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 <linux/mlx4/driver.h>
  44 #include "mlx4_ib.h"
  45
  46 enum {
  47         MLX4_IB_VENDOR_CLASS1 = 0x9,
  48         MLX4_IB_VENDOR_CLASS2 = 0xa
  49 };
  50
  51 #define MLX4_TUN_SEND_WRID_SHIFT 34
  52 #define MLX4_TUN_QPN_SHIFT 32
  53 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
  54 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
  55
  56 #define MLX4_TUN_IS_RECV(a)  (((a) >>  MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
  57 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
  58
  59  /* Port mgmt change event handling */
  60
  61 #define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
  62 #define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
  63 #define NUM_IDX_IN_PKEY_TBL_BLK 32
  64 #define GUID_TBL_ENTRY_SIZE 8      /* size in bytes */
  65 #define GUID_TBL_BLK_NUM_ENTRIES 8
  66 #define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
  67
  68 struct mlx4_mad_rcv_buf {
  69         struct ib_grh grh;
  70         u8 payload[256];
  71 } __packed;
  72
  73 struct mlx4_mad_snd_buf {
  74         u8 payload[256];
  75 } __packed;
  76
  77 struct mlx4_tunnel_mad {
  78         struct ib_grh grh;
  79         struct mlx4_ib_tunnel_header hdr;
  80         struct ib_mad mad;
  81 } __packed;
  82
  83 struct mlx4_rcv_tunnel_mad {
  84         struct mlx4_rcv_tunnel_hdr hdr;
  85         struct ib_grh grh;
  86         struct ib_mad mad;
  87 } __packed;
  88
  89 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
  90 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num);
  91 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
  92                                 int block, u32 change_bitmap);
  93
  94 __be64 mlx4_ib_gen_node_guid(void)
  95 {
  96 #define NODE_GUID_HI    ((u64) (((u64)IB_OPENIB_OUI) << 40))
  97         return cpu_to_be64(NODE_GUID_HI | prandom_u32());
  98 }
  99
 100 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
 101 {
 102         return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
 103                 cpu_to_be64(0xff00000000000000LL);
 104 }
 105
 106 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
 107                  int port, const struct ib_wc *in_wc,
 108                  const struct ib_grh *in_grh,
 109                  const void *in_mad, void *response_mad)
 110 {
 111         struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
 112         void *inbox;
 113         int err;
 114         u32 in_modifier = port;
 115         u8 op_modifier = 0;
 116
 117         inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
 118         if (IS_ERR(inmailbox))
 119                 return PTR_ERR(inmailbox);
 120         inbox = inmailbox->buf;
 121
 122         outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
 123         if (IS_ERR(outmailbox)) {
 124                 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
 125                 return PTR_ERR(outmailbox);
 126         }
 127
 128         memcpy(inbox, in_mad, 256);
 129
 130         /*
 131          * Key check traps can't be generated unless we have in_wc to
 132          * tell us where to send the trap.
 133          */
 134         if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
 135                 op_modifier |= 0x1;
 136         if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
 137                 op_modifier |= 0x2;
 138         if (mlx4_is_mfunc(dev->dev) &&
 139             (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
 140                 op_modifier |= 0x8;
 141
 142         if (in_wc) {
 143                 struct {
 144                         __be32          my_qpn;
 145                         u32             reserved1;
 146                         __be32          rqpn;
 147                         u8              sl;
 148                         u8              g_path;
 149                         u16             reserved2[2];
 150                         __be16          pkey;
 151                         u32             reserved3[11];
 152                         u8              grh[40];
 153                 } *ext_info;
 154
 155                 memset(inbox + 256, 0, 256);
 156                 ext_info = inbox + 256;
 157
 158                 ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
 159                 ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
 160                 ext_info->sl     = in_wc->sl << 4;
 161                 ext_info->g_path = in_wc->dlid_path_bits |
 162                         (in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
 163                 ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
 164
 165                 if (in_grh)
 166                         memcpy(ext_info->grh, in_grh, 40);
 167
 168                 op_modifier |= 0x4;
 169
 170                 in_modifier |= in_wc->slid << 16;
 171         }
 172
 173         err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
 174                            mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
 175                            MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
 176                            (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
 177
 178         if (!err)
 179                 memcpy(response_mad, outmailbox->buf, 256);
 180
 181         mlx4_free_cmd_mailbox(dev->dev, inmailbox);
 182         mlx4_free_cmd_mailbox(dev->dev, outmailbox);
 183
 184         return err;
 185 }
 186
 187 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
 188 {
 189         struct ib_ah *new_ah;
 190         struct ib_ah_attr ah_attr;
 191         unsigned long flags;
 192
 193         if (!dev->send_agent[port_num - 1][0])
 194                 return;
 195
 196         memset(&ah_attr, 0, sizeof ah_attr);
 197         ah_attr.dlid     = lid;
 198         ah_attr.sl       = sl;
 199         ah_attr.port_num = port_num;
 200
 201         new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
 202                               &ah_attr);
 203         if (IS_ERR(new_ah))
 204                 return;
 205
 206         spin_lock_irqsave(&dev->sm_lock, flags);
 207         if (dev->sm_ah[port_num - 1])
 208                 ib_destroy_ah(dev->sm_ah[port_num - 1]);
 209         dev->sm_ah[port_num - 1] = new_ah;
 210         spin_unlock_irqrestore(&dev->sm_lock, flags);
 211 }
 212
 213 /*
 214  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
 215  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
 216  */
 217 static void smp_snoop(struct ib_device *ibdev, u8 port_num, const struct ib_mad *mad,
 218                       u16 prev_lid)
 219 {
 220         struct ib_port_info *pinfo;
 221         u16 lid;
 222         __be16 *base;
 223         u32 bn, pkey_change_bitmap;
 224         int i;
 225
 226
 227         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 228         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 229              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 230             mad->mad_hdr.method == IB_MGMT_METHOD_SET)
 231                 switch (mad->mad_hdr.attr_id) {
 232                 case IB_SMP_ATTR_PORT_INFO:
 233                         pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
 234                         lid = be16_to_cpu(pinfo->lid);
 235
 236                         update_sm_ah(dev, port_num,
 237                                      be16_to_cpu(pinfo->sm_lid),
 238                                      pinfo->neighbormtu_mastersmsl & 0xf);
 239
 240                         if (pinfo->clientrereg_resv_subnetto & 0x80)
 241                                 handle_client_rereg_event(dev, port_num);
 242
 243                         if (prev_lid != lid)
 244                                 handle_lid_change_event(dev, port_num);
 245                         break;
 246
 247                 case IB_SMP_ATTR_PKEY_TABLE:
 248                         if (!mlx4_is_mfunc(dev->dev)) {
 249                                 mlx4_ib_dispatch_event(dev, port_num,
 250                                                        IB_EVENT_PKEY_CHANGE);
 251                                 break;
 252                         }
 253
 254                         /* at this point, we are running in the master.
 255                          * Slaves do not receive SMPs.
 256                          */
 257                         bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
 258                         base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
 259                         pkey_change_bitmap = 0;
 260                         for (i = 0; i < 32; i++) {
 261                                 pr_debug("PKEY[%d] = x%x\n",
 262                                          i + bn*32, be16_to_cpu(base[i]));
 263                                 if (be16_to_cpu(base[i]) !=
 264                                     dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
 265                                         pkey_change_bitmap |= (1 << i);
 266                                         dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
 267                                                 be16_to_cpu(base[i]);
 268                                 }
 269                         }
 270                         pr_debug("PKEY Change event: port=%d, "
 271                                  "block=0x%x, change_bitmap=0x%x\n",
 272                                  port_num, bn, pkey_change_bitmap);
 273
 274                         if (pkey_change_bitmap) {
 275                                 mlx4_ib_dispatch_event(dev, port_num,
 276                                                        IB_EVENT_PKEY_CHANGE);
 277                                 if (!dev->sriov.is_going_down)
 278                                         __propagate_pkey_ev(dev, port_num, bn,
 279                                                             pkey_change_bitmap);
 280                         }
 281                         break;
 282
 283                 case IB_SMP_ATTR_GUID_INFO:
 284                         /* paravirtualized master's guid is guid 0 -- does not change */
 285                         if (!mlx4_is_master(dev->dev))
 286                                 mlx4_ib_dispatch_event(dev, port_num,
 287                                                        IB_EVENT_GID_CHANGE);
 288                         /*if master, notify relevant slaves*/
 289                         if (mlx4_is_master(dev->dev) &&
 290                             !dev->sriov.is_going_down) {
 291                                 bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
 292                                 mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
 293                                                                     (u8 *)(&((struct ib_smp *)mad)->data));
 294                                 mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
 295                                                                      (u8 *)(&((struct ib_smp *)mad)->data));
 296                         }
 297                         break;
 298
 299                 default:
 300                         break;
 301                 }
 302 }
 303
 304 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
 305                                 int block, u32 change_bitmap)
 306 {
 307         int i, ix, slave, err;
 308         int have_event = 0;
 309
 310         for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
 311                 if (slave == mlx4_master_func_num(dev->dev))
 312                         continue;
 313                 if (!mlx4_is_slave_active(dev->dev, slave))
 314                         continue;
 315
 316                 have_event = 0;
 317                 for (i = 0; i < 32; i++) {
 318                         if (!(change_bitmap & (1 << i)))
 319                                 continue;
 320                         for (ix = 0;
 321                              ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
 322                                 if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
 323                                     [ix] == i + 32 * block) {
 324                                         err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
 325                                         pr_debug("propagate_pkey_ev: slave %d,"
 326                                                  " port %d, ix %d (%d)\n",
 327                                                  slave, port_num, ix, err);
 328                                         have_event = 1;
 329                                         break;
 330                                 }
 331                         }
 332                         if (have_event)
 333                                 break;
 334                 }
 335         }
 336 }
 337
 338 static void node_desc_override(struct ib_device *dev,
 339                                struct ib_mad *mad)
 340 {
 341         unsigned long flags;
 342
 343         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 344              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 345             mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
 346             mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
 347                 spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
 348                 memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
 349                 spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
 350         }
 351 }
 352
 353 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
 354 {
 355         int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
 356         struct ib_mad_send_buf *send_buf;
 357         struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
 358         int ret;
 359         unsigned long flags;
 360
 361         if (agent) {
 362                 send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
 363                                               IB_MGMT_MAD_DATA, GFP_ATOMIC,
 364                                               IB_MGMT_BASE_VERSION);
 365                 if (IS_ERR(send_buf))
 366                         return;
 367                 /*
 368                  * We rely here on the fact that MLX QPs don't use the
 369                  * address handle after the send is posted (this is
 370                  * wrong following the IB spec strictly, but we know
 371                  * it's OK for our devices).
 372                  */
 373                 spin_lock_irqsave(&dev->sm_lock, flags);
 374                 memcpy(send_buf->mad, mad, sizeof *mad);
 375                 if ((send_buf->ah = dev->sm_ah[port_num - 1]))
 376                         ret = ib_post_send_mad(send_buf, NULL);
 377                 else
 378                         ret = -EINVAL;
 379                 spin_unlock_irqrestore(&dev->sm_lock, flags);
 380
 381                 if (ret)
 382                         ib_free_send_mad(send_buf);
 383         }
 384 }
 385
 386 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
 387                                                              struct ib_sa_mad *sa_mad)
 388 {
 389         int ret = 0;
 390
 391         /* dispatch to different sa handlers */
 392         switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
 393         case IB_SA_ATTR_MC_MEMBER_REC:
 394                 ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
 395                 break;
 396         default:
 397                 break;
 398         }
 399         return ret;
 400 }
 401
 402 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
 403 {
 404         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 405         int i;
 406
 407         for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
 408                 if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
 409                         return i;
 410         }
 411         return -1;
 412 }
 413
 414
 415 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
 416                                    u8 port, u16 pkey, u16 *ix)
 417 {
 418         int i, ret;
 419         u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
 420         u16 slot_pkey;
 421
 422         if (slave == mlx4_master_func_num(dev->dev))
 423                 return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
 424
 425         unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
 426
 427         for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
 428                 if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
 429                         continue;
 430
 431                 pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
 432
 433                 ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
 434                 if (ret)
 435                         continue;
 436                 if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
 437                         if (slot_pkey & 0x8000) {
 438                                 *ix = (u16) pkey_ix;
 439                                 return 0;
 440                         } else {
 441                                 /* take first partial pkey index found */
 442                                 if (partial_ix == 0xFF)
 443                                         partial_ix = pkey_ix;
 444                         }
 445                 }
 446         }
 447
 448         if (partial_ix < 0xFF) {
 449                 *ix = (u16) partial_ix;
 450                 return 0;
 451         }
 452
 453         return -EINVAL;
 454 }
 455
 456 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
 457                           enum ib_qp_type dest_qpt, struct ib_wc *wc,
 458                           struct ib_grh *grh, struct ib_mad *mad)
 459 {
 460         struct ib_sge list;
 461         struct ib_ud_wr wr;
 462         struct ib_send_wr *bad_wr;
 463         struct mlx4_ib_demux_pv_ctx *tun_ctx;
 464         struct mlx4_ib_demux_pv_qp *tun_qp;
 465         struct mlx4_rcv_tunnel_mad *tun_mad;
 466         struct ib_ah_attr attr;
 467         struct ib_ah *ah;
 468         struct ib_qp *src_qp = NULL;
 469         unsigned tun_tx_ix = 0;
 470         int dqpn;
 471         int ret = 0;
 472         u16 tun_pkey_ix;
 473         u16 cached_pkey;
 474         u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
 475
 476         if (dest_qpt > IB_QPT_GSI)
 477                 return -EINVAL;
 478
 479         tun_ctx = dev->sriov.demux[port-1].tun[slave];
 480
 481         /* check if proxy qp created */
 482         if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
 483                 return -EAGAIN;
 484
 485         if (!dest_qpt)
 486                 tun_qp = &tun_ctx->qp[0];
 487         else
 488                 tun_qp = &tun_ctx->qp[1];
 489
 490         /* compute P_Key index to put in tunnel header for slave */
 491         if (dest_qpt) {
 492                 u16 pkey_ix;
 493                 ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
 494                 if (ret)
 495                         return -EINVAL;
 496
 497                 ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
 498                 if (ret)
 499                         return -EINVAL;
 500                 tun_pkey_ix = pkey_ix;
 501         } else
 502                 tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
 503
 504         dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
 505
 506         /* get tunnel tx data buf for slave */
 507         src_qp = tun_qp->qp;
 508
 509         /* create ah. Just need an empty one with the port num for the post send.
 510          * The driver will set the force loopback bit in post_send */
 511         memset(&attr, 0, sizeof attr);
 512         attr.port_num = port;
 513         if (is_eth) {
 514                 memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
 515                 attr.ah_flags = IB_AH_GRH;
 516         }
 517         ah = ib_create_ah(tun_ctx->pd, &attr);
 518         if (IS_ERR(ah))
 519                 return -ENOMEM;
 520
 521         /* allocate tunnel tx buf after pass failure returns */
 522         spin_lock(&tun_qp->tx_lock);
 523         if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
 524             (MLX4_NUM_TUNNEL_BUFS - 1))
 525                 ret = -EAGAIN;
 526         else
 527                 tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
 528         spin_unlock(&tun_qp->tx_lock);
 529         if (ret)
 530                 goto out;
 531
 532         tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
 533         if (tun_qp->tx_ring[tun_tx_ix].ah)
 534                 ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
 535         tun_qp->tx_ring[tun_tx_ix].ah = ah;
 536         ib_dma_sync_single_for_cpu(&dev->ib_dev,
 537                                    tun_qp->tx_ring[tun_tx_ix].buf.map,
 538                                    sizeof (struct mlx4_rcv_tunnel_mad),
 539                                    DMA_TO_DEVICE);
 540
 541         /* copy over to tunnel buffer */
 542         if (grh)
 543                 memcpy(&tun_mad->grh, grh, sizeof *grh);
 544         memcpy(&tun_mad->mad, mad, sizeof *mad);
 545
 546         /* adjust tunnel data */
 547         tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
 548         tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
 549         tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
 550
 551         if (is_eth) {
 552                 u16 vlan = 0;
 553                 if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
 554                                                 NULL)) {
 555                         /* VST mode */
 556                         if (vlan != wc->vlan_id)
 557                                 /* Packet vlan is not the VST-assigned vlan.
 558                                  * Drop the packet.
 559                                  */
 560                                 goto out;
 561                          else
 562                                 /* Remove the vlan tag before forwarding
 563                                  * the packet to the VF.
 564                                  */
 565                                 vlan = 0xffff;
 566                 } else {
 567                         vlan = wc->vlan_id;
 568                 }
 569
 570                 tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
 571                 memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
 572                 memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
 573         } else {
 574                 tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
 575                 tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
 576         }
 577
 578         ib_dma_sync_single_for_device(&dev->ib_dev,
 579                                       tun_qp->tx_ring[tun_tx_ix].buf.map,
 580                                       sizeof (struct mlx4_rcv_tunnel_mad),
 581                                       DMA_TO_DEVICE);
 582
 583         list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
 584         list.length = sizeof (struct mlx4_rcv_tunnel_mad);
 585         list.lkey = tun_ctx->pd->local_dma_lkey;
 586
 587         wr.ah = ah;
 588         wr.port_num = port;
 589         wr.remote_qkey = IB_QP_SET_QKEY;
 590         wr.remote_qpn = dqpn;
 591         wr.wr.next = NULL;
 592         wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
 593         wr.wr.sg_list = &list;
 594         wr.wr.num_sge = 1;
 595         wr.wr.opcode = IB_WR_SEND;
 596         wr.wr.send_flags = IB_SEND_SIGNALED;
 597
 598         ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
 599 out:
 600         if (ret)
 601                 ib_destroy_ah(ah);
 602         return ret;
 603 }
 604
 605 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
 606                         struct ib_wc *wc, struct ib_grh *grh,
 607                         struct ib_mad *mad)
 608 {
 609         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 610         int err, other_port;
 611         int slave = -1;
 612         u8 *slave_id;
 613         int is_eth = 0;
 614
 615         if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
 616                 is_eth = 0;
 617         else
 618                 is_eth = 1;
 619
 620         if (is_eth) {
 621                 if (!(wc->wc_flags & IB_WC_GRH)) {
 622                         mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
 623                         return -EINVAL;
 624                 }
 625                 if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
 626                         mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
 627                         return -EINVAL;
 628                 }
 629                 err = mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave);
 630                 if (err && mlx4_is_mf_bonded(dev->dev)) {
 631                         other_port = (port == 1) ? 2 : 1;
 632                         err = mlx4_get_slave_from_roce_gid(dev->dev, other_port, grh->dgid.raw, &slave);
 633                         if (!err) {
 634                                 port = other_port;
 635                                 pr_debug("resolved slave %d from gid %pI6 wire port %d other %d\n",
 636                                          slave, grh->dgid.raw, port, other_port);
 637                         }
 638                 }
 639                 if (err) {
 640                         mlx4_ib_warn(ibdev, "failed matching grh\n");
 641                         return -ENOENT;
 642                 }
 643                 if (slave >= dev->dev->caps.sqp_demux) {
 644                         mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
 645                                      slave, dev->dev->caps.sqp_demux);
 646                         return -ENOENT;
 647                 }
 648
 649                 if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
 650                         return 0;
 651
 652                 err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
 653                 if (err)
 654                         pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
 655                                  slave, err);
 656                 return 0;
 657         }
 658
 659         /* Initially assume that this mad is for us */
 660         slave = mlx4_master_func_num(dev->dev);
 661
 662         /* See if the slave id is encoded in a response mad */
 663         if (mad->mad_hdr.method & 0x80) {
 664                 slave_id = (u8 *) &mad->mad_hdr.tid;
 665                 slave = *slave_id;
 666                 if (slave != 255) /*255 indicates the dom0*/
 667                         *slave_id = 0; /* remap tid */
 668         }
 669
 670         /* If a grh is present, we demux according to it */
 671         if (wc->wc_flags & IB_WC_GRH) {
 672                 slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
 673                 if (slave < 0) {
 674                         mlx4_ib_warn(ibdev, "failed matching grh\n");
 675                         return -ENOENT;
 676                 }
 677         }
 678         /* Class-specific handling */
 679         switch (mad->mad_hdr.mgmt_class) {
 680         case IB_MGMT_CLASS_SUBN_LID_ROUTED:
 681         case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
 682                 /* 255 indicates the dom0 */
 683                 if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
 684                         if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
 685                                 return -EPERM;
 686                         /* for a VF. drop unsolicited MADs */
 687                         if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
 688                                 mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
 689                                              slave, mad->mad_hdr.mgmt_class,
 690                                              mad->mad_hdr.method);
 691                                 return -EINVAL;
 692                         }
 693                 }
 694                 break;
 695         case IB_MGMT_CLASS_SUBN_ADM:
 696                 if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
 697                                              (struct ib_sa_mad *) mad))
 698                         return 0;
 699                 break;
 700         case IB_MGMT_CLASS_CM:
 701                 if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
 702                         return 0;
 703                 break;
 704         case IB_MGMT_CLASS_DEVICE_MGMT:
 705                 if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
 706                         return 0;
 707                 break;
 708         default:
 709                 /* Drop unsupported classes for slaves in tunnel mode */
 710                 if (slave != mlx4_master_func_num(dev->dev)) {
 711                         pr_debug("dropping unsupported ingress mad from class:%d "
 712                                  "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
 713                         return 0;
 714                 }
 715         }
 716         /*make sure that no slave==255 was not handled yet.*/
 717         if (slave >= dev->dev->caps.sqp_demux) {
 718                 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
 719                              slave, dev->dev->caps.sqp_demux);
 720                 return -ENOENT;
 721         }
 722
 723         err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
 724         if (err)
 725                 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
 726                          slave, err);
 727         return 0;
 728 }
 729
 730 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 731                         const struct ib_wc *in_wc, const struct ib_grh *in_grh,
 732                         const struct ib_mad *in_mad, struct ib_mad *out_mad)
 733 {
 734         u16 slid, prev_lid = 0;
 735         int err;
 736         struct ib_port_attr pattr;
 737
 738         if (in_wc && in_wc->qp->qp_num) {
 739                 pr_debug("received MAD: slid:%d sqpn:%d "
 740                         "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
 741                         in_wc->slid, in_wc->src_qp,
 742                         in_wc->dlid_path_bits,
 743                         in_wc->qp->qp_num,
 744                         in_wc->wc_flags,
 745                         in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
 746                         be16_to_cpu(in_mad->mad_hdr.attr_id));
 747                 if (in_wc->wc_flags & IB_WC_GRH) {
 748                         pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
 749                                  be64_to_cpu(in_grh->sgid.global.subnet_prefix),
 750                                  be64_to_cpu(in_grh->sgid.global.interface_id));
 751                         pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
 752                                  be64_to_cpu(in_grh->dgid.global.subnet_prefix),
 753                                  be64_to_cpu(in_grh->dgid.global.interface_id));
 754                 }
 755         }
 756
 757         slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
 758
 759         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
 760                 forward_trap(to_mdev(ibdev), port_num, in_mad);
 761                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
 762         }
 763
 764         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 765             in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 766                 if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
 767                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
 768                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
 769                         return IB_MAD_RESULT_SUCCESS;
 770
 771                 /*
 772                  * Don't process SMInfo queries -- the SMA can't handle them.
 773                  */
 774                 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
 775                         return IB_MAD_RESULT_SUCCESS;
 776         } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
 777                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
 778                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
 779                    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
 780                 if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
 781                     in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
 782                         return IB_MAD_RESULT_SUCCESS;
 783         } else
 784                 return IB_MAD_RESULT_SUCCESS;
 785
 786         if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 787              in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 788             in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
 789             in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
 790             !ib_query_port(ibdev, port_num, &pattr))
 791                 prev_lid = pattr.lid;
 792
 793         err = mlx4_MAD_IFC(to_mdev(ibdev),
 794                            (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
 795                            (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
 796                            MLX4_MAD_IFC_NET_VIEW,
 797                            port_num, in_wc, in_grh, in_mad, out_mad);
 798         if (err)
 799                 return IB_MAD_RESULT_FAILURE;
 800
 801         if (!out_mad->mad_hdr.status) {
 802                 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
 803                         smp_snoop(ibdev, port_num, in_mad, prev_lid);
 804                 /* slaves get node desc from FW */
 805                 if (!mlx4_is_slave(to_mdev(ibdev)->dev))
 806                         node_desc_override(ibdev, out_mad);
 807         }
 808
 809         /* set return bit in status of directed route responses */
 810         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 811                 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
 812
 813         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
 814                 /* no response for trap repress */
 815                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
 816
 817         return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
 818 }
 819
 820 static void edit_counter(struct mlx4_counter *cnt,
 821                                         struct ib_pma_portcounters *pma_cnt)
 822 {
 823         ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
 824                              (be64_to_cpu(cnt->tx_bytes) >> 2));
 825         ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
 826                              (be64_to_cpu(cnt->rx_bytes) >> 2));
 827         ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
 828                              be64_to_cpu(cnt->tx_frames));
 829         ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
 830                              be64_to_cpu(cnt->rx_frames));
 831 }
 832
 833 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 834                         const struct ib_wc *in_wc, const struct ib_grh *in_grh,
 835                         const struct ib_mad *in_mad, struct ib_mad *out_mad)
 836 {
 837         struct mlx4_counter counter_stats;
 838         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 839         struct counter_index *tmp_counter;
 840         int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
 841
 842         if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
 843                 return -EINVAL;
 844
 845         memset(&counter_stats, 0, sizeof(counter_stats));
 846         mutex_lock(&dev->counters_table[port_num - 1].mutex);
 847         list_for_each_entry(tmp_counter,
 848                             &dev->counters_table[port_num - 1].counters_list,
 849                             list) {
 850                 err = mlx4_get_counter_stats(dev->dev,
 851                                              tmp_counter->index,
 852                                              &counter_stats, 0);
 853                 if (err) {
 854                         err = IB_MAD_RESULT_FAILURE;
 855                         stats_avail = 0;
 856                         break;
 857                 }
 858                 stats_avail = 1;
 859         }
 860         mutex_unlock(&dev->counters_table[port_num - 1].mutex);
 861         if (stats_avail) {
 862                 memset(out_mad->data, 0, sizeof out_mad->data);
 863                 switch (counter_stats.counter_mode & 0xf) {
 864                 case 0:
 865                         edit_counter(&counter_stats,
 866                                      (void *)(out_mad->data + 40));
 867                         err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
 868                         break;
 869                 default:
 870                         err = IB_MAD_RESULT_FAILURE;
 871                 }
 872         }
 873
 874         return err;
 875 }
 876
 877 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 878                         const struct ib_wc *in_wc, const struct ib_grh *in_grh,
 879                         const struct ib_mad_hdr *in, size_t in_mad_size,
 880                         struct ib_mad_hdr *out, size_t *out_mad_size,
 881                         u16 *out_mad_pkey_index)
 882 {
 883         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 884         const struct ib_mad *in_mad = (const struct ib_mad *)in;
 885         struct ib_mad *out_mad = (struct ib_mad *)out;
 886         enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
 887
 888         if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
 889                          *out_mad_size != sizeof(*out_mad)))
 890                 return IB_MAD_RESULT_FAILURE;
 891
 892         /* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
 893          * queries, should be called only by VFs and for that specific purpose
 894          */
 895         if (link == IB_LINK_LAYER_INFINIBAND) {
 896                 if (mlx4_is_slave(dev->dev) &&
 897                     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
 898                     in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS)
 899                         return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
 900                                                 in_grh, in_mad, out_mad);
 901
 902                 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
 903                                       in_grh, in_mad, out_mad);
 904         }
 905
 906         if (link == IB_LINK_LAYER_ETHERNET)
 907                 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
 908                                         in_grh, in_mad, out_mad);
 909
 910         return -EINVAL;
 911 }
 912
 913 static void send_handler(struct ib_mad_agent *agent,
 914                          struct ib_mad_send_wc *mad_send_wc)
 915 {
 916         if (mad_send_wc->send_buf->context[0])
 917                 ib_destroy_ah(mad_send_wc->send_buf->context[0]);
 918         ib_free_send_mad(mad_send_wc->send_buf);
 919 }
 920
 921 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
 922 {
 923         struct ib_mad_agent *agent;
 924         int p, q;
 925         int ret;
 926         enum rdma_link_layer ll;
 927
 928         for (p = 0; p < dev->num_ports; ++p) {
 929                 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
 930                 for (q = 0; q <= 1; ++q) {
 931                         if (ll == IB_LINK_LAYER_INFINIBAND) {
 932                                 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
 933                                                               q ? IB_QPT_GSI : IB_QPT_SMI,
 934                                                               NULL, 0, send_handler,
 935                                                               NULL, NULL, 0);
 936                                 if (IS_ERR(agent)) {
 937                                         ret = PTR_ERR(agent);
 938                                         goto err;
 939                                 }
 940                                 dev->send_agent[p][q] = agent;
 941                         } else
 942                                 dev->send_agent[p][q] = NULL;
 943                 }
 944         }
 945
 946         return 0;
 947
 948 err:
 949         for (p = 0; p < dev->num_ports; ++p)
 950                 for (q = 0; q <= 1; ++q)
 951                         if (dev->send_agent[p][q])
 952                                 ib_unregister_mad_agent(dev->send_agent[p][q]);
 953
 954         return ret;
 955 }
 956
 957 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
 958 {
 959         struct ib_mad_agent *agent;
 960         int p, q;
 961
 962         for (p = 0; p < dev->num_ports; ++p) {
 963                 for (q = 0; q <= 1; ++q) {
 964                         agent = dev->send_agent[p][q];
 965                         if (agent) {
 966                                 dev->send_agent[p][q] = NULL;
 967                                 ib_unregister_mad_agent(agent);
 968                         }
 969                 }
 970
 971                 if (dev->sm_ah[p])
 972                         ib_destroy_ah(dev->sm_ah[p]);
 973         }
 974 }
 975
 976 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
 977 {
 978         mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
 979
 980         if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
 981                 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
 982                                             MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
 983 }
 984
 985 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
 986 {
 987         /* re-configure the alias-guid and mcg's */
 988         if (mlx4_is_master(dev->dev)) {
 989                 mlx4_ib_invalidate_all_guid_record(dev, port_num);
 990
 991                 if (!dev->sriov.is_going_down) {
 992                         mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
 993                         mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
 994                                                     MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
 995                 }
 996         }
 997         mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
 998 }
 999
1000 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1001                               struct mlx4_eqe *eqe)
1002 {
1003         __propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1004                             GET_MASK_FROM_EQE(eqe));
1005 }
1006
1007 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1008                                       u32 guid_tbl_blk_num, u32 change_bitmap)
1009 {
1010         struct ib_smp *in_mad  = NULL;
1011         struct ib_smp *out_mad  = NULL;
1012         u16 i;
1013
1014         if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1015                 return;
1016
1017         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1018         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1019         if (!in_mad || !out_mad) {
1020                 mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
1021                 goto out;
1022         }
1023
1024         guid_tbl_blk_num  *= 4;
1025
1026         for (i = 0; i < 4; i++) {
1027                 if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1028                         continue;
1029                 memset(in_mad, 0, sizeof *in_mad);
1030                 memset(out_mad, 0, sizeof *out_mad);
1031
1032                 in_mad->base_version  = 1;
1033                 in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1034                 in_mad->class_version = 1;
1035                 in_mad->method        = IB_MGMT_METHOD_GET;
1036                 in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1037                 in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1038
1039                 if (mlx4_MAD_IFC(dev,
1040                                  MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1041                                  port_num, NULL, NULL, in_mad, out_mad)) {
1042                         mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1043                         goto out;
1044                 }
1045
1046                 mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1047                                                     port_num,
1048                                                     (u8 *)(&((struct ib_smp *)out_mad)->data));
1049                 mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1050                                                      port_num,
1051                                                      (u8 *)(&((struct ib_smp *)out_mad)->data));
1052         }
1053
1054 out:
1055         kfree(in_mad);
1056         kfree(out_mad);
1057         return;
1058 }
1059
1060 void handle_port_mgmt_change_event(struct work_struct *work)
1061 {
1062         struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1063         struct mlx4_ib_dev *dev = ew->ib_dev;
1064         struct mlx4_eqe *eqe = &(ew->ib_eqe);
1065         u8 port = eqe->event.port_mgmt_change.port;
1066         u32 changed_attr;
1067         u32 tbl_block;
1068         u32 change_bitmap;
1069
1070         switch (eqe->subtype) {
1071         case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1072                 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1073
1074                 /* Update the SM ah - This should be done before handling
1075                    the other changed attributes so that MADs can be sent to the SM */
1076                 if (changed_attr & MSTR_SM_CHANGE_MASK) {
1077                         u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1078                         u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1079                         update_sm_ah(dev, port, lid, sl);
1080                 }
1081
1082                 /* Check if it is a lid change event */
1083                 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1084                         handle_lid_change_event(dev, port);
1085
1086                 /* Generate GUID changed event */
1087                 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1088                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1089                         /*if master, notify all slaves*/
1090                         if (mlx4_is_master(dev->dev))
1091                                 mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1092                                                             MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1093                 }
1094
1095                 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1096                         handle_client_rereg_event(dev, port);
1097                 break;
1098
1099         case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1100                 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1101                 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1102                         propagate_pkey_ev(dev, port, eqe);
1103                 break;
1104         case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1105                 /* paravirtualized master's guid is guid 0 -- does not change */
1106                 if (!mlx4_is_master(dev->dev))
1107                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1108                 /*if master, notify relevant slaves*/
1109                 else if (!dev->sriov.is_going_down) {
1110                         tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1111                         change_bitmap = GET_MASK_FROM_EQE(eqe);
1112                         handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1113                 }
1114                 break;
1115         default:
1116                 pr_warn("Unsupported subtype 0x%x for "
1117                         "Port Management Change event\n", eqe->subtype);
1118         }
1119
1120         kfree(ew);
1121 }
1122
1123 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1124                             enum ib_event_type type)
1125 {
1126         struct ib_event event;
1127
1128         event.device            = &dev->ib_dev;
1129         event.element.port_num  = port_num;
1130         event.event             = type;
1131
1132         ib_dispatch_event(&event);
1133 }
1134
1135 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1136 {
1137         unsigned long flags;
1138         struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1139         struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1140         spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1141         if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1142                 queue_work(ctx->wq, &ctx->work);
1143         spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1144 }
1145
1146 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1147                                   struct mlx4_ib_demux_pv_qp *tun_qp,
1148                                   int index)
1149 {
1150         struct ib_sge sg_list;
1151         struct ib_recv_wr recv_wr, *bad_recv_wr;
1152         int size;
1153
1154         size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1155                 sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1156
1157         sg_list.addr = tun_qp->ring[index].map;
1158         sg_list.length = size;
1159         sg_list.lkey = ctx->pd->local_dma_lkey;
1160
1161         recv_wr.next = NULL;
1162         recv_wr.sg_list = &sg_list;
1163         recv_wr.num_sge = 1;
1164         recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1165                 MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1166         ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1167                                       size, DMA_FROM_DEVICE);
1168         return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1169 }
1170
1171 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1172                 int slave, struct ib_sa_mad *sa_mad)
1173 {
1174         int ret = 0;
1175
1176         /* dispatch to different sa handlers */
1177         switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1178         case IB_SA_ATTR_MC_MEMBER_REC:
1179                 ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1180                 break;
1181         default:
1182                 break;
1183         }
1184         return ret;
1185 }
1186
1187 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1188 {
1189         int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1190
1191         return (qpn >= proxy_start && qpn <= proxy_start + 1);
1192 }
1193
1194
1195 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1196                          enum ib_qp_type dest_qpt, u16 pkey_index,
1197                          u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1198                          u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1199 {
1200         struct ib_sge list;
1201         struct ib_ud_wr wr;
1202         struct ib_send_wr *bad_wr;
1203         struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1204         struct mlx4_ib_demux_pv_qp *sqp;
1205         struct mlx4_mad_snd_buf *sqp_mad;
1206         struct ib_ah *ah;
1207         struct ib_qp *send_qp = NULL;
1208         unsigned wire_tx_ix = 0;
1209         int ret = 0;
1210         u16 wire_pkey_ix;
1211         int src_qpnum;
1212         u8 sgid_index;
1213
1214
1215         sqp_ctx = dev->sriov.sqps[port-1];
1216
1217         /* check if proxy qp created */
1218         if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1219                 return -EAGAIN;
1220
1221         if (dest_qpt == IB_QPT_SMI) {
1222                 src_qpnum = 0;
1223                 sqp = &sqp_ctx->qp[0];
1224                 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1225         } else {
1226                 src_qpnum = 1;
1227                 sqp = &sqp_ctx->qp[1];
1228                 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1229         }
1230
1231         send_qp = sqp->qp;
1232
1233         /* create ah */
1234         sgid_index = attr->grh.sgid_index;
1235         attr->grh.sgid_index = 0;
1236         ah = ib_create_ah(sqp_ctx->pd, attr);
1237         if (IS_ERR(ah))
1238                 return -ENOMEM;
1239         attr->grh.sgid_index = sgid_index;
1240         to_mah(ah)->av.ib.gid_index = sgid_index;
1241         /* get rid of force-loopback bit */
1242         to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1243         spin_lock(&sqp->tx_lock);
1244         if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1245             (MLX4_NUM_TUNNEL_BUFS - 1))
1246                 ret = -EAGAIN;
1247         else
1248                 wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1249         spin_unlock(&sqp->tx_lock);
1250         if (ret)
1251                 goto out;
1252
1253         sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1254         if (sqp->tx_ring[wire_tx_ix].ah)
1255                 ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1256         sqp->tx_ring[wire_tx_ix].ah = ah;
1257         ib_dma_sync_single_for_cpu(&dev->ib_dev,
1258                                    sqp->tx_ring[wire_tx_ix].buf.map,
1259                                    sizeof (struct mlx4_mad_snd_buf),
1260                                    DMA_TO_DEVICE);
1261
1262         memcpy(&sqp_mad->payload, mad, sizeof *mad);
1263
1264         ib_dma_sync_single_for_device(&dev->ib_dev,
1265                                       sqp->tx_ring[wire_tx_ix].buf.map,
1266                                       sizeof (struct mlx4_mad_snd_buf),
1267                                       DMA_TO_DEVICE);
1268
1269         list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1270         list.length = sizeof (struct mlx4_mad_snd_buf);
1271         list.lkey = sqp_ctx->pd->local_dma_lkey;
1272
1273         wr.ah = ah;
1274         wr.port_num = port;
1275         wr.pkey_index = wire_pkey_ix;
1276         wr.remote_qkey = qkey;
1277         wr.remote_qpn = remote_qpn;
1278         wr.wr.next = NULL;
1279         wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1280         wr.wr.sg_list = &list;
1281         wr.wr.num_sge = 1;
1282         wr.wr.opcode = IB_WR_SEND;
1283         wr.wr.send_flags = IB_SEND_SIGNALED;
1284         if (s_mac)
1285                 memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1286         if (vlan_id < 0x1000)
1287                 vlan_id |= (attr->sl & 7) << 13;
1288         to_mah(ah)->av.eth.vlan = cpu_to_be16(vlan_id);
1289
1290
1291         ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1292 out:
1293         if (ret)
1294                 ib_destroy_ah(ah);
1295         return ret;
1296 }
1297
1298 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1299 {
1300         if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1301                 return slave;
1302         return mlx4_get_base_gid_ix(dev->dev, slave, port);
1303 }
1304
1305 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1306                                     struct ib_ah_attr *ah_attr)
1307 {
1308         if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1309                 ah_attr->grh.sgid_index = slave;
1310         else
1311                 ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1312 }
1313
1314 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1315 {
1316         struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1317         struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1318         int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1319         struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1320         struct mlx4_ib_ah ah;
1321         struct ib_ah_attr ah_attr;
1322         u8 *slave_id;
1323         int slave;
1324         int port;
1325         u16 vlan_id;
1326
1327         /* Get slave that sent this packet */
1328         if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1329             wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1330             (wc->src_qp & 0x1) != ctx->port - 1 ||
1331             wc->src_qp & 0x4) {
1332                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1333                 return;
1334         }
1335         slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1336         if (slave != ctx->slave) {
1337                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1338                              "belongs to another slave\n", wc->src_qp);
1339                 return;
1340         }
1341
1342         /* Map transaction ID */
1343         ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1344                                    sizeof (struct mlx4_tunnel_mad),
1345                                    DMA_FROM_DEVICE);
1346         switch (tunnel->mad.mad_hdr.method) {
1347         case IB_MGMT_METHOD_SET:
1348         case IB_MGMT_METHOD_GET:
1349         case IB_MGMT_METHOD_REPORT:
1350         case IB_SA_METHOD_GET_TABLE:
1351         case IB_SA_METHOD_DELETE:
1352         case IB_SA_METHOD_GET_MULTI:
1353         case IB_SA_METHOD_GET_TRACE_TBL:
1354                 slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1355                 if (*slave_id) {
1356                         mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1357                                      "class:%d slave:%d\n", *slave_id,
1358                                      tunnel->mad.mad_hdr.mgmt_class, slave);
1359                         return;
1360                 } else
1361                         *slave_id = slave;
1362         default:
1363                 /* nothing */;
1364         }
1365
1366         /* Class-specific handling */
1367         switch (tunnel->mad.mad_hdr.mgmt_class) {
1368         case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1369         case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1370                 if (slave != mlx4_master_func_num(dev->dev) &&
1371                     !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1372                         return;
1373                 break;
1374         case IB_MGMT_CLASS_SUBN_ADM:
1375                 if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1376                               (struct ib_sa_mad *) &tunnel->mad))
1377                         return;
1378                 break;
1379         case IB_MGMT_CLASS_CM:
1380                 if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1381                               (struct ib_mad *) &tunnel->mad))
1382                         return;
1383                 break;
1384         case IB_MGMT_CLASS_DEVICE_MGMT:
1385                 if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1386                     tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1387                         return;
1388                 break;
1389         default:
1390                 /* Drop unsupported classes for slaves in tunnel mode */
1391                 if (slave != mlx4_master_func_num(dev->dev)) {
1392                         mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1393                                      "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1394                         return;
1395                 }
1396         }
1397
1398         /* We are using standard ib_core services to send the mad, so generate a
1399          * stadard address handle by decoding the tunnelled mlx4_ah fields */
1400         memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1401         ah.ibah.device = ctx->ib_dev;
1402
1403         port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1404         port = mlx4_slave_convert_port(dev->dev, slave, port);
1405         if (port < 0)
1406                 return;
1407         ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1408
1409         mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1410         if (ah_attr.ah_flags & IB_AH_GRH)
1411                 fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1412
1413         memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1414         vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1415         /* if slave have default vlan use it */
1416         mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1417                                     &vlan_id, &ah_attr.sl);
1418
1419         mlx4_ib_send_to_wire(dev, slave, ctx->port,
1420                              is_proxy_qp0(dev, wc->src_qp, slave) ?
1421                              IB_QPT_SMI : IB_QPT_GSI,
1422                              be16_to_cpu(tunnel->hdr.pkey_index),
1423                              be32_to_cpu(tunnel->hdr.remote_qpn),
1424                              be32_to_cpu(tunnel->hdr.qkey),
1425                              &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1426 }
1427
1428 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1429                                  enum ib_qp_type qp_type, int is_tun)
1430 {
1431         int i;
1432         struct mlx4_ib_demux_pv_qp *tun_qp;
1433         int rx_buf_size, tx_buf_size;
1434
1435         if (qp_type > IB_QPT_GSI)
1436                 return -EINVAL;
1437
1438         tun_qp = &ctx->qp[qp_type];
1439
1440         tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1441                                GFP_KERNEL);
1442         if (!tun_qp->ring)
1443                 return -ENOMEM;
1444
1445         tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1446                                   sizeof (struct mlx4_ib_tun_tx_buf),
1447                                   GFP_KERNEL);
1448         if (!tun_qp->tx_ring) {
1449                 kfree(tun_qp->ring);
1450                 tun_qp->ring = NULL;
1451                 return -ENOMEM;
1452         }
1453
1454         if (is_tun) {
1455                 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1456                 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1457         } else {
1458                 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1459                 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1460         }
1461
1462         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1463                 tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1464                 if (!tun_qp->ring[i].addr)
1465                         goto err;
1466                 tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1467                                                         tun_qp->ring[i].addr,
1468                                                         rx_buf_size,
1469                                                         DMA_FROM_DEVICE);
1470                 if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1471                         kfree(tun_qp->ring[i].addr);
1472                         goto err;
1473                 }
1474         }
1475
1476         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1477                 tun_qp->tx_ring[i].buf.addr =
1478                         kmalloc(tx_buf_size, GFP_KERNEL);
1479                 if (!tun_qp->tx_ring[i].buf.addr)
1480                         goto tx_err;
1481                 tun_qp->tx_ring[i].buf.map =
1482                         ib_dma_map_single(ctx->ib_dev,
1483                                           tun_qp->tx_ring[i].buf.addr,
1484                                           tx_buf_size,
1485                                           DMA_TO_DEVICE);
1486                 if (ib_dma_mapping_error(ctx->ib_dev,
1487                                          tun_qp->tx_ring[i].buf.map)) {
1488                         kfree(tun_qp->tx_ring[i].buf.addr);
1489                         goto tx_err;
1490                 }
1491                 tun_qp->tx_ring[i].ah = NULL;
1492         }
1493         spin_lock_init(&tun_qp->tx_lock);
1494         tun_qp->tx_ix_head = 0;
1495         tun_qp->tx_ix_tail = 0;
1496         tun_qp->proxy_qpt = qp_type;
1497
1498         return 0;
1499
1500 tx_err:
1501         while (i > 0) {
1502                 --i;
1503                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1504                                     tx_buf_size, DMA_TO_DEVICE);
1505                 kfree(tun_qp->tx_ring[i].buf.addr);
1506         }
1507         kfree(tun_qp->tx_ring);
1508         tun_qp->tx_ring = NULL;
1509         i = MLX4_NUM_TUNNEL_BUFS;
1510 err:
1511         while (i > 0) {
1512                 --i;
1513                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1514                                     rx_buf_size, DMA_FROM_DEVICE);
1515                 kfree(tun_qp->ring[i].addr);
1516         }
1517         kfree(tun_qp->ring);
1518         tun_qp->ring = NULL;
1519         return -ENOMEM;
1520 }
1521
1522 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1523                                      enum ib_qp_type qp_type, int is_tun)
1524 {
1525         int i;
1526         struct mlx4_ib_demux_pv_qp *tun_qp;
1527         int rx_buf_size, tx_buf_size;
1528
1529         if (qp_type > IB_QPT_GSI)
1530                 return;
1531
1532         tun_qp = &ctx->qp[qp_type];
1533         if (is_tun) {
1534                 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1535                 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1536         } else {
1537                 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1538                 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1539         }
1540
1541
1542         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1543                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1544                                     rx_buf_size, DMA_FROM_DEVICE);
1545                 kfree(tun_qp->ring[i].addr);
1546         }
1547
1548         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1549                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1550                                     tx_buf_size, DMA_TO_DEVICE);
1551                 kfree(tun_qp->tx_ring[i].buf.addr);
1552                 if (tun_qp->tx_ring[i].ah)
1553                         ib_destroy_ah(tun_qp->tx_ring[i].ah);
1554         }
1555         kfree(tun_qp->tx_ring);
1556         kfree(tun_qp->ring);
1557 }
1558
1559 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1560 {
1561         struct mlx4_ib_demux_pv_ctx *ctx;
1562         struct mlx4_ib_demux_pv_qp *tun_qp;
1563         struct ib_wc wc;
1564         int ret;
1565         ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1566         ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1567
1568         while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1569                 tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1570                 if (wc.status == IB_WC_SUCCESS) {
1571                         switch (wc.opcode) {
1572                         case IB_WC_RECV:
1573                                 mlx4_ib_multiplex_mad(ctx, &wc);
1574                                 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1575                                                              wc.wr_id &
1576                                                              (MLX4_NUM_TUNNEL_BUFS - 1));
1577                                 if (ret)
1578                                         pr_err("Failed reposting tunnel "
1579                                                "buf:%lld\n", wc.wr_id);
1580                                 break;
1581                         case IB_WC_SEND:
1582                                 pr_debug("received tunnel send completion:"
1583                                          "wrid=0x%llx, status=0x%x\n",
1584                                          wc.wr_id, wc.status);
1585                                 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1586                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1587                                 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1588                                         = NULL;
1589                                 spin_lock(&tun_qp->tx_lock);
1590                                 tun_qp->tx_ix_tail++;
1591                                 spin_unlock(&tun_qp->tx_lock);
1592
1593                                 break;
1594                         default:
1595                                 break;
1596                         }
1597                 } else  {
1598                         pr_debug("mlx4_ib: completion error in tunnel: %d."
1599                                  " status = %d, wrid = 0x%llx\n",
1600                                  ctx->slave, wc.status, wc.wr_id);
1601                         if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1602                                 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1603                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1604                                 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1605                                         = NULL;
1606                                 spin_lock(&tun_qp->tx_lock);
1607                                 tun_qp->tx_ix_tail++;
1608                                 spin_unlock(&tun_qp->tx_lock);
1609                         }
1610                 }
1611         }
1612 }
1613
1614 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1615 {
1616         struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1617
1618         /* It's worse than that! He's dead, Jim! */
1619         pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1620                event->event, sqp->port);
1621 }
1622
1623 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1624                             enum ib_qp_type qp_type, int create_tun)
1625 {
1626         int i, ret;
1627         struct mlx4_ib_demux_pv_qp *tun_qp;
1628         struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1629         struct ib_qp_attr attr;
1630         int qp_attr_mask_INIT;
1631
1632         if (qp_type > IB_QPT_GSI)
1633                 return -EINVAL;
1634
1635         tun_qp = &ctx->qp[qp_type];
1636
1637         memset(&qp_init_attr, 0, sizeof qp_init_attr);
1638         qp_init_attr.init_attr.send_cq = ctx->cq;
1639         qp_init_attr.init_attr.recv_cq = ctx->cq;
1640         qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1641         qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1642         qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1643         qp_init_attr.init_attr.cap.max_send_sge = 1;
1644         qp_init_attr.init_attr.cap.max_recv_sge = 1;
1645         if (create_tun) {
1646                 qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1647                 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1648                 qp_init_attr.port = ctx->port;
1649                 qp_init_attr.slave = ctx->slave;
1650                 qp_init_attr.proxy_qp_type = qp_type;
1651                 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1652                            IB_QP_QKEY | IB_QP_PORT;
1653         } else {
1654                 qp_init_attr.init_attr.qp_type = qp_type;
1655                 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1656                 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1657         }
1658         qp_init_attr.init_attr.port_num = ctx->port;
1659         qp_init_attr.init_attr.qp_context = ctx;
1660         qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1661         tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1662         if (IS_ERR(tun_qp->qp)) {
1663                 ret = PTR_ERR(tun_qp->qp);
1664                 tun_qp->qp = NULL;
1665                 pr_err("Couldn't create %s QP (%d)\n",
1666                        create_tun ? "tunnel" : "special", ret);
1667                 return ret;
1668         }
1669
1670         memset(&attr, 0, sizeof attr);
1671         attr.qp_state = IB_QPS_INIT;
1672         ret = 0;
1673         if (create_tun)
1674                 ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1675                                               ctx->port, IB_DEFAULT_PKEY_FULL,
1676                                               &attr.pkey_index);
1677         if (ret || !create_tun)
1678                 attr.pkey_index =
1679                         to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1680         attr.qkey = IB_QP1_QKEY;
1681         attr.port_num = ctx->port;
1682         ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1683         if (ret) {
1684                 pr_err("Couldn't change %s qp state to INIT (%d)\n",
1685                        create_tun ? "tunnel" : "special", ret);
1686                 goto err_qp;
1687         }
1688         attr.qp_state = IB_QPS_RTR;
1689         ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1690         if (ret) {
1691                 pr_err("Couldn't change %s qp state to RTR (%d)\n",
1692                        create_tun ? "tunnel" : "special", ret);
1693                 goto err_qp;
1694         }
1695         attr.qp_state = IB_QPS_RTS;
1696         attr.sq_psn = 0;
1697         ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1698         if (ret) {
1699                 pr_err("Couldn't change %s qp state to RTS (%d)\n",
1700                        create_tun ? "tunnel" : "special", ret);
1701                 goto err_qp;
1702         }
1703
1704         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1705                 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1706                 if (ret) {
1707                         pr_err(" mlx4_ib_post_pv_buf error"
1708                                " (err = %d, i = %d)\n", ret, i);
1709                         goto err_qp;
1710                 }
1711         }
1712         return 0;
1713
1714 err_qp:
1715         ib_destroy_qp(tun_qp->qp);
1716         tun_qp->qp = NULL;
1717         return ret;
1718 }
1719
1720 /*
1721  * IB MAD completion callback for real SQPs
1722  */
1723 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1724 {
1725         struct mlx4_ib_demux_pv_ctx *ctx;
1726         struct mlx4_ib_demux_pv_qp *sqp;
1727         struct ib_wc wc;
1728         struct ib_grh *grh;
1729         struct ib_mad *mad;
1730
1731         ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1732         ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1733
1734         while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1735                 sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1736                 if (wc.status == IB_WC_SUCCESS) {
1737                         switch (wc.opcode) {
1738                         case IB_WC_SEND:
1739                                 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1740                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1741                                 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1742                                         = NULL;
1743                                 spin_lock(&sqp->tx_lock);
1744                                 sqp->tx_ix_tail++;
1745                                 spin_unlock(&sqp->tx_lock);
1746                                 break;
1747                         case IB_WC_RECV:
1748                                 mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1749                                                 (sqp->ring[wc.wr_id &
1750                                                 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1751                                 grh = &(((struct mlx4_mad_rcv_buf *)
1752                                                 (sqp->ring[wc.wr_id &
1753                                                 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1754                                 mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1755                                 if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1756                                                            (MLX4_NUM_TUNNEL_BUFS - 1)))
1757                                         pr_err("Failed reposting SQP "
1758                                                "buf:%lld\n", wc.wr_id);
1759                                 break;
1760                         default:
1761                                 BUG_ON(1);
1762                                 break;
1763                         }
1764                 } else  {
1765                         pr_debug("mlx4_ib: completion error in tunnel: %d."
1766                                  " status = %d, wrid = 0x%llx\n",
1767                                  ctx->slave, wc.status, wc.wr_id);
1768                         if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1769                                 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1770                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1771                                 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1772                                         = NULL;
1773                                 spin_lock(&sqp->tx_lock);
1774                                 sqp->tx_ix_tail++;
1775                                 spin_unlock(&sqp->tx_lock);
1776                         }
1777                 }
1778         }
1779 }
1780
1781 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1782                                struct mlx4_ib_demux_pv_ctx **ret_ctx)
1783 {
1784         struct mlx4_ib_demux_pv_ctx *ctx;
1785
1786         *ret_ctx = NULL;
1787         ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1788         if (!ctx) {
1789                 pr_err("failed allocating pv resource context "
1790                        "for port %d, slave %d\n", port, slave);
1791                 return -ENOMEM;
1792         }
1793
1794         ctx->ib_dev = &dev->ib_dev;
1795         ctx->port = port;
1796         ctx->slave = slave;
1797         *ret_ctx = ctx;
1798         return 0;
1799 }
1800
1801 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1802 {
1803         if (dev->sriov.demux[port - 1].tun[slave]) {
1804                 kfree(dev->sriov.demux[port - 1].tun[slave]);
1805                 dev->sriov.demux[port - 1].tun[slave] = NULL;
1806         }
1807 }
1808
1809 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1810                                int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1811 {
1812         int ret, cq_size;
1813         struct ib_cq_init_attr cq_attr = {};
1814
1815         if (ctx->state != DEMUX_PV_STATE_DOWN)
1816                 return -EEXIST;
1817
1818         ctx->state = DEMUX_PV_STATE_STARTING;
1819         /* have QP0 only if link layer is IB */
1820         if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1821             IB_LINK_LAYER_INFINIBAND)
1822                 ctx->has_smi = 1;
1823
1824         if (ctx->has_smi) {
1825                 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1826                 if (ret) {
1827                         pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1828                         goto err_out;
1829                 }
1830         }
1831
1832         ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1833         if (ret) {
1834                 pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1835                 goto err_out_qp0;
1836         }
1837
1838         cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1839         if (ctx->has_smi)
1840                 cq_size *= 2;
1841
1842         cq_attr.cqe = cq_size;
1843         ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1844                                NULL, ctx, &cq_attr);
1845         if (IS_ERR(ctx->cq)) {
1846                 ret = PTR_ERR(ctx->cq);
1847                 pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1848                 goto err_buf;
1849         }
1850
1851         ctx->pd = ib_alloc_pd(ctx->ib_dev);
1852         if (IS_ERR(ctx->pd)) {
1853                 ret = PTR_ERR(ctx->pd);
1854                 pr_err("Couldn't create tunnel PD (%d)\n", ret);
1855                 goto err_cq;
1856         }
1857
1858         if (ctx->has_smi) {
1859                 ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1860                 if (ret) {
1861                         pr_err("Couldn't create %s QP0 (%d)\n",
1862                                create_tun ? "tunnel for" : "",  ret);
1863                         goto err_pd;
1864                 }
1865         }
1866
1867         ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1868         if (ret) {
1869                 pr_err("Couldn't create %s QP1 (%d)\n",
1870                        create_tun ? "tunnel for" : "",  ret);
1871                 goto err_qp0;
1872         }
1873
1874         if (create_tun)
1875                 INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1876         else
1877                 INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1878
1879         ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1880
1881         ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1882         if (ret) {
1883                 pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1884                 goto err_wq;
1885         }
1886         ctx->state = DEMUX_PV_STATE_ACTIVE;
1887         return 0;
1888
1889 err_wq:
1890         ctx->wq = NULL;
1891         ib_destroy_qp(ctx->qp[1].qp);
1892         ctx->qp[1].qp = NULL;
1893
1894
1895 err_qp0:
1896         if (ctx->has_smi)
1897                 ib_destroy_qp(ctx->qp[0].qp);
1898         ctx->qp[0].qp = NULL;
1899
1900 err_pd:
1901         ib_dealloc_pd(ctx->pd);
1902         ctx->pd = NULL;
1903
1904 err_cq:
1905         ib_destroy_cq(ctx->cq);
1906         ctx->cq = NULL;
1907
1908 err_buf:
1909         mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1910
1911 err_out_qp0:
1912         if (ctx->has_smi)
1913                 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1914 err_out:
1915         ctx->state = DEMUX_PV_STATE_DOWN;
1916         return ret;
1917 }
1918
1919 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1920                                  struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1921 {
1922         if (!ctx)
1923                 return;
1924         if (ctx->state > DEMUX_PV_STATE_DOWN) {
1925                 ctx->state = DEMUX_PV_STATE_DOWNING;
1926                 if (flush)
1927                         flush_workqueue(ctx->wq);
1928                 if (ctx->has_smi) {
1929                         ib_destroy_qp(ctx->qp[0].qp);
1930                         ctx->qp[0].qp = NULL;
1931                         mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1932                 }
1933                 ib_destroy_qp(ctx->qp[1].qp);
1934                 ctx->qp[1].qp = NULL;
1935                 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1936                 ib_dealloc_pd(ctx->pd);
1937                 ctx->pd = NULL;
1938                 ib_destroy_cq(ctx->cq);
1939                 ctx->cq = NULL;
1940                 ctx->state = DEMUX_PV_STATE_DOWN;
1941         }
1942 }
1943
1944 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1945                                   int port, int do_init)
1946 {
1947         int ret = 0;
1948
1949         if (!do_init) {
1950                 clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1951                 /* for master, destroy real sqp resources */
1952                 if (slave == mlx4_master_func_num(dev->dev))
1953                         destroy_pv_resources(dev, slave, port,
1954                                              dev->sriov.sqps[port - 1], 1);
1955                 /* destroy the tunnel qp resources */
1956                 destroy_pv_resources(dev, slave, port,
1957                                      dev->sriov.demux[port - 1].tun[slave], 1);
1958                 return 0;
1959         }
1960
1961         /* create the tunnel qp resources */
1962         ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1963                                   dev->sriov.demux[port - 1].tun[slave]);
1964
1965         /* for master, create the real sqp resources */
1966         if (!ret && slave == mlx4_master_func_num(dev->dev))
1967                 ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1968                                           dev->sriov.sqps[port - 1]);
1969         return ret;
1970 }
1971
1972 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1973 {
1974         struct mlx4_ib_demux_work *dmxw;
1975
1976         dmxw = container_of(work, struct mlx4_ib_demux_work, work);
1977         mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
1978                                dmxw->do_init);
1979         kfree(dmxw);
1980         return;
1981 }
1982
1983 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
1984                                        struct mlx4_ib_demux_ctx *ctx,
1985                                        int port)
1986 {
1987         char name[12];
1988         int ret = 0;
1989         int i;
1990
1991         ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
1992                            sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
1993         if (!ctx->tun)
1994                 return -ENOMEM;
1995
1996         ctx->dev = dev;
1997         ctx->port = port;
1998         ctx->ib_dev = &dev->ib_dev;
1999
2000         for (i = 0;
2001              i < min(dev->dev->caps.sqp_demux,
2002              (u16)(dev->dev->persist->num_vfs + 1));
2003              i++) {
2004                 struct mlx4_active_ports actv_ports =
2005                         mlx4_get_active_ports(dev->dev, i);
2006
2007                 if (!test_bit(port - 1, actv_ports.ports))
2008                         continue;
2009
2010                 ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2011                 if (ret) {
2012                         ret = -ENOMEM;
2013                         goto err_mcg;
2014                 }
2015         }
2016
2017         ret = mlx4_ib_mcg_port_init(ctx);
2018         if (ret) {
2019                 pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2020                 goto err_mcg;
2021         }
2022
2023         snprintf(name, sizeof name, "mlx4_ibt%d", port);
2024         ctx->wq = create_singlethread_workqueue(name);
2025         if (!ctx->wq) {
2026                 pr_err("Failed to create tunnelling WQ for port %d\n", port);
2027                 ret = -ENOMEM;
2028                 goto err_wq;
2029         }
2030
2031         snprintf(name, sizeof name, "mlx4_ibud%d", port);
2032         ctx->ud_wq = create_singlethread_workqueue(name);
2033         if (!ctx->ud_wq) {
2034                 pr_err("Failed to create up/down WQ for port %d\n", port);
2035                 ret = -ENOMEM;
2036                 goto err_udwq;
2037         }
2038
2039         return 0;
2040
2041 err_udwq:
2042         destroy_workqueue(ctx->wq);
2043         ctx->wq = NULL;
2044
2045 err_wq:
2046         mlx4_ib_mcg_port_cleanup(ctx, 1);
2047 err_mcg:
2048         for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2049                 free_pv_object(dev, i, port);
2050         kfree(ctx->tun);
2051         ctx->tun = NULL;
2052         return ret;
2053 }
2054
2055 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2056 {
2057         if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2058                 sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2059                 flush_workqueue(sqp_ctx->wq);
2060                 if (sqp_ctx->has_smi) {
2061                         ib_destroy_qp(sqp_ctx->qp[0].qp);
2062                         sqp_ctx->qp[0].qp = NULL;
2063                         mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2064                 }
2065                 ib_destroy_qp(sqp_ctx->qp[1].qp);
2066                 sqp_ctx->qp[1].qp = NULL;
2067                 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2068                 ib_dealloc_pd(sqp_ctx->pd);
2069                 sqp_ctx->pd = NULL;
2070                 ib_destroy_cq(sqp_ctx->cq);
2071                 sqp_ctx->cq = NULL;
2072                 sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2073         }
2074 }
2075
2076 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2077 {
2078         int i;
2079         if (ctx) {
2080                 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2081                 mlx4_ib_mcg_port_cleanup(ctx, 1);
2082                 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2083                         if (!ctx->tun[i])
2084                                 continue;
2085                         if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2086                                 ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2087                 }
2088                 flush_workqueue(ctx->wq);
2089                 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2090                         destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2091                         free_pv_object(dev, i, ctx->port);
2092                 }
2093                 kfree(ctx->tun);
2094                 destroy_workqueue(ctx->ud_wq);
2095                 destroy_workqueue(ctx->wq);
2096         }
2097 }
2098
2099 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2100 {
2101         int i;
2102
2103         if (!mlx4_is_master(dev->dev))
2104                 return;
2105         /* initialize or tear down tunnel QPs for the master */
2106         for (i = 0; i < dev->dev->caps.num_ports; i++)
2107                 mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2108         return;
2109 }
2110
2111 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2112 {
2113         int i = 0;
2114         int err;
2115
2116         if (!mlx4_is_mfunc(dev->dev))
2117                 return 0;
2118
2119         dev->sriov.is_going_down = 0;
2120         spin_lock_init(&dev->sriov.going_down_lock);
2121         mlx4_ib_cm_paravirt_init(dev);
2122
2123         mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2124
2125         if (mlx4_is_slave(dev->dev)) {
2126                 mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2127                 return 0;
2128         }
2129
2130         for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2131                 if (i == mlx4_master_func_num(dev->dev))
2132                         mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2133                 else
2134                         mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2135         }
2136
2137         err = mlx4_ib_init_alias_guid_service(dev);
2138         if (err) {
2139                 mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2140                 goto paravirt_err;
2141         }
2142         err = mlx4_ib_device_register_sysfs(dev);
2143         if (err) {
2144                 mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2145                 goto sysfs_err;
2146         }
2147
2148         mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2149                      dev->dev->caps.sqp_demux);
2150         for (i = 0; i < dev->num_ports; i++) {
2151                 union ib_gid gid;
2152                 err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2153                 if (err)
2154                         goto demux_err;
2155                 dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2156                 err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2157                                       &dev->sriov.sqps[i]);
2158                 if (err)
2159                         goto demux_err;
2160                 err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2161                 if (err)
2162                         goto free_pv;
2163         }
2164         mlx4_ib_master_tunnels(dev, 1);
2165         return 0;
2166
2167 free_pv:
2168         free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2169 demux_err:
2170         while (--i >= 0) {
2171                 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2172                 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2173         }
2174         mlx4_ib_device_unregister_sysfs(dev);
2175
2176 sysfs_err:
2177         mlx4_ib_destroy_alias_guid_service(dev);
2178
2179 paravirt_err:
2180         mlx4_ib_cm_paravirt_clean(dev, -1);
2181
2182         return err;
2183 }
2184
2185 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2186 {
2187         int i;
2188         unsigned long flags;
2189
2190         if (!mlx4_is_mfunc(dev->dev))
2191                 return;
2192
2193         spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2194         dev->sriov.is_going_down = 1;
2195         spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2196         if (mlx4_is_master(dev->dev)) {
2197                 for (i = 0; i < dev->num_ports; i++) {
2198                         flush_workqueue(dev->sriov.demux[i].ud_wq);
2199                         mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2200                         kfree(dev->sriov.sqps[i]);
2201                         dev->sriov.sqps[i] = NULL;
2202                         mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2203                 }
2204
2205                 mlx4_ib_cm_paravirt_clean(dev, -1);
2206                 mlx4_ib_destroy_alias_guid_service(dev);
2207                 mlx4_ib_device_unregister_sysfs(dev);
2208         }
2209 }