drivers/infiniband/hw/mlx5/main.c

   1 /*
   2  * Copyright (c) 2013-2015, Mellanox Technologies. 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 <linux/highmem.h>
  34 #include <linux/module.h>
  35 #include <linux/init.h>
  36 #include <linux/errno.h>
  37 #include <linux/pci.h>
  38 #include <linux/dma-mapping.h>
  39 #include <linux/slab.h>
  40 #if defined(CONFIG_X86)
  41 #include <asm/pat.h>
  42 #endif
  43 #include <linux/sched.h>
  44 #include <linux/delay.h>
  45 #include <rdma/ib_user_verbs.h>
  46 #include <rdma/ib_addr.h>
  47 #include <rdma/ib_cache.h>
  48 #include <linux/mlx5/port.h>
  49 #include <linux/mlx5/vport.h>
  50 #include <linux/list.h>
  51 #include <rdma/ib_smi.h>
  52 #include <rdma/ib_umem.h>
  53 #include <linux/in.h>
  54 #include <linux/etherdevice.h>
  55 #include <linux/mlx5/fs.h>
  56 #include "user.h"
  57 #include "mlx5_ib.h"
  58
  59 #define DRIVER_NAME "mlx5_ib"
  60 #define DRIVER_VERSION "2.2-1"
  61 #define DRIVER_RELDATE  "Feb 2014"
  62
  63 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
  64 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
  65 MODULE_LICENSE("Dual BSD/GPL");
  66 MODULE_VERSION(DRIVER_VERSION);
  67
  68 static int deprecated_prof_sel = 2;
  69 module_param_named(prof_sel, deprecated_prof_sel, int, 0444);
  70 MODULE_PARM_DESC(prof_sel, "profile selector. Deprecated here. Moved to module mlx5_core");
  71
  72 static char mlx5_version[] =
  73         DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
  74         DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
  75
  76 enum {
  77         MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
  78 };
  79
  80 static enum rdma_link_layer
  81 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
  82 {
  83         switch (port_type_cap) {
  84         case MLX5_CAP_PORT_TYPE_IB:
  85                 return IB_LINK_LAYER_INFINIBAND;
  86         case MLX5_CAP_PORT_TYPE_ETH:
  87                 return IB_LINK_LAYER_ETHERNET;
  88         default:
  89                 return IB_LINK_LAYER_UNSPECIFIED;
  90         }
  91 }
  92
  93 static enum rdma_link_layer
  94 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
  95 {
  96         struct mlx5_ib_dev *dev = to_mdev(device);
  97         int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
  98
  99         return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
 100 }
 101
 102 static int mlx5_netdev_event(struct notifier_block *this,
 103                              unsigned long event, void *ptr)
 104 {
 105         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
 106         struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
 107                                                  roce.nb);
 108
 109         if ((event != NETDEV_UNREGISTER) && (event != NETDEV_REGISTER))
 110                 return NOTIFY_DONE;
 111
 112         write_lock(&ibdev->roce.netdev_lock);
 113         if (ndev->dev.parent == &ibdev->mdev->pdev->dev)
 114                 ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ? NULL : ndev;
 115         write_unlock(&ibdev->roce.netdev_lock);
 116
 117         return NOTIFY_DONE;
 118 }
 119
 120 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
 121                                              u8 port_num)
 122 {
 123         struct mlx5_ib_dev *ibdev = to_mdev(device);
 124         struct net_device *ndev;
 125
 126         /* Ensure ndev does not disappear before we invoke dev_hold()
 127          */
 128         read_lock(&ibdev->roce.netdev_lock);
 129         ndev = ibdev->roce.netdev;
 130         if (ndev)
 131                 dev_hold(ndev);
 132         read_unlock(&ibdev->roce.netdev_lock);
 133
 134         return ndev;
 135 }
 136
 137 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
 138                                 struct ib_port_attr *props)
 139 {
 140         struct mlx5_ib_dev *dev = to_mdev(device);
 141         struct net_device *ndev;
 142         enum ib_mtu ndev_ib_mtu;
 143         u16 qkey_viol_cntr;
 144
 145         memset(props, 0, sizeof(*props));
 146
 147         props->port_cap_flags  |= IB_PORT_CM_SUP;
 148         props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
 149
 150         props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
 151                                                 roce_address_table_size);
 152         props->max_mtu          = IB_MTU_4096;
 153         props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
 154         props->pkey_tbl_len     = 1;
 155         props->state            = IB_PORT_DOWN;
 156         props->phys_state       = 3;
 157
 158         mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
 159         props->qkey_viol_cntr = qkey_viol_cntr;
 160
 161         ndev = mlx5_ib_get_netdev(device, port_num);
 162         if (!ndev)
 163                 return 0;
 164
 165         if (netif_running(ndev) && netif_carrier_ok(ndev)) {
 166                 props->state      = IB_PORT_ACTIVE;
 167                 props->phys_state = 5;
 168         }
 169
 170         ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
 171
 172         dev_put(ndev);
 173
 174         props->active_mtu       = min(props->max_mtu, ndev_ib_mtu);
 175
 176         props->active_width     = IB_WIDTH_4X;  /* TODO */
 177         props->active_speed     = IB_SPEED_QDR; /* TODO */
 178
 179         return 0;
 180 }
 181
 182 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
 183                                      const struct ib_gid_attr *attr,
 184                                      void *mlx5_addr)
 185 {
 186 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
 187         char *mlx5_addr_l3_addr = MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
 188                                                source_l3_address);
 189         void *mlx5_addr_mac     = MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
 190                                                source_mac_47_32);
 191
 192         if (!gid)
 193                 return;
 194
 195         ether_addr_copy(mlx5_addr_mac, attr->ndev->dev_addr);
 196
 197         if (is_vlan_dev(attr->ndev)) {
 198                 MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
 199                 MLX5_SET_RA(mlx5_addr, vlan_id, vlan_dev_vlan_id(attr->ndev));
 200         }
 201
 202         switch (attr->gid_type) {
 203         case IB_GID_TYPE_IB:
 204                 MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
 205                 break;
 206         case IB_GID_TYPE_ROCE_UDP_ENCAP:
 207                 MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
 208                 break;
 209
 210         default:
 211                 WARN_ON(true);
 212         }
 213
 214         if (attr->gid_type != IB_GID_TYPE_IB) {
 215                 if (ipv6_addr_v4mapped((void *)gid))
 216                         MLX5_SET_RA(mlx5_addr, roce_l3_type,
 217                                     MLX5_ROCE_L3_TYPE_IPV4);
 218                 else
 219                         MLX5_SET_RA(mlx5_addr, roce_l3_type,
 220                                     MLX5_ROCE_L3_TYPE_IPV6);
 221         }
 222
 223         if ((attr->gid_type == IB_GID_TYPE_IB) ||
 224             !ipv6_addr_v4mapped((void *)gid))
 225                 memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
 226         else
 227                 memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
 228 }
 229
 230 static int set_roce_addr(struct ib_device *device, u8 port_num,
 231                          unsigned int index,
 232                          const union ib_gid *gid,
 233                          const struct ib_gid_attr *attr)
 234 {
 235         struct mlx5_ib_dev *dev = to_mdev(device);
 236         u32  in[MLX5_ST_SZ_DW(set_roce_address_in)];
 237         u32 out[MLX5_ST_SZ_DW(set_roce_address_out)];
 238         void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
 239         enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
 240
 241         if (ll != IB_LINK_LAYER_ETHERNET)
 242                 return -EINVAL;
 243
 244         memset(in, 0, sizeof(in));
 245
 246         ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
 247
 248         MLX5_SET(set_roce_address_in, in, roce_address_index, index);
 249         MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
 250
 251         memset(out, 0, sizeof(out));
 252         return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
 253 }
 254
 255 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
 256                            unsigned int index, const union ib_gid *gid,
 257                            const struct ib_gid_attr *attr,
 258                            __always_unused void **context)
 259 {
 260         return set_roce_addr(device, port_num, index, gid, attr);
 261 }
 262
 263 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
 264                            unsigned int index, __always_unused void **context)
 265 {
 266         return set_roce_addr(device, port_num, index, NULL, NULL);
 267 }
 268
 269 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
 270                                int index)
 271 {
 272         struct ib_gid_attr attr;
 273         union ib_gid gid;
 274
 275         if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
 276                 return 0;
 277
 278         if (!attr.ndev)
 279                 return 0;
 280
 281         dev_put(attr.ndev);
 282
 283         if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
 284                 return 0;
 285
 286         return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
 287 }
 288
 289 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
 290 {
 291         if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
 292                 return !MLX5_CAP_GEN(dev->mdev, ib_virt);
 293         return 0;
 294 }
 295
 296 enum {
 297         MLX5_VPORT_ACCESS_METHOD_MAD,
 298         MLX5_VPORT_ACCESS_METHOD_HCA,
 299         MLX5_VPORT_ACCESS_METHOD_NIC,
 300 };
 301
 302 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
 303 {
 304         if (mlx5_use_mad_ifc(to_mdev(ibdev)))
 305                 return MLX5_VPORT_ACCESS_METHOD_MAD;
 306
 307         if (mlx5_ib_port_link_layer(ibdev, 1) ==
 308             IB_LINK_LAYER_ETHERNET)
 309                 return MLX5_VPORT_ACCESS_METHOD_NIC;
 310
 311         return MLX5_VPORT_ACCESS_METHOD_HCA;
 312 }
 313
 314 static void get_atomic_caps(struct mlx5_ib_dev *dev,
 315                             struct ib_device_attr *props)
 316 {
 317         u8 tmp;
 318         u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
 319         u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
 320         u8 atomic_req_8B_endianness_mode =
 321                 MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
 322
 323         /* Check if HW supports 8 bytes standard atomic operations and capable
 324          * of host endianness respond
 325          */
 326         tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
 327         if (((atomic_operations & tmp) == tmp) &&
 328             (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
 329             (atomic_req_8B_endianness_mode)) {
 330                 props->atomic_cap = IB_ATOMIC_HCA;
 331         } else {
 332                 props->atomic_cap = IB_ATOMIC_NONE;
 333         }
 334 }
 335
 336 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
 337                                         __be64 *sys_image_guid)
 338 {
 339         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 340         struct mlx5_core_dev *mdev = dev->mdev;
 341         u64 tmp;
 342         int err;
 343
 344         switch (mlx5_get_vport_access_method(ibdev)) {
 345         case MLX5_VPORT_ACCESS_METHOD_MAD:
 346                 return mlx5_query_mad_ifc_system_image_guid(ibdev,
 347                                                             sys_image_guid);
 348
 349         case MLX5_VPORT_ACCESS_METHOD_HCA:
 350                 err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
 351                 break;
 352
 353         case MLX5_VPORT_ACCESS_METHOD_NIC:
 354                 err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
 355                 break;
 356
 357         default:
 358                 return -EINVAL;
 359         }
 360
 361         if (!err)
 362                 *sys_image_guid = cpu_to_be64(tmp);
 363
 364         return err;
 365
 366 }
 367
 368 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
 369                                 u16 *max_pkeys)
 370 {
 371         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 372         struct mlx5_core_dev *mdev = dev->mdev;
 373
 374         switch (mlx5_get_vport_access_method(ibdev)) {
 375         case MLX5_VPORT_ACCESS_METHOD_MAD:
 376                 return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
 377
 378         case MLX5_VPORT_ACCESS_METHOD_HCA:
 379         case MLX5_VPORT_ACCESS_METHOD_NIC:
 380                 *max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
 381                                                 pkey_table_size));
 382                 return 0;
 383
 384         default:
 385                 return -EINVAL;
 386         }
 387 }
 388
 389 static int mlx5_query_vendor_id(struct ib_device *ibdev,
 390                                 u32 *vendor_id)
 391 {
 392         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 393
 394         switch (mlx5_get_vport_access_method(ibdev)) {
 395         case MLX5_VPORT_ACCESS_METHOD_MAD:
 396                 return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
 397
 398         case MLX5_VPORT_ACCESS_METHOD_HCA:
 399         case MLX5_VPORT_ACCESS_METHOD_NIC:
 400                 return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
 401
 402         default:
 403                 return -EINVAL;
 404         }
 405 }
 406
 407 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
 408                                 __be64 *node_guid)
 409 {
 410         u64 tmp;
 411         int err;
 412
 413         switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
 414         case MLX5_VPORT_ACCESS_METHOD_MAD:
 415                 return mlx5_query_mad_ifc_node_guid(dev, node_guid);
 416
 417         case MLX5_VPORT_ACCESS_METHOD_HCA:
 418                 err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
 419                 break;
 420
 421         case MLX5_VPORT_ACCESS_METHOD_NIC:
 422                 err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
 423                 break;
 424
 425         default:
 426                 return -EINVAL;
 427         }
 428
 429         if (!err)
 430                 *node_guid = cpu_to_be64(tmp);
 431
 432         return err;
 433 }
 434
 435 struct mlx5_reg_node_desc {
 436         u8      desc[64];
 437 };
 438
 439 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
 440 {
 441         struct mlx5_reg_node_desc in;
 442
 443         if (mlx5_use_mad_ifc(dev))
 444                 return mlx5_query_mad_ifc_node_desc(dev, node_desc);
 445
 446         memset(&in, 0, sizeof(in));
 447
 448         return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
 449                                     sizeof(struct mlx5_reg_node_desc),
 450                                     MLX5_REG_NODE_DESC, 0, 0);
 451 }
 452
 453 static int mlx5_ib_query_device(struct ib_device *ibdev,
 454                                 struct ib_device_attr *props,
 455                                 struct ib_udata *uhw)
 456 {
 457         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 458         struct mlx5_core_dev *mdev = dev->mdev;
 459         int err = -ENOMEM;
 460         int max_rq_sg;
 461         int max_sq_sg;
 462         u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
 463         struct mlx5_ib_query_device_resp resp = {};
 464         size_t resp_len;
 465         u64 max_tso;
 466
 467         resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
 468         if (uhw->outlen && uhw->outlen < resp_len)
 469                 return -EINVAL;
 470         else
 471                 resp.response_length = resp_len;
 472
 473         if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
 474                 return -EINVAL;
 475
 476         memset(props, 0, sizeof(*props));
 477         err = mlx5_query_system_image_guid(ibdev,
 478                                            &props->sys_image_guid);
 479         if (err)
 480                 return err;
 481
 482         err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
 483         if (err)
 484                 return err;
 485
 486         err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
 487         if (err)
 488                 return err;
 489
 490         props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
 491                 (fw_rev_min(dev->mdev) << 16) |
 492                 fw_rev_sub(dev->mdev);
 493         props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
 494                 IB_DEVICE_PORT_ACTIVE_EVENT             |
 495                 IB_DEVICE_SYS_IMAGE_GUID                |
 496                 IB_DEVICE_RC_RNR_NAK_GEN;
 497
 498         if (MLX5_CAP_GEN(mdev, pkv))
 499                 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
 500         if (MLX5_CAP_GEN(mdev, qkv))
 501                 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
 502         if (MLX5_CAP_GEN(mdev, apm))
 503                 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
 504         if (MLX5_CAP_GEN(mdev, xrc))
 505                 props->device_cap_flags |= IB_DEVICE_XRC;
 506         if (MLX5_CAP_GEN(mdev, imaicl)) {
 507                 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
 508                                            IB_DEVICE_MEM_WINDOW_TYPE_2B;
 509                 props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
 510                 /* We support 'Gappy' memory registration too */
 511                 props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
 512         }
 513         props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
 514         if (MLX5_CAP_GEN(mdev, sho)) {
 515                 props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
 516                 /* At this stage no support for signature handover */
 517                 props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
 518                                       IB_PROT_T10DIF_TYPE_2 |
 519                                       IB_PROT_T10DIF_TYPE_3;
 520                 props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
 521                                        IB_GUARD_T10DIF_CSUM;
 522         }
 523         if (MLX5_CAP_GEN(mdev, block_lb_mc))
 524                 props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
 525
 526         if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
 527                 if (MLX5_CAP_ETH(mdev, csum_cap))
 528                         props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
 529
 530                 if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
 531                         max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
 532                         if (max_tso) {
 533                                 resp.tso_caps.max_tso = 1 << max_tso;
 534                                 resp.tso_caps.supported_qpts |=
 535                                         1 << IB_QPT_RAW_PACKET;
 536                                 resp.response_length += sizeof(resp.tso_caps);
 537                         }
 538                 }
 539         }
 540
 541         if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
 542                 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
 543                 props->device_cap_flags |= IB_DEVICE_UD_TSO;
 544         }
 545
 546         if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
 547             MLX5_CAP_ETH(dev->mdev, scatter_fcs))
 548                 props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
 549
 550         if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
 551                 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
 552
 553         props->vendor_part_id      = mdev->pdev->device;
 554         props->hw_ver              = mdev->pdev->revision;
 555
 556         props->max_mr_size         = ~0ull;
 557         props->page_size_cap       = ~(min_page_size - 1);
 558         props->max_qp              = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
 559         props->max_qp_wr           = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
 560         max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
 561                      sizeof(struct mlx5_wqe_data_seg);
 562         max_sq_sg = (MLX5_CAP_GEN(mdev, max_wqe_sz_sq) -
 563                      sizeof(struct mlx5_wqe_ctrl_seg)) /
 564                      sizeof(struct mlx5_wqe_data_seg);
 565         props->max_sge = min(max_rq_sg, max_sq_sg);
 566         props->max_sge_rd          = MLX5_MAX_SGE_RD;
 567         props->max_cq              = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
 568         props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
 569         props->max_mr              = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
 570         props->max_pd              = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
 571         props->max_qp_rd_atom      = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
 572         props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
 573         props->max_srq             = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
 574         props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
 575         props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
 576         props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
 577         props->max_srq_sge         = max_rq_sg - 1;
 578         props->max_fast_reg_page_list_len =
 579                 1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
 580         get_atomic_caps(dev, props);
 581         props->masked_atomic_cap   = IB_ATOMIC_NONE;
 582         props->max_mcast_grp       = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
 583         props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
 584         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
 585                                            props->max_mcast_grp;
 586         props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
 587         props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
 588         props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
 589
 590 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 591         if (MLX5_CAP_GEN(mdev, pg))
 592                 props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
 593         props->odp_caps = dev->odp_caps;
 594 #endif
 595
 596         if (MLX5_CAP_GEN(mdev, cd))
 597                 props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
 598
 599         if (!mlx5_core_is_pf(mdev))
 600                 props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
 601
 602         if (uhw->outlen) {
 603                 err = ib_copy_to_udata(uhw, &resp, resp.response_length);
 604
 605                 if (err)
 606                         return err;
 607         }
 608
 609         return 0;
 610 }
 611
 612 enum mlx5_ib_width {
 613         MLX5_IB_WIDTH_1X        = 1 << 0,
 614         MLX5_IB_WIDTH_2X        = 1 << 1,
 615         MLX5_IB_WIDTH_4X        = 1 << 2,
 616         MLX5_IB_WIDTH_8X        = 1 << 3,
 617         MLX5_IB_WIDTH_12X       = 1 << 4
 618 };
 619
 620 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
 621                                   u8 *ib_width)
 622 {
 623         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 624         int err = 0;
 625
 626         if (active_width & MLX5_IB_WIDTH_1X) {
 627                 *ib_width = IB_WIDTH_1X;
 628         } else if (active_width & MLX5_IB_WIDTH_2X) {
 629                 mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
 630                             (int)active_width);
 631                 err = -EINVAL;
 632         } else if (active_width & MLX5_IB_WIDTH_4X) {
 633                 *ib_width = IB_WIDTH_4X;
 634         } else if (active_width & MLX5_IB_WIDTH_8X) {
 635                 *ib_width = IB_WIDTH_8X;
 636         } else if (active_width & MLX5_IB_WIDTH_12X) {
 637                 *ib_width = IB_WIDTH_12X;
 638         } else {
 639                 mlx5_ib_dbg(dev, "Invalid active_width %d\n",
 640                             (int)active_width);
 641                 err = -EINVAL;
 642         }
 643
 644         return err;
 645 }
 646
 647 static int mlx5_mtu_to_ib_mtu(int mtu)
 648 {
 649         switch (mtu) {
 650         case 256: return 1;
 651         case 512: return 2;
 652         case 1024: return 3;
 653         case 2048: return 4;
 654         case 4096: return 5;
 655         default:
 656                 pr_warn("invalid mtu\n");
 657                 return -1;
 658         }
 659 }
 660
 661 enum ib_max_vl_num {
 662         __IB_MAX_VL_0           = 1,
 663         __IB_MAX_VL_0_1         = 2,
 664         __IB_MAX_VL_0_3         = 3,
 665         __IB_MAX_VL_0_7         = 4,
 666         __IB_MAX_VL_0_14        = 5,
 667 };
 668
 669 enum mlx5_vl_hw_cap {
 670         MLX5_VL_HW_0    = 1,
 671         MLX5_VL_HW_0_1  = 2,
 672         MLX5_VL_HW_0_2  = 3,
 673         MLX5_VL_HW_0_3  = 4,
 674         MLX5_VL_HW_0_4  = 5,
 675         MLX5_VL_HW_0_5  = 6,
 676         MLX5_VL_HW_0_6  = 7,
 677         MLX5_VL_HW_0_7  = 8,
 678         MLX5_VL_HW_0_14 = 15
 679 };
 680
 681 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
 682                                 u8 *max_vl_num)
 683 {
 684         switch (vl_hw_cap) {
 685         case MLX5_VL_HW_0:
 686                 *max_vl_num = __IB_MAX_VL_0;
 687                 break;
 688         case MLX5_VL_HW_0_1:
 689                 *max_vl_num = __IB_MAX_VL_0_1;
 690                 break;
 691         case MLX5_VL_HW_0_3:
 692                 *max_vl_num = __IB_MAX_VL_0_3;
 693                 break;
 694         case MLX5_VL_HW_0_7:
 695                 *max_vl_num = __IB_MAX_VL_0_7;
 696                 break;
 697         case MLX5_VL_HW_0_14:
 698                 *max_vl_num = __IB_MAX_VL_0_14;
 699                 break;
 700
 701         default:
 702                 return -EINVAL;
 703         }
 704
 705         return 0;
 706 }
 707
 708 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
 709                                struct ib_port_attr *props)
 710 {
 711         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 712         struct mlx5_core_dev *mdev = dev->mdev;
 713         struct mlx5_hca_vport_context *rep;
 714         u16 max_mtu;
 715         u16 oper_mtu;
 716         int err;
 717         u8 ib_link_width_oper;
 718         u8 vl_hw_cap;
 719
 720         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
 721         if (!rep) {
 722                 err = -ENOMEM;
 723                 goto out;
 724         }
 725
 726         memset(props, 0, sizeof(*props));
 727
 728         err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
 729         if (err)
 730                 goto out;
 731
 732         props->lid              = rep->lid;
 733         props->lmc              = rep->lmc;
 734         props->sm_lid           = rep->sm_lid;
 735         props->sm_sl            = rep->sm_sl;
 736         props->state            = rep->vport_state;
 737         props->phys_state       = rep->port_physical_state;
 738         props->port_cap_flags   = rep->cap_mask1;
 739         props->gid_tbl_len      = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
 740         props->max_msg_sz       = 1 << MLX5_CAP_GEN(mdev, log_max_msg);
 741         props->pkey_tbl_len     = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
 742         props->bad_pkey_cntr    = rep->pkey_violation_counter;
 743         props->qkey_viol_cntr   = rep->qkey_violation_counter;
 744         props->subnet_timeout   = rep->subnet_timeout;
 745         props->init_type_reply  = rep->init_type_reply;
 746         props->grh_required     = rep->grh_required;
 747
 748         err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
 749         if (err)
 750                 goto out;
 751
 752         err = translate_active_width(ibdev, ib_link_width_oper,
 753                                      &props->active_width);
 754         if (err)
 755                 goto out;
 756         err = mlx5_query_port_proto_oper(mdev, &props->active_speed, MLX5_PTYS_IB,
 757                                          port);
 758         if (err)
 759                 goto out;
 760
 761         mlx5_query_port_max_mtu(mdev, &max_mtu, port);
 762
 763         props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
 764
 765         mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
 766
 767         props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
 768
 769         err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
 770         if (err)
 771                 goto out;
 772
 773         err = translate_max_vl_num(ibdev, vl_hw_cap,
 774                                    &props->max_vl_num);
 775 out:
 776         kfree(rep);
 777         return err;
 778 }
 779
 780 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
 781                        struct ib_port_attr *props)
 782 {
 783         switch (mlx5_get_vport_access_method(ibdev)) {
 784         case MLX5_VPORT_ACCESS_METHOD_MAD:
 785                 return mlx5_query_mad_ifc_port(ibdev, port, props);
 786
 787         case MLX5_VPORT_ACCESS_METHOD_HCA:
 788                 return mlx5_query_hca_port(ibdev, port, props);
 789
 790         case MLX5_VPORT_ACCESS_METHOD_NIC:
 791                 return mlx5_query_port_roce(ibdev, port, props);
 792
 793         default:
 794                 return -EINVAL;
 795         }
 796 }
 797
 798 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
 799                              union ib_gid *gid)
 800 {
 801         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 802         struct mlx5_core_dev *mdev = dev->mdev;
 803
 804         switch (mlx5_get_vport_access_method(ibdev)) {
 805         case MLX5_VPORT_ACCESS_METHOD_MAD:
 806                 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
 807
 808         case MLX5_VPORT_ACCESS_METHOD_HCA:
 809                 return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
 810
 811         default:
 812                 return -EINVAL;
 813         }
 814
 815 }
 816
 817 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
 818                               u16 *pkey)
 819 {
 820         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 821         struct mlx5_core_dev *mdev = dev->mdev;
 822
 823         switch (mlx5_get_vport_access_method(ibdev)) {
 824         case MLX5_VPORT_ACCESS_METHOD_MAD:
 825                 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
 826
 827         case MLX5_VPORT_ACCESS_METHOD_HCA:
 828         case MLX5_VPORT_ACCESS_METHOD_NIC:
 829                 return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
 830                                                  pkey);
 831         default:
 832                 return -EINVAL;
 833         }
 834 }
 835
 836 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
 837                                  struct ib_device_modify *props)
 838 {
 839         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 840         struct mlx5_reg_node_desc in;
 841         struct mlx5_reg_node_desc out;
 842         int err;
 843
 844         if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
 845                 return -EOPNOTSUPP;
 846
 847         if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
 848                 return 0;
 849
 850         /*
 851          * If possible, pass node desc to FW, so it can generate
 852          * a 144 trap.  If cmd fails, just ignore.
 853          */
 854         memcpy(&in, props->node_desc, 64);
 855         err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
 856                                    sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
 857         if (err)
 858                 return err;
 859
 860         memcpy(ibdev->node_desc, props->node_desc, 64);
 861
 862         return err;
 863 }
 864
 865 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
 866                                struct ib_port_modify *props)
 867 {
 868         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 869         struct ib_port_attr attr;
 870         u32 tmp;
 871         int err;
 872
 873         mutex_lock(&dev->cap_mask_mutex);
 874
 875         err = mlx5_ib_query_port(ibdev, port, &attr);
 876         if (err)
 877                 goto out;
 878
 879         tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
 880                 ~props->clr_port_cap_mask;
 881
 882         err = mlx5_set_port_caps(dev->mdev, port, tmp);
 883
 884 out:
 885         mutex_unlock(&dev->cap_mask_mutex);
 886         return err;
 887 }
 888
 889 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
 890                                                   struct ib_udata *udata)
 891 {
 892         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 893         struct mlx5_ib_alloc_ucontext_req_v2 req = {};
 894         struct mlx5_ib_alloc_ucontext_resp resp = {};
 895         struct mlx5_ib_ucontext *context;
 896         struct mlx5_uuar_info *uuari;
 897         struct mlx5_uar *uars;
 898         int gross_uuars;
 899         int num_uars;
 900         int ver;
 901         int uuarn;
 902         int err;
 903         int i;
 904         size_t reqlen;
 905         size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
 906                                      max_cqe_version);
 907
 908         if (!dev->ib_active)
 909                 return ERR_PTR(-EAGAIN);
 910
 911         if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
 912                 return ERR_PTR(-EINVAL);
 913
 914         reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
 915         if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
 916                 ver = 0;
 917         else if (reqlen >= min_req_v2)
 918                 ver = 2;
 919         else
 920                 return ERR_PTR(-EINVAL);
 921
 922         err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
 923         if (err)
 924                 return ERR_PTR(err);
 925
 926         if (req.flags)
 927                 return ERR_PTR(-EINVAL);
 928
 929         if (req.total_num_uuars > MLX5_MAX_UUARS)
 930                 return ERR_PTR(-ENOMEM);
 931
 932         if (req.total_num_uuars == 0)
 933                 return ERR_PTR(-EINVAL);
 934
 935         if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
 936                 return ERR_PTR(-EOPNOTSUPP);
 937
 938         if (reqlen > sizeof(req) &&
 939             !ib_is_udata_cleared(udata, sizeof(req),
 940                                  reqlen - sizeof(req)))
 941                 return ERR_PTR(-EOPNOTSUPP);
 942
 943         req.total_num_uuars = ALIGN(req.total_num_uuars,
 944                                     MLX5_NON_FP_BF_REGS_PER_PAGE);
 945         if (req.num_low_latency_uuars > req.total_num_uuars - 1)
 946                 return ERR_PTR(-EINVAL);
 947
 948         num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
 949         gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
 950         resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
 951         if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
 952                 resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
 953         resp.cache_line_size = L1_CACHE_BYTES;
 954         resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
 955         resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
 956         resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
 957         resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
 958         resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
 959         resp.cqe_version = min_t(__u8,
 960                                  (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
 961                                  req.max_cqe_version);
 962         resp.response_length = min(offsetof(typeof(resp), response_length) +
 963                                    sizeof(resp.response_length), udata->outlen);
 964
 965         context = kzalloc(sizeof(*context), GFP_KERNEL);
 966         if (!context)
 967                 return ERR_PTR(-ENOMEM);
 968
 969         uuari = &context->uuari;
 970         mutex_init(&uuari->lock);
 971         uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
 972         if (!uars) {
 973                 err = -ENOMEM;
 974                 goto out_ctx;
 975         }
 976
 977         uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
 978                                 sizeof(*uuari->bitmap),
 979                                 GFP_KERNEL);
 980         if (!uuari->bitmap) {
 981                 err = -ENOMEM;
 982                 goto out_uar_ctx;
 983         }
 984         /*
 985          * clear all fast path uuars
 986          */
 987         for (i = 0; i < gross_uuars; i++) {
 988                 uuarn = i & 3;
 989                 if (uuarn == 2 || uuarn == 3)
 990                         set_bit(i, uuari->bitmap);
 991         }
 992
 993         uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
 994         if (!uuari->count) {
 995                 err = -ENOMEM;
 996                 goto out_bitmap;
 997         }
 998
 999         for (i = 0; i < num_uars; i++) {
1000                 err = mlx5_cmd_alloc_uar(dev->mdev, &uars[i].index);
1001                 if (err)
1002                         goto out_count;
1003         }
1004
1005 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1006         context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1007 #endif
1008
1009         if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1010                 err = mlx5_core_alloc_transport_domain(dev->mdev,
1011                                                        &context->tdn);
1012                 if (err)
1013                         goto out_uars;
1014         }
1015
1016         INIT_LIST_HEAD(&context->vma_private_list);
1017         INIT_LIST_HEAD(&context->db_page_list);
1018         mutex_init(&context->db_page_mutex);
1019
1020         resp.tot_uuars = req.total_num_uuars;
1021         resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1022
1023         if (field_avail(typeof(resp), cqe_version, udata->outlen))
1024                 resp.response_length += sizeof(resp.cqe_version);
1025
1026         if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1027                 resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE;
1028                 resp.response_length += sizeof(resp.cmds_supp_uhw);
1029         }
1030
1031         /*
1032          * We don't want to expose information from the PCI bar that is located
1033          * after 4096 bytes, so if the arch only supports larger pages, let's
1034          * pretend we don't support reading the HCA's core clock. This is also
1035          * forced by mmap function.
1036          */
1037         if (PAGE_SIZE <= 4096 &&
1038             field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1039                 resp.comp_mask |=
1040                         MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1041                 resp.hca_core_clock_offset =
1042                         offsetof(struct mlx5_init_seg, internal_timer_h) %
1043                         PAGE_SIZE;
1044                 resp.response_length += sizeof(resp.hca_core_clock_offset) +
1045                                         sizeof(resp.reserved2);
1046         }
1047
1048         err = ib_copy_to_udata(udata, &resp, resp.response_length);
1049         if (err)
1050                 goto out_td;
1051
1052         uuari->ver = ver;
1053         uuari->num_low_latency_uuars = req.num_low_latency_uuars;
1054         uuari->uars = uars;
1055         uuari->num_uars = num_uars;
1056         context->cqe_version = resp.cqe_version;
1057
1058         return &context->ibucontext;
1059
1060 out_td:
1061         if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1062                 mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1063
1064 out_uars:
1065         for (i--; i >= 0; i--)
1066                 mlx5_cmd_free_uar(dev->mdev, uars[i].index);
1067 out_count:
1068         kfree(uuari->count);
1069
1070 out_bitmap:
1071         kfree(uuari->bitmap);
1072
1073 out_uar_ctx:
1074         kfree(uars);
1075
1076 out_ctx:
1077         kfree(context);
1078         return ERR_PTR(err);
1079 }
1080
1081 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1082 {
1083         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1084         struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1085         struct mlx5_uuar_info *uuari = &context->uuari;
1086         int i;
1087
1088         if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1089                 mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1090
1091         for (i = 0; i < uuari->num_uars; i++) {
1092                 if (mlx5_cmd_free_uar(dev->mdev, uuari->uars[i].index))
1093                         mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
1094         }
1095
1096         kfree(uuari->count);
1097         kfree(uuari->bitmap);
1098         kfree(uuari->uars);
1099         kfree(context);
1100
1101         return 0;
1102 }
1103
1104 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
1105 {
1106         return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + index;
1107 }
1108
1109 static int get_command(unsigned long offset)
1110 {
1111         return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1112 }
1113
1114 static int get_arg(unsigned long offset)
1115 {
1116         return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1117 }
1118
1119 static int get_index(unsigned long offset)
1120 {
1121         return get_arg(offset);
1122 }
1123
1124 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1125 {
1126         /* vma_open is called when a new VMA is created on top of our VMA.  This
1127          * is done through either mremap flow or split_vma (usually due to
1128          * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1129          * as this VMA is strongly hardware related.  Therefore we set the
1130          * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1131          * calling us again and trying to do incorrect actions.  We assume that
1132          * the original VMA size is exactly a single page, and therefore all
1133          * "splitting" operation will not happen to it.
1134          */
1135         area->vm_ops = NULL;
1136 }
1137
1138 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1139 {
1140         struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1141
1142         /* It's guaranteed that all VMAs opened on a FD are closed before the
1143          * file itself is closed, therefore no sync is needed with the regular
1144          * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1145          * However need a sync with accessing the vma as part of
1146          * mlx5_ib_disassociate_ucontext.
1147          * The close operation is usually called under mm->mmap_sem except when
1148          * process is exiting.
1149          * The exiting case is handled explicitly as part of
1150          * mlx5_ib_disassociate_ucontext.
1151          */
1152         mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1153
1154         /* setting the vma context pointer to null in the mlx5_ib driver's
1155          * private data, to protect a race condition in
1156          * mlx5_ib_disassociate_ucontext().
1157          */
1158         mlx5_ib_vma_priv_data->vma = NULL;
1159         list_del(&mlx5_ib_vma_priv_data->list);
1160         kfree(mlx5_ib_vma_priv_data);
1161 }
1162
1163 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1164         .open = mlx5_ib_vma_open,
1165         .close = mlx5_ib_vma_close
1166 };
1167
1168 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1169                                 struct mlx5_ib_ucontext *ctx)
1170 {
1171         struct mlx5_ib_vma_private_data *vma_prv;
1172         struct list_head *vma_head = &ctx->vma_private_list;
1173
1174         vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1175         if (!vma_prv)
1176                 return -ENOMEM;
1177
1178         vma_prv->vma = vma;
1179         vma->vm_private_data = vma_prv;
1180         vma->vm_ops =  &mlx5_ib_vm_ops;
1181
1182         list_add(&vma_prv->list, vma_head);
1183
1184         return 0;
1185 }
1186
1187 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1188 {
1189         int ret;
1190         struct vm_area_struct *vma;
1191         struct mlx5_ib_vma_private_data *vma_private, *n;
1192         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1193         struct task_struct *owning_process  = NULL;
1194         struct mm_struct   *owning_mm       = NULL;
1195
1196         owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1197         if (!owning_process)
1198                 return;
1199
1200         owning_mm = get_task_mm(owning_process);
1201         if (!owning_mm) {
1202                 pr_info("no mm, disassociate ucontext is pending task termination\n");
1203                 while (1) {
1204                         put_task_struct(owning_process);
1205                         usleep_range(1000, 2000);
1206                         owning_process = get_pid_task(ibcontext->tgid,
1207                                                       PIDTYPE_PID);
1208                         if (!owning_process ||
1209                             owning_process->state == TASK_DEAD) {
1210                                 pr_info("disassociate ucontext done, task was terminated\n");
1211                                 /* in case task was dead need to release the
1212                                  * task struct.
1213                                  */
1214                                 if (owning_process)
1215                                         put_task_struct(owning_process);
1216                                 return;
1217                         }
1218                 }
1219         }
1220
1221         /* need to protect from a race on closing the vma as part of
1222          * mlx5_ib_vma_close.
1223          */
1224         down_read(&owning_mm->mmap_sem);
1225         list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1226                                  list) {
1227                 vma = vma_private->vma;
1228                 ret = zap_vma_ptes(vma, vma->vm_start,
1229                                    PAGE_SIZE);
1230                 WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1231                 /* context going to be destroyed, should
1232                  * not access ops any more.
1233                  */
1234                 vma->vm_ops = NULL;
1235                 list_del(&vma_private->list);
1236                 kfree(vma_private);
1237         }
1238         up_read(&owning_mm->mmap_sem);
1239         mmput(owning_mm);
1240         put_task_struct(owning_process);
1241 }
1242
1243 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1244 {
1245         switch (cmd) {
1246         case MLX5_IB_MMAP_WC_PAGE:
1247                 return "WC";
1248         case MLX5_IB_MMAP_REGULAR_PAGE:
1249                 return "best effort WC";
1250         case MLX5_IB_MMAP_NC_PAGE:
1251                 return "NC";
1252         default:
1253                 return NULL;
1254         }
1255 }
1256
1257 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1258                     struct vm_area_struct *vma,
1259                     struct mlx5_ib_ucontext *context)
1260 {
1261         struct mlx5_uuar_info *uuari = &context->uuari;
1262         int err;
1263         unsigned long idx;
1264         phys_addr_t pfn, pa;
1265         pgprot_t prot;
1266
1267         switch (cmd) {
1268         case MLX5_IB_MMAP_WC_PAGE:
1269 /* Some architectures don't support WC memory */
1270 #if defined(CONFIG_X86)
1271                 if (!pat_enabled())
1272                         return -EPERM;
1273 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1274                         return -EPERM;
1275 #endif
1276         /* fall through */
1277         case MLX5_IB_MMAP_REGULAR_PAGE:
1278                 /* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1279                 prot = pgprot_writecombine(vma->vm_page_prot);
1280                 break;
1281         case MLX5_IB_MMAP_NC_PAGE:
1282                 prot = pgprot_noncached(vma->vm_page_prot);
1283                 break;
1284         default:
1285                 return -EINVAL;
1286         }
1287
1288         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1289                 return -EINVAL;
1290
1291         idx = get_index(vma->vm_pgoff);
1292         if (idx >= uuari->num_uars)
1293                 return -EINVAL;
1294
1295         pfn = uar_index2pfn(dev, uuari->uars[idx].index);
1296         mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1297
1298         vma->vm_page_prot = prot;
1299         err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1300                                  PAGE_SIZE, vma->vm_page_prot);
1301         if (err) {
1302                 mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
1303                             err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
1304                 return -EAGAIN;
1305         }
1306
1307         pa = pfn << PAGE_SHIFT;
1308         mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
1309                     vma->vm_start, &pa);
1310
1311         return mlx5_ib_set_vma_data(vma, context);
1312 }
1313
1314 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1315 {
1316         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1317         struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1318         unsigned long command;
1319         phys_addr_t pfn;
1320
1321         command = get_command(vma->vm_pgoff);
1322         switch (command) {
1323         case MLX5_IB_MMAP_WC_PAGE:
1324         case MLX5_IB_MMAP_NC_PAGE:
1325         case MLX5_IB_MMAP_REGULAR_PAGE:
1326                 return uar_mmap(dev, command, vma, context);
1327
1328         case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1329                 return -ENOSYS;
1330
1331         case MLX5_IB_MMAP_CORE_CLOCK:
1332                 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1333                         return -EINVAL;
1334
1335                 if (vma->vm_flags & VM_WRITE)
1336                         return -EPERM;
1337
1338                 /* Don't expose to user-space information it shouldn't have */
1339                 if (PAGE_SIZE > 4096)
1340                         return -EOPNOTSUPP;
1341
1342                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1343                 pfn = (dev->mdev->iseg_base +
1344                        offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1345                         PAGE_SHIFT;
1346                 if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1347                                        PAGE_SIZE, vma->vm_page_prot))
1348                         return -EAGAIN;
1349
1350                 mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
1351                             vma->vm_start,
1352                             (unsigned long long)pfn << PAGE_SHIFT);
1353                 break;
1354
1355         default:
1356                 return -EINVAL;
1357         }
1358
1359         return 0;
1360 }
1361
1362 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1363                                       struct ib_ucontext *context,
1364                                       struct ib_udata *udata)
1365 {
1366         struct mlx5_ib_alloc_pd_resp resp;
1367         struct mlx5_ib_pd *pd;
1368         int err;
1369
1370         pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1371         if (!pd)
1372                 return ERR_PTR(-ENOMEM);
1373
1374         err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1375         if (err) {
1376                 kfree(pd);
1377                 return ERR_PTR(err);
1378         }
1379
1380         if (context) {
1381                 resp.pdn = pd->pdn;
1382                 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1383                         mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1384                         kfree(pd);
1385                         return ERR_PTR(-EFAULT);
1386                 }
1387         }
1388
1389         return &pd->ibpd;
1390 }
1391
1392 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1393 {
1394         struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1395         struct mlx5_ib_pd *mpd = to_mpd(pd);
1396
1397         mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1398         kfree(mpd);
1399
1400         return 0;
1401 }
1402
1403 static bool outer_header_zero(u32 *match_criteria)
1404 {
1405         int size = MLX5_ST_SZ_BYTES(fte_match_param);
1406         char *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_criteria,
1407                                              outer_headers);
1408
1409         return outer_headers_c[0] == 0 && !memcmp(outer_headers_c,
1410                                                   outer_headers_c + 1,
1411                                                   size - 1);
1412 }
1413
1414 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1415                            union ib_flow_spec *ib_spec)
1416 {
1417         void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1418                                              outer_headers);
1419         void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1420                                              outer_headers);
1421         switch (ib_spec->type) {
1422         case IB_FLOW_SPEC_ETH:
1423                 if (ib_spec->size != sizeof(ib_spec->eth))
1424                         return -EINVAL;
1425
1426                 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1427                                              dmac_47_16),
1428                                 ib_spec->eth.mask.dst_mac);
1429                 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1430                                              dmac_47_16),
1431                                 ib_spec->eth.val.dst_mac);
1432
1433                 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1434                                              smac_47_16),
1435                                 ib_spec->eth.mask.src_mac);
1436                 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1437                                              smac_47_16),
1438                                 ib_spec->eth.val.src_mac);
1439
1440                 if (ib_spec->eth.mask.vlan_tag) {
1441                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1442                                  vlan_tag, 1);
1443                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1444                                  vlan_tag, 1);
1445
1446                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1447                                  first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1448                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1449                                  first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1450
1451                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1452                                  first_cfi,
1453                                  ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1454                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1455                                  first_cfi,
1456                                  ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1457
1458                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1459                                  first_prio,
1460                                  ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1461                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1462                                  first_prio,
1463                                  ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1464                 }
1465                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1466                          ethertype, ntohs(ib_spec->eth.mask.ether_type));
1467                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1468                          ethertype, ntohs(ib_spec->eth.val.ether_type));
1469                 break;
1470         case IB_FLOW_SPEC_IPV4:
1471                 if (ib_spec->size != sizeof(ib_spec->ipv4))
1472                         return -EINVAL;
1473
1474                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1475                          ethertype, 0xffff);
1476                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1477                          ethertype, ETH_P_IP);
1478
1479                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1480                                     src_ipv4_src_ipv6.ipv4_layout.ipv4),
1481                        &ib_spec->ipv4.mask.src_ip,
1482                        sizeof(ib_spec->ipv4.mask.src_ip));
1483                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1484                                     src_ipv4_src_ipv6.ipv4_layout.ipv4),
1485                        &ib_spec->ipv4.val.src_ip,
1486                        sizeof(ib_spec->ipv4.val.src_ip));
1487                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1488                                     dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1489                        &ib_spec->ipv4.mask.dst_ip,
1490                        sizeof(ib_spec->ipv4.mask.dst_ip));
1491                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1492                                     dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1493                        &ib_spec->ipv4.val.dst_ip,
1494                        sizeof(ib_spec->ipv4.val.dst_ip));
1495                 break;
1496         case IB_FLOW_SPEC_IPV6:
1497                 if (ib_spec->size != sizeof(ib_spec->ipv6))
1498                         return -EINVAL;
1499
1500                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1501                          ethertype, 0xffff);
1502                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1503                          ethertype, ETH_P_IPV6);
1504
1505                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1506                                     src_ipv4_src_ipv6.ipv6_layout.ipv6),
1507                        &ib_spec->ipv6.mask.src_ip,
1508                        sizeof(ib_spec->ipv6.mask.src_ip));
1509                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1510                                     src_ipv4_src_ipv6.ipv6_layout.ipv6),
1511                        &ib_spec->ipv6.val.src_ip,
1512                        sizeof(ib_spec->ipv6.val.src_ip));
1513                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1514                                     dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1515                        &ib_spec->ipv6.mask.dst_ip,
1516                        sizeof(ib_spec->ipv6.mask.dst_ip));
1517                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1518                                     dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1519                        &ib_spec->ipv6.val.dst_ip,
1520                        sizeof(ib_spec->ipv6.val.dst_ip));
1521                 break;
1522         case IB_FLOW_SPEC_TCP:
1523                 if (ib_spec->size != sizeof(ib_spec->tcp_udp))
1524                         return -EINVAL;
1525
1526                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1527                          0xff);
1528                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1529                          IPPROTO_TCP);
1530
1531                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport,
1532                          ntohs(ib_spec->tcp_udp.mask.src_port));
1533                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport,
1534                          ntohs(ib_spec->tcp_udp.val.src_port));
1535
1536                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport,
1537                          ntohs(ib_spec->tcp_udp.mask.dst_port));
1538                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport,
1539                          ntohs(ib_spec->tcp_udp.val.dst_port));
1540                 break;
1541         case IB_FLOW_SPEC_UDP:
1542                 if (ib_spec->size != sizeof(ib_spec->tcp_udp))
1543                         return -EINVAL;
1544
1545                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1546                          0xff);
1547                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1548                          IPPROTO_UDP);
1549
1550                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport,
1551                          ntohs(ib_spec->tcp_udp.mask.src_port));
1552                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport,
1553                          ntohs(ib_spec->tcp_udp.val.src_port));
1554
1555                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport,
1556                          ntohs(ib_spec->tcp_udp.mask.dst_port));
1557                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport,
1558                          ntohs(ib_spec->tcp_udp.val.dst_port));
1559                 break;
1560         default:
1561                 return -EINVAL;
1562         }
1563
1564         return 0;
1565 }
1566
1567 /* If a flow could catch both multicast and unicast packets,
1568  * it won't fall into the multicast flow steering table and this rule
1569  * could steal other multicast packets.
1570  */
1571 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1572 {
1573         struct ib_flow_spec_eth *eth_spec;
1574
1575         if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1576             ib_attr->size < sizeof(struct ib_flow_attr) +
1577             sizeof(struct ib_flow_spec_eth) ||
1578             ib_attr->num_of_specs < 1)
1579                 return false;
1580
1581         eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1582         if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1583             eth_spec->size != sizeof(*eth_spec))
1584                 return false;
1585
1586         return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1587                is_multicast_ether_addr(eth_spec->val.dst_mac);
1588 }
1589
1590 static bool is_valid_attr(struct ib_flow_attr *flow_attr)
1591 {
1592         union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1593         bool has_ipv4_spec = false;
1594         bool eth_type_ipv4 = true;
1595         unsigned int spec_index;
1596
1597         /* Validate that ethertype is correct */
1598         for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1599                 if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1600                     ib_spec->eth.mask.ether_type) {
1601                         if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1602                               ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1603                                 eth_type_ipv4 = false;
1604                 } else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1605                         has_ipv4_spec = true;
1606                 }
1607                 ib_spec = (void *)ib_spec + ib_spec->size;
1608         }
1609         return !has_ipv4_spec || eth_type_ipv4;
1610 }
1611
1612 static void put_flow_table(struct mlx5_ib_dev *dev,
1613                            struct mlx5_ib_flow_prio *prio, bool ft_added)
1614 {
1615         prio->refcount -= !!ft_added;
1616         if (!prio->refcount) {
1617                 mlx5_destroy_flow_table(prio->flow_table);
1618                 prio->flow_table = NULL;
1619         }
1620 }
1621
1622 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
1623 {
1624         struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
1625         struct mlx5_ib_flow_handler *handler = container_of(flow_id,
1626                                                           struct mlx5_ib_flow_handler,
1627                                                           ibflow);
1628         struct mlx5_ib_flow_handler *iter, *tmp;
1629
1630         mutex_lock(&dev->flow_db.lock);
1631
1632         list_for_each_entry_safe(iter, tmp, &handler->list, list) {
1633                 mlx5_del_flow_rule(iter->rule);
1634                 list_del(&iter->list);
1635                 kfree(iter);
1636         }
1637
1638         mlx5_del_flow_rule(handler->rule);
1639         put_flow_table(dev, &dev->flow_db.prios[handler->prio], true);
1640         mutex_unlock(&dev->flow_db.lock);
1641
1642         kfree(handler);
1643
1644         return 0;
1645 }
1646
1647 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
1648 {
1649         priority *= 2;
1650         if (!dont_trap)
1651                 priority++;
1652         return priority;
1653 }
1654
1655 #define MLX5_FS_MAX_TYPES        10
1656 #define MLX5_FS_MAX_ENTRIES      32000UL
1657 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
1658                                                 struct ib_flow_attr *flow_attr)
1659 {
1660         bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
1661         struct mlx5_flow_namespace *ns = NULL;
1662         struct mlx5_ib_flow_prio *prio;
1663         struct mlx5_flow_table *ft;
1664         int num_entries;
1665         int num_groups;
1666         int priority;
1667         int err = 0;
1668
1669         if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1670                 if (flow_is_multicast_only(flow_attr) &&
1671                     !dont_trap)
1672                         priority = MLX5_IB_FLOW_MCAST_PRIO;
1673                 else
1674                         priority = ib_prio_to_core_prio(flow_attr->priority,
1675                                                         dont_trap);
1676                 ns = mlx5_get_flow_namespace(dev->mdev,
1677                                              MLX5_FLOW_NAMESPACE_BYPASS);
1678                 num_entries = MLX5_FS_MAX_ENTRIES;
1679                 num_groups = MLX5_FS_MAX_TYPES;
1680                 prio = &dev->flow_db.prios[priority];
1681         } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1682                    flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1683                 ns = mlx5_get_flow_namespace(dev->mdev,
1684                                              MLX5_FLOW_NAMESPACE_LEFTOVERS);
1685                 build_leftovers_ft_param(&priority,
1686                                          &num_entries,
1687                                          &num_groups);
1688                 prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
1689         }
1690
1691         if (!ns)
1692                 return ERR_PTR(-ENOTSUPP);
1693
1694         ft = prio->flow_table;
1695         if (!ft) {
1696                 ft = mlx5_create_auto_grouped_flow_table(ns, priority,
1697                                                          num_entries,
1698                                                          num_groups,
1699                                                          0);
1700
1701                 if (!IS_ERR(ft)) {
1702                         prio->refcount = 0;
1703                         prio->flow_table = ft;
1704                 } else {
1705                         err = PTR_ERR(ft);
1706                 }
1707         }
1708
1709         return err ? ERR_PTR(err) : prio;
1710 }
1711
1712 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
1713                                                      struct mlx5_ib_flow_prio *ft_prio,
1714                                                      struct ib_flow_attr *flow_attr,
1715                                                      struct mlx5_flow_destination *dst)
1716 {
1717         struct mlx5_flow_table  *ft = ft_prio->flow_table;
1718         struct mlx5_ib_flow_handler *handler;
1719         struct mlx5_flow_spec *spec;
1720         void *ib_flow = flow_attr + 1;
1721         unsigned int spec_index;
1722         u32 action;
1723         int err = 0;
1724
1725         if (!is_valid_attr(flow_attr))
1726                 return ERR_PTR(-EINVAL);
1727
1728         spec = mlx5_vzalloc(sizeof(*spec));
1729         handler = kzalloc(sizeof(*handler), GFP_KERNEL);
1730         if (!handler || !spec) {
1731                 err = -ENOMEM;
1732                 goto free;
1733         }
1734
1735         INIT_LIST_HEAD(&handler->list);
1736
1737         for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1738                 err = parse_flow_attr(spec->match_criteria,
1739                                       spec->match_value, ib_flow);
1740                 if (err < 0)
1741                         goto free;
1742
1743                 ib_flow += ((union ib_flow_spec *)ib_flow)->size;
1744         }
1745
1746         /* Outer header support only */
1747         spec->match_criteria_enable = (!outer_header_zero(spec->match_criteria))
1748                 << 0;
1749         action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
1750                 MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
1751         handler->rule = mlx5_add_flow_rule(ft, spec,
1752                                            action,
1753                                            MLX5_FS_DEFAULT_FLOW_TAG,
1754                                            dst);
1755
1756         if (IS_ERR(handler->rule)) {
1757                 err = PTR_ERR(handler->rule);
1758                 goto free;
1759         }
1760
1761         handler->prio = ft_prio - dev->flow_db.prios;
1762
1763         ft_prio->flow_table = ft;
1764 free:
1765         if (err)
1766                 kfree(handler);
1767         kvfree(spec);
1768         return err ? ERR_PTR(err) : handler;
1769 }
1770
1771 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
1772                                                           struct mlx5_ib_flow_prio *ft_prio,
1773                                                           struct ib_flow_attr *flow_attr,
1774                                                           struct mlx5_flow_destination *dst)
1775 {
1776         struct mlx5_ib_flow_handler *handler_dst = NULL;
1777         struct mlx5_ib_flow_handler *handler = NULL;
1778
1779         handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
1780         if (!IS_ERR(handler)) {
1781                 handler_dst = create_flow_rule(dev, ft_prio,
1782                                                flow_attr, dst);
1783                 if (IS_ERR(handler_dst)) {
1784                         mlx5_del_flow_rule(handler->rule);
1785                         kfree(handler);
1786                         handler = handler_dst;
1787                 } else {
1788                         list_add(&handler_dst->list, &handler->list);
1789                 }
1790         }
1791
1792         return handler;
1793 }
1794 enum {
1795         LEFTOVERS_MC,
1796         LEFTOVERS_UC,
1797 };
1798
1799 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
1800                                                           struct mlx5_ib_flow_prio *ft_prio,
1801                                                           struct ib_flow_attr *flow_attr,
1802                                                           struct mlx5_flow_destination *dst)
1803 {
1804         struct mlx5_ib_flow_handler *handler_ucast = NULL;
1805         struct mlx5_ib_flow_handler *handler = NULL;
1806
1807         static struct {
1808                 struct ib_flow_attr     flow_attr;
1809                 struct ib_flow_spec_eth eth_flow;
1810         } leftovers_specs[] = {
1811                 [LEFTOVERS_MC] = {
1812                         .flow_attr = {
1813                                 .num_of_specs = 1,
1814                                 .size = sizeof(leftovers_specs[0])
1815                         },
1816                         .eth_flow = {
1817                                 .type = IB_FLOW_SPEC_ETH,
1818                                 .size = sizeof(struct ib_flow_spec_eth),
1819                                 .mask = {.dst_mac = {0x1} },
1820                                 .val =  {.dst_mac = {0x1} }
1821                         }
1822                 },
1823                 [LEFTOVERS_UC] = {
1824                         .flow_attr = {
1825                                 .num_of_specs = 1,
1826                                 .size = sizeof(leftovers_specs[0])
1827                         },
1828                         .eth_flow = {
1829                                 .type = IB_FLOW_SPEC_ETH,
1830                                 .size = sizeof(struct ib_flow_spec_eth),
1831                                 .mask = {.dst_mac = {0x1} },
1832                                 .val = {.dst_mac = {} }
1833                         }
1834                 }
1835         };
1836
1837         handler = create_flow_rule(dev, ft_prio,
1838                                    &leftovers_specs[LEFTOVERS_MC].flow_attr,
1839                                    dst);
1840         if (!IS_ERR(handler) &&
1841             flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
1842                 handler_ucast = create_flow_rule(dev, ft_prio,
1843                                                  &leftovers_specs[LEFTOVERS_UC].flow_attr,
1844                                                  dst);
1845                 if (IS_ERR(handler_ucast)) {
1846                         kfree(handler);
1847                         handler = handler_ucast;
1848                 } else {
1849                         list_add(&handler_ucast->list, &handler->list);
1850                 }
1851         }
1852
1853         return handler;
1854 }
1855
1856 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
1857                                            struct ib_flow_attr *flow_attr,
1858                                            int domain)
1859 {
1860         struct mlx5_ib_dev *dev = to_mdev(qp->device);
1861         struct mlx5_ib_qp *mqp = to_mqp(qp);
1862         struct mlx5_ib_flow_handler *handler = NULL;
1863         struct mlx5_flow_destination *dst = NULL;
1864         struct mlx5_ib_flow_prio *ft_prio;
1865         int err;
1866
1867         if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
1868                 return ERR_PTR(-ENOSPC);
1869
1870         if (domain != IB_FLOW_DOMAIN_USER ||
1871             flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
1872             (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
1873                 return ERR_PTR(-EINVAL);
1874
1875         dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1876         if (!dst)
1877                 return ERR_PTR(-ENOMEM);
1878
1879         mutex_lock(&dev->flow_db.lock);
1880
1881         ft_prio = get_flow_table(dev, flow_attr);
1882         if (IS_ERR(ft_prio)) {
1883                 err = PTR_ERR(ft_prio);
1884                 goto unlock;
1885         }
1886
1887         dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
1888         if (mqp->flags & MLX5_IB_QP_RSS)
1889                 dst->tir_num = mqp->rss_qp.tirn;
1890         else
1891                 dst->tir_num = mqp->raw_packet_qp.rq.tirn;
1892
1893         if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1894                 if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
1895                         handler = create_dont_trap_rule(dev, ft_prio,
1896                                                         flow_attr, dst);
1897                 } else {
1898                         handler = create_flow_rule(dev, ft_prio, flow_attr,
1899                                                    dst);
1900                 }
1901         } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1902                    flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1903                 handler = create_leftovers_rule(dev, ft_prio, flow_attr,
1904                                                 dst);
1905         } else {
1906                 err = -EINVAL;
1907                 goto destroy_ft;
1908         }
1909
1910         if (IS_ERR(handler)) {
1911                 err = PTR_ERR(handler);
1912                 handler = NULL;
1913                 goto destroy_ft;
1914         }
1915
1916         ft_prio->refcount++;
1917         mutex_unlock(&dev->flow_db.lock);
1918         kfree(dst);
1919
1920         return &handler->ibflow;
1921
1922 destroy_ft:
1923         put_flow_table(dev, ft_prio, false);
1924 unlock:
1925         mutex_unlock(&dev->flow_db.lock);
1926         kfree(dst);
1927         kfree(handler);
1928         return ERR_PTR(err);
1929 }
1930
1931 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
1932 {
1933         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1934         int err;
1935
1936         err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
1937         if (err)
1938                 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
1939                              ibqp->qp_num, gid->raw);
1940
1941         return err;
1942 }
1943
1944 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
1945 {
1946         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1947         int err;
1948
1949         err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
1950         if (err)
1951                 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
1952                              ibqp->qp_num, gid->raw);
1953
1954         return err;
1955 }
1956
1957 static int init_node_data(struct mlx5_ib_dev *dev)
1958 {
1959         int err;
1960
1961         err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
1962         if (err)
1963                 return err;
1964
1965         dev->mdev->rev_id = dev->mdev->pdev->revision;
1966
1967         return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
1968 }
1969
1970 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
1971                              char *buf)
1972 {
1973         struct mlx5_ib_dev *dev =
1974                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1975
1976         return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
1977 }
1978
1979 static ssize_t show_reg_pages(struct device *device,
1980                               struct device_attribute *attr, char *buf)
1981 {
1982         struct mlx5_ib_dev *dev =
1983                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1984
1985         return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
1986 }
1987
1988 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
1989                         char *buf)
1990 {
1991         struct mlx5_ib_dev *dev =
1992                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
1993         return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
1994 }
1995
1996 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
1997                         char *buf)
1998 {
1999         struct mlx5_ib_dev *dev =
2000                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2001         return sprintf(buf, "%x\n", dev->mdev->rev_id);
2002 }
2003
2004 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2005                           char *buf)
2006 {
2007         struct mlx5_ib_dev *dev =
2008                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2009         return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2010                        dev->mdev->board_id);
2011 }
2012
2013 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2014 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2015 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2016 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2017 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2018
2019 static struct device_attribute *mlx5_class_attributes[] = {
2020         &dev_attr_hw_rev,
2021         &dev_attr_hca_type,
2022         &dev_attr_board_id,
2023         &dev_attr_fw_pages,
2024         &dev_attr_reg_pages,
2025 };
2026
2027 static void pkey_change_handler(struct work_struct *work)
2028 {
2029         struct mlx5_ib_port_resources *ports =
2030                 container_of(work, struct mlx5_ib_port_resources,
2031                              pkey_change_work);
2032
2033         mutex_lock(&ports->devr->mutex);
2034         mlx5_ib_gsi_pkey_change(ports->gsi);
2035         mutex_unlock(&ports->devr->mutex);
2036 }
2037
2038 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2039 {
2040         struct mlx5_ib_qp *mqp;
2041         struct mlx5_ib_cq *send_mcq, *recv_mcq;
2042         struct mlx5_core_cq *mcq;
2043         struct list_head cq_armed_list;
2044         unsigned long flags_qp;
2045         unsigned long flags_cq;
2046         unsigned long flags;
2047
2048         INIT_LIST_HEAD(&cq_armed_list);
2049
2050         /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2051         spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2052         list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2053                 spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2054                 if (mqp->sq.tail != mqp->sq.head) {
2055                         send_mcq = to_mcq(mqp->ibqp.send_cq);
2056                         spin_lock_irqsave(&send_mcq->lock, flags_cq);
2057                         if (send_mcq->mcq.comp &&
2058                             mqp->ibqp.send_cq->comp_handler) {
2059                                 if (!send_mcq->mcq.reset_notify_added) {
2060                                         send_mcq->mcq.reset_notify_added = 1;
2061                                         list_add_tail(&send_mcq->mcq.reset_notify,
2062                                                       &cq_armed_list);
2063                                 }
2064                         }
2065                         spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2066                 }
2067                 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2068                 spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2069                 /* no handling is needed for SRQ */
2070                 if (!mqp->ibqp.srq) {
2071                         if (mqp->rq.tail != mqp->rq.head) {
2072                                 recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2073                                 spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2074                                 if (recv_mcq->mcq.comp &&
2075                                     mqp->ibqp.recv_cq->comp_handler) {
2076                                         if (!recv_mcq->mcq.reset_notify_added) {
2077                                                 recv_mcq->mcq.reset_notify_added = 1;
2078                                                 list_add_tail(&recv_mcq->mcq.reset_notify,
2079                                                               &cq_armed_list);
2080                                         }
2081                                 }
2082                                 spin_unlock_irqrestore(&recv_mcq->lock,
2083                                                        flags_cq);
2084                         }
2085                 }
2086                 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2087         }
2088         /*At that point all inflight post send were put to be executed as of we
2089          * lock/unlock above locks Now need to arm all involved CQs.
2090          */
2091         list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2092                 mcq->comp(mcq);
2093         }
2094         spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2095 }
2096
2097 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2098                           enum mlx5_dev_event event, unsigned long param)
2099 {
2100         struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2101         struct ib_event ibev;
2102
2103         u8 port = 0;
2104
2105         switch (event) {
2106         case MLX5_DEV_EVENT_SYS_ERROR:
2107                 ibdev->ib_active = false;
2108                 ibev.event = IB_EVENT_DEVICE_FATAL;
2109                 mlx5_ib_handle_internal_error(ibdev);
2110                 break;
2111
2112         case MLX5_DEV_EVENT_PORT_UP:
2113                 ibev.event = IB_EVENT_PORT_ACTIVE;
2114                 port = (u8)param;
2115                 break;
2116
2117         case MLX5_DEV_EVENT_PORT_DOWN:
2118         case MLX5_DEV_EVENT_PORT_INITIALIZED:
2119                 ibev.event = IB_EVENT_PORT_ERR;
2120                 port = (u8)param;
2121                 break;
2122
2123         case MLX5_DEV_EVENT_LID_CHANGE:
2124                 ibev.event = IB_EVENT_LID_CHANGE;
2125                 port = (u8)param;
2126                 break;
2127
2128         case MLX5_DEV_EVENT_PKEY_CHANGE:
2129                 ibev.event = IB_EVENT_PKEY_CHANGE;
2130                 port = (u8)param;
2131
2132                 schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2133                 break;
2134
2135         case MLX5_DEV_EVENT_GUID_CHANGE:
2136                 ibev.event = IB_EVENT_GID_CHANGE;
2137                 port = (u8)param;
2138                 break;
2139
2140         case MLX5_DEV_EVENT_CLIENT_REREG:
2141                 ibev.event = IB_EVENT_CLIENT_REREGISTER;
2142                 port = (u8)param;
2143                 break;
2144         }
2145
2146         ibev.device           = &ibdev->ib_dev;
2147         ibev.element.port_num = port;
2148
2149         if (port < 1 || port > ibdev->num_ports) {
2150                 mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2151                 return;
2152         }
2153
2154         if (ibdev->ib_active)
2155                 ib_dispatch_event(&ibev);
2156 }
2157
2158 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2159 {
2160         int port;
2161
2162         for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2163                 mlx5_query_ext_port_caps(dev, port);
2164 }
2165
2166 static int get_port_caps(struct mlx5_ib_dev *dev)
2167 {
2168         struct ib_device_attr *dprops = NULL;
2169         struct ib_port_attr *pprops = NULL;
2170         int err = -ENOMEM;
2171         int port;
2172         struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2173
2174         pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2175         if (!pprops)
2176                 goto out;
2177
2178         dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2179         if (!dprops)
2180                 goto out;
2181
2182         err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2183         if (err) {
2184                 mlx5_ib_warn(dev, "query_device failed %d\n", err);
2185                 goto out;
2186         }
2187
2188         for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2189                 err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2190                 if (err) {
2191                         mlx5_ib_warn(dev, "query_port %d failed %d\n",
2192                                      port, err);
2193                         break;
2194                 }
2195                 dev->mdev->port_caps[port - 1].pkey_table_len =
2196                                                 dprops->max_pkeys;
2197                 dev->mdev->port_caps[port - 1].gid_table_len =
2198                                                 pprops->gid_tbl_len;
2199                 mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2200                             dprops->max_pkeys, pprops->gid_tbl_len);
2201         }
2202
2203 out:
2204         kfree(pprops);
2205         kfree(dprops);
2206
2207         return err;
2208 }
2209
2210 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2211 {
2212         int err;
2213
2214         err = mlx5_mr_cache_cleanup(dev);
2215         if (err)
2216                 mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2217
2218         mlx5_ib_destroy_qp(dev->umrc.qp);
2219         ib_free_cq(dev->umrc.cq);
2220         ib_dealloc_pd(dev->umrc.pd);
2221 }
2222
2223 enum {
2224         MAX_UMR_WR = 128,
2225 };
2226
2227 static int create_umr_res(struct mlx5_ib_dev *dev)
2228 {
2229         struct ib_qp_init_attr *init_attr = NULL;
2230         struct ib_qp_attr *attr = NULL;
2231         struct ib_pd *pd;
2232         struct ib_cq *cq;
2233         struct ib_qp *qp;
2234         int ret;
2235
2236         attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2237         init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2238         if (!attr || !init_attr) {
2239                 ret = -ENOMEM;
2240                 goto error_0;
2241         }
2242
2243         pd = ib_alloc_pd(&dev->ib_dev);
2244         if (IS_ERR(pd)) {
2245                 mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2246                 ret = PTR_ERR(pd);
2247                 goto error_0;
2248         }
2249
2250         cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2251         if (IS_ERR(cq)) {
2252                 mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2253                 ret = PTR_ERR(cq);
2254                 goto error_2;
2255         }
2256
2257         init_attr->send_cq = cq;
2258         init_attr->recv_cq = cq;
2259         init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2260         init_attr->cap.max_send_wr = MAX_UMR_WR;
2261         init_attr->cap.max_send_sge = 1;
2262         init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2263         init_attr->port_num = 1;
2264         qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2265         if (IS_ERR(qp)) {
2266                 mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2267                 ret = PTR_ERR(qp);
2268                 goto error_3;
2269         }
2270         qp->device     = &dev->ib_dev;
2271         qp->real_qp    = qp;
2272         qp->uobject    = NULL;
2273         qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2274
2275         attr->qp_state = IB_QPS_INIT;
2276         attr->port_num = 1;
2277         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2278                                 IB_QP_PORT, NULL);
2279         if (ret) {
2280                 mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2281                 goto error_4;
2282         }
2283
2284         memset(attr, 0, sizeof(*attr));
2285         attr->qp_state = IB_QPS_RTR;
2286         attr->path_mtu = IB_MTU_256;
2287
2288         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2289         if (ret) {
2290                 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2291                 goto error_4;
2292         }
2293
2294         memset(attr, 0, sizeof(*attr));
2295         attr->qp_state = IB_QPS_RTS;
2296         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2297         if (ret) {
2298                 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2299                 goto error_4;
2300         }
2301
2302         dev->umrc.qp = qp;
2303         dev->umrc.cq = cq;
2304         dev->umrc.pd = pd;
2305
2306         sema_init(&dev->umrc.sem, MAX_UMR_WR);
2307         ret = mlx5_mr_cache_init(dev);
2308         if (ret) {
2309                 mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2310                 goto error_4;
2311         }
2312
2313         kfree(attr);
2314         kfree(init_attr);
2315
2316         return 0;
2317
2318 error_4:
2319         mlx5_ib_destroy_qp(qp);
2320
2321 error_3:
2322         ib_free_cq(cq);
2323
2324 error_2:
2325         ib_dealloc_pd(pd);
2326
2327 error_0:
2328         kfree(attr);
2329         kfree(init_attr);
2330         return ret;
2331 }
2332
2333 static int create_dev_resources(struct mlx5_ib_resources *devr)
2334 {
2335         struct ib_srq_init_attr attr;
2336         struct mlx5_ib_dev *dev;
2337         struct ib_cq_init_attr cq_attr = {.cqe = 1};
2338         int port;
2339         int ret = 0;
2340
2341         dev = container_of(devr, struct mlx5_ib_dev, devr);
2342
2343         mutex_init(&devr->mutex);
2344
2345         devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2346         if (IS_ERR(devr->p0)) {
2347                 ret = PTR_ERR(devr->p0);
2348                 goto error0;
2349         }
2350         devr->p0->device  = &dev->ib_dev;
2351         devr->p0->uobject = NULL;
2352         atomic_set(&devr->p0->usecnt, 0);
2353
2354         devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
2355         if (IS_ERR(devr->c0)) {
2356                 ret = PTR_ERR(devr->c0);
2357                 goto error1;
2358         }
2359         devr->c0->device        = &dev->ib_dev;
2360         devr->c0->uobject       = NULL;
2361         devr->c0->comp_handler  = NULL;
2362         devr->c0->event_handler = NULL;
2363         devr->c0->cq_context    = NULL;
2364         atomic_set(&devr->c0->usecnt, 0);
2365
2366         devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2367         if (IS_ERR(devr->x0)) {
2368                 ret = PTR_ERR(devr->x0);
2369                 goto error2;
2370         }
2371         devr->x0->device = &dev->ib_dev;
2372         devr->x0->inode = NULL;
2373         atomic_set(&devr->x0->usecnt, 0);
2374         mutex_init(&devr->x0->tgt_qp_mutex);
2375         INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2376
2377         devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2378         if (IS_ERR(devr->x1)) {
2379                 ret = PTR_ERR(devr->x1);
2380                 goto error3;
2381         }
2382         devr->x1->device = &dev->ib_dev;
2383         devr->x1->inode = NULL;
2384         atomic_set(&devr->x1->usecnt, 0);
2385         mutex_init(&devr->x1->tgt_qp_mutex);
2386         INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2387
2388         memset(&attr, 0, sizeof(attr));
2389         attr.attr.max_sge = 1;
2390         attr.attr.max_wr = 1;
2391         attr.srq_type = IB_SRQT_XRC;
2392         attr.ext.xrc.cq = devr->c0;
2393         attr.ext.xrc.xrcd = devr->x0;
2394
2395         devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2396         if (IS_ERR(devr->s0)) {
2397                 ret = PTR_ERR(devr->s0);
2398                 goto error4;
2399         }
2400         devr->s0->device        = &dev->ib_dev;
2401         devr->s0->pd            = devr->p0;
2402         devr->s0->uobject       = NULL;
2403         devr->s0->event_handler = NULL;
2404         devr->s0->srq_context   = NULL;
2405         devr->s0->srq_type      = IB_SRQT_XRC;
2406         devr->s0->ext.xrc.xrcd  = devr->x0;
2407         devr->s0->ext.xrc.cq    = devr->c0;
2408         atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2409         atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2410         atomic_inc(&devr->p0->usecnt);
2411         atomic_set(&devr->s0->usecnt, 0);
2412
2413         memset(&attr, 0, sizeof(attr));
2414         attr.attr.max_sge = 1;
2415         attr.attr.max_wr = 1;
2416         attr.srq_type = IB_SRQT_BASIC;
2417         devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2418         if (IS_ERR(devr->s1)) {
2419                 ret = PTR_ERR(devr->s1);
2420                 goto error5;
2421         }
2422         devr->s1->device        = &dev->ib_dev;
2423         devr->s1->pd            = devr->p0;
2424         devr->s1->uobject       = NULL;
2425         devr->s1->event_handler = NULL;
2426         devr->s1->srq_context   = NULL;
2427         devr->s1->srq_type      = IB_SRQT_BASIC;
2428         devr->s1->ext.xrc.cq    = devr->c0;
2429         atomic_inc(&devr->p0->usecnt);
2430         atomic_set(&devr->s0->usecnt, 0);
2431
2432         for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2433                 INIT_WORK(&devr->ports[port].pkey_change_work,
2434                           pkey_change_handler);
2435                 devr->ports[port].devr = devr;
2436         }
2437
2438         return 0;
2439
2440 error5:
2441         mlx5_ib_destroy_srq(devr->s0);
2442 error4:
2443         mlx5_ib_dealloc_xrcd(devr->x1);
2444 error3:
2445         mlx5_ib_dealloc_xrcd(devr->x0);
2446 error2:
2447         mlx5_ib_destroy_cq(devr->c0);
2448 error1:
2449         mlx5_ib_dealloc_pd(devr->p0);
2450 error0:
2451         return ret;
2452 }
2453
2454 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2455 {
2456         struct mlx5_ib_dev *dev =
2457                 container_of(devr, struct mlx5_ib_dev, devr);
2458         int port;
2459
2460         mlx5_ib_destroy_srq(devr->s1);
2461         mlx5_ib_destroy_srq(devr->s0);
2462         mlx5_ib_dealloc_xrcd(devr->x0);
2463         mlx5_ib_dealloc_xrcd(devr->x1);
2464         mlx5_ib_destroy_cq(devr->c0);
2465         mlx5_ib_dealloc_pd(devr->p0);
2466
2467         /* Make sure no change P_Key work items are still executing */
2468         for (port = 0; port < dev->num_ports; ++port)
2469                 cancel_work_sync(&devr->ports[port].pkey_change_work);
2470 }
2471
2472 static u32 get_core_cap_flags(struct ib_device *ibdev)
2473 {
2474         struct mlx5_ib_dev *dev = to_mdev(ibdev);
2475         enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2476         u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2477         u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2478         u32 ret = 0;
2479
2480         if (ll == IB_LINK_LAYER_INFINIBAND)
2481                 return RDMA_CORE_PORT_IBA_IB;
2482
2483         if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2484                 return 0;
2485
2486         if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2487                 return 0;
2488
2489         if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2490                 ret |= RDMA_CORE_PORT_IBA_ROCE;
2491
2492         if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2493                 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2494
2495         return ret;
2496 }
2497
2498 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
2499                                struct ib_port_immutable *immutable)
2500 {
2501         struct ib_port_attr attr;
2502         int err;
2503
2504         err = mlx5_ib_query_port(ibdev, port_num, &attr);
2505         if (err)
2506                 return err;
2507
2508         immutable->pkey_tbl_len = attr.pkey_tbl_len;
2509         immutable->gid_tbl_len = attr.gid_tbl_len;
2510         immutable->core_cap_flags = get_core_cap_flags(ibdev);
2511         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2512
2513         return 0;
2514 }
2515
2516 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
2517                            size_t str_len)
2518 {
2519         struct mlx5_ib_dev *dev =
2520                 container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2521         snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
2522                        fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
2523 }
2524
2525 static int mlx5_enable_roce(struct mlx5_ib_dev *dev)
2526 {
2527         int err;
2528
2529         dev->roce.nb.notifier_call = mlx5_netdev_event;
2530         err = register_netdevice_notifier(&dev->roce.nb);
2531         if (err)
2532                 return err;
2533
2534         err = mlx5_nic_vport_enable_roce(dev->mdev);
2535         if (err)
2536                 goto err_unregister_netdevice_notifier;
2537
2538         return 0;
2539
2540 err_unregister_netdevice_notifier:
2541         unregister_netdevice_notifier(&dev->roce.nb);
2542         return err;
2543 }
2544
2545 static void mlx5_disable_roce(struct mlx5_ib_dev *dev)
2546 {
2547         mlx5_nic_vport_disable_roce(dev->mdev);
2548         unregister_netdevice_notifier(&dev->roce.nb);
2549 }
2550
2551 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
2552 {
2553         unsigned int i;
2554
2555         for (i = 0; i < dev->num_ports; i++)
2556                 mlx5_core_dealloc_q_counter(dev->mdev,
2557                                             dev->port[i].q_cnt_id);
2558 }
2559
2560 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
2561 {
2562         int i;
2563         int ret;
2564
2565         for (i = 0; i < dev->num_ports; i++) {
2566                 ret = mlx5_core_alloc_q_counter(dev->mdev,
2567                                                 &dev->port[i].q_cnt_id);
2568                 if (ret) {
2569                         mlx5_ib_warn(dev,
2570                                      "couldn't allocate queue counter for port %d, err %d\n",
2571                                      i + 1, ret);
2572                         goto dealloc_counters;
2573                 }
2574         }
2575
2576         return 0;
2577
2578 dealloc_counters:
2579         while (--i >= 0)
2580                 mlx5_core_dealloc_q_counter(dev->mdev,
2581                                             dev->port[i].q_cnt_id);
2582
2583         return ret;
2584 }
2585
2586 static const char * const names[] = {
2587         "rx_write_requests",
2588         "rx_read_requests",
2589         "rx_atomic_requests",
2590         "out_of_buffer",
2591         "out_of_sequence",
2592         "duplicate_request",
2593         "rnr_nak_retry_err",
2594         "packet_seq_err",
2595         "implied_nak_seq_err",
2596         "local_ack_timeout_err",
2597 };
2598
2599 static const size_t stats_offsets[] = {
2600         MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests),
2601         MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests),
2602         MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests),
2603         MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer),
2604         MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence),
2605         MLX5_BYTE_OFF(query_q_counter_out, duplicate_request),
2606         MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err),
2607         MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err),
2608         MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err),
2609         MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err),
2610 };
2611
2612 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
2613                                                     u8 port_num)
2614 {
2615         BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets));
2616
2617         /* We support only per port stats */
2618         if (port_num == 0)
2619                 return NULL;
2620
2621         return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names),
2622                                           RDMA_HW_STATS_DEFAULT_LIFESPAN);
2623 }
2624
2625 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
2626                                 struct rdma_hw_stats *stats,
2627                                 u8 port, int index)
2628 {
2629         struct mlx5_ib_dev *dev = to_mdev(ibdev);
2630         int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
2631         void *out;
2632         __be32 val;
2633         int ret;
2634         int i;
2635
2636         if (!port || !stats)
2637                 return -ENOSYS;
2638
2639         out = mlx5_vzalloc(outlen);
2640         if (!out)
2641                 return -ENOMEM;
2642
2643         ret = mlx5_core_query_q_counter(dev->mdev,
2644                                         dev->port[port - 1].q_cnt_id, 0,
2645                                         out, outlen);
2646         if (ret)
2647                 goto free;
2648
2649         for (i = 0; i < ARRAY_SIZE(names); i++) {
2650                 val = *(__be32 *)(out + stats_offsets[i]);
2651                 stats->value[i] = (u64)be32_to_cpu(val);
2652         }
2653 free:
2654         kvfree(out);
2655         return ARRAY_SIZE(names);
2656 }
2657
2658 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
2659 {
2660         struct mlx5_ib_dev *dev;
2661         enum rdma_link_layer ll;
2662         int port_type_cap;
2663         int err;
2664         int i;
2665
2666         port_type_cap = MLX5_CAP_GEN(mdev, port_type);
2667         ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
2668
2669         if ((ll == IB_LINK_LAYER_ETHERNET) && !MLX5_CAP_GEN(mdev, roce))
2670                 return NULL;
2671
2672         printk_once(KERN_INFO "%s", mlx5_version);
2673
2674         dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
2675         if (!dev)
2676                 return NULL;
2677
2678         dev->mdev = mdev;
2679
2680         dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
2681                             GFP_KERNEL);
2682         if (!dev->port)
2683                 goto err_dealloc;
2684
2685         rwlock_init(&dev->roce.netdev_lock);
2686         err = get_port_caps(dev);
2687         if (err)
2688                 goto err_free_port;
2689
2690         if (mlx5_use_mad_ifc(dev))
2691                 get_ext_port_caps(dev);
2692
2693         MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
2694
2695         strlcpy(dev->ib_dev.name, "mlx5_%d", IB_DEVICE_NAME_MAX);
2696         dev->ib_dev.owner               = THIS_MODULE;
2697         dev->ib_dev.node_type           = RDMA_NODE_IB_CA;
2698         dev->ib_dev.local_dma_lkey      = 0 /* not supported for now */;
2699         dev->num_ports          = MLX5_CAP_GEN(mdev, num_ports);
2700         dev->ib_dev.phys_port_cnt     = dev->num_ports;
2701         dev->ib_dev.num_comp_vectors    =
2702                 dev->mdev->priv.eq_table.num_comp_vectors;
2703         dev->ib_dev.dma_device  = &mdev->pdev->dev;
2704
2705         dev->ib_dev.uverbs_abi_ver      = MLX5_IB_UVERBS_ABI_VERSION;
2706         dev->ib_dev.uverbs_cmd_mask     =
2707                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
2708                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
2709                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
2710                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
2711                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
2712                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
2713                 (1ull << IB_USER_VERBS_CMD_REREG_MR)            |
2714                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
2715                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2716                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
2717                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
2718                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
2719                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
2720                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
2721                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
2722                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
2723                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
2724                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
2725                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
2726                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
2727                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
2728                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
2729                 (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)         |
2730                 (1ull << IB_USER_VERBS_CMD_OPEN_QP);
2731         dev->ib_dev.uverbs_ex_cmd_mask =
2732                 (1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)     |
2733                 (1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)        |
2734                 (1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
2735
2736         dev->ib_dev.query_device        = mlx5_ib_query_device;
2737         dev->ib_dev.query_port          = mlx5_ib_query_port;
2738         dev->ib_dev.get_link_layer      = mlx5_ib_port_link_layer;
2739         if (ll == IB_LINK_LAYER_ETHERNET)
2740                 dev->ib_dev.get_netdev  = mlx5_ib_get_netdev;
2741         dev->ib_dev.query_gid           = mlx5_ib_query_gid;
2742         dev->ib_dev.add_gid             = mlx5_ib_add_gid;
2743         dev->ib_dev.del_gid             = mlx5_ib_del_gid;
2744         dev->ib_dev.query_pkey          = mlx5_ib_query_pkey;
2745         dev->ib_dev.modify_device       = mlx5_ib_modify_device;
2746         dev->ib_dev.modify_port         = mlx5_ib_modify_port;
2747         dev->ib_dev.alloc_ucontext      = mlx5_ib_alloc_ucontext;
2748         dev->ib_dev.dealloc_ucontext    = mlx5_ib_dealloc_ucontext;
2749         dev->ib_dev.mmap                = mlx5_ib_mmap;
2750         dev->ib_dev.alloc_pd            = mlx5_ib_alloc_pd;
2751         dev->ib_dev.dealloc_pd          = mlx5_ib_dealloc_pd;
2752         dev->ib_dev.create_ah           = mlx5_ib_create_ah;
2753         dev->ib_dev.query_ah            = mlx5_ib_query_ah;
2754         dev->ib_dev.destroy_ah          = mlx5_ib_destroy_ah;
2755         dev->ib_dev.create_srq          = mlx5_ib_create_srq;
2756         dev->ib_dev.modify_srq          = mlx5_ib_modify_srq;
2757         dev->ib_dev.query_srq           = mlx5_ib_query_srq;
2758         dev->ib_dev.destroy_srq         = mlx5_ib_destroy_srq;
2759         dev->ib_dev.post_srq_recv       = mlx5_ib_post_srq_recv;
2760         dev->ib_dev.create_qp           = mlx5_ib_create_qp;
2761         dev->ib_dev.modify_qp           = mlx5_ib_modify_qp;
2762         dev->ib_dev.query_qp            = mlx5_ib_query_qp;
2763         dev->ib_dev.destroy_qp          = mlx5_ib_destroy_qp;
2764         dev->ib_dev.post_send           = mlx5_ib_post_send;
2765         dev->ib_dev.post_recv           = mlx5_ib_post_recv;
2766         dev->ib_dev.create_cq           = mlx5_ib_create_cq;
2767         dev->ib_dev.modify_cq           = mlx5_ib_modify_cq;
2768         dev->ib_dev.resize_cq           = mlx5_ib_resize_cq;
2769         dev->ib_dev.destroy_cq          = mlx5_ib_destroy_cq;
2770         dev->ib_dev.poll_cq             = mlx5_ib_poll_cq;
2771         dev->ib_dev.req_notify_cq       = mlx5_ib_arm_cq;
2772         dev->ib_dev.get_dma_mr          = mlx5_ib_get_dma_mr;
2773         dev->ib_dev.reg_user_mr         = mlx5_ib_reg_user_mr;
2774         dev->ib_dev.rereg_user_mr       = mlx5_ib_rereg_user_mr;
2775         dev->ib_dev.dereg_mr            = mlx5_ib_dereg_mr;
2776         dev->ib_dev.attach_mcast        = mlx5_ib_mcg_attach;
2777         dev->ib_dev.detach_mcast        = mlx5_ib_mcg_detach;
2778         dev->ib_dev.process_mad         = mlx5_ib_process_mad;
2779         dev->ib_dev.alloc_mr            = mlx5_ib_alloc_mr;
2780         dev->ib_dev.map_mr_sg           = mlx5_ib_map_mr_sg;
2781         dev->ib_dev.check_mr_status     = mlx5_ib_check_mr_status;
2782         dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
2783         dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
2784         if (mlx5_core_is_pf(mdev)) {
2785                 dev->ib_dev.get_vf_config       = mlx5_ib_get_vf_config;
2786                 dev->ib_dev.set_vf_link_state   = mlx5_ib_set_vf_link_state;
2787                 dev->ib_dev.get_vf_stats        = mlx5_ib_get_vf_stats;
2788                 dev->ib_dev.set_vf_guid         = mlx5_ib_set_vf_guid;
2789         }
2790
2791         dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
2792
2793         mlx5_ib_internal_fill_odp_caps(dev);
2794
2795         if (MLX5_CAP_GEN(mdev, imaicl)) {
2796                 dev->ib_dev.alloc_mw            = mlx5_ib_alloc_mw;
2797                 dev->ib_dev.dealloc_mw          = mlx5_ib_dealloc_mw;
2798                 dev->ib_dev.uverbs_cmd_mask |=
2799                         (1ull << IB_USER_VERBS_CMD_ALLOC_MW)    |
2800                         (1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
2801         }
2802
2803         if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) &&
2804             MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
2805                 dev->ib_dev.get_hw_stats        = mlx5_ib_get_hw_stats;
2806                 dev->ib_dev.alloc_hw_stats      = mlx5_ib_alloc_hw_stats;
2807         }
2808
2809         if (MLX5_CAP_GEN(mdev, xrc)) {
2810                 dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
2811                 dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
2812                 dev->ib_dev.uverbs_cmd_mask |=
2813                         (1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
2814                         (1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
2815         }
2816
2817         if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
2818             IB_LINK_LAYER_ETHERNET) {
2819                 dev->ib_dev.create_flow = mlx5_ib_create_flow;
2820                 dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
2821                 dev->ib_dev.create_wq    = mlx5_ib_create_wq;
2822                 dev->ib_dev.modify_wq    = mlx5_ib_modify_wq;
2823                 dev->ib_dev.destroy_wq   = mlx5_ib_destroy_wq;
2824                 dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
2825                 dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
2826                 dev->ib_dev.uverbs_ex_cmd_mask |=
2827                         (1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
2828                         (1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
2829                         (1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
2830                         (1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
2831                         (1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
2832                         (1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
2833                         (1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
2834         }
2835         err = init_node_data(dev);
2836         if (err)
2837                 goto err_dealloc;
2838
2839         mutex_init(&dev->flow_db.lock);
2840         mutex_init(&dev->cap_mask_mutex);
2841         INIT_LIST_HEAD(&dev->qp_list);
2842         spin_lock_init(&dev->reset_flow_resource_lock);
2843
2844         if (ll == IB_LINK_LAYER_ETHERNET) {
2845                 err = mlx5_enable_roce(dev);
2846                 if (err)
2847                         goto err_dealloc;
2848         }
2849
2850         err = create_dev_resources(&dev->devr);
2851         if (err)
2852                 goto err_disable_roce;
2853
2854         err = mlx5_ib_odp_init_one(dev);
2855         if (err)
2856                 goto err_rsrc;
2857
2858         err = mlx5_ib_alloc_q_counters(dev);
2859         if (err)
2860                 goto err_odp;
2861
2862         err = ib_register_device(&dev->ib_dev, NULL);
2863         if (err)
2864                 goto err_q_cnt;
2865
2866         err = create_umr_res(dev);
2867         if (err)
2868                 goto err_dev;
2869
2870         for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
2871                 err = device_create_file(&dev->ib_dev.dev,
2872                                          mlx5_class_attributes[i]);
2873                 if (err)
2874                         goto err_umrc;
2875         }
2876
2877         dev->ib_active = true;
2878
2879         return dev;
2880
2881 err_umrc:
2882         destroy_umrc_res(dev);
2883
2884 err_dev:
2885         ib_unregister_device(&dev->ib_dev);
2886
2887 err_q_cnt:
2888         mlx5_ib_dealloc_q_counters(dev);
2889
2890 err_odp:
2891         mlx5_ib_odp_remove_one(dev);
2892
2893 err_rsrc:
2894         destroy_dev_resources(&dev->devr);
2895
2896 err_disable_roce:
2897         if (ll == IB_LINK_LAYER_ETHERNET)
2898                 mlx5_disable_roce(dev);
2899
2900 err_free_port:
2901         kfree(dev->port);
2902
2903 err_dealloc:
2904         ib_dealloc_device((struct ib_device *)dev);
2905
2906         return NULL;
2907 }
2908
2909 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
2910 {
2911         struct mlx5_ib_dev *dev = context;
2912         enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
2913
2914         ib_unregister_device(&dev->ib_dev);
2915         mlx5_ib_dealloc_q_counters(dev);
2916         destroy_umrc_res(dev);
2917         mlx5_ib_odp_remove_one(dev);
2918         destroy_dev_resources(&dev->devr);
2919         if (ll == IB_LINK_LAYER_ETHERNET)
2920                 mlx5_disable_roce(dev);
2921         kfree(dev->port);
2922         ib_dealloc_device(&dev->ib_dev);
2923 }
2924
2925 static struct mlx5_interface mlx5_ib_interface = {
2926         .add            = mlx5_ib_add,
2927         .remove         = mlx5_ib_remove,
2928         .event          = mlx5_ib_event,
2929         .protocol       = MLX5_INTERFACE_PROTOCOL_IB,
2930 };
2931
2932 static int __init mlx5_ib_init(void)
2933 {
2934         int err;
2935
2936         if (deprecated_prof_sel != 2)
2937                 pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
2938
2939         err = mlx5_ib_odp_init();
2940         if (err)
2941                 return err;
2942
2943         err = mlx5_register_interface(&mlx5_ib_interface);
2944         if (err)
2945                 goto clean_odp;
2946
2947         return err;
2948
2949 clean_odp:
2950         mlx5_ib_odp_cleanup();
2951         return err;
2952 }
2953
2954 static void __exit mlx5_ib_cleanup(void)
2955 {
2956         mlx5_unregister_interface(&mlx5_ib_interface);
2957         mlx5_ib_odp_cleanup();
2958 }
2959
2960 module_init(mlx5_ib_init);
2961 module_exit(mlx5_ib_cleanup);