1868 err_qp0:
1869         if (ctx->has_smi)
1870                 ib_destroy_qp(ctx->qp[0].qp);
1871         ctx->qp[0].qp = NULL;
1872
1873 err_mr:
1874         ib_dereg_mr(ctx->mr);
1875         ctx->mr = NULL;
1876
1877 err_pd:
1878         ib_dealloc_pd(ctx->pd);
1879         ctx->pd = NULL;
1880
1881 err_cq:
1882         ib_destroy_cq(ctx->cq);
1883         ctx->cq = NULL;
1884
1885 err_buf:
1886         mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1887
1888 err_out_qp0:
1889         if (ctx->has_smi)
1890                 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1891 err_out:
1892         ctx->state = DEMUX_PV_STATE_DOWN;
1893         return ret;
1894 }
1895
1896 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1897                                  struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1898 {
1899         if (!ctx)
1900                 return;
1901         if (ctx->state > DEMUX_PV_STATE_DOWN) {
1902                 ctx->state = DEMUX_PV_STATE_DOWNING;
1903                 if (flush)
1904                         flush_workqueue(ctx->wq);
1905                 if (ctx->has_smi) {
1906                         ib_destroy_qp(ctx->qp[0].qp);
1907                         ctx->qp[0].qp = NULL;
1908                         mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1909                 }
1910                 ib_destroy_qp(ctx->qp[1].qp);
1911                 ctx->qp[1].qp = NULL;
1912                 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1913                 ib_dereg_mr(ctx->mr);
1914                 ctx->mr = NULL;
1915                 ib_dealloc_pd(ctx->pd);
1916                 ctx->pd = NULL;
1917                 ib_destroy_cq(ctx->cq);
1918                 ctx->cq = NULL;
1919                 ctx->state = DEMUX_PV_STATE_DOWN;
1920         }
1921 }
1922
1923 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1924                                   int port, int do_init)
1925 {
1926         int ret = 0;
1927
1928         if (!do_init) {
1929                 clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1930                 /* for master, destroy real sqp resources */
1931                 if (slave == mlx4_master_func_num(dev->dev))
1932                         destroy_pv_resources(dev, slave, port,
1933                                              dev->sriov.sqps[port - 1], 1);
1934                 /* destroy the tunnel qp resources */
1935                 destroy_pv_resources(dev, slave, port,
1936                                      dev->sriov.demux[port - 1].tun[slave], 1);
1937                 return 0;
1938         }
1939
1940         /* create the tunnel qp resources */
1941         ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1942                                   dev->sriov.demux[port - 1].tun[slave]);
1943
1944         /* for master, create the real sqp resources */
1945         if (!ret && slave == mlx4_master_func_num(dev->dev))
1946                 ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1947                                           dev->sriov.sqps[port - 1]);
1948         return ret;
1949 }
1950
1951 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1952 {
1953         struct mlx4_ib_demux_work *dmxw;
1954
1955         dmxw = container_of(work, struct mlx4_ib_demux_work, work);
1956         mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
1957                                dmxw->do_init);
1958         kfree(dmxw);
1959         return;
1960 }
1961
1962 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
1963                                        struct mlx4_ib_demux_ctx *ctx,
1964                                        int port)
1965 {
1966         char name[12];
1967         int ret = 0;
1968         int i;
1969
1970         ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
1971                            sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
1972         if (!ctx->tun)
1973                 return -ENOMEM;
1974
1975         ctx->dev = dev;
1976         ctx->port = port;
1977         ctx->ib_dev = &dev->ib_dev;
1978
1979         for (i = 0;
1980              i < min(dev->dev->caps.sqp_demux,
1981              (u16)(dev->dev->persist->num_vfs + 1));
1982              i++) {
1983                 struct mlx4_active_ports actv_ports =
1984                         mlx4_get_active_ports(dev->dev, i);
1985
1986                 if (!test_bit(port - 1, actv_ports.ports))
1987                         continue;
1988
1989                 ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
1990                 if (ret) {
1991                         ret = -ENOMEM;
1992                         goto err_mcg;
1993                 }
1994         }
1995
1996         ret = mlx4_ib_mcg_port_init(ctx);
1997         if (ret) {
1998                 pr_err("Failed initializing mcg para-virt (%d)\n", ret);
1999                 goto err_mcg;
2000         }
2001
2002         snprintf(name, sizeof name, "mlx4_ibt%d", port);
2003         ctx->wq = create_singlethread_workqueue(name);
2004         if (!ctx->wq) {
2005                 pr_err("Failed to create tunnelling WQ for port %d\n", port);
2006                 ret = -ENOMEM;
2007                 goto err_wq;
2008         }
2009
2010         snprintf(name, sizeof name, "mlx4_ibud%d", port);
2011         ctx->ud_wq = create_singlethread_workqueue(name);
2012         if (!ctx->ud_wq) {
2013                 pr_err("Failed to create up/down WQ for port %d\n", port);
2014                 ret = -ENOMEM;
2015                 goto err_udwq;
2016         }
2017
2018         return 0;
2019
2020 err_udwq:
2021         destroy_workqueue(ctx->wq);
2022         ctx->wq = NULL;
2023
2024 err_wq:
2025         mlx4_ib_mcg_port_cleanup(ctx, 1);
2026 err_mcg:
2027         for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2028                 free_pv_object(dev, i, port);
2029         kfree(ctx->tun);
2030         ctx->tun = NULL;
2031         return ret;
2032 }
2033
2034 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2035 {
2036         if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2037                 sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2038                 flush_workqueue(sqp_ctx->wq);
2039                 if (sqp_ctx->has_smi) {
2040                         ib_destroy_qp(sqp_ctx->qp[0].qp);
2041                         sqp_ctx->qp[0].qp = NULL;
2042                         mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2043                 }
2044                 ib_destroy_qp(sqp_ctx->qp[1].qp);
2045                 sqp_ctx->qp[1].qp = NULL;
2046                 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2047                 ib_dereg_mr(sqp_ctx->mr);
2048                 sqp_ctx->mr = NULL;
2049                 ib_dealloc_pd(sqp_ctx->pd);
2050                 sqp_ctx->pd = NULL;
2051                 ib_destroy_cq(sqp_ctx->cq);
2052                 sqp_ctx->cq = NULL;
2053                 sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2054         }
2055 }
2056
2057 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2058 {
2059         int i;
2060         if (ctx) {
2061                 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2062                 mlx4_ib_mcg_port_cleanup(ctx, 1);
2063                 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2064                         if (!ctx->tun[i])
2065                                 continue;
2066                         if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2067                                 ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2068                 }
2069                 flush_workqueue(ctx->wq);
2070                 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2071                         destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2072                         free_pv_object(dev, i, ctx->port);
2073                 }
2074                 kfree(ctx->tun);
2075                 destroy_workqueue(ctx->ud_wq);
2076                 destroy_workqueue(ctx->wq);
2077         }
2078 }
2079
2080 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2081 {
2082         int i;
2083
2084         if (!mlx4_is_master(dev->dev))
2085                 return;
2086         /* initialize or tear down tunnel QPs for the master */
2087         for (i = 0; i < dev->dev->caps.num_ports; i++)
2088                 mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2089         return;
2090 }
2091
2092 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2093 {
2094         int i = 0;
2095         int err;
2096
2097         if (!mlx4_is_mfunc(dev->dev))
2098                 return 0;
2099
2100         dev->sriov.is_going_down = 0;
2101         spin_lock_init(&dev->sriov.going_down_lock);
2102         mlx4_ib_cm_paravirt_init(dev);
2103
2104         mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2105
2106         if (mlx4_is_slave(dev->dev)) {
2107                 mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2108                 return 0;
2109         }
2110
2111         for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2112                 if (i == mlx4_master_func_num(dev->dev))
2113                         mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2114                 else
2115                         mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2116         }
2117
2118         err = mlx4_ib_init_alias_guid_service(dev);
2119         if (err) {
2120                 mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2121                 goto paravirt_err;
2122         }
2123         err = mlx4_ib_device_register_sysfs(dev);
2124         if (err) {
2125                 mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2126                 goto sysfs_err;
2127         }
2128
2129         mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2130                      dev->dev->caps.sqp_demux);
2131         for (i = 0; i < dev->num_ports; i++) {
2132                 union ib_gid gid;
2133                 err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2134                 if (err)
2135                         goto demux_err;
2136                 dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2137                 err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2138                                       &dev->sriov.sqps[i]);
2139                 if (err)
2140                         goto demux_err;
2141                 err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2142                 if (err)
2143                         goto free_pv;
2144         }
2145         mlx4_ib_master_tunnels(dev, 1);
2146         return 0;
2147
2148 free_pv:
2149         free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2150 demux_err:
2151         while (--i >= 0) {
2152                 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2153                 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2154         }
2155         mlx4_ib_device_unregister_sysfs(dev);
2156
2157 sysfs_err:
2158         mlx4_ib_destroy_alias_guid_service(dev);
2159
2160 paravirt_err:
2161         mlx4_ib_cm_paravirt_clean(dev, -1);
2162
2163         return err;
2164 }
2165
2166 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2167 {
2168         int i;
2169         unsigned long flags;
2170
2171         if (!mlx4_is_mfunc(dev->dev))
2172                 return;
2173
2174         spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2175         dev->sriov.is_going_down = 1;
2176         spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2177         if (mlx4_is_master(dev->dev)) {
2178                 for (i = 0; i < dev->num_ports; i++) {
2179                         flush_workqueue(dev->sriov.demux[i].ud_wq);
2180                         mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2181                         kfree(dev->sriov.sqps[i]);
2182                         dev->sriov.sqps[i] = NULL;
2183                         mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2184                 }
2185
2186                 mlx4_ib_cm_paravirt_clean(dev, -1);
2187                 mlx4_ib_destroy_alias_guid_service(dev);
2188                 mlx4_ib_device_unregister_sysfs(dev);
2189         }
2190 }