drivers/infiniband/hw/mlx5/main.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
   4  * Copyright (c) 2020, Intel Corporation. All rights reserved.
   5  */
   6
   7 #include <linux/debugfs.h>
   8 #include <linux/highmem.h>
   9 #include <linux/module.h>
  10 #include <linux/init.h>
  11 #include <linux/errno.h>
  12 #include <linux/pci.h>
  13 #include <linux/dma-mapping.h>
  14 #include <linux/slab.h>
  15 #include <linux/bitmap.h>
  16 #include <linux/sched.h>
  17 #include <linux/sched/mm.h>
  18 #include <linux/sched/task.h>
  19 #include <linux/delay.h>
  20 #include <rdma/ib_user_verbs.h>
  21 #include <rdma/ib_addr.h>
  22 #include <rdma/ib_cache.h>
  23 #include <linux/mlx5/port.h>
  24 #include <linux/mlx5/vport.h>
  25 #include <linux/mlx5/fs.h>
  26 #include <linux/mlx5/eswitch.h>
  27 #include <linux/list.h>
  28 #include <rdma/ib_smi.h>
  29 #include <rdma/ib_umem.h>
  30 #include <rdma/lag.h>
  31 #include <linux/in.h>
  32 #include <linux/etherdevice.h>
  33 #include "mlx5_ib.h"
  34 #include "ib_rep.h"
  35 #include "cmd.h"
  36 #include "devx.h"
  37 #include "dm.h"
  38 #include "fs.h"
  39 #include "srq.h"
  40 #include "qp.h"
  41 #include "wr.h"
  42 #include "restrack.h"
  43 #include "counters.h"
  44 #include <linux/mlx5/accel.h>
  45 #include <rdma/uverbs_std_types.h>
  46 #include <rdma/uverbs_ioctl.h>
  47 #include <rdma/mlx5_user_ioctl_verbs.h>
  48 #include <rdma/mlx5_user_ioctl_cmds.h>
  49 #include <rdma/ib_umem_odp.h>
  50
  51 #define UVERBS_MODULE_NAME mlx5_ib
  52 #include <rdma/uverbs_named_ioctl.h>
  53
  54 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>");
  55 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver");
  56 MODULE_LICENSE("Dual BSD/GPL");
  57
  58 struct mlx5_ib_event_work {
  59         struct work_struct      work;
  60         union {
  61                 struct mlx5_ib_dev            *dev;
  62                 struct mlx5_ib_multiport_info *mpi;
  63         };
  64         bool                    is_slave;
  65         unsigned int            event;
  66         void                    *param;
  67 };
  68
  69 enum {
  70         MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3,
  71 };
  72
  73 static struct workqueue_struct *mlx5_ib_event_wq;
  74 static LIST_HEAD(mlx5_ib_unaffiliated_port_list);
  75 static LIST_HEAD(mlx5_ib_dev_list);
  76 /*
  77  * This mutex should be held when accessing either of the above lists
  78  */
  79 static DEFINE_MUTEX(mlx5_ib_multiport_mutex);
  80
  81 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi)
  82 {
  83         struct mlx5_ib_dev *dev;
  84
  85         mutex_lock(&mlx5_ib_multiport_mutex);
  86         dev = mpi->ibdev;
  87         mutex_unlock(&mlx5_ib_multiport_mutex);
  88         return dev;
  89 }
  90
  91 static enum rdma_link_layer
  92 mlx5_port_type_cap_to_rdma_ll(int port_type_cap)
  93 {
  94         switch (port_type_cap) {
  95         case MLX5_CAP_PORT_TYPE_IB:
  96                 return IB_LINK_LAYER_INFINIBAND;
  97         case MLX5_CAP_PORT_TYPE_ETH:
  98                 return IB_LINK_LAYER_ETHERNET;
  99         default:
 100                 return IB_LINK_LAYER_UNSPECIFIED;
 101         }
 102 }
 103
 104 static enum rdma_link_layer
 105 mlx5_ib_port_link_layer(struct ib_device *device, u32 port_num)
 106 {
 107         struct mlx5_ib_dev *dev = to_mdev(device);
 108         int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type);
 109
 110         return mlx5_port_type_cap_to_rdma_ll(port_type_cap);
 111 }
 112
 113 static int get_port_state(struct ib_device *ibdev,
 114                           u32 port_num,
 115                           enum ib_port_state *state)
 116 {
 117         struct ib_port_attr attr;
 118         int ret;
 119
 120         memset(&attr, 0, sizeof(attr));
 121         ret = ibdev->ops.query_port(ibdev, port_num, &attr);
 122         if (!ret)
 123                 *state = attr.state;
 124         return ret;
 125 }
 126
 127 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
 128                                            struct net_device *ndev,
 129                                            u32 *port_num)
 130 {
 131         struct net_device *rep_ndev;
 132         struct mlx5_ib_port *port;
 133         int i;
 134
 135         for (i = 0; i < dev->num_ports; i++) {
 136                 port  = &dev->port[i];
 137                 if (!port->rep)
 138                         continue;
 139
 140                 read_lock(&port->roce.netdev_lock);
 141                 rep_ndev = mlx5_ib_get_rep_netdev(port->rep->esw,
 142                                                   port->rep->vport);
 143                 if (rep_ndev == ndev) {
 144                         read_unlock(&port->roce.netdev_lock);
 145                         *port_num = i + 1;
 146                         return &port->roce;
 147                 }
 148                 read_unlock(&port->roce.netdev_lock);
 149         }
 150
 151         return NULL;
 152 }
 153
 154 static int mlx5_netdev_event(struct notifier_block *this,
 155                              unsigned long event, void *ptr)
 156 {
 157         struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
 158         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
 159         u32 port_num = roce->native_port_num;
 160         struct mlx5_core_dev *mdev;
 161         struct mlx5_ib_dev *ibdev;
 162
 163         ibdev = roce->dev;
 164         mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
 165         if (!mdev)
 166                 return NOTIFY_DONE;
 167
 168         switch (event) {
 169         case NETDEV_REGISTER:
 170                 /* Should already be registered during the load */
 171                 if (ibdev->is_rep)
 172                         break;
 173                 write_lock(&roce->netdev_lock);
 174                 if (ndev->dev.parent == mdev->device)
 175                         roce->netdev = ndev;
 176                 write_unlock(&roce->netdev_lock);
 177                 break;
 178
 179         case NETDEV_UNREGISTER:
 180                 /* In case of reps, ib device goes away before the netdevs */
 181                 write_lock(&roce->netdev_lock);
 182                 if (roce->netdev == ndev)
 183                         roce->netdev = NULL;
 184                 write_unlock(&roce->netdev_lock);
 185                 break;
 186
 187         case NETDEV_CHANGE:
 188         case NETDEV_UP:
 189         case NETDEV_DOWN: {
 190                 struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
 191                 struct net_device *upper = NULL;
 192
 193                 if (lag_ndev) {
 194                         upper = netdev_master_upper_dev_get(lag_ndev);
 195                         dev_put(lag_ndev);
 196                 }
 197
 198                 if (ibdev->is_rep)
 199                         roce = mlx5_get_rep_roce(ibdev, ndev, &port_num);
 200                 if (!roce)
 201                         return NOTIFY_DONE;
 202                 if ((upper == ndev || (!upper && ndev == roce->netdev))
 203                     && ibdev->ib_active) {
 204                         struct ib_event ibev = { };
 205                         enum ib_port_state port_state;
 206
 207                         if (get_port_state(&ibdev->ib_dev, port_num,
 208                                            &port_state))
 209                                 goto done;
 210
 211                         if (roce->last_port_state == port_state)
 212                                 goto done;
 213
 214                         roce->last_port_state = port_state;
 215                         ibev.device = &ibdev->ib_dev;
 216                         if (port_state == IB_PORT_DOWN)
 217                                 ibev.event = IB_EVENT_PORT_ERR;
 218                         else if (port_state == IB_PORT_ACTIVE)
 219                                 ibev.event = IB_EVENT_PORT_ACTIVE;
 220                         else
 221                                 goto done;
 222
 223                         ibev.element.port_num = port_num;
 224                         ib_dispatch_event(&ibev);
 225                 }
 226                 break;
 227         }
 228
 229         default:
 230                 break;
 231         }
 232 done:
 233         mlx5_ib_put_native_port_mdev(ibdev, port_num);
 234         return NOTIFY_DONE;
 235 }
 236
 237 static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
 238                                              u32 port_num)
 239 {
 240         struct mlx5_ib_dev *ibdev = to_mdev(device);
 241         struct net_device *ndev;
 242         struct mlx5_core_dev *mdev;
 243
 244         mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL);
 245         if (!mdev)
 246                 return NULL;
 247
 248         ndev = mlx5_lag_get_roce_netdev(mdev);
 249         if (ndev)
 250                 goto out;
 251
 252         /* Ensure ndev does not disappear before we invoke dev_hold()
 253          */
 254         read_lock(&ibdev->port[port_num - 1].roce.netdev_lock);
 255         ndev = ibdev->port[port_num - 1].roce.netdev;
 256         if (ndev)
 257                 dev_hold(ndev);
 258         read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock);
 259
 260 out:
 261         mlx5_ib_put_native_port_mdev(ibdev, port_num);
 262         return ndev;
 263 }
 264
 265 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev,
 266                                                    u32 ib_port_num,
 267                                                    u32 *native_port_num)
 268 {
 269         enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
 270                                                           ib_port_num);
 271         struct mlx5_core_dev *mdev = NULL;
 272         struct mlx5_ib_multiport_info *mpi;
 273         struct mlx5_ib_port *port;
 274
 275         if (!mlx5_core_mp_enabled(ibdev->mdev) ||
 276             ll != IB_LINK_LAYER_ETHERNET) {
 277                 if (native_port_num)
 278                         *native_port_num = ib_port_num;
 279                 return ibdev->mdev;
 280         }
 281
 282         if (native_port_num)
 283                 *native_port_num = 1;
 284
 285         port = &ibdev->port[ib_port_num - 1];
 286         spin_lock(&port->mp.mpi_lock);
 287         mpi = ibdev->port[ib_port_num - 1].mp.mpi;
 288         if (mpi && !mpi->unaffiliate) {
 289                 mdev = mpi->mdev;
 290                 /* If it's the master no need to refcount, it'll exist
 291                  * as long as the ib_dev exists.
 292                  */
 293                 if (!mpi->is_master)
 294                         mpi->mdev_refcnt++;
 295         }
 296         spin_unlock(&port->mp.mpi_lock);
 297
 298         return mdev;
 299 }
 300
 301 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 port_num)
 302 {
 303         enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev,
 304                                                           port_num);
 305         struct mlx5_ib_multiport_info *mpi;
 306         struct mlx5_ib_port *port;
 307
 308         if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
 309                 return;
 310
 311         port = &ibdev->port[port_num - 1];
 312
 313         spin_lock(&port->mp.mpi_lock);
 314         mpi = ibdev->port[port_num - 1].mp.mpi;
 315         if (mpi->is_master)
 316                 goto out;
 317
 318         mpi->mdev_refcnt--;
 319         if (mpi->unaffiliate)
 320                 complete(&mpi->unref_comp);
 321 out:
 322         spin_unlock(&port->mp.mpi_lock);
 323 }
 324
 325 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper,
 326                                            u16 *active_speed, u8 *active_width)
 327 {
 328         switch (eth_proto_oper) {
 329         case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII):
 330         case MLX5E_PROT_MASK(MLX5E_1000BASE_KX):
 331         case MLX5E_PROT_MASK(MLX5E_100BASE_TX):
 332         case MLX5E_PROT_MASK(MLX5E_1000BASE_T):
 333                 *active_width = IB_WIDTH_1X;
 334                 *active_speed = IB_SPEED_SDR;
 335                 break;
 336         case MLX5E_PROT_MASK(MLX5E_10GBASE_T):
 337         case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4):
 338         case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4):
 339         case MLX5E_PROT_MASK(MLX5E_10GBASE_KR):
 340         case MLX5E_PROT_MASK(MLX5E_10GBASE_CR):
 341         case MLX5E_PROT_MASK(MLX5E_10GBASE_SR):
 342         case MLX5E_PROT_MASK(MLX5E_10GBASE_ER):
 343                 *active_width = IB_WIDTH_1X;
 344                 *active_speed = IB_SPEED_QDR;
 345                 break;
 346         case MLX5E_PROT_MASK(MLX5E_25GBASE_CR):
 347         case MLX5E_PROT_MASK(MLX5E_25GBASE_KR):
 348         case MLX5E_PROT_MASK(MLX5E_25GBASE_SR):
 349                 *active_width = IB_WIDTH_1X;
 350                 *active_speed = IB_SPEED_EDR;
 351                 break;
 352         case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4):
 353         case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4):
 354         case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4):
 355         case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4):
 356                 *active_width = IB_WIDTH_4X;
 357                 *active_speed = IB_SPEED_QDR;
 358                 break;
 359         case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2):
 360         case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2):
 361         case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2):
 362                 *active_width = IB_WIDTH_1X;
 363                 *active_speed = IB_SPEED_HDR;
 364                 break;
 365         case MLX5E_PROT_MASK(MLX5E_56GBASE_R4):
 366                 *active_width = IB_WIDTH_4X;
 367                 *active_speed = IB_SPEED_FDR;
 368                 break;
 369         case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4):
 370         case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4):
 371         case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4):
 372         case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4):
 373                 *active_width = IB_WIDTH_4X;
 374                 *active_speed = IB_SPEED_EDR;
 375                 break;
 376         default:
 377                 return -EINVAL;
 378         }
 379
 380         return 0;
 381 }
 382
 383 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed,
 384                                         u8 *active_width)
 385 {
 386         switch (eth_proto_oper) {
 387         case MLX5E_PROT_MASK(MLX5E_SGMII_100M):
 388         case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII):
 389                 *active_width = IB_WIDTH_1X;
 390                 *active_speed = IB_SPEED_SDR;
 391                 break;
 392         case MLX5E_PROT_MASK(MLX5E_5GBASE_R):
 393                 *active_width = IB_WIDTH_1X;
 394                 *active_speed = IB_SPEED_DDR;
 395                 break;
 396         case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1):
 397                 *active_width = IB_WIDTH_1X;
 398                 *active_speed = IB_SPEED_QDR;
 399                 break;
 400         case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4):
 401                 *active_width = IB_WIDTH_4X;
 402                 *active_speed = IB_SPEED_QDR;
 403                 break;
 404         case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR):
 405                 *active_width = IB_WIDTH_1X;
 406                 *active_speed = IB_SPEED_EDR;
 407                 break;
 408         case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2):
 409                 *active_width = IB_WIDTH_2X;
 410                 *active_speed = IB_SPEED_EDR;
 411                 break;
 412         case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR):
 413                 *active_width = IB_WIDTH_1X;
 414                 *active_speed = IB_SPEED_HDR;
 415                 break;
 416         case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4):
 417                 *active_width = IB_WIDTH_4X;
 418                 *active_speed = IB_SPEED_EDR;
 419                 break;
 420         case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2):
 421                 *active_width = IB_WIDTH_2X;
 422                 *active_speed = IB_SPEED_HDR;
 423                 break;
 424         case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR):
 425                 *active_width = IB_WIDTH_1X;
 426                 *active_speed = IB_SPEED_NDR;
 427                 break;
 428         case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4):
 429                 *active_width = IB_WIDTH_4X;
 430                 *active_speed = IB_SPEED_HDR;
 431                 break;
 432         case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2):
 433                 *active_width = IB_WIDTH_2X;
 434                 *active_speed = IB_SPEED_NDR;
 435                 break;
 436         case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4):
 437                 *active_width = IB_WIDTH_4X;
 438                 *active_speed = IB_SPEED_NDR;
 439                 break;
 440         default:
 441                 return -EINVAL;
 442         }
 443
 444         return 0;
 445 }
 446
 447 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed,
 448                                     u8 *active_width, bool ext)
 449 {
 450         return ext ?
 451                 translate_eth_ext_proto_oper(eth_proto_oper, active_speed,
 452                                              active_width) :
 453                 translate_eth_legacy_proto_oper(eth_proto_oper, active_speed,
 454                                                 active_width);
 455 }
 456
 457 static int mlx5_query_port_roce(struct ib_device *device, u32 port_num,
 458                                 struct ib_port_attr *props)
 459 {
 460         struct mlx5_ib_dev *dev = to_mdev(device);
 461         u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0};
 462         struct mlx5_core_dev *mdev;
 463         struct net_device *ndev, *upper;
 464         enum ib_mtu ndev_ib_mtu;
 465         bool put_mdev = true;
 466         u32 eth_prot_oper;
 467         u32 mdev_port_num;
 468         bool ext;
 469         int err;
 470
 471         mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
 472         if (!mdev) {
 473                 /* This means the port isn't affiliated yet. Get the
 474                  * info for the master port instead.
 475                  */
 476                 put_mdev = false;
 477                 mdev = dev->mdev;
 478                 mdev_port_num = 1;
 479                 port_num = 1;
 480         }
 481
 482         /* Possible bad flows are checked before filling out props so in case
 483          * of an error it will still be zeroed out.
 484          * Use native port in case of reps
 485          */
 486         if (dev->is_rep)
 487                 err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
 488                                            1);
 489         else
 490                 err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN,
 491                                            mdev_port_num);
 492         if (err)
 493                 goto out;
 494         ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability);
 495         eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper);
 496
 497         props->active_width     = IB_WIDTH_4X;
 498         props->active_speed     = IB_SPEED_QDR;
 499
 500         translate_eth_proto_oper(eth_prot_oper, &props->active_speed,
 501                                  &props->active_width, ext);
 502
 503         if (!dev->is_rep && dev->mdev->roce.roce_en) {
 504                 u16 qkey_viol_cntr;
 505
 506                 props->port_cap_flags |= IB_PORT_CM_SUP;
 507                 props->ip_gids = true;
 508                 props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev,
 509                                                    roce_address_table_size);
 510                 mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr);
 511                 props->qkey_viol_cntr = qkey_viol_cntr;
 512         }
 513         props->max_mtu          = IB_MTU_4096;
 514         props->max_msg_sz       = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg);
 515         props->pkey_tbl_len     = 1;
 516         props->state            = IB_PORT_DOWN;
 517         props->phys_state       = IB_PORT_PHYS_STATE_DISABLED;
 518
 519         /* If this is a stub query for an unaffiliated port stop here */
 520         if (!put_mdev)
 521                 goto out;
 522
 523         ndev = mlx5_ib_get_netdev(device, port_num);
 524         if (!ndev)
 525                 goto out;
 526
 527         if (dev->lag_active) {
 528                 rcu_read_lock();
 529                 upper = netdev_master_upper_dev_get_rcu(ndev);
 530                 if (upper) {
 531                         dev_put(ndev);
 532                         ndev = upper;
 533                         dev_hold(ndev);
 534                 }
 535                 rcu_read_unlock();
 536         }
 537
 538         if (netif_running(ndev) && netif_carrier_ok(ndev)) {
 539                 props->state      = IB_PORT_ACTIVE;
 540                 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 541         }
 542
 543         ndev_ib_mtu = iboe_get_mtu(ndev->mtu);
 544
 545         dev_put(ndev);
 546
 547         props->active_mtu       = min(props->max_mtu, ndev_ib_mtu);
 548 out:
 549         if (put_mdev)
 550                 mlx5_ib_put_native_port_mdev(dev, port_num);
 551         return err;
 552 }
 553
 554 static int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num,
 555                          unsigned int index, const union ib_gid *gid,
 556                          const struct ib_gid_attr *attr)
 557 {
 558         enum ib_gid_type gid_type;
 559         u16 vlan_id = 0xffff;
 560         u8 roce_version = 0;
 561         u8 roce_l3_type = 0;
 562         u8 mac[ETH_ALEN];
 563         int ret;
 564
 565         gid_type = attr->gid_type;
 566         if (gid) {
 567                 ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]);
 568                 if (ret)
 569                         return ret;
 570         }
 571
 572         switch (gid_type) {
 573         case IB_GID_TYPE_ROCE:
 574                 roce_version = MLX5_ROCE_VERSION_1;
 575                 break;
 576         case IB_GID_TYPE_ROCE_UDP_ENCAP:
 577                 roce_version = MLX5_ROCE_VERSION_2;
 578                 if (gid && ipv6_addr_v4mapped((void *)gid))
 579                         roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4;
 580                 else
 581                         roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6;
 582                 break;
 583
 584         default:
 585                 mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type);
 586         }
 587
 588         return mlx5_core_roce_gid_set(dev->mdev, index, roce_version,
 589                                       roce_l3_type, gid->raw, mac,
 590                                       vlan_id < VLAN_CFI_MASK, vlan_id,
 591                                       port_num);
 592 }
 593
 594 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr,
 595                            __always_unused void **context)
 596 {
 597         return set_roce_addr(to_mdev(attr->device), attr->port_num,
 598                              attr->index, &attr->gid, attr);
 599 }
 600
 601 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr,
 602                            __always_unused void **context)
 603 {
 604         return set_roce_addr(to_mdev(attr->device), attr->port_num,
 605                              attr->index, NULL, attr);
 606 }
 607
 608 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev,
 609                                    const struct ib_gid_attr *attr)
 610 {
 611         if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP)
 612                 return 0;
 613
 614         return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port));
 615 }
 616
 617 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev)
 618 {
 619         if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB)
 620                 return !MLX5_CAP_GEN(dev->mdev, ib_virt);
 621         return 0;
 622 }
 623
 624 enum {
 625         MLX5_VPORT_ACCESS_METHOD_MAD,
 626         MLX5_VPORT_ACCESS_METHOD_HCA,
 627         MLX5_VPORT_ACCESS_METHOD_NIC,
 628 };
 629
 630 static int mlx5_get_vport_access_method(struct ib_device *ibdev)
 631 {
 632         if (mlx5_use_mad_ifc(to_mdev(ibdev)))
 633                 return MLX5_VPORT_ACCESS_METHOD_MAD;
 634
 635         if (mlx5_ib_port_link_layer(ibdev, 1) ==
 636             IB_LINK_LAYER_ETHERNET)
 637                 return MLX5_VPORT_ACCESS_METHOD_NIC;
 638
 639         return MLX5_VPORT_ACCESS_METHOD_HCA;
 640 }
 641
 642 static void get_atomic_caps(struct mlx5_ib_dev *dev,
 643                             u8 atomic_size_qp,
 644                             struct ib_device_attr *props)
 645 {
 646         u8 tmp;
 647         u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
 648         u8 atomic_req_8B_endianness_mode =
 649                 MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode);
 650
 651         /* Check if HW supports 8 bytes standard atomic operations and capable
 652          * of host endianness respond
 653          */
 654         tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
 655         if (((atomic_operations & tmp) == tmp) &&
 656             (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) &&
 657             (atomic_req_8B_endianness_mode)) {
 658                 props->atomic_cap = IB_ATOMIC_HCA;
 659         } else {
 660                 props->atomic_cap = IB_ATOMIC_NONE;
 661         }
 662 }
 663
 664 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev,
 665                                struct ib_device_attr *props)
 666 {
 667         u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
 668
 669         get_atomic_caps(dev, atomic_size_qp, props);
 670 }
 671
 672 static int mlx5_query_system_image_guid(struct ib_device *ibdev,
 673                                         __be64 *sys_image_guid)
 674 {
 675         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 676         struct mlx5_core_dev *mdev = dev->mdev;
 677         u64 tmp;
 678         int err;
 679
 680         switch (mlx5_get_vport_access_method(ibdev)) {
 681         case MLX5_VPORT_ACCESS_METHOD_MAD:
 682                 return mlx5_query_mad_ifc_system_image_guid(ibdev,
 683                                                             sys_image_guid);
 684
 685         case MLX5_VPORT_ACCESS_METHOD_HCA:
 686                 err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp);
 687                 break;
 688
 689         case MLX5_VPORT_ACCESS_METHOD_NIC:
 690                 err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp);
 691                 break;
 692
 693         default:
 694                 return -EINVAL;
 695         }
 696
 697         if (!err)
 698                 *sys_image_guid = cpu_to_be64(tmp);
 699
 700         return err;
 701
 702 }
 703
 704 static int mlx5_query_max_pkeys(struct ib_device *ibdev,
 705                                 u16 *max_pkeys)
 706 {
 707         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 708         struct mlx5_core_dev *mdev = dev->mdev;
 709
 710         switch (mlx5_get_vport_access_method(ibdev)) {
 711         case MLX5_VPORT_ACCESS_METHOD_MAD:
 712                 return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys);
 713
 714         case MLX5_VPORT_ACCESS_METHOD_HCA:
 715         case MLX5_VPORT_ACCESS_METHOD_NIC:
 716                 *max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev,
 717                                                 pkey_table_size));
 718                 return 0;
 719
 720         default:
 721                 return -EINVAL;
 722         }
 723 }
 724
 725 static int mlx5_query_vendor_id(struct ib_device *ibdev,
 726                                 u32 *vendor_id)
 727 {
 728         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 729
 730         switch (mlx5_get_vport_access_method(ibdev)) {
 731         case MLX5_VPORT_ACCESS_METHOD_MAD:
 732                 return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id);
 733
 734         case MLX5_VPORT_ACCESS_METHOD_HCA:
 735         case MLX5_VPORT_ACCESS_METHOD_NIC:
 736                 return mlx5_core_query_vendor_id(dev->mdev, vendor_id);
 737
 738         default:
 739                 return -EINVAL;
 740         }
 741 }
 742
 743 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev,
 744                                 __be64 *node_guid)
 745 {
 746         u64 tmp;
 747         int err;
 748
 749         switch (mlx5_get_vport_access_method(&dev->ib_dev)) {
 750         case MLX5_VPORT_ACCESS_METHOD_MAD:
 751                 return mlx5_query_mad_ifc_node_guid(dev, node_guid);
 752
 753         case MLX5_VPORT_ACCESS_METHOD_HCA:
 754                 err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp);
 755                 break;
 756
 757         case MLX5_VPORT_ACCESS_METHOD_NIC:
 758                 err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp);
 759                 break;
 760
 761         default:
 762                 return -EINVAL;
 763         }
 764
 765         if (!err)
 766                 *node_guid = cpu_to_be64(tmp);
 767
 768         return err;
 769 }
 770
 771 struct mlx5_reg_node_desc {
 772         u8      desc[IB_DEVICE_NODE_DESC_MAX];
 773 };
 774
 775 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc)
 776 {
 777         struct mlx5_reg_node_desc in;
 778
 779         if (mlx5_use_mad_ifc(dev))
 780                 return mlx5_query_mad_ifc_node_desc(dev, node_desc);
 781
 782         memset(&in, 0, sizeof(in));
 783
 784         return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc,
 785                                     sizeof(struct mlx5_reg_node_desc),
 786                                     MLX5_REG_NODE_DESC, 0, 0);
 787 }
 788
 789 static int mlx5_ib_query_device(struct ib_device *ibdev,
 790                                 struct ib_device_attr *props,
 791                                 struct ib_udata *uhw)
 792 {
 793         size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
 794         struct mlx5_ib_dev *dev = to_mdev(ibdev);
 795         struct mlx5_core_dev *mdev = dev->mdev;
 796         int err = -ENOMEM;
 797         int max_sq_desc;
 798         int max_rq_sg;
 799         int max_sq_sg;
 800         u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz);
 801         bool raw_support = !mlx5_core_mp_enabled(mdev);
 802         struct mlx5_ib_query_device_resp resp = {};
 803         size_t resp_len;
 804         u64 max_tso;
 805
 806         resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
 807         if (uhw_outlen && uhw_outlen < resp_len)
 808                 return -EINVAL;
 809
 810         resp.response_length = resp_len;
 811
 812         if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
 813                 return -EINVAL;
 814
 815         memset(props, 0, sizeof(*props));
 816         err = mlx5_query_system_image_guid(ibdev,
 817                                            &props->sys_image_guid);
 818         if (err)
 819                 return err;
 820
 821         props->max_pkeys = dev->pkey_table_len;
 822
 823         err = mlx5_query_vendor_id(ibdev, &props->vendor_id);
 824         if (err)
 825                 return err;
 826
 827         props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) |
 828                 (fw_rev_min(dev->mdev) << 16) |
 829                 fw_rev_sub(dev->mdev);
 830         props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
 831                 IB_DEVICE_PORT_ACTIVE_EVENT             |
 832                 IB_DEVICE_SYS_IMAGE_GUID                |
 833                 IB_DEVICE_RC_RNR_NAK_GEN;
 834
 835         if (MLX5_CAP_GEN(mdev, pkv))
 836                 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
 837         if (MLX5_CAP_GEN(mdev, qkv))
 838                 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
 839         if (MLX5_CAP_GEN(mdev, apm))
 840                 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
 841         if (MLX5_CAP_GEN(mdev, xrc))
 842                 props->device_cap_flags |= IB_DEVICE_XRC;
 843         if (MLX5_CAP_GEN(mdev, imaicl)) {
 844                 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW |
 845                                            IB_DEVICE_MEM_WINDOW_TYPE_2B;
 846                 props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
 847                 /* We support 'Gappy' memory registration too */
 848                 props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG;
 849         }
 850         /* IB_WR_REG_MR always requires changing the entity size with UMR */
 851         if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled))
 852                 props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
 853         if (MLX5_CAP_GEN(mdev, sho)) {
 854                 props->device_cap_flags |= IB_DEVICE_INTEGRITY_HANDOVER;
 855                 /* At this stage no support for signature handover */
 856                 props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 |
 857                                       IB_PROT_T10DIF_TYPE_2 |
 858                                       IB_PROT_T10DIF_TYPE_3;
 859                 props->sig_guard_cap = IB_GUARD_T10DIF_CRC |
 860                                        IB_GUARD_T10DIF_CSUM;
 861         }
 862         if (MLX5_CAP_GEN(mdev, block_lb_mc))
 863                 props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
 864
 865         if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) {
 866                 if (MLX5_CAP_ETH(mdev, csum_cap)) {
 867                         /* Legacy bit to support old userspace libraries */
 868                         props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
 869                         props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM;
 870                 }
 871
 872                 if (MLX5_CAP_ETH(dev->mdev, vlan_cap))
 873                         props->raw_packet_caps |=
 874                                 IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
 875
 876                 if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) {
 877                         max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
 878                         if (max_tso) {
 879                                 resp.tso_caps.max_tso = 1 << max_tso;
 880                                 resp.tso_caps.supported_qpts |=
 881                                         1 << IB_QPT_RAW_PACKET;
 882                                 resp.response_length += sizeof(resp.tso_caps);
 883                         }
 884                 }
 885
 886                 if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) {
 887                         resp.rss_caps.rx_hash_function =
 888                                                 MLX5_RX_HASH_FUNC_TOEPLITZ;
 889                         resp.rss_caps.rx_hash_fields_mask =
 890                                                 MLX5_RX_HASH_SRC_IPV4 |
 891                                                 MLX5_RX_HASH_DST_IPV4 |
 892                                                 MLX5_RX_HASH_SRC_IPV6 |
 893                                                 MLX5_RX_HASH_DST_IPV6 |
 894                                                 MLX5_RX_HASH_SRC_PORT_TCP |
 895                                                 MLX5_RX_HASH_DST_PORT_TCP |
 896                                                 MLX5_RX_HASH_SRC_PORT_UDP |
 897                                                 MLX5_RX_HASH_DST_PORT_UDP |
 898                                                 MLX5_RX_HASH_INNER;
 899                         if (mlx5_accel_ipsec_device_caps(dev->mdev) &
 900                             MLX5_ACCEL_IPSEC_CAP_DEVICE)
 901                                 resp.rss_caps.rx_hash_fields_mask |=
 902                                         MLX5_RX_HASH_IPSEC_SPI;
 903                         resp.response_length += sizeof(resp.rss_caps);
 904                 }
 905         } else {
 906                 if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen)
 907                         resp.response_length += sizeof(resp.tso_caps);
 908                 if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen)
 909                         resp.response_length += sizeof(resp.rss_caps);
 910         }
 911
 912         if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
 913                 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
 914                 props->device_cap_flags |= IB_DEVICE_UD_TSO;
 915         }
 916
 917         if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) &&
 918             MLX5_CAP_GEN(dev->mdev, general_notification_event) &&
 919             raw_support)
 920                 props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP;
 921
 922         if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
 923             MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap))
 924                 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
 925
 926         if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
 927             MLX5_CAP_ETH(dev->mdev, scatter_fcs) &&
 928             raw_support) {
 929                 /* Legacy bit to support old userspace libraries */
 930                 props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS;
 931                 props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS;
 932         }
 933
 934         if (MLX5_CAP_DEV_MEM(mdev, memic)) {
 935                 props->max_dm_size =
 936                         MLX5_CAP_DEV_MEM(mdev, max_memic_size);
 937         }
 938
 939         if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS))
 940                 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
 941
 942         if (MLX5_CAP_GEN(mdev, end_pad))
 943                 props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING;
 944
 945         props->vendor_part_id      = mdev->pdev->device;
 946         props->hw_ver              = mdev->pdev->revision;
 947
 948         props->max_mr_size         = ~0ull;
 949         props->page_size_cap       = ~(min_page_size - 1);
 950         props->max_qp              = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
 951         props->max_qp_wr           = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
 952         max_rq_sg =  MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
 953                      sizeof(struct mlx5_wqe_data_seg);
 954         max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
 955         max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) -
 956                      sizeof(struct mlx5_wqe_raddr_seg)) /
 957                 sizeof(struct mlx5_wqe_data_seg);
 958         props->max_send_sge = max_sq_sg;
 959         props->max_recv_sge = max_rq_sg;
 960         props->max_sge_rd          = MLX5_MAX_SGE_RD;
 961         props->max_cq              = 1 << MLX5_CAP_GEN(mdev, log_max_cq);
 962         props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1;
 963         props->max_mr              = 1 << MLX5_CAP_GEN(mdev, log_max_mkey);
 964         props->max_pd              = 1 << MLX5_CAP_GEN(mdev, log_max_pd);
 965         props->max_qp_rd_atom      = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp);
 966         props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp);
 967         props->max_srq             = 1 << MLX5_CAP_GEN(mdev, log_max_srq);
 968         props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1;
 969         props->local_ca_ack_delay  = MLX5_CAP_GEN(mdev, local_ca_ack_delay);
 970         props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
 971         props->max_srq_sge         = max_rq_sg - 1;
 972         props->max_fast_reg_page_list_len =
 973                 1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
 974         props->max_pi_fast_reg_page_list_len =
 975                 props->max_fast_reg_page_list_len / 2;
 976         props->max_sgl_rd =
 977                 MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
 978         get_atomic_caps_qp(dev, props);
 979         props->masked_atomic_cap   = IB_ATOMIC_NONE;
 980         props->max_mcast_grp       = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);
 981         props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg);
 982         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
 983                                            props->max_mcast_grp;
 984         props->max_ah = INT_MAX;
 985         props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz);
 986         props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL;
 987
 988         if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
 989                 if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
 990                         props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
 991                 props->odp_caps = dev->odp_caps;
 992                 if (!uhw) {
 993                         /* ODP for kernel QPs is not implemented for receive
 994                          * WQEs and SRQ WQEs
 995                          */
 996                         props->odp_caps.per_transport_caps.rc_odp_caps &=
 997                                 ~(IB_ODP_SUPPORT_READ |
 998                                   IB_ODP_SUPPORT_SRQ_RECV);
 999                         props->odp_caps.per_transport_caps.uc_odp_caps &=
1000                                 ~(IB_ODP_SUPPORT_READ |
1001                                   IB_ODP_SUPPORT_SRQ_RECV);
1002                         props->odp_caps.per_transport_caps.ud_odp_caps &=
1003                                 ~(IB_ODP_SUPPORT_READ |
1004                                   IB_ODP_SUPPORT_SRQ_RECV);
1005                         props->odp_caps.per_transport_caps.xrc_odp_caps &=
1006                                 ~(IB_ODP_SUPPORT_READ |
1007                                   IB_ODP_SUPPORT_SRQ_RECV);
1008                 }
1009         }
1010
1011         if (MLX5_CAP_GEN(mdev, cd))
1012                 props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL;
1013
1014         if (mlx5_core_is_vf(mdev))
1015                 props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION;
1016
1017         if (mlx5_ib_port_link_layer(ibdev, 1) ==
1018             IB_LINK_LAYER_ETHERNET && raw_support) {
1019                 props->rss_caps.max_rwq_indirection_tables =
1020                         1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt);
1021                 props->rss_caps.max_rwq_indirection_table_size =
1022                         1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size);
1023                 props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET;
1024                 props->max_wq_type_rq =
1025                         1 << MLX5_CAP_GEN(dev->mdev, log_max_rq);
1026         }
1027
1028         if (MLX5_CAP_GEN(mdev, tag_matching)) {
1029                 props->tm_caps.max_num_tags =
1030                         (1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1;
1031                 props->tm_caps.max_ops =
1032                         1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1033                 props->tm_caps.max_sge = MLX5_TM_MAX_SGE;
1034         }
1035
1036         if (MLX5_CAP_GEN(mdev, tag_matching) &&
1037             MLX5_CAP_GEN(mdev, rndv_offload_rc)) {
1038                 props->tm_caps.flags = IB_TM_CAP_RNDV_RC;
1039                 props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE;
1040         }
1041
1042         if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) {
1043                 props->cq_caps.max_cq_moderation_count =
1044                                                 MLX5_MAX_CQ_COUNT;
1045                 props->cq_caps.max_cq_moderation_period =
1046                                                 MLX5_MAX_CQ_PERIOD;
1047         }
1048
1049         if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) {
1050                 resp.response_length += sizeof(resp.cqe_comp_caps);
1051
1052                 if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
1053                         resp.cqe_comp_caps.max_num =
1054                                 MLX5_CAP_GEN(dev->mdev,
1055                                              cqe_compression_max_num);
1056
1057                         resp.cqe_comp_caps.supported_format =
1058                                 MLX5_IB_CQE_RES_FORMAT_HASH |
1059                                 MLX5_IB_CQE_RES_FORMAT_CSUM;
1060
1061                         if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index))
1062                                 resp.cqe_comp_caps.supported_format |=
1063                                         MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX;
1064                 }
1065         }
1066
1067         if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen &&
1068             raw_support) {
1069                 if (MLX5_CAP_QOS(mdev, packet_pacing) &&
1070                     MLX5_CAP_GEN(mdev, qos)) {
1071                         resp.packet_pacing_caps.qp_rate_limit_max =
1072                                 MLX5_CAP_QOS(mdev, packet_pacing_max_rate);
1073                         resp.packet_pacing_caps.qp_rate_limit_min =
1074                                 MLX5_CAP_QOS(mdev, packet_pacing_min_rate);
1075                         resp.packet_pacing_caps.supported_qpts |=
1076                                 1 << IB_QPT_RAW_PACKET;
1077                         if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) &&
1078                             MLX5_CAP_QOS(mdev, packet_pacing_typical_size))
1079                                 resp.packet_pacing_caps.cap_flags |=
1080                                         MLX5_IB_PP_SUPPORT_BURST;
1081                 }
1082                 resp.response_length += sizeof(resp.packet_pacing_caps);
1083         }
1084
1085         if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <=
1086             uhw_outlen) {
1087                 if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
1088                         resp.mlx5_ib_support_multi_pkt_send_wqes =
1089                                 MLX5_IB_ALLOW_MPW;
1090
1091                 if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe))
1092                         resp.mlx5_ib_support_multi_pkt_send_wqes |=
1093                                 MLX5_IB_SUPPORT_EMPW;
1094
1095                 resp.response_length +=
1096                         sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
1097         }
1098
1099         if (offsetofend(typeof(resp), flags) <= uhw_outlen) {
1100                 resp.response_length += sizeof(resp.flags);
1101
1102                 if (MLX5_CAP_GEN(mdev, cqe_compression_128))
1103                         resp.flags |=
1104                                 MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP;
1105
1106                 if (MLX5_CAP_GEN(mdev, cqe_128_always))
1107                         resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD;
1108                 if (MLX5_CAP_GEN(mdev, qp_packet_based))
1109                         resp.flags |=
1110                                 MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE;
1111
1112                 resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
1113         }
1114
1115         if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) {
1116                 resp.response_length += sizeof(resp.sw_parsing_caps);
1117                 if (MLX5_CAP_ETH(mdev, swp)) {
1118                         resp.sw_parsing_caps.sw_parsing_offloads |=
1119                                 MLX5_IB_SW_PARSING;
1120
1121                         if (MLX5_CAP_ETH(mdev, swp_csum))
1122                                 resp.sw_parsing_caps.sw_parsing_offloads |=
1123                                         MLX5_IB_SW_PARSING_CSUM;
1124
1125                         if (MLX5_CAP_ETH(mdev, swp_lso))
1126                                 resp.sw_parsing_caps.sw_parsing_offloads |=
1127                                         MLX5_IB_SW_PARSING_LSO;
1128
1129                         if (resp.sw_parsing_caps.sw_parsing_offloads)
1130                                 resp.sw_parsing_caps.supported_qpts =
1131                                         BIT(IB_QPT_RAW_PACKET);
1132                 }
1133         }
1134
1135         if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen &&
1136             raw_support) {
1137                 resp.response_length += sizeof(resp.striding_rq_caps);
1138                 if (MLX5_CAP_GEN(mdev, striding_rq)) {
1139                         resp.striding_rq_caps.min_single_stride_log_num_of_bytes =
1140                                 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1141                         resp.striding_rq_caps.max_single_stride_log_num_of_bytes =
1142                                 MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES;
1143                         if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range))
1144                                 resp.striding_rq_caps
1145                                         .min_single_wqe_log_num_of_strides =
1146                                         MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1147                         else
1148                                 resp.striding_rq_caps
1149                                         .min_single_wqe_log_num_of_strides =
1150                                         MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1151                         resp.striding_rq_caps.max_single_wqe_log_num_of_strides =
1152                                 MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES;
1153                         resp.striding_rq_caps.supported_qpts =
1154                                 BIT(IB_QPT_RAW_PACKET);
1155                 }
1156         }
1157
1158         if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) {
1159                 resp.response_length += sizeof(resp.tunnel_offloads_caps);
1160                 if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
1161                         resp.tunnel_offloads_caps |=
1162                                 MLX5_IB_TUNNELED_OFFLOADS_VXLAN;
1163                 if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx))
1164                         resp.tunnel_offloads_caps |=
1165                                 MLX5_IB_TUNNELED_OFFLOADS_GENEVE;
1166                 if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre))
1167                         resp.tunnel_offloads_caps |=
1168                                 MLX5_IB_TUNNELED_OFFLOADS_GRE;
1169                 if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre))
1170                         resp.tunnel_offloads_caps |=
1171                                 MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE;
1172                 if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp))
1173                         resp.tunnel_offloads_caps |=
1174                                 MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
1175         }
1176
1177         if (uhw_outlen) {
1178                 err = ib_copy_to_udata(uhw, &resp, resp.response_length);
1179
1180                 if (err)
1181                         return err;
1182         }
1183
1184         return 0;
1185 }
1186
1187 static void translate_active_width(struct ib_device *ibdev, u16 active_width,
1188                                    u8 *ib_width)
1189 {
1190         struct mlx5_ib_dev *dev = to_mdev(ibdev);
1191
1192         if (active_width & MLX5_PTYS_WIDTH_1X)
1193                 *ib_width = IB_WIDTH_1X;
1194         else if (active_width & MLX5_PTYS_WIDTH_2X)
1195                 *ib_width = IB_WIDTH_2X;
1196         else if (active_width & MLX5_PTYS_WIDTH_4X)
1197                 *ib_width = IB_WIDTH_4X;
1198         else if (active_width & MLX5_PTYS_WIDTH_8X)
1199                 *ib_width = IB_WIDTH_8X;
1200         else if (active_width & MLX5_PTYS_WIDTH_12X)
1201                 *ib_width = IB_WIDTH_12X;
1202         else {
1203                 mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n",
1204                             active_width);
1205                 *ib_width = IB_WIDTH_4X;
1206         }
1207
1208         return;
1209 }
1210
1211 static int mlx5_mtu_to_ib_mtu(int mtu)
1212 {
1213         switch (mtu) {
1214         case 256: return 1;
1215         case 512: return 2;
1216         case 1024: return 3;
1217         case 2048: return 4;
1218         case 4096: return 5;
1219         default:
1220                 pr_warn("invalid mtu\n");
1221                 return -1;
1222         }
1223 }
1224
1225 enum ib_max_vl_num {
1226         __IB_MAX_VL_0           = 1,
1227         __IB_MAX_VL_0_1         = 2,
1228         __IB_MAX_VL_0_3         = 3,
1229         __IB_MAX_VL_0_7         = 4,
1230         __IB_MAX_VL_0_14        = 5,
1231 };
1232
1233 enum mlx5_vl_hw_cap {
1234         MLX5_VL_HW_0    = 1,
1235         MLX5_VL_HW_0_1  = 2,
1236         MLX5_VL_HW_0_2  = 3,
1237         MLX5_VL_HW_0_3  = 4,
1238         MLX5_VL_HW_0_4  = 5,
1239         MLX5_VL_HW_0_5  = 6,
1240         MLX5_VL_HW_0_6  = 7,
1241         MLX5_VL_HW_0_7  = 8,
1242         MLX5_VL_HW_0_14 = 15
1243 };
1244
1245 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap,
1246                                 u8 *max_vl_num)
1247 {
1248         switch (vl_hw_cap) {
1249         case MLX5_VL_HW_0:
1250                 *max_vl_num = __IB_MAX_VL_0;
1251                 break;
1252         case MLX5_VL_HW_0_1:
1253                 *max_vl_num = __IB_MAX_VL_0_1;
1254                 break;
1255         case MLX5_VL_HW_0_3:
1256                 *max_vl_num = __IB_MAX_VL_0_3;
1257                 break;
1258         case MLX5_VL_HW_0_7:
1259                 *max_vl_num = __IB_MAX_VL_0_7;
1260                 break;
1261         case MLX5_VL_HW_0_14:
1262                 *max_vl_num = __IB_MAX_VL_0_14;
1263                 break;
1264
1265         default:
1266                 return -EINVAL;
1267         }
1268
1269         return 0;
1270 }
1271
1272 static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port,
1273                                struct ib_port_attr *props)
1274 {
1275         struct mlx5_ib_dev *dev = to_mdev(ibdev);
1276         struct mlx5_core_dev *mdev = dev->mdev;
1277         struct mlx5_hca_vport_context *rep;
1278         u16 max_mtu;
1279         u16 oper_mtu;
1280         int err;
1281         u16 ib_link_width_oper;
1282         u8 vl_hw_cap;
1283
1284         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1285         if (!rep) {
1286                 err = -ENOMEM;
1287                 goto out;
1288         }
1289
1290         /* props being zeroed by the caller, avoid zeroing it here */
1291
1292         err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep);
1293         if (err)
1294                 goto out;
1295
1296         props->lid              = rep->lid;
1297         props->lmc              = rep->lmc;
1298         props->sm_lid           = rep->sm_lid;
1299         props->sm_sl            = rep->sm_sl;
1300         props->state            = rep->vport_state;
1301         props->phys_state       = rep->port_physical_state;
1302         props->port_cap_flags   = rep->cap_mask1;
1303         props->gid_tbl_len      = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size));
1304         props->max_msg_sz       = 1 << MLX5_CAP_GEN(mdev, log_max_msg);
1305         props->pkey_tbl_len     = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size));
1306         props->bad_pkey_cntr    = rep->pkey_violation_counter;
1307         props->qkey_viol_cntr   = rep->qkey_violation_counter;
1308         props->subnet_timeout   = rep->subnet_timeout;
1309         props->init_type_reply  = rep->init_type_reply;
1310
1311         if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP)
1312                 props->port_cap_flags2 = rep->cap_mask2;
1313
1314         err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper,
1315                                       &props->active_speed, port);
1316         if (err)
1317                 goto out;
1318
1319         translate_active_width(ibdev, ib_link_width_oper, &props->active_width);
1320
1321         mlx5_query_port_max_mtu(mdev, &max_mtu, port);
1322
1323         props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu);
1324
1325         mlx5_query_port_oper_mtu(mdev, &oper_mtu, port);
1326
1327         props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu);
1328
1329         err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port);
1330         if (err)
1331                 goto out;
1332
1333         err = translate_max_vl_num(ibdev, vl_hw_cap,
1334                                    &props->max_vl_num);
1335 out:
1336         kfree(rep);
1337         return err;
1338 }
1339
1340 int mlx5_ib_query_port(struct ib_device *ibdev, u32 port,
1341                        struct ib_port_attr *props)
1342 {
1343         unsigned int count;
1344         int ret;
1345
1346         switch (mlx5_get_vport_access_method(ibdev)) {
1347         case MLX5_VPORT_ACCESS_METHOD_MAD:
1348                 ret = mlx5_query_mad_ifc_port(ibdev, port, props);
1349                 break;
1350
1351         case MLX5_VPORT_ACCESS_METHOD_HCA:
1352                 ret = mlx5_query_hca_port(ibdev, port, props);
1353                 break;
1354
1355         case MLX5_VPORT_ACCESS_METHOD_NIC:
1356                 ret = mlx5_query_port_roce(ibdev, port, props);
1357                 break;
1358
1359         default:
1360                 ret = -EINVAL;
1361         }
1362
1363         if (!ret && props) {
1364                 struct mlx5_ib_dev *dev = to_mdev(ibdev);
1365                 struct mlx5_core_dev *mdev;
1366                 bool put_mdev = true;
1367
1368                 mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL);
1369                 if (!mdev) {
1370                         /* If the port isn't affiliated yet query the master.
1371                          * The master and slave will have the same values.
1372                          */
1373                         mdev = dev->mdev;
1374                         port = 1;
1375                         put_mdev = false;
1376                 }
1377                 count = mlx5_core_reserved_gids_count(mdev);
1378                 if (put_mdev)
1379                         mlx5_ib_put_native_port_mdev(dev, port);
1380                 props->gid_tbl_len -= count;
1381         }
1382         return ret;
1383 }
1384
1385 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port,
1386                                   struct ib_port_attr *props)
1387 {
1388         return mlx5_query_port_roce(ibdev, port, props);
1389 }
1390
1391 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1392                                   u16 *pkey)
1393 {
1394         /* Default special Pkey for representor device port as per the
1395          * IB specification 1.3 section 10.9.1.2.
1396          */
1397         *pkey = 0xffff;
1398         return 0;
1399 }
1400
1401 static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
1402                              union ib_gid *gid)
1403 {
1404         struct mlx5_ib_dev *dev = to_mdev(ibdev);
1405         struct mlx5_core_dev *mdev = dev->mdev;
1406
1407         switch (mlx5_get_vport_access_method(ibdev)) {
1408         case MLX5_VPORT_ACCESS_METHOD_MAD:
1409                 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid);
1410
1411         case MLX5_VPORT_ACCESS_METHOD_HCA:
1412                 return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid);
1413
1414         default:
1415                 return -EINVAL;
1416         }
1417
1418 }
1419
1420 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port,
1421                                    u16 index, u16 *pkey)
1422 {
1423         struct mlx5_ib_dev *dev = to_mdev(ibdev);
1424         struct mlx5_core_dev *mdev;
1425         bool put_mdev = true;
1426         u32 mdev_port_num;
1427         int err;
1428
1429         mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num);
1430         if (!mdev) {
1431                 /* The port isn't affiliated yet, get the PKey from the master
1432                  * port. For RoCE the PKey tables will be the same.
1433                  */
1434                 put_mdev = false;
1435                 mdev = dev->mdev;
1436                 mdev_port_num = 1;
1437         }
1438
1439         err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0,
1440                                         index, pkey);
1441         if (put_mdev)
1442                 mlx5_ib_put_native_port_mdev(dev, port);
1443
1444         return err;
1445 }
1446
1447 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1448                               u16 *pkey)
1449 {
1450         switch (mlx5_get_vport_access_method(ibdev)) {
1451         case MLX5_VPORT_ACCESS_METHOD_MAD:
1452                 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey);
1453
1454         case MLX5_VPORT_ACCESS_METHOD_HCA:
1455         case MLX5_VPORT_ACCESS_METHOD_NIC:
1456                 return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey);
1457         default:
1458                 return -EINVAL;
1459         }
1460 }
1461
1462 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask,
1463                                  struct ib_device_modify *props)
1464 {
1465         struct mlx5_ib_dev *dev = to_mdev(ibdev);
1466         struct mlx5_reg_node_desc in;
1467         struct mlx5_reg_node_desc out;
1468         int err;
1469
1470         if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
1471                 return -EOPNOTSUPP;
1472
1473         if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
1474                 return 0;
1475
1476         /*
1477          * If possible, pass node desc to FW, so it can generate
1478          * a 144 trap.  If cmd fails, just ignore.
1479          */
1480         memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1481         err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out,
1482                                    sizeof(out), MLX5_REG_NODE_DESC, 0, 1);
1483         if (err)
1484                 return err;
1485
1486         memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
1487
1488         return err;
1489 }
1490
1491 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask,
1492                                 u32 value)
1493 {
1494         struct mlx5_hca_vport_context ctx = {};
1495         struct mlx5_core_dev *mdev;
1496         u32 mdev_port_num;
1497         int err;
1498
1499         mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num);
1500         if (!mdev)
1501                 return -ENODEV;
1502
1503         err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx);
1504         if (err)
1505                 goto out;
1506
1507         if (~ctx.cap_mask1_perm & mask) {
1508                 mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n",
1509                              mask, ctx.cap_mask1_perm);
1510                 err = -EINVAL;
1511                 goto out;
1512         }
1513
1514         ctx.cap_mask1 = value;
1515         ctx.cap_mask1_perm = mask;
1516         err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num,
1517                                                  0, &ctx);
1518
1519 out:
1520         mlx5_ib_put_native_port_mdev(dev, port_num);
1521
1522         return err;
1523 }
1524
1525 static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask,
1526                                struct ib_port_modify *props)
1527 {
1528         struct mlx5_ib_dev *dev = to_mdev(ibdev);
1529         struct ib_port_attr attr;
1530         u32 tmp;
1531         int err;
1532         u32 change_mask;
1533         u32 value;
1534         bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) ==
1535                       IB_LINK_LAYER_INFINIBAND);
1536
1537         /* CM layer calls ib_modify_port() regardless of the link layer. For
1538          * Ethernet ports, qkey violation and Port capabilities are meaningless.
1539          */
1540         if (!is_ib)
1541                 return 0;
1542
1543         if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) {
1544                 change_mask = props->clr_port_cap_mask | props->set_port_cap_mask;
1545                 value = ~props->clr_port_cap_mask | props->set_port_cap_mask;
1546                 return set_port_caps_atomic(dev, port, change_mask, value);
1547         }
1548
1549         mutex_lock(&dev->cap_mask_mutex);
1550
1551         err = ib_query_port(ibdev, port, &attr);
1552         if (err)
1553                 goto out;
1554
1555         tmp = (attr.port_cap_flags | props->set_port_cap_mask) &
1556                 ~props->clr_port_cap_mask;
1557
1558         err = mlx5_set_port_caps(dev->mdev, port, tmp);
1559
1560 out:
1561         mutex_unlock(&dev->cap_mask_mutex);
1562         return err;
1563 }
1564
1565 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
1566 {
1567         mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n",
1568                     caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
1569 }
1570
1571 static u16 calc_dynamic_bfregs(int uars_per_sys_page)
1572 {
1573         /* Large page with non 4k uar support might limit the dynamic size */
1574         if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
1575                 return MLX5_MIN_DYN_BFREGS;
1576
1577         return MLX5_MAX_DYN_BFREGS;
1578 }
1579
1580 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
1581                              struct mlx5_ib_alloc_ucontext_req_v2 *req,
1582                              struct mlx5_bfreg_info *bfregi)
1583 {
1584         int uars_per_sys_page;
1585         int bfregs_per_sys_page;
1586         int ref_bfregs = req->total_num_bfregs;
1587
1588         if (req->total_num_bfregs == 0)
1589                 return -EINVAL;
1590
1591         BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE);
1592         BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE);
1593
1594         if (req->total_num_bfregs > MLX5_MAX_BFREGS)
1595                 return -ENOMEM;
1596
1597         uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
1598         bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
1599         /* This holds the required static allocation asked by the user */
1600         req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
1601         if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
1602                 return -EINVAL;
1603
1604         bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
1605         bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
1606         bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
1607         bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
1608
1609         mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, allocated %d, total bfregs %d, using %d sys pages\n",
1610                     MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
1611                     lib_uar_4k ? "yes" : "no", ref_bfregs,
1612                     req->total_num_bfregs, bfregi->total_num_bfregs,
1613                     bfregi->num_sys_pages);
1614
1615         return 0;
1616 }
1617
1618 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context)
1619 {
1620         struct mlx5_bfreg_info *bfregi;
1621         int err;
1622         int i;
1623
1624         bfregi = &context->bfregi;
1625         for (i = 0; i < bfregi->num_static_sys_pages; i++) {
1626                 err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
1627                 if (err)
1628                         goto error;
1629
1630                 mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
1631         }
1632
1633         for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
1634                 bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
1635
1636         return 0;
1637
1638 error:
1639         for (--i; i >= 0; i--)
1640                 if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]))
1641                         mlx5_ib_warn(dev, "failed to free uar %d\n", i);
1642
1643         return err;
1644 }
1645
1646 static void deallocate_uars(struct mlx5_ib_dev *dev,
1647                             struct mlx5_ib_ucontext *context)
1648 {
1649         struct mlx5_bfreg_info *bfregi;
1650         int i;
1651
1652         bfregi = &context->bfregi;
1653         for (i = 0; i < bfregi->num_sys_pages; i++)
1654                 if (i < bfregi->num_static_sys_pages ||
1655                     bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX)
1656                         mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
1657 }
1658
1659 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1660 {
1661         int err = 0;
1662
1663         mutex_lock(&dev->lb.mutex);
1664         if (td)
1665                 dev->lb.user_td++;
1666         if (qp)
1667                 dev->lb.qps++;
1668
1669         if (dev->lb.user_td == 2 ||
1670             dev->lb.qps == 1) {
1671                 if (!dev->lb.enabled) {
1672                         err = mlx5_nic_vport_update_local_lb(dev->mdev, true);
1673                         dev->lb.enabled = true;
1674                 }
1675         }
1676
1677         mutex_unlock(&dev->lb.mutex);
1678
1679         return err;
1680 }
1681
1682 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
1683 {
1684         mutex_lock(&dev->lb.mutex);
1685         if (td)
1686                 dev->lb.user_td--;
1687         if (qp)
1688                 dev->lb.qps--;
1689
1690         if (dev->lb.user_td == 1 &&
1691             dev->lb.qps == 0) {
1692                 if (dev->lb.enabled) {
1693                         mlx5_nic_vport_update_local_lb(dev->mdev, false);
1694                         dev->lb.enabled = false;
1695                 }
1696         }
1697
1698         mutex_unlock(&dev->lb.mutex);
1699 }
1700
1701 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn,
1702                                           u16 uid)
1703 {
1704         int err;
1705
1706         if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1707                 return 0;
1708
1709         err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid);
1710         if (err)
1711                 return err;
1712
1713         if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1714             (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1715              !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1716                 return err;
1717
1718         return mlx5_ib_enable_lb(dev, true, false);
1719 }
1720
1721 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn,
1722                                              u16 uid)
1723 {
1724         if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain))
1725                 return;
1726
1727         mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid);
1728
1729         if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) ||
1730             (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) &&
1731              !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
1732                 return;
1733
1734         mlx5_ib_disable_lb(dev, true, false);
1735 }
1736
1737 static int set_ucontext_resp(struct ib_ucontext *uctx,
1738                              struct mlx5_ib_alloc_ucontext_resp *resp)
1739 {
1740         struct ib_device *ibdev = uctx->device;
1741         struct mlx5_ib_dev *dev = to_mdev(ibdev);
1742         struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1743         struct mlx5_bfreg_info *bfregi = &context->bfregi;
1744         int err;
1745
1746         if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) {
1747                 err = mlx5_cmd_dump_fill_mkey(dev->mdev,
1748                                               &resp->dump_fill_mkey);
1749                 if (err)
1750                         return err;
1751                 resp->comp_mask |=
1752                         MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY;
1753         }
1754
1755         resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp);
1756         if (dev->wc_support)
1757                 resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev,
1758                                                       log_bf_reg_size);
1759         resp->cache_line_size = cache_line_size();
1760         resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq);
1761         resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq);
1762         resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1763         resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz);
1764         resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
1765         resp->cqe_version = context->cqe_version;
1766         resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1767                                 MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT;
1768         resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
1769                                         MLX5_CAP_GEN(dev->mdev,
1770                                                      num_of_uars_per_page) : 1;
1771
1772         if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1773                                 MLX5_ACCEL_IPSEC_CAP_DEVICE) {
1774                 if (mlx5_get_flow_namespace(dev->mdev,
1775                                 MLX5_FLOW_NAMESPACE_EGRESS))
1776                         resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM;
1777                 if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1778                                 MLX5_ACCEL_IPSEC_CAP_REQUIRED_METADATA)
1779                         resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_REQ_METADATA;
1780                 if (MLX5_CAP_FLOWTABLE(dev->mdev, flow_table_properties_nic_receive.ft_field_support.outer_esp_spi))
1781                         resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_SPI_STEERING;
1782                 if (mlx5_accel_ipsec_device_caps(dev->mdev) &
1783                                 MLX5_ACCEL_IPSEC_CAP_TX_IV_IS_ESN)
1784                         resp->flow_action_flags |= MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_TX_IV_IS_ESN;
1785                 /* MLX5_USER_ALLOC_UCONTEXT_FLOW_ACTION_FLAGS_ESP_AES_GCM_FULL_OFFLOAD is currently always 0 */
1786         }
1787
1788         resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 :
1789                         bfregi->total_num_bfregs - bfregi->num_dyn_bfregs;
1790         resp->num_ports = dev->num_ports;
1791         resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE |
1792                                       MLX5_USER_CMDS_SUPP_UHW_CREATE_AH;
1793
1794         if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) {
1795                 mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline);
1796                 resp->eth_min_inline++;
1797         }
1798
1799         if (dev->mdev->clock_info)
1800                 resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1);
1801
1802         /*
1803          * We don't want to expose information from the PCI bar that is located
1804          * after 4096 bytes, so if the arch only supports larger pages, let's
1805          * pretend we don't support reading the HCA's core clock. This is also
1806          * forced by mmap function.
1807          */
1808         if (PAGE_SIZE <= 4096) {
1809                 resp->comp_mask |=
1810                         MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET;
1811                 resp->hca_core_clock_offset =
1812                         offsetof(struct mlx5_init_seg,
1813                                  internal_timer_h) % PAGE_SIZE;
1814         }
1815
1816         if (MLX5_CAP_GEN(dev->mdev, ece_support))
1817                 resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
1818
1819         resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
1820
1821         if (MLX5_CAP_GEN(dev->mdev, drain_sigerr))
1822                 resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS;
1823
1824         return 0;
1825 }
1826
1827 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx,
1828                                   struct ib_udata *udata)
1829 {
1830         struct ib_device *ibdev = uctx->device;
1831         struct mlx5_ib_dev *dev = to_mdev(ibdev);
1832         struct mlx5_ib_alloc_ucontext_req_v2 req = {};
1833         struct mlx5_ib_alloc_ucontext_resp resp = {};
1834         struct mlx5_ib_ucontext *context = to_mucontext(uctx);
1835         struct mlx5_bfreg_info *bfregi;
1836         int ver;
1837         int err;
1838         size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2,
1839                                      max_cqe_version);
1840         bool lib_uar_4k;
1841         bool lib_uar_dyn;
1842
1843         if (!dev->ib_active)
1844                 return -EAGAIN;
1845
1846         if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req))
1847                 ver = 0;
1848         else if (udata->inlen >= min_req_v2)
1849                 ver = 2;
1850         else
1851                 return -EINVAL;
1852
1853         err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1854         if (err)
1855                 return err;
1856
1857         if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX)
1858                 return -EOPNOTSUPP;
1859
1860         if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2)
1861                 return -EOPNOTSUPP;
1862
1863         req.total_num_bfregs = ALIGN(req.total_num_bfregs,
1864                                     MLX5_NON_FP_BFREGS_PER_UAR);
1865         if (req.num_low_latency_bfregs > req.total_num_bfregs - 1)
1866                 return -EINVAL;
1867
1868         lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR;
1869         lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR;
1870         bfregi = &context->bfregi;
1871
1872         if (lib_uar_dyn) {
1873                 bfregi->lib_uar_dyn = lib_uar_dyn;
1874                 goto uar_done;
1875         }
1876
1877         /* updates req->total_num_bfregs */
1878         err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
1879         if (err)
1880                 goto out_ctx;
1881
1882         mutex_init(&bfregi->lock);
1883         bfregi->lib_uar_4k = lib_uar_4k;
1884         bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
1885                                 GFP_KERNEL);
1886         if (!bfregi->count) {
1887                 err = -ENOMEM;
1888                 goto out_ctx;
1889         }
1890
1891         bfregi->sys_pages = kcalloc(bfregi->num_sys_pages,
1892                                     sizeof(*bfregi->sys_pages),
1893                                     GFP_KERNEL);
1894         if (!bfregi->sys_pages) {
1895                 err = -ENOMEM;
1896                 goto out_count;
1897         }
1898
1899         err = allocate_uars(dev, context);
1900         if (err)
1901                 goto out_sys_pages;
1902
1903 uar_done:
1904         if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) {
1905                 err = mlx5_ib_devx_create(dev, true);
1906                 if (err < 0)
1907                         goto out_uars;
1908                 context->devx_uid = err;
1909         }
1910
1911         err = mlx5_ib_alloc_transport_domain(dev, &context->tdn,
1912                                              context->devx_uid);
1913         if (err)
1914                 goto out_devx;
1915
1916         INIT_LIST_HEAD(&context->db_page_list);
1917         mutex_init(&context->db_page_mutex);
1918
1919         context->cqe_version = min_t(__u8,
1920                                  (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version),
1921                                  req.max_cqe_version);
1922
1923         err = set_ucontext_resp(uctx, &resp);
1924         if (err)
1925                 goto out_mdev;
1926
1927         resp.response_length = min(udata->outlen, sizeof(resp));
1928         err = ib_copy_to_udata(udata, &resp, resp.response_length);
1929         if (err)
1930                 goto out_mdev;
1931
1932         bfregi->ver = ver;
1933         bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs;
1934         context->lib_caps = req.lib_caps;
1935         print_lib_caps(dev, context->lib_caps);
1936
1937         if (mlx5_ib_lag_should_assign_affinity(dev)) {
1938                 u32 port = mlx5_core_native_port_num(dev->mdev) - 1;
1939
1940                 atomic_set(&context->tx_port_affinity,
1941                            atomic_add_return(
1942                                    1, &dev->port[port].roce.tx_port_affinity));
1943         }
1944
1945         return 0;
1946
1947 out_mdev:
1948         mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1949 out_devx:
1950         if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX)
1951                 mlx5_ib_devx_destroy(dev, context->devx_uid);
1952
1953 out_uars:
1954         deallocate_uars(dev, context);
1955
1956 out_sys_pages:
1957         kfree(bfregi->sys_pages);
1958
1959 out_count:
1960         kfree(bfregi->count);
1961
1962 out_ctx:
1963         return err;
1964 }
1965
1966 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext,
1967                                   struct uverbs_attr_bundle *attrs)
1968 {
1969         struct mlx5_ib_alloc_ucontext_resp uctx_resp = {};
1970         int ret;
1971
1972         ret = set_ucontext_resp(ibcontext, &uctx_resp);
1973         if (ret)
1974                 return ret;
1975
1976         uctx_resp.response_length =
1977                 min_t(size_t,
1978                       uverbs_attr_get_len(attrs,
1979                                 MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX),
1980                       sizeof(uctx_resp));
1981
1982         ret = uverbs_copy_to_struct_or_zero(attrs,
1983                                         MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
1984                                         &uctx_resp,
1985                                         sizeof(uctx_resp));
1986         return ret;
1987 }
1988
1989 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1990 {
1991         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
1992         struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
1993         struct mlx5_bfreg_info *bfregi;
1994
1995         bfregi = &context->bfregi;
1996         mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid);
1997
1998         if (context->devx_uid)
1999                 mlx5_ib_devx_destroy(dev, context->devx_uid);
2000
2001         deallocate_uars(dev, context);
2002         kfree(bfregi->sys_pages);
2003         kfree(bfregi->count);
2004 }
2005
2006 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
2007                                  int uar_idx)
2008 {
2009         int fw_uars_per_page;
2010
2011         fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
2012
2013         return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
2014 }
2015
2016 static u64 uar_index2paddress(struct mlx5_ib_dev *dev,
2017                                  int uar_idx)
2018 {
2019         unsigned int fw_uars_per_page;
2020
2021         fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ?
2022                                 MLX5_UARS_IN_PAGE : 1;
2023
2024         return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE);
2025 }
2026
2027 static int get_command(unsigned long offset)
2028 {
2029         return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK;
2030 }
2031
2032 static int get_arg(unsigned long offset)
2033 {
2034         return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1);
2035 }
2036
2037 static int get_index(unsigned long offset)
2038 {
2039         return get_arg(offset);
2040 }
2041
2042 /* Index resides in an extra byte to enable larger values than 255 */
2043 static int get_extended_index(unsigned long offset)
2044 {
2045         return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
2046 }
2047
2048
2049 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
2050 {
2051 }
2052
2053 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd)
2054 {
2055         switch (cmd) {
2056         case MLX5_IB_MMAP_WC_PAGE:
2057                 return "WC";
2058         case MLX5_IB_MMAP_REGULAR_PAGE:
2059                 return "best effort WC";
2060         case MLX5_IB_MMAP_NC_PAGE:
2061                 return "NC";
2062         case MLX5_IB_MMAP_DEVICE_MEM:
2063                 return "Device Memory";
2064         default:
2065                 return NULL;
2066         }
2067 }
2068
2069 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev,
2070                                         struct vm_area_struct *vma,
2071                                         struct mlx5_ib_ucontext *context)
2072 {
2073         if ((vma->vm_end - vma->vm_start != PAGE_SIZE) ||
2074             !(vma->vm_flags & VM_SHARED))
2075                 return -EINVAL;
2076
2077         if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1)
2078                 return -EOPNOTSUPP;
2079
2080         if (vma->vm_flags & (VM_WRITE | VM_EXEC))
2081                 return -EPERM;
2082         vma->vm_flags &= ~VM_MAYWRITE;
2083
2084         if (!dev->mdev->clock_info)
2085                 return -EOPNOTSUPP;
2086
2087         return vm_insert_page(vma, vma->vm_start,
2088                               virt_to_page(dev->mdev->clock_info));
2089 }
2090
2091 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry)
2092 {
2093         struct mlx5_user_mmap_entry *mentry = to_mmmap(entry);
2094         struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device);
2095         struct mlx5_var_table *var_table = &dev->var_table;
2096
2097         switch (mentry->mmap_flag) {
2098         case MLX5_IB_MMAP_TYPE_MEMIC:
2099         case MLX5_IB_MMAP_TYPE_MEMIC_OP:
2100                 mlx5_ib_dm_mmap_free(dev, mentry);
2101                 break;
2102         case MLX5_IB_MMAP_TYPE_VAR:
2103                 mutex_lock(&var_table->bitmap_lock);
2104                 clear_bit(mentry->page_idx, var_table->bitmap);
2105                 mutex_unlock(&var_table->bitmap_lock);
2106                 kfree(mentry);
2107                 break;
2108         case MLX5_IB_MMAP_TYPE_UAR_WC:
2109         case MLX5_IB_MMAP_TYPE_UAR_NC:
2110                 mlx5_cmd_free_uar(dev->mdev, mentry->page_idx);
2111                 kfree(mentry);
2112                 break;
2113         default:
2114                 WARN_ON(true);
2115         }
2116 }
2117
2118 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
2119                     struct vm_area_struct *vma,
2120                     struct mlx5_ib_ucontext *context)
2121 {
2122         struct mlx5_bfreg_info *bfregi = &context->bfregi;
2123         int err;
2124         unsigned long idx;
2125         phys_addr_t pfn;
2126         pgprot_t prot;
2127         u32 bfreg_dyn_idx = 0;
2128         u32 uar_index;
2129         int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
2130         int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
2131                                 bfregi->num_static_sys_pages;
2132
2133         if (bfregi->lib_uar_dyn)
2134                 return -EINVAL;
2135
2136         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2137                 return -EINVAL;
2138
2139         if (dyn_uar)
2140                 idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
2141         else
2142                 idx = get_index(vma->vm_pgoff);
2143
2144         if (idx >= max_valid_idx) {
2145                 mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
2146                              idx, max_valid_idx);
2147                 return -EINVAL;
2148         }
2149
2150         switch (cmd) {
2151         case MLX5_IB_MMAP_WC_PAGE:
2152         case MLX5_IB_MMAP_ALLOC_WC:
2153         case MLX5_IB_MMAP_REGULAR_PAGE:
2154                 /* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */
2155                 prot = pgprot_writecombine(vma->vm_page_prot);
2156                 break;
2157         case MLX5_IB_MMAP_NC_PAGE:
2158                 prot = pgprot_noncached(vma->vm_page_prot);
2159                 break;
2160         default:
2161                 return -EINVAL;
2162         }
2163
2164         if (dyn_uar) {
2165                 int uars_per_page;
2166
2167                 uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
2168                 bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
2169                 if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
2170                         mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
2171                                      bfreg_dyn_idx, bfregi->total_num_bfregs);
2172                         return -EINVAL;
2173                 }
2174
2175                 mutex_lock(&bfregi->lock);
2176                 /* Fail if uar already allocated, first bfreg index of each
2177                  * page holds its count.
2178                  */
2179                 if (bfregi->count[bfreg_dyn_idx]) {
2180                         mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
2181                         mutex_unlock(&bfregi->lock);
2182                         return -EINVAL;
2183                 }
2184
2185                 bfregi->count[bfreg_dyn_idx]++;
2186                 mutex_unlock(&bfregi->lock);
2187
2188                 err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
2189                 if (err) {
2190                         mlx5_ib_warn(dev, "UAR alloc failed\n");
2191                         goto free_bfreg;
2192                 }
2193         } else {
2194                 uar_index = bfregi->sys_pages[idx];
2195         }
2196
2197         pfn = uar_index2pfn(dev, uar_index);
2198         mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
2199
2200         err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE,
2201                                 prot, NULL);
2202         if (err) {
2203                 mlx5_ib_err(dev,
2204                             "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n",
2205                             err, mmap_cmd2str(cmd));
2206                 goto err;
2207         }
2208
2209         if (dyn_uar)
2210                 bfregi->sys_pages[idx] = uar_index;
2211         return 0;
2212
2213 err:
2214         if (!dyn_uar)
2215                 return err;
2216
2217         mlx5_cmd_free_uar(dev->mdev, idx);
2218
2219 free_bfreg:
2220         mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
2221
2222         return err;
2223 }
2224
2225 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma)
2226 {
2227         unsigned long idx;
2228         u8 command;
2229
2230         command = get_command(vma->vm_pgoff);
2231         idx = get_extended_index(vma->vm_pgoff);
2232
2233         return (command << 16 | idx);
2234 }
2235
2236 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev,
2237                                struct vm_area_struct *vma,
2238                                struct ib_ucontext *ucontext)
2239 {
2240         struct mlx5_user_mmap_entry *mentry;
2241         struct rdma_user_mmap_entry *entry;
2242         unsigned long pgoff;
2243         pgprot_t prot;
2244         phys_addr_t pfn;
2245         int ret;
2246
2247         pgoff = mlx5_vma_to_pgoff(vma);
2248         entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff);
2249         if (!entry)
2250                 return -EINVAL;
2251
2252         mentry = to_mmmap(entry);
2253         pfn = (mentry->address >> PAGE_SHIFT);
2254         if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR ||
2255             mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC)
2256                 prot = pgprot_noncached(vma->vm_page_prot);
2257         else
2258                 prot = pgprot_writecombine(vma->vm_page_prot);
2259         ret = rdma_user_mmap_io(ucontext, vma, pfn,
2260                                 entry->npages * PAGE_SIZE,
2261                                 prot,
2262                                 entry);
2263         rdma_user_mmap_entry_put(&mentry->rdma_entry);
2264         return ret;
2265 }
2266
2267 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry)
2268 {
2269         u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF;
2270         u64 index = entry->rdma_entry.start_pgoff & 0xFFFF;
2271
2272         return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) |
2273                 (index & 0xFF)) << PAGE_SHIFT;
2274 }
2275
2276 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
2277 {
2278         struct mlx5_ib_ucontext *context = to_mucontext(ibcontext);
2279         struct mlx5_ib_dev *dev = to_mdev(ibcontext->device);
2280         unsigned long command;
2281         phys_addr_t pfn;
2282
2283         command = get_command(vma->vm_pgoff);
2284         switch (command) {
2285         case MLX5_IB_MMAP_WC_PAGE:
2286         case MLX5_IB_MMAP_ALLOC_WC:
2287                 if (!dev->wc_support)
2288                         return -EPERM;
2289                 fallthrough;
2290         case MLX5_IB_MMAP_NC_PAGE:
2291         case MLX5_IB_MMAP_REGULAR_PAGE:
2292                 return uar_mmap(dev, command, vma, context);
2293
2294         case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:
2295                 return -ENOSYS;
2296
2297         case MLX5_IB_MMAP_CORE_CLOCK:
2298                 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
2299                         return -EINVAL;
2300
2301                 if (vma->vm_flags & VM_WRITE)
2302                         return -EPERM;
2303                 vma->vm_flags &= ~VM_MAYWRITE;
2304
2305                 /* Don't expose to user-space information it shouldn't have */
2306                 if (PAGE_SIZE > 4096)
2307                         return -EOPNOTSUPP;
2308
2309                 pfn = (dev->mdev->iseg_base +
2310                        offsetof(struct mlx5_init_seg, internal_timer_h)) >>
2311                         PAGE_SHIFT;
2312                 return rdma_user_mmap_io(&context->ibucontext, vma, pfn,
2313                                          PAGE_SIZE,
2314                                          pgprot_noncached(vma->vm_page_prot),
2315                                          NULL);
2316         case MLX5_IB_MMAP_CLOCK_INFO:
2317                 return mlx5_ib_mmap_clock_info_page(dev, vma, context);
2318
2319         default:
2320                 return mlx5_ib_mmap_offset(dev, vma, ibcontext);
2321         }
2322
2323         return 0;
2324 }
2325
2326 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
2327 {
2328         struct mlx5_ib_pd *pd = to_mpd(ibpd);
2329         struct ib_device *ibdev = ibpd->device;
2330         struct mlx5_ib_alloc_pd_resp resp;
2331         int err;
2332         u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
2333         u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
2334         u16 uid = 0;
2335         struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
2336                 udata, struct mlx5_ib_ucontext, ibucontext);
2337
2338         uid = context ? context->devx_uid : 0;
2339         MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD);
2340         MLX5_SET(alloc_pd_in, in, uid, uid);
2341         err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out);
2342         if (err)
2343                 return err;
2344
2345         pd->pdn = MLX5_GET(alloc_pd_out, out, pd);
2346         pd->uid = uid;
2347         if (udata) {
2348                 resp.pdn = pd->pdn;
2349                 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) {
2350                         mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid);
2351                         return -EFAULT;
2352                 }
2353         }
2354
2355         return 0;
2356 }
2357
2358 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
2359 {
2360         struct mlx5_ib_dev *mdev = to_mdev(pd->device);
2361         struct mlx5_ib_pd *mpd = to_mpd(pd);
2362
2363         return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid);
2364 }
2365
2366 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2367 {
2368         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2369         struct mlx5_ib_qp *mqp = to_mqp(ibqp);
2370         int err;
2371         u16 uid;
2372
2373         uid = ibqp->pd ?
2374                 to_mpd(ibqp->pd)->uid : 0;
2375
2376         if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) {
2377                 mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n");
2378                 return -EOPNOTSUPP;
2379         }
2380
2381         err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2382         if (err)
2383                 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n",
2384                              ibqp->qp_num, gid->raw);
2385
2386         return err;
2387 }
2388
2389 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2390 {
2391         struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2392         int err;
2393         u16 uid;
2394
2395         uid = ibqp->pd ?
2396                 to_mpd(ibqp->pd)->uid : 0;
2397         err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid);
2398         if (err)
2399                 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n",
2400                              ibqp->qp_num, gid->raw);
2401
2402         return err;
2403 }
2404
2405 static int init_node_data(struct mlx5_ib_dev *dev)
2406 {
2407         int err;
2408
2409         err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc);
2410         if (err)
2411                 return err;
2412
2413         dev->mdev->rev_id = dev->mdev->pdev->revision;
2414
2415         return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid);
2416 }
2417
2418 static ssize_t fw_pages_show(struct device *device,
2419                              struct device_attribute *attr, char *buf)
2420 {
2421         struct mlx5_ib_dev *dev =
2422                 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2423
2424         return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages);
2425 }
2426 static DEVICE_ATTR_RO(fw_pages);
2427
2428 static ssize_t reg_pages_show(struct device *device,
2429                               struct device_attribute *attr, char *buf)
2430 {
2431         struct mlx5_ib_dev *dev =
2432                 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2433
2434         return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
2435 }
2436 static DEVICE_ATTR_RO(reg_pages);
2437
2438 static ssize_t hca_type_show(struct device *device,
2439                              struct device_attribute *attr, char *buf)
2440 {
2441         struct mlx5_ib_dev *dev =
2442                 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2443
2444         return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device);
2445 }
2446 static DEVICE_ATTR_RO(hca_type);
2447
2448 static ssize_t hw_rev_show(struct device *device,
2449                            struct device_attribute *attr, char *buf)
2450 {
2451         struct mlx5_ib_dev *dev =
2452                 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2453
2454         return sysfs_emit(buf, "%x\n", dev->mdev->rev_id);
2455 }
2456 static DEVICE_ATTR_RO(hw_rev);
2457
2458 static ssize_t board_id_show(struct device *device,
2459                              struct device_attribute *attr, char *buf)
2460 {
2461         struct mlx5_ib_dev *dev =
2462                 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev);
2463
2464         return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN,
2465                           dev->mdev->board_id);
2466 }
2467 static DEVICE_ATTR_RO(board_id);
2468
2469 static struct attribute *mlx5_class_attributes[] = {
2470         &dev_attr_hw_rev.attr,
2471         &dev_attr_hca_type.attr,
2472         &dev_attr_board_id.attr,
2473         &dev_attr_fw_pages.attr,
2474         &dev_attr_reg_pages.attr,
2475         NULL,
2476 };
2477
2478 static const struct attribute_group mlx5_attr_group = {
2479         .attrs = mlx5_class_attributes,
2480 };
2481
2482 static void pkey_change_handler(struct work_struct *work)
2483 {
2484         struct mlx5_ib_port_resources *ports =
2485                 container_of(work, struct mlx5_ib_port_resources,
2486                              pkey_change_work);
2487
2488         mlx5_ib_gsi_pkey_change(ports->gsi);
2489 }
2490
2491 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
2492 {
2493         struct mlx5_ib_qp *mqp;
2494         struct mlx5_ib_cq *send_mcq, *recv_mcq;
2495         struct mlx5_core_cq *mcq;
2496         struct list_head cq_armed_list;
2497         unsigned long flags_qp;
2498         unsigned long flags_cq;
2499         unsigned long flags;
2500
2501         INIT_LIST_HEAD(&cq_armed_list);
2502
2503         /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
2504         spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
2505         list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
2506                 spin_lock_irqsave(&mqp->sq.lock, flags_qp);
2507                 if (mqp->sq.tail != mqp->sq.head) {
2508                         send_mcq = to_mcq(mqp->ibqp.send_cq);
2509                         spin_lock_irqsave(&send_mcq->lock, flags_cq);
2510                         if (send_mcq->mcq.comp &&
2511                             mqp->ibqp.send_cq->comp_handler) {
2512                                 if (!send_mcq->mcq.reset_notify_added) {
2513                                         send_mcq->mcq.reset_notify_added = 1;
2514                                         list_add_tail(&send_mcq->mcq.reset_notify,
2515                                                       &cq_armed_list);
2516                                 }
2517                         }
2518                         spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
2519                 }
2520                 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
2521                 spin_lock_irqsave(&mqp->rq.lock, flags_qp);
2522                 /* no handling is needed for SRQ */
2523                 if (!mqp->ibqp.srq) {
2524                         if (mqp->rq.tail != mqp->rq.head) {
2525                                 recv_mcq = to_mcq(mqp->ibqp.recv_cq);
2526                                 spin_lock_irqsave(&recv_mcq->lock, flags_cq);
2527                                 if (recv_mcq->mcq.comp &&
2528                                     mqp->ibqp.recv_cq->comp_handler) {
2529                                         if (!recv_mcq->mcq.reset_notify_added) {
2530                                                 recv_mcq->mcq.reset_notify_added = 1;
2531                                                 list_add_tail(&recv_mcq->mcq.reset_notify,
2532                                                               &cq_armed_list);
2533                                         }
2534                                 }
2535                                 spin_unlock_irqrestore(&recv_mcq->lock,
2536                                                        flags_cq);
2537                         }
2538                 }
2539                 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
2540         }
2541         /*At that point all inflight post send were put to be executed as of we
2542          * lock/unlock above locks Now need to arm all involved CQs.
2543          */
2544         list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
2545                 mcq->comp(mcq, NULL);
2546         }
2547         spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
2548 }
2549
2550 static void delay_drop_handler(struct work_struct *work)
2551 {
2552         int err;
2553         struct mlx5_ib_delay_drop *delay_drop =
2554                 container_of(work, struct mlx5_ib_delay_drop,
2555                              delay_drop_work);
2556
2557         atomic_inc(&delay_drop->events_cnt);
2558
2559         mutex_lock(&delay_drop->lock);
2560         err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout);
2561         if (err) {
2562                 mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n",
2563                              delay_drop->timeout);
2564                 delay_drop->activate = false;
2565         }
2566         mutex_unlock(&delay_drop->lock);
2567 }
2568
2569 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2570                                  struct ib_event *ibev)
2571 {
2572         u32 port = (eqe->data.port.port >> 4) & 0xf;
2573
2574         switch (eqe->sub_type) {
2575         case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT:
2576                 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2577                                             IB_LINK_LAYER_ETHERNET)
2578                         schedule_work(&ibdev->delay_drop.delay_drop_work);
2579                 break;
2580         default: /* do nothing */
2581                 return;
2582         }
2583 }
2584
2585 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe,
2586                               struct ib_event *ibev)
2587 {
2588         u32 port = (eqe->data.port.port >> 4) & 0xf;
2589
2590         ibev->element.port_num = port;
2591
2592         switch (eqe->sub_type) {
2593         case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
2594         case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
2595         case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
2596                 /* In RoCE, port up/down events are handled in
2597                  * mlx5_netdev_event().
2598                  */
2599                 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) ==
2600                                             IB_LINK_LAYER_ETHERNET)
2601                         return -EINVAL;
2602
2603                 ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ?
2604                                 IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
2605                 break;
2606
2607         case MLX5_PORT_CHANGE_SUBTYPE_LID:
2608                 ibev->event = IB_EVENT_LID_CHANGE;
2609                 break;
2610
2611         case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
2612                 ibev->event = IB_EVENT_PKEY_CHANGE;
2613                 schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work);
2614                 break;
2615
2616         case MLX5_PORT_CHANGE_SUBTYPE_GUID:
2617                 ibev->event = IB_EVENT_GID_CHANGE;
2618                 break;
2619
2620         case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
2621                 ibev->event = IB_EVENT_CLIENT_REREGISTER;
2622                 break;
2623         default:
2624                 return -EINVAL;
2625         }
2626
2627         return 0;
2628 }
2629
2630 static void mlx5_ib_handle_event(struct work_struct *_work)
2631 {
2632         struct mlx5_ib_event_work *work =
2633                 container_of(_work, struct mlx5_ib_event_work, work);
2634         struct mlx5_ib_dev *ibdev;
2635         struct ib_event ibev;
2636         bool fatal = false;
2637
2638         if (work->is_slave) {
2639                 ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi);
2640                 if (!ibdev)
2641                         goto out;
2642         } else {
2643                 ibdev = work->dev;
2644         }
2645
2646         switch (work->event) {
2647         case MLX5_DEV_EVENT_SYS_ERROR:
2648                 ibev.event = IB_EVENT_DEVICE_FATAL;
2649                 mlx5_ib_handle_internal_error(ibdev);
2650                 ibev.element.port_num  = (u8)(unsigned long)work->param;
2651                 fatal = true;
2652                 break;
2653         case MLX5_EVENT_TYPE_PORT_CHANGE:
2654                 if (handle_port_change(ibdev, work->param, &ibev))
2655                         goto out;
2656                 break;
2657         case MLX5_EVENT_TYPE_GENERAL_EVENT:
2658                 handle_general_event(ibdev, work->param, &ibev);
2659                 fallthrough;
2660         default:
2661                 goto out;
2662         }
2663
2664         ibev.device = &ibdev->ib_dev;
2665
2666         if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) {
2667                 mlx5_ib_warn(ibdev, "warning: event on port %d\n",  ibev.element.port_num);
2668                 goto out;
2669         }
2670
2671         if (ibdev->ib_active)
2672                 ib_dispatch_event(&ibev);
2673
2674         if (fatal)
2675                 ibdev->ib_active = false;
2676 out:
2677         kfree(work);
2678 }
2679
2680 static int mlx5_ib_event(struct notifier_block *nb,
2681                          unsigned long event, void *param)
2682 {
2683         struct mlx5_ib_event_work *work;
2684
2685         work = kmalloc(sizeof(*work), GFP_ATOMIC);
2686         if (!work)
2687                 return NOTIFY_DONE;
2688
2689         INIT_WORK(&work->work, mlx5_ib_handle_event);
2690         work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events);
2691         work->is_slave = false;
2692         work->param = param;
2693         work->event = event;
2694
2695         queue_work(mlx5_ib_event_wq, &work->work);
2696
2697         return NOTIFY_OK;
2698 }
2699
2700 static int mlx5_ib_event_slave_port(struct notifier_block *nb,
2701                                     unsigned long event, void *param)
2702 {
2703         struct mlx5_ib_event_work *work;
2704
2705         work = kmalloc(sizeof(*work), GFP_ATOMIC);
2706         if (!work)
2707                 return NOTIFY_DONE;
2708
2709         INIT_WORK(&work->work, mlx5_ib_handle_event);
2710         work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events);
2711         work->is_slave = true;
2712         work->param = param;
2713         work->event = event;
2714         queue_work(mlx5_ib_event_wq, &work->work);
2715
2716         return NOTIFY_OK;
2717 }
2718
2719 static int set_has_smi_cap(struct mlx5_ib_dev *dev)
2720 {
2721         struct mlx5_hca_vport_context vport_ctx;
2722         int err;
2723         int port;
2724
2725         for (port = 1; port <= ARRAY_SIZE(dev->port_caps); port++) {
2726                 dev->port_caps[port - 1].has_smi = false;
2727                 if (MLX5_CAP_GEN(dev->mdev, port_type) ==
2728                     MLX5_CAP_PORT_TYPE_IB) {
2729                         if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
2730                                 err = mlx5_query_hca_vport_context(dev->mdev, 0,
2731                                                                    port, 0,
2732                                                                    &vport_ctx);
2733                                 if (err) {
2734                                         mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
2735                                                     port, err);
2736                                         return err;
2737                                 }
2738                                 dev->port_caps[port - 1].has_smi =
2739                                         vport_ctx.has_smi;
2740                         } else {
2741                                 dev->port_caps[port - 1].has_smi = true;
2742                         }
2743                 }
2744         }
2745         return 0;
2746 }
2747
2748 static void get_ext_port_caps(struct mlx5_ib_dev *dev)
2749 {
2750         unsigned int port;
2751
2752         rdma_for_each_port (&dev->ib_dev, port)
2753                 mlx5_query_ext_port_caps(dev, port);
2754 }
2755
2756 static u8 mlx5_get_umr_fence(u8 umr_fence_cap)
2757 {
2758         switch (umr_fence_cap) {
2759         case MLX5_CAP_UMR_FENCE_NONE:
2760                 return MLX5_FENCE_MODE_NONE;
2761         case MLX5_CAP_UMR_FENCE_SMALL:
2762                 return MLX5_FENCE_MODE_INITIATOR_SMALL;
2763         default:
2764                 return MLX5_FENCE_MODE_STRONG_ORDERING;
2765         }
2766 }
2767
2768 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev)
2769 {
2770         struct mlx5_ib_resources *devr = &dev->devr;
2771         struct ib_srq_init_attr attr;
2772         struct ib_device *ibdev;
2773         struct ib_cq_init_attr cq_attr = {.cqe = 1};
2774         int port;
2775         int ret = 0;
2776
2777         ibdev = &dev->ib_dev;
2778
2779         if (!MLX5_CAP_GEN(dev->mdev, xrc))
2780                 return -EOPNOTSUPP;
2781
2782         mutex_init(&devr->mutex);
2783
2784         devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd);
2785         if (!devr->p0)
2786                 return -ENOMEM;
2787
2788         devr->p0->device  = ibdev;
2789         devr->p0->uobject = NULL;
2790         atomic_set(&devr->p0->usecnt, 0);
2791
2792         ret = mlx5_ib_alloc_pd(devr->p0, NULL);
2793         if (ret)
2794                 goto error0;
2795
2796         devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq);
2797         if (!devr->c0) {
2798                 ret = -ENOMEM;
2799                 goto error1;
2800         }
2801
2802         devr->c0->device = &dev->ib_dev;
2803         atomic_set(&devr->c0->usecnt, 0);
2804
2805         ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL);
2806         if (ret)
2807                 goto err_create_cq;
2808
2809         ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0);
2810         if (ret)
2811                 goto error2;
2812
2813         ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0);
2814         if (ret)
2815                 goto error3;
2816
2817         memset(&attr, 0, sizeof(attr));
2818         attr.attr.max_sge = 1;
2819         attr.attr.max_wr = 1;
2820         attr.srq_type = IB_SRQT_XRC;
2821         attr.ext.cq = devr->c0;
2822
2823         devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq);
2824         if (!devr->s0) {
2825                 ret = -ENOMEM;
2826                 goto error4;
2827         }
2828
2829         devr->s0->device        = &dev->ib_dev;
2830         devr->s0->pd            = devr->p0;
2831         devr->s0->srq_type      = IB_SRQT_XRC;
2832         devr->s0->ext.cq        = devr->c0;
2833         ret = mlx5_ib_create_srq(devr->s0, &attr, NULL);
2834         if (ret)
2835                 goto err_create;
2836
2837         atomic_inc(&devr->s0->ext.cq->usecnt);
2838         atomic_inc(&devr->p0->usecnt);
2839         atomic_set(&devr->s0->usecnt, 0);
2840
2841         memset(&attr, 0, sizeof(attr));
2842         attr.attr.max_sge = 1;
2843         attr.attr.max_wr = 1;
2844         attr.srq_type = IB_SRQT_BASIC;
2845         devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq);
2846         if (!devr->s1) {
2847                 ret = -ENOMEM;
2848                 goto error5;
2849         }
2850
2851         devr->s1->device        = &dev->ib_dev;
2852         devr->s1->pd            = devr->p0;
2853         devr->s1->srq_type      = IB_SRQT_BASIC;
2854         devr->s1->ext.cq        = devr->c0;
2855
2856         ret = mlx5_ib_create_srq(devr->s1, &attr, NULL);
2857         if (ret)
2858                 goto error6;
2859
2860         atomic_inc(&devr->p0->usecnt);
2861         atomic_set(&devr->s1->usecnt, 0);
2862
2863         for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2864                 INIT_WORK(&devr->ports[port].pkey_change_work,
2865                           pkey_change_handler);
2866
2867         return 0;
2868
2869 error6:
2870         kfree(devr->s1);
2871 error5:
2872         mlx5_ib_destroy_srq(devr->s0, NULL);
2873 err_create:
2874         kfree(devr->s0);
2875 error4:
2876         mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2877 error3:
2878         mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2879 error2:
2880         mlx5_ib_destroy_cq(devr->c0, NULL);
2881 err_create_cq:
2882         kfree(devr->c0);
2883 error1:
2884         mlx5_ib_dealloc_pd(devr->p0, NULL);
2885 error0:
2886         kfree(devr->p0);
2887         return ret;
2888 }
2889
2890 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev)
2891 {
2892         struct mlx5_ib_resources *devr = &dev->devr;
2893         int port;
2894
2895         mlx5_ib_destroy_srq(devr->s1, NULL);
2896         kfree(devr->s1);
2897         mlx5_ib_destroy_srq(devr->s0, NULL);
2898         kfree(devr->s0);
2899         mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0);
2900         mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0);
2901         mlx5_ib_destroy_cq(devr->c0, NULL);
2902         kfree(devr->c0);
2903         mlx5_ib_dealloc_pd(devr->p0, NULL);
2904         kfree(devr->p0);
2905
2906         /* Make sure no change P_Key work items are still executing */
2907         for (port = 0; port < ARRAY_SIZE(devr->ports); ++port)
2908                 cancel_work_sync(&devr->ports[port].pkey_change_work);
2909 }
2910
2911 static u32 get_core_cap_flags(struct ib_device *ibdev,
2912                               struct mlx5_hca_vport_context *rep)
2913 {
2914         struct mlx5_ib_dev *dev = to_mdev(ibdev);
2915         enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1);
2916         u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type);
2917         u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version);
2918         bool raw_support = !mlx5_core_mp_enabled(dev->mdev);
2919         u32 ret = 0;
2920
2921         if (rep->grh_required)
2922                 ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED;
2923
2924         if (ll == IB_LINK_LAYER_INFINIBAND)
2925                 return ret | RDMA_CORE_PORT_IBA_IB;
2926
2927         if (raw_support)
2928                 ret |= RDMA_CORE_PORT_RAW_PACKET;
2929
2930         if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP))
2931                 return ret;
2932
2933         if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP))
2934                 return ret;
2935
2936         if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP)
2937                 ret |= RDMA_CORE_PORT_IBA_ROCE;
2938
2939         if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP)
2940                 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2941
2942         return ret;
2943 }
2944
2945 static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num,
2946                                struct ib_port_immutable *immutable)
2947 {
2948         struct ib_port_attr attr;
2949         struct mlx5_ib_dev *dev = to_mdev(ibdev);
2950         enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num);
2951         struct mlx5_hca_vport_context rep = {0};
2952         int err;
2953
2954         err = ib_query_port(ibdev, port_num, &attr);
2955         if (err)
2956                 return err;
2957
2958         if (ll == IB_LINK_LAYER_INFINIBAND) {
2959                 err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0,
2960                                                    &rep);
2961                 if (err)
2962                         return err;
2963         }
2964
2965         immutable->pkey_tbl_len = attr.pkey_tbl_len;
2966         immutable->gid_tbl_len = attr.gid_tbl_len;
2967         immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep);
2968         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2969
2970         return 0;
2971 }
2972
2973 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num,
2974                                    struct ib_port_immutable *immutable)
2975 {
2976         struct ib_port_attr attr;
2977         int err;
2978
2979         immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2980
2981         err = ib_query_port(ibdev, port_num, &attr);
2982         if (err)
2983                 return err;
2984
2985         immutable->pkey_tbl_len = attr.pkey_tbl_len;
2986         immutable->gid_tbl_len = attr.gid_tbl_len;
2987         immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
2988
2989         return 0;
2990 }
2991
2992 static void get_dev_fw_str(struct ib_device *ibdev, char *str)
2993 {
2994         struct mlx5_ib_dev *dev =
2995                 container_of(ibdev, struct mlx5_ib_dev, ib_dev);
2996         snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d",
2997                  fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev),
2998                  fw_rev_sub(dev->mdev));
2999 }
3000
3001 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev)
3002 {
3003         struct mlx5_core_dev *mdev = dev->mdev;
3004         struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev,
3005                                                                  MLX5_FLOW_NAMESPACE_LAG);
3006         struct mlx5_flow_table *ft;
3007         int err;
3008
3009         if (!ns || !mlx5_lag_is_roce(mdev))
3010                 return 0;
3011
3012         err = mlx5_cmd_create_vport_lag(mdev);
3013         if (err)
3014                 return err;
3015
3016         ft = mlx5_create_lag_demux_flow_table(ns, 0, 0);
3017         if (IS_ERR(ft)) {
3018                 err = PTR_ERR(ft);
3019                 goto err_destroy_vport_lag;
3020         }
3021
3022         dev->flow_db->lag_demux_ft = ft;
3023         dev->lag_active = true;
3024         return 0;
3025
3026 err_destroy_vport_lag:
3027         mlx5_cmd_destroy_vport_lag(mdev);
3028         return err;
3029 }
3030
3031 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
3032 {
3033         struct mlx5_core_dev *mdev = dev->mdev;
3034
3035         if (dev->lag_active) {
3036                 dev->lag_active = false;
3037
3038                 mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft);
3039                 dev->flow_db->lag_demux_ft = NULL;
3040
3041                 mlx5_cmd_destroy_vport_lag(mdev);
3042         }
3043 }
3044
3045 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3046 {
3047         int err;
3048
3049         dev->port[port_num].roce.nb.notifier_call = mlx5_netdev_event;
3050         err = register_netdevice_notifier(&dev->port[port_num].roce.nb);
3051         if (err) {
3052                 dev->port[port_num].roce.nb.notifier_call = NULL;
3053                 return err;
3054         }
3055
3056         return 0;
3057 }
3058
3059 static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u32 port_num)
3060 {
3061         if (dev->port[port_num].roce.nb.notifier_call) {
3062                 unregister_netdevice_notifier(&dev->port[port_num].roce.nb);
3063                 dev->port[port_num].roce.nb.notifier_call = NULL;
3064         }
3065 }
3066
3067 static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
3068 {
3069         int err;
3070
3071         err = mlx5_nic_vport_enable_roce(dev->mdev);
3072         if (err)
3073                 return err;
3074
3075         err = mlx5_eth_lag_init(dev);
3076         if (err)
3077                 goto err_disable_roce;
3078
3079         return 0;
3080
3081 err_disable_roce:
3082         mlx5_nic_vport_disable_roce(dev->mdev);
3083
3084         return err;
3085 }
3086
3087 static void mlx5_disable_eth(struct mlx5_ib_dev *dev)
3088 {
3089         mlx5_eth_lag_cleanup(dev);
3090         mlx5_nic_vport_disable_roce(dev->mdev);
3091 }
3092
3093 static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num,
3094                                  enum rdma_netdev_t type,
3095                                  struct rdma_netdev_alloc_params *params)
3096 {
3097         if (type != RDMA_NETDEV_IPOIB)
3098                 return -EOPNOTSUPP;
3099
3100         return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params);
3101 }
3102
3103 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf,
3104                                        size_t count, loff_t *pos)
3105 {
3106         struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3107         char lbuf[20];
3108         int len;
3109
3110         len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout);
3111         return simple_read_from_buffer(buf, count, pos, lbuf, len);
3112 }
3113
3114 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf,
3115                                         size_t count, loff_t *pos)
3116 {
3117         struct mlx5_ib_delay_drop *delay_drop = filp->private_data;
3118         u32 timeout;
3119         u32 var;
3120
3121         if (kstrtouint_from_user(buf, count, 0, &var))
3122                 return -EFAULT;
3123
3124         timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS *
3125                         1000);
3126         if (timeout != var)
3127                 mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n",
3128                             timeout);
3129
3130         delay_drop->timeout = timeout;
3131
3132         return count;
3133 }
3134
3135 static const struct file_operations fops_delay_drop_timeout = {
3136         .owner  = THIS_MODULE,
3137         .open   = simple_open,
3138         .write  = delay_drop_timeout_write,
3139         .read   = delay_drop_timeout_read,
3140 };
3141
3142 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
3143                                       struct mlx5_ib_multiport_info *mpi)
3144 {
3145         u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3146         struct mlx5_ib_port *port = &ibdev->port[port_num];
3147         int comps;
3148         int err;
3149         int i;
3150
3151         lockdep_assert_held(&mlx5_ib_multiport_mutex);
3152
3153         mlx5_ib_cleanup_cong_debugfs(ibdev, port_num);
3154
3155         spin_lock(&port->mp.mpi_lock);
3156         if (!mpi->ibdev) {
3157                 spin_unlock(&port->mp.mpi_lock);
3158                 return;
3159         }
3160
3161         mpi->ibdev = NULL;
3162
3163         spin_unlock(&port->mp.mpi_lock);
3164         if (mpi->mdev_events.notifier_call)
3165                 mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events);
3166         mpi->mdev_events.notifier_call = NULL;
3167         mlx5_remove_netdev_notifier(ibdev, port_num);
3168         spin_lock(&port->mp.mpi_lock);
3169
3170         comps = mpi->mdev_refcnt;
3171         if (comps) {
3172                 mpi->unaffiliate = true;
3173                 init_completion(&mpi->unref_comp);
3174                 spin_unlock(&port->mp.mpi_lock);
3175
3176                 for (i = 0; i < comps; i++)
3177                         wait_for_completion(&mpi->unref_comp);
3178
3179                 spin_lock(&port->mp.mpi_lock);
3180                 mpi->unaffiliate = false;
3181         }
3182
3183         port->mp.mpi = NULL;
3184
3185         spin_unlock(&port->mp.mpi_lock);
3186
3187         err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev);
3188
3189         mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1);
3190         /* Log an error, still needed to cleanup the pointers and add
3191          * it back to the list.
3192          */
3193         if (err)
3194                 mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n",
3195                             port_num + 1);
3196
3197         ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN;
3198 }
3199
3200 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
3201                                     struct mlx5_ib_multiport_info *mpi)
3202 {
3203         u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1;
3204         int err;
3205
3206         lockdep_assert_held(&mlx5_ib_multiport_mutex);
3207
3208         spin_lock(&ibdev->port[port_num].mp.mpi_lock);
3209         if (ibdev->port[port_num].mp.mpi) {
3210                 mlx5_ib_dbg(ibdev, "port %u already affiliated.\n",
3211                             port_num + 1);
3212                 spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3213                 return false;
3214         }
3215
3216         ibdev->port[port_num].mp.mpi = mpi;
3217         mpi->ibdev = ibdev;
3218         mpi->mdev_events.notifier_call = NULL;
3219         spin_unlock(&ibdev->port[port_num].mp.mpi_lock);
3220
3221         err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev);
3222         if (err)
3223                 goto unbind;
3224
3225         err = mlx5_add_netdev_notifier(ibdev, port_num);
3226         if (err) {
3227                 mlx5_ib_err(ibdev, "failed adding netdev notifier for port %u\n",
3228                             port_num + 1);
3229                 goto unbind;
3230         }
3231
3232         mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port;
3233         mlx5_notifier_register(mpi->mdev, &mpi->mdev_events);
3234
3235         mlx5_ib_init_cong_debugfs(ibdev, port_num);
3236
3237         return true;
3238
3239 unbind:
3240         mlx5_ib_unbind_slave_port(ibdev, mpi);
3241         return false;
3242 }
3243
3244 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev)
3245 {
3246         u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3247         enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3248                                                           port_num + 1);
3249         struct mlx5_ib_multiport_info *mpi;
3250         int err;
3251         u32 i;
3252
3253         if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3254                 return 0;
3255
3256         err = mlx5_query_nic_vport_system_image_guid(dev->mdev,
3257                                                      &dev->sys_image_guid);
3258         if (err)
3259                 return err;
3260
3261         err = mlx5_nic_vport_enable_roce(dev->mdev);
3262         if (err)
3263                 return err;
3264
3265         mutex_lock(&mlx5_ib_multiport_mutex);
3266         for (i = 0; i < dev->num_ports; i++) {
3267                 bool bound = false;
3268
3269                 /* build a stub multiport info struct for the native port. */
3270                 if (i == port_num) {
3271                         mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
3272                         if (!mpi) {
3273                                 mutex_unlock(&mlx5_ib_multiport_mutex);
3274                                 mlx5_nic_vport_disable_roce(dev->mdev);
3275                                 return -ENOMEM;
3276                         }
3277
3278                         mpi->is_master = true;
3279                         mpi->mdev = dev->mdev;
3280                         mpi->sys_image_guid = dev->sys_image_guid;
3281                         dev->port[i].mp.mpi = mpi;
3282                         mpi->ibdev = dev;
3283                         mpi = NULL;
3284                         continue;
3285                 }
3286
3287                 list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list,
3288                                     list) {
3289                         if (dev->sys_image_guid == mpi->sys_image_guid &&
3290                             (mlx5_core_native_port_num(mpi->mdev) - 1) == i) {
3291                                 bound = mlx5_ib_bind_slave_port(dev, mpi);
3292                         }
3293
3294                         if (bound) {
3295                                 dev_dbg(mpi->mdev->device,
3296                                         "removing port from unaffiliated list.\n");
3297                                 mlx5_ib_dbg(dev, "port %d bound\n", i + 1);
3298                                 list_del(&mpi->list);
3299                                 break;
3300                         }
3301                 }
3302                 if (!bound)
3303                         mlx5_ib_dbg(dev, "no free port found for port %d\n",
3304                                     i + 1);
3305         }
3306
3307         list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list);
3308         mutex_unlock(&mlx5_ib_multiport_mutex);
3309         return err;
3310 }
3311
3312 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev)
3313 {
3314         u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3315         enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev,
3316                                                           port_num + 1);
3317         u32 i;
3318
3319         if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET)
3320                 return;
3321
3322         mutex_lock(&mlx5_ib_multiport_mutex);
3323         for (i = 0; i < dev->num_ports; i++) {
3324                 if (dev->port[i].mp.mpi) {
3325                         /* Destroy the native port stub */
3326                         if (i == port_num) {
3327                                 kfree(dev->port[i].mp.mpi);
3328                                 dev->port[i].mp.mpi = NULL;
3329                         } else {
3330                                 mlx5_ib_dbg(dev, "unbinding port_num: %u\n",
3331                                             i + 1);
3332                                 mlx5_ib_unbind_slave_port(dev, dev->port[i].mp.mpi);
3333                                 list_add_tail(&dev->port[i].mp.mpi->list,
3334                                               &mlx5_ib_unaffiliated_port_list);
3335                         }
3336                 }
3337         }
3338
3339         mlx5_ib_dbg(dev, "removing from devlist\n");
3340         list_del(&dev->ib_dev_list);
3341         mutex_unlock(&mlx5_ib_multiport_mutex);
3342
3343         mlx5_nic_vport_disable_roce(dev->mdev);
3344 }
3345
3346 static int mmap_obj_cleanup(struct ib_uobject *uobject,
3347                             enum rdma_remove_reason why,
3348                             struct uverbs_attr_bundle *attrs)
3349 {
3350         struct mlx5_user_mmap_entry *obj = uobject->object;
3351
3352         rdma_user_mmap_entry_remove(&obj->rdma_entry);
3353         return 0;
3354 }
3355
3356 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c,
3357                                             struct mlx5_user_mmap_entry *entry,
3358                                             size_t length)
3359 {
3360         return rdma_user_mmap_entry_insert_range(
3361                 &c->ibucontext, &entry->rdma_entry, length,
3362                 (MLX5_IB_MMAP_OFFSET_START << 16),
3363                 ((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1));
3364 }
3365
3366 static struct mlx5_user_mmap_entry *
3367 alloc_var_entry(struct mlx5_ib_ucontext *c)
3368 {
3369         struct mlx5_user_mmap_entry *entry;
3370         struct mlx5_var_table *var_table;
3371         u32 page_idx;
3372         int err;
3373
3374         var_table = &to_mdev(c->ibucontext.device)->var_table;
3375         entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3376         if (!entry)
3377                 return ERR_PTR(-ENOMEM);
3378
3379         mutex_lock(&var_table->bitmap_lock);
3380         page_idx = find_first_zero_bit(var_table->bitmap,
3381                                        var_table->num_var_hw_entries);
3382         if (page_idx >= var_table->num_var_hw_entries) {
3383                 err = -ENOSPC;
3384                 mutex_unlock(&var_table->bitmap_lock);
3385                 goto end;
3386         }
3387
3388         set_bit(page_idx, var_table->bitmap);
3389         mutex_unlock(&var_table->bitmap_lock);
3390
3391         entry->address = var_table->hw_start_addr +
3392                                 (page_idx * var_table->stride_size);
3393         entry->page_idx = page_idx;
3394         entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR;
3395
3396         err = mlx5_rdma_user_mmap_entry_insert(c, entry,
3397                                                var_table->stride_size);
3398         if (err)
3399                 goto err_insert;
3400
3401         return entry;
3402
3403 err_insert:
3404         mutex_lock(&var_table->bitmap_lock);
3405         clear_bit(page_idx, var_table->bitmap);
3406         mutex_unlock(&var_table->bitmap_lock);
3407 end:
3408         kfree(entry);
3409         return ERR_PTR(err);
3410 }
3411
3412 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)(
3413         struct uverbs_attr_bundle *attrs)
3414 {
3415         struct ib_uobject *uobj = uverbs_attr_get_uobject(
3416                 attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3417         struct mlx5_ib_ucontext *c;
3418         struct mlx5_user_mmap_entry *entry;
3419         u64 mmap_offset;
3420         u32 length;
3421         int err;
3422
3423         c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3424         if (IS_ERR(c))
3425                 return PTR_ERR(c);
3426
3427         entry = alloc_var_entry(c);
3428         if (IS_ERR(entry))
3429                 return PTR_ERR(entry);
3430
3431         mmap_offset = mlx5_entry_to_mmap_offset(entry);
3432         length = entry->rdma_entry.npages * PAGE_SIZE;
3433         uobj->object = entry;
3434         uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE);
3435
3436         err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3437                              &mmap_offset, sizeof(mmap_offset));
3438         if (err)
3439                 return err;
3440
3441         err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3442                              &entry->page_idx, sizeof(entry->page_idx));
3443         if (err)
3444                 return err;
3445
3446         err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3447                              &length, sizeof(length));
3448         return err;
3449 }
3450
3451 DECLARE_UVERBS_NAMED_METHOD(
3452         MLX5_IB_METHOD_VAR_OBJ_ALLOC,
3453         UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE,
3454                         MLX5_IB_OBJECT_VAR,
3455                         UVERBS_ACCESS_NEW,
3456                         UA_MANDATORY),
3457         UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID,
3458                            UVERBS_ATTR_TYPE(u32),
3459                            UA_MANDATORY),
3460         UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH,
3461                            UVERBS_ATTR_TYPE(u32),
3462                            UA_MANDATORY),
3463         UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET,
3464                             UVERBS_ATTR_TYPE(u64),
3465                             UA_MANDATORY));
3466
3467 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3468         MLX5_IB_METHOD_VAR_OBJ_DESTROY,
3469         UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE,
3470                         MLX5_IB_OBJECT_VAR,
3471                         UVERBS_ACCESS_DESTROY,
3472                         UA_MANDATORY));
3473
3474 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR,
3475                             UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3476                             &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC),
3477                             &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY));
3478
3479 static bool var_is_supported(struct ib_device *device)
3480 {
3481         struct mlx5_ib_dev *dev = to_mdev(device);
3482
3483         return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3484                         MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q);
3485 }
3486
3487 static struct mlx5_user_mmap_entry *
3488 alloc_uar_entry(struct mlx5_ib_ucontext *c,
3489                 enum mlx5_ib_uapi_uar_alloc_type alloc_type)
3490 {
3491         struct mlx5_user_mmap_entry *entry;
3492         struct mlx5_ib_dev *dev;
3493         u32 uar_index;
3494         int err;
3495
3496         entry = kzalloc(sizeof(*entry), GFP_KERNEL);
3497         if (!entry)
3498                 return ERR_PTR(-ENOMEM);
3499
3500         dev = to_mdev(c->ibucontext.device);
3501         err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
3502         if (err)
3503                 goto end;
3504
3505         entry->page_idx = uar_index;
3506         entry->address = uar_index2paddress(dev, uar_index);
3507         if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3508                 entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC;
3509         else
3510                 entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC;
3511
3512         err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE);
3513         if (err)
3514                 goto err_insert;
3515
3516         return entry;
3517
3518 err_insert:
3519         mlx5_cmd_free_uar(dev->mdev, uar_index);
3520 end:
3521         kfree(entry);
3522         return ERR_PTR(err);
3523 }
3524
3525 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)(
3526         struct uverbs_attr_bundle *attrs)
3527 {
3528         struct ib_uobject *uobj = uverbs_attr_get_uobject(
3529                 attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3530         enum mlx5_ib_uapi_uar_alloc_type alloc_type;
3531         struct mlx5_ib_ucontext *c;
3532         struct mlx5_user_mmap_entry *entry;
3533         u64 mmap_offset;
3534         u32 length;
3535         int err;
3536
3537         c = to_mucontext(ib_uverbs_get_ucontext(attrs));
3538         if (IS_ERR(c))
3539                 return PTR_ERR(c);
3540
3541         err = uverbs_get_const(&alloc_type, attrs,
3542                                MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE);
3543         if (err)
3544                 return err;
3545
3546         if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF &&
3547             alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC)
3548                 return -EOPNOTSUPP;
3549
3550         if (!to_mdev(c->ibucontext.device)->wc_support &&
3551             alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF)
3552                 return -EOPNOTSUPP;
3553
3554         entry = alloc_uar_entry(c, alloc_type);
3555         if (IS_ERR(entry))
3556                 return PTR_ERR(entry);
3557
3558         mmap_offset = mlx5_entry_to_mmap_offset(entry);
3559         length = entry->rdma_entry.npages * PAGE_SIZE;
3560         uobj->object = entry;
3561         uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE);
3562
3563         err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3564                              &mmap_offset, sizeof(mmap_offset));
3565         if (err)
3566                 return err;
3567
3568         err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3569                              &entry->page_idx, sizeof(entry->page_idx));
3570         if (err)
3571                 return err;
3572
3573         err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3574                              &length, sizeof(length));
3575         return err;
3576 }
3577
3578 DECLARE_UVERBS_NAMED_METHOD(
3579         MLX5_IB_METHOD_UAR_OBJ_ALLOC,
3580         UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE,
3581                         MLX5_IB_OBJECT_UAR,
3582                         UVERBS_ACCESS_NEW,
3583                         UA_MANDATORY),
3584         UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE,
3585                              enum mlx5_ib_uapi_uar_alloc_type,
3586                              UA_MANDATORY),
3587         UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID,
3588                            UVERBS_ATTR_TYPE(u32),
3589                            UA_MANDATORY),
3590         UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH,
3591                            UVERBS_ATTR_TYPE(u32),
3592                            UA_MANDATORY),
3593         UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET,
3594                             UVERBS_ATTR_TYPE(u64),
3595                             UA_MANDATORY));
3596
3597 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
3598         MLX5_IB_METHOD_UAR_OBJ_DESTROY,
3599         UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE,
3600                         MLX5_IB_OBJECT_UAR,
3601                         UVERBS_ACCESS_DESTROY,
3602                         UA_MANDATORY));
3603
3604 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR,
3605                             UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup),
3606                             &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC),
3607                             &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY));
3608
3609 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3610         mlx5_ib_flow_action,
3611         UVERBS_OBJECT_FLOW_ACTION,
3612         UVERBS_METHOD_FLOW_ACTION_ESP_CREATE,
3613         UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_CREATE_FLOW_ACTION_FLAGS,
3614                              enum mlx5_ib_uapi_flow_action_flags));
3615
3616 ADD_UVERBS_ATTRIBUTES_SIMPLE(
3617         mlx5_ib_query_context,
3618         UVERBS_OBJECT_DEVICE,
3619         UVERBS_METHOD_QUERY_CONTEXT,
3620         UVERBS_ATTR_PTR_OUT(
3621                 MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX,
3622                 UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp,
3623                                    dump_fill_mkey),
3624                 UA_MANDATORY));
3625
3626 static const struct uapi_definition mlx5_ib_defs[] = {
3627         UAPI_DEF_CHAIN(mlx5_ib_devx_defs),
3628         UAPI_DEF_CHAIN(mlx5_ib_flow_defs),
3629         UAPI_DEF_CHAIN(mlx5_ib_qos_defs),
3630         UAPI_DEF_CHAIN(mlx5_ib_std_types_defs),
3631         UAPI_DEF_CHAIN(mlx5_ib_dm_defs),
3632
3633         UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
3634                                 &mlx5_ib_flow_action),
3635         UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context),
3636         UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR,
3637                                 UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)),
3638         UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR),
3639         {}
3640 };
3641
3642 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev)
3643 {
3644         mlx5_ib_cleanup_multiport_master(dev);
3645         WARN_ON(!xa_empty(&dev->odp_mkeys));
3646         mutex_destroy(&dev->cap_mask_mutex);
3647         WARN_ON(!xa_empty(&dev->sig_mrs));
3648         WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES));
3649 }
3650
3651 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
3652 {
3653         struct mlx5_core_dev *mdev = dev->mdev;
3654         int err;
3655         int i;
3656
3657         dev->ib_dev.node_type = RDMA_NODE_IB_CA;
3658         dev->ib_dev.local_dma_lkey = 0 /* not supported for now */;
3659         dev->ib_dev.phys_port_cnt = dev->num_ports;
3660         dev->ib_dev.dev.parent = mdev->device;
3661         dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
3662
3663         for (i = 0; i < dev->num_ports; i++) {
3664                 spin_lock_init(&dev->port[i].mp.mpi_lock);
3665                 rwlock_init(&dev->port[i].roce.netdev_lock);
3666                 dev->port[i].roce.dev = dev;
3667                 dev->port[i].roce.native_port_num = i + 1;
3668                 dev->port[i].roce.last_port_state = IB_PORT_DOWN;
3669         }
3670
3671         err = mlx5_ib_init_multiport_master(dev);
3672         if (err)
3673                 return err;
3674
3675         err = set_has_smi_cap(dev);
3676         if (err)
3677                 goto err_mp;
3678
3679         err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len);
3680         if (err)
3681                 goto err_mp;
3682
3683         if (mlx5_use_mad_ifc(dev))
3684                 get_ext_port_caps(dev);
3685
3686         dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
3687
3688         mutex_init(&dev->cap_mask_mutex);
3689         INIT_LIST_HEAD(&dev->qp_list);
3690         spin_lock_init(&dev->reset_flow_resource_lock);
3691         xa_init(&dev->odp_mkeys);
3692         xa_init(&dev->sig_mrs);
3693         atomic_set(&dev->mkey_var, 0);
3694
3695         spin_lock_init(&dev->dm.lock);
3696         dev->dm.dev = mdev;
3697         return 0;
3698
3699 err_mp:
3700         mlx5_ib_cleanup_multiport_master(dev);
3701         return err;
3702 }
3703
3704 static int mlx5_ib_enable_driver(struct ib_device *dev)
3705 {
3706         struct mlx5_ib_dev *mdev = to_mdev(dev);
3707         int ret;
3708
3709         ret = mlx5_ib_test_wc(mdev);
3710         mlx5_ib_dbg(mdev, "Write-Combining %s",
3711                     mdev->wc_support ? "supported" : "not supported");
3712
3713         return ret;
3714 }
3715
3716 static const struct ib_device_ops mlx5_ib_dev_ops = {
3717         .owner = THIS_MODULE,
3718         .driver_id = RDMA_DRIVER_MLX5,
3719         .uverbs_abi_ver = MLX5_IB_UVERBS_ABI_VERSION,
3720
3721         .add_gid = mlx5_ib_add_gid,
3722         .alloc_mr = mlx5_ib_alloc_mr,
3723         .alloc_mr_integrity = mlx5_ib_alloc_mr_integrity,
3724         .alloc_pd = mlx5_ib_alloc_pd,
3725         .alloc_ucontext = mlx5_ib_alloc_ucontext,
3726         .attach_mcast = mlx5_ib_mcg_attach,
3727         .check_mr_status = mlx5_ib_check_mr_status,
3728         .create_ah = mlx5_ib_create_ah,
3729         .create_cq = mlx5_ib_create_cq,
3730         .create_qp = mlx5_ib_create_qp,
3731         .create_srq = mlx5_ib_create_srq,
3732         .create_user_ah = mlx5_ib_create_ah,
3733         .dealloc_pd = mlx5_ib_dealloc_pd,
3734         .dealloc_ucontext = mlx5_ib_dealloc_ucontext,
3735         .del_gid = mlx5_ib_del_gid,
3736         .dereg_mr = mlx5_ib_dereg_mr,
3737         .destroy_ah = mlx5_ib_destroy_ah,
3738         .destroy_cq = mlx5_ib_destroy_cq,
3739         .destroy_qp = mlx5_ib_destroy_qp,
3740         .destroy_srq = mlx5_ib_destroy_srq,
3741         .detach_mcast = mlx5_ib_mcg_detach,
3742         .disassociate_ucontext = mlx5_ib_disassociate_ucontext,
3743         .drain_rq = mlx5_ib_drain_rq,
3744         .drain_sq = mlx5_ib_drain_sq,
3745         .device_group = &mlx5_attr_group,
3746         .enable_driver = mlx5_ib_enable_driver,
3747         .get_dev_fw_str = get_dev_fw_str,
3748         .get_dma_mr = mlx5_ib_get_dma_mr,
3749         .get_link_layer = mlx5_ib_port_link_layer,
3750         .map_mr_sg = mlx5_ib_map_mr_sg,
3751         .map_mr_sg_pi = mlx5_ib_map_mr_sg_pi,
3752         .mmap = mlx5_ib_mmap,
3753         .mmap_free = mlx5_ib_mmap_free,
3754         .modify_cq = mlx5_ib_modify_cq,
3755         .modify_device = mlx5_ib_modify_device,
3756         .modify_port = mlx5_ib_modify_port,
3757         .modify_qp = mlx5_ib_modify_qp,
3758         .modify_srq = mlx5_ib_modify_srq,
3759         .poll_cq = mlx5_ib_poll_cq,
3760         .post_recv = mlx5_ib_post_recv_nodrain,
3761         .post_send = mlx5_ib_post_send_nodrain,
3762         .post_srq_recv = mlx5_ib_post_srq_recv,
3763         .process_mad = mlx5_ib_process_mad,
3764         .query_ah = mlx5_ib_query_ah,
3765         .query_device = mlx5_ib_query_device,
3766         .query_gid = mlx5_ib_query_gid,
3767         .query_pkey = mlx5_ib_query_pkey,
3768         .query_qp = mlx5_ib_query_qp,
3769         .query_srq = mlx5_ib_query_srq,
3770         .query_ucontext = mlx5_ib_query_ucontext,
3771         .reg_user_mr = mlx5_ib_reg_user_mr,
3772         .reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
3773         .req_notify_cq = mlx5_ib_arm_cq,
3774         .rereg_user_mr = mlx5_ib_rereg_user_mr,
3775         .resize_cq = mlx5_ib_resize_cq,
3776
3777         INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah),
3778         INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs),
3779         INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq),
3780         INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd),
3781         INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq),
3782         INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext),
3783 };
3784
3785 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = {
3786         .rdma_netdev_get_params = mlx5_ib_rn_get_params,
3787 };
3788
3789 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = {
3790         .get_vf_config = mlx5_ib_get_vf_config,
3791         .get_vf_guid = mlx5_ib_get_vf_guid,
3792         .get_vf_stats = mlx5_ib_get_vf_stats,
3793         .set_vf_guid = mlx5_ib_set_vf_guid,
3794         .set_vf_link_state = mlx5_ib_set_vf_link_state,
3795 };
3796
3797 static const struct ib_device_ops mlx5_ib_dev_mw_ops = {
3798         .alloc_mw = mlx5_ib_alloc_mw,
3799         .dealloc_mw = mlx5_ib_dealloc_mw,
3800
3801         INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw),
3802 };
3803
3804 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = {
3805         .alloc_xrcd = mlx5_ib_alloc_xrcd,
3806         .dealloc_xrcd = mlx5_ib_dealloc_xrcd,
3807
3808         INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd),
3809 };
3810
3811 static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev)
3812 {
3813         struct mlx5_core_dev *mdev = dev->mdev;
3814         struct mlx5_var_table *var_table = &dev->var_table;
3815         u8 log_doorbell_bar_size;
3816         u8 log_doorbell_stride;
3817         u64 bar_size;
3818
3819         log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3820                                         log_doorbell_bar_size);
3821         log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev,
3822                                         log_doorbell_stride);
3823         var_table->hw_start_addr = dev->mdev->bar_addr +
3824                                 MLX5_CAP64_DEV_VDPA_EMULATION(mdev,
3825                                         doorbell_bar_offset);
3826         bar_size = (1ULL << log_doorbell_bar_size) * 4096;
3827         var_table->stride_size = 1ULL << log_doorbell_stride;
3828         var_table->num_var_hw_entries = div_u64(bar_size,
3829                                                 var_table->stride_size);
3830         mutex_init(&var_table->bitmap_lock);
3831         var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries,
3832                                           GFP_KERNEL);
3833         return (var_table->bitmap) ? 0 : -ENOMEM;
3834 }
3835
3836 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev)
3837 {
3838         bitmap_free(dev->var_table.bitmap);
3839 }
3840
3841 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev)
3842 {
3843         struct mlx5_core_dev *mdev = dev->mdev;
3844         int err;
3845
3846         if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) &&
3847             IS_ENABLED(CONFIG_MLX5_CORE_IPOIB))
3848                 ib_set_device_ops(&dev->ib_dev,
3849                                   &mlx5_ib_dev_ipoib_enhanced_ops);
3850
3851         if (mlx5_core_is_pf(mdev))
3852                 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops);
3853
3854         dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence));
3855
3856         if (MLX5_CAP_GEN(mdev, imaicl))
3857                 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops);
3858
3859         if (MLX5_CAP_GEN(mdev, xrc))
3860                 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops);
3861
3862         if (MLX5_CAP_DEV_MEM(mdev, memic) ||
3863             MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3864             MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM)
3865                 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops);
3866
3867         ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops);
3868
3869         if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS))
3870                 dev->ib_dev.driver_def = mlx5_ib_defs;
3871
3872         err = init_node_data(dev);
3873         if (err)
3874                 return err;
3875
3876         if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) &&
3877             (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) ||
3878              MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc)))
3879                 mutex_init(&dev->lb.mutex);
3880
3881         if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) &
3882                         MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) {
3883                 err = mlx5_ib_init_var_table(dev);
3884                 if (err)
3885                         return err;
3886         }
3887
3888         dev->ib_dev.use_cq_dim = true;
3889
3890         return 0;
3891 }
3892
3893 static const struct ib_device_ops mlx5_ib_dev_port_ops = {
3894         .get_port_immutable = mlx5_port_immutable,
3895         .query_port = mlx5_ib_query_port,
3896 };
3897
3898 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev)
3899 {
3900         ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops);
3901         return 0;
3902 }
3903
3904 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = {
3905         .get_port_immutable = mlx5_port_rep_immutable,
3906         .query_port = mlx5_ib_rep_query_port,
3907         .query_pkey = mlx5_ib_rep_query_pkey,
3908 };
3909
3910 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev)
3911 {
3912         ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops);
3913         return 0;
3914 }
3915
3916 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
3917         .create_rwq_ind_table = mlx5_ib_create_rwq_ind_table,
3918         .create_wq = mlx5_ib_create_wq,
3919         .destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table,
3920         .destroy_wq = mlx5_ib_destroy_wq,
3921         .get_netdev = mlx5_ib_get_netdev,
3922         .modify_wq = mlx5_ib_modify_wq,
3923
3924         INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table,
3925                            ib_rwq_ind_tbl),
3926 };
3927
3928 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev)
3929 {
3930         struct mlx5_core_dev *mdev = dev->mdev;
3931         enum rdma_link_layer ll;
3932         int port_type_cap;
3933         u32 port_num = 0;
3934         int err;
3935
3936         port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3937         ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3938
3939         if (ll == IB_LINK_LAYER_ETHERNET) {
3940                 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops);
3941
3942                 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3943
3944                 /* Register only for native ports */
3945                 err = mlx5_add_netdev_notifier(dev, port_num);
3946                 if (err || dev->is_rep || !mlx5_is_roce_init_enabled(mdev))
3947                         /*
3948                          * We don't enable ETH interface for
3949                          * 1. IB representors
3950                          * 2. User disabled ROCE through devlink interface
3951                          */
3952                         return err;
3953
3954                 err = mlx5_enable_eth(dev);
3955                 if (err)
3956                         goto cleanup;
3957         }
3958
3959         return 0;
3960 cleanup:
3961         mlx5_remove_netdev_notifier(dev, port_num);
3962         return err;
3963 }
3964
3965 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev)
3966 {
3967         struct mlx5_core_dev *mdev = dev->mdev;
3968         enum rdma_link_layer ll;
3969         int port_type_cap;
3970         u32 port_num;
3971
3972         port_type_cap = MLX5_CAP_GEN(mdev, port_type);
3973         ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
3974
3975         if (ll == IB_LINK_LAYER_ETHERNET) {
3976                 if (!dev->is_rep)
3977                         mlx5_disable_eth(dev);
3978
3979                 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3980                 mlx5_remove_netdev_notifier(dev, port_num);
3981         }
3982 }
3983
3984 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev)
3985 {
3986         mlx5_ib_init_cong_debugfs(dev,
3987                                   mlx5_core_native_port_num(dev->mdev) - 1);
3988         return 0;
3989 }
3990
3991 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev)
3992 {
3993         mlx5_ib_cleanup_cong_debugfs(dev,
3994                                      mlx5_core_native_port_num(dev->mdev) - 1);
3995 }
3996
3997 static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev)
3998 {
3999         dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev);
4000         return PTR_ERR_OR_ZERO(dev->mdev->priv.uar);
4001 }
4002
4003 static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev)
4004 {
4005         mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar);
4006 }
4007
4008 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev)
4009 {
4010         int err;
4011
4012         err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false);
4013         if (err)
4014                 return err;
4015
4016         err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true);
4017         if (err)
4018                 mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4019
4020         return err;
4021 }
4022
4023 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev)
4024 {
4025         mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg);
4026         mlx5_free_bfreg(dev->mdev, &dev->bfreg);
4027 }
4028
4029 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev)
4030 {
4031         const char *name;
4032
4033         if (!mlx5_lag_is_roce(dev->mdev))
4034                 name = "mlx5_%d";
4035         else
4036                 name = "mlx5_bond_%d";
4037         return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev);
4038 }
4039
4040 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
4041 {
4042         int err;
4043
4044         err = mlx5_mr_cache_cleanup(dev);
4045         if (err)
4046                 mlx5_ib_warn(dev, "mr cache cleanup failed\n");
4047
4048         if (dev->umrc.qp)
4049                 mlx5_ib_destroy_qp(dev->umrc.qp, NULL);
4050         if (dev->umrc.cq)
4051                 ib_free_cq(dev->umrc.cq);
4052         if (dev->umrc.pd)
4053                 ib_dealloc_pd(dev->umrc.pd);
4054 }
4055
4056 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev)
4057 {
4058         ib_unregister_device(&dev->ib_dev);
4059 }
4060
4061 enum {
4062         MAX_UMR_WR = 128,
4063 };
4064
4065 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
4066 {
4067         struct ib_qp_init_attr *init_attr = NULL;
4068         struct ib_qp_attr *attr = NULL;
4069         struct ib_pd *pd;
4070         struct ib_cq *cq;
4071         struct ib_qp *qp;
4072         int ret;
4073
4074         attr = kzalloc(sizeof(*attr), GFP_KERNEL);
4075         init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL);
4076         if (!attr || !init_attr) {
4077                 ret = -ENOMEM;
4078                 goto error_0;
4079         }
4080
4081         pd = ib_alloc_pd(&dev->ib_dev, 0);
4082         if (IS_ERR(pd)) {
4083                 mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n");
4084                 ret = PTR_ERR(pd);
4085                 goto error_0;
4086         }
4087
4088         cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ);
4089         if (IS_ERR(cq)) {
4090                 mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n");
4091                 ret = PTR_ERR(cq);
4092                 goto error_2;
4093         }
4094
4095         init_attr->send_cq = cq;
4096         init_attr->recv_cq = cq;
4097         init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
4098         init_attr->cap.max_send_wr = MAX_UMR_WR;
4099         init_attr->cap.max_send_sge = 1;
4100         init_attr->qp_type = MLX5_IB_QPT_REG_UMR;
4101         init_attr->port_num = 1;
4102         qp = mlx5_ib_create_qp(pd, init_attr, NULL);
4103         if (IS_ERR(qp)) {
4104                 mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n");
4105                 ret = PTR_ERR(qp);
4106                 goto error_3;
4107         }
4108         qp->device     = &dev->ib_dev;
4109         qp->real_qp    = qp;
4110         qp->uobject    = NULL;
4111         qp->qp_type    = MLX5_IB_QPT_REG_UMR;
4112         qp->send_cq    = init_attr->send_cq;
4113         qp->recv_cq    = init_attr->recv_cq;
4114
4115         attr->qp_state = IB_QPS_INIT;
4116         attr->port_num = 1;
4117         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX |
4118                                 IB_QP_PORT, NULL);
4119         if (ret) {
4120                 mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
4121                 goto error_4;
4122         }
4123
4124         memset(attr, 0, sizeof(*attr));
4125         attr->qp_state = IB_QPS_RTR;
4126         attr->path_mtu = IB_MTU_256;
4127
4128         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4129         if (ret) {
4130                 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n");
4131                 goto error_4;
4132         }
4133
4134         memset(attr, 0, sizeof(*attr));
4135         attr->qp_state = IB_QPS_RTS;
4136         ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL);
4137         if (ret) {
4138                 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n");
4139                 goto error_4;
4140         }
4141
4142         dev->umrc.qp = qp;
4143         dev->umrc.cq = cq;
4144         dev->umrc.pd = pd;
4145
4146         sema_init(&dev->umrc.sem, MAX_UMR_WR);
4147         ret = mlx5_mr_cache_init(dev);
4148         if (ret) {
4149                 mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
4150                 goto error_4;
4151         }
4152
4153         kfree(attr);
4154         kfree(init_attr);
4155
4156         return 0;
4157
4158 error_4:
4159         mlx5_ib_destroy_qp(qp, NULL);
4160         dev->umrc.qp = NULL;
4161
4162 error_3:
4163         ib_free_cq(cq);
4164         dev->umrc.cq = NULL;
4165
4166 error_2:
4167         ib_dealloc_pd(pd);
4168         dev->umrc.pd = NULL;
4169
4170 error_0:
4171         kfree(attr);
4172         kfree(init_attr);
4173         return ret;
4174 }
4175
4176 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev)
4177 {
4178         struct dentry *root;
4179
4180         if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4181                 return 0;
4182
4183         mutex_init(&dev->delay_drop.lock);
4184         dev->delay_drop.dev = dev;
4185         dev->delay_drop.activate = false;
4186         dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000;
4187         INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler);
4188         atomic_set(&dev->delay_drop.rqs_cnt, 0);
4189         atomic_set(&dev->delay_drop.events_cnt, 0);
4190
4191         if (!mlx5_debugfs_root)
4192                 return 0;
4193
4194         root = debugfs_create_dir("delay_drop", dev->mdev->priv.dbg_root);
4195         dev->delay_drop.dir_debugfs = root;
4196
4197         debugfs_create_atomic_t("num_timeout_events", 0400, root,
4198                                 &dev->delay_drop.events_cnt);
4199         debugfs_create_atomic_t("num_rqs", 0400, root,
4200                                 &dev->delay_drop.rqs_cnt);
4201         debugfs_create_file("timeout", 0600, root, &dev->delay_drop,
4202                             &fops_delay_drop_timeout);
4203         return 0;
4204 }
4205
4206 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev)
4207 {
4208         if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP))
4209                 return;
4210
4211         cancel_work_sync(&dev->delay_drop.delay_drop_work);
4212         if (!dev->delay_drop.dir_debugfs)
4213                 return;
4214
4215         debugfs_remove_recursive(dev->delay_drop.dir_debugfs);
4216         dev->delay_drop.dir_debugfs = NULL;
4217 }
4218
4219 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev)
4220 {
4221         dev->mdev_events.notifier_call = mlx5_ib_event;
4222         mlx5_notifier_register(dev->mdev, &dev->mdev_events);
4223         return 0;
4224 }
4225
4226 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev)
4227 {
4228         mlx5_notifier_unregister(dev->mdev, &dev->mdev_events);
4229 }
4230
4231 void __mlx5_ib_remove(struct mlx5_ib_dev *dev,
4232                       const struct mlx5_ib_profile *profile,
4233                       int stage)
4234 {
4235         dev->ib_active = false;
4236
4237         /* Number of stages to cleanup */
4238         while (stage) {
4239                 stage--;
4240                 if (profile->stage[stage].cleanup)
4241                         profile->stage[stage].cleanup(dev);
4242         }
4243
4244         kfree(dev->port);
4245         ib_dealloc_device(&dev->ib_dev);
4246 }
4247
4248 int __mlx5_ib_add(struct mlx5_ib_dev *dev,
4249                   const struct mlx5_ib_profile *profile)
4250 {
4251         int err;
4252         int i;
4253
4254         dev->profile = profile;
4255
4256         for (i = 0; i < MLX5_IB_STAGE_MAX; i++) {
4257                 if (profile->stage[i].init) {
4258                         err = profile->stage[i].init(dev);
4259                         if (err)
4260                                 goto err_out;
4261                 }
4262         }
4263
4264         dev->ib_active = true;
4265         return 0;
4266
4267 err_out:
4268         /* Clean up stages which were initialized */
4269         while (i) {
4270                 i--;
4271                 if (profile->stage[i].cleanup)
4272                         profile->stage[i].cleanup(dev);
4273         }
4274         return -ENOMEM;
4275 }
4276
4277 static const struct mlx5_ib_profile pf_profile = {
4278         STAGE_CREATE(MLX5_IB_STAGE_INIT,
4279                      mlx5_ib_stage_init_init,
4280                      mlx5_ib_stage_init_cleanup),
4281         STAGE_CREATE(MLX5_IB_STAGE_FS,
4282                      mlx5_ib_fs_init,
4283                      mlx5_ib_fs_cleanup),
4284         STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4285                      mlx5_ib_stage_caps_init,
4286                      mlx5_ib_stage_caps_cleanup),
4287         STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4288                      mlx5_ib_stage_non_default_cb,
4289                      NULL),
4290         STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4291                      mlx5_ib_roce_init,
4292                      mlx5_ib_roce_cleanup),
4293         STAGE_CREATE(MLX5_IB_STAGE_QP,
4294                      mlx5_init_qp_table,
4295                      mlx5_cleanup_qp_table),
4296         STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4297                      mlx5_init_srq_table,
4298                      mlx5_cleanup_srq_table),
4299         STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4300                      mlx5_ib_dev_res_init,
4301                      mlx5_ib_dev_res_cleanup),
4302         STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4303                      mlx5_ib_stage_dev_notifier_init,
4304                      mlx5_ib_stage_dev_notifier_cleanup),
4305         STAGE_CREATE(MLX5_IB_STAGE_ODP,
4306                      mlx5_ib_odp_init_one,
4307                      mlx5_ib_odp_cleanup_one),
4308         STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4309                      mlx5_ib_counters_init,
4310                      mlx5_ib_counters_cleanup),
4311         STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4312                      mlx5_ib_stage_cong_debugfs_init,
4313                      mlx5_ib_stage_cong_debugfs_cleanup),
4314         STAGE_CREATE(MLX5_IB_STAGE_UAR,
4315                      mlx5_ib_stage_uar_init,
4316                      mlx5_ib_stage_uar_cleanup),
4317         STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4318                      mlx5_ib_stage_bfrag_init,
4319                      mlx5_ib_stage_bfrag_cleanup),
4320         STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4321                      NULL,
4322                      mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4323         STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4324                      mlx5_ib_devx_init,
4325                      mlx5_ib_devx_cleanup),
4326         STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4327                      mlx5_ib_stage_ib_reg_init,
4328                      mlx5_ib_stage_ib_reg_cleanup),
4329         STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4330                      mlx5_ib_stage_post_ib_reg_umr_init,
4331                      NULL),
4332         STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP,
4333                      mlx5_ib_stage_delay_drop_init,
4334                      mlx5_ib_stage_delay_drop_cleanup),
4335         STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4336                      mlx5_ib_restrack_init,
4337                      NULL),
4338 };
4339
4340 const struct mlx5_ib_profile raw_eth_profile = {
4341         STAGE_CREATE(MLX5_IB_STAGE_INIT,
4342                      mlx5_ib_stage_init_init,
4343                      mlx5_ib_stage_init_cleanup),
4344         STAGE_CREATE(MLX5_IB_STAGE_FS,
4345                      mlx5_ib_fs_init,
4346                      mlx5_ib_fs_cleanup),
4347         STAGE_CREATE(MLX5_IB_STAGE_CAPS,
4348                      mlx5_ib_stage_caps_init,
4349                      mlx5_ib_stage_caps_cleanup),
4350         STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB,
4351                      mlx5_ib_stage_raw_eth_non_default_cb,
4352                      NULL),
4353         STAGE_CREATE(MLX5_IB_STAGE_ROCE,
4354                      mlx5_ib_roce_init,
4355                      mlx5_ib_roce_cleanup),
4356         STAGE_CREATE(MLX5_IB_STAGE_QP,
4357                      mlx5_init_qp_table,
4358                      mlx5_cleanup_qp_table),
4359         STAGE_CREATE(MLX5_IB_STAGE_SRQ,
4360                      mlx5_init_srq_table,
4361                      mlx5_cleanup_srq_table),
4362         STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES,
4363                      mlx5_ib_dev_res_init,
4364                      mlx5_ib_dev_res_cleanup),
4365         STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER,
4366                      mlx5_ib_stage_dev_notifier_init,
4367                      mlx5_ib_stage_dev_notifier_cleanup),
4368         STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
4369                      mlx5_ib_counters_init,
4370                      mlx5_ib_counters_cleanup),
4371         STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS,
4372                      mlx5_ib_stage_cong_debugfs_init,
4373                      mlx5_ib_stage_cong_debugfs_cleanup),
4374         STAGE_CREATE(MLX5_IB_STAGE_UAR,
4375                      mlx5_ib_stage_uar_init,
4376                      mlx5_ib_stage_uar_cleanup),
4377         STAGE_CREATE(MLX5_IB_STAGE_BFREG,
4378                      mlx5_ib_stage_bfrag_init,
4379                      mlx5_ib_stage_bfrag_cleanup),
4380         STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR,
4381                      NULL,
4382                      mlx5_ib_stage_pre_ib_reg_umr_cleanup),
4383         STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID,
4384                      mlx5_ib_devx_init,
4385                      mlx5_ib_devx_cleanup),
4386         STAGE_CREATE(MLX5_IB_STAGE_IB_REG,
4387                      mlx5_ib_stage_ib_reg_init,
4388                      mlx5_ib_stage_ib_reg_cleanup),
4389         STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR,
4390                      mlx5_ib_stage_post_ib_reg_umr_init,
4391                      NULL),
4392         STAGE_CREATE(MLX5_IB_STAGE_RESTRACK,
4393                      mlx5_ib_restrack_init,
4394                      NULL),
4395 };
4396
4397 static int mlx5r_mp_probe(struct auxiliary_device *adev,
4398                           const struct auxiliary_device_id *id)
4399 {
4400         struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4401         struct mlx5_core_dev *mdev = idev->mdev;
4402         struct mlx5_ib_multiport_info *mpi;
4403         struct mlx5_ib_dev *dev;
4404         bool bound = false;
4405         int err;
4406
4407         mpi = kzalloc(sizeof(*mpi), GFP_KERNEL);
4408         if (!mpi)
4409                 return -ENOMEM;
4410
4411         mpi->mdev = mdev;
4412         err = mlx5_query_nic_vport_system_image_guid(mdev,
4413                                                      &mpi->sys_image_guid);
4414         if (err) {
4415                 kfree(mpi);
4416                 return err;
4417         }
4418
4419         mutex_lock(&mlx5_ib_multiport_mutex);
4420         list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
4421                 if (dev->sys_image_guid == mpi->sys_image_guid)
4422                         bound = mlx5_ib_bind_slave_port(dev, mpi);
4423
4424                 if (bound) {
4425                         rdma_roce_rescan_device(&dev->ib_dev);
4426                         mpi->ibdev->ib_active = true;
4427                         break;
4428                 }
4429         }
4430
4431         if (!bound) {
4432                 list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list);
4433                 dev_dbg(mdev->device,
4434                         "no suitable IB device found to bind to, added to unaffiliated list.\n");
4435         }
4436         mutex_unlock(&mlx5_ib_multiport_mutex);
4437
4438         dev_set_drvdata(&adev->dev, mpi);
4439         return 0;
4440 }
4441
4442 static void mlx5r_mp_remove(struct auxiliary_device *adev)
4443 {
4444         struct mlx5_ib_multiport_info *mpi;
4445
4446         mpi = dev_get_drvdata(&adev->dev);
4447         mutex_lock(&mlx5_ib_multiport_mutex);
4448         if (mpi->ibdev)
4449                 mlx5_ib_unbind_slave_port(mpi->ibdev, mpi);
4450         list_del(&mpi->list);
4451         mutex_unlock(&mlx5_ib_multiport_mutex);
4452         kfree(mpi);
4453 }
4454
4455 static int mlx5r_probe(struct auxiliary_device *adev,
4456                        const struct auxiliary_device_id *id)
4457 {
4458         struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
4459         struct mlx5_core_dev *mdev = idev->mdev;
4460         const struct mlx5_ib_profile *profile;
4461         int port_type_cap, num_ports, ret;
4462         enum rdma_link_layer ll;
4463         struct mlx5_ib_dev *dev;
4464
4465         port_type_cap = MLX5_CAP_GEN(mdev, port_type);
4466         ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
4467
4468         num_ports = max(MLX5_CAP_GEN(mdev, num_ports),
4469                         MLX5_CAP_GEN(mdev, num_vhca_ports));
4470         dev = ib_alloc_device(mlx5_ib_dev, ib_dev);
4471         if (!dev)
4472                 return -ENOMEM;
4473         dev->port = kcalloc(num_ports, sizeof(*dev->port),
4474                              GFP_KERNEL);
4475         if (!dev->port) {
4476                 ib_dealloc_device(&dev->ib_dev);
4477                 return -ENOMEM;
4478         }
4479
4480         dev->mdev = mdev;
4481         dev->num_ports = num_ports;
4482
4483         if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_is_roce_init_enabled(mdev))
4484                 profile = &raw_eth_profile;
4485         else
4486                 profile = &pf_profile;
4487
4488         ret = __mlx5_ib_add(dev, profile);
4489         if (ret) {
4490                 kfree(dev->port);
4491                 ib_dealloc_device(&dev->ib_dev);
4492                 return ret;
4493         }
4494
4495         dev_set_drvdata(&adev->dev, dev);
4496         return 0;
4497 }
4498
4499 static void mlx5r_remove(struct auxiliary_device *adev)
4500 {
4501         struct mlx5_ib_dev *dev;
4502
4503         dev = dev_get_drvdata(&adev->dev);
4504         __mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
4505 }
4506
4507 static const struct auxiliary_device_id mlx5r_mp_id_table[] = {
4508         { .name = MLX5_ADEV_NAME ".multiport", },
4509         {},
4510 };
4511
4512 static const struct auxiliary_device_id mlx5r_id_table[] = {
4513         { .name = MLX5_ADEV_NAME ".rdma", },
4514         {},
4515 };
4516
4517 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table);
4518 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table);
4519
4520 static struct auxiliary_driver mlx5r_mp_driver = {
4521         .name = "multiport",
4522         .probe = mlx5r_mp_probe,
4523         .remove = mlx5r_mp_remove,
4524         .id_table = mlx5r_mp_id_table,
4525 };
4526
4527 static struct auxiliary_driver mlx5r_driver = {
4528         .name = "rdma",
4529         .probe = mlx5r_probe,
4530         .remove = mlx5r_remove,
4531         .id_table = mlx5r_id_table,
4532 };
4533
4534 static int __init mlx5_ib_init(void)
4535 {
4536         int ret;
4537
4538         xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL);
4539         if (!xlt_emergency_page)
4540                 return -ENOMEM;
4541
4542         mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0);
4543         if (!mlx5_ib_event_wq) {
4544                 free_page((unsigned long)xlt_emergency_page);
4545                 return -ENOMEM;
4546         }
4547
4548         mlx5_ib_odp_init();
4549         ret = mlx5r_rep_init();
4550         if (ret)
4551                 goto rep_err;
4552         ret = auxiliary_driver_register(&mlx5r_mp_driver);
4553         if (ret)
4554                 goto mp_err;
4555         ret = auxiliary_driver_register(&mlx5r_driver);
4556         if (ret)
4557                 goto drv_err;
4558         return 0;
4559
4560 drv_err:
4561         auxiliary_driver_unregister(&mlx5r_mp_driver);
4562 mp_err:
4563         mlx5r_rep_cleanup();
4564 rep_err:
4565         destroy_workqueue(mlx5_ib_event_wq);
4566         free_page((unsigned long)xlt_emergency_page);
4567         return ret;
4568 }
4569
4570 static void __exit mlx5_ib_cleanup(void)
4571 {
4572         auxiliary_driver_unregister(&mlx5r_driver);
4573         auxiliary_driver_unregister(&mlx5r_mp_driver);
4574         mlx5r_rep_cleanup();
4575
4576         destroy_workqueue(mlx5_ib_event_wq);
4577         free_page((unsigned long)xlt_emergency_page);
4578 }
4579
4580 module_init(mlx5_ib_init);
4581 module_exit(mlx5_ib_cleanup);