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, void *counters,
 821                          __be16 attr_id)
 822 {
 823         switch (attr_id) {
 824         case IB_PMA_PORT_COUNTERS:
 825         {
 826                 struct ib_pma_portcounters *pma_cnt =
 827                         (struct ib_pma_portcounters *)counters;
 828
 829                 ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
 830                                      (be64_to_cpu(cnt->tx_bytes) >> 2));
 831                 ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
 832                                      (be64_to_cpu(cnt->rx_bytes) >> 2));
 833                 ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
 834                                      be64_to_cpu(cnt->tx_frames));
 835                 ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
 836                                      be64_to_cpu(cnt->rx_frames));
 837                 break;
 838         }
 839         case IB_PMA_PORT_COUNTERS_EXT:
 840         {
 841                 struct ib_pma_portcounters_ext *pma_cnt_ext =
 842                         (struct ib_pma_portcounters_ext *)counters;
 843
 844                 pma_cnt_ext->port_xmit_data =
 845                         cpu_to_be64(be64_to_cpu(cnt->tx_bytes) >> 2);
 846                 pma_cnt_ext->port_rcv_data =
 847                         cpu_to_be64(be64_to_cpu(cnt->rx_bytes) >> 2);
 848                 pma_cnt_ext->port_xmit_packets = cnt->tx_frames;
 849                 pma_cnt_ext->port_rcv_packets = cnt->rx_frames;
 850                 break;
 851         }
 852         }
 853 }
 854
 855 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 856                         const struct ib_wc *in_wc, const struct ib_grh *in_grh,
 857                         const struct ib_mad *in_mad, struct ib_mad *out_mad)
 858 {
 859         struct mlx4_counter counter_stats;
 860         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 861         struct counter_index *tmp_counter;
 862         int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
 863
 864         if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
 865                 return -EINVAL;
 866
 867         memset(&counter_stats, 0, sizeof(counter_stats));
 868         mutex_lock(&dev->counters_table[port_num - 1].mutex);
 869         list_for_each_entry(tmp_counter,
 870                             &dev->counters_table[port_num - 1].counters_list,
 871                             list) {
 872                 err = mlx4_get_counter_stats(dev->dev,
 873                                              tmp_counter->index,
 874                                              &counter_stats, 0);
 875                 if (err) {
 876                         err = IB_MAD_RESULT_FAILURE;
 877                         stats_avail = 0;
 878                         break;
 879                 }
 880                 stats_avail = 1;
 881         }
 882         mutex_unlock(&dev->counters_table[port_num - 1].mutex);
 883         if (stats_avail) {
 884                 memset(out_mad->data, 0, sizeof out_mad->data);
 885                 switch (counter_stats.counter_mode & 0xf) {
 886                 case 0:
 887                         edit_counter(&counter_stats,
 888                                      (void *)(out_mad->data + 40),
 889                                      in_mad->mad_hdr.attr_id);
 890                         err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
 891                         break;
 892                 default:
 893                         err = IB_MAD_RESULT_FAILURE;
 894                 }
 895         }
 896
 897         return err;
 898 }
 899
 900 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 901                         const struct ib_wc *in_wc, const struct ib_grh *in_grh,
 902                         const struct ib_mad_hdr *in, size_t in_mad_size,
 903                         struct ib_mad_hdr *out, size_t *out_mad_size,
 904                         u16 *out_mad_pkey_index)
 905 {
 906         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 907         const struct ib_mad *in_mad = (const struct ib_mad *)in;
 908         struct ib_mad *out_mad = (struct ib_mad *)out;
 909         enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
 910
 911         if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
 912                          *out_mad_size != sizeof(*out_mad)))
 913                 return IB_MAD_RESULT_FAILURE;
 914
 915         /* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
 916          * queries, should be called only by VFs and for that specific purpose
 917          */
 918         if (link == IB_LINK_LAYER_INFINIBAND) {
 919                 if (mlx4_is_slave(dev->dev) &&
 920                     (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
 921                      (in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
 922                       in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT)))
 923                         return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
 924                                                 in_grh, in_mad, out_mad);
 925
 926                 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
 927                                       in_grh, in_mad, out_mad);
 928         }
 929
 930         if (link == IB_LINK_LAYER_ETHERNET)
 931                 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
 932                                         in_grh, in_mad, out_mad);
 933
 934         return -EINVAL;
 935 }
 936
 937 static void send_handler(struct ib_mad_agent *agent,
 938                          struct ib_mad_send_wc *mad_send_wc)
 939 {
 940         if (mad_send_wc->send_buf->context[0])
 941                 ib_destroy_ah(mad_send_wc->send_buf->context[0]);
 942         ib_free_send_mad(mad_send_wc->send_buf);
 943 }
 944
 945 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
 946 {
 947         struct ib_mad_agent *agent;
 948         int p, q;
 949         int ret;
 950         enum rdma_link_layer ll;
 951
 952         for (p = 0; p < dev->num_ports; ++p) {
 953                 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
 954                 for (q = 0; q <= 1; ++q) {
 955                         if (ll == IB_LINK_LAYER_INFINIBAND) {
 956                                 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
 957                                                               q ? IB_QPT_GSI : IB_QPT_SMI,
 958                                                               NULL, 0, send_handler,
 959                                                               NULL, NULL, 0);
 960                                 if (IS_ERR(agent)) {
 961                                         ret = PTR_ERR(agent);
 962                                         goto err;
 963                                 }
 964                                 dev->send_agent[p][q] = agent;
 965                         } else
 966                                 dev->send_agent[p][q] = NULL;
 967                 }
 968         }
 969
 970         return 0;
 971
 972 err:
 973         for (p = 0; p < dev->num_ports; ++p)
 974                 for (q = 0; q <= 1; ++q)
 975                         if (dev->send_agent[p][q])
 976                                 ib_unregister_mad_agent(dev->send_agent[p][q]);
 977
 978         return ret;
 979 }
 980
 981 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
 982 {
 983         struct ib_mad_agent *agent;
 984         int p, q;
 985
 986         for (p = 0; p < dev->num_ports; ++p) {
 987                 for (q = 0; q <= 1; ++q) {
 988                         agent = dev->send_agent[p][q];
 989                         if (agent) {
 990                                 dev->send_agent[p][q] = NULL;
 991                                 ib_unregister_mad_agent(agent);
 992                         }
 993                 }
 994
 995                 if (dev->sm_ah[p])
 996                         ib_destroy_ah(dev->sm_ah[p]);
 997         }
 998 }
 999
1000 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
1001 {
1002         mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1003
1004         if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1005                 mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1006                                             MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
1007 }
1008
1009 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
1010 {
1011         /* re-configure the alias-guid and mcg's */
1012         if (mlx4_is_master(dev->dev)) {
1013                 mlx4_ib_invalidate_all_guid_record(dev, port_num);
1014
1015                 if (!dev->sriov.is_going_down) {
1016                         mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1017                         mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1018                                                     MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
1019                 }
1020         }
1021         mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
1022 }
1023
1024 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1025                               struct mlx4_eqe *eqe)
1026 {
1027         __propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1028                             GET_MASK_FROM_EQE(eqe));
1029 }
1030
1031 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1032                                       u32 guid_tbl_blk_num, u32 change_bitmap)
1033 {
1034         struct ib_smp *in_mad  = NULL;
1035         struct ib_smp *out_mad  = NULL;
1036         u16 i;
1037
1038         if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1039                 return;
1040
1041         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1042         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1043         if (!in_mad || !out_mad) {
1044                 mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
1045                 goto out;
1046         }
1047
1048         guid_tbl_blk_num  *= 4;
1049
1050         for (i = 0; i < 4; i++) {
1051                 if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1052                         continue;
1053                 memset(in_mad, 0, sizeof *in_mad);
1054                 memset(out_mad, 0, sizeof *out_mad);
1055
1056                 in_mad->base_version  = 1;
1057                 in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1058                 in_mad->class_version = 1;
1059                 in_mad->method        = IB_MGMT_METHOD_GET;
1060                 in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1061                 in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1062
1063                 if (mlx4_MAD_IFC(dev,
1064                                  MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1065                                  port_num, NULL, NULL, in_mad, out_mad)) {
1066                         mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1067                         goto out;
1068                 }
1069
1070                 mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1071                                                     port_num,
1072                                                     (u8 *)(&((struct ib_smp *)out_mad)->data));
1073                 mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1074                                                      port_num,
1075                                                      (u8 *)(&((struct ib_smp *)out_mad)->data));
1076         }
1077
1078 out:
1079         kfree(in_mad);
1080         kfree(out_mad);
1081         return;
1082 }
1083
1084 void handle_port_mgmt_change_event(struct work_struct *work)
1085 {
1086         struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1087         struct mlx4_ib_dev *dev = ew->ib_dev;
1088         struct mlx4_eqe *eqe = &(ew->ib_eqe);
1089         u8 port = eqe->event.port_mgmt_change.port;
1090         u32 changed_attr;
1091         u32 tbl_block;
1092         u32 change_bitmap;
1093
1094         switch (eqe->subtype) {
1095         case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1096                 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1097
1098                 /* Update the SM ah - This should be done before handling
1099                    the other changed attributes so that MADs can be sent to the SM */
1100                 if (changed_attr & MSTR_SM_CHANGE_MASK) {
1101                         u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1102                         u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1103                         update_sm_ah(dev, port, lid, sl);
1104                 }
1105
1106                 /* Check if it is a lid change event */
1107                 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1108                         handle_lid_change_event(dev, port);
1109
1110                 /* Generate GUID changed event */
1111                 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1112                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1113                         /*if master, notify all slaves*/
1114                         if (mlx4_is_master(dev->dev))
1115                                 mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1116                                                             MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1117                 }
1118
1119                 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1120                         handle_client_rereg_event(dev, port);
1121                 break;
1122
1123         case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1124                 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1125                 if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1126                         propagate_pkey_ev(dev, port, eqe);
1127                 break;
1128         case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1129                 /* paravirtualized master's guid is guid 0 -- does not change */
1130                 if (!mlx4_is_master(dev->dev))
1131                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1132                 /*if master, notify relevant slaves*/
1133                 else if (!dev->sriov.is_going_down) {
1134                         tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1135                         change_bitmap = GET_MASK_FROM_EQE(eqe);
1136                         handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1137                 }
1138                 break;
1139         default:
1140                 pr_warn("Unsupported subtype 0x%x for "
1141                         "Port Management Change event\n", eqe->subtype);
1142         }
1143
1144         kfree(ew);
1145 }
1146
1147 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1148                             enum ib_event_type type)
1149 {
1150         struct ib_event event;
1151
1152         event.device            = &dev->ib_dev;
1153         event.element.port_num  = port_num;
1154         event.event             = type;
1155
1156         ib_dispatch_event(&event);
1157 }
1158
1159 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1160 {
1161         unsigned long flags;
1162         struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1163         struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1164         spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1165         if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1166                 queue_work(ctx->wq, &ctx->work);
1167         spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1168 }
1169
1170 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1171                                   struct mlx4_ib_demux_pv_qp *tun_qp,
1172                                   int index)
1173 {
1174         struct ib_sge sg_list;
1175         struct ib_recv_wr recv_wr, *bad_recv_wr;
1176         int size;
1177
1178         size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1179                 sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1180
1181         sg_list.addr = tun_qp->ring[index].map;
1182         sg_list.length = size;
1183         sg_list.lkey = ctx->pd->local_dma_lkey;
1184
1185         recv_wr.next = NULL;
1186         recv_wr.sg_list = &sg_list;
1187         recv_wr.num_sge = 1;
1188         recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1189                 MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1190         ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1191                                       size, DMA_FROM_DEVICE);
1192         return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1193 }
1194
1195 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1196                 int slave, struct ib_sa_mad *sa_mad)
1197 {
1198         int ret = 0;
1199
1200         /* dispatch to different sa handlers */
1201         switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1202         case IB_SA_ATTR_MC_MEMBER_REC:
1203                 ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1204                 break;
1205         default:
1206                 break;
1207         }
1208         return ret;
1209 }
1210
1211 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1212 {
1213         int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1214
1215         return (qpn >= proxy_start && qpn <= proxy_start + 1);
1216 }
1217
1218
1219 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1220                          enum ib_qp_type dest_qpt, u16 pkey_index,
1221                          u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1222                          u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1223 {
1224         struct ib_sge list;
1225         struct ib_ud_wr wr;
1226         struct ib_send_wr *bad_wr;
1227         struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1228         struct mlx4_ib_demux_pv_qp *sqp;
1229         struct mlx4_mad_snd_buf *sqp_mad;
1230         struct ib_ah *ah;
1231         struct ib_qp *send_qp = NULL;
1232         unsigned wire_tx_ix = 0;
1233         int ret = 0;
1234         u16 wire_pkey_ix;
1235         int src_qpnum;
1236         u8 sgid_index;
1237
1238
1239         sqp_ctx = dev->sriov.sqps[port-1];
1240
1241         /* check if proxy qp created */
1242         if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1243                 return -EAGAIN;
1244
1245         if (dest_qpt == IB_QPT_SMI) {
1246                 src_qpnum = 0;
1247                 sqp = &sqp_ctx->qp[0];
1248                 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1249         } else {
1250                 src_qpnum = 1;
1251                 sqp = &sqp_ctx->qp[1];
1252                 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1253         }
1254
1255         send_qp = sqp->qp;
1256
1257         /* create ah */
1258         sgid_index = attr->grh.sgid_index;
1259         attr->grh.sgid_index = 0;
1260         ah = ib_create_ah(sqp_ctx->pd, attr);
1261         if (IS_ERR(ah))
1262                 return -ENOMEM;
1263         attr->grh.sgid_index = sgid_index;
1264         to_mah(ah)->av.ib.gid_index = sgid_index;
1265         /* get rid of force-loopback bit */
1266         to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1267         spin_lock(&sqp->tx_lock);
1268         if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1269             (MLX4_NUM_TUNNEL_BUFS - 1))
1270                 ret = -EAGAIN;
1271         else
1272                 wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1273         spin_unlock(&sqp->tx_lock);
1274         if (ret)
1275                 goto out;
1276
1277         sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1278         if (sqp->tx_ring[wire_tx_ix].ah)
1279                 ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1280         sqp->tx_ring[wire_tx_ix].ah = ah;
1281         ib_dma_sync_single_for_cpu(&dev->ib_dev,
1282                                    sqp->tx_ring[wire_tx_ix].buf.map,
1283                                    sizeof (struct mlx4_mad_snd_buf),
1284                                    DMA_TO_DEVICE);
1285
1286         memcpy(&sqp_mad->payload, mad, sizeof *mad);
1287
1288         ib_dma_sync_single_for_device(&dev->ib_dev,
1289                                       sqp->tx_ring[wire_tx_ix].buf.map,
1290                                       sizeof (struct mlx4_mad_snd_buf),
1291                                       DMA_TO_DEVICE);
1292
1293         list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1294         list.length = sizeof (struct mlx4_mad_snd_buf);
1295         list.lkey = sqp_ctx->pd->local_dma_lkey;
1296
1297         wr.ah = ah;
1298         wr.port_num = port;
1299         wr.pkey_index = wire_pkey_ix;
1300         wr.remote_qkey = qkey;
1301         wr.remote_qpn = remote_qpn;
1302         wr.wr.next = NULL;
1303         wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1304         wr.wr.sg_list = &list;
1305         wr.wr.num_sge = 1;
1306         wr.wr.opcode = IB_WR_SEND;
1307         wr.wr.send_flags = IB_SEND_SIGNALED;
1308         if (s_mac)
1309                 memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1310         if (vlan_id < 0x1000)
1311                 vlan_id |= (attr->sl & 7) << 13;
1312         to_mah(ah)->av.eth.vlan = cpu_to_be16(vlan_id);
1313
1314
1315         ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1316 out:
1317         if (ret)
1318                 ib_destroy_ah(ah);
1319         return ret;
1320 }
1321
1322 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1323 {
1324         if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1325                 return slave;
1326         return mlx4_get_base_gid_ix(dev->dev, slave, port);
1327 }
1328
1329 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1330                                     struct ib_ah_attr *ah_attr)
1331 {
1332         if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1333                 ah_attr->grh.sgid_index = slave;
1334         else
1335                 ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1336 }
1337
1338 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1339 {
1340         struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1341         struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1342         int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1343         struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1344         struct mlx4_ib_ah ah;
1345         struct ib_ah_attr ah_attr;
1346         u8 *slave_id;
1347         int slave;
1348         int port;
1349         u16 vlan_id;
1350
1351         /* Get slave that sent this packet */
1352         if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1353             wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1354             (wc->src_qp & 0x1) != ctx->port - 1 ||
1355             wc->src_qp & 0x4) {
1356                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1357                 return;
1358         }
1359         slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1360         if (slave != ctx->slave) {
1361                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1362                              "belongs to another slave\n", wc->src_qp);
1363                 return;
1364         }
1365
1366         /* Map transaction ID */
1367         ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1368                                    sizeof (struct mlx4_tunnel_mad),
1369                                    DMA_FROM_DEVICE);
1370         switch (tunnel->mad.mad_hdr.method) {
1371         case IB_MGMT_METHOD_SET:
1372         case IB_MGMT_METHOD_GET:
1373         case IB_MGMT_METHOD_REPORT:
1374         case IB_SA_METHOD_GET_TABLE:
1375         case IB_SA_METHOD_DELETE:
1376         case IB_SA_METHOD_GET_MULTI:
1377         case IB_SA_METHOD_GET_TRACE_TBL:
1378                 slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1379                 if (*slave_id) {
1380                         mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1381                                      "class:%d slave:%d\n", *slave_id,
1382                                      tunnel->mad.mad_hdr.mgmt_class, slave);
1383                         return;
1384                 } else
1385                         *slave_id = slave;
1386         default:
1387                 /* nothing */;
1388         }
1389
1390         /* Class-specific handling */
1391         switch (tunnel->mad.mad_hdr.mgmt_class) {
1392         case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1393         case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1394                 if (slave != mlx4_master_func_num(dev->dev) &&
1395                     !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1396                         return;
1397                 break;
1398         case IB_MGMT_CLASS_SUBN_ADM:
1399                 if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1400                               (struct ib_sa_mad *) &tunnel->mad))
1401                         return;
1402                 break;
1403         case IB_MGMT_CLASS_CM:
1404                 if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1405                               (struct ib_mad *) &tunnel->mad))
1406                         return;
1407                 break;
1408         case IB_MGMT_CLASS_DEVICE_MGMT:
1409                 if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1410                     tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1411                         return;
1412                 break;
1413         default:
1414                 /* Drop unsupported classes for slaves in tunnel mode */
1415                 if (slave != mlx4_master_func_num(dev->dev)) {
1416                         mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1417                                      "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1418                         return;
1419                 }
1420         }
1421
1422         /* We are using standard ib_core services to send the mad, so generate a
1423          * stadard address handle by decoding the tunnelled mlx4_ah fields */
1424         memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1425         ah.ibah.device = ctx->ib_dev;
1426
1427         port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1428         port = mlx4_slave_convert_port(dev->dev, slave, port);
1429         if (port < 0)
1430                 return;
1431         ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1432
1433         mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1434         if (ah_attr.ah_flags & IB_AH_GRH)
1435                 fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1436
1437         memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1438         vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1439         /* if slave have default vlan use it */
1440         mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1441                                     &vlan_id, &ah_attr.sl);
1442
1443         mlx4_ib_send_to_wire(dev, slave, ctx->port,
1444                              is_proxy_qp0(dev, wc->src_qp, slave) ?
1445                              IB_QPT_SMI : IB_QPT_GSI,
1446                              be16_to_cpu(tunnel->hdr.pkey_index),
1447                              be32_to_cpu(tunnel->hdr.remote_qpn),
1448                              be32_to_cpu(tunnel->hdr.qkey),
1449                              &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1450 }
1451
1452 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1453                                  enum ib_qp_type qp_type, int is_tun)
1454 {
1455         int i;
1456         struct mlx4_ib_demux_pv_qp *tun_qp;
1457         int rx_buf_size, tx_buf_size;
1458
1459         if (qp_type > IB_QPT_GSI)
1460                 return -EINVAL;
1461
1462         tun_qp = &ctx->qp[qp_type];
1463
1464         tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1465                                GFP_KERNEL);
1466         if (!tun_qp->ring)
1467                 return -ENOMEM;
1468
1469         tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1470                                   sizeof (struct mlx4_ib_tun_tx_buf),
1471                                   GFP_KERNEL);
1472         if (!tun_qp->tx_ring) {
1473                 kfree(tun_qp->ring);
1474                 tun_qp->ring = NULL;
1475                 return -ENOMEM;
1476         }
1477
1478         if (is_tun) {
1479                 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1480                 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1481         } else {
1482                 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1483                 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1484         }
1485
1486         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1487                 tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1488                 if (!tun_qp->ring[i].addr)
1489                         goto err;
1490                 tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1491                                                         tun_qp->ring[i].addr,
1492                                                         rx_buf_size,
1493                                                         DMA_FROM_DEVICE);
1494                 if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1495                         kfree(tun_qp->ring[i].addr);
1496                         goto err;
1497                 }
1498         }
1499
1500         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1501                 tun_qp->tx_ring[i].buf.addr =
1502                         kmalloc(tx_buf_size, GFP_KERNEL);
1503                 if (!tun_qp->tx_ring[i].buf.addr)
1504                         goto tx_err;
1505                 tun_qp->tx_ring[i].buf.map =
1506                         ib_dma_map_single(ctx->ib_dev,
1507                                           tun_qp->tx_ring[i].buf.addr,
1508                                           tx_buf_size,
1509                                           DMA_TO_DEVICE);
1510                 if (ib_dma_mapping_error(ctx->ib_dev,
1511                                          tun_qp->tx_ring[i].buf.map)) {
1512                         kfree(tun_qp->tx_ring[i].buf.addr);
1513                         goto tx_err;
1514                 }
1515                 tun_qp->tx_ring[i].ah = NULL;
1516         }
1517         spin_lock_init(&tun_qp->tx_lock);
1518         tun_qp->tx_ix_head = 0;
1519         tun_qp->tx_ix_tail = 0;
1520         tun_qp->proxy_qpt = qp_type;
1521
1522         return 0;
1523
1524 tx_err:
1525         while (i > 0) {
1526                 --i;
1527                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1528                                     tx_buf_size, DMA_TO_DEVICE);
1529                 kfree(tun_qp->tx_ring[i].buf.addr);
1530         }
1531         kfree(tun_qp->tx_ring);
1532         tun_qp->tx_ring = NULL;
1533         i = MLX4_NUM_TUNNEL_BUFS;
1534 err:
1535         while (i > 0) {
1536                 --i;
1537                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1538                                     rx_buf_size, DMA_FROM_DEVICE);
1539                 kfree(tun_qp->ring[i].addr);
1540         }
1541         kfree(tun_qp->ring);
1542         tun_qp->ring = NULL;
1543         return -ENOMEM;
1544 }
1545
1546 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1547                                      enum ib_qp_type qp_type, int is_tun)
1548 {
1549         int i;
1550         struct mlx4_ib_demux_pv_qp *tun_qp;
1551         int rx_buf_size, tx_buf_size;
1552
1553         if (qp_type > IB_QPT_GSI)
1554                 return;
1555
1556         tun_qp = &ctx->qp[qp_type];
1557         if (is_tun) {
1558                 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1559                 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1560         } else {
1561                 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1562                 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1563         }
1564
1565
1566         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1567                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1568                                     rx_buf_size, DMA_FROM_DEVICE);
1569                 kfree(tun_qp->ring[i].addr);
1570         }
1571
1572         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1573                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1574                                     tx_buf_size, DMA_TO_DEVICE);
1575                 kfree(tun_qp->tx_ring[i].buf.addr);
1576                 if (tun_qp->tx_ring[i].ah)
1577                         ib_destroy_ah(tun_qp->tx_ring[i].ah);
1578         }
1579         kfree(tun_qp->tx_ring);
1580         kfree(tun_qp->ring);
1581 }
1582
1583 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1584 {
1585         struct mlx4_ib_demux_pv_ctx *ctx;
1586         struct mlx4_ib_demux_pv_qp *tun_qp;
1587         struct ib_wc wc;
1588         int ret;
1589         ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1590         ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1591
1592         while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1593                 tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1594                 if (wc.status == IB_WC_SUCCESS) {
1595                         switch (wc.opcode) {
1596                         case IB_WC_RECV:
1597                                 mlx4_ib_multiplex_mad(ctx, &wc);
1598                                 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1599                                                              wc.wr_id &
1600                                                              (MLX4_NUM_TUNNEL_BUFS - 1));
1601                                 if (ret)
1602                                         pr_err("Failed reposting tunnel "
1603                                                "buf:%lld\n", wc.wr_id);
1604                                 break;
1605                         case IB_WC_SEND:
1606                                 pr_debug("received tunnel send completion:"
1607                                          "wrid=0x%llx, status=0x%x\n",
1608                                          wc.wr_id, wc.status);
1609                                 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1610                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1611                                 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1612                                         = NULL;
1613                                 spin_lock(&tun_qp->tx_lock);
1614                                 tun_qp->tx_ix_tail++;
1615                                 spin_unlock(&tun_qp->tx_lock);
1616
1617                                 break;
1618                         default:
1619                                 break;
1620                         }
1621                 } else  {
1622                         pr_debug("mlx4_ib: completion error in tunnel: %d."
1623                                  " status = %d, wrid = 0x%llx\n",
1624                                  ctx->slave, wc.status, wc.wr_id);
1625                         if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1626                                 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1627                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1628                                 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1629                                         = NULL;
1630                                 spin_lock(&tun_qp->tx_lock);
1631                                 tun_qp->tx_ix_tail++;
1632                                 spin_unlock(&tun_qp->tx_lock);
1633                         }
1634                 }
1635         }
1636 }
1637
1638 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1639 {
1640         struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1641
1642         /* It's worse than that! He's dead, Jim! */
1643         pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1644                event->event, sqp->port);
1645 }
1646
1647 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1648                             enum ib_qp_type qp_type, int create_tun)
1649 {
1650         int i, ret;
1651         struct mlx4_ib_demux_pv_qp *tun_qp;
1652         struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1653         struct ib_qp_attr attr;
1654         int qp_attr_mask_INIT;
1655
1656         if (qp_type > IB_QPT_GSI)
1657                 return -EINVAL;
1658
1659         tun_qp = &ctx->qp[qp_type];
1660
1661         memset(&qp_init_attr, 0, sizeof qp_init_attr);
1662         qp_init_attr.init_attr.send_cq = ctx->cq;
1663         qp_init_attr.init_attr.recv_cq = ctx->cq;
1664         qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1665         qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1666         qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1667         qp_init_attr.init_attr.cap.max_send_sge = 1;
1668         qp_init_attr.init_attr.cap.max_recv_sge = 1;
1669         if (create_tun) {
1670                 qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1671                 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1672                 qp_init_attr.port = ctx->port;
1673                 qp_init_attr.slave = ctx->slave;
1674                 qp_init_attr.proxy_qp_type = qp_type;
1675                 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1676                            IB_QP_QKEY | IB_QP_PORT;
1677         } else {
1678                 qp_init_attr.init_attr.qp_type = qp_type;
1679                 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1680                 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1681         }
1682         qp_init_attr.init_attr.port_num = ctx->port;
1683         qp_init_attr.init_attr.qp_context = ctx;
1684         qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1685         tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1686         if (IS_ERR(tun_qp->qp)) {
1687                 ret = PTR_ERR(tun_qp->qp);
1688                 tun_qp->qp = NULL;
1689                 pr_err("Couldn't create %s QP (%d)\n",
1690                        create_tun ? "tunnel" : "special", ret);
1691                 return ret;
1692         }
1693
1694         memset(&attr, 0, sizeof attr);
1695         attr.qp_state = IB_QPS_INIT;
1696         ret = 0;
1697         if (create_tun)
1698                 ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1699                                               ctx->port, IB_DEFAULT_PKEY_FULL,
1700                                               &attr.pkey_index);
1701         if (ret || !create_tun)
1702                 attr.pkey_index =
1703                         to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1704         attr.qkey = IB_QP1_QKEY;
1705         attr.port_num = ctx->port;
1706         ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1707         if (ret) {
1708                 pr_err("Couldn't change %s qp state to INIT (%d)\n",
1709                        create_tun ? "tunnel" : "special", ret);
1710                 goto err_qp;
1711         }
1712         attr.qp_state = IB_QPS_RTR;
1713         ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1714         if (ret) {
1715                 pr_err("Couldn't change %s qp state to RTR (%d)\n",
1716                        create_tun ? "tunnel" : "special", ret);
1717                 goto err_qp;
1718         }
1719         attr.qp_state = IB_QPS_RTS;
1720         attr.sq_psn = 0;
1721         ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1722         if (ret) {
1723                 pr_err("Couldn't change %s qp state to RTS (%d)\n",
1724                        create_tun ? "tunnel" : "special", ret);
1725                 goto err_qp;
1726         }
1727
1728         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1729                 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1730                 if (ret) {
1731                         pr_err(" mlx4_ib_post_pv_buf error"
1732                                " (err = %d, i = %d)\n", ret, i);
1733                         goto err_qp;
1734                 }
1735         }
1736         return 0;
1737
1738 err_qp:
1739         ib_destroy_qp(tun_qp->qp);
1740         tun_qp->qp = NULL;
1741         return ret;
1742 }
1743
1744 /*
1745  * IB MAD completion callback for real SQPs
1746  */
1747 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1748 {
1749         struct mlx4_ib_demux_pv_ctx *ctx;
1750         struct mlx4_ib_demux_pv_qp *sqp;
1751         struct ib_wc wc;
1752         struct ib_grh *grh;
1753         struct ib_mad *mad;
1754
1755         ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1756         ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1757
1758         while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1759                 sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1760                 if (wc.status == IB_WC_SUCCESS) {
1761                         switch (wc.opcode) {
1762                         case IB_WC_SEND:
1763                                 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1764                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1765                                 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1766                                         = NULL;
1767                                 spin_lock(&sqp->tx_lock);
1768                                 sqp->tx_ix_tail++;
1769                                 spin_unlock(&sqp->tx_lock);
1770                                 break;
1771                         case IB_WC_RECV:
1772                                 mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1773                                                 (sqp->ring[wc.wr_id &
1774                                                 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1775                                 grh = &(((struct mlx4_mad_rcv_buf *)
1776                                                 (sqp->ring[wc.wr_id &
1777                                                 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1778                                 mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1779                                 if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1780                                                            (MLX4_NUM_TUNNEL_BUFS - 1)))
1781                                         pr_err("Failed reposting SQP "
1782                                                "buf:%lld\n", wc.wr_id);
1783                                 break;
1784                         default:
1785                                 BUG_ON(1);
1786                                 break;
1787                         }
1788                 } else  {
1789                         pr_debug("mlx4_ib: completion error in tunnel: %d."
1790                                  " status = %d, wrid = 0x%llx\n",
1791                                  ctx->slave, wc.status, wc.wr_id);
1792                         if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1793                                 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1794                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1795                                 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1796                                         = NULL;
1797                                 spin_lock(&sqp->tx_lock);
1798                                 sqp->tx_ix_tail++;
1799                                 spin_unlock(&sqp->tx_lock);
1800                         }
1801                 }
1802         }
1803 }
1804
1805 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1806                                struct mlx4_ib_demux_pv_ctx **ret_ctx)
1807 {
1808         struct mlx4_ib_demux_pv_ctx *ctx;
1809
1810         *ret_ctx = NULL;
1811         ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1812         if (!ctx) {
1813                 pr_err("failed allocating pv resource context "
1814                        "for port %d, slave %d\n", port, slave);
1815                 return -ENOMEM;
1816         }
1817
1818         ctx->ib_dev = &dev->ib_dev;
1819         ctx->port = port;
1820         ctx->slave = slave;
1821         *ret_ctx = ctx;
1822         return 0;
1823 }
1824
1825 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1826 {
1827         if (dev->sriov.demux[port - 1].tun[slave]) {
1828                 kfree(dev->sriov.demux[port - 1].tun[slave]);
1829                 dev->sriov.demux[port - 1].tun[slave] = NULL;
1830         }
1831 }
1832
1833 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1834                                int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1835 {
1836         int ret, cq_size;
1837         struct ib_cq_init_attr cq_attr = {};
1838
1839         if (ctx->state != DEMUX_PV_STATE_DOWN)
1840                 return -EEXIST;
1841
1842         ctx->state = DEMUX_PV_STATE_STARTING;
1843         /* have QP0 only if link layer is IB */
1844         if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1845             IB_LINK_LAYER_INFINIBAND)
1846                 ctx->has_smi = 1;
1847
1848         if (ctx->has_smi) {
1849                 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1850                 if (ret) {
1851                         pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1852                         goto err_out;
1853                 }
1854         }
1855
1856         ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1857         if (ret) {
1858                 pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1859                 goto err_out_qp0;
1860         }
1861
1862         cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1863         if (ctx->has_smi)
1864                 cq_size *= 2;
1865
1866         cq_attr.cqe = cq_size;
1867         ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1868                                NULL, ctx, &cq_attr);
1869         if (IS_ERR(ctx->cq)) {
1870                 ret = PTR_ERR(ctx->cq);
1871                 pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1872                 goto err_buf;
1873         }
1874
1875         ctx->pd = ib_alloc_pd(ctx->ib_dev);
1876         if (IS_ERR(ctx->pd)) {
1877                 ret = PTR_ERR(ctx->pd);
1878                 pr_err("Couldn't create tunnel PD (%d)\n", ret);
1879                 goto err_cq;
1880         }
1881
1882         if (ctx->has_smi) {
1883                 ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1884                 if (ret) {
1885                         pr_err("Couldn't create %s QP0 (%d)\n",
1886                                create_tun ? "tunnel for" : "",  ret);
1887                         goto err_pd;
1888                 }
1889         }
1890
1891         ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1892         if (ret) {
1893                 pr_err("Couldn't create %s QP1 (%d)\n",
1894                        create_tun ? "tunnel for" : "",  ret);
1895                 goto err_qp0;
1896         }
1897
1898         if (create_tun)
1899                 INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1900         else
1901                 INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1902
1903         ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1904
1905         ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1906         if (ret) {
1907                 pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1908                 goto err_wq;
1909         }
1910         ctx->state = DEMUX_PV_STATE_ACTIVE;
1911         return 0;
1912
1913 err_wq:
1914         ctx->wq = NULL;
1915         ib_destroy_qp(ctx->qp[1].qp);
1916         ctx->qp[1].qp = NULL;
1917
1918
1919 err_qp0:
1920         if (ctx->has_smi)
1921                 ib_destroy_qp(ctx->qp[0].qp);
1922         ctx->qp[0].qp = NULL;
1923
1924 err_pd:
1925         ib_dealloc_pd(ctx->pd);
1926         ctx->pd = NULL;
1927
1928 err_cq:
1929         ib_destroy_cq(ctx->cq);
1930         ctx->cq = NULL;
1931
1932 err_buf:
1933         mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1934
1935 err_out_qp0:
1936         if (ctx->has_smi)
1937                 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1938 err_out:
1939         ctx->state = DEMUX_PV_STATE_DOWN;
1940         return ret;
1941 }
1942
1943 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1944                                  struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1945 {
1946         if (!ctx)
1947                 return;
1948         if (ctx->state > DEMUX_PV_STATE_DOWN) {
1949                 ctx->state = DEMUX_PV_STATE_DOWNING;
1950                 if (flush)
1951                         flush_workqueue(ctx->wq);
1952                 if (ctx->has_smi) {
1953                         ib_destroy_qp(ctx->qp[0].qp);
1954                         ctx->qp[0].qp = NULL;
1955                         mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1956                 }
1957                 ib_destroy_qp(ctx->qp[1].qp);
1958                 ctx->qp[1].qp = NULL;
1959                 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1960                 ib_dealloc_pd(ctx->pd);
1961                 ctx->pd = NULL;
1962                 ib_destroy_cq(ctx->cq);
1963                 ctx->cq = NULL;
1964                 ctx->state = DEMUX_PV_STATE_DOWN;
1965         }
1966 }
1967
1968 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1969                                   int port, int do_init)
1970 {
1971         int ret = 0;
1972
1973         if (!do_init) {
1974                 clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1975                 /* for master, destroy real sqp resources */
1976                 if (slave == mlx4_master_func_num(dev->dev))
1977                         destroy_pv_resources(dev, slave, port,
1978                                              dev->sriov.sqps[port - 1], 1);
1979                 /* destroy the tunnel qp resources */
1980                 destroy_pv_resources(dev, slave, port,
1981                                      dev->sriov.demux[port - 1].tun[slave], 1);
1982                 return 0;
1983         }
1984
1985         /* create the tunnel qp resources */
1986         ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1987                                   dev->sriov.demux[port - 1].tun[slave]);
1988
1989         /* for master, create the real sqp resources */
1990         if (!ret && slave == mlx4_master_func_num(dev->dev))
1991                 ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1992                                           dev->sriov.sqps[port - 1]);
1993         return ret;
1994 }
1995
1996 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1997 {
1998         struct mlx4_ib_demux_work *dmxw;
1999
2000         dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2001         mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2002                                dmxw->do_init);
2003         kfree(dmxw);
2004         return;
2005 }
2006
2007 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2008                                        struct mlx4_ib_demux_ctx *ctx,
2009                                        int port)
2010 {
2011         char name[12];
2012         int ret = 0;
2013         int i;
2014
2015         ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2016                            sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2017         if (!ctx->tun)
2018                 return -ENOMEM;
2019
2020         ctx->dev = dev;
2021         ctx->port = port;
2022         ctx->ib_dev = &dev->ib_dev;
2023
2024         for (i = 0;
2025              i < min(dev->dev->caps.sqp_demux,
2026              (u16)(dev->dev->persist->num_vfs + 1));
2027              i++) {
2028                 struct mlx4_active_ports actv_ports =
2029                         mlx4_get_active_ports(dev->dev, i);
2030
2031                 if (!test_bit(port - 1, actv_ports.ports))
2032                         continue;
2033
2034                 ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2035                 if (ret) {
2036                         ret = -ENOMEM;
2037                         goto err_mcg;
2038                 }
2039         }
2040
2041         ret = mlx4_ib_mcg_port_init(ctx);
2042         if (ret) {
2043                 pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2044                 goto err_mcg;
2045         }
2046
2047         snprintf(name, sizeof name, "mlx4_ibt%d", port);
2048         ctx->wq = create_singlethread_workqueue(name);
2049         if (!ctx->wq) {
2050                 pr_err("Failed to create tunnelling WQ for port %d\n", port);
2051                 ret = -ENOMEM;
2052                 goto err_wq;
2053         }
2054
2055         snprintf(name, sizeof name, "mlx4_ibud%d", port);
2056         ctx->ud_wq = create_singlethread_workqueue(name);
2057         if (!ctx->ud_wq) {
2058                 pr_err("Failed to create up/down WQ for port %d\n", port);
2059                 ret = -ENOMEM;
2060                 goto err_udwq;
2061         }
2062
2063         return 0;
2064
2065 err_udwq:
2066         destroy_workqueue(ctx->wq);
2067         ctx->wq = NULL;
2068
2069 err_wq:
2070         mlx4_ib_mcg_port_cleanup(ctx, 1);
2071 err_mcg:
2072         for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2073                 free_pv_object(dev, i, port);
2074         kfree(ctx->tun);
2075         ctx->tun = NULL;
2076         return ret;
2077 }
2078
2079 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2080 {
2081         if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2082                 sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2083                 flush_workqueue(sqp_ctx->wq);
2084                 if (sqp_ctx->has_smi) {
2085                         ib_destroy_qp(sqp_ctx->qp[0].qp);
2086                         sqp_ctx->qp[0].qp = NULL;
2087                         mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2088                 }
2089                 ib_destroy_qp(sqp_ctx->qp[1].qp);
2090                 sqp_ctx->qp[1].qp = NULL;
2091                 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2092                 ib_dealloc_pd(sqp_ctx->pd);
2093                 sqp_ctx->pd = NULL;
2094                 ib_destroy_cq(sqp_ctx->cq);
2095                 sqp_ctx->cq = NULL;
2096                 sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2097         }
2098 }
2099
2100 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2101 {
2102         int i;
2103         if (ctx) {
2104                 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2105                 mlx4_ib_mcg_port_cleanup(ctx, 1);
2106                 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2107                         if (!ctx->tun[i])
2108                                 continue;
2109                         if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2110                                 ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2111                 }
2112                 flush_workqueue(ctx->wq);
2113                 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2114                         destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2115                         free_pv_object(dev, i, ctx->port);
2116                 }
2117                 kfree(ctx->tun);
2118                 destroy_workqueue(ctx->ud_wq);
2119                 destroy_workqueue(ctx->wq);
2120         }
2121 }
2122
2123 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2124 {
2125         int i;
2126
2127         if (!mlx4_is_master(dev->dev))
2128                 return;
2129         /* initialize or tear down tunnel QPs for the master */
2130         for (i = 0; i < dev->dev->caps.num_ports; i++)
2131                 mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2132         return;
2133 }
2134
2135 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2136 {
2137         int i = 0;
2138         int err;
2139
2140         if (!mlx4_is_mfunc(dev->dev))
2141                 return 0;
2142
2143         dev->sriov.is_going_down = 0;
2144         spin_lock_init(&dev->sriov.going_down_lock);
2145         mlx4_ib_cm_paravirt_init(dev);
2146
2147         mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2148
2149         if (mlx4_is_slave(dev->dev)) {
2150                 mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2151                 return 0;
2152         }
2153
2154         for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2155                 if (i == mlx4_master_func_num(dev->dev))
2156                         mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2157                 else
2158                         mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2159         }
2160
2161         err = mlx4_ib_init_alias_guid_service(dev);
2162         if (err) {
2163                 mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2164                 goto paravirt_err;
2165         }
2166         err = mlx4_ib_device_register_sysfs(dev);
2167         if (err) {
2168                 mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2169                 goto sysfs_err;
2170         }
2171
2172         mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2173                      dev->dev->caps.sqp_demux);
2174         for (i = 0; i < dev->num_ports; i++) {
2175                 union ib_gid gid;
2176                 err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2177                 if (err)
2178                         goto demux_err;
2179                 dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2180                 err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2181                                       &dev->sriov.sqps[i]);
2182                 if (err)
2183                         goto demux_err;
2184                 err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2185                 if (err)
2186                         goto free_pv;
2187         }
2188         mlx4_ib_master_tunnels(dev, 1);
2189         return 0;
2190
2191 free_pv:
2192         free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2193 demux_err:
2194         while (--i >= 0) {
2195                 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2196                 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2197         }
2198         mlx4_ib_device_unregister_sysfs(dev);
2199
2200 sysfs_err:
2201         mlx4_ib_destroy_alias_guid_service(dev);
2202
2203 paravirt_err:
2204         mlx4_ib_cm_paravirt_clean(dev, -1);
2205
2206         return err;
2207 }
2208
2209 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2210 {
2211         int i;
2212         unsigned long flags;
2213
2214         if (!mlx4_is_mfunc(dev->dev))
2215                 return;
2216
2217         spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2218         dev->sriov.is_going_down = 1;
2219         spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2220         if (mlx4_is_master(dev->dev)) {
2221                 for (i = 0; i < dev->num_ports; i++) {
2222                         flush_workqueue(dev->sriov.demux[i].ud_wq);
2223                         mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2224                         kfree(dev->sriov.sqps[i]);
2225                         dev->sriov.sqps[i] = NULL;
2226                         mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2227                 }
2228
2229                 mlx4_ib_cm_paravirt_clean(dev, -1);
2230                 mlx4_ib_destroy_alias_guid_service(dev);
2231                 mlx4_ib_device_unregister_sysfs(dev);
2232         }
2233 }