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 #include <linux/io-mapping.h>
  41 #if defined(CONFIG_X86)
  42 #include <asm/pat.h>
  43 #endif
  44 #include <linux/sched.h>
  45 #include <linux/delay.h>
  46 #include <rdma/ib_user_verbs.h>
  47 #include <rdma/ib_addr.h>
  48 #include <rdma/ib_cache.h>
  49 #include <linux/mlx5/port.h>
  50 #include <linux/mlx5/vport.h>
  51 #include <linux/list.h>
  52 #include <rdma/ib_smi.h>
  53 #include <rdma/ib_umem.h>
  54 #include <linux/in.h>
  55 #include <linux/etherdevice.h>
  56 #include <linux/mlx5/fs.h>
  57 #include "user.h"
  58 #include "mlx5_ib.h"
  59
  60 #define DRIVER_NAME "mlx5_ib"
  61 #define DRIVER_VERSION "2.2-1"
  62 #define DRIVER_RELDATE  "Feb 2014"
  63
  64 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
  65 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
  66 MODULE_LICENSE("Dual BSD/GPL");
  67 MODULE_VERSION(DRIVER_VERSION);
  68
  69 static int deprecated_prof_sel = 2;
  70 module_param_named(prof_sel, deprecated_prof_sel, int, 0444);
  71 MODULE_PARM_DESC(prof_sel, "profile selector. Deprecated here. Moved to module mlx5_core");
  72
  73 static char mlx5_version[] =
  74         DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
  75         DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
  76
  77 enum {
  78         MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
  79 };
  80
  81 static enum rdma_link_layer
  82 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
  83 {
  84         switch (port_type_cap) {
  85         case MLX5_CAP_PORT_TYPE_IB:
  86                 return IB_LINK_LAYER_INFINIBAND;
  87         case MLX5_CAP_PORT_TYPE_ETH:
  88                 return IB_LINK_LAYER_ETHERNET;
  89         default:
  90                 return IB_LINK_LAYER_UNSPECIFIED;
  91         }
  92 }
  93
  94 static enum rdma_link_layer
  95 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
  96 {
  97         struct mlx5_ib_dev *dev = to_mdev(device);
  98         int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
  99
 100         return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
 101 }
 102
 103 static int mlx5_netdev_event(struct notifier_block *this,
 104                              unsigned long event, void *ptr)
 105 {
 106         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
 107         struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
 108                                                  roce.nb);
 109
 110         if ((event != NETDEV_UNREGISTER) && (event != NETDEV_REGISTER))
 111                 return NOTIFY_DONE;
 112
 113         write_lock(&ibdev->roce.netdev_lock);
 114         if (ndev->dev.parent == &ibdev->mdev->pdev->dev)
 115                 ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ? NULL : ndev;
 116         write_unlock(&ibdev->roce.netdev_lock);
 117
 118         return NOTIFY_DONE;
 119 }
 120
 121 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
 122                                              u8 port_num)
 123 {
 124         struct mlx5_ib_dev *ibdev = to_mdev(device);
 125         struct net_device *ndev;
 126
 127         /* Ensure ndev does not disappear before we invoke dev_hold()
 128          */
 129         read_lock(&ibdev->roce.netdev_lock);
 130         ndev = ibdev->roce.netdev;
 131         if (ndev)
 132                 dev_hold(ndev);
 133         read_unlock(&ibdev->roce.netdev_lock);
 134
 135         return ndev;
 136 }
 137
 138 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
 139                                 struct ib_port_attr *props)
 140 {
 141         struct mlx5_ib_dev *dev = to_mdev(device);
 142         struct net_device *ndev;
 143         enum ib_mtu ndev_ib_mtu;
 144         u16 qkey_viol_cntr;
 145
 146         memset(props, 0, sizeof(*props));
 147
 148         props->port_cap_flags  |= IB_PORT_CM_SUP;
 149         props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
 150
 151         props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
 152                                                 roce_address_table_size);
 153         props->max_mtu          = IB_MTU_4096;
 154         props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
 155         props->pkey_tbl_len     = 1;
 156         props->state            = IB_PORT_DOWN;
 157         props->phys_state       = 3;
 158
 159         mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
 160         props->qkey_viol_cntr = qkey_viol_cntr;
 161
 162         ndev = mlx5_ib_get_netdev(device, port_num);
 163         if (!ndev)
 164                 return 0;
 165
 166         if (netif_running(ndev) && netif_carrier_ok(ndev)) {
 167                 props->state      = IB_PORT_ACTIVE;
 168                 props->phys_state = 5;
 169         }
 170
 171         ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
 172
 173         dev_put(ndev);
 174
 175         props->active_mtu       = min(props->max_mtu, ndev_ib_mtu);
 176
 177         props->active_width     = IB_WIDTH_4X;  /* TODO */
 178         props->active_speed     = IB_SPEED_QDR; /* TODO */
 179
 180         return 0;
 181 }
 182
 183 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
 184                                      const struct ib_gid_attr *attr,
 185                                      void *mlx5_addr)
 186 {
 187 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
 188         char *mlx5_addr_l3_addr = MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
 189                                                source_l3_address);
 190         void *mlx5_addr_mac     = MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
 191                                                source_mac_47_32);
 192
 193         if (!gid)
 194                 return;
 195
 196         ether_addr_copy(mlx5_addr_mac, attr->ndev->dev_addr);
 197
 198         if (is_vlan_dev(attr->ndev)) {
 199                 MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
 200                 MLX5_SET_RA(mlx5_addr, vlan_id, vlan_dev_vlan_id(attr->ndev));
 201         }
 202
 203         switch (attr->gid_type) {
 204         case IB_GID_TYPE_IB:
 205                 MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
 206                 break;
 207         case IB_GID_TYPE_ROCE_UDP_ENCAP:
 208                 MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
 209                 break;
 210
 211         default:
 212                 WARN_ON(true);
 213         }
 214
 215         if (attr->gid_type != IB_GID_TYPE_IB) {
 216                 if (ipv6_addr_v4mapped((void *)gid))
 217                         MLX5_SET_RA(mlx5_addr, roce_l3_type,
 218                                     MLX5_ROCE_L3_TYPE_IPV4);
 219                 else
 220                         MLX5_SET_RA(mlx5_addr, roce_l3_type,
 221                                     MLX5_ROCE_L3_TYPE_IPV6);
 222         }
 223
 224         if ((attr->gid_type == IB_GID_TYPE_IB) ||
 225             !ipv6_addr_v4mapped((void *)gid))
 226                 memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
 227         else
 228                 memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
 229 }
 230
 231 static int set_roce_addr(struct ib_device *device, u8 port_num,
 232                          unsigned int index,
 233                          const union ib_gid *gid,
 234                          const struct ib_gid_attr *attr)
 235 {
 236         struct mlx5_ib_dev *dev = to_mdev(device);
 237         u32  in[MLX5_ST_SZ_DW(set_roce_address_in)]  = {0};
 238         u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0};
 239         void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
 240         enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
 241
 242         if (ll != IB_LINK_LAYER_ETHERNET)
 243                 return -EINVAL;
 244
 245         ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
 246
 247         MLX5_SET(set_roce_address_in, in, roce_address_index, index);
 248         MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
 249         return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
 250 }
 251
 252 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
 253                            unsigned int index, const union ib_gid *gid,
 254                            const struct ib_gid_attr *attr,
 255                            __always_unused void **context)
 256 {
 257         return set_roce_addr(device, port_num, index, gid, attr);
 258 }
 259
 260 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
 261                            unsigned int index, __always_unused void **context)
 262 {
 263         return set_roce_addr(device, port_num, index, NULL, NULL);
 264 }
 265
 266 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
 267                                int index)
 268 {
 269         struct ib_gid_attr attr;
 270         union ib_gid gid;
 271
 272         if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
 273                 return 0;
 274
 275         if (!attr.ndev)
 276                 return 0;
 277
 278         dev_put(attr.ndev);
 279
 280         if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
 281                 return 0;
 282
 283         return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
 284 }
 285
 286 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
 287 {
 288         return !MLX5_CAP_GEN(dev->mdev, ib_virt);
 289 }
 290
 291 enum {
 292         MLX5_VPORT_ACCESS_METHOD_MAD,
 293         MLX5_VPORT_ACCESS_METHOD_HCA,
 294         MLX5_VPORT_ACCESS_METHOD_NIC,
 295 };
 296
 297 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
 298 {
 299         if (mlx5_use_mad_ifc(to_mdev(ibdev)))
 300                 return MLX5_VPORT_ACCESS_METHOD_MAD;
 301
 302         if (mlx5_ib_port_link_layer(ibdev, 1) ==
 303             IB_LINK_LAYER_ETHERNET)
 304                 return MLX5_VPORT_ACCESS_METHOD_NIC;
 305
 306         return MLX5_VPORT_ACCESS_METHOD_HCA;
 307 }
 308
 309 static void get_atomic_caps(struct mlx5_ib_dev *dev,
 310                             struct ib_device_attr *props)
 311 {
 312         u8 tmp;
 313         u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
 314         u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
 315         u8 atomic_req_8B_endianness_mode =
 316                 MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
 317
 318         /* Check if HW supports 8 bytes standard atomic operations and capable
 319          * of host endianness respond
 320          */
 321         tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
 322         if (((atomic_operations & tmp) == tmp) &&
 323             (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
 324             (atomic_req_8B_endianness_mode)) {
 325                 props->atomic_cap = IB_ATOMIC_HCA;
 326         } else {
 327                 props->atomic_cap = IB_ATOMIC_NONE;
 328         }
 329 }
 330
 331 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
 332                                         __be64 *sys_image_guid)
 333 {
 334         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 335         struct mlx5_core_dev *mdev = dev->mdev;
 336         u64 tmp;
 337         int err;
 338
 339         switch (mlx5_get_vport_access_method(ibdev)) {
 340         case MLX5_VPORT_ACCESS_METHOD_MAD:
 341                 return mlx5_query_mad_ifc_system_image_guid(ibdev,
 342                                                             sys_image_guid);
 343
 344         case MLX5_VPORT_ACCESS_METHOD_HCA:
 345                 err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
 346                 break;
 347
 348         case MLX5_VPORT_ACCESS_METHOD_NIC:
 349                 err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
 350                 break;
 351
 352         default:
 353                 return -EINVAL;
 354         }
 355
 356         if (!err)
 357                 *sys_image_guid = cpu_to_be64(tmp);
 358
 359         return err;
 360
 361 }
 362
 363 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
 364                                 u16 *max_pkeys)
 365 {
 366         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 367         struct mlx5_core_dev *mdev = dev->mdev;
 368
 369         switch (mlx5_get_vport_access_method(ibdev)) {
 370         case MLX5_VPORT_ACCESS_METHOD_MAD:
 371                 return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
 372
 373         case MLX5_VPORT_ACCESS_METHOD_HCA:
 374         case MLX5_VPORT_ACCESS_METHOD_NIC:
 375                 *max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
 376                                                 pkey_table_size));
 377                 return 0;
 378
 379         default:
 380                 return -EINVAL;
 381         }
 382 }
 383
 384 static int mlx5_query_vendor_id(struct ib_device *ibdev,
 385                                 u32 *vendor_id)
 386 {
 387         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 388
 389         switch (mlx5_get_vport_access_method(ibdev)) {
 390         case MLX5_VPORT_ACCESS_METHOD_MAD:
 391                 return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
 392
 393         case MLX5_VPORT_ACCESS_METHOD_HCA:
 394         case MLX5_VPORT_ACCESS_METHOD_NIC:
 395                 return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
 396
 397         default:
 398                 return -EINVAL;
 399         }
 400 }
 401
 402 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
 403                                 __be64 *node_guid)
 404 {
 405         u64 tmp;
 406         int err;
 407
 408         switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
 409         case MLX5_VPORT_ACCESS_METHOD_MAD:
 410                 return mlx5_query_mad_ifc_node_guid(dev, node_guid);
 411
 412         case MLX5_VPORT_ACCESS_METHOD_HCA:
 413                 err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
 414                 break;
 415
 416         case MLX5_VPORT_ACCESS_METHOD_NIC:
 417                 err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
 418                 break;
 419
 420         default:
 421                 return -EINVAL;
 422         }
 423
 424         if (!err)
 425                 *node_guid = cpu_to_be64(tmp);
 426
 427         return err;
 428 }
 429
 430 struct mlx5_reg_node_desc {
 431         u8      desc[64];
 432 };
 433
 434 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
 435 {
 436         struct mlx5_reg_node_desc in;
 437
 438         if (mlx5_use_mad_ifc(dev))
 439                 return mlx5_query_mad_ifc_node_desc(dev, node_desc);
 440
 441         memset(&in, 0, sizeof(in));
 442
 443         return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
 444                                     sizeof(struct mlx5_reg_node_desc),
 445                                     MLX5_REG_NODE_DESC, 0, 0);
 446 }
 447
 448 static int mlx5_ib_query_device(struct ib_device *ibdev,
 449                                 struct ib_device_attr *props,
 450                                 struct ib_udata *uhw)
 451 {
 452         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 453         struct mlx5_core_dev *mdev = dev->mdev;
 454         int err = -ENOMEM;
 455         int max_rq_sg;
 456         int max_sq_sg;
 457         u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
 458         struct mlx5_ib_query_device_resp resp = {};
 459         size_t resp_len;
 460         u64 max_tso;
 461
 462         resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
 463         if (uhw->outlen && uhw->outlen < resp_len)
 464                 return -EINVAL;
 465         else
 466                 resp.response_length = resp_len;
 467
 468         if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
 469                 return -EINVAL;
 470
 471         memset(props, 0, sizeof(*props));
 472         err = mlx5_query_system_image_guid(ibdev,
 473                                            &props->sys_image_guid);
 474         if (err)
 475                 return err;
 476
 477         err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
 478         if (err)
 479                 return err;
 480
 481         err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
 482         if (err)
 483                 return err;
 484
 485         props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
 486                 (fw_rev_min(dev->mdev) << 16) |
 487                 fw_rev_sub(dev->mdev);
 488         props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
 489                 IB_DEVICE_PORT_ACTIVE_EVENT             |
 490                 IB_DEVICE_SYS_IMAGE_GUID                |
 491                 IB_DEVICE_RC_RNR_NAK_GEN;
 492
 493         if (MLX5_CAP_GEN(mdev, pkv))
 494                 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
 495         if (MLX5_CAP_GEN(mdev, qkv))
 496                 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
 497         if (MLX5_CAP_GEN(mdev, apm))
 498                 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
 499         if (MLX5_CAP_GEN(mdev, xrc))
 500                 props->device_cap_flags |= IB_DEVICE_XRC;
 501         if (MLX5_CAP_GEN(mdev, imaicl)) {
 502                 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
 503                                            IB_DEVICE_MEM_WINDOW_TYPE_2B;
 504                 props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
 505                 /* We support 'Gappy' memory registration too */
 506                 props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
 507         }
 508         props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
 509         if (MLX5_CAP_GEN(mdev, sho)) {
 510                 props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
 511                 /* At this stage no support for signature handover */
 512                 props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
 513                                       IB_PROT_T10DIF_TYPE_2 |
 514                                       IB_PROT_T10DIF_TYPE_3;
 515                 props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
 516                                        IB_GUARD_T10DIF_CSUM;
 517         }
 518         if (MLX5_CAP_GEN(mdev, block_lb_mc))
 519                 props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
 520
 521         if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
 522                 if (MLX5_CAP_ETH(mdev, csum_cap))
 523                         props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
 524
 525                 if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
 526                         max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
 527                         if (max_tso) {
 528                                 resp.tso_caps.max_tso = 1 << max_tso;
 529                                 resp.tso_caps.supported_qpts |=
 530                                         1 << IB_QPT_RAW_PACKET;
 531                                 resp.response_length += sizeof(resp.tso_caps);
 532                         }
 533                 }
 534
 535                 if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
 536                         resp.rss_caps.rx_hash_function =
 537                                                 MLX5_RX_HASH_FUNC_TOEPLITZ;
 538                         resp.rss_caps.rx_hash_fields_mask =
 539                                                 MLX5_RX_HASH_SRC_IPV4 |
 540                                                 MLX5_RX_HASH_DST_IPV4 |
 541                                                 MLX5_RX_HASH_SRC_IPV6 |
 542                                                 MLX5_RX_HASH_DST_IPV6 |
 543                                                 MLX5_RX_HASH_SRC_PORT_TCP |
 544                                                 MLX5_RX_HASH_DST_PORT_TCP |
 545                                                 MLX5_RX_HASH_SRC_PORT_UDP |
 546                                                 MLX5_RX_HASH_DST_PORT_UDP;
 547                         resp.response_length += sizeof(resp.rss_caps);
 548                 }
 549         } else {
 550                 if (field_avail(typeof(resp), tso_caps, uhw->outlen))
 551                         resp.response_length += sizeof(resp.tso_caps);
 552                 if (field_avail(typeof(resp), rss_caps, uhw->outlen))
 553                         resp.response_length += sizeof(resp.rss_caps);
 554         }
 555
 556         if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
 557                 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
 558                 props->device_cap_flags |= IB_DEVICE_UD_TSO;
 559         }
 560
 561         if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
 562             MLX5_CAP_ETH(dev->mdev, scatter_fcs))
 563                 props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
 564
 565         if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
 566                 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
 567
 568         props->vendor_part_id      = mdev->pdev->device;
 569         props->hw_ver              = mdev->pdev->revision;
 570
 571         props->max_mr_size         = ~0ull;
 572         props->page_size_cap       = ~(min_page_size - 1);
 573         props->max_qp              = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
 574         props->max_qp_wr           = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
 575         max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
 576                      sizeof(struct mlx5_wqe_data_seg);
 577         max_sq_sg = (MLX5_CAP_GEN(mdev, max_wqe_sz_sq) -
 578                      sizeof(struct mlx5_wqe_ctrl_seg)) /
 579                      sizeof(struct mlx5_wqe_data_seg);
 580         props->max_sge = min(max_rq_sg, max_sq_sg);
 581         props->max_sge_rd          = MLX5_MAX_SGE_RD;
 582         props->max_cq              = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
 583         props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
 584         props->max_mr              = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
 585         props->max_pd              = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
 586         props->max_qp_rd_atom      = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
 587         props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
 588         props->max_srq             = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
 589         props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
 590         props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
 591         props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
 592         props->max_srq_sge         = max_rq_sg - 1;
 593         props->max_fast_reg_page_list_len =
 594                 1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
 595         get_atomic_caps(dev, props);
 596         props->masked_atomic_cap   = IB_ATOMIC_NONE;
 597         props->max_mcast_grp       = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
 598         props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
 599         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
 600                                            props->max_mcast_grp;
 601         props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
 602         props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
 603         props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
 604
 605 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 606         if (MLX5_CAP_GEN(mdev, pg))
 607                 props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
 608         props->odp_caps = dev->odp_caps;
 609 #endif
 610
 611         if (MLX5_CAP_GEN(mdev, cd))
 612                 props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
 613
 614         if (!mlx5_core_is_pf(mdev))
 615                 props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
 616
 617         if (mlx5_ib_port_link_layer(ibdev, 1) ==
 618             IB_LINK_LAYER_ETHERNET) {
 619                 props->rss_caps.max_rwq_indirection_tables =
 620                         1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
 621                 props->rss_caps.max_rwq_indirection_table_size =
 622                         1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
 623                 props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
 624                 props->max_wq_type_rq =
 625                         1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
 626         }
 627
 628         if (uhw->outlen) {
 629                 err = ib_copy_to_udata(uhw, &resp, resp.response_length);
 630
 631                 if (err)
 632                         return err;
 633         }
 634
 635         return 0;
 636 }
 637
 638 enum mlx5_ib_width {
 639         MLX5_IB_WIDTH_1X        = 1 << 0,
 640         MLX5_IB_WIDTH_2X        = 1 << 1,
 641         MLX5_IB_WIDTH_4X        = 1 << 2,
 642         MLX5_IB_WIDTH_8X        = 1 << 3,
 643         MLX5_IB_WIDTH_12X       = 1 << 4
 644 };
 645
 646 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
 647                                   u8 *ib_width)
 648 {
 649         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 650         int err = 0;
 651
 652         if (active_width & MLX5_IB_WIDTH_1X) {
 653                 *ib_width = IB_WIDTH_1X;
 654         } else if (active_width & MLX5_IB_WIDTH_2X) {
 655                 mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
 656                             (int)active_width);
 657                 err = -EINVAL;
 658         } else if (active_width & MLX5_IB_WIDTH_4X) {
 659                 *ib_width = IB_WIDTH_4X;
 660         } else if (active_width & MLX5_IB_WIDTH_8X) {
 661                 *ib_width = IB_WIDTH_8X;
 662         } else if (active_width & MLX5_IB_WIDTH_12X) {
 663                 *ib_width = IB_WIDTH_12X;
 664         } else {
 665                 mlx5_ib_dbg(dev, "Invalid active_width %d\n",
 666                             (int)active_width);
 667                 err = -EINVAL;
 668         }
 669
 670         return err;
 671 }
 672
 673 static int mlx5_mtu_to_ib_mtu(int mtu)
 674 {
 675         switch (mtu) {
 676         case 256: return 1;
 677         case 512: return 2;
 678         case 1024: return 3;
 679         case 2048: return 4;
 680         case 4096: return 5;
 681         default:
 682                 pr_warn("invalid mtu\n");
 683                 return -1;
 684         }
 685 }
 686
 687 enum ib_max_vl_num {
 688         __IB_MAX_VL_0           = 1,
 689         __IB_MAX_VL_0_1         = 2,
 690         __IB_MAX_VL_0_3         = 3,
 691         __IB_MAX_VL_0_7         = 4,
 692         __IB_MAX_VL_0_14        = 5,
 693 };
 694
 695 enum mlx5_vl_hw_cap {
 696         MLX5_VL_HW_0    = 1,
 697         MLX5_VL_HW_0_1  = 2,
 698         MLX5_VL_HW_0_2  = 3,
 699         MLX5_VL_HW_0_3  = 4,
 700         MLX5_VL_HW_0_4  = 5,
 701         MLX5_VL_HW_0_5  = 6,
 702         MLX5_VL_HW_0_6  = 7,
 703         MLX5_VL_HW_0_7  = 8,
 704         MLX5_VL_HW_0_14 = 15
 705 };
 706
 707 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
 708                                 u8 *max_vl_num)
 709 {
 710         switch (vl_hw_cap) {
 711         case MLX5_VL_HW_0:
 712                 *max_vl_num = __IB_MAX_VL_0;
 713                 break;
 714         case MLX5_VL_HW_0_1:
 715                 *max_vl_num = __IB_MAX_VL_0_1;
 716                 break;
 717         case MLX5_VL_HW_0_3:
 718                 *max_vl_num = __IB_MAX_VL_0_3;
 719                 break;
 720         case MLX5_VL_HW_0_7:
 721                 *max_vl_num = __IB_MAX_VL_0_7;
 722                 break;
 723         case MLX5_VL_HW_0_14:
 724                 *max_vl_num = __IB_MAX_VL_0_14;
 725                 break;
 726
 727         default:
 728                 return -EINVAL;
 729         }
 730
 731         return 0;
 732 }
 733
 734 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
 735                                struct ib_port_attr *props)
 736 {
 737         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 738         struct mlx5_core_dev *mdev = dev->mdev;
 739         struct mlx5_hca_vport_context *rep;
 740         u16 max_mtu;
 741         u16 oper_mtu;
 742         int err;
 743         u8 ib_link_width_oper;
 744         u8 vl_hw_cap;
 745
 746         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
 747         if (!rep) {
 748                 err = -ENOMEM;
 749                 goto out;
 750         }
 751
 752         memset(props, 0, sizeof(*props));
 753
 754         err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
 755         if (err)
 756                 goto out;
 757
 758         props->lid              = rep->lid;
 759         props->lmc              = rep->lmc;
 760         props->sm_lid           = rep->sm_lid;
 761         props->sm_sl            = rep->sm_sl;
 762         props->state            = rep->vport_state;
 763         props->phys_state       = rep->port_physical_state;
 764         props->port_cap_flags   = rep->cap_mask1;
 765         props->gid_tbl_len      = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
 766         props->max_msg_sz       = 1 << MLX5_CAP_GEN(mdev, log_max_msg);
 767         props->pkey_tbl_len     = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
 768         props->bad_pkey_cntr    = rep->pkey_violation_counter;
 769         props->qkey_viol_cntr   = rep->qkey_violation_counter;
 770         props->subnet_timeout   = rep->subnet_timeout;
 771         props->init_type_reply  = rep->init_type_reply;
 772         props->grh_required     = rep->grh_required;
 773
 774         err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
 775         if (err)
 776                 goto out;
 777
 778         err = translate_active_width(ibdev, ib_link_width_oper,
 779                                      &props->active_width);
 780         if (err)
 781                 goto out;
 782         err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
 783         if (err)
 784                 goto out;
 785
 786         mlx5_query_port_max_mtu(mdev, &max_mtu, port);
 787
 788         props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
 789
 790         mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
 791
 792         props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
 793
 794         err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
 795         if (err)
 796                 goto out;
 797
 798         err = translate_max_vl_num(ibdev, vl_hw_cap,
 799                                    &props->max_vl_num);
 800 out:
 801         kfree(rep);
 802         return err;
 803 }
 804
 805 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
 806                        struct ib_port_attr *props)
 807 {
 808         switch (mlx5_get_vport_access_method(ibdev)) {
 809         case MLX5_VPORT_ACCESS_METHOD_MAD:
 810                 return mlx5_query_mad_ifc_port(ibdev, port, props);
 811
 812         case MLX5_VPORT_ACCESS_METHOD_HCA:
 813                 return mlx5_query_hca_port(ibdev, port, props);
 814
 815         case MLX5_VPORT_ACCESS_METHOD_NIC:
 816                 return mlx5_query_port_roce(ibdev, port, props);
 817
 818         default:
 819                 return -EINVAL;
 820         }
 821 }
 822
 823 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
 824                              union ib_gid *gid)
 825 {
 826         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 827         struct mlx5_core_dev *mdev = dev->mdev;
 828
 829         switch (mlx5_get_vport_access_method(ibdev)) {
 830         case MLX5_VPORT_ACCESS_METHOD_MAD:
 831                 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
 832
 833         case MLX5_VPORT_ACCESS_METHOD_HCA:
 834                 return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
 835
 836         default:
 837                 return -EINVAL;
 838         }
 839
 840 }
 841
 842 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
 843                               u16 *pkey)
 844 {
 845         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 846         struct mlx5_core_dev *mdev = dev->mdev;
 847
 848         switch (mlx5_get_vport_access_method(ibdev)) {
 849         case MLX5_VPORT_ACCESS_METHOD_MAD:
 850                 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
 851
 852         case MLX5_VPORT_ACCESS_METHOD_HCA:
 853         case MLX5_VPORT_ACCESS_METHOD_NIC:
 854                 return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
 855                                                  pkey);
 856         default:
 857                 return -EINVAL;
 858         }
 859 }
 860
 861 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
 862                                  struct ib_device_modify *props)
 863 {
 864         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 865         struct mlx5_reg_node_desc in;
 866         struct mlx5_reg_node_desc out;
 867         int err;
 868
 869         if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
 870                 return -EOPNOTSUPP;
 871
 872         if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
 873                 return 0;
 874
 875         /*
 876          * If possible, pass node desc to FW, so it can generate
 877          * a 144 trap.  If cmd fails, just ignore.
 878          */
 879         memcpy(&in, props->node_desc, 64);
 880         err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
 881                                    sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
 882         if (err)
 883                 return err;
 884
 885         memcpy(ibdev->node_desc, props->node_desc, 64);
 886
 887         return err;
 888 }
 889
 890 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
 891                                struct ib_port_modify *props)
 892 {
 893         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 894         struct ib_port_attr attr;
 895         u32 tmp;
 896         int err;
 897
 898         mutex_lock(&dev->cap_mask_mutex);
 899
 900         err = mlx5_ib_query_port(ibdev, port, &attr);
 901         if (err)
 902                 goto out;
 903
 904         tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
 905                 ~props->clr_port_cap_mask;
 906
 907         err = mlx5_set_port_caps(dev->mdev, port, tmp);
 908
 909 out:
 910         mutex_unlock(&dev->cap_mask_mutex);
 911         return err;
 912 }
 913
 914 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
 915                                                   struct ib_udata *udata)
 916 {
 917         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 918         struct mlx5_ib_alloc_ucontext_req_v2 req = {};
 919         struct mlx5_ib_alloc_ucontext_resp resp = {};
 920         struct mlx5_ib_ucontext *context;
 921         struct mlx5_uuar_info *uuari;
 922         struct mlx5_uar *uars;
 923         int gross_uuars;
 924         int num_uars;
 925         int ver;
 926         int uuarn;
 927         int err;
 928         int i;
 929         size_t reqlen;
 930         size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
 931                                      max_cqe_version);
 932
 933         if (!dev->ib_active)
 934                 return ERR_PTR(-EAGAIN);
 935
 936         if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
 937                 return ERR_PTR(-EINVAL);
 938
 939         reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
 940         if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
 941                 ver = 0;
 942         else if (reqlen >= min_req_v2)
 943                 ver = 2;
 944         else
 945                 return ERR_PTR(-EINVAL);
 946
 947         err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
 948         if (err)
 949                 return ERR_PTR(err);
 950
 951         if (req.flags)
 952                 return ERR_PTR(-EINVAL);
 953
 954         if (req.total_num_uuars > MLX5_MAX_UUARS)
 955                 return ERR_PTR(-ENOMEM);
 956
 957         if (req.total_num_uuars == 0)
 958                 return ERR_PTR(-EINVAL);
 959
 960         if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
 961                 return ERR_PTR(-EOPNOTSUPP);
 962
 963         if (reqlen > sizeof(req) &&
 964             !ib_is_udata_cleared(udata, sizeof(req),
 965                                  reqlen - sizeof(req)))
 966                 return ERR_PTR(-EOPNOTSUPP);
 967
 968         req.total_num_uuars = ALIGN(req.total_num_uuars,
 969                                     MLX5_NON_FP_BF_REGS_PER_PAGE);
 970         if (req.num_low_latency_uuars > req.total_num_uuars - 1)
 971                 return ERR_PTR(-EINVAL);
 972
 973         num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
 974         gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
 975         resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
 976         if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
 977                 resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
 978         resp.cache_line_size = L1_CACHE_BYTES;
 979         resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
 980         resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
 981         resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
 982         resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
 983         resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
 984         resp.cqe_version = min_t(__u8,
 985                                  (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
 986                                  req.max_cqe_version);
 987         resp.response_length = min(offsetof(typeof(resp), response_length) +
 988                                    sizeof(resp.response_length), udata->outlen);
 989
 990         context = kzalloc(sizeof(*context), GFP_KERNEL);
 991         if (!context)
 992                 return ERR_PTR(-ENOMEM);
 993
 994         uuari = &context->uuari;
 995         mutex_init(&uuari->lock);
 996         uars = kcalloc(num_uars, sizeof(*uars), GFP_KERNEL);
 997         if (!uars) {
 998                 err = -ENOMEM;
 999                 goto out_ctx;
1000         }
1001
1002         uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
1003                                 sizeof(*uuari->bitmap),
1004                                 GFP_KERNEL);
1005         if (!uuari->bitmap) {
1006                 err = -ENOMEM;
1007                 goto out_uar_ctx;
1008         }
1009         /*
1010          * clear all fast path uuars
1011          */
1012         for (i = 0; i < gross_uuars; i++) {
1013                 uuarn = i & 3;
1014                 if (uuarn == 2 || uuarn == 3)
1015                         set_bit(i, uuari->bitmap);
1016         }
1017
1018         uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
1019         if (!uuari->count) {
1020                 err = -ENOMEM;
1021                 goto out_bitmap;
1022         }
1023
1024         for (i = 0; i < num_uars; i++) {
1025                 err = mlx5_cmd_alloc_uar(dev->mdev, &uars[i].index);
1026                 if (err)
1027                         goto out_count;
1028         }
1029
1030 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1031         context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1032 #endif
1033
1034         if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1035                 err = mlx5_core_alloc_transport_domain(dev->mdev,
1036                                                        &context->tdn);
1037                 if (err)
1038                         goto out_uars;
1039         }
1040
1041         INIT_LIST_HEAD(&context->vma_private_list);
1042         INIT_LIST_HEAD(&context->db_page_list);
1043         mutex_init(&context->db_page_mutex);
1044
1045         resp.tot_uuars = req.total_num_uuars;
1046         resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1047
1048         if (field_avail(typeof(resp), cqe_version, udata->outlen))
1049                 resp.response_length += sizeof(resp.cqe_version);
1050
1051         if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1052                 resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE;
1053                 resp.response_length += sizeof(resp.cmds_supp_uhw);
1054         }
1055
1056         /*
1057          * We don't want to expose information from the PCI bar that is located
1058          * after 4096 bytes, so if the arch only supports larger pages, let's
1059          * pretend we don't support reading the HCA's core clock. This is also
1060          * forced by mmap function.
1061          */
1062         if (PAGE_SIZE <= 4096 &&
1063             field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1064                 resp.comp_mask |=
1065                         MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1066                 resp.hca_core_clock_offset =
1067                         offsetof(struct mlx5_init_seg, internal_timer_h) %
1068                         PAGE_SIZE;
1069                 resp.response_length += sizeof(resp.hca_core_clock_offset) +
1070                                         sizeof(resp.reserved2);
1071         }
1072
1073         err = ib_copy_to_udata(udata, &resp, resp.response_length);
1074         if (err)
1075                 goto out_td;
1076
1077         uuari->ver = ver;
1078         uuari->num_low_latency_uuars = req.num_low_latency_uuars;
1079         uuari->uars = uars;
1080         uuari->num_uars = num_uars;
1081         context->cqe_version = resp.cqe_version;
1082
1083         return &context->ibucontext;
1084
1085 out_td:
1086         if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1087                 mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1088
1089 out_uars:
1090         for (i--; i >= 0; i--)
1091                 mlx5_cmd_free_uar(dev->mdev, uars[i].index);
1092 out_count:
1093         kfree(uuari->count);
1094
1095 out_bitmap:
1096         kfree(uuari->bitmap);
1097
1098 out_uar_ctx:
1099         kfree(uars);
1100
1101 out_ctx:
1102         kfree(context);
1103         return ERR_PTR(err);
1104 }
1105
1106 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1107 {
1108         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1109         struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1110         struct mlx5_uuar_info *uuari = &context->uuari;
1111         int i;
1112
1113         if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1114                 mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1115
1116         for (i = 0; i < uuari->num_uars; i++) {
1117                 if (mlx5_cmd_free_uar(dev->mdev, uuari->uars[i].index))
1118                         mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index);
1119         }
1120
1121         kfree(uuari->count);
1122         kfree(uuari->bitmap);
1123         kfree(uuari->uars);
1124         kfree(context);
1125
1126         return 0;
1127 }
1128
1129 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, int index)
1130 {
1131         return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + index;
1132 }
1133
1134 static int get_command(unsigned long offset)
1135 {
1136         return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1137 }
1138
1139 static int get_arg(unsigned long offset)
1140 {
1141         return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1142 }
1143
1144 static int get_index(unsigned long offset)
1145 {
1146         return get_arg(offset);
1147 }
1148
1149 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1150 {
1151         /* vma_open is called when a new VMA is created on top of our VMA.  This
1152          * is done through either mremap flow or split_vma (usually due to
1153          * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1154          * as this VMA is strongly hardware related.  Therefore we set the
1155          * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1156          * calling us again and trying to do incorrect actions.  We assume that
1157          * the original VMA size is exactly a single page, and therefore all
1158          * "splitting" operation will not happen to it.
1159          */
1160         area->vm_ops = NULL;
1161 }
1162
1163 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1164 {
1165         struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1166
1167         /* It's guaranteed that all VMAs opened on a FD are closed before the
1168          * file itself is closed, therefore no sync is needed with the regular
1169          * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1170          * However need a sync with accessing the vma as part of
1171          * mlx5_ib_disassociate_ucontext.
1172          * The close operation is usually called under mm->mmap_sem except when
1173          * process is exiting.
1174          * The exiting case is handled explicitly as part of
1175          * mlx5_ib_disassociate_ucontext.
1176          */
1177         mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1178
1179         /* setting the vma context pointer to null in the mlx5_ib driver's
1180          * private data, to protect a race condition in
1181          * mlx5_ib_disassociate_ucontext().
1182          */
1183         mlx5_ib_vma_priv_data->vma = NULL;
1184         list_del(&mlx5_ib_vma_priv_data->list);
1185         kfree(mlx5_ib_vma_priv_data);
1186 }
1187
1188 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1189         .open = mlx5_ib_vma_open,
1190         .close = mlx5_ib_vma_close
1191 };
1192
1193 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1194                                 struct mlx5_ib_ucontext *ctx)
1195 {
1196         struct mlx5_ib_vma_private_data *vma_prv;
1197         struct list_head *vma_head = &ctx->vma_private_list;
1198
1199         vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1200         if (!vma_prv)
1201                 return -ENOMEM;
1202
1203         vma_prv->vma = vma;
1204         vma->vm_private_data = vma_prv;
1205         vma->vm_ops =  &mlx5_ib_vm_ops;
1206
1207         list_add(&vma_prv->list, vma_head);
1208
1209         return 0;
1210 }
1211
1212 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1213 {
1214         int ret;
1215         struct vm_area_struct *vma;
1216         struct mlx5_ib_vma_private_data *vma_private, *n;
1217         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1218         struct task_struct *owning_process  = NULL;
1219         struct mm_struct   *owning_mm       = NULL;
1220
1221         owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1222         if (!owning_process)
1223                 return;
1224
1225         owning_mm = get_task_mm(owning_process);
1226         if (!owning_mm) {
1227                 pr_info("no mm, disassociate ucontext is pending task termination\n");
1228                 while (1) {
1229                         put_task_struct(owning_process);
1230                         usleep_range(1000, 2000);
1231                         owning_process = get_pid_task(ibcontext->tgid,
1232                                                       PIDTYPE_PID);
1233                         if (!owning_process ||
1234                             owning_process->state == TASK_DEAD) {
1235                                 pr_info("disassociate ucontext done, task was terminated\n");
1236                                 /* in case task was dead need to release the
1237                                  * task struct.
1238                                  */
1239                                 if (owning_process)
1240                                         put_task_struct(owning_process);
1241                                 return;
1242                         }
1243                 }
1244         }
1245
1246         /* need to protect from a race on closing the vma as part of
1247          * mlx5_ib_vma_close.
1248          */
1249         down_read(&owning_mm->mmap_sem);
1250         list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1251                                  list) {
1252                 vma = vma_private->vma;
1253                 ret = zap_vma_ptes(vma, vma->vm_start,
1254                                    PAGE_SIZE);
1255                 WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1256                 /* context going to be destroyed, should
1257                  * not access ops any more.
1258                  */
1259                 vma->vm_ops = NULL;
1260                 list_del(&vma_private->list);
1261                 kfree(vma_private);
1262         }
1263         up_read(&owning_mm->mmap_sem);
1264         mmput(owning_mm);
1265         put_task_struct(owning_process);
1266 }
1267
1268 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1269 {
1270         switch (cmd) {
1271         case MLX5_IB_MMAP_WC_PAGE:
1272                 return "WC";
1273         case MLX5_IB_MMAP_REGULAR_PAGE:
1274                 return "best effort WC";
1275         case MLX5_IB_MMAP_NC_PAGE:
1276                 return "NC";
1277         default:
1278                 return NULL;
1279         }
1280 }
1281
1282 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1283                     struct vm_area_struct *vma,
1284                     struct mlx5_ib_ucontext *context)
1285 {
1286         struct mlx5_uuar_info *uuari = &context->uuari;
1287         int err;
1288         unsigned long idx;
1289         phys_addr_t pfn, pa;
1290         pgprot_t prot;
1291
1292         switch (cmd) {
1293         case MLX5_IB_MMAP_WC_PAGE:
1294 /* Some architectures don't support WC memory */
1295 #if defined(CONFIG_X86)
1296                 if (!pat_enabled())
1297                         return -EPERM;
1298 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1299                         return -EPERM;
1300 #endif
1301         /* fall through */
1302         case MLX5_IB_MMAP_REGULAR_PAGE:
1303                 /* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1304                 prot = pgprot_writecombine(vma->vm_page_prot);
1305                 break;
1306         case MLX5_IB_MMAP_NC_PAGE:
1307                 prot = pgprot_noncached(vma->vm_page_prot);
1308                 break;
1309         default:
1310                 return -EINVAL;
1311         }
1312
1313         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1314                 return -EINVAL;
1315
1316         idx = get_index(vma->vm_pgoff);
1317         if (idx >= uuari->num_uars)
1318                 return -EINVAL;
1319
1320         pfn = uar_index2pfn(dev, uuari->uars[idx].index);
1321         mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1322
1323         vma->vm_page_prot = prot;
1324         err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1325                                  PAGE_SIZE, vma->vm_page_prot);
1326         if (err) {
1327                 mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
1328                             err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
1329                 return -EAGAIN;
1330         }
1331
1332         pa = pfn << PAGE_SHIFT;
1333         mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
1334                     vma->vm_start, &pa);
1335
1336         return mlx5_ib_set_vma_data(vma, context);
1337 }
1338
1339 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1340 {
1341         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1342         struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1343         unsigned long command;
1344         phys_addr_t pfn;
1345
1346         command = get_command(vma->vm_pgoff);
1347         switch (command) {
1348         case MLX5_IB_MMAP_WC_PAGE:
1349         case MLX5_IB_MMAP_NC_PAGE:
1350         case MLX5_IB_MMAP_REGULAR_PAGE:
1351                 return uar_mmap(dev, command, vma, context);
1352
1353         case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1354                 return -ENOSYS;
1355
1356         case MLX5_IB_MMAP_CORE_CLOCK:
1357                 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1358                         return -EINVAL;
1359
1360                 if (vma->vm_flags & VM_WRITE)
1361                         return -EPERM;
1362
1363                 /* Don't expose to user-space information it shouldn't have */
1364                 if (PAGE_SIZE > 4096)
1365                         return -EOPNOTSUPP;
1366
1367                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1368                 pfn = (dev->mdev->iseg_base +
1369                        offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1370                         PAGE_SHIFT;
1371                 if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1372                                        PAGE_SIZE, vma->vm_page_prot))
1373                         return -EAGAIN;
1374
1375                 mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
1376                             vma->vm_start,
1377                             (unsigned long long)pfn << PAGE_SHIFT);
1378                 break;
1379
1380         default:
1381                 return -EINVAL;
1382         }
1383
1384         return 0;
1385 }
1386
1387 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1388                                       struct ib_ucontext *context,
1389                                       struct ib_udata *udata)
1390 {
1391         struct mlx5_ib_alloc_pd_resp resp;
1392         struct mlx5_ib_pd *pd;
1393         int err;
1394
1395         pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1396         if (!pd)
1397                 return ERR_PTR(-ENOMEM);
1398
1399         err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1400         if (err) {
1401                 kfree(pd);
1402                 return ERR_PTR(err);
1403         }
1404
1405         if (context) {
1406                 resp.pdn = pd->pdn;
1407                 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1408                         mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1409                         kfree(pd);
1410                         return ERR_PTR(-EFAULT);
1411                 }
1412         }
1413
1414         return &pd->ibpd;
1415 }
1416
1417 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1418 {
1419         struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1420         struct mlx5_ib_pd *mpd = to_mpd(pd);
1421
1422         mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1423         kfree(mpd);
1424
1425         return 0;
1426 }
1427
1428 static bool outer_header_zero(u32 *match_criteria)
1429 {
1430         int size = MLX5_ST_SZ_BYTES(fte_match_param);
1431         char *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_criteria,
1432                                              outer_headers);
1433
1434         return outer_headers_c[0] == 0 && !memcmp(outer_headers_c,
1435                                                   outer_headers_c + 1,
1436                                                   size - 1);
1437 }
1438
1439 #define LAST_ETH_FIELD vlan_tag
1440 #define LAST_IB_FIELD sl
1441 #define LAST_IPV4_FIELD dst_ip
1442 #define LAST_IPV6_FIELD dst_ip
1443 #define LAST_TCP_UDP_FIELD src_port
1444
1445 /* Field is the last supported field */
1446 #define FIELDS_NOT_SUPPORTED(filter, field)\
1447         memchr_inv((void *)&filter.field  +\
1448                    sizeof(filter.field), 0,\
1449                    sizeof(filter) -\
1450                    offsetof(typeof(filter), field) -\
1451                    sizeof(filter.field))
1452
1453 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1454                            const union ib_flow_spec *ib_spec)
1455 {
1456         void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1457                                              outer_headers);
1458         void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1459                                              outer_headers);
1460         switch (ib_spec->type) {
1461         case IB_FLOW_SPEC_ETH:
1462                 if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1463                         return -ENOTSUPP;
1464
1465                 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1466                                              dmac_47_16),
1467                                 ib_spec->eth.mask.dst_mac);
1468                 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1469                                              dmac_47_16),
1470                                 ib_spec->eth.val.dst_mac);
1471
1472                 if (ib_spec->eth.mask.vlan_tag) {
1473                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1474                                  vlan_tag, 1);
1475                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1476                                  vlan_tag, 1);
1477
1478                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1479                                  first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1480                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1481                                  first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1482
1483                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1484                                  first_cfi,
1485                                  ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1486                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1487                                  first_cfi,
1488                                  ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1489
1490                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1491                                  first_prio,
1492                                  ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1493                         MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1494                                  first_prio,
1495                                  ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1496                 }
1497                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1498                          ethertype, ntohs(ib_spec->eth.mask.ether_type));
1499                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1500                          ethertype, ntohs(ib_spec->eth.val.ether_type));
1501                 break;
1502         case IB_FLOW_SPEC_IPV4:
1503                 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1504                         return -ENOTSUPP;
1505
1506                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1507                          ethertype, 0xffff);
1508                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1509                          ethertype, ETH_P_IP);
1510
1511                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1512                                     src_ipv4_src_ipv6.ipv4_layout.ipv4),
1513                        &ib_spec->ipv4.mask.src_ip,
1514                        sizeof(ib_spec->ipv4.mask.src_ip));
1515                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1516                                     src_ipv4_src_ipv6.ipv4_layout.ipv4),
1517                        &ib_spec->ipv4.val.src_ip,
1518                        sizeof(ib_spec->ipv4.val.src_ip));
1519                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1520                                     dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1521                        &ib_spec->ipv4.mask.dst_ip,
1522                        sizeof(ib_spec->ipv4.mask.dst_ip));
1523                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1524                                     dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1525                        &ib_spec->ipv4.val.dst_ip,
1526                        sizeof(ib_spec->ipv4.val.dst_ip));
1527                 break;
1528         case IB_FLOW_SPEC_IPV6:
1529                 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1530                         return -ENOTSUPP;
1531
1532                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c,
1533                          ethertype, 0xffff);
1534                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v,
1535                          ethertype, ETH_P_IPV6);
1536
1537                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1538                                     src_ipv4_src_ipv6.ipv6_layout.ipv6),
1539                        &ib_spec->ipv6.mask.src_ip,
1540                        sizeof(ib_spec->ipv6.mask.src_ip));
1541                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1542                                     src_ipv4_src_ipv6.ipv6_layout.ipv6),
1543                        &ib_spec->ipv6.val.src_ip,
1544                        sizeof(ib_spec->ipv6.val.src_ip));
1545                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c,
1546                                     dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1547                        &ib_spec->ipv6.mask.dst_ip,
1548                        sizeof(ib_spec->ipv6.mask.dst_ip));
1549                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v,
1550                                     dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1551                        &ib_spec->ipv6.val.dst_ip,
1552                        sizeof(ib_spec->ipv6.val.dst_ip));
1553                 break;
1554         case IB_FLOW_SPEC_TCP:
1555                 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1556                                          LAST_TCP_UDP_FIELD))
1557                         return -ENOTSUPP;
1558
1559                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1560                          0xff);
1561                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1562                          IPPROTO_TCP);
1563
1564                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport,
1565                          ntohs(ib_spec->tcp_udp.mask.src_port));
1566                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport,
1567                          ntohs(ib_spec->tcp_udp.val.src_port));
1568
1569                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport,
1570                          ntohs(ib_spec->tcp_udp.mask.dst_port));
1571                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport,
1572                          ntohs(ib_spec->tcp_udp.val.dst_port));
1573                 break;
1574         case IB_FLOW_SPEC_UDP:
1575                 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1576                                          LAST_TCP_UDP_FIELD))
1577                         return -ENOTSUPP;
1578
1579                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol,
1580                          0xff);
1581                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol,
1582                          IPPROTO_UDP);
1583
1584                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport,
1585                          ntohs(ib_spec->tcp_udp.mask.src_port));
1586                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport,
1587                          ntohs(ib_spec->tcp_udp.val.src_port));
1588
1589                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport,
1590                          ntohs(ib_spec->tcp_udp.mask.dst_port));
1591                 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport,
1592                          ntohs(ib_spec->tcp_udp.val.dst_port));
1593                 break;
1594         default:
1595                 return -EINVAL;
1596         }
1597
1598         return 0;
1599 }
1600
1601 /* If a flow could catch both multicast and unicast packets,
1602  * it won't fall into the multicast flow steering table and this rule
1603  * could steal other multicast packets.
1604  */
1605 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1606 {
1607         struct ib_flow_spec_eth *eth_spec;
1608
1609         if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1610             ib_attr->size < sizeof(struct ib_flow_attr) +
1611             sizeof(struct ib_flow_spec_eth) ||
1612             ib_attr->num_of_specs < 1)
1613                 return false;
1614
1615         eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1616         if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1617             eth_spec->size != sizeof(*eth_spec))
1618                 return false;
1619
1620         return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1621                is_multicast_ether_addr(eth_spec->val.dst_mac);
1622 }
1623
1624 static bool is_valid_attr(const struct ib_flow_attr *flow_attr)
1625 {
1626         union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1627         bool has_ipv4_spec = false;
1628         bool eth_type_ipv4 = true;
1629         unsigned int spec_index;
1630
1631         /* Validate that ethertype is correct */
1632         for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1633                 if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1634                     ib_spec->eth.mask.ether_type) {
1635                         if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1636                               ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1637                                 eth_type_ipv4 = false;
1638                 } else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1639                         has_ipv4_spec = true;
1640                 }
1641                 ib_spec = (void *)ib_spec + ib_spec->size;
1642         }
1643         return !has_ipv4_spec || eth_type_ipv4;
1644 }
1645
1646 static void put_flow_table(struct mlx5_ib_dev *dev,
1647                            struct mlx5_ib_flow_prio *prio, bool ft_added)
1648 {
1649         prio->refcount -= !!ft_added;
1650         if (!prio->refcount) {
1651                 mlx5_destroy_flow_table(prio->flow_table);
1652                 prio->flow_table = NULL;
1653         }
1654 }
1655
1656 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
1657 {
1658         struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
1659         struct mlx5_ib_flow_handler *handler = container_of(flow_id,
1660                                                           struct mlx5_ib_flow_handler,
1661                                                           ibflow);
1662         struct mlx5_ib_flow_handler *iter, *tmp;
1663
1664         mutex_lock(&dev->flow_db.lock);
1665
1666         list_for_each_entry_safe(iter, tmp, &handler->list, list) {
1667                 mlx5_del_flow_rule(iter->rule);
1668                 put_flow_table(dev, iter->prio, true);
1669                 list_del(&iter->list);
1670                 kfree(iter);
1671         }
1672
1673         mlx5_del_flow_rule(handler->rule);
1674         put_flow_table(dev, handler->prio, true);
1675         mutex_unlock(&dev->flow_db.lock);
1676
1677         kfree(handler);
1678
1679         return 0;
1680 }
1681
1682 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
1683 {
1684         priority *= 2;
1685         if (!dont_trap)
1686                 priority++;
1687         return priority;
1688 }
1689
1690 enum flow_table_type {
1691         MLX5_IB_FT_RX,
1692         MLX5_IB_FT_TX
1693 };
1694
1695 #define MLX5_FS_MAX_TYPES        10
1696 #define MLX5_FS_MAX_ENTRIES      32000UL
1697 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
1698                                                 struct ib_flow_attr *flow_attr,
1699                                                 enum flow_table_type ft_type)
1700 {
1701         bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
1702         struct mlx5_flow_namespace *ns = NULL;
1703         struct mlx5_ib_flow_prio *prio;
1704         struct mlx5_flow_table *ft;
1705         int num_entries;
1706         int num_groups;
1707         int priority;
1708         int err = 0;
1709
1710         if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1711                 if (flow_is_multicast_only(flow_attr) &&
1712                     !dont_trap)
1713                         priority = MLX5_IB_FLOW_MCAST_PRIO;
1714                 else
1715                         priority = ib_prio_to_core_prio(flow_attr->priority,
1716                                                         dont_trap);
1717                 ns = mlx5_get_flow_namespace(dev->mdev,
1718                                              MLX5_FLOW_NAMESPACE_BYPASS);
1719                 num_entries = MLX5_FS_MAX_ENTRIES;
1720                 num_groups = MLX5_FS_MAX_TYPES;
1721                 prio = &dev->flow_db.prios[priority];
1722         } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1723                    flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1724                 ns = mlx5_get_flow_namespace(dev->mdev,
1725                                              MLX5_FLOW_NAMESPACE_LEFTOVERS);
1726                 build_leftovers_ft_param(&priority,
1727                                          &num_entries,
1728                                          &num_groups);
1729                 prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
1730         } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
1731                 if (!MLX5_CAP_FLOWTABLE(dev->mdev,
1732                                         allow_sniffer_and_nic_rx_shared_tir))
1733                         return ERR_PTR(-ENOTSUPP);
1734
1735                 ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
1736                                              MLX5_FLOW_NAMESPACE_SNIFFER_RX :
1737                                              MLX5_FLOW_NAMESPACE_SNIFFER_TX);
1738
1739                 prio = &dev->flow_db.sniffer[ft_type];
1740                 priority = 0;
1741                 num_entries = 1;
1742                 num_groups = 1;
1743         }
1744
1745         if (!ns)
1746                 return ERR_PTR(-ENOTSUPP);
1747
1748         ft = prio->flow_table;
1749         if (!ft) {
1750                 ft = mlx5_create_auto_grouped_flow_table(ns, priority,
1751                                                          num_entries,
1752                                                          num_groups,
1753                                                          0);
1754
1755                 if (!IS_ERR(ft)) {
1756                         prio->refcount = 0;
1757                         prio->flow_table = ft;
1758                 } else {
1759                         err = PTR_ERR(ft);
1760                 }
1761         }
1762
1763         return err ? ERR_PTR(err) : prio;
1764 }
1765
1766 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
1767                                                      struct mlx5_ib_flow_prio *ft_prio,
1768                                                      const struct ib_flow_attr *flow_attr,
1769                                                      struct mlx5_flow_destination *dst)
1770 {
1771         struct mlx5_flow_table  *ft = ft_prio->flow_table;
1772         struct mlx5_ib_flow_handler *handler;
1773         struct mlx5_flow_spec *spec;
1774         const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
1775         unsigned int spec_index;
1776         u32 action;
1777         int err = 0;
1778
1779         if (!is_valid_attr(flow_attr))
1780                 return ERR_PTR(-EINVAL);
1781
1782         spec = mlx5_vzalloc(sizeof(*spec));
1783         handler = kzalloc(sizeof(*handler), GFP_KERNEL);
1784         if (!handler || !spec) {
1785                 err = -ENOMEM;
1786                 goto free;
1787         }
1788
1789         INIT_LIST_HEAD(&handler->list);
1790
1791         for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1792                 err = parse_flow_attr(spec->match_criteria,
1793                                       spec->match_value, ib_flow);
1794                 if (err < 0)
1795                         goto free;
1796
1797                 ib_flow += ((union ib_flow_spec *)ib_flow)->size;
1798         }
1799
1800         /* Outer header support only */
1801         spec->match_criteria_enable = (!outer_header_zero(spec->match_criteria))
1802                 << 0;
1803         action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
1804                 MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
1805         handler->rule = mlx5_add_flow_rule(ft, spec,
1806                                            action,
1807                                            MLX5_FS_DEFAULT_FLOW_TAG,
1808                                            dst);
1809
1810         if (IS_ERR(handler->rule)) {
1811                 err = PTR_ERR(handler->rule);
1812                 goto free;
1813         }
1814
1815         ft_prio->refcount++;
1816         handler->prio = ft_prio;
1817
1818         ft_prio->flow_table = ft;
1819 free:
1820         if (err)
1821                 kfree(handler);
1822         kvfree(spec);
1823         return err ? ERR_PTR(err) : handler;
1824 }
1825
1826 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
1827                                                           struct mlx5_ib_flow_prio *ft_prio,
1828                                                           struct ib_flow_attr *flow_attr,
1829                                                           struct mlx5_flow_destination *dst)
1830 {
1831         struct mlx5_ib_flow_handler *handler_dst = NULL;
1832         struct mlx5_ib_flow_handler *handler = NULL;
1833
1834         handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
1835         if (!IS_ERR(handler)) {
1836                 handler_dst = create_flow_rule(dev, ft_prio,
1837                                                flow_attr, dst);
1838                 if (IS_ERR(handler_dst)) {
1839                         mlx5_del_flow_rule(handler->rule);
1840                         ft_prio->refcount--;
1841                         kfree(handler);
1842                         handler = handler_dst;
1843                 } else {
1844                         list_add(&handler_dst->list, &handler->list);
1845                 }
1846         }
1847
1848         return handler;
1849 }
1850 enum {
1851         LEFTOVERS_MC,
1852         LEFTOVERS_UC,
1853 };
1854
1855 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
1856                                                           struct mlx5_ib_flow_prio *ft_prio,
1857                                                           struct ib_flow_attr *flow_attr,
1858                                                           struct mlx5_flow_destination *dst)
1859 {
1860         struct mlx5_ib_flow_handler *handler_ucast = NULL;
1861         struct mlx5_ib_flow_handler *handler = NULL;
1862
1863         static struct {
1864                 struct ib_flow_attr     flow_attr;
1865                 struct ib_flow_spec_eth eth_flow;
1866         } leftovers_specs[] = {
1867                 [LEFTOVERS_MC] = {
1868                         .flow_attr = {
1869                                 .num_of_specs = 1,
1870                                 .size = sizeof(leftovers_specs[0])
1871                         },
1872                         .eth_flow = {
1873                                 .type = IB_FLOW_SPEC_ETH,
1874                                 .size = sizeof(struct ib_flow_spec_eth),
1875                                 .mask = {.dst_mac = {0x1} },
1876                                 .val =  {.dst_mac = {0x1} }
1877                         }
1878                 },
1879                 [LEFTOVERS_UC] = {
1880                         .flow_attr = {
1881                                 .num_of_specs = 1,
1882                                 .size = sizeof(leftovers_specs[0])
1883                         },
1884                         .eth_flow = {
1885                                 .type = IB_FLOW_SPEC_ETH,
1886                                 .size = sizeof(struct ib_flow_spec_eth),
1887                                 .mask = {.dst_mac = {0x1} },
1888                                 .val = {.dst_mac = {} }
1889                         }
1890                 }
1891         };
1892
1893         handler = create_flow_rule(dev, ft_prio,
1894                                    &leftovers_specs[LEFTOVERS_MC].flow_attr,
1895                                    dst);
1896         if (!IS_ERR(handler) &&
1897             flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
1898                 handler_ucast = create_flow_rule(dev, ft_prio,
1899                                                  &leftovers_specs[LEFTOVERS_UC].flow_attr,
1900                                                  dst);
1901                 if (IS_ERR(handler_ucast)) {
1902                         mlx5_del_flow_rule(handler->rule);
1903                         ft_prio->refcount--;
1904                         kfree(handler);
1905                         handler = handler_ucast;
1906                 } else {
1907                         list_add(&handler_ucast->list, &handler->list);
1908                 }
1909         }
1910
1911         return handler;
1912 }
1913
1914 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
1915                                                         struct mlx5_ib_flow_prio *ft_rx,
1916                                                         struct mlx5_ib_flow_prio *ft_tx,
1917                                                         struct mlx5_flow_destination *dst)
1918 {
1919         struct mlx5_ib_flow_handler *handler_rx;
1920         struct mlx5_ib_flow_handler *handler_tx;
1921         int err;
1922         static const struct ib_flow_attr flow_attr  = {
1923                 .num_of_specs = 0,
1924                 .size = sizeof(flow_attr)
1925         };
1926
1927         handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
1928         if (IS_ERR(handler_rx)) {
1929                 err = PTR_ERR(handler_rx);
1930                 goto err;
1931         }
1932
1933         handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
1934         if (IS_ERR(handler_tx)) {
1935                 err = PTR_ERR(handler_tx);
1936                 goto err_tx;
1937         }
1938
1939         list_add(&handler_tx->list, &handler_rx->list);
1940
1941         return handler_rx;
1942
1943 err_tx:
1944         mlx5_del_flow_rule(handler_rx->rule);
1945         ft_rx->refcount--;
1946         kfree(handler_rx);
1947 err:
1948         return ERR_PTR(err);
1949 }
1950
1951 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
1952                                            struct ib_flow_attr *flow_attr,
1953                                            int domain)
1954 {
1955         struct mlx5_ib_dev *dev = to_mdev(qp->device);
1956         struct mlx5_ib_flow_handler *handler = NULL;
1957         struct mlx5_flow_destination *dst = NULL;
1958         struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
1959         struct mlx5_ib_flow_prio *ft_prio;
1960         int err;
1961
1962         if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
1963                 return ERR_PTR(-ENOSPC);
1964
1965         if (domain != IB_FLOW_DOMAIN_USER ||
1966             flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
1967             (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
1968                 return ERR_PTR(-EINVAL);
1969
1970         dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1971         if (!dst)
1972                 return ERR_PTR(-ENOMEM);
1973
1974         mutex_lock(&dev->flow_db.lock);
1975
1976         ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
1977         if (IS_ERR(ft_prio)) {
1978                 err = PTR_ERR(ft_prio);
1979                 goto unlock;
1980         }
1981         if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
1982                 ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
1983                 if (IS_ERR(ft_prio_tx)) {
1984                         err = PTR_ERR(ft_prio_tx);
1985                         ft_prio_tx = NULL;
1986                         goto destroy_ft;
1987                 }
1988         }
1989
1990         dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
1991         dst->tir_num = to_mqp(qp)->raw_packet_qp.rq.tirn;
1992
1993         if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1994                 if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
1995                         handler = create_dont_trap_rule(dev, ft_prio,
1996                                                         flow_attr, dst);
1997                 } else {
1998                         handler = create_flow_rule(dev, ft_prio, flow_attr,
1999                                                    dst);
2000                 }
2001         } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2002                    flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2003                 handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2004                                                 dst);
2005         } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2006                 handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2007         } else {
2008                 err = -EINVAL;
2009                 goto destroy_ft;
2010         }
2011
2012         if (IS_ERR(handler)) {
2013                 err = PTR_ERR(handler);
2014                 handler = NULL;
2015                 goto destroy_ft;
2016         }
2017
2018         mutex_unlock(&dev->flow_db.lock);
2019         kfree(dst);
2020
2021         return &handler->ibflow;
2022
2023 destroy_ft:
2024         put_flow_table(dev, ft_prio, false);
2025         if (ft_prio_tx)
2026                 put_flow_table(dev, ft_prio_tx, false);
2027 unlock:
2028         mutex_unlock(&dev->flow_db.lock);
2029         kfree(dst);
2030         kfree(handler);
2031         return ERR_PTR(err);
2032 }
2033
2034 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2035 {
2036         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2037         int err;
2038
2039         err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2040         if (err)
2041                 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2042                              ibqp->qp_num, gid->raw);
2043
2044         return err;
2045 }
2046
2047 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2048 {
2049         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2050         int err;
2051
2052         err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2053         if (err)
2054                 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2055                              ibqp->qp_num, gid->raw);
2056
2057         return err;
2058 }
2059
2060 static int init_node_data(struct mlx5_ib_dev *dev)
2061 {
2062         int err;
2063
2064         err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2065         if (err)
2066                 return err;
2067
2068         dev->mdev->rev_id = dev->mdev->pdev->revision;
2069
2070         return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2071 }
2072
2073 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2074                              char *buf)
2075 {
2076         struct mlx5_ib_dev *dev =
2077                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2078
2079         return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
2080 }
2081
2082 static ssize_t show_reg_pages(struct device *device,
2083                               struct device_attribute *attr, char *buf)
2084 {
2085         struct mlx5_ib_dev *dev =
2086                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2087
2088         return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2089 }
2090
2091 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2092                         char *buf)
2093 {
2094         struct mlx5_ib_dev *dev =
2095                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2096         return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2097 }
2098
2099 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2100                         char *buf)
2101 {
2102         struct mlx5_ib_dev *dev =
2103                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2104         return sprintf(buf, "%x\n", dev->mdev->rev_id);
2105 }
2106
2107 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2108                           char *buf)
2109 {
2110         struct mlx5_ib_dev *dev =
2111                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2112         return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2113                        dev->mdev->board_id);
2114 }
2115
2116 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2117 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2118 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2119 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2120 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2121
2122 static struct device_attribute *mlx5_class_attributes[] = {
2123         &dev_attr_hw_rev,
2124         &dev_attr_hca_type,
2125         &dev_attr_board_id,
2126         &dev_attr_fw_pages,
2127         &dev_attr_reg_pages,
2128 };
2129
2130 static void pkey_change_handler(struct work_struct *work)
2131 {
2132         struct mlx5_ib_port_resources *ports =
2133                 container_of(work, struct mlx5_ib_port_resources,
2134                              pkey_change_work);
2135
2136         mutex_lock(&ports->devr->mutex);
2137         mlx5_ib_gsi_pkey_change(ports->gsi);
2138         mutex_unlock(&ports->devr->mutex);
2139 }
2140
2141 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2142 {
2143         struct mlx5_ib_qp *mqp;
2144         struct mlx5_ib_cq *send_mcq, *recv_mcq;
2145         struct mlx5_core_cq *mcq;
2146         struct list_head cq_armed_list;
2147         unsigned long flags_qp;
2148         unsigned long flags_cq;
2149         unsigned long flags;
2150
2151         INIT_LIST_HEAD(&cq_armed_list);
2152
2153         /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2154         spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2155         list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2156                 spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2157                 if (mqp->sq.tail != mqp->sq.head) {
2158                         send_mcq = to_mcq(mqp->ibqp.send_cq);
2159                         spin_lock_irqsave(&send_mcq->lock, flags_cq);
2160                         if (send_mcq->mcq.comp &&
2161                             mqp->ibqp.send_cq->comp_handler) {
2162                                 if (!send_mcq->mcq.reset_notify_added) {
2163                                         send_mcq->mcq.reset_notify_added = 1;
2164                                         list_add_tail(&send_mcq->mcq.reset_notify,
2165                                                       &cq_armed_list);
2166                                 }
2167                         }
2168                         spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2169                 }
2170                 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2171                 spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2172                 /* no handling is needed for SRQ */
2173                 if (!mqp->ibqp.srq) {
2174                         if (mqp->rq.tail != mqp->rq.head) {
2175                                 recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2176                                 spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2177                                 if (recv_mcq->mcq.comp &&
2178                                     mqp->ibqp.recv_cq->comp_handler) {
2179                                         if (!recv_mcq->mcq.reset_notify_added) {
2180                                                 recv_mcq->mcq.reset_notify_added = 1;
2181                                                 list_add_tail(&recv_mcq->mcq.reset_notify,
2182                                                               &cq_armed_list);
2183                                         }
2184                                 }
2185                                 spin_unlock_irqrestore(&recv_mcq->lock,
2186                                                        flags_cq);
2187                         }
2188                 }
2189                 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2190         }
2191         /*At that point all inflight post send were put to be executed as of we
2192          * lock/unlock above locks Now need to arm all involved CQs.
2193          */
2194         list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2195                 mcq->comp(mcq);
2196         }
2197         spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2198 }
2199
2200 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2201                           enum mlx5_dev_event event, unsigned long param)
2202 {
2203         struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2204         struct ib_event ibev;
2205
2206         u8 port = 0;
2207
2208         switch (event) {
2209         case MLX5_DEV_EVENT_SYS_ERROR:
2210                 ibdev->ib_active = false;
2211                 ibev.event = IB_EVENT_DEVICE_FATAL;
2212                 mlx5_ib_handle_internal_error(ibdev);
2213                 break;
2214
2215         case MLX5_DEV_EVENT_PORT_UP:
2216                 ibev.event = IB_EVENT_PORT_ACTIVE;
2217                 port = (u8)param;
2218                 break;
2219
2220         case MLX5_DEV_EVENT_PORT_DOWN:
2221         case MLX5_DEV_EVENT_PORT_INITIALIZED:
2222                 ibev.event = IB_EVENT_PORT_ERR;
2223                 port = (u8)param;
2224                 break;
2225
2226         case MLX5_DEV_EVENT_LID_CHANGE:
2227                 ibev.event = IB_EVENT_LID_CHANGE;
2228                 port = (u8)param;
2229                 break;
2230
2231         case MLX5_DEV_EVENT_PKEY_CHANGE:
2232                 ibev.event = IB_EVENT_PKEY_CHANGE;
2233                 port = (u8)param;
2234
2235                 schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2236                 break;
2237
2238         case MLX5_DEV_EVENT_GUID_CHANGE:
2239                 ibev.event = IB_EVENT_GID_CHANGE;
2240                 port = (u8)param;
2241                 break;
2242
2243         case MLX5_DEV_EVENT_CLIENT_REREG:
2244                 ibev.event = IB_EVENT_CLIENT_REREGISTER;
2245                 port = (u8)param;
2246                 break;
2247         }
2248
2249         ibev.device           = &ibdev->ib_dev;
2250         ibev.element.port_num = port;
2251
2252         if (port < 1 || port > ibdev->num_ports) {
2253                 mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2254                 return;
2255         }
2256
2257         if (ibdev->ib_active)
2258                 ib_dispatch_event(&ibev);
2259 }
2260
2261 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2262 {
2263         int port;
2264
2265         for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2266                 mlx5_query_ext_port_caps(dev, port);
2267 }
2268
2269 static int get_port_caps(struct mlx5_ib_dev *dev)
2270 {
2271         struct ib_device_attr *dprops = NULL;
2272         struct ib_port_attr *pprops = NULL;
2273         int err = -ENOMEM;
2274         int port;
2275         struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2276
2277         pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2278         if (!pprops)
2279                 goto out;
2280
2281         dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2282         if (!dprops)
2283                 goto out;
2284
2285         err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2286         if (err) {
2287                 mlx5_ib_warn(dev, "query_device failed %d\n", err);
2288                 goto out;
2289         }
2290
2291         for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2292                 err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2293                 if (err) {
2294                         mlx5_ib_warn(dev, "query_port %d failed %d\n",
2295                                      port, err);
2296                         break;
2297                 }
2298                 dev->mdev->port_caps[port - 1].pkey_table_len =
2299                                                 dprops->max_pkeys;
2300                 dev->mdev->port_caps[port - 1].gid_table_len =
2301                                                 pprops->gid_tbl_len;
2302                 mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2303                             dprops->max_pkeys, pprops->gid_tbl_len);
2304         }
2305
2306 out:
2307         kfree(pprops);
2308         kfree(dprops);
2309
2310         return err;
2311 }
2312
2313 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2314 {
2315         int err;
2316
2317         err = mlx5_mr_cache_cleanup(dev);
2318         if (err)
2319                 mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2320
2321         mlx5_ib_destroy_qp(dev->umrc.qp);
2322         ib_free_cq(dev->umrc.cq);
2323         ib_dealloc_pd(dev->umrc.pd);
2324 }
2325
2326 enum {
2327         MAX_UMR_WR = 128,
2328 };
2329
2330 static int create_umr_res(struct mlx5_ib_dev *dev)
2331 {
2332         struct ib_qp_init_attr *init_attr = NULL;
2333         struct ib_qp_attr *attr = NULL;
2334         struct ib_pd *pd;
2335         struct ib_cq *cq;
2336         struct ib_qp *qp;
2337         int ret;
2338
2339         attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2340         init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2341         if (!attr || !init_attr) {
2342                 ret = -ENOMEM;
2343                 goto error_0;
2344         }
2345
2346         pd = ib_alloc_pd(&dev->ib_dev, 0);
2347         if (IS_ERR(pd)) {
2348                 mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2349                 ret = PTR_ERR(pd);
2350                 goto error_0;
2351         }
2352
2353         cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2354         if (IS_ERR(cq)) {
2355                 mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2356                 ret = PTR_ERR(cq);
2357                 goto error_2;
2358         }
2359
2360         init_attr->send_cq = cq;
2361         init_attr->recv_cq = cq;
2362         init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2363         init_attr->cap.max_send_wr = MAX_UMR_WR;
2364         init_attr->cap.max_send_sge = 1;
2365         init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2366         init_attr->port_num = 1;
2367         qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2368         if (IS_ERR(qp)) {
2369                 mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2370                 ret = PTR_ERR(qp);
2371                 goto error_3;
2372         }
2373         qp->device     = &dev->ib_dev;
2374         qp->real_qp    = qp;
2375         qp->uobject    = NULL;
2376         qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2377
2378         attr->qp_state = IB_QPS_INIT;
2379         attr->port_num = 1;
2380         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2381                                 IB_QP_PORT, NULL);
2382         if (ret) {
2383                 mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2384                 goto error_4;
2385         }
2386
2387         memset(attr, 0, sizeof(*attr));
2388         attr->qp_state = IB_QPS_RTR;
2389         attr->path_mtu = IB_MTU_256;
2390
2391         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2392         if (ret) {
2393                 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2394                 goto error_4;
2395         }
2396
2397         memset(attr, 0, sizeof(*attr));
2398         attr->qp_state = IB_QPS_RTS;
2399         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2400         if (ret) {
2401                 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2402                 goto error_4;
2403         }
2404
2405         dev->umrc.qp = qp;
2406         dev->umrc.cq = cq;
2407         dev->umrc.pd = pd;
2408
2409         sema_init(&dev->umrc.sem, MAX_UMR_WR);
2410         ret = mlx5_mr_cache_init(dev);
2411         if (ret) {
2412                 mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2413                 goto error_4;
2414         }
2415
2416         kfree(attr);
2417         kfree(init_attr);
2418
2419         return 0;
2420
2421 error_4:
2422         mlx5_ib_destroy_qp(qp);
2423
2424 error_3:
2425         ib_free_cq(cq);
2426
2427 error_2:
2428         ib_dealloc_pd(pd);
2429
2430 error_0:
2431         kfree(attr);
2432         kfree(init_attr);
2433         return ret;
2434 }
2435
2436 static int create_dev_resources(struct mlx5_ib_resources *devr)
2437 {
2438         struct ib_srq_init_attr attr;
2439         struct mlx5_ib_dev *dev;
2440         struct ib_cq_init_attr cq_attr = {.cqe = 1};
2441         int port;
2442         int ret = 0;
2443
2444         dev = container_of(devr, struct mlx5_ib_dev, devr);
2445
2446         mutex_init(&devr->mutex);
2447
2448         devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2449         if (IS_ERR(devr->p0)) {
2450                 ret = PTR_ERR(devr->p0);
2451                 goto error0;
2452         }
2453         devr->p0->device  = &dev->ib_dev;
2454         devr->p0->uobject = NULL;
2455         atomic_set(&devr->p0->usecnt, 0);
2456
2457         devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
2458         if (IS_ERR(devr->c0)) {
2459                 ret = PTR_ERR(devr->c0);
2460                 goto error1;
2461         }
2462         devr->c0->device        = &dev->ib_dev;
2463         devr->c0->uobject       = NULL;
2464         devr->c0->comp_handler  = NULL;
2465         devr->c0->event_handler = NULL;
2466         devr->c0->cq_context    = NULL;
2467         atomic_set(&devr->c0->usecnt, 0);
2468
2469         devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2470         if (IS_ERR(devr->x0)) {
2471                 ret = PTR_ERR(devr->x0);
2472                 goto error2;
2473         }
2474         devr->x0->device = &dev->ib_dev;
2475         devr->x0->inode = NULL;
2476         atomic_set(&devr->x0->usecnt, 0);
2477         mutex_init(&devr->x0->tgt_qp_mutex);
2478         INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2479
2480         devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2481         if (IS_ERR(devr->x1)) {
2482                 ret = PTR_ERR(devr->x1);
2483                 goto error3;
2484         }
2485         devr->x1->device = &dev->ib_dev;
2486         devr->x1->inode = NULL;
2487         atomic_set(&devr->x1->usecnt, 0);
2488         mutex_init(&devr->x1->tgt_qp_mutex);
2489         INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2490
2491         memset(&attr, 0, sizeof(attr));
2492         attr.attr.max_sge = 1;
2493         attr.attr.max_wr = 1;
2494         attr.srq_type = IB_SRQT_XRC;
2495         attr.ext.xrc.cq = devr->c0;
2496         attr.ext.xrc.xrcd = devr->x0;
2497
2498         devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2499         if (IS_ERR(devr->s0)) {
2500                 ret = PTR_ERR(devr->s0);
2501                 goto error4;
2502         }
2503         devr->s0->device        = &dev->ib_dev;
2504         devr->s0->pd            = devr->p0;
2505         devr->s0->uobject       = NULL;
2506         devr->s0->event_handler = NULL;
2507         devr->s0->srq_context   = NULL;
2508         devr->s0->srq_type      = IB_SRQT_XRC;
2509         devr->s0->ext.xrc.xrcd  = devr->x0;
2510         devr->s0->ext.xrc.cq    = devr->c0;
2511         atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2512         atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2513         atomic_inc(&devr->p0->usecnt);
2514         atomic_set(&devr->s0->usecnt, 0);
2515
2516         memset(&attr, 0, sizeof(attr));
2517         attr.attr.max_sge = 1;
2518         attr.attr.max_wr = 1;
2519         attr.srq_type = IB_SRQT_BASIC;
2520         devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2521         if (IS_ERR(devr->s1)) {
2522                 ret = PTR_ERR(devr->s1);
2523                 goto error5;
2524         }
2525         devr->s1->device        = &dev->ib_dev;
2526         devr->s1->pd            = devr->p0;
2527         devr->s1->uobject       = NULL;
2528         devr->s1->event_handler = NULL;
2529         devr->s1->srq_context   = NULL;
2530         devr->s1->srq_type      = IB_SRQT_BASIC;
2531         devr->s1->ext.xrc.cq    = devr->c0;
2532         atomic_inc(&devr->p0->usecnt);
2533         atomic_set(&devr->s0->usecnt, 0);
2534
2535         for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2536                 INIT_WORK(&devr->ports[port].pkey_change_work,
2537                           pkey_change_handler);
2538                 devr->ports[port].devr = devr;
2539         }
2540
2541         return 0;
2542
2543 error5:
2544         mlx5_ib_destroy_srq(devr->s0);
2545 error4:
2546         mlx5_ib_dealloc_xrcd(devr->x1);
2547 error3:
2548         mlx5_ib_dealloc_xrcd(devr->x0);
2549 error2:
2550         mlx5_ib_destroy_cq(devr->c0);
2551 error1:
2552         mlx5_ib_dealloc_pd(devr->p0);
2553 error0:
2554         return ret;
2555 }
2556
2557 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2558 {
2559         struct mlx5_ib_dev *dev =
2560                 container_of(devr, struct mlx5_ib_dev, devr);
2561         int port;
2562
2563         mlx5_ib_destroy_srq(devr->s1);
2564         mlx5_ib_destroy_srq(devr->s0);
2565         mlx5_ib_dealloc_xrcd(devr->x0);
2566         mlx5_ib_dealloc_xrcd(devr->x1);
2567         mlx5_ib_destroy_cq(devr->c0);
2568         mlx5_ib_dealloc_pd(devr->p0);
2569
2570         /* Make sure no change P_Key work items are still executing */
2571         for (port = 0; port < dev->num_ports; ++port)
2572                 cancel_work_sync(&devr->ports[port].pkey_change_work);
2573 }
2574
2575 static u32 get_core_cap_flags(struct ib_device *ibdev)
2576 {
2577         struct mlx5_ib_dev *dev = to_mdev(ibdev);
2578         enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2579         u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2580         u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2581         u32 ret = 0;
2582
2583         if (ll == IB_LINK_LAYER_INFINIBAND)
2584                 return RDMA_CORE_PORT_IBA_IB;
2585
2586         if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2587                 return 0;
2588
2589         if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2590                 return 0;
2591
2592         if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2593                 ret |= RDMA_CORE_PORT_IBA_ROCE;
2594
2595         if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2596                 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2597
2598         return ret;
2599 }
2600
2601 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
2602                                struct ib_port_immutable *immutable)
2603 {
2604         struct ib_port_attr attr;
2605         int err;
2606
2607         err = mlx5_ib_query_port(ibdev, port_num, &attr);
2608         if (err)
2609                 return err;
2610
2611         immutable->pkey_tbl_len = attr.pkey_tbl_len;
2612         immutable->gid_tbl_len = attr.gid_tbl_len;
2613         immutable->core_cap_flags = get_core_cap_flags(ibdev);
2614         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2615
2616         return 0;
2617 }
2618
2619 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
2620                            size_t str_len)
2621 {
2622         struct mlx5_ib_dev *dev =
2623                 container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2624         snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
2625                        fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
2626 }
2627
2628 static int mlx5_enable_roce(struct mlx5_ib_dev *dev)
2629 {
2630         int err;
2631
2632         dev->roce.nb.notifier_call = mlx5_netdev_event;
2633         err = register_netdevice_notifier(&dev->roce.nb);
2634         if (err)
2635                 return err;
2636
2637         err = mlx5_nic_vport_enable_roce(dev->mdev);
2638         if (err)
2639                 goto err_unregister_netdevice_notifier;
2640
2641         return 0;
2642
2643 err_unregister_netdevice_notifier:
2644         unregister_netdevice_notifier(&dev->roce.nb);
2645         return err;
2646 }
2647
2648 static void mlx5_disable_roce(struct mlx5_ib_dev *dev)
2649 {
2650         mlx5_nic_vport_disable_roce(dev->mdev);
2651         unregister_netdevice_notifier(&dev->roce.nb);
2652 }
2653
2654 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
2655 {
2656         unsigned int i;
2657
2658         for (i = 0; i < dev->num_ports; i++)
2659                 mlx5_core_dealloc_q_counter(dev->mdev,
2660                                             dev->port[i].q_cnt_id);
2661 }
2662
2663 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
2664 {
2665         int i;
2666         int ret;
2667
2668         for (i = 0; i < dev->num_ports; i++) {
2669                 ret = mlx5_core_alloc_q_counter(dev->mdev,
2670                                                 &dev->port[i].q_cnt_id);
2671                 if (ret) {
2672                         mlx5_ib_warn(dev,
2673                                      "couldn't allocate queue counter for port %d, err %d\n",
2674                                      i + 1, ret);
2675                         goto dealloc_counters;
2676                 }
2677         }
2678
2679         return 0;
2680
2681 dealloc_counters:
2682         while (--i >= 0)
2683                 mlx5_core_dealloc_q_counter(dev->mdev,
2684                                             dev->port[i].q_cnt_id);
2685
2686         return ret;
2687 }
2688
2689 static const char * const names[] = {
2690         "rx_write_requests",
2691         "rx_read_requests",
2692         "rx_atomic_requests",
2693         "out_of_buffer",
2694         "out_of_sequence",
2695         "duplicate_request",
2696         "rnr_nak_retry_err",
2697         "packet_seq_err",
2698         "implied_nak_seq_err",
2699         "local_ack_timeout_err",
2700 };
2701
2702 static const size_t stats_offsets[] = {
2703         MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests),
2704         MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests),
2705         MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests),
2706         MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer),
2707         MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence),
2708         MLX5_BYTE_OFF(query_q_counter_out, duplicate_request),
2709         MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err),
2710         MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err),
2711         MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err),
2712         MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err),
2713 };
2714
2715 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
2716                                                     u8 port_num)
2717 {
2718         BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets));
2719
2720         /* We support only per port stats */
2721         if (port_num == 0)
2722                 return NULL;
2723
2724         return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names),
2725                                           RDMA_HW_STATS_DEFAULT_LIFESPAN);
2726 }
2727
2728 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
2729                                 struct rdma_hw_stats *stats,
2730                                 u8 port, int index)
2731 {
2732         struct mlx5_ib_dev *dev = to_mdev(ibdev);
2733         int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
2734         void *out;
2735         __be32 val;
2736         int ret;
2737         int i;
2738
2739         if (!port || !stats)
2740                 return -ENOSYS;
2741
2742         out = mlx5_vzalloc(outlen);
2743         if (!out)
2744                 return -ENOMEM;
2745
2746         ret = mlx5_core_query_q_counter(dev->mdev,
2747                                         dev->port[port - 1].q_cnt_id, 0,
2748                                         out, outlen);
2749         if (ret)
2750                 goto free;
2751
2752         for (i = 0; i < ARRAY_SIZE(names); i++) {
2753                 val = *(__be32 *)(out + stats_offsets[i]);
2754                 stats->value[i] = (u64)be32_to_cpu(val);
2755         }
2756 free:
2757         kvfree(out);
2758         return ARRAY_SIZE(names);
2759 }
2760
2761 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
2762 {
2763         struct mlx5_ib_dev *dev;
2764         enum rdma_link_layer ll;
2765         int port_type_cap;
2766         int err;
2767         int i;
2768
2769         port_type_cap = MLX5_CAP_GEN(mdev, port_type);
2770         ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
2771
2772         if ((ll == IB_LINK_LAYER_ETHERNET) && !MLX5_CAP_GEN(mdev, roce))
2773                 return NULL;
2774
2775         printk_once(KERN_INFO "%s", mlx5_version);
2776
2777         dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
2778         if (!dev)
2779                 return NULL;
2780
2781         dev->mdev = mdev;
2782
2783         dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
2784                             GFP_KERNEL);
2785         if (!dev->port)
2786                 goto err_dealloc;
2787
2788         rwlock_init(&dev->roce.netdev_lock);
2789         err = get_port_caps(dev);
2790         if (err)
2791                 goto err_free_port;
2792
2793         if (mlx5_use_mad_ifc(dev))
2794                 get_ext_port_caps(dev);
2795
2796         MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock);
2797
2798         strlcpy(dev->ib_dev.name, "mlx5_%d", IB_DEVICE_NAME_MAX);
2799         dev->ib_dev.owner               = THIS_MODULE;
2800         dev->ib_dev.node_type           = RDMA_NODE_IB_CA;
2801         dev->ib_dev.local_dma_lkey      = 0 /* not supported for now */;
2802         dev->num_ports          = MLX5_CAP_GEN(mdev, num_ports);
2803         dev->ib_dev.phys_port_cnt     = dev->num_ports;
2804         dev->ib_dev.num_comp_vectors    =
2805                 dev->mdev->priv.eq_table.num_comp_vectors;
2806         dev->ib_dev.dma_device  = &mdev->pdev->dev;
2807
2808         dev->ib_dev.uverbs_abi_ver      = MLX5_IB_UVERBS_ABI_VERSION;
2809         dev->ib_dev.uverbs_cmd_mask     =
2810                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
2811                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
2812                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
2813                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
2814                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
2815                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
2816                 (1ull << IB_USER_VERBS_CMD_REREG_MR)            |
2817                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
2818                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2819                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
2820                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
2821                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
2822                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
2823                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
2824                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
2825                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
2826                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
2827                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
2828                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
2829                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
2830                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
2831                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
2832                 (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)         |
2833                 (1ull << IB_USER_VERBS_CMD_OPEN_QP);
2834         dev->ib_dev.uverbs_ex_cmd_mask =
2835                 (1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)     |
2836                 (1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)        |
2837                 (1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
2838
2839         dev->ib_dev.query_device        = mlx5_ib_query_device;
2840         dev->ib_dev.query_port          = mlx5_ib_query_port;
2841         dev->ib_dev.get_link_layer      = mlx5_ib_port_link_layer;
2842         if (ll == IB_LINK_LAYER_ETHERNET)
2843                 dev->ib_dev.get_netdev  = mlx5_ib_get_netdev;
2844         dev->ib_dev.query_gid           = mlx5_ib_query_gid;
2845         dev->ib_dev.add_gid             = mlx5_ib_add_gid;
2846         dev->ib_dev.del_gid             = mlx5_ib_del_gid;
2847         dev->ib_dev.query_pkey          = mlx5_ib_query_pkey;
2848         dev->ib_dev.modify_device       = mlx5_ib_modify_device;
2849         dev->ib_dev.modify_port         = mlx5_ib_modify_port;
2850         dev->ib_dev.alloc_ucontext      = mlx5_ib_alloc_ucontext;
2851         dev->ib_dev.dealloc_ucontext    = mlx5_ib_dealloc_ucontext;
2852         dev->ib_dev.mmap                = mlx5_ib_mmap;
2853         dev->ib_dev.alloc_pd            = mlx5_ib_alloc_pd;
2854         dev->ib_dev.dealloc_pd          = mlx5_ib_dealloc_pd;
2855         dev->ib_dev.create_ah           = mlx5_ib_create_ah;
2856         dev->ib_dev.query_ah            = mlx5_ib_query_ah;
2857         dev->ib_dev.destroy_ah          = mlx5_ib_destroy_ah;
2858         dev->ib_dev.create_srq          = mlx5_ib_create_srq;
2859         dev->ib_dev.modify_srq          = mlx5_ib_modify_srq;
2860         dev->ib_dev.query_srq           = mlx5_ib_query_srq;
2861         dev->ib_dev.destroy_srq         = mlx5_ib_destroy_srq;
2862         dev->ib_dev.post_srq_recv       = mlx5_ib_post_srq_recv;
2863         dev->ib_dev.create_qp           = mlx5_ib_create_qp;
2864         dev->ib_dev.modify_qp           = mlx5_ib_modify_qp;
2865         dev->ib_dev.query_qp            = mlx5_ib_query_qp;
2866         dev->ib_dev.destroy_qp          = mlx5_ib_destroy_qp;
2867         dev->ib_dev.post_send           = mlx5_ib_post_send;
2868         dev->ib_dev.post_recv           = mlx5_ib_post_recv;
2869         dev->ib_dev.create_cq           = mlx5_ib_create_cq;
2870         dev->ib_dev.modify_cq           = mlx5_ib_modify_cq;
2871         dev->ib_dev.resize_cq           = mlx5_ib_resize_cq;
2872         dev->ib_dev.destroy_cq          = mlx5_ib_destroy_cq;
2873         dev->ib_dev.poll_cq             = mlx5_ib_poll_cq;
2874         dev->ib_dev.req_notify_cq       = mlx5_ib_arm_cq;
2875         dev->ib_dev.get_dma_mr          = mlx5_ib_get_dma_mr;
2876         dev->ib_dev.reg_user_mr         = mlx5_ib_reg_user_mr;
2877         dev->ib_dev.rereg_user_mr       = mlx5_ib_rereg_user_mr;
2878         dev->ib_dev.dereg_mr            = mlx5_ib_dereg_mr;
2879         dev->ib_dev.attach_mcast        = mlx5_ib_mcg_attach;
2880         dev->ib_dev.detach_mcast        = mlx5_ib_mcg_detach;
2881         dev->ib_dev.process_mad         = mlx5_ib_process_mad;
2882         dev->ib_dev.alloc_mr            = mlx5_ib_alloc_mr;
2883         dev->ib_dev.map_mr_sg           = mlx5_ib_map_mr_sg;
2884         dev->ib_dev.check_mr_status     = mlx5_ib_check_mr_status;
2885         dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
2886         dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
2887         if (mlx5_core_is_pf(mdev)) {
2888                 dev->ib_dev.get_vf_config       = mlx5_ib_get_vf_config;
2889                 dev->ib_dev.set_vf_link_state   = mlx5_ib_set_vf_link_state;
2890                 dev->ib_dev.get_vf_stats        = mlx5_ib_get_vf_stats;
2891                 dev->ib_dev.set_vf_guid         = mlx5_ib_set_vf_guid;
2892         }
2893
2894         dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
2895
2896         mlx5_ib_internal_fill_odp_caps(dev);
2897
2898         if (MLX5_CAP_GEN(mdev, imaicl)) {
2899                 dev->ib_dev.alloc_mw            = mlx5_ib_alloc_mw;
2900                 dev->ib_dev.dealloc_mw          = mlx5_ib_dealloc_mw;
2901                 dev->ib_dev.uverbs_cmd_mask |=
2902                         (1ull << IB_USER_VERBS_CMD_ALLOC_MW)    |
2903                         (1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
2904         }
2905
2906         if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) &&
2907             MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
2908                 dev->ib_dev.get_hw_stats        = mlx5_ib_get_hw_stats;
2909                 dev->ib_dev.alloc_hw_stats      = mlx5_ib_alloc_hw_stats;
2910         }
2911
2912         if (MLX5_CAP_GEN(mdev, xrc)) {
2913                 dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
2914                 dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
2915                 dev->ib_dev.uverbs_cmd_mask |=
2916                         (1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
2917                         (1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
2918         }
2919
2920         if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
2921             IB_LINK_LAYER_ETHERNET) {
2922                 dev->ib_dev.create_flow = mlx5_ib_create_flow;
2923                 dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
2924                 dev->ib_dev.create_wq    = mlx5_ib_create_wq;
2925                 dev->ib_dev.modify_wq    = mlx5_ib_modify_wq;
2926                 dev->ib_dev.destroy_wq   = mlx5_ib_destroy_wq;
2927                 dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
2928                 dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
2929                 dev->ib_dev.uverbs_ex_cmd_mask |=
2930                         (1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
2931                         (1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
2932                         (1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
2933                         (1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
2934                         (1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
2935                         (1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
2936                         (1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
2937         }
2938         err = init_node_data(dev);
2939         if (err)
2940                 goto err_dealloc;
2941
2942         mutex_init(&dev->flow_db.lock);
2943         mutex_init(&dev->cap_mask_mutex);
2944         INIT_LIST_HEAD(&dev->qp_list);
2945         spin_lock_init(&dev->reset_flow_resource_lock);
2946
2947         if (ll == IB_LINK_LAYER_ETHERNET) {
2948                 err = mlx5_enable_roce(dev);
2949                 if (err)
2950                         goto err_dealloc;
2951         }
2952
2953         err = create_dev_resources(&dev->devr);
2954         if (err)
2955                 goto err_disable_roce;
2956
2957         err = mlx5_ib_odp_init_one(dev);
2958         if (err)
2959                 goto err_rsrc;
2960
2961         err = mlx5_ib_alloc_q_counters(dev);
2962         if (err)
2963                 goto err_odp;
2964
2965         err = ib_register_device(&dev->ib_dev, NULL);
2966         if (err)
2967                 goto err_q_cnt;
2968
2969         err = create_umr_res(dev);
2970         if (err)
2971                 goto err_dev;
2972
2973         for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
2974                 err = device_create_file(&dev->ib_dev.dev,
2975                                          mlx5_class_attributes[i]);
2976                 if (err)
2977                         goto err_umrc;
2978         }
2979
2980         dev->ib_active = true;
2981
2982         return dev;
2983
2984 err_umrc:
2985         destroy_umrc_res(dev);
2986
2987 err_dev:
2988         ib_unregister_device(&dev->ib_dev);
2989
2990 err_q_cnt:
2991         mlx5_ib_dealloc_q_counters(dev);
2992
2993 err_odp:
2994         mlx5_ib_odp_remove_one(dev);
2995
2996 err_rsrc:
2997         destroy_dev_resources(&dev->devr);
2998
2999 err_disable_roce:
3000         if (ll == IB_LINK_LAYER_ETHERNET)
3001                 mlx5_disable_roce(dev);
3002
3003 err_free_port:
3004         kfree(dev->port);
3005
3006 err_dealloc:
3007         ib_dealloc_device((struct ib_device *)dev);
3008
3009         return NULL;
3010 }
3011
3012 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3013 {
3014         struct mlx5_ib_dev *dev = context;
3015         enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3016
3017         ib_unregister_device(&dev->ib_dev);
3018         mlx5_ib_dealloc_q_counters(dev);
3019         destroy_umrc_res(dev);
3020         mlx5_ib_odp_remove_one(dev);
3021         destroy_dev_resources(&dev->devr);
3022         if (ll == IB_LINK_LAYER_ETHERNET)
3023                 mlx5_disable_roce(dev);
3024         kfree(dev->port);
3025         ib_dealloc_device(&dev->ib_dev);
3026 }
3027
3028 static struct mlx5_interface mlx5_ib_interface = {
3029         .add            = mlx5_ib_add,
3030         .remove         = mlx5_ib_remove,
3031         .event          = mlx5_ib_event,
3032         .protocol       = MLX5_INTERFACE_PROTOCOL_IB,
3033 };
3034
3035 static int __init mlx5_ib_init(void)
3036 {
3037         int err;
3038
3039         if (deprecated_prof_sel != 2)
3040                 pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
3041
3042         err = mlx5_ib_odp_init();
3043         if (err)
3044                 return err;
3045
3046         err = mlx5_register_interface(&mlx5_ib_interface);
3047         if (err)
3048                 goto clean_odp;
3049
3050         return err;
3051
3052 clean_odp:
3053         mlx5_ib_odp_cleanup();
3054         return err;
3055 }
3056
3057 static void __exit mlx5_ib_cleanup(void)
3058 {
3059         mlx5_unregister_interface(&mlx5_ib_interface);
3060         mlx5_ib_odp_cleanup();
3061 }
3062
3063 module_init(mlx5_ib_init);
3064 module_exit(mlx5_ib_cleanup);