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