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/mlx4/cmd.h>
  39 #include <linux/gfp.h>
  40 #include <rdma/ib_pma.h>
  41
  42 #include "mlx4_ib.h"
  43
  44 enum {
  45         MLX4_IB_VENDOR_CLASS1 = 0x9,
  46         MLX4_IB_VENDOR_CLASS2 = 0xa
  47 };
  48
  49 #define MLX4_TUN_SEND_WRID_SHIFT 34
  50 #define MLX4_TUN_QPN_SHIFT 32
  51 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
  52 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
  53
  54 #define MLX4_TUN_IS_RECV(a)  (((a) >>  MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
  55 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
  56
  57 struct mlx4_mad_rcv_buf {
  58         struct ib_grh grh;
  59         u8 payload[256];
  60 } __packed;
  61
  62 struct mlx4_mad_snd_buf {
  63         u8 payload[256];
  64 } __packed;
  65
  66 struct mlx4_tunnel_mad {
  67         struct ib_grh grh;
  68         struct mlx4_ib_tunnel_header hdr;
  69         struct ib_mad mad;
  70 } __packed;
  71
  72 struct mlx4_rcv_tunnel_mad {
  73         struct mlx4_rcv_tunnel_hdr hdr;
  74         struct ib_grh grh;
  75         struct ib_mad mad;
  76 } __packed;
  77
  78 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
  79
  80 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
  81 {
  82         return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
  83                 cpu_to_be64(0xff00000000000000LL);
  84 }
  85
  86 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
  87                  int port, struct ib_wc *in_wc, struct ib_grh *in_grh,
  88                  void *in_mad, void *response_mad)
  89 {
  90         struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
  91         void *inbox;
  92         int err;
  93         u32 in_modifier = port;
  94         u8 op_modifier = 0;
  95
  96         inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
  97         if (IS_ERR(inmailbox))
  98                 return PTR_ERR(inmailbox);
  99         inbox = inmailbox->buf;
 100
 101         outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
 102         if (IS_ERR(outmailbox)) {
 103                 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
 104                 return PTR_ERR(outmailbox);
 105         }
 106
 107         memcpy(inbox, in_mad, 256);
 108
 109         /*
 110          * Key check traps can't be generated unless we have in_wc to
 111          * tell us where to send the trap.
 112          */
 113         if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
 114                 op_modifier |= 0x1;
 115         if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
 116                 op_modifier |= 0x2;
 117         if (mlx4_is_mfunc(dev->dev) &&
 118             (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
 119                 op_modifier |= 0x8;
 120
 121         if (in_wc) {
 122                 struct {
 123                         __be32          my_qpn;
 124                         u32             reserved1;
 125                         __be32          rqpn;
 126                         u8              sl;
 127                         u8              g_path;
 128                         u16             reserved2[2];
 129                         __be16          pkey;
 130                         u32             reserved3[11];
 131                         u8              grh[40];
 132                 } *ext_info;
 133
 134                 memset(inbox + 256, 0, 256);
 135                 ext_info = inbox + 256;
 136
 137                 ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
 138                 ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
 139                 ext_info->sl     = in_wc->sl << 4;
 140                 ext_info->g_path = in_wc->dlid_path_bits |
 141                         (in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
 142                 ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
 143
 144                 if (in_grh)
 145                         memcpy(ext_info->grh, in_grh, 40);
 146
 147                 op_modifier |= 0x4;
 148
 149                 in_modifier |= in_wc->slid << 16;
 150         }
 151
 152         err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
 153                            mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
 154                            MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
 155                            (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
 156
 157         if (!err)
 158                 memcpy(response_mad, outmailbox->buf, 256);
 159
 160         mlx4_free_cmd_mailbox(dev->dev, inmailbox);
 161         mlx4_free_cmd_mailbox(dev->dev, outmailbox);
 162
 163         return err;
 164 }
 165
 166 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
 167 {
 168         struct ib_ah *new_ah;
 169         struct ib_ah_attr ah_attr;
 170         unsigned long flags;
 171
 172         if (!dev->send_agent[port_num - 1][0])
 173                 return;
 174
 175         memset(&ah_attr, 0, sizeof ah_attr);
 176         ah_attr.dlid     = lid;
 177         ah_attr.sl       = sl;
 178         ah_attr.port_num = port_num;
 179
 180         new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
 181                               &ah_attr);
 182         if (IS_ERR(new_ah))
 183                 return;
 184
 185         spin_lock_irqsave(&dev->sm_lock, flags);
 186         if (dev->sm_ah[port_num - 1])
 187                 ib_destroy_ah(dev->sm_ah[port_num - 1]);
 188         dev->sm_ah[port_num - 1] = new_ah;
 189         spin_unlock_irqrestore(&dev->sm_lock, flags);
 190 }
 191
 192 /*
 193  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
 194  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
 195  */
 196 static void smp_snoop(struct ib_device *ibdev, u8 port_num, struct ib_mad *mad,
 197                       u16 prev_lid)
 198 {
 199         struct ib_port_info *pinfo;
 200         u16 lid;
 201         __be16 *base;
 202         u32 bn, pkey_change_bitmap;
 203         int i;
 204
 205
 206         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 207         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 208              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 209             mad->mad_hdr.method == IB_MGMT_METHOD_SET)
 210                 switch (mad->mad_hdr.attr_id) {
 211                 case IB_SMP_ATTR_PORT_INFO:
 212                         pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
 213                         lid = be16_to_cpu(pinfo->lid);
 214
 215                         update_sm_ah(dev, port_num,
 216                                      be16_to_cpu(pinfo->sm_lid),
 217                                      pinfo->neighbormtu_mastersmsl & 0xf);
 218
 219                         if (pinfo->clientrereg_resv_subnetto & 0x80)
 220                                 handle_client_rereg_event(dev, port_num);
 221
 222                         if (prev_lid != lid)
 223                                 mlx4_ib_dispatch_event(dev, port_num,
 224                                                        IB_EVENT_LID_CHANGE);
 225                         break;
 226
 227                 case IB_SMP_ATTR_PKEY_TABLE:
 228                         if (!mlx4_is_mfunc(dev->dev)) {
 229                                 mlx4_ib_dispatch_event(dev, port_num,
 230                                                        IB_EVENT_PKEY_CHANGE);
 231                                 break;
 232                         }
 233
 234                         bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
 235                         base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
 236                         pkey_change_bitmap = 0;
 237                         for (i = 0; i < 32; i++) {
 238                                 pr_debug("PKEY[%d] = x%x\n",
 239                                          i + bn*32, be16_to_cpu(base[i]));
 240                                 if (be16_to_cpu(base[i]) !=
 241                                     dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
 242                                         pkey_change_bitmap |= (1 << i);
 243                                         dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
 244                                                 be16_to_cpu(base[i]);
 245                                 }
 246                         }
 247                         pr_debug("PKEY Change event: port=%d, "
 248                                  "block=0x%x, change_bitmap=0x%x\n",
 249                                  port_num, bn, pkey_change_bitmap);
 250
 251                         if (pkey_change_bitmap)
 252                                 mlx4_ib_dispatch_event(dev, port_num,
 253                                                        IB_EVENT_PKEY_CHANGE);
 254
 255                         break;
 256
 257                 case IB_SMP_ATTR_GUID_INFO:
 258                         /* paravirtualized master's guid is guid 0 -- does not change */
 259                         if (!mlx4_is_master(dev->dev))
 260                                 mlx4_ib_dispatch_event(dev, port_num,
 261                                                        IB_EVENT_GID_CHANGE);
 262                         break;
 263                 default:
 264                         break;
 265                 }
 266 }
 267
 268 static void node_desc_override(struct ib_device *dev,
 269                                struct ib_mad *mad)
 270 {
 271         unsigned long flags;
 272
 273         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 274              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 275             mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
 276             mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
 277                 spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
 278                 memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
 279                 spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
 280         }
 281 }
 282
 283 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, struct ib_mad *mad)
 284 {
 285         int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
 286         struct ib_mad_send_buf *send_buf;
 287         struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
 288         int ret;
 289         unsigned long flags;
 290
 291         if (agent) {
 292                 send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
 293                                               IB_MGMT_MAD_DATA, GFP_ATOMIC);
 294                 if (IS_ERR(send_buf))
 295                         return;
 296                 /*
 297                  * We rely here on the fact that MLX QPs don't use the
 298                  * address handle after the send is posted (this is
 299                  * wrong following the IB spec strictly, but we know
 300                  * it's OK for our devices).
 301                  */
 302                 spin_lock_irqsave(&dev->sm_lock, flags);
 303                 memcpy(send_buf->mad, mad, sizeof *mad);
 304                 if ((send_buf->ah = dev->sm_ah[port_num - 1]))
 305                         ret = ib_post_send_mad(send_buf, NULL);
 306                 else
 307                         ret = -EINVAL;
 308                 spin_unlock_irqrestore(&dev->sm_lock, flags);
 309
 310                 if (ret)
 311                         ib_free_send_mad(send_buf);
 312         }
 313 }
 314
 315 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
 316                                                              struct ib_sa_mad *sa_mad)
 317 {
 318         int ret = 0;
 319
 320         /* dispatch to different sa handlers */
 321         switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
 322         case IB_SA_ATTR_MC_MEMBER_REC:
 323                 ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
 324                 break;
 325         default:
 326                 break;
 327         }
 328         return ret;
 329 }
 330
 331 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
 332 {
 333         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 334         int i;
 335
 336         for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
 337                 if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
 338                         return i;
 339         }
 340         return -1;
 341 }
 342
 343
 344 static int get_pkey_phys_indices(struct mlx4_ib_dev *ibdev, u8 port, u8 ph_pkey_ix,
 345                                  u8 *full_pk_ix, u8 *partial_pk_ix,
 346                                  int *is_full_member)
 347 {
 348         u16 search_pkey;
 349         int fm;
 350         int err = 0;
 351         u16 pk;
 352
 353         err = ib_get_cached_pkey(&ibdev->ib_dev, port, ph_pkey_ix, &search_pkey);
 354         if (err)
 355                 return err;
 356
 357         fm = (search_pkey & 0x8000) ? 1 : 0;
 358         if (fm) {
 359                 *full_pk_ix = ph_pkey_ix;
 360                 search_pkey &= 0x7FFF;
 361         } else {
 362                 *partial_pk_ix = ph_pkey_ix;
 363                 search_pkey |= 0x8000;
 364         }
 365
 366         if (ib_find_exact_cached_pkey(&ibdev->ib_dev, port, search_pkey, &pk))
 367                 pk = 0xFFFF;
 368
 369         if (fm)
 370                 *partial_pk_ix = (pk & 0xFF);
 371         else
 372                 *full_pk_ix = (pk & 0xFF);
 373
 374         *is_full_member = fm;
 375         return err;
 376 }
 377
 378 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
 379                           enum ib_qp_type dest_qpt, struct ib_wc *wc,
 380                           struct ib_grh *grh, struct ib_mad *mad)
 381 {
 382         struct ib_sge list;
 383         struct ib_send_wr wr, *bad_wr;
 384         struct mlx4_ib_demux_pv_ctx *tun_ctx;
 385         struct mlx4_ib_demux_pv_qp *tun_qp;
 386         struct mlx4_rcv_tunnel_mad *tun_mad;
 387         struct ib_ah_attr attr;
 388         struct ib_ah *ah;
 389         struct ib_qp *src_qp = NULL;
 390         unsigned tun_tx_ix = 0;
 391         int dqpn;
 392         int ret = 0;
 393         int i;
 394         int is_full_member = 0;
 395         u16 tun_pkey_ix;
 396         u8 ph_pkey_ix, full_pk_ix = 0, partial_pk_ix = 0;
 397
 398         if (dest_qpt > IB_QPT_GSI)
 399                 return -EINVAL;
 400
 401         tun_ctx = dev->sriov.demux[port-1].tun[slave];
 402
 403         /* check if proxy qp created */
 404         if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
 405                 return -EAGAIN;
 406
 407         /* QP0 forwarding only for Dom0 */
 408         if (!dest_qpt && (mlx4_master_func_num(dev->dev) != slave))
 409                 return -EINVAL;
 410
 411         if (!dest_qpt)
 412                 tun_qp = &tun_ctx->qp[0];
 413         else
 414                 tun_qp = &tun_ctx->qp[1];
 415
 416         /* compute pkey index for slave */
 417         /* get physical pkey -- virtualized Dom0 pkey to phys*/
 418         if (dest_qpt) {
 419                 ph_pkey_ix =
 420                         dev->pkeys.virt2phys_pkey[mlx4_master_func_num(dev->dev)][port - 1][wc->pkey_index];
 421
 422                 /* now, translate this to the slave pkey index */
 423                 ret = get_pkey_phys_indices(dev, port, ph_pkey_ix, &full_pk_ix,
 424                                             &partial_pk_ix, &is_full_member);
 425                 if (ret)
 426                         return -EINVAL;
 427
 428                 for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
 429                         if ((dev->pkeys.virt2phys_pkey[slave][port - 1][i] == full_pk_ix) ||
 430                             (is_full_member &&
 431                              (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == partial_pk_ix)))
 432                                 break;
 433                 }
 434                 if (i == dev->dev->caps.pkey_table_len[port])
 435                         return -EINVAL;
 436                 tun_pkey_ix = i;
 437         } else
 438                 tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
 439
 440         dqpn = dev->dev->caps.sqp_start + 8 * slave + port + (dest_qpt * 2) - 1;
 441
 442         /* get tunnel tx data buf for slave */
 443         src_qp = tun_qp->qp;
 444
 445         /* create ah. Just need an empty one with the port num for the post send.
 446          * The driver will set the force loopback bit in post_send */
 447         memset(&attr, 0, sizeof attr);
 448         attr.port_num = port;
 449         ah = ib_create_ah(tun_ctx->pd, &attr);
 450         if (IS_ERR(ah))
 451                 return -ENOMEM;
 452
 453         /* allocate tunnel tx buf after pass failure returns */
 454         spin_lock(&tun_qp->tx_lock);
 455         if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
 456             (MLX4_NUM_TUNNEL_BUFS - 1))
 457                 ret = -EAGAIN;
 458         else
 459                 tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
 460         spin_unlock(&tun_qp->tx_lock);
 461         if (ret)
 462                 goto out;
 463
 464         tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
 465         if (tun_qp->tx_ring[tun_tx_ix].ah)
 466                 ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
 467         tun_qp->tx_ring[tun_tx_ix].ah = ah;
 468         ib_dma_sync_single_for_cpu(&dev->ib_dev,
 469                                    tun_qp->tx_ring[tun_tx_ix].buf.map,
 470                                    sizeof (struct mlx4_rcv_tunnel_mad),
 471                                    DMA_TO_DEVICE);
 472
 473         /* copy over to tunnel buffer */
 474         if (grh)
 475                 memcpy(&tun_mad->grh, grh, sizeof *grh);
 476         memcpy(&tun_mad->mad, mad, sizeof *mad);
 477
 478         /* adjust tunnel data */
 479         tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
 480         tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
 481         tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
 482         tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
 483         tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
 484
 485         ib_dma_sync_single_for_device(&dev->ib_dev,
 486                                       tun_qp->tx_ring[tun_tx_ix].buf.map,
 487                                       sizeof (struct mlx4_rcv_tunnel_mad),
 488                                       DMA_TO_DEVICE);
 489
 490         list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
 491         list.length = sizeof (struct mlx4_rcv_tunnel_mad);
 492         list.lkey = tun_ctx->mr->lkey;
 493
 494         wr.wr.ud.ah = ah;
 495         wr.wr.ud.port_num = port;
 496         wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
 497         wr.wr.ud.remote_qpn = dqpn;
 498         wr.next = NULL;
 499         wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
 500         wr.sg_list = &list;
 501         wr.num_sge = 1;
 502         wr.opcode = IB_WR_SEND;
 503         wr.send_flags = IB_SEND_SIGNALED;
 504
 505         ret = ib_post_send(src_qp, &wr, &bad_wr);
 506 out:
 507         if (ret)
 508                 ib_destroy_ah(ah);
 509         return ret;
 510 }
 511
 512 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
 513                         struct ib_wc *wc, struct ib_grh *grh,
 514                         struct ib_mad *mad)
 515 {
 516         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 517         int err;
 518         int slave;
 519         u8 *slave_id;
 520
 521         /* Initially assume that this mad is for us */
 522         slave = mlx4_master_func_num(dev->dev);
 523
 524         /* See if the slave id is encoded in a response mad */
 525         if (mad->mad_hdr.method & 0x80) {
 526                 slave_id = (u8 *) &mad->mad_hdr.tid;
 527                 slave = *slave_id;
 528                 if (slave != 255) /*255 indicates the dom0*/
 529                         *slave_id = 0; /* remap tid */
 530         }
 531
 532         /* If a grh is present, we demux according to it */
 533         if (wc->wc_flags & IB_WC_GRH) {
 534                 slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
 535                 if (slave < 0) {
 536                         mlx4_ib_warn(ibdev, "failed matching grh\n");
 537                         return -ENOENT;
 538                 }
 539         }
 540         /* Class-specific handling */
 541         switch (mad->mad_hdr.mgmt_class) {
 542         case IB_MGMT_CLASS_SUBN_ADM:
 543                 if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
 544                                              (struct ib_sa_mad *) mad))
 545                         return 0;
 546                 break;
 547         case IB_MGMT_CLASS_CM:
 548                 if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
 549                         return 0;
 550                 break;
 551         case IB_MGMT_CLASS_DEVICE_MGMT:
 552                 if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
 553                         return 0;
 554                 break;
 555         default:
 556                 /* Drop unsupported classes for slaves in tunnel mode */
 557                 if (slave != mlx4_master_func_num(dev->dev)) {
 558                         pr_debug("dropping unsupported ingress mad from class:%d "
 559                                  "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
 560                         return 0;
 561                 }
 562         }
 563         /*make sure that no slave==255 was not handled yet.*/
 564         if (slave >= dev->dev->caps.sqp_demux) {
 565                 mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
 566                              slave, dev->dev->caps.sqp_demux);
 567                 return -ENOENT;
 568         }
 569
 570         err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
 571         if (err)
 572                 pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
 573                          slave, err);
 574         return 0;
 575 }
 576
 577 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 578                         struct ib_wc *in_wc, struct ib_grh *in_grh,
 579                         struct ib_mad *in_mad, struct ib_mad *out_mad)
 580 {
 581         u16 slid, prev_lid = 0;
 582         int err;
 583         struct ib_port_attr pattr;
 584
 585         if (in_wc && in_wc->qp->qp_num) {
 586                 pr_debug("received MAD: slid:%d sqpn:%d "
 587                         "dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
 588                         in_wc->slid, in_wc->src_qp,
 589                         in_wc->dlid_path_bits,
 590                         in_wc->qp->qp_num,
 591                         in_wc->wc_flags,
 592                         in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
 593                         be16_to_cpu(in_mad->mad_hdr.attr_id));
 594                 if (in_wc->wc_flags & IB_WC_GRH) {
 595                         pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
 596                                  be64_to_cpu(in_grh->sgid.global.subnet_prefix),
 597                                  be64_to_cpu(in_grh->sgid.global.interface_id));
 598                         pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
 599                                  be64_to_cpu(in_grh->dgid.global.subnet_prefix),
 600                                  be64_to_cpu(in_grh->dgid.global.interface_id));
 601                 }
 602         }
 603
 604         slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
 605
 606         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
 607                 forward_trap(to_mdev(ibdev), port_num, in_mad);
 608                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
 609         }
 610
 611         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 612             in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 613                 if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
 614                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
 615                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
 616                         return IB_MAD_RESULT_SUCCESS;
 617
 618                 /*
 619                  * Don't process SMInfo queries -- the SMA can't handle them.
 620                  */
 621                 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
 622                         return IB_MAD_RESULT_SUCCESS;
 623         } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
 624                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
 625                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
 626                    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
 627                 if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
 628                     in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
 629                         return IB_MAD_RESULT_SUCCESS;
 630         } else
 631                 return IB_MAD_RESULT_SUCCESS;
 632
 633         if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 634              in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 635             in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
 636             in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
 637             !ib_query_port(ibdev, port_num, &pattr))
 638                 prev_lid = pattr.lid;
 639
 640         err = mlx4_MAD_IFC(to_mdev(ibdev),
 641                            (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
 642                            (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
 643                            MLX4_MAD_IFC_NET_VIEW,
 644                            port_num, in_wc, in_grh, in_mad, out_mad);
 645         if (err)
 646                 return IB_MAD_RESULT_FAILURE;
 647
 648         if (!out_mad->mad_hdr.status) {
 649                 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
 650                         smp_snoop(ibdev, port_num, in_mad, prev_lid);
 651                 node_desc_override(ibdev, out_mad);
 652         }
 653
 654         /* set return bit in status of directed route responses */
 655         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 656                 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
 657
 658         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
 659                 /* no response for trap repress */
 660                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
 661
 662         return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
 663 }
 664
 665 static void edit_counter(struct mlx4_counter *cnt,
 666                                         struct ib_pma_portcounters *pma_cnt)
 667 {
 668         pma_cnt->port_xmit_data = cpu_to_be32((be64_to_cpu(cnt->tx_bytes)>>2));
 669         pma_cnt->port_rcv_data  = cpu_to_be32((be64_to_cpu(cnt->rx_bytes)>>2));
 670         pma_cnt->port_xmit_packets = cpu_to_be32(be64_to_cpu(cnt->tx_frames));
 671         pma_cnt->port_rcv_packets  = cpu_to_be32(be64_to_cpu(cnt->rx_frames));
 672 }
 673
 674 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 675                         struct ib_wc *in_wc, struct ib_grh *in_grh,
 676                         struct ib_mad *in_mad, struct ib_mad *out_mad)
 677 {
 678         struct mlx4_cmd_mailbox *mailbox;
 679         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 680         int err;
 681         u32 inmod = dev->counters[port_num - 1] & 0xffff;
 682         u8 mode;
 683
 684         if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
 685                 return -EINVAL;
 686
 687         mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
 688         if (IS_ERR(mailbox))
 689                 return IB_MAD_RESULT_FAILURE;
 690
 691         err = mlx4_cmd_box(dev->dev, 0, mailbox->dma, inmod, 0,
 692                            MLX4_CMD_QUERY_IF_STAT, MLX4_CMD_TIME_CLASS_C,
 693                            MLX4_CMD_WRAPPED);
 694         if (err)
 695                 err = IB_MAD_RESULT_FAILURE;
 696         else {
 697                 memset(out_mad->data, 0, sizeof out_mad->data);
 698                 mode = ((struct mlx4_counter *)mailbox->buf)->counter_mode;
 699                 switch (mode & 0xf) {
 700                 case 0:
 701                         edit_counter(mailbox->buf,
 702                                                 (void *)(out_mad->data + 40));
 703                         err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
 704                         break;
 705                 default:
 706                         err = IB_MAD_RESULT_FAILURE;
 707                 }
 708         }
 709
 710         mlx4_free_cmd_mailbox(dev->dev, mailbox);
 711
 712         return err;
 713 }
 714
 715 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 716                         struct ib_wc *in_wc, struct ib_grh *in_grh,
 717                         struct ib_mad *in_mad, struct ib_mad *out_mad)
 718 {
 719         switch (rdma_port_get_link_layer(ibdev, port_num)) {
 720         case IB_LINK_LAYER_INFINIBAND:
 721                 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
 722                                       in_grh, in_mad, out_mad);
 723         case IB_LINK_LAYER_ETHERNET:
 724                 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
 725                                           in_grh, in_mad, out_mad);
 726         default:
 727                 return -EINVAL;
 728         }
 729 }
 730
 731 static void send_handler(struct ib_mad_agent *agent,
 732                          struct ib_mad_send_wc *mad_send_wc)
 733 {
 734         ib_free_send_mad(mad_send_wc->send_buf);
 735 }
 736
 737 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
 738 {
 739         struct ib_mad_agent *agent;
 740         int p, q;
 741         int ret;
 742         enum rdma_link_layer ll;
 743
 744         for (p = 0; p < dev->num_ports; ++p) {
 745                 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
 746                 for (q = 0; q <= 1; ++q) {
 747                         if (ll == IB_LINK_LAYER_INFINIBAND) {
 748                                 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
 749                                                               q ? IB_QPT_GSI : IB_QPT_SMI,
 750                                                               NULL, 0, send_handler,
 751                                                               NULL, NULL);
 752                                 if (IS_ERR(agent)) {
 753                                         ret = PTR_ERR(agent);
 754                                         goto err;
 755                                 }
 756                                 dev->send_agent[p][q] = agent;
 757                         } else
 758                                 dev->send_agent[p][q] = NULL;
 759                 }
 760         }
 761
 762         return 0;
 763
 764 err:
 765         for (p = 0; p < dev->num_ports; ++p)
 766                 for (q = 0; q <= 1; ++q)
 767                         if (dev->send_agent[p][q])
 768                                 ib_unregister_mad_agent(dev->send_agent[p][q]);
 769
 770         return ret;
 771 }
 772
 773 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
 774 {
 775         struct ib_mad_agent *agent;
 776         int p, q;
 777
 778         for (p = 0; p < dev->num_ports; ++p) {
 779                 for (q = 0; q <= 1; ++q) {
 780                         agent = dev->send_agent[p][q];
 781                         if (agent) {
 782                                 dev->send_agent[p][q] = NULL;
 783                                 ib_unregister_mad_agent(agent);
 784                         }
 785                 }
 786
 787                 if (dev->sm_ah[p])
 788                         ib_destroy_ah(dev->sm_ah[p]);
 789         }
 790 }
 791
 792 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
 793 {
 794         /* re-configure the mcg's */
 795         if (mlx4_is_master(dev->dev)) {
 796                 if (!dev->sriov.is_going_down)
 797                         mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
 798         }
 799         mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
 800 }
 801
 802 void handle_port_mgmt_change_event(struct work_struct *work)
 803 {
 804         struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
 805         struct mlx4_ib_dev *dev = ew->ib_dev;
 806         struct mlx4_eqe *eqe = &(ew->ib_eqe);
 807         u8 port = eqe->event.port_mgmt_change.port;
 808         u32 changed_attr;
 809
 810         switch (eqe->subtype) {
 811         case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
 812                 changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
 813
 814                 /* Update the SM ah - This should be done before handling
 815                    the other changed attributes so that MADs can be sent to the SM */
 816                 if (changed_attr & MSTR_SM_CHANGE_MASK) {
 817                         u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
 818                         u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
 819                         update_sm_ah(dev, port, lid, sl);
 820                 }
 821
 822                 /* Check if it is a lid change event */
 823                 if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
 824                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_LID_CHANGE);
 825
 826                 /* Generate GUID changed event */
 827                 if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK)
 828                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
 829
 830                 if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
 831                         handle_client_rereg_event(dev, port);
 832                 break;
 833
 834         case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
 835                 mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
 836                 break;
 837         case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
 838                 /* paravirtualized master's guid is guid 0 -- does not change */
 839                 if (!mlx4_is_master(dev->dev))
 840                         mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
 841                 break;
 842         default:
 843                 pr_warn("Unsupported subtype 0x%x for "
 844                         "Port Management Change event\n", eqe->subtype);
 845         }
 846
 847         kfree(ew);
 848 }
 849
 850 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
 851                             enum ib_event_type type)
 852 {
 853         struct ib_event event;
 854
 855         event.device            = &dev->ib_dev;
 856         event.element.port_num  = port_num;
 857         event.event             = type;
 858
 859         ib_dispatch_event(&event);
 860 }
 861
 862 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
 863 {
 864         unsigned long flags;
 865         struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
 866         struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
 867         spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
 868         if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
 869                 queue_work(ctx->wq, &ctx->work);
 870         spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
 871 }
 872
 873 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
 874                                   struct mlx4_ib_demux_pv_qp *tun_qp,
 875                                   int index)
 876 {
 877         struct ib_sge sg_list;
 878         struct ib_recv_wr recv_wr, *bad_recv_wr;
 879         int size;
 880
 881         size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
 882                 sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
 883
 884         sg_list.addr = tun_qp->ring[index].map;
 885         sg_list.length = size;
 886         sg_list.lkey = ctx->mr->lkey;
 887
 888         recv_wr.next = NULL;
 889         recv_wr.sg_list = &sg_list;
 890         recv_wr.num_sge = 1;
 891         recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
 892                 MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
 893         ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
 894                                       size, DMA_FROM_DEVICE);
 895         return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
 896 }
 897
 898 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
 899                 int slave, struct ib_sa_mad *sa_mad)
 900 {
 901         int ret = 0;
 902
 903         /* dispatch to different sa handlers */
 904         switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
 905         case IB_SA_ATTR_MC_MEMBER_REC:
 906                 ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
 907                 break;
 908         default:
 909                 break;
 910         }
 911         return ret;
 912 }
 913
 914 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
 915 {
 916         int slave_start = dev->dev->caps.sqp_start + 8 * slave;
 917
 918         return (qpn >= slave_start && qpn <= slave_start + 1);
 919 }
 920
 921
 922 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
 923                          enum ib_qp_type dest_qpt, u16 pkey_index, u32 remote_qpn,
 924                          u32 qkey, struct ib_ah_attr *attr, struct ib_mad *mad)
 925 {
 926         struct ib_sge list;
 927         struct ib_send_wr wr, *bad_wr;
 928         struct mlx4_ib_demux_pv_ctx *sqp_ctx;
 929         struct mlx4_ib_demux_pv_qp *sqp;
 930         struct mlx4_mad_snd_buf *sqp_mad;
 931         struct ib_ah *ah;
 932         struct ib_qp *send_qp = NULL;
 933         unsigned wire_tx_ix = 0;
 934         int ret = 0;
 935         u16 wire_pkey_ix;
 936         int src_qpnum;
 937         u8 sgid_index;
 938
 939
 940         sqp_ctx = dev->sriov.sqps[port-1];
 941
 942         /* check if proxy qp created */
 943         if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
 944                 return -EAGAIN;
 945
 946         /* QP0 forwarding only for Dom0 */
 947         if (dest_qpt == IB_QPT_SMI && (mlx4_master_func_num(dev->dev) != slave))
 948                 return -EINVAL;
 949
 950         if (dest_qpt == IB_QPT_SMI) {
 951                 src_qpnum = 0;
 952                 sqp = &sqp_ctx->qp[0];
 953                 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
 954         } else {
 955                 src_qpnum = 1;
 956                 sqp = &sqp_ctx->qp[1];
 957                 wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
 958         }
 959
 960         send_qp = sqp->qp;
 961
 962         /* create ah */
 963         sgid_index = attr->grh.sgid_index;
 964         attr->grh.sgid_index = 0;
 965         ah = ib_create_ah(sqp_ctx->pd, attr);
 966         if (IS_ERR(ah))
 967                 return -ENOMEM;
 968         attr->grh.sgid_index = sgid_index;
 969         to_mah(ah)->av.ib.gid_index = sgid_index;
 970         /* get rid of force-loopback bit */
 971         to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
 972         spin_lock(&sqp->tx_lock);
 973         if (sqp->tx_ix_head - sqp->tx_ix_tail >=
 974             (MLX4_NUM_TUNNEL_BUFS - 1))
 975                 ret = -EAGAIN;
 976         else
 977                 wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
 978         spin_unlock(&sqp->tx_lock);
 979         if (ret)
 980                 goto out;
 981
 982         sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
 983         if (sqp->tx_ring[wire_tx_ix].ah)
 984                 ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
 985         sqp->tx_ring[wire_tx_ix].ah = ah;
 986         ib_dma_sync_single_for_cpu(&dev->ib_dev,
 987                                    sqp->tx_ring[wire_tx_ix].buf.map,
 988                                    sizeof (struct mlx4_mad_snd_buf),
 989                                    DMA_TO_DEVICE);
 990
 991         memcpy(&sqp_mad->payload, mad, sizeof *mad);
 992
 993         ib_dma_sync_single_for_device(&dev->ib_dev,
 994                                       sqp->tx_ring[wire_tx_ix].buf.map,
 995                                       sizeof (struct mlx4_mad_snd_buf),
 996                                       DMA_TO_DEVICE);
 997
 998         list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
 999         list.length = sizeof (struct mlx4_mad_snd_buf);
1000         list.lkey = sqp_ctx->mr->lkey;
1001
1002         wr.wr.ud.ah = ah;
1003         wr.wr.ud.port_num = port;
1004         wr.wr.ud.pkey_index = wire_pkey_ix;
1005         wr.wr.ud.remote_qkey = qkey;
1006         wr.wr.ud.remote_qpn = remote_qpn;
1007         wr.next = NULL;
1008         wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1009         wr.sg_list = &list;
1010         wr.num_sge = 1;
1011         wr.opcode = IB_WR_SEND;
1012         wr.send_flags = IB_SEND_SIGNALED;
1013
1014         ret = ib_post_send(send_qp, &wr, &bad_wr);
1015 out:
1016         if (ret)
1017                 ib_destroy_ah(ah);
1018         return ret;
1019 }
1020
1021 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1022 {
1023         struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1024         struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1025         int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1026         struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1027         struct mlx4_ib_ah ah;
1028         struct ib_ah_attr ah_attr;
1029         u8 *slave_id;
1030         int slave;
1031
1032         /* Get slave that sent this packet */
1033         if (wc->src_qp < dev->dev->caps.sqp_start ||
1034             wc->src_qp >= dev->dev->caps.base_tunnel_sqpn ||
1035             (wc->src_qp & 0x1) != ctx->port - 1 ||
1036             wc->src_qp & 0x4) {
1037                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1038                 return;
1039         }
1040         slave = ((wc->src_qp & ~0x7) - dev->dev->caps.sqp_start) / 8;
1041         if (slave != ctx->slave) {
1042                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1043                              "belongs to another slave\n", wc->src_qp);
1044                 return;
1045         }
1046         if (slave != mlx4_master_func_num(dev->dev) && !(wc->src_qp & 0x2)) {
1047                 mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1048                              "non-master trying to send QP0 packets\n", wc->src_qp);
1049                 return;
1050         }
1051
1052         /* Map transaction ID */
1053         ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1054                                    sizeof (struct mlx4_tunnel_mad),
1055                                    DMA_FROM_DEVICE);
1056         switch (tunnel->mad.mad_hdr.method) {
1057         case IB_MGMT_METHOD_SET:
1058         case IB_MGMT_METHOD_GET:
1059         case IB_MGMT_METHOD_REPORT:
1060         case IB_SA_METHOD_GET_TABLE:
1061         case IB_SA_METHOD_DELETE:
1062         case IB_SA_METHOD_GET_MULTI:
1063         case IB_SA_METHOD_GET_TRACE_TBL:
1064                 slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1065                 if (*slave_id) {
1066                         mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1067                                      "class:%d slave:%d\n", *slave_id,
1068                                      tunnel->mad.mad_hdr.mgmt_class, slave);
1069                         return;
1070                 } else
1071                         *slave_id = slave;
1072         default:
1073                 /* nothing */;
1074         }
1075
1076         /* Class-specific handling */
1077         switch (tunnel->mad.mad_hdr.mgmt_class) {
1078         case IB_MGMT_CLASS_SUBN_ADM:
1079                 if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1080                               (struct ib_sa_mad *) &tunnel->mad))
1081                         return;
1082                 break;
1083         case IB_MGMT_CLASS_CM:
1084                 if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1085                               (struct ib_mad *) &tunnel->mad))
1086                         return;
1087                 break;
1088         case IB_MGMT_CLASS_DEVICE_MGMT:
1089                 if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1090                     tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1091                         return;
1092                 break;
1093         default:
1094                 /* Drop unsupported classes for slaves in tunnel mode */
1095                 if (slave != mlx4_master_func_num(dev->dev)) {
1096                         mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1097                                      "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1098                         return;
1099                 }
1100         }
1101
1102         /* We are using standard ib_core services to send the mad, so generate a
1103          * stadard address handle by decoding the tunnelled mlx4_ah fields */
1104         memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1105         ah.ibah.device = ctx->ib_dev;
1106         mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1107         if ((ah_attr.ah_flags & IB_AH_GRH) &&
1108             (ah_attr.grh.sgid_index != slave)) {
1109                 mlx4_ib_warn(ctx->ib_dev, "slave:%d accessed invalid sgid_index:%d\n",
1110                              slave, ah_attr.grh.sgid_index);
1111                 return;
1112         }
1113
1114         mlx4_ib_send_to_wire(dev, slave, ctx->port,
1115                              is_proxy_qp0(dev, wc->src_qp, slave) ?
1116                              IB_QPT_SMI : IB_QPT_GSI,
1117                              be16_to_cpu(tunnel->hdr.pkey_index),
1118                              be32_to_cpu(tunnel->hdr.remote_qpn),
1119                              be32_to_cpu(tunnel->hdr.qkey),
1120                              &ah_attr, &tunnel->mad);
1121 }
1122
1123 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1124                                  enum ib_qp_type qp_type, int is_tun)
1125 {
1126         int i;
1127         struct mlx4_ib_demux_pv_qp *tun_qp;
1128         int rx_buf_size, tx_buf_size;
1129
1130         if (qp_type > IB_QPT_GSI)
1131                 return -EINVAL;
1132
1133         tun_qp = &ctx->qp[qp_type];
1134
1135         tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1136                                GFP_KERNEL);
1137         if (!tun_qp->ring)
1138                 return -ENOMEM;
1139
1140         tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1141                                   sizeof (struct mlx4_ib_tun_tx_buf),
1142                                   GFP_KERNEL);
1143         if (!tun_qp->tx_ring) {
1144                 kfree(tun_qp->ring);
1145                 tun_qp->ring = NULL;
1146                 return -ENOMEM;
1147         }
1148
1149         if (is_tun) {
1150                 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1151                 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1152         } else {
1153                 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1154                 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1155         }
1156
1157         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1158                 tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1159                 if (!tun_qp->ring[i].addr)
1160                         goto err;
1161                 tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1162                                                         tun_qp->ring[i].addr,
1163                                                         rx_buf_size,
1164                                                         DMA_FROM_DEVICE);
1165         }
1166
1167         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1168                 tun_qp->tx_ring[i].buf.addr =
1169                         kmalloc(tx_buf_size, GFP_KERNEL);
1170                 if (!tun_qp->tx_ring[i].buf.addr)
1171                         goto tx_err;
1172                 tun_qp->tx_ring[i].buf.map =
1173                         ib_dma_map_single(ctx->ib_dev,
1174                                           tun_qp->tx_ring[i].buf.addr,
1175                                           tx_buf_size,
1176                                           DMA_TO_DEVICE);
1177                 tun_qp->tx_ring[i].ah = NULL;
1178         }
1179         spin_lock_init(&tun_qp->tx_lock);
1180         tun_qp->tx_ix_head = 0;
1181         tun_qp->tx_ix_tail = 0;
1182         tun_qp->proxy_qpt = qp_type;
1183
1184         return 0;
1185
1186 tx_err:
1187         while (i > 0) {
1188                 --i;
1189                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1190                                     tx_buf_size, DMA_TO_DEVICE);
1191                 kfree(tun_qp->tx_ring[i].buf.addr);
1192         }
1193         kfree(tun_qp->tx_ring);
1194         tun_qp->tx_ring = NULL;
1195         i = MLX4_NUM_TUNNEL_BUFS;
1196 err:
1197         while (i > 0) {
1198                 --i;
1199                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1200                                     rx_buf_size, DMA_FROM_DEVICE);
1201                 kfree(tun_qp->ring[i].addr);
1202         }
1203         kfree(tun_qp->ring);
1204         tun_qp->ring = NULL;
1205         return -ENOMEM;
1206 }
1207
1208 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1209                                      enum ib_qp_type qp_type, int is_tun)
1210 {
1211         int i;
1212         struct mlx4_ib_demux_pv_qp *tun_qp;
1213         int rx_buf_size, tx_buf_size;
1214
1215         if (qp_type > IB_QPT_GSI)
1216                 return;
1217
1218         tun_qp = &ctx->qp[qp_type];
1219         if (is_tun) {
1220                 rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1221                 tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1222         } else {
1223                 rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1224                 tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1225         }
1226
1227
1228         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1229                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1230                                     rx_buf_size, DMA_FROM_DEVICE);
1231                 kfree(tun_qp->ring[i].addr);
1232         }
1233
1234         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1235                 ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1236                                     tx_buf_size, DMA_TO_DEVICE);
1237                 kfree(tun_qp->tx_ring[i].buf.addr);
1238                 if (tun_qp->tx_ring[i].ah)
1239                         ib_destroy_ah(tun_qp->tx_ring[i].ah);
1240         }
1241         kfree(tun_qp->tx_ring);
1242         kfree(tun_qp->ring);
1243 }
1244
1245 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1246 {
1247         struct mlx4_ib_demux_pv_ctx *ctx;
1248         struct mlx4_ib_demux_pv_qp *tun_qp;
1249         struct ib_wc wc;
1250         int ret;
1251         ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1252         ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1253
1254         while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1255                 tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1256                 if (wc.status == IB_WC_SUCCESS) {
1257                         switch (wc.opcode) {
1258                         case IB_WC_RECV:
1259                                 mlx4_ib_multiplex_mad(ctx, &wc);
1260                                 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1261                                                              wc.wr_id &
1262                                                              (MLX4_NUM_TUNNEL_BUFS - 1));
1263                                 if (ret)
1264                                         pr_err("Failed reposting tunnel "
1265                                                "buf:%lld\n", wc.wr_id);
1266                                 break;
1267                         case IB_WC_SEND:
1268                                 pr_debug("received tunnel send completion:"
1269                                          "wrid=0x%llx, status=0x%x\n",
1270                                          wc.wr_id, wc.status);
1271                                 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1272                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1273                                 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1274                                         = NULL;
1275                                 spin_lock(&tun_qp->tx_lock);
1276                                 tun_qp->tx_ix_tail++;
1277                                 spin_unlock(&tun_qp->tx_lock);
1278
1279                                 break;
1280                         default:
1281                                 break;
1282                         }
1283                 } else  {
1284                         pr_debug("mlx4_ib: completion error in tunnel: %d."
1285                                  " status = %d, wrid = 0x%llx\n",
1286                                  ctx->slave, wc.status, wc.wr_id);
1287                         if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1288                                 ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1289                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1290                                 tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1291                                         = NULL;
1292                                 spin_lock(&tun_qp->tx_lock);
1293                                 tun_qp->tx_ix_tail++;
1294                                 spin_unlock(&tun_qp->tx_lock);
1295                         }
1296                 }
1297         }
1298 }
1299
1300 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1301 {
1302         struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1303
1304         /* It's worse than that! He's dead, Jim! */
1305         pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1306                event->event, sqp->port);
1307 }
1308
1309 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1310                             enum ib_qp_type qp_type, int create_tun)
1311 {
1312         int i, ret;
1313         struct mlx4_ib_demux_pv_qp *tun_qp;
1314         struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1315         struct ib_qp_attr attr;
1316         int qp_attr_mask_INIT;
1317
1318         if (qp_type > IB_QPT_GSI)
1319                 return -EINVAL;
1320
1321         tun_qp = &ctx->qp[qp_type];
1322
1323         memset(&qp_init_attr, 0, sizeof qp_init_attr);
1324         qp_init_attr.init_attr.send_cq = ctx->cq;
1325         qp_init_attr.init_attr.recv_cq = ctx->cq;
1326         qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1327         qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1328         qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1329         qp_init_attr.init_attr.cap.max_send_sge = 1;
1330         qp_init_attr.init_attr.cap.max_recv_sge = 1;
1331         if (create_tun) {
1332                 qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1333                 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1334                 qp_init_attr.port = ctx->port;
1335                 qp_init_attr.slave = ctx->slave;
1336                 qp_init_attr.proxy_qp_type = qp_type;
1337                 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1338                            IB_QP_QKEY | IB_QP_PORT;
1339         } else {
1340                 qp_init_attr.init_attr.qp_type = qp_type;
1341                 qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1342                 qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1343         }
1344         qp_init_attr.init_attr.port_num = ctx->port;
1345         qp_init_attr.init_attr.qp_context = ctx;
1346         qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1347         tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1348         if (IS_ERR(tun_qp->qp)) {
1349                 ret = PTR_ERR(tun_qp->qp);
1350                 tun_qp->qp = NULL;
1351                 pr_err("Couldn't create %s QP (%d)\n",
1352                        create_tun ? "tunnel" : "special", ret);
1353                 return ret;
1354         }
1355
1356         memset(&attr, 0, sizeof attr);
1357         attr.qp_state = IB_QPS_INIT;
1358         attr.pkey_index =
1359                 to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1360         attr.qkey = IB_QP1_QKEY;
1361         attr.port_num = ctx->port;
1362         ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1363         if (ret) {
1364                 pr_err("Couldn't change %s qp state to INIT (%d)\n",
1365                        create_tun ? "tunnel" : "special", ret);
1366                 goto err_qp;
1367         }
1368         attr.qp_state = IB_QPS_RTR;
1369         ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1370         if (ret) {
1371                 pr_err("Couldn't change %s qp state to RTR (%d)\n",
1372                        create_tun ? "tunnel" : "special", ret);
1373                 goto err_qp;
1374         }
1375         attr.qp_state = IB_QPS_RTS;
1376         attr.sq_psn = 0;
1377         ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1378         if (ret) {
1379                 pr_err("Couldn't change %s qp state to RTS (%d)\n",
1380                        create_tun ? "tunnel" : "special", ret);
1381                 goto err_qp;
1382         }
1383
1384         for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1385                 ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1386                 if (ret) {
1387                         pr_err(" mlx4_ib_post_pv_buf error"
1388                                " (err = %d, i = %d)\n", ret, i);
1389                         goto err_qp;
1390                 }
1391         }
1392         return 0;
1393
1394 err_qp:
1395         ib_destroy_qp(tun_qp->qp);
1396         tun_qp->qp = NULL;
1397         return ret;
1398 }
1399
1400 /*
1401  * IB MAD completion callback for real SQPs
1402  */
1403 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1404 {
1405         struct mlx4_ib_demux_pv_ctx *ctx;
1406         struct mlx4_ib_demux_pv_qp *sqp;
1407         struct ib_wc wc;
1408         struct ib_grh *grh;
1409         struct ib_mad *mad;
1410
1411         ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1412         ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1413
1414         while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1415                 sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1416                 if (wc.status == IB_WC_SUCCESS) {
1417                         switch (wc.opcode) {
1418                         case IB_WC_SEND:
1419                                 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1420                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1421                                 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1422                                         = NULL;
1423                                 spin_lock(&sqp->tx_lock);
1424                                 sqp->tx_ix_tail++;
1425                                 spin_unlock(&sqp->tx_lock);
1426                                 break;
1427                         case IB_WC_RECV:
1428                                 mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1429                                                 (sqp->ring[wc.wr_id &
1430                                                 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1431                                 grh = &(((struct mlx4_mad_rcv_buf *)
1432                                                 (sqp->ring[wc.wr_id &
1433                                                 (MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1434                                 mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1435                                 if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1436                                                            (MLX4_NUM_TUNNEL_BUFS - 1)))
1437                                         pr_err("Failed reposting SQP "
1438                                                "buf:%lld\n", wc.wr_id);
1439                                 break;
1440                         default:
1441                                 BUG_ON(1);
1442                                 break;
1443                         }
1444                 } else  {
1445                         pr_debug("mlx4_ib: completion error in tunnel: %d."
1446                                  " status = %d, wrid = 0x%llx\n",
1447                                  ctx->slave, wc.status, wc.wr_id);
1448                         if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1449                                 ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1450                                               (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1451                                 sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1452                                         = NULL;
1453                                 spin_lock(&sqp->tx_lock);
1454                                 sqp->tx_ix_tail++;
1455                                 spin_unlock(&sqp->tx_lock);
1456                         }
1457                 }
1458         }
1459 }
1460
1461 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1462                                struct mlx4_ib_demux_pv_ctx **ret_ctx)
1463 {
1464         struct mlx4_ib_demux_pv_ctx *ctx;
1465
1466         *ret_ctx = NULL;
1467         ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1468         if (!ctx) {
1469                 pr_err("failed allocating pv resource context "
1470                        "for port %d, slave %d\n", port, slave);
1471                 return -ENOMEM;
1472         }
1473
1474         ctx->ib_dev = &dev->ib_dev;
1475         ctx->port = port;
1476         ctx->slave = slave;
1477         *ret_ctx = ctx;
1478         return 0;
1479 }
1480
1481 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1482 {
1483         if (dev->sriov.demux[port - 1].tun[slave]) {
1484                 kfree(dev->sriov.demux[port - 1].tun[slave]);
1485                 dev->sriov.demux[port - 1].tun[slave] = NULL;
1486         }
1487 }
1488
1489 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1490                                int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1491 {
1492         int ret, cq_size;
1493
1494         ctx->state = DEMUX_PV_STATE_STARTING;
1495         /* have QP0 only on port owner, and only if link layer is IB */
1496         if (ctx->slave == mlx4_master_func_num(to_mdev(ctx->ib_dev)->dev) &&
1497             rdma_port_get_link_layer(ibdev, ctx->port) == IB_LINK_LAYER_INFINIBAND)
1498                 ctx->has_smi = 1;
1499
1500         if (ctx->has_smi) {
1501                 ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1502                 if (ret) {
1503                         pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1504                         goto err_out;
1505                 }
1506         }
1507
1508         ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1509         if (ret) {
1510                 pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1511                 goto err_out_qp0;
1512         }
1513
1514         cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1515         if (ctx->has_smi)
1516                 cq_size *= 2;
1517
1518         ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1519                                NULL, ctx, cq_size, 0);
1520         if (IS_ERR(ctx->cq)) {
1521                 ret = PTR_ERR(ctx->cq);
1522                 pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1523                 goto err_buf;
1524         }
1525
1526         ctx->pd = ib_alloc_pd(ctx->ib_dev);
1527         if (IS_ERR(ctx->pd)) {
1528                 ret = PTR_ERR(ctx->pd);
1529                 pr_err("Couldn't create tunnel PD (%d)\n", ret);
1530                 goto err_cq;
1531         }
1532
1533         ctx->mr = ib_get_dma_mr(ctx->pd, IB_ACCESS_LOCAL_WRITE);
1534         if (IS_ERR(ctx->mr)) {
1535                 ret = PTR_ERR(ctx->mr);
1536                 pr_err("Couldn't get tunnel DMA MR (%d)\n", ret);
1537                 goto err_pd;
1538         }
1539
1540         if (ctx->has_smi) {
1541                 ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1542                 if (ret) {
1543                         pr_err("Couldn't create %s QP0 (%d)\n",
1544                                create_tun ? "tunnel for" : "",  ret);
1545                         goto err_mr;
1546                 }
1547         }
1548
1549         ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1550         if (ret) {
1551                 pr_err("Couldn't create %s QP1 (%d)\n",
1552                        create_tun ? "tunnel for" : "",  ret);
1553                 goto err_qp0;
1554         }
1555
1556         if (create_tun)
1557                 INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1558         else
1559                 INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1560
1561         ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1562
1563         ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1564         if (ret) {
1565                 pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1566                 goto err_wq;
1567         }
1568         ctx->state = DEMUX_PV_STATE_ACTIVE;
1569         return 0;
1570
1571 err_wq:
1572         ctx->wq = NULL;
1573         ib_destroy_qp(ctx->qp[1].qp);
1574         ctx->qp[1].qp = NULL;
1575
1576
1577 err_qp0:
1578         if (ctx->has_smi)
1579                 ib_destroy_qp(ctx->qp[0].qp);
1580         ctx->qp[0].qp = NULL;
1581
1582 err_mr:
1583         ib_dereg_mr(ctx->mr);
1584         ctx->mr = NULL;
1585
1586 err_pd:
1587         ib_dealloc_pd(ctx->pd);
1588         ctx->pd = NULL;
1589
1590 err_cq:
1591         ib_destroy_cq(ctx->cq);
1592         ctx->cq = NULL;
1593
1594 err_buf:
1595         mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1596
1597 err_out_qp0:
1598         if (ctx->has_smi)
1599                 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1600 err_out:
1601         ctx->state = DEMUX_PV_STATE_DOWN;
1602         return ret;
1603 }
1604
1605 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1606                                  struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1607 {
1608         if (!ctx)
1609                 return;
1610         if (ctx->state > DEMUX_PV_STATE_DOWN) {
1611                 ctx->state = DEMUX_PV_STATE_DOWNING;
1612                 if (flush)
1613                         flush_workqueue(ctx->wq);
1614                 if (ctx->has_smi) {
1615                         ib_destroy_qp(ctx->qp[0].qp);
1616                         ctx->qp[0].qp = NULL;
1617                         mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1618                 }
1619                 ib_destroy_qp(ctx->qp[1].qp);
1620                 ctx->qp[1].qp = NULL;
1621                 mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1622                 ib_dereg_mr(ctx->mr);
1623                 ctx->mr = NULL;
1624                 ib_dealloc_pd(ctx->pd);
1625                 ctx->pd = NULL;
1626                 ib_destroy_cq(ctx->cq);
1627                 ctx->cq = NULL;
1628                 ctx->state = DEMUX_PV_STATE_DOWN;
1629         }
1630 }
1631
1632 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1633                                   int port, int do_init)
1634 {
1635         int ret = 0;
1636
1637         if (!do_init) {
1638                 clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1639                 /* for master, destroy real sqp resources */
1640                 if (slave == mlx4_master_func_num(dev->dev))
1641                         destroy_pv_resources(dev, slave, port,
1642                                              dev->sriov.sqps[port - 1], 1);
1643                 /* destroy the tunnel qp resources */
1644                 destroy_pv_resources(dev, slave, port,
1645                                      dev->sriov.demux[port - 1].tun[slave], 1);
1646                 return 0;
1647         }
1648
1649         /* create the tunnel qp resources */
1650         ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1651                                   dev->sriov.demux[port - 1].tun[slave]);
1652
1653         /* for master, create the real sqp resources */
1654         if (!ret && slave == mlx4_master_func_num(dev->dev))
1655                 ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1656                                           dev->sriov.sqps[port - 1]);
1657         return ret;
1658 }
1659
1660 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1661 {
1662         struct mlx4_ib_demux_work *dmxw;
1663
1664         dmxw = container_of(work, struct mlx4_ib_demux_work, work);
1665         mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
1666                                dmxw->do_init);
1667         kfree(dmxw);
1668         return;
1669 }
1670
1671 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
1672                                        struct mlx4_ib_demux_ctx *ctx,
1673                                        int port)
1674 {
1675         char name[12];
1676         int ret = 0;
1677         int i;
1678
1679         ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
1680                            sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
1681         if (!ctx->tun)
1682                 return -ENOMEM;
1683
1684         ctx->dev = dev;
1685         ctx->port = port;
1686         ctx->ib_dev = &dev->ib_dev;
1687
1688         for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
1689                 ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
1690                 if (ret) {
1691                         ret = -ENOMEM;
1692                         goto err_mcg;
1693                 }
1694         }
1695
1696         ret = mlx4_ib_mcg_port_init(ctx);
1697         if (ret) {
1698                 pr_err("Failed initializing mcg para-virt (%d)\n", ret);
1699                 goto err_mcg;
1700         }
1701
1702         snprintf(name, sizeof name, "mlx4_ibt%d", port);
1703         ctx->wq = create_singlethread_workqueue(name);
1704         if (!ctx->wq) {
1705                 pr_err("Failed to create tunnelling WQ for port %d\n", port);
1706                 ret = -ENOMEM;
1707                 goto err_wq;
1708         }
1709
1710         snprintf(name, sizeof name, "mlx4_ibud%d", port);
1711         ctx->ud_wq = create_singlethread_workqueue(name);
1712         if (!ctx->ud_wq) {
1713                 pr_err("Failed to create up/down WQ for port %d\n", port);
1714                 ret = -ENOMEM;
1715                 goto err_udwq;
1716         }
1717
1718         return 0;
1719
1720 err_udwq:
1721         destroy_workqueue(ctx->wq);
1722         ctx->wq = NULL;
1723
1724 err_wq:
1725         mlx4_ib_mcg_port_cleanup(ctx, 1);
1726 err_mcg:
1727         for (i = 0; i < dev->dev->caps.sqp_demux; i++)
1728                 free_pv_object(dev, i, port);
1729         kfree(ctx->tun);
1730         ctx->tun = NULL;
1731         return ret;
1732 }
1733
1734 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
1735 {
1736         if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
1737                 sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
1738                 flush_workqueue(sqp_ctx->wq);
1739                 if (sqp_ctx->has_smi) {
1740                         ib_destroy_qp(sqp_ctx->qp[0].qp);
1741                         sqp_ctx->qp[0].qp = NULL;
1742                         mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
1743                 }
1744                 ib_destroy_qp(sqp_ctx->qp[1].qp);
1745                 sqp_ctx->qp[1].qp = NULL;
1746                 mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
1747                 ib_dereg_mr(sqp_ctx->mr);
1748                 sqp_ctx->mr = NULL;
1749                 ib_dealloc_pd(sqp_ctx->pd);
1750                 sqp_ctx->pd = NULL;
1751                 ib_destroy_cq(sqp_ctx->cq);
1752                 sqp_ctx->cq = NULL;
1753                 sqp_ctx->state = DEMUX_PV_STATE_DOWN;
1754         }
1755 }
1756
1757 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
1758 {
1759         int i;
1760         if (ctx) {
1761                 struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1762                 mlx4_ib_mcg_port_cleanup(ctx, 1);
1763                 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
1764                         if (!ctx->tun[i])
1765                                 continue;
1766                         if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
1767                                 ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
1768                 }
1769                 flush_workqueue(ctx->wq);
1770                 for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
1771                         destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
1772                         free_pv_object(dev, i, ctx->port);
1773                 }
1774                 kfree(ctx->tun);
1775                 destroy_workqueue(ctx->ud_wq);
1776                 destroy_workqueue(ctx->wq);
1777         }
1778 }
1779
1780 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
1781 {
1782         int i;
1783
1784         if (!mlx4_is_master(dev->dev))
1785                 return;
1786         /* initialize or tear down tunnel QPs for the master */
1787         for (i = 0; i < dev->dev->caps.num_ports; i++)
1788                 mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
1789         return;
1790 }
1791
1792 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
1793 {
1794         int i = 0;
1795         int err;
1796
1797         if (!mlx4_is_mfunc(dev->dev))
1798                 return 0;
1799
1800         dev->sriov.is_going_down = 0;
1801         spin_lock_init(&dev->sriov.going_down_lock);
1802         mlx4_ib_cm_paravirt_init(dev);
1803
1804         mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
1805
1806         if (mlx4_is_slave(dev->dev)) {
1807                 mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
1808                 return 0;
1809         }
1810
1811         mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
1812                      dev->dev->caps.sqp_demux);
1813         for (i = 0; i < dev->num_ports; i++) {
1814                 err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
1815                                       &dev->sriov.sqps[i]);
1816                 if (err)
1817                         goto demux_err;
1818                 err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
1819                 if (err)
1820                         goto demux_err;
1821         }
1822         mlx4_ib_master_tunnels(dev, 1);
1823         return 0;
1824
1825 demux_err:
1826         while (i > 0) {
1827                 free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
1828                 mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
1829                 --i;
1830         }
1831         mlx4_ib_cm_paravirt_clean(dev, -1);
1832
1833         return err;
1834 }
1835
1836 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
1837 {
1838         int i;
1839         unsigned long flags;
1840
1841         if (!mlx4_is_mfunc(dev->dev))
1842                 return;
1843
1844         spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1845         dev->sriov.is_going_down = 1;
1846         spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1847         if (mlx4_is_master(dev->dev)) {
1848                 for (i = 0; i < dev->num_ports; i++) {
1849                         flush_workqueue(dev->sriov.demux[i].ud_wq);
1850                         mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
1851                         kfree(dev->sriov.sqps[i]);
1852                         dev->sriov.sqps[i] = NULL;
1853                         mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
1854                 }
1855
1856                 mlx4_ib_cm_paravirt_clean(dev, -1);
1857         }
1858 }