drivers/infiniband/hw/mlx5/main.c

   1 /*
   2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/highmem.h>
  34 #include <linux/module.h>
  35 #include <linux/init.h>
  36 #include <linux/errno.h>
  37 #include <linux/pci.h>
  38 #include <linux/dma-mapping.h>
  39 #include <linux/slab.h>
  40 #if defined(CONFIG_X86)
  41 #include <asm/pat.h>
  42 #endif
  43 #include <linux/sched.h>
  44 #include <linux/delay.h>
  45 #include <rdma/ib_user_verbs.h>
  46 #include <rdma/ib_addr.h>
  47 #include <rdma/ib_cache.h>
  48 #include <linux/mlx5/port.h>
  49 #include <linux/mlx5/vport.h>
  50 #include <linux/list.h>
  51 #include <rdma/ib_smi.h>
  52 #include <rdma/ib_umem.h>
  53 #include <linux/in.h>
  54 #include <linux/etherdevice.h>
  55 #include <linux/mlx5/fs.h>
  56 #include "mlx5_ib.h"
  57
  58 #define DRIVER_NAME "mlx5_ib"
  59 #define DRIVER_VERSION "2.2-1"
  60 #define DRIVER_RELDATE  "Feb 2014"
  61
  62 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
  63 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver");
  64 MODULE_LICENSE("Dual BSD/GPL");
  65 MODULE_VERSION(DRIVER_VERSION);
  66
  67 static int deprecated_prof_sel = 2;
  68 module_param_named(prof_sel, deprecated_prof_sel, int, 0444);
  69 MODULE_PARM_DESC(prof_sel, "profile selector. Deprecated here. Moved to module mlx5_core");
  70
  71 static char mlx5_version[] =
  72         DRIVER_NAME ": Mellanox Connect-IB Infiniband driver v"
  73         DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
  74
  75 enum {
  76         MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
  77 };
  78
  79 static enum rdma_link_layer
  80 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
  81 {
  82         switch (port_type_cap) {
  83         case MLX5_CAP_PORT_TYPE_IB:
  84                 return IB_LINK_LAYER_INFINIBAND;
  85         case MLX5_CAP_PORT_TYPE_ETH:
  86                 return IB_LINK_LAYER_ETHERNET;
  87         default:
  88                 return IB_LINK_LAYER_UNSPECIFIED;
  89         }
  90 }
  91
  92 static enum rdma_link_layer
  93 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num)
  94 {
  95         struct mlx5_ib_dev *dev = to_mdev(device);
  96         int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
  97
  98         return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
  99 }
 100
 101 static int mlx5_netdev_event(struct notifier_block *this,
 102                              unsigned long event, void *ptr)
 103 {
 104         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
 105         struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
 106                                                  roce.nb);
 107
 108         switch (event) {
 109         case NETDEV_REGISTER:
 110         case NETDEV_UNREGISTER:
 111                 write_lock(&ibdev->roce.netdev_lock);
 112                 if (ndev->dev.parent == &ibdev->mdev->pdev->dev)
 113                         ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
 114                                              NULL : ndev;
 115                 write_unlock(&ibdev->roce.netdev_lock);
 116                 break;
 117
 118         case NETDEV_UP:
 119         case NETDEV_DOWN: {
 120                 struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
 121                 struct net_device *upper = NULL;
 122
 123                 if (lag_ndev) {
 124                         upper = netdev_master_upper_dev_get(lag_ndev);
 125                         dev_put(lag_ndev);
 126                 }
 127
 128                 if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
 129                     && ibdev->ib_active) {
 130                         struct ib_event ibev = { };
 131
 132                         ibev.device = &ibdev->ib_dev;
 133                         ibev.event = (event == NETDEV_UP) ?
 134                                      IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
 135                         ibev.element.port_num = 1;
 136                         ib_dispatch_event(&ibev);
 137                 }
 138                 break;
 139         }
 140
 141         default:
 142                 break;
 143         }
 144
 145         return NOTIFY_DONE;
 146 }
 147
 148 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
 149                                              u8 port_num)
 150 {
 151         struct mlx5_ib_dev *ibdev = to_mdev(device);
 152         struct net_device *ndev;
 153
 154         ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
 155         if (ndev)
 156                 return ndev;
 157
 158         /* Ensure ndev does not disappear before we invoke dev_hold()
 159          */
 160         read_lock(&ibdev->roce.netdev_lock);
 161         ndev = ibdev->roce.netdev;
 162         if (ndev)
 163                 dev_hold(ndev);
 164         read_unlock(&ibdev->roce.netdev_lock);
 165
 166         return ndev;
 167 }
 168
 169 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,
 170                                 struct ib_port_attr *props)
 171 {
 172         struct mlx5_ib_dev *dev = to_mdev(device);
 173         struct net_device *ndev, *upper;
 174         enum ib_mtu ndev_ib_mtu;
 175         u16 qkey_viol_cntr;
 176
 177         memset(props, 0, sizeof(*props));
 178
 179         props->port_cap_flags  |= IB_PORT_CM_SUP;
 180         props->port_cap_flags  |= IB_PORT_IP_BASED_GIDS;
 181
 182         props->gid_tbl_len      = MLX5_CAP_ROCE(dev->mdev,
 183                                                 roce_address_table_size);
 184         props->max_mtu          = IB_MTU_4096;
 185         props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
 186         props->pkey_tbl_len     = 1;
 187         props->state            = IB_PORT_DOWN;
 188         props->phys_state       = 3;
 189
 190         mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr);
 191         props->qkey_viol_cntr = qkey_viol_cntr;
 192
 193         ndev = mlx5_ib_get_netdev(device, port_num);
 194         if (!ndev)
 195                 return 0;
 196
 197         if (mlx5_lag_is_active(dev->mdev)) {
 198                 rcu_read_lock();
 199                 upper = netdev_master_upper_dev_get_rcu(ndev);
 200                 if (upper) {
 201                         dev_put(ndev);
 202                         ndev = upper;
 203                         dev_hold(ndev);
 204                 }
 205                 rcu_read_unlock();
 206         }
 207
 208         if (netif_running(ndev) && netif_carrier_ok(ndev)) {
 209                 props->state      = IB_PORT_ACTIVE;
 210                 props->phys_state = 5;
 211         }
 212
 213         ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
 214
 215         dev_put(ndev);
 216
 217         props->active_mtu       = min(props->max_mtu, ndev_ib_mtu);
 218
 219         props->active_width     = IB_WIDTH_4X;  /* TODO */
 220         props->active_speed     = IB_SPEED_QDR; /* TODO */
 221
 222         return 0;
 223 }
 224
 225 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid,
 226                                      const struct ib_gid_attr *attr,
 227                                      void *mlx5_addr)
 228 {
 229 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v)
 230         char *mlx5_addr_l3_addr = MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
 231                                                source_l3_address);
 232         void *mlx5_addr_mac     = MLX5_ADDR_OF(roce_addr_layout, mlx5_addr,
 233                                                source_mac_47_32);
 234
 235         if (!gid)
 236                 return;
 237
 238         ether_addr_copy(mlx5_addr_mac, attr->ndev->dev_addr);
 239
 240         if (is_vlan_dev(attr->ndev)) {
 241                 MLX5_SET_RA(mlx5_addr, vlan_valid, 1);
 242                 MLX5_SET_RA(mlx5_addr, vlan_id, vlan_dev_vlan_id(attr->ndev));
 243         }
 244
 245         switch (attr->gid_type) {
 246         case IB_GID_TYPE_IB:
 247                 MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1);
 248                 break;
 249         case IB_GID_TYPE_ROCE_UDP_ENCAP:
 250                 MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2);
 251                 break;
 252
 253         default:
 254                 WARN_ON(true);
 255         }
 256
 257         if (attr->gid_type != IB_GID_TYPE_IB) {
 258                 if (ipv6_addr_v4mapped((void *)gid))
 259                         MLX5_SET_RA(mlx5_addr, roce_l3_type,
 260                                     MLX5_ROCE_L3_TYPE_IPV4);
 261                 else
 262                         MLX5_SET_RA(mlx5_addr, roce_l3_type,
 263                                     MLX5_ROCE_L3_TYPE_IPV6);
 264         }
 265
 266         if ((attr->gid_type == IB_GID_TYPE_IB) ||
 267             !ipv6_addr_v4mapped((void *)gid))
 268                 memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid));
 269         else
 270                 memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4);
 271 }
 272
 273 static int set_roce_addr(struct ib_device *device, u8 port_num,
 274                          unsigned int index,
 275                          const union ib_gid *gid,
 276                          const struct ib_gid_attr *attr)
 277 {
 278         struct mlx5_ib_dev *dev = to_mdev(device);
 279         u32  in[MLX5_ST_SZ_DW(set_roce_address_in)]  = {0};
 280         u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0};
 281         void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address);
 282         enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num);
 283
 284         if (ll != IB_LINK_LAYER_ETHERNET)
 285                 return -EINVAL;
 286
 287         ib_gid_to_mlx5_roce_addr(gid, attr, in_addr);
 288
 289         MLX5_SET(set_roce_address_in, in, roce_address_index, index);
 290         MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS);
 291         return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
 292 }
 293
 294 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num,
 295                            unsigned int index, const union ib_gid *gid,
 296                            const struct ib_gid_attr *attr,
 297                            __always_unused void **context)
 298 {
 299         return set_roce_addr(device, port_num, index, gid, attr);
 300 }
 301
 302 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num,
 303                            unsigned int index, __always_unused void **context)
 304 {
 305         return set_roce_addr(device, port_num, index, NULL, NULL);
 306 }
 307
 308 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num,
 309                                int index)
 310 {
 311         struct ib_gid_attr attr;
 312         union ib_gid gid;
 313
 314         if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr))
 315                 return 0;
 316
 317         if (!attr.ndev)
 318                 return 0;
 319
 320         dev_put(attr.ndev);
 321
 322         if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
 323                 return 0;
 324
 325         return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
 326 }
 327
 328 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
 329 {
 330         if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
 331                 return !MLX5_CAP_GEN(dev->mdev, ib_virt);
 332         return 0;
 333 }
 334
 335 enum {
 336         MLX5_VPORT_ACCESS_METHOD_MAD,
 337         MLX5_VPORT_ACCESS_METHOD_HCA,
 338         MLX5_VPORT_ACCESS_METHOD_NIC,
 339 };
 340
 341 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
 342 {
 343         if (mlx5_use_mad_ifc(to_mdev(ibdev)))
 344                 return MLX5_VPORT_ACCESS_METHOD_MAD;
 345
 346         if (mlx5_ib_port_link_layer(ibdev, 1) ==
 347             IB_LINK_LAYER_ETHERNET)
 348                 return MLX5_VPORT_ACCESS_METHOD_NIC;
 349
 350         return MLX5_VPORT_ACCESS_METHOD_HCA;
 351 }
 352
 353 static void get_atomic_caps(struct mlx5_ib_dev *dev,
 354                             struct ib_device_attr *props)
 355 {
 356         u8 tmp;
 357         u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
 358         u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
 359         u8 atomic_req_8B_endianness_mode =
 360                 MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode);
 361
 362         /* Check if HW supports 8 bytes standard atomic operations and capable
 363          * of host endianness respond
 364          */
 365         tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
 366         if (((atomic_operations & tmp) == tmp) &&
 367             (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
 368             (atomic_req_8B_endianness_mode)) {
 369                 props->atomic_cap = IB_ATOMIC_HCA;
 370         } else {
 371                 props->atomic_cap = IB_ATOMIC_NONE;
 372         }
 373 }
 374
 375 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
 376                                         __be64 *sys_image_guid)
 377 {
 378         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 379         struct mlx5_core_dev *mdev = dev->mdev;
 380         u64 tmp;
 381         int err;
 382
 383         switch (mlx5_get_vport_access_method(ibdev)) {
 384         case MLX5_VPORT_ACCESS_METHOD_MAD:
 385                 return mlx5_query_mad_ifc_system_image_guid(ibdev,
 386                                                             sys_image_guid);
 387
 388         case MLX5_VPORT_ACCESS_METHOD_HCA:
 389                 err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
 390                 break;
 391
 392         case MLX5_VPORT_ACCESS_METHOD_NIC:
 393                 err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
 394                 break;
 395
 396         default:
 397                 return -EINVAL;
 398         }
 399
 400         if (!err)
 401                 *sys_image_guid = cpu_to_be64(tmp);
 402
 403         return err;
 404
 405 }
 406
 407 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
 408                                 u16 *max_pkeys)
 409 {
 410         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 411         struct mlx5_core_dev *mdev = dev->mdev;
 412
 413         switch (mlx5_get_vport_access_method(ibdev)) {
 414         case MLX5_VPORT_ACCESS_METHOD_MAD:
 415                 return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
 416
 417         case MLX5_VPORT_ACCESS_METHOD_HCA:
 418         case MLX5_VPORT_ACCESS_METHOD_NIC:
 419                 *max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
 420                                                 pkey_table_size));
 421                 return 0;
 422
 423         default:
 424                 return -EINVAL;
 425         }
 426 }
 427
 428 static int mlx5_query_vendor_id(struct ib_device *ibdev,
 429                                 u32 *vendor_id)
 430 {
 431         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 432
 433         switch (mlx5_get_vport_access_method(ibdev)) {
 434         case MLX5_VPORT_ACCESS_METHOD_MAD:
 435                 return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
 436
 437         case MLX5_VPORT_ACCESS_METHOD_HCA:
 438         case MLX5_VPORT_ACCESS_METHOD_NIC:
 439                 return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
 440
 441         default:
 442                 return -EINVAL;
 443         }
 444 }
 445
 446 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
 447                                 __be64 *node_guid)
 448 {
 449         u64 tmp;
 450         int err;
 451
 452         switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
 453         case MLX5_VPORT_ACCESS_METHOD_MAD:
 454                 return mlx5_query_mad_ifc_node_guid(dev, node_guid);
 455
 456         case MLX5_VPORT_ACCESS_METHOD_HCA:
 457                 err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
 458                 break;
 459
 460         case MLX5_VPORT_ACCESS_METHOD_NIC:
 461                 err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
 462                 break;
 463
 464         default:
 465                 return -EINVAL;
 466         }
 467
 468         if (!err)
 469                 *node_guid = cpu_to_be64(tmp);
 470
 471         return err;
 472 }
 473
 474 struct mlx5_reg_node_desc {
 475         u8      desc[IB_DEVICE_NODE_DESC_MAX];
 476 };
 477
 478 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
 479 {
 480         struct mlx5_reg_node_desc in;
 481
 482         if (mlx5_use_mad_ifc(dev))
 483                 return mlx5_query_mad_ifc_node_desc(dev, node_desc);
 484
 485         memset(&in, 0, sizeof(in));
 486
 487         return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
 488                                     sizeof(struct mlx5_reg_node_desc),
 489                                     MLX5_REG_NODE_DESC, 0, 0);
 490 }
 491
 492 static int mlx5_ib_query_device(struct ib_device *ibdev,
 493                                 struct ib_device_attr *props,
 494                                 struct ib_udata *uhw)
 495 {
 496         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 497         struct mlx5_core_dev *mdev = dev->mdev;
 498         int err = -ENOMEM;
 499         int max_sq_desc;
 500         int max_rq_sg;
 501         int max_sq_sg;
 502         u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
 503         struct mlx5_ib_query_device_resp resp = {};
 504         size_t resp_len;
 505         u64 max_tso;
 506
 507         resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
 508         if (uhw->outlen && uhw->outlen < resp_len)
 509                 return -EINVAL;
 510         else
 511                 resp.response_length = resp_len;
 512
 513         if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
 514                 return -EINVAL;
 515
 516         memset(props, 0, sizeof(*props));
 517         err = mlx5_query_system_image_guid(ibdev,
 518                                            &props->sys_image_guid);
 519         if (err)
 520                 return err;
 521
 522         err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys);
 523         if (err)
 524                 return err;
 525
 526         err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
 527         if (err)
 528                 return err;
 529
 530         props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
 531                 (fw_rev_min(dev->mdev) << 16) |
 532                 fw_rev_sub(dev->mdev);
 533         props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
 534                 IB_DEVICE_PORT_ACTIVE_EVENT             |
 535                 IB_DEVICE_SYS_IMAGE_GUID                |
 536                 IB_DEVICE_RC_RNR_NAK_GEN;
 537
 538         if (MLX5_CAP_GEN(mdev, pkv))
 539                 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
 540         if (MLX5_CAP_GEN(mdev, qkv))
 541                 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
 542         if (MLX5_CAP_GEN(mdev, apm))
 543                 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
 544         if (MLX5_CAP_GEN(mdev, xrc))
 545                 props->device_cap_flags |= IB_DEVICE_XRC;
 546         if (MLX5_CAP_GEN(mdev, imaicl)) {
 547                 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
 548                                            IB_DEVICE_MEM_WINDOW_TYPE_2B;
 549                 props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
 550                 /* We support 'Gappy' memory registration too */
 551                 props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
 552         }
 553         props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
 554         if (MLX5_CAP_GEN(mdev, sho)) {
 555                 props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER;
 556                 /* At this stage no support for signature handover */
 557                 props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
 558                                       IB_PROT_T10DIF_TYPE_2 |
 559                                       IB_PROT_T10DIF_TYPE_3;
 560                 props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
 561                                        IB_GUARD_T10DIF_CSUM;
 562         }
 563         if (MLX5_CAP_GEN(mdev, block_lb_mc))
 564                 props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
 565
 566         if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) {
 567                 if (MLX5_CAP_ETH(mdev, csum_cap))
 568                         props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
 569
 570                 if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
 571                         max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
 572                         if (max_tso) {
 573                                 resp.tso_caps.max_tso = 1 << max_tso;
 574                                 resp.tso_caps.supported_qpts |=
 575                                         1 << IB_QPT_RAW_PACKET;
 576                                 resp.response_length += sizeof(resp.tso_caps);
 577                         }
 578                 }
 579
 580                 if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
 581                         resp.rss_caps.rx_hash_function =
 582                                                 MLX5_RX_HASH_FUNC_TOEPLITZ;
 583                         resp.rss_caps.rx_hash_fields_mask =
 584                                                 MLX5_RX_HASH_SRC_IPV4 |
 585                                                 MLX5_RX_HASH_DST_IPV4 |
 586                                                 MLX5_RX_HASH_SRC_IPV6 |
 587                                                 MLX5_RX_HASH_DST_IPV6 |
 588                                                 MLX5_RX_HASH_SRC_PORT_TCP |
 589                                                 MLX5_RX_HASH_DST_PORT_TCP |
 590                                                 MLX5_RX_HASH_SRC_PORT_UDP |
 591                                                 MLX5_RX_HASH_DST_PORT_UDP;
 592                         resp.response_length += sizeof(resp.rss_caps);
 593                 }
 594         } else {
 595                 if (field_avail(typeof(resp), tso_caps, uhw->outlen))
 596                         resp.response_length += sizeof(resp.tso_caps);
 597                 if (field_avail(typeof(resp), rss_caps, uhw->outlen))
 598                         resp.response_length += sizeof(resp.rss_caps);
 599         }
 600
 601         if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
 602                 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
 603                 props->device_cap_flags |= IB_DEVICE_UD_TSO;
 604         }
 605
 606         if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
 607             MLX5_CAP_ETH(dev->mdev, scatter_fcs))
 608                 props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
 609
 610         if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
 611                 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
 612
 613         props->vendor_part_id      = mdev->pdev->device;
 614         props->hw_ver              = mdev->pdev->revision;
 615
 616         props->max_mr_size         = ~0ull;
 617         props->page_size_cap       = ~(min_page_size - 1);
 618         props->max_qp              = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
 619         props->max_qp_wr           = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
 620         max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
 621                      sizeof(struct mlx5_wqe_data_seg);
 622         max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
 623         max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
 624                      sizeof(struct mlx5_wqe_raddr_seg)) /
 625                 sizeof(struct mlx5_wqe_data_seg);
 626         props->max_sge = min(max_rq_sg, max_sq_sg);
 627         props->max_sge_rd          = MLX5_MAX_SGE_RD;
 628         props->max_cq              = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
 629         props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
 630         props->max_mr              = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
 631         props->max_pd              = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
 632         props->max_qp_rd_atom      = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
 633         props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
 634         props->max_srq             = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
 635         props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
 636         props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
 637         props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
 638         props->max_srq_sge         = max_rq_sg - 1;
 639         props->max_fast_reg_page_list_len =
 640                 1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
 641         get_atomic_caps(dev, props);
 642         props->masked_atomic_cap   = IB_ATOMIC_NONE;
 643         props->max_mcast_grp       = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
 644         props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
 645         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
 646                                            props->max_mcast_grp;
 647         props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */
 648         props->max_ah = INT_MAX;
 649         props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
 650         props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
 651
 652 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 653         if (MLX5_CAP_GEN(mdev, pg))
 654                 props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
 655         props->odp_caps = dev->odp_caps;
 656 #endif
 657
 658         if (MLX5_CAP_GEN(mdev, cd))
 659                 props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
 660
 661         if (!mlx5_core_is_pf(mdev))
 662                 props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
 663
 664         if (mlx5_ib_port_link_layer(ibdev, 1) ==
 665             IB_LINK_LAYER_ETHERNET) {
 666                 props->rss_caps.max_rwq_indirection_tables =
 667                         1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
 668                 props->rss_caps.max_rwq_indirection_table_size =
 669                         1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
 670                 props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
 671                 props->max_wq_type_rq =
 672                         1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
 673         }
 674
 675         if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
 676                 resp.cqe_comp_caps.max_num =
 677                         MLX5_CAP_GEN(dev->mdev, cqe_compression) ?
 678                         MLX5_CAP_GEN(dev->mdev, cqe_compression_max_num) : 0;
 679                 resp.cqe_comp_caps.supported_format =
 680                         MLX5_IB_CQE_RES_FORMAT_HASH |
 681                         MLX5_IB_CQE_RES_FORMAT_CSUM;
 682                 resp.response_length += sizeof(resp.cqe_comp_caps);
 683         }
 684
 685         if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen)) {
 686                 if (MLX5_CAP_QOS(mdev, packet_pacing) &&
 687                     MLX5_CAP_GEN(mdev, qos)) {
 688                         resp.packet_pacing_caps.qp_rate_limit_max =
 689                                 MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
 690                         resp.packet_pacing_caps.qp_rate_limit_min =
 691                                 MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
 692                         resp.packet_pacing_caps.supported_qpts |=
 693                                 1 << IB_QPT_RAW_PACKET;
 694                 }
 695                 resp.response_length += sizeof(resp.packet_pacing_caps);
 696         }
 697
 698         if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
 699                         uhw->outlen)) {
 700                 resp.mlx5_ib_support_multi_pkt_send_wqes =
 701                         MLX5_CAP_ETH(mdev, multi_pkt_send_wqe);
 702                 resp.response_length +=
 703                         sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
 704         }
 705
 706         if (field_avail(typeof(resp), reserved, uhw->outlen))
 707                 resp.response_length += sizeof(resp.reserved);
 708
 709         if (uhw->outlen) {
 710                 err = ib_copy_to_udata(uhw, &resp, resp.response_length);
 711
 712                 if (err)
 713                         return err;
 714         }
 715
 716         return 0;
 717 }
 718
 719 enum mlx5_ib_width {
 720         MLX5_IB_WIDTH_1X        = 1 << 0,
 721         MLX5_IB_WIDTH_2X        = 1 << 1,
 722         MLX5_IB_WIDTH_4X        = 1 << 2,
 723         MLX5_IB_WIDTH_8X        = 1 << 3,
 724         MLX5_IB_WIDTH_12X       = 1 << 4
 725 };
 726
 727 static int translate_active_width(struct ib_device *ibdev, u8 active_width,
 728                                   u8 *ib_width)
 729 {
 730         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 731         int err = 0;
 732
 733         if (active_width & MLX5_IB_WIDTH_1X) {
 734                 *ib_width = IB_WIDTH_1X;
 735         } else if (active_width & MLX5_IB_WIDTH_2X) {
 736                 mlx5_ib_dbg(dev, "active_width %d is not supported by IB spec\n",
 737                             (int)active_width);
 738                 err = -EINVAL;
 739         } else if (active_width & MLX5_IB_WIDTH_4X) {
 740                 *ib_width = IB_WIDTH_4X;
 741         } else if (active_width & MLX5_IB_WIDTH_8X) {
 742                 *ib_width = IB_WIDTH_8X;
 743         } else if (active_width & MLX5_IB_WIDTH_12X) {
 744                 *ib_width = IB_WIDTH_12X;
 745         } else {
 746                 mlx5_ib_dbg(dev, "Invalid active_width %d\n",
 747                             (int)active_width);
 748                 err = -EINVAL;
 749         }
 750
 751         return err;
 752 }
 753
 754 static int mlx5_mtu_to_ib_mtu(int mtu)
 755 {
 756         switch (mtu) {
 757         case 256: return 1;
 758         case 512: return 2;
 759         case 1024: return 3;
 760         case 2048: return 4;
 761         case 4096: return 5;
 762         default:
 763                 pr_warn("invalid mtu\n");
 764                 return -1;
 765         }
 766 }
 767
 768 enum ib_max_vl_num {
 769         __IB_MAX_VL_0           = 1,
 770         __IB_MAX_VL_0_1         = 2,
 771         __IB_MAX_VL_0_3         = 3,
 772         __IB_MAX_VL_0_7         = 4,
 773         __IB_MAX_VL_0_14        = 5,
 774 };
 775
 776 enum mlx5_vl_hw_cap {
 777         MLX5_VL_HW_0    = 1,
 778         MLX5_VL_HW_0_1  = 2,
 779         MLX5_VL_HW_0_2  = 3,
 780         MLX5_VL_HW_0_3  = 4,
 781         MLX5_VL_HW_0_4  = 5,
 782         MLX5_VL_HW_0_5  = 6,
 783         MLX5_VL_HW_0_6  = 7,
 784         MLX5_VL_HW_0_7  = 8,
 785         MLX5_VL_HW_0_14 = 15
 786 };
 787
 788 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
 789                                 u8 *max_vl_num)
 790 {
 791         switch (vl_hw_cap) {
 792         case MLX5_VL_HW_0:
 793                 *max_vl_num = __IB_MAX_VL_0;
 794                 break;
 795         case MLX5_VL_HW_0_1:
 796                 *max_vl_num = __IB_MAX_VL_0_1;
 797                 break;
 798         case MLX5_VL_HW_0_3:
 799                 *max_vl_num = __IB_MAX_VL_0_3;
 800                 break;
 801         case MLX5_VL_HW_0_7:
 802                 *max_vl_num = __IB_MAX_VL_0_7;
 803                 break;
 804         case MLX5_VL_HW_0_14:
 805                 *max_vl_num = __IB_MAX_VL_0_14;
 806                 break;
 807
 808         default:
 809                 return -EINVAL;
 810         }
 811
 812         return 0;
 813 }
 814
 815 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port,
 816                                struct ib_port_attr *props)
 817 {
 818         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 819         struct mlx5_core_dev *mdev = dev->mdev;
 820         struct mlx5_hca_vport_context *rep;
 821         u16 max_mtu;
 822         u16 oper_mtu;
 823         int err;
 824         u8 ib_link_width_oper;
 825         u8 vl_hw_cap;
 826
 827         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
 828         if (!rep) {
 829                 err = -ENOMEM;
 830                 goto out;
 831         }
 832
 833         memset(props, 0, sizeof(*props));
 834
 835         err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
 836         if (err)
 837                 goto out;
 838
 839         props->lid              = rep->lid;
 840         props->lmc              = rep->lmc;
 841         props->sm_lid           = rep->sm_lid;
 842         props->sm_sl            = rep->sm_sl;
 843         props->state            = rep->vport_state;
 844         props->phys_state       = rep->port_physical_state;
 845         props->port_cap_flags   = rep->cap_mask1;
 846         props->gid_tbl_len      = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
 847         props->max_msg_sz       = 1 << MLX5_CAP_GEN(mdev, log_max_msg);
 848         props->pkey_tbl_len     = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
 849         props->bad_pkey_cntr    = rep->pkey_violation_counter;
 850         props->qkey_viol_cntr   = rep->qkey_violation_counter;
 851         props->subnet_timeout   = rep->subnet_timeout;
 852         props->init_type_reply  = rep->init_type_reply;
 853         props->grh_required     = rep->grh_required;
 854
 855         err = mlx5_query_port_link_width_oper(mdev, &ib_link_width_oper, port);
 856         if (err)
 857                 goto out;
 858
 859         err = translate_active_width(ibdev, ib_link_width_oper,
 860                                      &props->active_width);
 861         if (err)
 862                 goto out;
 863         err = mlx5_query_port_ib_proto_oper(mdev, &props->active_speed, port);
 864         if (err)
 865                 goto out;
 866
 867         mlx5_query_port_max_mtu(mdev, &max_mtu, port);
 868
 869         props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
 870
 871         mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
 872
 873         props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
 874
 875         err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
 876         if (err)
 877                 goto out;
 878
 879         err = translate_max_vl_num(ibdev, vl_hw_cap,
 880                                    &props->max_vl_num);
 881 out:
 882         kfree(rep);
 883         return err;
 884 }
 885
 886 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port,
 887                        struct ib_port_attr *props)
 888 {
 889         switch (mlx5_get_vport_access_method(ibdev)) {
 890         case MLX5_VPORT_ACCESS_METHOD_MAD:
 891                 return mlx5_query_mad_ifc_port(ibdev, port, props);
 892
 893         case MLX5_VPORT_ACCESS_METHOD_HCA:
 894                 return mlx5_query_hca_port(ibdev, port, props);
 895
 896         case MLX5_VPORT_ACCESS_METHOD_NIC:
 897                 return mlx5_query_port_roce(ibdev, port, props);
 898
 899         default:
 900                 return -EINVAL;
 901         }
 902 }
 903
 904 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
 905                              union ib_gid *gid)
 906 {
 907         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 908         struct mlx5_core_dev *mdev = dev->mdev;
 909
 910         switch (mlx5_get_vport_access_method(ibdev)) {
 911         case MLX5_VPORT_ACCESS_METHOD_MAD:
 912                 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
 913
 914         case MLX5_VPORT_ACCESS_METHOD_HCA:
 915                 return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
 916
 917         default:
 918                 return -EINVAL;
 919         }
 920
 921 }
 922
 923 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
 924                               u16 *pkey)
 925 {
 926         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 927         struct mlx5_core_dev *mdev = dev->mdev;
 928
 929         switch (mlx5_get_vport_access_method(ibdev)) {
 930         case MLX5_VPORT_ACCESS_METHOD_MAD:
 931                 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
 932
 933         case MLX5_VPORT_ACCESS_METHOD_HCA:
 934         case MLX5_VPORT_ACCESS_METHOD_NIC:
 935                 return mlx5_query_hca_vport_pkey(mdev, 0, port,  0, index,
 936                                                  pkey);
 937         default:
 938                 return -EINVAL;
 939         }
 940 }
 941
 942 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
 943                                  struct ib_device_modify *props)
 944 {
 945         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 946         struct mlx5_reg_node_desc in;
 947         struct mlx5_reg_node_desc out;
 948         int err;
 949
 950         if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
 951                 return -EOPNOTSUPP;
 952
 953         if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
 954                 return 0;
 955
 956         /*
 957          * If possible, pass node desc to FW, so it can generate
 958          * a 144 trap.  If cmd fails, just ignore.
 959          */
 960         memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
 961         err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
 962                                    sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
 963         if (err)
 964                 return err;
 965
 966         memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
 967
 968         return err;
 969 }
 970
 971 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
 972                                struct ib_port_modify *props)
 973 {
 974         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 975         struct ib_port_attr attr;
 976         u32 tmp;
 977         int err;
 978
 979         mutex_lock(&dev->cap_mask_mutex);
 980
 981         err = mlx5_ib_query_port(ibdev, port, &attr);
 982         if (err)
 983                 goto out;
 984
 985         tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
 986                 ~props->clr_port_cap_mask;
 987
 988         err = mlx5_set_port_caps(dev->mdev, port, tmp);
 989
 990 out:
 991         mutex_unlock(&dev->cap_mask_mutex);
 992         return err;
 993 }
 994
 995 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
 996 {
 997         mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
 998                     caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
 999 }
1000
1001 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1002                              struct mlx5_ib_alloc_ucontext_req_v2 *req,
1003                              u32 *num_sys_pages)
1004 {
1005         int uars_per_sys_page;
1006         int bfregs_per_sys_page;
1007         int ref_bfregs = req->total_num_bfregs;
1008
1009         if (req->total_num_bfregs == 0)
1010                 return -EINVAL;
1011
1012         BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1013         BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1014
1015         if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1016                 return -ENOMEM;
1017
1018         uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1019         bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1020         req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1021         *num_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1022
1023         if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1024                 return -EINVAL;
1025
1026         mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, alloated %d, using %d sys pages\n",
1027                     MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1028                     lib_uar_4k ? "yes" : "no", ref_bfregs,
1029                     req->total_num_bfregs, *num_sys_pages);
1030
1031         return 0;
1032 }
1033
1034 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1035 {
1036         struct mlx5_bfreg_info *bfregi;
1037         int err;
1038         int i;
1039
1040         bfregi = &context->bfregi;
1041         for (i = 0; i < bfregi->num_sys_pages; i++) {
1042                 err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1043                 if (err)
1044                         goto error;
1045
1046                 mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1047         }
1048         return 0;
1049
1050 error:
1051         for (--i; i >= 0; i--)
1052                 if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1053                         mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1054
1055         return err;
1056 }
1057
1058 static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1059 {
1060         struct mlx5_bfreg_info *bfregi;
1061         int err;
1062         int i;
1063
1064         bfregi = &context->bfregi;
1065         for (i = 0; i < bfregi->num_sys_pages; i++) {
1066                 err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1067                 if (err) {
1068                         mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1069                         return err;
1070                 }
1071         }
1072         return 0;
1073 }
1074
1075 static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
1076                                                   struct ib_udata *udata)
1077 {
1078         struct mlx5_ib_dev *dev = to_mdev(ibdev);
1079         struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1080         struct mlx5_ib_alloc_ucontext_resp resp = {};
1081         struct mlx5_ib_ucontext *context;
1082         struct mlx5_bfreg_info *bfregi;
1083         int ver;
1084         int err;
1085         size_t reqlen;
1086         size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1087                                      max_cqe_version);
1088         bool lib_uar_4k;
1089
1090         if (!dev->ib_active)
1091                 return ERR_PTR(-EAGAIN);
1092
1093         if (udata->inlen < sizeof(struct ib_uverbs_cmd_hdr))
1094                 return ERR_PTR(-EINVAL);
1095
1096         reqlen = udata->inlen - sizeof(struct ib_uverbs_cmd_hdr);
1097         if (reqlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1098                 ver = 0;
1099         else if (reqlen >= min_req_v2)
1100                 ver = 2;
1101         else
1102                 return ERR_PTR(-EINVAL);
1103
1104         err = ib_copy_from_udata(&req, udata, min(reqlen, sizeof(req)));
1105         if (err)
1106                 return ERR_PTR(err);
1107
1108         if (req.flags)
1109                 return ERR_PTR(-EINVAL);
1110
1111         if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1112                 return ERR_PTR(-EOPNOTSUPP);
1113
1114         req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1115                                     MLX5_NON_FP_BFREGS_PER_UAR);
1116         if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1117                 return ERR_PTR(-EINVAL);
1118
1119         resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1120         if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf))
1121                 resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size);
1122         resp.cache_line_size = cache_line_size();
1123         resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1124         resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1125         resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1126         resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1127         resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1128         resp.cqe_version = min_t(__u8,
1129                                  (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1130                                  req.max_cqe_version);
1131         resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1132                                 MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1133         resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1134                                         MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1;
1135         resp.response_length = min(offsetof(typeof(resp), response_length) +
1136                                    sizeof(resp.response_length), udata->outlen);
1137
1138         context = kzalloc(sizeof(*context), GFP_KERNEL);
1139         if (!context)
1140                 return ERR_PTR(-ENOMEM);
1141
1142         lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1143         bfregi = &context->bfregi;
1144
1145         /* updates req->total_num_bfregs */
1146         err = calc_total_bfregs(dev, lib_uar_4k, &req, &bfregi->num_sys_pages);
1147         if (err)
1148                 goto out_ctx;
1149
1150         mutex_init(&bfregi->lock);
1151         bfregi->lib_uar_4k = lib_uar_4k;
1152         bfregi->count = kcalloc(req.total_num_bfregs, sizeof(*bfregi->count),
1153                                 GFP_KERNEL);
1154         if (!bfregi->count) {
1155                 err = -ENOMEM;
1156                 goto out_ctx;
1157         }
1158
1159         bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1160                                     sizeof(*bfregi->sys_pages),
1161                                     GFP_KERNEL);
1162         if (!bfregi->sys_pages) {
1163                 err = -ENOMEM;
1164                 goto out_count;
1165         }
1166
1167         err = allocate_uars(dev, context);
1168         if (err)
1169                 goto out_sys_pages;
1170
1171 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
1172         context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range;
1173 #endif
1174
1175         context->upd_xlt_page = __get_free_page(GFP_KERNEL);
1176         if (!context->upd_xlt_page) {
1177                 err = -ENOMEM;
1178                 goto out_uars;
1179         }
1180         mutex_init(&context->upd_xlt_page_mutex);
1181
1182         if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) {
1183                 err = mlx5_core_alloc_transport_domain(dev->mdev,
1184                                                        &context->tdn);
1185                 if (err)
1186                         goto out_page;
1187         }
1188
1189         INIT_LIST_HEAD(&context->vma_private_list);
1190         INIT_LIST_HEAD(&context->db_page_list);
1191         mutex_init(&context->db_page_mutex);
1192
1193         resp.tot_bfregs = req.total_num_bfregs;
1194         resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports);
1195
1196         if (field_avail(typeof(resp), cqe_version, udata->outlen))
1197                 resp.response_length += sizeof(resp.cqe_version);
1198
1199         if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) {
1200                 resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1201                                       MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1202                 resp.response_length += sizeof(resp.cmds_supp_uhw);
1203         }
1204
1205         /*
1206          * We don't want to expose information from the PCI bar that is located
1207          * after 4096 bytes, so if the arch only supports larger pages, let's
1208          * pretend we don't support reading the HCA's core clock. This is also
1209          * forced by mmap function.
1210          */
1211         if (field_avail(typeof(resp), hca_core_clock_offset, udata->outlen)) {
1212                 if (PAGE_SIZE <= 4096) {
1213                         resp.comp_mask |=
1214                                 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1215                         resp.hca_core_clock_offset =
1216                                 offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE;
1217                 }
1218                 resp.response_length += sizeof(resp.hca_core_clock_offset) +
1219                                         sizeof(resp.reserved2);
1220         }
1221
1222         if (field_avail(typeof(resp), log_uar_size, udata->outlen))
1223                 resp.response_length += sizeof(resp.log_uar_size);
1224
1225         if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
1226                 resp.response_length += sizeof(resp.num_uars_per_page);
1227
1228         err = ib_copy_to_udata(udata, &resp, resp.response_length);
1229         if (err)
1230                 goto out_td;
1231
1232         bfregi->ver = ver;
1233         bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1234         context->cqe_version = resp.cqe_version;
1235         context->lib_caps = req.lib_caps;
1236         print_lib_caps(dev, context->lib_caps);
1237
1238         return &context->ibucontext;
1239
1240 out_td:
1241         if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1242                 mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1243
1244 out_page:
1245         free_page(context->upd_xlt_page);
1246
1247 out_uars:
1248         deallocate_uars(dev, context);
1249
1250 out_sys_pages:
1251         kfree(bfregi->sys_pages);
1252
1253 out_count:
1254         kfree(bfregi->count);
1255
1256 out_ctx:
1257         kfree(context);
1258
1259         return ERR_PTR(err);
1260 }
1261
1262 static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1263 {
1264         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1265         struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1266         struct mlx5_bfreg_info *bfregi;
1267
1268         bfregi = &context->bfregi;
1269         if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1270                 mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn);
1271
1272         free_page(context->upd_xlt_page);
1273         deallocate_uars(dev, context);
1274         kfree(bfregi->sys_pages);
1275         kfree(bfregi->count);
1276         kfree(context);
1277
1278         return 0;
1279 }
1280
1281 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
1282                                  struct mlx5_bfreg_info *bfregi,
1283                                  int idx)
1284 {
1285         int fw_uars_per_page;
1286
1287         fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
1288
1289         return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) +
1290                         bfregi->sys_pages[idx] / fw_uars_per_page;
1291 }
1292
1293 static int get_command(unsigned long offset)
1294 {
1295         return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
1296 }
1297
1298 static int get_arg(unsigned long offset)
1299 {
1300         return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
1301 }
1302
1303 static int get_index(unsigned long offset)
1304 {
1305         return get_arg(offset);
1306 }
1307
1308 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
1309 {
1310         /* vma_open is called when a new VMA is created on top of our VMA.  This
1311          * is done through either mremap flow or split_vma (usually due to
1312          * mlock, madvise, munmap, etc.) We do not support a clone of the VMA,
1313          * as this VMA is strongly hardware related.  Therefore we set the
1314          * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1315          * calling us again and trying to do incorrect actions.  We assume that
1316          * the original VMA size is exactly a single page, and therefore all
1317          * "splitting" operation will not happen to it.
1318          */
1319         area->vm_ops = NULL;
1320 }
1321
1322 static void  mlx5_ib_vma_close(struct vm_area_struct *area)
1323 {
1324         struct mlx5_ib_vma_private_data *mlx5_ib_vma_priv_data;
1325
1326         /* It's guaranteed that all VMAs opened on a FD are closed before the
1327          * file itself is closed, therefore no sync is needed with the regular
1328          * closing flow. (e.g. mlx5 ib_dealloc_ucontext)
1329          * However need a sync with accessing the vma as part of
1330          * mlx5_ib_disassociate_ucontext.
1331          * The close operation is usually called under mm->mmap_sem except when
1332          * process is exiting.
1333          * The exiting case is handled explicitly as part of
1334          * mlx5_ib_disassociate_ucontext.
1335          */
1336         mlx5_ib_vma_priv_data = (struct mlx5_ib_vma_private_data *)area->vm_private_data;
1337
1338         /* setting the vma context pointer to null in the mlx5_ib driver's
1339          * private data, to protect a race condition in
1340          * mlx5_ib_disassociate_ucontext().
1341          */
1342         mlx5_ib_vma_priv_data->vma = NULL;
1343         list_del(&mlx5_ib_vma_priv_data->list);
1344         kfree(mlx5_ib_vma_priv_data);
1345 }
1346
1347 static const struct vm_operations_struct mlx5_ib_vm_ops = {
1348         .open = mlx5_ib_vma_open,
1349         .close = mlx5_ib_vma_close
1350 };
1351
1352 static int mlx5_ib_set_vma_data(struct vm_area_struct *vma,
1353                                 struct mlx5_ib_ucontext *ctx)
1354 {
1355         struct mlx5_ib_vma_private_data *vma_prv;
1356         struct list_head *vma_head = &ctx->vma_private_list;
1357
1358         vma_prv = kzalloc(sizeof(*vma_prv), GFP_KERNEL);
1359         if (!vma_prv)
1360                 return -ENOMEM;
1361
1362         vma_prv->vma = vma;
1363         vma->vm_private_data = vma_prv;
1364         vma->vm_ops =  &mlx5_ib_vm_ops;
1365
1366         list_add(&vma_prv->list, vma_head);
1367
1368         return 0;
1369 }
1370
1371 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1372 {
1373         int ret;
1374         struct vm_area_struct *vma;
1375         struct mlx5_ib_vma_private_data *vma_private, *n;
1376         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1377         struct task_struct *owning_process  = NULL;
1378         struct mm_struct   *owning_mm       = NULL;
1379
1380         owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1381         if (!owning_process)
1382                 return;
1383
1384         owning_mm = get_task_mm(owning_process);
1385         if (!owning_mm) {
1386                 pr_info("no mm, disassociate ucontext is pending task termination\n");
1387                 while (1) {
1388                         put_task_struct(owning_process);
1389                         usleep_range(1000, 2000);
1390                         owning_process = get_pid_task(ibcontext->tgid,
1391                                                       PIDTYPE_PID);
1392                         if (!owning_process ||
1393                             owning_process->state == TASK_DEAD) {
1394                                 pr_info("disassociate ucontext done, task was terminated\n");
1395                                 /* in case task was dead need to release the
1396                                  * task struct.
1397                                  */
1398                                 if (owning_process)
1399                                         put_task_struct(owning_process);
1400                                 return;
1401                         }
1402                 }
1403         }
1404
1405         /* need to protect from a race on closing the vma as part of
1406          * mlx5_ib_vma_close.
1407          */
1408         down_read(&owning_mm->mmap_sem);
1409         list_for_each_entry_safe(vma_private, n, &context->vma_private_list,
1410                                  list) {
1411                 vma = vma_private->vma;
1412                 ret = zap_vma_ptes(vma, vma->vm_start,
1413                                    PAGE_SIZE);
1414                 WARN_ONCE(ret, "%s: zap_vma_ptes failed", __func__);
1415                 /* context going to be destroyed, should
1416                  * not access ops any more.
1417                  */
1418                 vma->vm_ops = NULL;
1419                 list_del(&vma_private->list);
1420                 kfree(vma_private);
1421         }
1422         up_read(&owning_mm->mmap_sem);
1423         mmput(owning_mm);
1424         put_task_struct(owning_process);
1425 }
1426
1427 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
1428 {
1429         switch (cmd) {
1430         case MLX5_IB_MMAP_WC_PAGE:
1431                 return "WC";
1432         case MLX5_IB_MMAP_REGULAR_PAGE:
1433                 return "best effort WC";
1434         case MLX5_IB_MMAP_NC_PAGE:
1435                 return "NC";
1436         default:
1437                 return NULL;
1438         }
1439 }
1440
1441 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
1442                     struct vm_area_struct *vma,
1443                     struct mlx5_ib_ucontext *context)
1444 {
1445         struct mlx5_bfreg_info *bfregi = &context->bfregi;
1446         int err;
1447         unsigned long idx;
1448         phys_addr_t pfn, pa;
1449         pgprot_t prot;
1450         int uars_per_page;
1451
1452         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1453                 return -EINVAL;
1454
1455         uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
1456         idx = get_index(vma->vm_pgoff);
1457         if (idx % uars_per_page ||
1458             idx * uars_per_page >= bfregi->num_sys_pages) {
1459                 mlx5_ib_warn(dev, "invalid uar index %lu\n", idx);
1460                 return -EINVAL;
1461         }
1462
1463         switch (cmd) {
1464         case MLX5_IB_MMAP_WC_PAGE:
1465 /* Some architectures don't support WC memory */
1466 #if defined(CONFIG_X86)
1467                 if (!pat_enabled())
1468                         return -EPERM;
1469 #elif !(defined(CONFIG_PPC) || (defined(CONFIG_ARM) && defined(CONFIG_MMU)))
1470                         return -EPERM;
1471 #endif
1472         /* fall through */
1473         case MLX5_IB_MMAP_REGULAR_PAGE:
1474                 /* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
1475                 prot = pgprot_writecombine(vma->vm_page_prot);
1476                 break;
1477         case MLX5_IB_MMAP_NC_PAGE:
1478                 prot = pgprot_noncached(vma->vm_page_prot);
1479                 break;
1480         default:
1481                 return -EINVAL;
1482         }
1483
1484         pfn = uar_index2pfn(dev, bfregi, idx);
1485         mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
1486
1487         vma->vm_page_prot = prot;
1488         err = io_remap_pfn_range(vma, vma->vm_start, pfn,
1489                                  PAGE_SIZE, vma->vm_page_prot);
1490         if (err) {
1491                 mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
1492                             err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
1493                 return -EAGAIN;
1494         }
1495
1496         pa = pfn << PAGE_SHIFT;
1497         mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
1498                     vma->vm_start, &pa);
1499
1500         return mlx5_ib_set_vma_data(vma, context);
1501 }
1502
1503 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
1504 {
1505         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1506         struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1507         unsigned long command;
1508         phys_addr_t pfn;
1509
1510         command = get_command(vma->vm_pgoff);
1511         switch (command) {
1512         case MLX5_IB_MMAP_WC_PAGE:
1513         case MLX5_IB_MMAP_NC_PAGE:
1514         case MLX5_IB_MMAP_REGULAR_PAGE:
1515                 return uar_mmap(dev, command, vma, context);
1516
1517         case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
1518                 return -ENOSYS;
1519
1520         case MLX5_IB_MMAP_CORE_CLOCK:
1521                 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1522                         return -EINVAL;
1523
1524                 if (vma->vm_flags & VM_WRITE)
1525                         return -EPERM;
1526
1527                 /* Don't expose to user-space information it shouldn't have */
1528                 if (PAGE_SIZE > 4096)
1529                         return -EOPNOTSUPP;
1530
1531                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1532                 pfn = (dev->mdev->iseg_base +
1533                        offsetof(struct mlx5_init_seg, internal_timer_h)) >>
1534                         PAGE_SHIFT;
1535                 if (io_remap_pfn_range(vma, vma->vm_start, pfn,
1536                                        PAGE_SIZE, vma->vm_page_prot))
1537                         return -EAGAIN;
1538
1539                 mlx5_ib_dbg(dev, "mapped internal timer at 0x%lx, PA 0x%llx\n",
1540                             vma->vm_start,
1541                             (unsigned long long)pfn << PAGE_SHIFT);
1542                 break;
1543
1544         default:
1545                 return -EINVAL;
1546         }
1547
1548         return 0;
1549 }
1550
1551 static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
1552                                       struct ib_ucontext *context,
1553                                       struct ib_udata *udata)
1554 {
1555         struct mlx5_ib_alloc_pd_resp resp;
1556         struct mlx5_ib_pd *pd;
1557         int err;
1558
1559         pd = kmalloc(sizeof(*pd), GFP_KERNEL);
1560         if (!pd)
1561                 return ERR_PTR(-ENOMEM);
1562
1563         err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn);
1564         if (err) {
1565                 kfree(pd);
1566                 return ERR_PTR(err);
1567         }
1568
1569         if (context) {
1570                 resp.pdn = pd->pdn;
1571                 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
1572                         mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn);
1573                         kfree(pd);
1574                         return ERR_PTR(-EFAULT);
1575                 }
1576         }
1577
1578         return &pd->ibpd;
1579 }
1580
1581 static int mlx5_ib_dealloc_pd(struct ib_pd *pd)
1582 {
1583         struct mlx5_ib_dev *mdev = to_mdev(pd->device);
1584         struct mlx5_ib_pd *mpd = to_mpd(pd);
1585
1586         mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn);
1587         kfree(mpd);
1588
1589         return 0;
1590 }
1591
1592 enum {
1593         MATCH_CRITERIA_ENABLE_OUTER_BIT,
1594         MATCH_CRITERIA_ENABLE_MISC_BIT,
1595         MATCH_CRITERIA_ENABLE_INNER_BIT
1596 };
1597
1598 #define HEADER_IS_ZERO(match_criteria, headers)                            \
1599         !(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \
1600                     0, MLX5_FLD_SZ_BYTES(fte_match_param, headers)))       \
1601
1602 static u8 get_match_criteria_enable(u32 *match_criteria)
1603 {
1604         u8 match_criteria_enable;
1605
1606         match_criteria_enable =
1607                 (!HEADER_IS_ZERO(match_criteria, outer_headers)) <<
1608                 MATCH_CRITERIA_ENABLE_OUTER_BIT;
1609         match_criteria_enable |=
1610                 (!HEADER_IS_ZERO(match_criteria, misc_parameters)) <<
1611                 MATCH_CRITERIA_ENABLE_MISC_BIT;
1612         match_criteria_enable |=
1613                 (!HEADER_IS_ZERO(match_criteria, inner_headers)) <<
1614                 MATCH_CRITERIA_ENABLE_INNER_BIT;
1615
1616         return match_criteria_enable;
1617 }
1618
1619 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val)
1620 {
1621         MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask);
1622         MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val);
1623 }
1624
1625 static void set_flow_label(void *misc_c, void *misc_v, u8 mask, u8 val,
1626                            bool inner)
1627 {
1628         if (inner) {
1629                 MLX5_SET(fte_match_set_misc,
1630                          misc_c, inner_ipv6_flow_label, mask);
1631                 MLX5_SET(fte_match_set_misc,
1632                          misc_v, inner_ipv6_flow_label, val);
1633         } else {
1634                 MLX5_SET(fte_match_set_misc,
1635                          misc_c, outer_ipv6_flow_label, mask);
1636                 MLX5_SET(fte_match_set_misc,
1637                          misc_v, outer_ipv6_flow_label, val);
1638         }
1639 }
1640
1641 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val)
1642 {
1643         MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask);
1644         MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val);
1645         MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2);
1646         MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2);
1647 }
1648
1649 #define LAST_ETH_FIELD vlan_tag
1650 #define LAST_IB_FIELD sl
1651 #define LAST_IPV4_FIELD tos
1652 #define LAST_IPV6_FIELD traffic_class
1653 #define LAST_TCP_UDP_FIELD src_port
1654 #define LAST_TUNNEL_FIELD tunnel_id
1655
1656 /* Field is the last supported field */
1657 #define FIELDS_NOT_SUPPORTED(filter, field)\
1658         memchr_inv((void *)&filter.field  +\
1659                    sizeof(filter.field), 0,\
1660                    sizeof(filter) -\
1661                    offsetof(typeof(filter), field) -\
1662                    sizeof(filter.field))
1663
1664 static int parse_flow_attr(u32 *match_c, u32 *match_v,
1665                            const union ib_flow_spec *ib_spec)
1666 {
1667         void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c,
1668                                            misc_parameters);
1669         void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v,
1670                                            misc_parameters);
1671         void *headers_c;
1672         void *headers_v;
1673
1674         if (ib_spec->type & IB_FLOW_SPEC_INNER) {
1675                 headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1676                                          inner_headers);
1677                 headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1678                                          inner_headers);
1679         } else {
1680                 headers_c = MLX5_ADDR_OF(fte_match_param, match_c,
1681                                          outer_headers);
1682                 headers_v = MLX5_ADDR_OF(fte_match_param, match_v,
1683                                          outer_headers);
1684         }
1685
1686         switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
1687         case IB_FLOW_SPEC_ETH:
1688                 if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1689                         return -ENOTSUPP;
1690
1691                 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1692                                              dmac_47_16),
1693                                 ib_spec->eth.mask.dst_mac);
1694                 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1695                                              dmac_47_16),
1696                                 ib_spec->eth.val.dst_mac);
1697
1698                 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1699                                              smac_47_16),
1700                                 ib_spec->eth.mask.src_mac);
1701                 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1702                                              smac_47_16),
1703                                 ib_spec->eth.val.src_mac);
1704
1705                 if (ib_spec->eth.mask.vlan_tag) {
1706                         MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1707                                  vlan_tag, 1);
1708                         MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1709                                  vlan_tag, 1);
1710
1711                         MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1712                                  first_vid, ntohs(ib_spec->eth.mask.vlan_tag));
1713                         MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1714                                  first_vid, ntohs(ib_spec->eth.val.vlan_tag));
1715
1716                         MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1717                                  first_cfi,
1718                                  ntohs(ib_spec->eth.mask.vlan_tag) >> 12);
1719                         MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1720                                  first_cfi,
1721                                  ntohs(ib_spec->eth.val.vlan_tag) >> 12);
1722
1723                         MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1724                                  first_prio,
1725                                  ntohs(ib_spec->eth.mask.vlan_tag) >> 13);
1726                         MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1727                                  first_prio,
1728                                  ntohs(ib_spec->eth.val.vlan_tag) >> 13);
1729                 }
1730                 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1731                          ethertype, ntohs(ib_spec->eth.mask.ether_type));
1732                 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1733                          ethertype, ntohs(ib_spec->eth.val.ether_type));
1734                 break;
1735         case IB_FLOW_SPEC_IPV4:
1736                 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1737                         return -ENOTSUPP;
1738
1739                 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1740                          ethertype, 0xffff);
1741                 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1742                          ethertype, ETH_P_IP);
1743
1744                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1745                                     src_ipv4_src_ipv6.ipv4_layout.ipv4),
1746                        &ib_spec->ipv4.mask.src_ip,
1747                        sizeof(ib_spec->ipv4.mask.src_ip));
1748                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1749                                     src_ipv4_src_ipv6.ipv4_layout.ipv4),
1750                        &ib_spec->ipv4.val.src_ip,
1751                        sizeof(ib_spec->ipv4.val.src_ip));
1752                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1753                                     dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1754                        &ib_spec->ipv4.mask.dst_ip,
1755                        sizeof(ib_spec->ipv4.mask.dst_ip));
1756                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1757                                     dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
1758                        &ib_spec->ipv4.val.dst_ip,
1759                        sizeof(ib_spec->ipv4.val.dst_ip));
1760
1761                 set_tos(headers_c, headers_v,
1762                         ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos);
1763
1764                 set_proto(headers_c, headers_v,
1765                           ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto);
1766                 break;
1767         case IB_FLOW_SPEC_IPV6:
1768                 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD))
1769                         return -ENOTSUPP;
1770
1771                 MLX5_SET(fte_match_set_lyr_2_4, headers_c,
1772                          ethertype, 0xffff);
1773                 MLX5_SET(fte_match_set_lyr_2_4, headers_v,
1774                          ethertype, ETH_P_IPV6);
1775
1776                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1777                                     src_ipv4_src_ipv6.ipv6_layout.ipv6),
1778                        &ib_spec->ipv6.mask.src_ip,
1779                        sizeof(ib_spec->ipv6.mask.src_ip));
1780                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1781                                     src_ipv4_src_ipv6.ipv6_layout.ipv6),
1782                        &ib_spec->ipv6.val.src_ip,
1783                        sizeof(ib_spec->ipv6.val.src_ip));
1784                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_c,
1785                                     dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1786                        &ib_spec->ipv6.mask.dst_ip,
1787                        sizeof(ib_spec->ipv6.mask.dst_ip));
1788                 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
1789                                     dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
1790                        &ib_spec->ipv6.val.dst_ip,
1791                        sizeof(ib_spec->ipv6.val.dst_ip));
1792
1793                 set_tos(headers_c, headers_v,
1794                         ib_spec->ipv6.mask.traffic_class,
1795                         ib_spec->ipv6.val.traffic_class);
1796
1797                 set_proto(headers_c, headers_v,
1798                           ib_spec->ipv6.mask.next_hdr,
1799                           ib_spec->ipv6.val.next_hdr);
1800
1801                 set_flow_label(misc_params_c, misc_params_v,
1802                                ntohl(ib_spec->ipv6.mask.flow_label),
1803                                ntohl(ib_spec->ipv6.val.flow_label),
1804                                ib_spec->type & IB_FLOW_SPEC_INNER);
1805
1806                 break;
1807         case IB_FLOW_SPEC_TCP:
1808                 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1809                                          LAST_TCP_UDP_FIELD))
1810                         return -ENOTSUPP;
1811
1812                 MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
1813                          0xff);
1814                 MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
1815                          IPPROTO_TCP);
1816
1817                 MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_sport,
1818                          ntohs(ib_spec->tcp_udp.mask.src_port));
1819                 MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_sport,
1820                          ntohs(ib_spec->tcp_udp.val.src_port));
1821
1822                 MLX5_SET(fte_match_set_lyr_2_4, headers_c, tcp_dport,
1823                          ntohs(ib_spec->tcp_udp.mask.dst_port));
1824                 MLX5_SET(fte_match_set_lyr_2_4, headers_v, tcp_dport,
1825                          ntohs(ib_spec->tcp_udp.val.dst_port));
1826                 break;
1827         case IB_FLOW_SPEC_UDP:
1828                 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask,
1829                                          LAST_TCP_UDP_FIELD))
1830                         return -ENOTSUPP;
1831
1832                 MLX5_SET(fte_match_set_lyr_2_4, headers_c, ip_protocol,
1833                          0xff);
1834                 MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_protocol,
1835                          IPPROTO_UDP);
1836
1837                 MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_sport,
1838                          ntohs(ib_spec->tcp_udp.mask.src_port));
1839                 MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_sport,
1840                          ntohs(ib_spec->tcp_udp.val.src_port));
1841
1842                 MLX5_SET(fte_match_set_lyr_2_4, headers_c, udp_dport,
1843                          ntohs(ib_spec->tcp_udp.mask.dst_port));
1844                 MLX5_SET(fte_match_set_lyr_2_4, headers_v, udp_dport,
1845                          ntohs(ib_spec->tcp_udp.val.dst_port));
1846                 break;
1847         case IB_FLOW_SPEC_VXLAN_TUNNEL:
1848                 if (FIELDS_NOT_SUPPORTED(ib_spec->tunnel.mask,
1849                                          LAST_TUNNEL_FIELD))
1850                         return -ENOTSUPP;
1851
1852                 MLX5_SET(fte_match_set_misc, misc_params_c, vxlan_vni,
1853                          ntohl(ib_spec->tunnel.mask.tunnel_id));
1854                 MLX5_SET(fte_match_set_misc, misc_params_v, vxlan_vni,
1855                          ntohl(ib_spec->tunnel.val.tunnel_id));
1856                 break;
1857         default:
1858                 return -EINVAL;
1859         }
1860
1861         return 0;
1862 }
1863
1864 /* If a flow could catch both multicast and unicast packets,
1865  * it won't fall into the multicast flow steering table and this rule
1866  * could steal other multicast packets.
1867  */
1868 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr)
1869 {
1870         struct ib_flow_spec_eth *eth_spec;
1871
1872         if (ib_attr->type != IB_FLOW_ATTR_NORMAL ||
1873             ib_attr->size < sizeof(struct ib_flow_attr) +
1874             sizeof(struct ib_flow_spec_eth) ||
1875             ib_attr->num_of_specs < 1)
1876                 return false;
1877
1878         eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1);
1879         if (eth_spec->type != IB_FLOW_SPEC_ETH ||
1880             eth_spec->size != sizeof(*eth_spec))
1881                 return false;
1882
1883         return is_multicast_ether_addr(eth_spec->mask.dst_mac) &&
1884                is_multicast_ether_addr(eth_spec->val.dst_mac);
1885 }
1886
1887 static bool is_valid_attr(const struct ib_flow_attr *flow_attr)
1888 {
1889         union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1890         bool has_ipv4_spec = false;
1891         bool eth_type_ipv4 = true;
1892         unsigned int spec_index;
1893
1894         /* Validate that ethertype is correct */
1895         for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
1896                 if (ib_spec->type == IB_FLOW_SPEC_ETH &&
1897                     ib_spec->eth.mask.ether_type) {
1898                         if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) &&
1899                               ib_spec->eth.val.ether_type == htons(ETH_P_IP)))
1900                                 eth_type_ipv4 = false;
1901                 } else if (ib_spec->type == IB_FLOW_SPEC_IPV4) {
1902                         has_ipv4_spec = true;
1903                 }
1904                 ib_spec = (void *)ib_spec + ib_spec->size;
1905         }
1906         return !has_ipv4_spec || eth_type_ipv4;
1907 }
1908
1909 static void put_flow_table(struct mlx5_ib_dev *dev,
1910                            struct mlx5_ib_flow_prio *prio, bool ft_added)
1911 {
1912         prio->refcount -= !!ft_added;
1913         if (!prio->refcount) {
1914                 mlx5_destroy_flow_table(prio->flow_table);
1915                 prio->flow_table = NULL;
1916         }
1917 }
1918
1919 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id)
1920 {
1921         struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device);
1922         struct mlx5_ib_flow_handler *handler = container_of(flow_id,
1923                                                           struct mlx5_ib_flow_handler,
1924                                                           ibflow);
1925         struct mlx5_ib_flow_handler *iter, *tmp;
1926
1927         mutex_lock(&dev->flow_db.lock);
1928
1929         list_for_each_entry_safe(iter, tmp, &handler->list, list) {
1930                 mlx5_del_flow_rules(iter->rule);
1931                 put_flow_table(dev, iter->prio, true);
1932                 list_del(&iter->list);
1933                 kfree(iter);
1934         }
1935
1936         mlx5_del_flow_rules(handler->rule);
1937         put_flow_table(dev, handler->prio, true);
1938         mutex_unlock(&dev->flow_db.lock);
1939
1940         kfree(handler);
1941
1942         return 0;
1943 }
1944
1945 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap)
1946 {
1947         priority *= 2;
1948         if (!dont_trap)
1949                 priority++;
1950         return priority;
1951 }
1952
1953 enum flow_table_type {
1954         MLX5_IB_FT_RX,
1955         MLX5_IB_FT_TX
1956 };
1957
1958 #define MLX5_FS_MAX_TYPES        10
1959 #define MLX5_FS_MAX_ENTRIES      32000UL
1960 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
1961                                                 struct ib_flow_attr *flow_attr,
1962                                                 enum flow_table_type ft_type)
1963 {
1964         bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP;
1965         struct mlx5_flow_namespace *ns = NULL;
1966         struct mlx5_ib_flow_prio *prio;
1967         struct mlx5_flow_table *ft;
1968         int num_entries;
1969         int num_groups;
1970         int priority;
1971         int err = 0;
1972
1973         if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1974                 if (flow_is_multicast_only(flow_attr) &&
1975                     !dont_trap)
1976                         priority = MLX5_IB_FLOW_MCAST_PRIO;
1977                 else
1978                         priority = ib_prio_to_core_prio(flow_attr->priority,
1979                                                         dont_trap);
1980                 ns = mlx5_get_flow_namespace(dev->mdev,
1981                                              MLX5_FLOW_NAMESPACE_BYPASS);
1982                 num_entries = MLX5_FS_MAX_ENTRIES;
1983                 num_groups = MLX5_FS_MAX_TYPES;
1984                 prio = &dev->flow_db.prios[priority];
1985         } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
1986                    flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
1987                 ns = mlx5_get_flow_namespace(dev->mdev,
1988                                              MLX5_FLOW_NAMESPACE_LEFTOVERS);
1989                 build_leftovers_ft_param(&priority,
1990                                          &num_entries,
1991                                          &num_groups);
1992                 prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO];
1993         } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
1994                 if (!MLX5_CAP_FLOWTABLE(dev->mdev,
1995                                         allow_sniffer_and_nic_rx_shared_tir))
1996                         return ERR_PTR(-ENOTSUPP);
1997
1998                 ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ?
1999                                              MLX5_FLOW_NAMESPACE_SNIFFER_RX :
2000                                              MLX5_FLOW_NAMESPACE_SNIFFER_TX);
2001
2002                 prio = &dev->flow_db.sniffer[ft_type];
2003                 priority = 0;
2004                 num_entries = 1;
2005                 num_groups = 1;
2006         }
2007
2008         if (!ns)
2009                 return ERR_PTR(-ENOTSUPP);
2010
2011         ft = prio->flow_table;
2012         if (!ft) {
2013                 ft = mlx5_create_auto_grouped_flow_table(ns, priority,
2014                                                          num_entries,
2015                                                          num_groups,
2016                                                          0, 0);
2017
2018                 if (!IS_ERR(ft)) {
2019                         prio->refcount = 0;
2020                         prio->flow_table = ft;
2021                 } else {
2022                         err = PTR_ERR(ft);
2023                 }
2024         }
2025
2026         return err ? ERR_PTR(err) : prio;
2027 }
2028
2029 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev,
2030                                                      struct mlx5_ib_flow_prio *ft_prio,
2031                                                      const struct ib_flow_attr *flow_attr,
2032                                                      struct mlx5_flow_destination *dst)
2033 {
2034         struct mlx5_flow_table  *ft = ft_prio->flow_table;
2035         struct mlx5_ib_flow_handler *handler;
2036         struct mlx5_flow_act flow_act = {0};
2037         struct mlx5_flow_spec *spec;
2038         const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr);
2039         unsigned int spec_index;
2040         int err = 0;
2041
2042         if (!is_valid_attr(flow_attr))
2043                 return ERR_PTR(-EINVAL);
2044
2045         spec = mlx5_vzalloc(sizeof(*spec));
2046         handler = kzalloc(sizeof(*handler), GFP_KERNEL);
2047         if (!handler || !spec) {
2048                 err = -ENOMEM;
2049                 goto free;
2050         }
2051
2052         INIT_LIST_HEAD(&handler->list);
2053
2054         for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) {
2055                 err = parse_flow_attr(spec->match_criteria,
2056                                       spec->match_value, ib_flow);
2057                 if (err < 0)
2058                         goto free;
2059
2060                 ib_flow += ((union ib_flow_spec *)ib_flow)->size;
2061         }
2062
2063         spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria);
2064         flow_act.action = dst ? MLX5_FLOW_CONTEXT_ACTION_FWD_DEST :
2065                 MLX5_FLOW_CONTEXT_ACTION_FWD_NEXT_PRIO;
2066         flow_act.flow_tag = MLX5_FS_DEFAULT_FLOW_TAG;
2067         handler->rule = mlx5_add_flow_rules(ft, spec,
2068                                             &flow_act,
2069                                             dst, 1);
2070
2071         if (IS_ERR(handler->rule)) {
2072                 err = PTR_ERR(handler->rule);
2073                 goto free;
2074         }
2075
2076         ft_prio->refcount++;
2077         handler->prio = ft_prio;
2078
2079         ft_prio->flow_table = ft;
2080 free:
2081         if (err)
2082                 kfree(handler);
2083         kvfree(spec);
2084         return err ? ERR_PTR(err) : handler;
2085 }
2086
2087 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev,
2088                                                           struct mlx5_ib_flow_prio *ft_prio,
2089                                                           struct ib_flow_attr *flow_attr,
2090                                                           struct mlx5_flow_destination *dst)
2091 {
2092         struct mlx5_ib_flow_handler *handler_dst = NULL;
2093         struct mlx5_ib_flow_handler *handler = NULL;
2094
2095         handler = create_flow_rule(dev, ft_prio, flow_attr, NULL);
2096         if (!IS_ERR(handler)) {
2097                 handler_dst = create_flow_rule(dev, ft_prio,
2098                                                flow_attr, dst);
2099                 if (IS_ERR(handler_dst)) {
2100                         mlx5_del_flow_rules(handler->rule);
2101                         ft_prio->refcount--;
2102                         kfree(handler);
2103                         handler = handler_dst;
2104                 } else {
2105                         list_add(&handler_dst->list, &handler->list);
2106                 }
2107         }
2108
2109         return handler;
2110 }
2111 enum {
2112         LEFTOVERS_MC,
2113         LEFTOVERS_UC,
2114 };
2115
2116 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev,
2117                                                           struct mlx5_ib_flow_prio *ft_prio,
2118                                                           struct ib_flow_attr *flow_attr,
2119                                                           struct mlx5_flow_destination *dst)
2120 {
2121         struct mlx5_ib_flow_handler *handler_ucast = NULL;
2122         struct mlx5_ib_flow_handler *handler = NULL;
2123
2124         static struct {
2125                 struct ib_flow_attr     flow_attr;
2126                 struct ib_flow_spec_eth eth_flow;
2127         } leftovers_specs[] = {
2128                 [LEFTOVERS_MC] = {
2129                         .flow_attr = {
2130                                 .num_of_specs = 1,
2131                                 .size = sizeof(leftovers_specs[0])
2132                         },
2133                         .eth_flow = {
2134                                 .type = IB_FLOW_SPEC_ETH,
2135                                 .size = sizeof(struct ib_flow_spec_eth),
2136                                 .mask = {.dst_mac = {0x1} },
2137                                 .val =  {.dst_mac = {0x1} }
2138                         }
2139                 },
2140                 [LEFTOVERS_UC] = {
2141                         .flow_attr = {
2142                                 .num_of_specs = 1,
2143                                 .size = sizeof(leftovers_specs[0])
2144                         },
2145                         .eth_flow = {
2146                                 .type = IB_FLOW_SPEC_ETH,
2147                                 .size = sizeof(struct ib_flow_spec_eth),
2148                                 .mask = {.dst_mac = {0x1} },
2149                                 .val = {.dst_mac = {} }
2150                         }
2151                 }
2152         };
2153
2154         handler = create_flow_rule(dev, ft_prio,
2155                                    &leftovers_specs[LEFTOVERS_MC].flow_attr,
2156                                    dst);
2157         if (!IS_ERR(handler) &&
2158             flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) {
2159                 handler_ucast = create_flow_rule(dev, ft_prio,
2160                                                  &leftovers_specs[LEFTOVERS_UC].flow_attr,
2161                                                  dst);
2162                 if (IS_ERR(handler_ucast)) {
2163                         mlx5_del_flow_rules(handler->rule);
2164                         ft_prio->refcount--;
2165                         kfree(handler);
2166                         handler = handler_ucast;
2167                 } else {
2168                         list_add(&handler_ucast->list, &handler->list);
2169                 }
2170         }
2171
2172         return handler;
2173 }
2174
2175 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev,
2176                                                         struct mlx5_ib_flow_prio *ft_rx,
2177                                                         struct mlx5_ib_flow_prio *ft_tx,
2178                                                         struct mlx5_flow_destination *dst)
2179 {
2180         struct mlx5_ib_flow_handler *handler_rx;
2181         struct mlx5_ib_flow_handler *handler_tx;
2182         int err;
2183         static const struct ib_flow_attr flow_attr  = {
2184                 .num_of_specs = 0,
2185                 .size = sizeof(flow_attr)
2186         };
2187
2188         handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst);
2189         if (IS_ERR(handler_rx)) {
2190                 err = PTR_ERR(handler_rx);
2191                 goto err;
2192         }
2193
2194         handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst);
2195         if (IS_ERR(handler_tx)) {
2196                 err = PTR_ERR(handler_tx);
2197                 goto err_tx;
2198         }
2199
2200         list_add(&handler_tx->list, &handler_rx->list);
2201
2202         return handler_rx;
2203
2204 err_tx:
2205         mlx5_del_flow_rules(handler_rx->rule);
2206         ft_rx->refcount--;
2207         kfree(handler_rx);
2208 err:
2209         return ERR_PTR(err);
2210 }
2211
2212 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp,
2213                                            struct ib_flow_attr *flow_attr,
2214                                            int domain)
2215 {
2216         struct mlx5_ib_dev *dev = to_mdev(qp->device);
2217         struct mlx5_ib_qp *mqp = to_mqp(qp);
2218         struct mlx5_ib_flow_handler *handler = NULL;
2219         struct mlx5_flow_destination *dst = NULL;
2220         struct mlx5_ib_flow_prio *ft_prio_tx = NULL;
2221         struct mlx5_ib_flow_prio *ft_prio;
2222         int err;
2223
2224         if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO)
2225                 return ERR_PTR(-ENOSPC);
2226
2227         if (domain != IB_FLOW_DOMAIN_USER ||
2228             flow_attr->port > MLX5_CAP_GEN(dev->mdev, num_ports) ||
2229             (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP))
2230                 return ERR_PTR(-EINVAL);
2231
2232         dst = kzalloc(sizeof(*dst), GFP_KERNEL);
2233         if (!dst)
2234                 return ERR_PTR(-ENOMEM);
2235
2236         mutex_lock(&dev->flow_db.lock);
2237
2238         ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX);
2239         if (IS_ERR(ft_prio)) {
2240                 err = PTR_ERR(ft_prio);
2241                 goto unlock;
2242         }
2243         if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2244                 ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX);
2245                 if (IS_ERR(ft_prio_tx)) {
2246                         err = PTR_ERR(ft_prio_tx);
2247                         ft_prio_tx = NULL;
2248                         goto destroy_ft;
2249                 }
2250         }
2251
2252         dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR;
2253         if (mqp->flags & MLX5_IB_QP_RSS)
2254                 dst->tir_num = mqp->rss_qp.tirn;
2255         else
2256                 dst->tir_num = mqp->raw_packet_qp.rq.tirn;
2257
2258         if (flow_attr->type == IB_FLOW_ATTR_NORMAL) {
2259                 if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)  {
2260                         handler = create_dont_trap_rule(dev, ft_prio,
2261                                                         flow_attr, dst);
2262                 } else {
2263                         handler = create_flow_rule(dev, ft_prio, flow_attr,
2264                                                    dst);
2265                 }
2266         } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT ||
2267                    flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) {
2268                 handler = create_leftovers_rule(dev, ft_prio, flow_attr,
2269                                                 dst);
2270         } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) {
2271                 handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst);
2272         } else {
2273                 err = -EINVAL;
2274                 goto destroy_ft;
2275         }
2276
2277         if (IS_ERR(handler)) {
2278                 err = PTR_ERR(handler);
2279                 handler = NULL;
2280                 goto destroy_ft;
2281         }
2282
2283         mutex_unlock(&dev->flow_db.lock);
2284         kfree(dst);
2285
2286         return &handler->ibflow;
2287
2288 destroy_ft:
2289         put_flow_table(dev, ft_prio, false);
2290         if (ft_prio_tx)
2291                 put_flow_table(dev, ft_prio_tx, false);
2292 unlock:
2293         mutex_unlock(&dev->flow_db.lock);
2294         kfree(dst);
2295         kfree(handler);
2296         return ERR_PTR(err);
2297 }
2298
2299 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2300 {
2301         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2302         int err;
2303
2304         err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num);
2305         if (err)
2306                 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2307                              ibqp->qp_num, gid->raw);
2308
2309         return err;
2310 }
2311
2312 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2313 {
2314         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2315         int err;
2316
2317         err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num);
2318         if (err)
2319                 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2320                              ibqp->qp_num, gid->raw);
2321
2322         return err;
2323 }
2324
2325 static int init_node_data(struct mlx5_ib_dev *dev)
2326 {
2327         int err;
2328
2329         err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2330         if (err)
2331                 return err;
2332
2333         dev->mdev->rev_id = dev->mdev->pdev->revision;
2334
2335         return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2336 }
2337
2338 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr,
2339                              char *buf)
2340 {
2341         struct mlx5_ib_dev *dev =
2342                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2343
2344         return sprintf(buf, "%d\n", dev->mdev->priv.fw_pages);
2345 }
2346
2347 static ssize_t show_reg_pages(struct device *device,
2348                               struct device_attribute *attr, char *buf)
2349 {
2350         struct mlx5_ib_dev *dev =
2351                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2352
2353         return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2354 }
2355
2356 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2357                         char *buf)
2358 {
2359         struct mlx5_ib_dev *dev =
2360                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2361         return sprintf(buf, "MT%d\n", dev->mdev->pdev->device);
2362 }
2363
2364 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2365                         char *buf)
2366 {
2367         struct mlx5_ib_dev *dev =
2368                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2369         return sprintf(buf, "%x\n", dev->mdev->rev_id);
2370 }
2371
2372 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2373                           char *buf)
2374 {
2375         struct mlx5_ib_dev *dev =
2376                 container_of(device, struct mlx5_ib_dev, ib_dev.dev);
2377         return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2378                        dev->mdev->board_id);
2379 }
2380
2381 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
2382 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
2383 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
2384 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL);
2385 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL);
2386
2387 static struct device_attribute *mlx5_class_attributes[] = {
2388         &dev_attr_hw_rev,
2389         &dev_attr_hca_type,
2390         &dev_attr_board_id,
2391         &dev_attr_fw_pages,
2392         &dev_attr_reg_pages,
2393 };
2394
2395 static void pkey_change_handler(struct work_struct *work)
2396 {
2397         struct mlx5_ib_port_resources *ports =
2398                 container_of(work, struct mlx5_ib_port_resources,
2399                              pkey_change_work);
2400
2401         mutex_lock(&ports->devr->mutex);
2402         mlx5_ib_gsi_pkey_change(ports->gsi);
2403         mutex_unlock(&ports->devr->mutex);
2404 }
2405
2406 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2407 {
2408         struct mlx5_ib_qp *mqp;
2409         struct mlx5_ib_cq *send_mcq, *recv_mcq;
2410         struct mlx5_core_cq *mcq;
2411         struct list_head cq_armed_list;
2412         unsigned long flags_qp;
2413         unsigned long flags_cq;
2414         unsigned long flags;
2415
2416         INIT_LIST_HEAD(&cq_armed_list);
2417
2418         /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2419         spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2420         list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2421                 spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2422                 if (mqp->sq.tail != mqp->sq.head) {
2423                         send_mcq = to_mcq(mqp->ibqp.send_cq);
2424                         spin_lock_irqsave(&send_mcq->lock, flags_cq);
2425                         if (send_mcq->mcq.comp &&
2426                             mqp->ibqp.send_cq->comp_handler) {
2427                                 if (!send_mcq->mcq.reset_notify_added) {
2428                                         send_mcq->mcq.reset_notify_added = 1;
2429                                         list_add_tail(&send_mcq->mcq.reset_notify,
2430                                                       &cq_armed_list);
2431                                 }
2432                         }
2433                         spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2434                 }
2435                 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2436                 spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2437                 /* no handling is needed for SRQ */
2438                 if (!mqp->ibqp.srq) {
2439                         if (mqp->rq.tail != mqp->rq.head) {
2440                                 recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2441                                 spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2442                                 if (recv_mcq->mcq.comp &&
2443                                     mqp->ibqp.recv_cq->comp_handler) {
2444                                         if (!recv_mcq->mcq.reset_notify_added) {
2445                                                 recv_mcq->mcq.reset_notify_added = 1;
2446                                                 list_add_tail(&recv_mcq->mcq.reset_notify,
2447                                                               &cq_armed_list);
2448                                         }
2449                                 }
2450                                 spin_unlock_irqrestore(&recv_mcq->lock,
2451                                                        flags_cq);
2452                         }
2453                 }
2454                 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2455         }
2456         /*At that point all inflight post send were put to be executed as of we
2457          * lock/unlock above locks Now need to arm all involved CQs.
2458          */
2459         list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2460                 mcq->comp(mcq);
2461         }
2462         spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2463 }
2464
2465 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
2466                           enum mlx5_dev_event event, unsigned long param)
2467 {
2468         struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context;
2469         struct ib_event ibev;
2470         bool fatal = false;
2471         u8 port = 0;
2472
2473         switch (event) {
2474         case MLX5_DEV_EVENT_SYS_ERROR:
2475                 ibev.event = IB_EVENT_DEVICE_FATAL;
2476                 mlx5_ib_handle_internal_error(ibdev);
2477                 fatal = true;
2478                 break;
2479
2480         case MLX5_DEV_EVENT_PORT_UP:
2481         case MLX5_DEV_EVENT_PORT_DOWN:
2482         case MLX5_DEV_EVENT_PORT_INITIALIZED:
2483                 port = (u8)param;
2484
2485                 /* In RoCE, port up/down events are handled in
2486                  * mlx5_netdev_event().
2487                  */
2488                 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2489                         IB_LINK_LAYER_ETHERNET)
2490                         return;
2491
2492                 ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ?
2493                              IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2494                 break;
2495
2496         case MLX5_DEV_EVENT_LID_CHANGE:
2497                 ibev.event = IB_EVENT_LID_CHANGE;
2498                 port = (u8)param;
2499                 break;
2500
2501         case MLX5_DEV_EVENT_PKEY_CHANGE:
2502                 ibev.event = IB_EVENT_PKEY_CHANGE;
2503                 port = (u8)param;
2504
2505                 schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2506                 break;
2507
2508         case MLX5_DEV_EVENT_GUID_CHANGE:
2509                 ibev.event = IB_EVENT_GID_CHANGE;
2510                 port = (u8)param;
2511                 break;
2512
2513         case MLX5_DEV_EVENT_CLIENT_REREG:
2514                 ibev.event = IB_EVENT_CLIENT_REREGISTER;
2515                 port = (u8)param;
2516                 break;
2517         default:
2518                 return;
2519         }
2520
2521         ibev.device           = &ibdev->ib_dev;
2522         ibev.element.port_num = port;
2523
2524         if (port < 1 || port > ibdev->num_ports) {
2525                 mlx5_ib_warn(ibdev, "warning: event on port %d\n", port);
2526                 return;
2527         }
2528
2529         if (ibdev->ib_active)
2530                 ib_dispatch_event(&ibev);
2531
2532         if (fatal)
2533                 ibdev->ib_active = false;
2534 }
2535
2536 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2537 {
2538         struct mlx5_hca_vport_context vport_ctx;
2539         int err;
2540         int port;
2541
2542         for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2543                 dev->mdev->port_caps[port - 1].has_smi = false;
2544                 if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2545                     MLX5_CAP_PORT_TYPE_IB) {
2546                         if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2547                                 err = mlx5_query_hca_vport_context(dev->mdev, 0,
2548                                                                    port, 0,
2549                                                                    &vport_ctx);
2550                                 if (err) {
2551                                         mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2552                                                     port, err);
2553                                         return err;
2554                                 }
2555                                 dev->mdev->port_caps[port - 1].has_smi =
2556                                         vport_ctx.has_smi;
2557                         } else {
2558                                 dev->mdev->port_caps[port - 1].has_smi = true;
2559                         }
2560                 }
2561         }
2562         return 0;
2563 }
2564
2565 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2566 {
2567         int port;
2568
2569         for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++)
2570                 mlx5_query_ext_port_caps(dev, port);
2571 }
2572
2573 static int get_port_caps(struct mlx5_ib_dev *dev)
2574 {
2575         struct ib_device_attr *dprops = NULL;
2576         struct ib_port_attr *pprops = NULL;
2577         int err = -ENOMEM;
2578         int port;
2579         struct ib_udata uhw = {.inlen = 0, .outlen = 0};
2580
2581         pprops = kmalloc(sizeof(*pprops), GFP_KERNEL);
2582         if (!pprops)
2583                 goto out;
2584
2585         dprops = kmalloc(sizeof(*dprops), GFP_KERNEL);
2586         if (!dprops)
2587                 goto out;
2588
2589         err = set_has_smi_cap(dev);
2590         if (err)
2591                 goto out;
2592
2593         err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
2594         if (err) {
2595                 mlx5_ib_warn(dev, "query_device failed %d\n", err);
2596                 goto out;
2597         }
2598
2599         for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) {
2600                 err = mlx5_ib_query_port(&dev->ib_dev, port, pprops);
2601                 if (err) {
2602                         mlx5_ib_warn(dev, "query_port %d failed %d\n",
2603                                      port, err);
2604                         break;
2605                 }
2606                 dev->mdev->port_caps[port - 1].pkey_table_len =
2607                                                 dprops->max_pkeys;
2608                 dev->mdev->port_caps[port - 1].gid_table_len =
2609                                                 pprops->gid_tbl_len;
2610                 mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n",
2611                             dprops->max_pkeys, pprops->gid_tbl_len);
2612         }
2613
2614 out:
2615         kfree(pprops);
2616         kfree(dprops);
2617
2618         return err;
2619 }
2620
2621 static void destroy_umrc_res(struct mlx5_ib_dev *dev)
2622 {
2623         int err;
2624
2625         err = mlx5_mr_cache_cleanup(dev);
2626         if (err)
2627                 mlx5_ib_warn(dev, "mr cache cleanup failed\n");
2628
2629         mlx5_ib_destroy_qp(dev->umrc.qp);
2630         ib_free_cq(dev->umrc.cq);
2631         ib_dealloc_pd(dev->umrc.pd);
2632 }
2633
2634 enum {
2635         MAX_UMR_WR = 128,
2636 };
2637
2638 static int create_umr_res(struct mlx5_ib_dev *dev)
2639 {
2640         struct ib_qp_init_attr *init_attr = NULL;
2641         struct ib_qp_attr *attr = NULL;
2642         struct ib_pd *pd;
2643         struct ib_cq *cq;
2644         struct ib_qp *qp;
2645         int ret;
2646
2647         attr = kzalloc(sizeof(*attr), GFP_KERNEL);
2648         init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
2649         if (!attr || !init_attr) {
2650                 ret = -ENOMEM;
2651                 goto error_0;
2652         }
2653
2654         pd = ib_alloc_pd(&dev->ib_dev, 0);
2655         if (IS_ERR(pd)) {
2656                 mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
2657                 ret = PTR_ERR(pd);
2658                 goto error_0;
2659         }
2660
2661         cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
2662         if (IS_ERR(cq)) {
2663                 mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
2664                 ret = PTR_ERR(cq);
2665                 goto error_2;
2666         }
2667
2668         init_attr->send_cq = cq;
2669         init_attr->recv_cq = cq;
2670         init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
2671         init_attr->cap.max_send_wr = MAX_UMR_WR;
2672         init_attr->cap.max_send_sge = 1;
2673         init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
2674         init_attr->port_num = 1;
2675         qp = mlx5_ib_create_qp(pd, init_attr, NULL);
2676         if (IS_ERR(qp)) {
2677                 mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
2678                 ret = PTR_ERR(qp);
2679                 goto error_3;
2680         }
2681         qp->device     = &dev->ib_dev;
2682         qp->real_qp    = qp;
2683         qp->uobject    = NULL;
2684         qp->qp_type    = MLX5_IB_QPT_REG_UMR;
2685
2686         attr->qp_state = IB_QPS_INIT;
2687         attr->port_num = 1;
2688         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
2689                                 IB_QP_PORT, NULL);
2690         if (ret) {
2691                 mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
2692                 goto error_4;
2693         }
2694
2695         memset(attr, 0, sizeof(*attr));
2696         attr->qp_state = IB_QPS_RTR;
2697         attr->path_mtu = IB_MTU_256;
2698
2699         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2700         if (ret) {
2701                 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
2702                 goto error_4;
2703         }
2704
2705         memset(attr, 0, sizeof(*attr));
2706         attr->qp_state = IB_QPS_RTS;
2707         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
2708         if (ret) {
2709                 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
2710                 goto error_4;
2711         }
2712
2713         dev->umrc.qp = qp;
2714         dev->umrc.cq = cq;
2715         dev->umrc.pd = pd;
2716
2717         sema_init(&dev->umrc.sem, MAX_UMR_WR);
2718         ret = mlx5_mr_cache_init(dev);
2719         if (ret) {
2720                 mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
2721                 goto error_4;
2722         }
2723
2724         kfree(attr);
2725         kfree(init_attr);
2726
2727         return 0;
2728
2729 error_4:
2730         mlx5_ib_destroy_qp(qp);
2731
2732 error_3:
2733         ib_free_cq(cq);
2734
2735 error_2:
2736         ib_dealloc_pd(pd);
2737
2738 error_0:
2739         kfree(attr);
2740         kfree(init_attr);
2741         return ret;
2742 }
2743
2744 static int create_dev_resources(struct mlx5_ib_resources *devr)
2745 {
2746         struct ib_srq_init_attr attr;
2747         struct mlx5_ib_dev *dev;
2748         struct ib_cq_init_attr cq_attr = {.cqe = 1};
2749         int port;
2750         int ret = 0;
2751
2752         dev = container_of(devr, struct mlx5_ib_dev, devr);
2753
2754         mutex_init(&devr->mutex);
2755
2756         devr->p0 = mlx5_ib_alloc_pd(&dev->ib_dev, NULL, NULL);
2757         if (IS_ERR(devr->p0)) {
2758                 ret = PTR_ERR(devr->p0);
2759                 goto error0;
2760         }
2761         devr->p0->device  = &dev->ib_dev;
2762         devr->p0->uobject = NULL;
2763         atomic_set(&devr->p0->usecnt, 0);
2764
2765         devr->c0 = mlx5_ib_create_cq(&dev->ib_dev, &cq_attr, NULL, NULL);
2766         if (IS_ERR(devr->c0)) {
2767                 ret = PTR_ERR(devr->c0);
2768                 goto error1;
2769         }
2770         devr->c0->device        = &dev->ib_dev;
2771         devr->c0->uobject       = NULL;
2772         devr->c0->comp_handler  = NULL;
2773         devr->c0->event_handler = NULL;
2774         devr->c0->cq_context    = NULL;
2775         atomic_set(&devr->c0->usecnt, 0);
2776
2777         devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2778         if (IS_ERR(devr->x0)) {
2779                 ret = PTR_ERR(devr->x0);
2780                 goto error2;
2781         }
2782         devr->x0->device = &dev->ib_dev;
2783         devr->x0->inode = NULL;
2784         atomic_set(&devr->x0->usecnt, 0);
2785         mutex_init(&devr->x0->tgt_qp_mutex);
2786         INIT_LIST_HEAD(&devr->x0->tgt_qp_list);
2787
2788         devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL, NULL);
2789         if (IS_ERR(devr->x1)) {
2790                 ret = PTR_ERR(devr->x1);
2791                 goto error3;
2792         }
2793         devr->x1->device = &dev->ib_dev;
2794         devr->x1->inode = NULL;
2795         atomic_set(&devr->x1->usecnt, 0);
2796         mutex_init(&devr->x1->tgt_qp_mutex);
2797         INIT_LIST_HEAD(&devr->x1->tgt_qp_list);
2798
2799         memset(&attr, 0, sizeof(attr));
2800         attr.attr.max_sge = 1;
2801         attr.attr.max_wr = 1;
2802         attr.srq_type = IB_SRQT_XRC;
2803         attr.ext.xrc.cq = devr->c0;
2804         attr.ext.xrc.xrcd = devr->x0;
2805
2806         devr->s0 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2807         if (IS_ERR(devr->s0)) {
2808                 ret = PTR_ERR(devr->s0);
2809                 goto error4;
2810         }
2811         devr->s0->device        = &dev->ib_dev;
2812         devr->s0->pd            = devr->p0;
2813         devr->s0->uobject       = NULL;
2814         devr->s0->event_handler = NULL;
2815         devr->s0->srq_context   = NULL;
2816         devr->s0->srq_type      = IB_SRQT_XRC;
2817         devr->s0->ext.xrc.xrcd  = devr->x0;
2818         devr->s0->ext.xrc.cq    = devr->c0;
2819         atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt);
2820         atomic_inc(&devr->s0->ext.xrc.cq->usecnt);
2821         atomic_inc(&devr->p0->usecnt);
2822         atomic_set(&devr->s0->usecnt, 0);
2823
2824         memset(&attr, 0, sizeof(attr));
2825         attr.attr.max_sge = 1;
2826         attr.attr.max_wr = 1;
2827         attr.srq_type = IB_SRQT_BASIC;
2828         devr->s1 = mlx5_ib_create_srq(devr->p0, &attr, NULL);
2829         if (IS_ERR(devr->s1)) {
2830                 ret = PTR_ERR(devr->s1);
2831                 goto error5;
2832         }
2833         devr->s1->device        = &dev->ib_dev;
2834         devr->s1->pd            = devr->p0;
2835         devr->s1->uobject       = NULL;
2836         devr->s1->event_handler = NULL;
2837         devr->s1->srq_context   = NULL;
2838         devr->s1->srq_type      = IB_SRQT_BASIC;
2839         devr->s1->ext.xrc.cq    = devr->c0;
2840         atomic_inc(&devr->p0->usecnt);
2841         atomic_set(&devr->s0->usecnt, 0);
2842
2843         for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) {
2844                 INIT_WORK(&devr->ports[port].pkey_change_work,
2845                           pkey_change_handler);
2846                 devr->ports[port].devr = devr;
2847         }
2848
2849         return 0;
2850
2851 error5:
2852         mlx5_ib_destroy_srq(devr->s0);
2853 error4:
2854         mlx5_ib_dealloc_xrcd(devr->x1);
2855 error3:
2856         mlx5_ib_dealloc_xrcd(devr->x0);
2857 error2:
2858         mlx5_ib_destroy_cq(devr->c0);
2859 error1:
2860         mlx5_ib_dealloc_pd(devr->p0);
2861 error0:
2862         return ret;
2863 }
2864
2865 static void destroy_dev_resources(struct mlx5_ib_resources *devr)
2866 {
2867         struct mlx5_ib_dev *dev =
2868                 container_of(devr, struct mlx5_ib_dev, devr);
2869         int port;
2870
2871         mlx5_ib_destroy_srq(devr->s1);
2872         mlx5_ib_destroy_srq(devr->s0);
2873         mlx5_ib_dealloc_xrcd(devr->x0);
2874         mlx5_ib_dealloc_xrcd(devr->x1);
2875         mlx5_ib_destroy_cq(devr->c0);
2876         mlx5_ib_dealloc_pd(devr->p0);
2877
2878         /* Make sure no change P_Key work items are still executing */
2879         for (port = 0; port < dev->num_ports; ++port)
2880                 cancel_work_sync(&devr->ports[port].pkey_change_work);
2881 }
2882
2883 static u32 get_core_cap_flags(struct ib_device *ibdev)
2884 {
2885         struct mlx5_ib_dev *dev = to_mdev(ibdev);
2886         enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2887         u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2888         u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2889         u32 ret = 0;
2890
2891         if (ll == IB_LINK_LAYER_INFINIBAND)
2892                 return RDMA_CORE_PORT_IBA_IB;
2893
2894         if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2895                 return 0;
2896
2897         if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2898                 return 0;
2899
2900         if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2901                 ret |= RDMA_CORE_PORT_IBA_ROCE;
2902
2903         if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2904                 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2905
2906         return ret;
2907 }
2908
2909 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num,
2910                                struct ib_port_immutable *immutable)
2911 {
2912         struct ib_port_attr attr;
2913         struct mlx5_ib_dev *dev = to_mdev(ibdev);
2914         enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
2915         int err;
2916
2917         err = mlx5_ib_query_port(ibdev, port_num, &attr);
2918         if (err)
2919                 return err;
2920
2921         immutable->pkey_tbl_len = attr.pkey_tbl_len;
2922         immutable->gid_tbl_len = attr.gid_tbl_len;
2923         immutable->core_cap_flags = get_core_cap_flags(ibdev);
2924         if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce))
2925                 immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2926
2927         return 0;
2928 }
2929
2930 static void get_dev_fw_str(struct ib_device *ibdev, char *str,
2931                            size_t str_len)
2932 {
2933         struct mlx5_ib_dev *dev =
2934                 container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2935         snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev),
2936                        fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev));
2937 }
2938
2939 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
2940 {
2941         struct mlx5_core_dev *mdev = dev->mdev;
2942         struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
2943                                                                  MLX5_FLOW_NAMESPACE_LAG);
2944         struct mlx5_flow_table *ft;
2945         int err;
2946
2947         if (!ns || !mlx5_lag_is_active(mdev))
2948                 return 0;
2949
2950         err = mlx5_cmd_create_vport_lag(mdev);
2951         if (err)
2952                 return err;
2953
2954         ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
2955         if (IS_ERR(ft)) {
2956                 err = PTR_ERR(ft);
2957                 goto err_destroy_vport_lag;
2958         }
2959
2960         dev->flow_db.lag_demux_ft = ft;
2961         return 0;
2962
2963 err_destroy_vport_lag:
2964         mlx5_cmd_destroy_vport_lag(mdev);
2965         return err;
2966 }
2967
2968 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
2969 {
2970         struct mlx5_core_dev *mdev = dev->mdev;
2971
2972         if (dev->flow_db.lag_demux_ft) {
2973                 mlx5_destroy_flow_table(dev->flow_db.lag_demux_ft);
2974                 dev->flow_db.lag_demux_ft = NULL;
2975
2976                 mlx5_cmd_destroy_vport_lag(mdev);
2977         }
2978 }
2979
2980 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev)
2981 {
2982         int err;
2983
2984         dev->roce.nb.notifier_call = mlx5_netdev_event;
2985         err = register_netdevice_notifier(&dev->roce.nb);
2986         if (err) {
2987                 dev->roce.nb.notifier_call = NULL;
2988                 return err;
2989         }
2990
2991         return 0;
2992 }
2993
2994 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev)
2995 {
2996         if (dev->roce.nb.notifier_call) {
2997                 unregister_netdevice_notifier(&dev->roce.nb);
2998                 dev->roce.nb.notifier_call = NULL;
2999         }
3000 }
3001
3002 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3003 {
3004         int err;
3005
3006         err = mlx5_add_netdev_notifier(dev);
3007         if (err)
3008                 return err;
3009
3010         if (MLX5_CAP_GEN(dev->mdev, roce)) {
3011                 err = mlx5_nic_vport_enable_roce(dev->mdev);
3012                 if (err)
3013                         goto err_unregister_netdevice_notifier;
3014         }
3015
3016         err = mlx5_eth_lag_init(dev);
3017         if (err)
3018                 goto err_disable_roce;
3019
3020         return 0;
3021
3022 err_disable_roce:
3023         if (MLX5_CAP_GEN(dev->mdev, roce))
3024                 mlx5_nic_vport_disable_roce(dev->mdev);
3025
3026 err_unregister_netdevice_notifier:
3027         mlx5_remove_netdev_notifier(dev);
3028         return err;
3029 }
3030
3031 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3032 {
3033         mlx5_eth_lag_cleanup(dev);
3034         if (MLX5_CAP_GEN(dev->mdev, roce))
3035                 mlx5_nic_vport_disable_roce(dev->mdev);
3036 }
3037
3038 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev)
3039 {
3040         unsigned int i;
3041
3042         for (i = 0; i < dev->num_ports; i++)
3043                 mlx5_core_dealloc_q_counter(dev->mdev,
3044                                             dev->port[i].q_cnt_id);
3045 }
3046
3047 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev)
3048 {
3049         int i;
3050         int ret;
3051
3052         for (i = 0; i < dev->num_ports; i++) {
3053                 ret = mlx5_core_alloc_q_counter(dev->mdev,
3054                                                 &dev->port[i].q_cnt_id);
3055                 if (ret) {
3056                         mlx5_ib_warn(dev,
3057                                      "couldn't allocate queue counter for port %d, err %d\n",
3058                                      i + 1, ret);
3059                         goto dealloc_counters;
3060                 }
3061         }
3062
3063         return 0;
3064
3065 dealloc_counters:
3066         while (--i >= 0)
3067                 mlx5_core_dealloc_q_counter(dev->mdev,
3068                                             dev->port[i].q_cnt_id);
3069
3070         return ret;
3071 }
3072
3073 static const char * const names[] = {
3074         "rx_write_requests",
3075         "rx_read_requests",
3076         "rx_atomic_requests",
3077         "out_of_buffer",
3078         "out_of_sequence",
3079         "duplicate_request",
3080         "rnr_nak_retry_err",
3081         "packet_seq_err",
3082         "implied_nak_seq_err",
3083         "local_ack_timeout_err",
3084 };
3085
3086 static const size_t stats_offsets[] = {
3087         MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests),
3088         MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests),
3089         MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests),
3090         MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer),
3091         MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence),
3092         MLX5_BYTE_OFF(query_q_counter_out, duplicate_request),
3093         MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err),
3094         MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err),
3095         MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err),
3096         MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err),
3097 };
3098
3099 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev,
3100                                                     u8 port_num)
3101 {
3102         BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets));
3103
3104         /* We support only per port stats */
3105         if (port_num == 0)
3106                 return NULL;
3107
3108         return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names),
3109                                           RDMA_HW_STATS_DEFAULT_LIFESPAN);
3110 }
3111
3112 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev,
3113                                 struct rdma_hw_stats *stats,
3114                                 u8 port, int index)
3115 {
3116         struct mlx5_ib_dev *dev = to_mdev(ibdev);
3117         int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out);
3118         void *out;
3119         __be32 val;
3120         int ret;
3121         int i;
3122
3123         if (!port || !stats)
3124                 return -ENOSYS;
3125
3126         out = mlx5_vzalloc(outlen);
3127         if (!out)
3128                 return -ENOMEM;
3129
3130         ret = mlx5_core_query_q_counter(dev->mdev,
3131                                         dev->port[port - 1].q_cnt_id, 0,
3132                                         out, outlen);
3133         if (ret)
3134                 goto free;
3135
3136         for (i = 0; i < ARRAY_SIZE(names); i++) {
3137                 val = *(__be32 *)(out + stats_offsets[i]);
3138                 stats->value[i] = (u64)be32_to_cpu(val);
3139         }
3140 free:
3141         kvfree(out);
3142         return ARRAY_SIZE(names);
3143 }
3144
3145 static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
3146 {
3147         struct mlx5_ib_dev *dev;
3148         enum rdma_link_layer ll;
3149         int port_type_cap;
3150         const char *name;
3151         int err;
3152         int i;
3153
3154         port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3155         ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3156
3157         printk_once(KERN_INFO "%s", mlx5_version);
3158
3159         dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev));
3160         if (!dev)
3161                 return NULL;
3162
3163         dev->mdev = mdev;
3164
3165         dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port),
3166                             GFP_KERNEL);
3167         if (!dev->port)
3168                 goto err_dealloc;
3169
3170         rwlock_init(&dev->roce.netdev_lock);
3171         err = get_port_caps(dev);
3172         if (err)
3173                 goto err_free_port;
3174
3175         if (mlx5_use_mad_ifc(dev))
3176                 get_ext_port_caps(dev);
3177
3178         if (!mlx5_lag_is_active(mdev))
3179                 name = "mlx5_%d";
3180         else
3181                 name = "mlx5_bond_%d";
3182
3183         strlcpy(dev->ib_dev.name, name, IB_DEVICE_NAME_MAX);
3184         dev->ib_dev.owner               = THIS_MODULE;
3185         dev->ib_dev.node_type           = RDMA_NODE_IB_CA;
3186         dev->ib_dev.local_dma_lkey      = 0 /* not supported for now */;
3187         dev->num_ports          = MLX5_CAP_GEN(mdev, num_ports);
3188         dev->ib_dev.phys_port_cnt     = dev->num_ports;
3189         dev->ib_dev.num_comp_vectors    =
3190                 dev->mdev->priv.eq_table.num_comp_vectors;
3191         dev->ib_dev.dma_device  = &mdev->pdev->dev;
3192
3193         dev->ib_dev.uverbs_abi_ver      = MLX5_IB_UVERBS_ABI_VERSION;
3194         dev->ib_dev.uverbs_cmd_mask     =
3195                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
3196                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
3197                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
3198                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
3199                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
3200                 (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
3201                 (1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
3202                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
3203                 (1ull << IB_USER_VERBS_CMD_REREG_MR)            |
3204                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
3205                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
3206                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
3207                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
3208                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
3209                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
3210                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
3211                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
3212                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
3213                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
3214                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
3215                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
3216                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
3217                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
3218                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
3219                 (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)         |
3220                 (1ull << IB_USER_VERBS_CMD_OPEN_QP);
3221         dev->ib_dev.uverbs_ex_cmd_mask =
3222                 (1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE)     |
3223                 (1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ)        |
3224                 (1ull << IB_USER_VERBS_EX_CMD_CREATE_QP)        |
3225                 (1ull << IB_USER_VERBS_EX_CMD_MODIFY_QP);
3226
3227         dev->ib_dev.query_device        = mlx5_ib_query_device;
3228         dev->ib_dev.query_port          = mlx5_ib_query_port;
3229         dev->ib_dev.get_link_layer      = mlx5_ib_port_link_layer;
3230         if (ll == IB_LINK_LAYER_ETHERNET)
3231                 dev->ib_dev.get_netdev  = mlx5_ib_get_netdev;
3232         dev->ib_dev.query_gid           = mlx5_ib_query_gid;
3233         dev->ib_dev.add_gid             = mlx5_ib_add_gid;
3234         dev->ib_dev.del_gid             = mlx5_ib_del_gid;
3235         dev->ib_dev.query_pkey          = mlx5_ib_query_pkey;
3236         dev->ib_dev.modify_device       = mlx5_ib_modify_device;
3237         dev->ib_dev.modify_port         = mlx5_ib_modify_port;
3238         dev->ib_dev.alloc_ucontext      = mlx5_ib_alloc_ucontext;
3239         dev->ib_dev.dealloc_ucontext    = mlx5_ib_dealloc_ucontext;
3240         dev->ib_dev.mmap                = mlx5_ib_mmap;
3241         dev->ib_dev.alloc_pd            = mlx5_ib_alloc_pd;
3242         dev->ib_dev.dealloc_pd          = mlx5_ib_dealloc_pd;
3243         dev->ib_dev.create_ah           = mlx5_ib_create_ah;
3244         dev->ib_dev.query_ah            = mlx5_ib_query_ah;
3245         dev->ib_dev.destroy_ah          = mlx5_ib_destroy_ah;
3246         dev->ib_dev.create_srq          = mlx5_ib_create_srq;
3247         dev->ib_dev.modify_srq          = mlx5_ib_modify_srq;
3248         dev->ib_dev.query_srq           = mlx5_ib_query_srq;
3249         dev->ib_dev.destroy_srq         = mlx5_ib_destroy_srq;
3250         dev->ib_dev.post_srq_recv       = mlx5_ib_post_srq_recv;
3251         dev->ib_dev.create_qp           = mlx5_ib_create_qp;
3252         dev->ib_dev.modify_qp           = mlx5_ib_modify_qp;
3253         dev->ib_dev.query_qp            = mlx5_ib_query_qp;
3254         dev->ib_dev.destroy_qp          = mlx5_ib_destroy_qp;
3255         dev->ib_dev.post_send           = mlx5_ib_post_send;
3256         dev->ib_dev.post_recv           = mlx5_ib_post_recv;
3257         dev->ib_dev.create_cq           = mlx5_ib_create_cq;
3258         dev->ib_dev.modify_cq           = mlx5_ib_modify_cq;
3259         dev->ib_dev.resize_cq           = mlx5_ib_resize_cq;
3260         dev->ib_dev.destroy_cq          = mlx5_ib_destroy_cq;
3261         dev->ib_dev.poll_cq             = mlx5_ib_poll_cq;
3262         dev->ib_dev.req_notify_cq       = mlx5_ib_arm_cq;
3263         dev->ib_dev.get_dma_mr          = mlx5_ib_get_dma_mr;
3264         dev->ib_dev.reg_user_mr         = mlx5_ib_reg_user_mr;
3265         dev->ib_dev.rereg_user_mr       = mlx5_ib_rereg_user_mr;
3266         dev->ib_dev.dereg_mr            = mlx5_ib_dereg_mr;
3267         dev->ib_dev.attach_mcast        = mlx5_ib_mcg_attach;
3268         dev->ib_dev.detach_mcast        = mlx5_ib_mcg_detach;
3269         dev->ib_dev.process_mad         = mlx5_ib_process_mad;
3270         dev->ib_dev.alloc_mr            = mlx5_ib_alloc_mr;
3271         dev->ib_dev.map_mr_sg           = mlx5_ib_map_mr_sg;
3272         dev->ib_dev.check_mr_status     = mlx5_ib_check_mr_status;
3273         dev->ib_dev.get_port_immutable  = mlx5_port_immutable;
3274         dev->ib_dev.get_dev_fw_str      = get_dev_fw_str;
3275         if (mlx5_core_is_pf(mdev)) {
3276                 dev->ib_dev.get_vf_config       = mlx5_ib_get_vf_config;
3277                 dev->ib_dev.set_vf_link_state   = mlx5_ib_set_vf_link_state;
3278                 dev->ib_dev.get_vf_stats        = mlx5_ib_get_vf_stats;
3279                 dev->ib_dev.set_vf_guid         = mlx5_ib_set_vf_guid;
3280         }
3281
3282         dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext;
3283
3284         mlx5_ib_internal_fill_odp_caps(dev);
3285
3286         if (MLX5_CAP_GEN(mdev, imaicl)) {
3287                 dev->ib_dev.alloc_mw            = mlx5_ib_alloc_mw;
3288                 dev->ib_dev.dealloc_mw          = mlx5_ib_dealloc_mw;
3289                 dev->ib_dev.uverbs_cmd_mask |=
3290                         (1ull << IB_USER_VERBS_CMD_ALLOC_MW)    |
3291                         (1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
3292         }
3293
3294         if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) &&
3295             MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) {
3296                 dev->ib_dev.get_hw_stats        = mlx5_ib_get_hw_stats;
3297                 dev->ib_dev.alloc_hw_stats      = mlx5_ib_alloc_hw_stats;
3298         }
3299
3300         if (MLX5_CAP_GEN(mdev, xrc)) {
3301                 dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd;
3302                 dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd;
3303                 dev->ib_dev.uverbs_cmd_mask |=
3304                         (1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
3305                         (1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
3306         }
3307
3308         if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) ==
3309             IB_LINK_LAYER_ETHERNET) {
3310                 dev->ib_dev.create_flow = mlx5_ib_create_flow;
3311                 dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow;
3312                 dev->ib_dev.create_wq    = mlx5_ib_create_wq;
3313                 dev->ib_dev.modify_wq    = mlx5_ib_modify_wq;
3314                 dev->ib_dev.destroy_wq   = mlx5_ib_destroy_wq;
3315                 dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table;
3316                 dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table;
3317                 dev->ib_dev.uverbs_ex_cmd_mask |=
3318                         (1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
3319                         (1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) |
3320                         (1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
3321                         (1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) |
3322                         (1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
3323                         (1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
3324                         (1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
3325         }
3326         err = init_node_data(dev);
3327         if (err)
3328                 goto err_free_port;
3329
3330         mutex_init(&dev->flow_db.lock);
3331         mutex_init(&dev->cap_mask_mutex);
3332         INIT_LIST_HEAD(&dev->qp_list);
3333         spin_lock_init(&dev->reset_flow_resource_lock);
3334
3335         if (ll == IB_LINK_LAYER_ETHERNET) {
3336                 err = mlx5_enable_eth(dev);
3337                 if (err)
3338                         goto err_free_port;
3339         }
3340
3341         err = create_dev_resources(&dev->devr);
3342         if (err)
3343                 goto err_disable_eth;
3344
3345         err = mlx5_ib_odp_init_one(dev);
3346         if (err)
3347                 goto err_rsrc;
3348
3349         if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
3350                 err = mlx5_ib_alloc_q_counters(dev);
3351                 if (err)
3352                         goto err_odp;
3353         }
3354
3355         dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
3356         if (!dev->mdev->priv.uar)
3357                 goto err_q_cnt;
3358
3359         err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
3360         if (err)
3361                 goto err_uar_page;
3362
3363         err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
3364         if (err)
3365                 goto err_bfreg;
3366
3367         err = ib_register_device(&dev->ib_dev, NULL);
3368         if (err)
3369                 goto err_fp_bfreg;
3370
3371         err = create_umr_res(dev);
3372         if (err)
3373                 goto err_dev;
3374
3375         for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) {
3376                 err = device_create_file(&dev->ib_dev.dev,
3377                                          mlx5_class_attributes[i]);
3378                 if (err)
3379                         goto err_umrc;
3380         }
3381
3382         dev->ib_active = true;
3383
3384         return dev;
3385
3386 err_umrc:
3387         destroy_umrc_res(dev);
3388
3389 err_dev:
3390         ib_unregister_device(&dev->ib_dev);
3391
3392 err_fp_bfreg:
3393         mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3394
3395 err_bfreg:
3396         mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3397
3398 err_uar_page:
3399         mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
3400
3401 err_q_cnt:
3402         if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
3403                 mlx5_ib_dealloc_q_counters(dev);
3404
3405 err_odp:
3406         mlx5_ib_odp_remove_one(dev);
3407
3408 err_rsrc:
3409         destroy_dev_resources(&dev->devr);
3410
3411 err_disable_eth:
3412         if (ll == IB_LINK_LAYER_ETHERNET) {
3413                 mlx5_disable_eth(dev);
3414                 mlx5_remove_netdev_notifier(dev);
3415         }
3416
3417 err_free_port:
3418         kfree(dev->port);
3419
3420 err_dealloc:
3421         ib_dealloc_device((struct ib_device *)dev);
3422
3423         return NULL;
3424 }
3425
3426 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
3427 {
3428         struct mlx5_ib_dev *dev = context;
3429         enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1);
3430
3431         mlx5_remove_netdev_notifier(dev);
3432         ib_unregister_device(&dev->ib_dev);
3433         mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
3434         mlx5_free_bfreg(dev->mdev, &dev->bfreg);
3435         mlx5_put_uars_page(dev->mdev, mdev->priv.uar);
3436         if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt))
3437                 mlx5_ib_dealloc_q_counters(dev);
3438         destroy_umrc_res(dev);
3439         mlx5_ib_odp_remove_one(dev);
3440         destroy_dev_resources(&dev->devr);
3441         if (ll == IB_LINK_LAYER_ETHERNET)
3442                 mlx5_disable_eth(dev);
3443         kfree(dev->port);
3444         ib_dealloc_device(&dev->ib_dev);
3445 }
3446
3447 static struct mlx5_interface mlx5_ib_interface = {
3448         .add            = mlx5_ib_add,
3449         .remove         = mlx5_ib_remove,
3450         .event          = mlx5_ib_event,
3451 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
3452         .pfault         = mlx5_ib_pfault,
3453 #endif
3454         .protocol       = MLX5_INTERFACE_PROTOCOL_IB,
3455 };
3456
3457 static int __init mlx5_ib_init(void)
3458 {
3459         int err;
3460
3461         if (deprecated_prof_sel != 2)
3462                 pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
3463
3464         err = mlx5_register_interface(&mlx5_ib_interface);
3465
3466         return err;
3467 }
3468
3469 static void __exit mlx5_ib_cleanup(void)
3470 {
3471         mlx5_unregister_interface(&mlx5_ib_interface);
3472 }
3473
3474 module_init(mlx5_ib_init);
3475 module_exit(mlx5_ib_cleanup);