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/mlx5/driver.h> 28#include <linux/list.h> 29#include <rdma/ib_smi.h> 30#include <rdma/ib_umem_odp.h> 31#include <rdma/lag.h> 32#include <linux/in.h> 33#include <linux/etherdevice.h> 34#include "mlx5_ib.h" 35#include "ib_rep.h" 36#include "cmd.h" 37#include "devx.h" 38#include "dm.h" 39#include "fs.h" 40#include "srq.h" 41#include "qp.h" 42#include "wr.h" 43#include "restrack.h" 44#include "counters.h" 45#include "umr.h" 46#include <rdma/uverbs_std_types.h> 47#include <rdma/uverbs_ioctl.h> 48#include <rdma/mlx5_user_ioctl_verbs.h> 49#include <rdma/mlx5_user_ioctl_cmds.h> 50#include "macsec.h" 51 52#define UVERBS_MODULE_NAME mlx5_ib 53#include <rdma/uverbs_named_ioctl.h> 54 55MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>"); 56MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver"); 57MODULE_LICENSE("Dual BSD/GPL"); 58 59struct mlx5_ib_event_work { 60 struct work_struct work; 61 union { 62 struct mlx5_ib_dev *dev; 63 struct mlx5_ib_multiport_info *mpi; 64 }; 65 bool is_slave; 66 unsigned int event; 67 void *param; 68}; 69 70enum { 71 MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3, 72}; 73 74static struct workqueue_struct *mlx5_ib_event_wq; 75static LIST_HEAD(mlx5_ib_unaffiliated_port_list); 76static LIST_HEAD(mlx5_ib_dev_list); 77/* 78 * This mutex should be held when accessing either of the above lists 79 */ 80static DEFINE_MUTEX(mlx5_ib_multiport_mutex); 81 82struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi) 83{ 84 struct mlx5_ib_dev *dev; 85 86 mutex_lock(&mlx5_ib_multiport_mutex); 87 dev = mpi->ibdev; 88 mutex_unlock(&mlx5_ib_multiport_mutex); 89 return dev; 90} 91 92static enum rdma_link_layer 93mlx5_port_type_cap_to_rdma_ll(int port_type_cap) 94{ 95 switch (port_type_cap) { 96 case MLX5_CAP_PORT_TYPE_IB: 97 return IB_LINK_LAYER_INFINIBAND; 98 case MLX5_CAP_PORT_TYPE_ETH: 99 return IB_LINK_LAYER_ETHERNET; 100 default: 101 return IB_LINK_LAYER_UNSPECIFIED; 102 } 103} 104 105static enum rdma_link_layer 106mlx5_ib_port_link_layer(struct ib_device *device, u32 port_num) 107{ 108 struct mlx5_ib_dev *dev = to_mdev(device); 109 int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type); 110 111 return mlx5_port_type_cap_to_rdma_ll(port_type_cap); 112} 113 114static int get_port_state(struct ib_device *ibdev, 115 u32 port_num, 116 enum ib_port_state *state) 117{ 118 struct ib_port_attr attr; 119 int ret; 120 121 memset(&attr, 0, sizeof(attr)); 122 ret = ibdev->ops.query_port(ibdev, port_num, &attr); 123 if (!ret) 124 *state = attr.state; 125 return ret; 126} 127 128static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev, 129 struct net_device *ndev, 130 struct net_device *upper, 131 u32 *port_num) 132{ 133 struct net_device *rep_ndev; 134 struct mlx5_ib_port *port; 135 int i; 136 137 for (i = 0; i < dev->num_ports; i++) { 138 port = &dev->port[i]; 139 if (!port->rep) 140 continue; 141 142 if (upper == ndev && port->rep->vport == MLX5_VPORT_UPLINK) { 143 *port_num = i + 1; 144 return &port->roce; 145 } 146 147 if (upper && port->rep->vport == MLX5_VPORT_UPLINK) 148 continue; 149 150 read_lock(&port->roce.netdev_lock); 151 rep_ndev = mlx5_ib_get_rep_netdev(port->rep->esw, 152 port->rep->vport); 153 if (rep_ndev == ndev) { 154 read_unlock(&port->roce.netdev_lock); 155 *port_num = i + 1; 156 return &port->roce; 157 } 158 read_unlock(&port->roce.netdev_lock); 159 } 160 161 return NULL; 162} 163 164static int mlx5_netdev_event(struct notifier_block *this, 165 unsigned long event, void *ptr) 166{ 167 struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb); 168 struct net_device *ndev = netdev_notifier_info_to_dev(ptr); 169 u32 port_num = roce->native_port_num; 170 struct mlx5_core_dev *mdev; 171 struct mlx5_ib_dev *ibdev; 172 173 ibdev = roce->dev; 174 mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL); 175 if (!mdev) 176 return NOTIFY_DONE; 177 178 switch (event) { 179 case NETDEV_REGISTER: 180 /* Should already be registered during the load */ 181 if (ibdev->is_rep) 182 break; 183 write_lock(&roce->netdev_lock); 184 if (ndev->dev.parent == mdev->device) 185 roce->netdev = ndev; 186 write_unlock(&roce->netdev_lock); 187 break; 188 189 case NETDEV_UNREGISTER: 190 /* In case of reps, ib device goes away before the netdevs */ 191 write_lock(&roce->netdev_lock); 192 if (roce->netdev == ndev) 193 roce->netdev = NULL; 194 write_unlock(&roce->netdev_lock); 195 break; 196 197 case NETDEV_CHANGE: 198 case NETDEV_UP: 199 case NETDEV_DOWN: { 200 struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev); 201 struct net_device *upper = NULL; 202 203 if (lag_ndev) { 204 upper = netdev_master_upper_dev_get(lag_ndev); 205 dev_put(lag_ndev); 206 } 207 208 if (ibdev->is_rep) 209 roce = mlx5_get_rep_roce(ibdev, ndev, upper, &port_num); 210 if (!roce) 211 return NOTIFY_DONE; 212 if ((upper == ndev || 213 ((!upper || ibdev->is_rep) && ndev == roce->netdev)) && 214 ibdev->ib_active) { 215 struct ib_event ibev = { }; 216 enum ib_port_state port_state; 217 218 if (get_port_state(&ibdev->ib_dev, port_num, 219 &port_state)) 220 goto done; 221 222 if (roce->last_port_state == port_state) 223 goto done; 224 225 roce->last_port_state = port_state; 226 ibev.device = &ibdev->ib_dev; 227 if (port_state == IB_PORT_DOWN) 228 ibev.event = IB_EVENT_PORT_ERR; 229 else if (port_state == IB_PORT_ACTIVE) 230 ibev.event = IB_EVENT_PORT_ACTIVE; 231 else 232 goto done; 233 234 ibev.element.port_num = port_num; 235 ib_dispatch_event(&ibev); 236 } 237 break; 238 } 239 240 default: 241 break; 242 } 243done: 244 mlx5_ib_put_native_port_mdev(ibdev, port_num); 245 return NOTIFY_DONE; 246} 247 248static struct net_device *mlx5_ib_get_netdev(struct ib_device *device, 249 u32 port_num) 250{ 251 struct mlx5_ib_dev *ibdev = to_mdev(device); 252 struct net_device *ndev; 253 struct mlx5_core_dev *mdev; 254 255 mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL); 256 if (!mdev) 257 return NULL; 258 259 ndev = mlx5_lag_get_roce_netdev(mdev); 260 if (ndev) 261 goto out; 262 263 /* Ensure ndev does not disappear before we invoke dev_hold() 264 */ 265 read_lock(&ibdev->port[port_num - 1].roce.netdev_lock); 266 ndev = ibdev->port[port_num - 1].roce.netdev; 267 if (ndev) 268 dev_hold(ndev); 269 read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock); 270 271out: 272 mlx5_ib_put_native_port_mdev(ibdev, port_num); 273 return ndev; 274} 275 276struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev, 277 u32 ib_port_num, 278 u32 *native_port_num) 279{ 280 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev, 281 ib_port_num); 282 struct mlx5_core_dev *mdev = NULL; 283 struct mlx5_ib_multiport_info *mpi; 284 struct mlx5_ib_port *port; 285 286 if (!mlx5_core_mp_enabled(ibdev->mdev) || 287 ll != IB_LINK_LAYER_ETHERNET) { 288 if (native_port_num) 289 *native_port_num = ib_port_num; 290 return ibdev->mdev; 291 } 292 293 if (native_port_num) 294 *native_port_num = 1; 295 296 port = &ibdev->port[ib_port_num - 1]; 297 spin_lock(&port->mp.mpi_lock); 298 mpi = ibdev->port[ib_port_num - 1].mp.mpi; 299 if (mpi && !mpi->unaffiliate) { 300 mdev = mpi->mdev; 301 /* If it's the master no need to refcount, it'll exist 302 * as long as the ib_dev exists. 303 */ 304 if (!mpi->is_master) 305 mpi->mdev_refcnt++; 306 } 307 spin_unlock(&port->mp.mpi_lock); 308 309 return mdev; 310} 311 312void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 port_num) 313{ 314 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev, 315 port_num); 316 struct mlx5_ib_multiport_info *mpi; 317 struct mlx5_ib_port *port; 318 319 if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET) 320 return; 321 322 port = &ibdev->port[port_num - 1]; 323 324 spin_lock(&port->mp.mpi_lock); 325 mpi = ibdev->port[port_num - 1].mp.mpi; 326 if (mpi->is_master) 327 goto out; 328 329 mpi->mdev_refcnt--; 330 if (mpi->unaffiliate) 331 complete(&mpi->unref_comp); 332out: 333 spin_unlock(&port->mp.mpi_lock); 334} 335 336static int translate_eth_legacy_proto_oper(u32 eth_proto_oper, 337 u16 *active_speed, u8 *active_width) 338{ 339 switch (eth_proto_oper) { 340 case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII): 341 case MLX5E_PROT_MASK(MLX5E_1000BASE_KX): 342 case MLX5E_PROT_MASK(MLX5E_100BASE_TX): 343 case MLX5E_PROT_MASK(MLX5E_1000BASE_T): 344 *active_width = IB_WIDTH_1X; 345 *active_speed = IB_SPEED_SDR; 346 break; 347 case MLX5E_PROT_MASK(MLX5E_10GBASE_T): 348 case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4): 349 case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4): 350 case MLX5E_PROT_MASK(MLX5E_10GBASE_KR): 351 case MLX5E_PROT_MASK(MLX5E_10GBASE_CR): 352 case MLX5E_PROT_MASK(MLX5E_10GBASE_SR): 353 case MLX5E_PROT_MASK(MLX5E_10GBASE_ER): 354 *active_width = IB_WIDTH_1X; 355 *active_speed = IB_SPEED_QDR; 356 break; 357 case MLX5E_PROT_MASK(MLX5E_25GBASE_CR): 358 case MLX5E_PROT_MASK(MLX5E_25GBASE_KR): 359 case MLX5E_PROT_MASK(MLX5E_25GBASE_SR): 360 *active_width = IB_WIDTH_1X; 361 *active_speed = IB_SPEED_EDR; 362 break; 363 case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4): 364 case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4): 365 case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4): 366 case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4): 367 *active_width = IB_WIDTH_4X; 368 *active_speed = IB_SPEED_QDR; 369 break; 370 case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2): 371 case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2): 372 case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2): 373 *active_width = IB_WIDTH_1X; 374 *active_speed = IB_SPEED_HDR; 375 break; 376 case MLX5E_PROT_MASK(MLX5E_56GBASE_R4): 377 *active_width = IB_WIDTH_4X; 378 *active_speed = IB_SPEED_FDR; 379 break; 380 case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4): 381 case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4): 382 case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4): 383 case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4): 384 *active_width = IB_WIDTH_4X; 385 *active_speed = IB_SPEED_EDR; 386 break; 387 default: 388 return -EINVAL; 389 } 390 391 return 0; 392} 393 394static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed, 395 u8 *active_width) 396{ 397 switch (eth_proto_oper) { 398 case MLX5E_PROT_MASK(MLX5E_SGMII_100M): 399 case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII): 400 *active_width = IB_WIDTH_1X; 401 *active_speed = IB_SPEED_SDR; 402 break; 403 case MLX5E_PROT_MASK(MLX5E_5GBASE_R): 404 *active_width = IB_WIDTH_1X; 405 *active_speed = IB_SPEED_DDR; 406 break; 407 case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1): 408 *active_width = IB_WIDTH_1X; 409 *active_speed = IB_SPEED_QDR; 410 break; 411 case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4): 412 *active_width = IB_WIDTH_4X; 413 *active_speed = IB_SPEED_QDR; 414 break; 415 case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR): 416 *active_width = IB_WIDTH_1X; 417 *active_speed = IB_SPEED_EDR; 418 break; 419 case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2): 420 *active_width = IB_WIDTH_2X; 421 *active_speed = IB_SPEED_EDR; 422 break; 423 case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR): 424 *active_width = IB_WIDTH_1X; 425 *active_speed = IB_SPEED_HDR; 426 break; 427 case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4): 428 *active_width = IB_WIDTH_4X; 429 *active_speed = IB_SPEED_EDR; 430 break; 431 case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2): 432 *active_width = IB_WIDTH_2X; 433 *active_speed = IB_SPEED_HDR; 434 break; 435 case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR): 436 *active_width = IB_WIDTH_1X; 437 *active_speed = IB_SPEED_NDR; 438 break; 439 case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4): 440 *active_width = IB_WIDTH_4X; 441 *active_speed = IB_SPEED_HDR; 442 break; 443 case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2): 444 *active_width = IB_WIDTH_2X; 445 *active_speed = IB_SPEED_NDR; 446 break; 447 case MLX5E_PROT_MASK(MLX5E_400GAUI_8): 448 *active_width = IB_WIDTH_8X; 449 *active_speed = IB_SPEED_HDR; 450 break; 451 case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4): 452 *active_width = IB_WIDTH_4X; 453 *active_speed = IB_SPEED_NDR; 454 break; 455 default: 456 return -EINVAL; 457 } 458 459 return 0; 460} 461 462static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed, 463 u8 *active_width, bool ext) 464{ 465 return ext ? 466 translate_eth_ext_proto_oper(eth_proto_oper, active_speed, 467 active_width) : 468 translate_eth_legacy_proto_oper(eth_proto_oper, active_speed, 469 active_width); 470} 471 472static int mlx5_query_port_roce(struct ib_device *device, u32 port_num, 473 struct ib_port_attr *props) 474{ 475 struct mlx5_ib_dev *dev = to_mdev(device); 476 u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0}; 477 struct mlx5_core_dev *mdev; 478 struct net_device *ndev, *upper; 479 enum ib_mtu ndev_ib_mtu; 480 bool put_mdev = true; 481 u32 eth_prot_oper; 482 u32 mdev_port_num; 483 bool ext; 484 int err; 485 486 mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num); 487 if (!mdev) { 488 /* This means the port isn't affiliated yet. Get the 489 * info for the master port instead. 490 */ 491 put_mdev = false; 492 mdev = dev->mdev; 493 mdev_port_num = 1; 494 port_num = 1; 495 } 496 497 /* Possible bad flows are checked before filling out props so in case 498 * of an error it will still be zeroed out. 499 * Use native port in case of reps 500 */ 501 if (dev->is_rep) 502 err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN, 503 1); 504 else 505 err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN, 506 mdev_port_num); 507 if (err) 508 goto out; 509 ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability); 510 eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper); 511 512 props->active_width = IB_WIDTH_4X; 513 props->active_speed = IB_SPEED_QDR; 514 515 translate_eth_proto_oper(eth_prot_oper, &props->active_speed, 516 &props->active_width, ext); 517 518 if (!dev->is_rep && dev->mdev->roce.roce_en) { 519 u16 qkey_viol_cntr; 520 521 props->port_cap_flags |= IB_PORT_CM_SUP; 522 props->ip_gids = true; 523 props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev, 524 roce_address_table_size); 525 mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr); 526 props->qkey_viol_cntr = qkey_viol_cntr; 527 } 528 props->max_mtu = IB_MTU_4096; 529 props->max_msg_sz = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg); 530 props->pkey_tbl_len = 1; 531 props->state = IB_PORT_DOWN; 532 props->phys_state = IB_PORT_PHYS_STATE_DISABLED; 533 534 /* If this is a stub query for an unaffiliated port stop here */ 535 if (!put_mdev) 536 goto out; 537 538 ndev = mlx5_ib_get_netdev(device, port_num); 539 if (!ndev) 540 goto out; 541 542 if (dev->lag_active) { 543 rcu_read_lock(); 544 upper = netdev_master_upper_dev_get_rcu(ndev); 545 if (upper) { 546 dev_put(ndev); 547 ndev = upper; 548 dev_hold(ndev); 549 } 550 rcu_read_unlock(); 551 } 552 553 if (netif_running(ndev) && netif_carrier_ok(ndev)) { 554 props->state = IB_PORT_ACTIVE; 555 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP; 556 } 557 558 ndev_ib_mtu = iboe_get_mtu(ndev->mtu); 559 560 dev_put(ndev); 561 562 props->active_mtu = min(props->max_mtu, ndev_ib_mtu); 563out: 564 if (put_mdev) 565 mlx5_ib_put_native_port_mdev(dev, port_num); 566 return err; 567} 568 569int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num, 570 unsigned int index, const union ib_gid *gid, 571 const struct ib_gid_attr *attr) 572{ 573 enum ib_gid_type gid_type; 574 u16 vlan_id = 0xffff; 575 u8 roce_version = 0; 576 u8 roce_l3_type = 0; 577 u8 mac[ETH_ALEN]; 578 int ret; 579 580 gid_type = attr->gid_type; 581 if (gid) { 582 ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]); 583 if (ret) 584 return ret; 585 } 586 587 switch (gid_type) { 588 case IB_GID_TYPE_ROCE: 589 roce_version = MLX5_ROCE_VERSION_1; 590 break; 591 case IB_GID_TYPE_ROCE_UDP_ENCAP: 592 roce_version = MLX5_ROCE_VERSION_2; 593 if (gid && ipv6_addr_v4mapped((void *)gid)) 594 roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4; 595 else 596 roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6; 597 break; 598 599 default: 600 mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type); 601 } 602 603 return mlx5_core_roce_gid_set(dev->mdev, index, roce_version, 604 roce_l3_type, gid->raw, mac, 605 vlan_id < VLAN_CFI_MASK, vlan_id, 606 port_num); 607} 608 609static int mlx5_ib_add_gid(const struct ib_gid_attr *attr, 610 __always_unused void **context) 611{ 612 int ret; 613 614 ret = mlx5r_add_gid_macsec_operations(attr); 615 if (ret) 616 return ret; 617 618 return set_roce_addr(to_mdev(attr->device), attr->port_num, 619 attr->index, &attr->gid, attr); 620} 621 622static int mlx5_ib_del_gid(const struct ib_gid_attr *attr, 623 __always_unused void **context) 624{ 625 int ret; 626 627 ret = set_roce_addr(to_mdev(attr->device), attr->port_num, 628 attr->index, NULL, attr); 629 if (ret) 630 return ret; 631 632 mlx5r_del_gid_macsec_operations(attr); 633 return 0; 634} 635 636__be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev, 637 const struct ib_gid_attr *attr) 638{ 639 if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP) 640 return 0; 641 642 return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port)); 643} 644 645static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev) 646{ 647 if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB) 648 return !MLX5_CAP_GEN(dev->mdev, ib_virt); 649 return 0; 650} 651 652enum { 653 MLX5_VPORT_ACCESS_METHOD_MAD, 654 MLX5_VPORT_ACCESS_METHOD_HCA, 655 MLX5_VPORT_ACCESS_METHOD_NIC, 656}; 657 658static int mlx5_get_vport_access_method(struct ib_device *ibdev) 659{ 660 if (mlx5_use_mad_ifc(to_mdev(ibdev))) 661 return MLX5_VPORT_ACCESS_METHOD_MAD; 662 663 if (mlx5_ib_port_link_layer(ibdev, 1) == 664 IB_LINK_LAYER_ETHERNET) 665 return MLX5_VPORT_ACCESS_METHOD_NIC; 666 667 return MLX5_VPORT_ACCESS_METHOD_HCA; 668} 669 670static void get_atomic_caps(struct mlx5_ib_dev *dev, 671 u8 atomic_size_qp, 672 struct ib_device_attr *props) 673{ 674 u8 tmp; 675 u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations); 676 u8 atomic_req_8B_endianness_mode = 677 MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode); 678 679 /* Check if HW supports 8 bytes standard atomic operations and capable 680 * of host endianness respond 681 */ 682 tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD; 683 if (((atomic_operations & tmp) == tmp) && 684 (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) && 685 (atomic_req_8B_endianness_mode)) { 686 props->atomic_cap = IB_ATOMIC_HCA; 687 } else { 688 props->atomic_cap = IB_ATOMIC_NONE; 689 } 690} 691 692static void get_atomic_caps_qp(struct mlx5_ib_dev *dev, 693 struct ib_device_attr *props) 694{ 695 u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp); 696 697 get_atomic_caps(dev, atomic_size_qp, props); 698} 699 700static int mlx5_query_system_image_guid(struct ib_device *ibdev, 701 __be64 *sys_image_guid) 702{ 703 struct mlx5_ib_dev *dev = to_mdev(ibdev); 704 struct mlx5_core_dev *mdev = dev->mdev; 705 u64 tmp; 706 int err; 707 708 switch (mlx5_get_vport_access_method(ibdev)) { 709 case MLX5_VPORT_ACCESS_METHOD_MAD: 710 return mlx5_query_mad_ifc_system_image_guid(ibdev, 711 sys_image_guid); 712 713 case MLX5_VPORT_ACCESS_METHOD_HCA: 714 err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp); 715 break; 716 717 case MLX5_VPORT_ACCESS_METHOD_NIC: 718 err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp); 719 break; 720 721 default: 722 return -EINVAL; 723 } 724 725 if (!err) 726 *sys_image_guid = cpu_to_be64(tmp); 727 728 return err; 729 730} 731 732static int mlx5_query_max_pkeys(struct ib_device *ibdev, 733 u16 *max_pkeys) 734{ 735 struct mlx5_ib_dev *dev = to_mdev(ibdev); 736 struct mlx5_core_dev *mdev = dev->mdev; 737 738 switch (mlx5_get_vport_access_method(ibdev)) { 739 case MLX5_VPORT_ACCESS_METHOD_MAD: 740 return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys); 741 742 case MLX5_VPORT_ACCESS_METHOD_HCA: 743 case MLX5_VPORT_ACCESS_METHOD_NIC: 744 *max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, 745 pkey_table_size)); 746 return 0; 747 748 default: 749 return -EINVAL; 750 } 751} 752 753static int mlx5_query_vendor_id(struct ib_device *ibdev, 754 u32 *vendor_id) 755{ 756 struct mlx5_ib_dev *dev = to_mdev(ibdev); 757 758 switch (mlx5_get_vport_access_method(ibdev)) { 759 case MLX5_VPORT_ACCESS_METHOD_MAD: 760 return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id); 761 762 case MLX5_VPORT_ACCESS_METHOD_HCA: 763 case MLX5_VPORT_ACCESS_METHOD_NIC: 764 return mlx5_core_query_vendor_id(dev->mdev, vendor_id); 765 766 default: 767 return -EINVAL; 768 } 769} 770 771static int mlx5_query_node_guid(struct mlx5_ib_dev *dev, 772 __be64 *node_guid) 773{ 774 u64 tmp; 775 int err; 776 777 switch (mlx5_get_vport_access_method(&dev->ib_dev)) { 778 case MLX5_VPORT_ACCESS_METHOD_MAD: 779 return mlx5_query_mad_ifc_node_guid(dev, node_guid); 780 781 case MLX5_VPORT_ACCESS_METHOD_HCA: 782 err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp); 783 break; 784 785 case MLX5_VPORT_ACCESS_METHOD_NIC: 786 err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp); 787 break; 788 789 default: 790 return -EINVAL; 791 } 792 793 if (!err) 794 *node_guid = cpu_to_be64(tmp); 795 796 return err; 797} 798 799struct mlx5_reg_node_desc { 800 u8 desc[IB_DEVICE_NODE_DESC_MAX]; 801}; 802 803static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc) 804{ 805 struct mlx5_reg_node_desc in; 806 807 if (mlx5_use_mad_ifc(dev)) 808 return mlx5_query_mad_ifc_node_desc(dev, node_desc); 809 810 memset(&in, 0, sizeof(in)); 811 812 return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc, 813 sizeof(struct mlx5_reg_node_desc), 814 MLX5_REG_NODE_DESC, 0, 0); 815} 816 817static int mlx5_ib_query_device(struct ib_device *ibdev, 818 struct ib_device_attr *props, 819 struct ib_udata *uhw) 820{ 821 size_t uhw_outlen = (uhw) ? uhw->outlen : 0; 822 struct mlx5_ib_dev *dev = to_mdev(ibdev); 823 struct mlx5_core_dev *mdev = dev->mdev; 824 int err = -ENOMEM; 825 int max_sq_desc; 826 int max_rq_sg; 827 int max_sq_sg; 828 u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz); 829 bool raw_support = !mlx5_core_mp_enabled(mdev); 830 struct mlx5_ib_query_device_resp resp = {}; 831 size_t resp_len; 832 u64 max_tso; 833 834 resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length); 835 if (uhw_outlen && uhw_outlen < resp_len) 836 return -EINVAL; 837 838 resp.response_length = resp_len; 839 840 if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen)) 841 return -EINVAL; 842 843 memset(props, 0, sizeof(*props)); 844 err = mlx5_query_system_image_guid(ibdev, 845 &props->sys_image_guid); 846 if (err) 847 return err; 848 849 props->max_pkeys = dev->pkey_table_len; 850 851 err = mlx5_query_vendor_id(ibdev, &props->vendor_id); 852 if (err) 853 return err; 854 855 props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) | 856 (fw_rev_min(dev->mdev) << 16) | 857 fw_rev_sub(dev->mdev); 858 props->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT | 859 IB_DEVICE_PORT_ACTIVE_EVENT | 860 IB_DEVICE_SYS_IMAGE_GUID | 861 IB_DEVICE_RC_RNR_NAK_GEN; 862 863 if (MLX5_CAP_GEN(mdev, pkv)) 864 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR; 865 if (MLX5_CAP_GEN(mdev, qkv)) 866 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR; 867 if (MLX5_CAP_GEN(mdev, apm)) 868 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG; 869 if (MLX5_CAP_GEN(mdev, xrc)) 870 props->device_cap_flags |= IB_DEVICE_XRC; 871 if (MLX5_CAP_GEN(mdev, imaicl)) { 872 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW | 873 IB_DEVICE_MEM_WINDOW_TYPE_2B; 874 props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey); 875 /* We support 'Gappy' memory registration too */ 876 props->kernel_cap_flags |= IBK_SG_GAPS_REG; 877 } 878 /* IB_WR_REG_MR always requires changing the entity size with UMR */ 879 if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled)) 880 props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS; 881 if (MLX5_CAP_GEN(mdev, sho)) { 882 props->kernel_cap_flags |= IBK_INTEGRITY_HANDOVER; 883 /* At this stage no support for signature handover */ 884 props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 | 885 IB_PROT_T10DIF_TYPE_2 | 886 IB_PROT_T10DIF_TYPE_3; 887 props->sig_guard_cap = IB_GUARD_T10DIF_CRC | 888 IB_GUARD_T10DIF_CSUM; 889 } 890 if (MLX5_CAP_GEN(mdev, block_lb_mc)) 891 props->kernel_cap_flags |= IBK_BLOCK_MULTICAST_LOOPBACK; 892 893 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) { 894 if (MLX5_CAP_ETH(mdev, csum_cap)) { 895 /* Legacy bit to support old userspace libraries */ 896 props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM; 897 props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM; 898 } 899 900 if (MLX5_CAP_ETH(dev->mdev, vlan_cap)) 901 props->raw_packet_caps |= 902 IB_RAW_PACKET_CAP_CVLAN_STRIPPING; 903 904 if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) { 905 max_tso = MLX5_CAP_ETH(mdev, max_lso_cap); 906 if (max_tso) { 907 resp.tso_caps.max_tso = 1 << max_tso; 908 resp.tso_caps.supported_qpts |= 909 1 << IB_QPT_RAW_PACKET; 910 resp.response_length += sizeof(resp.tso_caps); 911 } 912 } 913 914 if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) { 915 resp.rss_caps.rx_hash_function = 916 MLX5_RX_HASH_FUNC_TOEPLITZ; 917 resp.rss_caps.rx_hash_fields_mask = 918 MLX5_RX_HASH_SRC_IPV4 | 919 MLX5_RX_HASH_DST_IPV4 | 920 MLX5_RX_HASH_SRC_IPV6 | 921 MLX5_RX_HASH_DST_IPV6 | 922 MLX5_RX_HASH_SRC_PORT_TCP | 923 MLX5_RX_HASH_DST_PORT_TCP | 924 MLX5_RX_HASH_SRC_PORT_UDP | 925 MLX5_RX_HASH_DST_PORT_UDP | 926 MLX5_RX_HASH_INNER; 927 resp.response_length += sizeof(resp.rss_caps); 928 } 929 } else { 930 if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) 931 resp.response_length += sizeof(resp.tso_caps); 932 if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) 933 resp.response_length += sizeof(resp.rss_caps); 934 } 935 936 if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) { 937 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM; 938 props->kernel_cap_flags |= IBK_UD_TSO; 939 } 940 941 if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) && 942 MLX5_CAP_GEN(dev->mdev, general_notification_event) && 943 raw_support) 944 props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP; 945 946 if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) && 947 MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap)) 948 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM; 949 950 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && 951 MLX5_CAP_ETH(dev->mdev, scatter_fcs) && 952 raw_support) { 953 /* Legacy bit to support old userspace libraries */ 954 props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS; 955 props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS; 956 } 957 958 if (MLX5_CAP_DEV_MEM(mdev, memic)) { 959 props->max_dm_size = 960 MLX5_CAP_DEV_MEM(mdev, max_memic_size); 961 } 962 963 if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS)) 964 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING; 965 966 if (MLX5_CAP_GEN(mdev, end_pad)) 967 props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING; 968 969 props->vendor_part_id = mdev->pdev->device; 970 props->hw_ver = mdev->pdev->revision; 971 972 props->max_mr_size = ~0ull; 973 props->page_size_cap = ~(min_page_size - 1); 974 props->max_qp = 1 << MLX5_CAP_GEN(mdev, log_max_qp); 975 props->max_qp_wr = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz); 976 max_rq_sg = MLX5_CAP_GEN(mdev, max_wqe_sz_rq) / 977 sizeof(struct mlx5_wqe_data_seg); 978 max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512); 979 max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) - 980 sizeof(struct mlx5_wqe_raddr_seg)) / 981 sizeof(struct mlx5_wqe_data_seg); 982 props->max_send_sge = max_sq_sg; 983 props->max_recv_sge = max_rq_sg; 984 props->max_sge_rd = MLX5_MAX_SGE_RD; 985 props->max_cq = 1 << MLX5_CAP_GEN(mdev, log_max_cq); 986 props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1; 987 props->max_mr = 1 << MLX5_CAP_GEN(mdev, log_max_mkey); 988 props->max_pd = 1 << MLX5_CAP_GEN(mdev, log_max_pd); 989 props->max_qp_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp); 990 props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp); 991 props->max_srq = 1 << MLX5_CAP_GEN(mdev, log_max_srq); 992 props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1; 993 props->local_ca_ack_delay = MLX5_CAP_GEN(mdev, local_ca_ack_delay); 994 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp; 995 props->max_srq_sge = max_rq_sg - 1; 996 props->max_fast_reg_page_list_len = 997 1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size); 998 props->max_pi_fast_reg_page_list_len = 999 props->max_fast_reg_page_list_len / 2; 1000 props->max_sgl_rd = 1001 MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance); 1002 get_atomic_caps_qp(dev, props); 1003 props->masked_atomic_cap = IB_ATOMIC_NONE; 1004 props->max_mcast_grp = 1 << MLX5_CAP_GEN(mdev, log_max_mcg); 1005 props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg); 1006 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach * 1007 props->max_mcast_grp; 1008 props->max_ah = INT_MAX; 1009 props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz); 1010 props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL; 1011 1012 if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) { 1013 if (dev->odp_caps.general_caps & IB_ODP_SUPPORT) 1014 props->kernel_cap_flags |= IBK_ON_DEMAND_PAGING; 1015 props->odp_caps = dev->odp_caps; 1016 if (!uhw) { 1017 /* ODP for kernel QPs is not implemented for receive 1018 * WQEs and SRQ WQEs 1019 */ 1020 props->odp_caps.per_transport_caps.rc_odp_caps &= 1021 ~(IB_ODP_SUPPORT_READ | 1022 IB_ODP_SUPPORT_SRQ_RECV); 1023 props->odp_caps.per_transport_caps.uc_odp_caps &= 1024 ~(IB_ODP_SUPPORT_READ | 1025 IB_ODP_SUPPORT_SRQ_RECV); 1026 props->odp_caps.per_transport_caps.ud_odp_caps &= 1027 ~(IB_ODP_SUPPORT_READ | 1028 IB_ODP_SUPPORT_SRQ_RECV); 1029 props->odp_caps.per_transport_caps.xrc_odp_caps &= 1030 ~(IB_ODP_SUPPORT_READ | 1031 IB_ODP_SUPPORT_SRQ_RECV); 1032 } 1033 } 1034 1035 if (mlx5_core_is_vf(mdev)) 1036 props->kernel_cap_flags |= IBK_VIRTUAL_FUNCTION; 1037 1038 if (mlx5_ib_port_link_layer(ibdev, 1) == 1039 IB_LINK_LAYER_ETHERNET && raw_support) { 1040 props->rss_caps.max_rwq_indirection_tables = 1041 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt); 1042 props->rss_caps.max_rwq_indirection_table_size = 1043 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size); 1044 props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET; 1045 props->max_wq_type_rq = 1046 1 << MLX5_CAP_GEN(dev->mdev, log_max_rq); 1047 } 1048 1049 if (MLX5_CAP_GEN(mdev, tag_matching)) { 1050 props->tm_caps.max_num_tags = 1051 (1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1; 1052 props->tm_caps.max_ops = 1053 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz); 1054 props->tm_caps.max_sge = MLX5_TM_MAX_SGE; 1055 } 1056 1057 if (MLX5_CAP_GEN(mdev, tag_matching) && 1058 MLX5_CAP_GEN(mdev, rndv_offload_rc)) { 1059 props->tm_caps.flags = IB_TM_CAP_RNDV_RC; 1060 props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE; 1061 } 1062 1063 if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) { 1064 props->cq_caps.max_cq_moderation_count = 1065 MLX5_MAX_CQ_COUNT; 1066 props->cq_caps.max_cq_moderation_period = 1067 MLX5_MAX_CQ_PERIOD; 1068 } 1069 1070 if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) { 1071 resp.response_length += sizeof(resp.cqe_comp_caps); 1072 1073 if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) { 1074 resp.cqe_comp_caps.max_num = 1075 MLX5_CAP_GEN(dev->mdev, 1076 cqe_compression_max_num); 1077 1078 resp.cqe_comp_caps.supported_format = 1079 MLX5_IB_CQE_RES_FORMAT_HASH | 1080 MLX5_IB_CQE_RES_FORMAT_CSUM; 1081 1082 if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index)) 1083 resp.cqe_comp_caps.supported_format |= 1084 MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX; 1085 } 1086 } 1087 1088 if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen && 1089 raw_support) { 1090 if (MLX5_CAP_QOS(mdev, packet_pacing) && 1091 MLX5_CAP_GEN(mdev, qos)) { 1092 resp.packet_pacing_caps.qp_rate_limit_max = 1093 MLX5_CAP_QOS(mdev, packet_pacing_max_rate); 1094 resp.packet_pacing_caps.qp_rate_limit_min = 1095 MLX5_CAP_QOS(mdev, packet_pacing_min_rate); 1096 resp.packet_pacing_caps.supported_qpts |= 1097 1 << IB_QPT_RAW_PACKET; 1098 if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) && 1099 MLX5_CAP_QOS(mdev, packet_pacing_typical_size)) 1100 resp.packet_pacing_caps.cap_flags |= 1101 MLX5_IB_PP_SUPPORT_BURST; 1102 } 1103 resp.response_length += sizeof(resp.packet_pacing_caps); 1104 } 1105 1106 if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <= 1107 uhw_outlen) { 1108 if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe)) 1109 resp.mlx5_ib_support_multi_pkt_send_wqes = 1110 MLX5_IB_ALLOW_MPW; 1111 1112 if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe)) 1113 resp.mlx5_ib_support_multi_pkt_send_wqes |= 1114 MLX5_IB_SUPPORT_EMPW; 1115 1116 resp.response_length += 1117 sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes); 1118 } 1119 1120 if (offsetofend(typeof(resp), flags) <= uhw_outlen) { 1121 resp.response_length += sizeof(resp.flags); 1122 1123 if (MLX5_CAP_GEN(mdev, cqe_compression_128)) 1124 resp.flags |= 1125 MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP; 1126 1127 if (MLX5_CAP_GEN(mdev, cqe_128_always)) 1128 resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD; 1129 if (MLX5_CAP_GEN(mdev, qp_packet_based)) 1130 resp.flags |= 1131 MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE; 1132 1133 resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT; 1134 } 1135 1136 if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) { 1137 resp.response_length += sizeof(resp.sw_parsing_caps); 1138 if (MLX5_CAP_ETH(mdev, swp)) { 1139 resp.sw_parsing_caps.sw_parsing_offloads |= 1140 MLX5_IB_SW_PARSING; 1141 1142 if (MLX5_CAP_ETH(mdev, swp_csum)) 1143 resp.sw_parsing_caps.sw_parsing_offloads |= 1144 MLX5_IB_SW_PARSING_CSUM; 1145 1146 if (MLX5_CAP_ETH(mdev, swp_lso)) 1147 resp.sw_parsing_caps.sw_parsing_offloads |= 1148 MLX5_IB_SW_PARSING_LSO; 1149 1150 if (resp.sw_parsing_caps.sw_parsing_offloads) 1151 resp.sw_parsing_caps.supported_qpts = 1152 BIT(IB_QPT_RAW_PACKET); 1153 } 1154 } 1155 1156 if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen && 1157 raw_support) { 1158 resp.response_length += sizeof(resp.striding_rq_caps); 1159 if (MLX5_CAP_GEN(mdev, striding_rq)) { 1160 resp.striding_rq_caps.min_single_stride_log_num_of_bytes = 1161 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES; 1162 resp.striding_rq_caps.max_single_stride_log_num_of_bytes = 1163 MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES; 1164 if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range)) 1165 resp.striding_rq_caps 1166 .min_single_wqe_log_num_of_strides = 1167 MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES; 1168 else 1169 resp.striding_rq_caps 1170 .min_single_wqe_log_num_of_strides = 1171 MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES; 1172 resp.striding_rq_caps.max_single_wqe_log_num_of_strides = 1173 MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES; 1174 resp.striding_rq_caps.supported_qpts = 1175 BIT(IB_QPT_RAW_PACKET); 1176 } 1177 } 1178 1179 if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) { 1180 resp.response_length += sizeof(resp.tunnel_offloads_caps); 1181 if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan)) 1182 resp.tunnel_offloads_caps |= 1183 MLX5_IB_TUNNELED_OFFLOADS_VXLAN; 1184 if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx)) 1185 resp.tunnel_offloads_caps |= 1186 MLX5_IB_TUNNELED_OFFLOADS_GENEVE; 1187 if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre)) 1188 resp.tunnel_offloads_caps |= 1189 MLX5_IB_TUNNELED_OFFLOADS_GRE; 1190 if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre)) 1191 resp.tunnel_offloads_caps |= 1192 MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE; 1193 if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp)) 1194 resp.tunnel_offloads_caps |= 1195 MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP; 1196 } 1197 1198 if (offsetofend(typeof(resp), dci_streams_caps) <= uhw_outlen) { 1199 resp.response_length += sizeof(resp.dci_streams_caps); 1200 1201 resp.dci_streams_caps.max_log_num_concurent = 1202 MLX5_CAP_GEN(mdev, log_max_dci_stream_channels); 1203 1204 resp.dci_streams_caps.max_log_num_errored = 1205 MLX5_CAP_GEN(mdev, log_max_dci_errored_streams); 1206 } 1207 1208 if (uhw_outlen) { 1209 err = ib_copy_to_udata(uhw, &resp, resp.response_length); 1210 1211 if (err) 1212 return err; 1213 } 1214 1215 return 0; 1216} 1217 1218static void translate_active_width(struct ib_device *ibdev, u16 active_width, 1219 u8 *ib_width) 1220{ 1221 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1222 1223 if (active_width & MLX5_PTYS_WIDTH_1X) 1224 *ib_width = IB_WIDTH_1X; 1225 else if (active_width & MLX5_PTYS_WIDTH_2X) 1226 *ib_width = IB_WIDTH_2X; 1227 else if (active_width & MLX5_PTYS_WIDTH_4X) 1228 *ib_width = IB_WIDTH_4X; 1229 else if (active_width & MLX5_PTYS_WIDTH_8X) 1230 *ib_width = IB_WIDTH_8X; 1231 else if (active_width & MLX5_PTYS_WIDTH_12X) 1232 *ib_width = IB_WIDTH_12X; 1233 else { 1234 mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n", 1235 active_width); 1236 *ib_width = IB_WIDTH_4X; 1237 } 1238 1239 return; 1240} 1241 1242static int mlx5_mtu_to_ib_mtu(int mtu) 1243{ 1244 switch (mtu) { 1245 case 256: return 1; 1246 case 512: return 2; 1247 case 1024: return 3; 1248 case 2048: return 4; 1249 case 4096: return 5; 1250 default: 1251 pr_warn("invalid mtu\n"); 1252 return -1; 1253 } 1254} 1255 1256enum ib_max_vl_num { 1257 __IB_MAX_VL_0 = 1, 1258 __IB_MAX_VL_0_1 = 2, 1259 __IB_MAX_VL_0_3 = 3, 1260 __IB_MAX_VL_0_7 = 4, 1261 __IB_MAX_VL_0_14 = 5, 1262}; 1263 1264enum mlx5_vl_hw_cap { 1265 MLX5_VL_HW_0 = 1, 1266 MLX5_VL_HW_0_1 = 2, 1267 MLX5_VL_HW_0_2 = 3, 1268 MLX5_VL_HW_0_3 = 4, 1269 MLX5_VL_HW_0_4 = 5, 1270 MLX5_VL_HW_0_5 = 6, 1271 MLX5_VL_HW_0_6 = 7, 1272 MLX5_VL_HW_0_7 = 8, 1273 MLX5_VL_HW_0_14 = 15 1274}; 1275 1276static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap, 1277 u8 *max_vl_num) 1278{ 1279 switch (vl_hw_cap) { 1280 case MLX5_VL_HW_0: 1281 *max_vl_num = __IB_MAX_VL_0; 1282 break; 1283 case MLX5_VL_HW_0_1: 1284 *max_vl_num = __IB_MAX_VL_0_1; 1285 break; 1286 case MLX5_VL_HW_0_3: 1287 *max_vl_num = __IB_MAX_VL_0_3; 1288 break; 1289 case MLX5_VL_HW_0_7: 1290 *max_vl_num = __IB_MAX_VL_0_7; 1291 break; 1292 case MLX5_VL_HW_0_14: 1293 *max_vl_num = __IB_MAX_VL_0_14; 1294 break; 1295 1296 default: 1297 return -EINVAL; 1298 } 1299 1300 return 0; 1301} 1302 1303static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port, 1304 struct ib_port_attr *props) 1305{ 1306 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1307 struct mlx5_core_dev *mdev = dev->mdev; 1308 struct mlx5_hca_vport_context *rep; 1309 u16 max_mtu; 1310 u16 oper_mtu; 1311 int err; 1312 u16 ib_link_width_oper; 1313 u8 vl_hw_cap; 1314 1315 rep = kzalloc(sizeof(*rep), GFP_KERNEL); 1316 if (!rep) { 1317 err = -ENOMEM; 1318 goto out; 1319 } 1320 1321 /* props being zeroed by the caller, avoid zeroing it here */ 1322 1323 err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep); 1324 if (err) 1325 goto out; 1326 1327 props->lid = rep->lid; 1328 props->lmc = rep->lmc; 1329 props->sm_lid = rep->sm_lid; 1330 props->sm_sl = rep->sm_sl; 1331 props->state = rep->vport_state; 1332 props->phys_state = rep->port_physical_state; 1333 props->port_cap_flags = rep->cap_mask1; 1334 props->gid_tbl_len = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size)); 1335 props->max_msg_sz = 1 << MLX5_CAP_GEN(mdev, log_max_msg); 1336 props->pkey_tbl_len = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size)); 1337 props->bad_pkey_cntr = rep->pkey_violation_counter; 1338 props->qkey_viol_cntr = rep->qkey_violation_counter; 1339 props->subnet_timeout = rep->subnet_timeout; 1340 props->init_type_reply = rep->init_type_reply; 1341 1342 if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP) 1343 props->port_cap_flags2 = rep->cap_mask2; 1344 1345 err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper, 1346 &props->active_speed, port); 1347 if (err) 1348 goto out; 1349 1350 translate_active_width(ibdev, ib_link_width_oper, &props->active_width); 1351 1352 mlx5_query_port_max_mtu(mdev, &max_mtu, port); 1353 1354 props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu); 1355 1356 mlx5_query_port_oper_mtu(mdev, &oper_mtu, port); 1357 1358 props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu); 1359 1360 err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port); 1361 if (err) 1362 goto out; 1363 1364 err = translate_max_vl_num(ibdev, vl_hw_cap, 1365 &props->max_vl_num); 1366out: 1367 kfree(rep); 1368 return err; 1369} 1370 1371int mlx5_ib_query_port(struct ib_device *ibdev, u32 port, 1372 struct ib_port_attr *props) 1373{ 1374 unsigned int count; 1375 int ret; 1376 1377 switch (mlx5_get_vport_access_method(ibdev)) { 1378 case MLX5_VPORT_ACCESS_METHOD_MAD: 1379 ret = mlx5_query_mad_ifc_port(ibdev, port, props); 1380 break; 1381 1382 case MLX5_VPORT_ACCESS_METHOD_HCA: 1383 ret = mlx5_query_hca_port(ibdev, port, props); 1384 break; 1385 1386 case MLX5_VPORT_ACCESS_METHOD_NIC: 1387 ret = mlx5_query_port_roce(ibdev, port, props); 1388 break; 1389 1390 default: 1391 ret = -EINVAL; 1392 } 1393 1394 if (!ret && props) { 1395 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1396 struct mlx5_core_dev *mdev; 1397 bool put_mdev = true; 1398 1399 mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL); 1400 if (!mdev) { 1401 /* If the port isn't affiliated yet query the master. 1402 * The master and slave will have the same values. 1403 */ 1404 mdev = dev->mdev; 1405 port = 1; 1406 put_mdev = false; 1407 } 1408 count = mlx5_core_reserved_gids_count(mdev); 1409 if (put_mdev) 1410 mlx5_ib_put_native_port_mdev(dev, port); 1411 props->gid_tbl_len -= count; 1412 } 1413 return ret; 1414} 1415 1416static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port, 1417 struct ib_port_attr *props) 1418{ 1419 return mlx5_query_port_roce(ibdev, port, props); 1420} 1421 1422static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index, 1423 u16 *pkey) 1424{ 1425 /* Default special Pkey for representor device port as per the 1426 * IB specification 1.3 section 10.9.1.2. 1427 */ 1428 *pkey = 0xffff; 1429 return 0; 1430} 1431 1432static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index, 1433 union ib_gid *gid) 1434{ 1435 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1436 struct mlx5_core_dev *mdev = dev->mdev; 1437 1438 switch (mlx5_get_vport_access_method(ibdev)) { 1439 case MLX5_VPORT_ACCESS_METHOD_MAD: 1440 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid); 1441 1442 case MLX5_VPORT_ACCESS_METHOD_HCA: 1443 return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid); 1444 1445 default: 1446 return -EINVAL; 1447 } 1448 1449} 1450 1451static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port, 1452 u16 index, u16 *pkey) 1453{ 1454 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1455 struct mlx5_core_dev *mdev; 1456 bool put_mdev = true; 1457 u32 mdev_port_num; 1458 int err; 1459 1460 mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num); 1461 if (!mdev) { 1462 /* The port isn't affiliated yet, get the PKey from the master 1463 * port. For RoCE the PKey tables will be the same. 1464 */ 1465 put_mdev = false; 1466 mdev = dev->mdev; 1467 mdev_port_num = 1; 1468 } 1469 1470 err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0, 1471 index, pkey); 1472 if (put_mdev) 1473 mlx5_ib_put_native_port_mdev(dev, port); 1474 1475 return err; 1476} 1477 1478static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index, 1479 u16 *pkey) 1480{ 1481 switch (mlx5_get_vport_access_method(ibdev)) { 1482 case MLX5_VPORT_ACCESS_METHOD_MAD: 1483 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey); 1484 1485 case MLX5_VPORT_ACCESS_METHOD_HCA: 1486 case MLX5_VPORT_ACCESS_METHOD_NIC: 1487 return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey); 1488 default: 1489 return -EINVAL; 1490 } 1491} 1492 1493static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask, 1494 struct ib_device_modify *props) 1495{ 1496 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1497 struct mlx5_reg_node_desc in; 1498 struct mlx5_reg_node_desc out; 1499 int err; 1500 1501 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC) 1502 return -EOPNOTSUPP; 1503 1504 if (!(mask & IB_DEVICE_MODIFY_NODE_DESC)) 1505 return 0; 1506 1507 /* 1508 * If possible, pass node desc to FW, so it can generate 1509 * a 144 trap. If cmd fails, just ignore. 1510 */ 1511 memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX); 1512 err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out, 1513 sizeof(out), MLX5_REG_NODE_DESC, 0, 1); 1514 if (err) 1515 return err; 1516 1517 memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX); 1518 1519 return err; 1520} 1521 1522static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask, 1523 u32 value) 1524{ 1525 struct mlx5_hca_vport_context ctx = {}; 1526 struct mlx5_core_dev *mdev; 1527 u32 mdev_port_num; 1528 int err; 1529 1530 mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num); 1531 if (!mdev) 1532 return -ENODEV; 1533 1534 err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx); 1535 if (err) 1536 goto out; 1537 1538 if (~ctx.cap_mask1_perm & mask) { 1539 mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n", 1540 mask, ctx.cap_mask1_perm); 1541 err = -EINVAL; 1542 goto out; 1543 } 1544 1545 ctx.cap_mask1 = value; 1546 ctx.cap_mask1_perm = mask; 1547 err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num, 1548 0, &ctx); 1549 1550out: 1551 mlx5_ib_put_native_port_mdev(dev, port_num); 1552 1553 return err; 1554} 1555 1556static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask, 1557 struct ib_port_modify *props) 1558{ 1559 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1560 struct ib_port_attr attr; 1561 u32 tmp; 1562 int err; 1563 u32 change_mask; 1564 u32 value; 1565 bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) == 1566 IB_LINK_LAYER_INFINIBAND); 1567 1568 /* CM layer calls ib_modify_port() regardless of the link layer. For 1569 * Ethernet ports, qkey violation and Port capabilities are meaningless. 1570 */ 1571 if (!is_ib) 1572 return 0; 1573 1574 if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) { 1575 change_mask = props->clr_port_cap_mask | props->set_port_cap_mask; 1576 value = ~props->clr_port_cap_mask | props->set_port_cap_mask; 1577 return set_port_caps_atomic(dev, port, change_mask, value); 1578 } 1579 1580 mutex_lock(&dev->cap_mask_mutex); 1581 1582 err = ib_query_port(ibdev, port, &attr); 1583 if (err) 1584 goto out; 1585 1586 tmp = (attr.port_cap_flags | props->set_port_cap_mask) & 1587 ~props->clr_port_cap_mask; 1588 1589 err = mlx5_set_port_caps(dev->mdev, port, tmp); 1590 1591out: 1592 mutex_unlock(&dev->cap_mask_mutex); 1593 return err; 1594} 1595 1596static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps) 1597{ 1598 mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n", 1599 caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n"); 1600} 1601 1602static u16 calc_dynamic_bfregs(int uars_per_sys_page) 1603{ 1604 /* Large page with non 4k uar support might limit the dynamic size */ 1605 if (uars_per_sys_page == 1 && PAGE_SIZE > 4096) 1606 return MLX5_MIN_DYN_BFREGS; 1607 1608 return MLX5_MAX_DYN_BFREGS; 1609} 1610 1611static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k, 1612 struct mlx5_ib_alloc_ucontext_req_v2 *req, 1613 struct mlx5_bfreg_info *bfregi) 1614{ 1615 int uars_per_sys_page; 1616 int bfregs_per_sys_page; 1617 int ref_bfregs = req->total_num_bfregs; 1618 1619 if (req->total_num_bfregs == 0) 1620 return -EINVAL; 1621 1622 BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE); 1623 BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE); 1624 1625 if (req->total_num_bfregs > MLX5_MAX_BFREGS) 1626 return -ENOMEM; 1627 1628 uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k); 1629 bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR; 1630 /* This holds the required static allocation asked by the user */ 1631 req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page); 1632 if (req->num_low_latency_bfregs > req->total_num_bfregs - 1) 1633 return -EINVAL; 1634 1635 bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page; 1636 bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page); 1637 bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs; 1638 bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page; 1639 1640 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", 1641 MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no", 1642 lib_uar_4k ? "yes" : "no", ref_bfregs, 1643 req->total_num_bfregs, bfregi->total_num_bfregs, 1644 bfregi->num_sys_pages); 1645 1646 return 0; 1647} 1648 1649static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context) 1650{ 1651 struct mlx5_bfreg_info *bfregi; 1652 int err; 1653 int i; 1654 1655 bfregi = &context->bfregi; 1656 for (i = 0; i < bfregi->num_static_sys_pages; i++) { 1657 err = mlx5_cmd_uar_alloc(dev->mdev, &bfregi->sys_pages[i], 1658 context->devx_uid); 1659 if (err) 1660 goto error; 1661 1662 mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]); 1663 } 1664 1665 for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++) 1666 bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX; 1667 1668 return 0; 1669 1670error: 1671 for (--i; i >= 0; i--) 1672 if (mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i], 1673 context->devx_uid)) 1674 mlx5_ib_warn(dev, "failed to free uar %d\n", i); 1675 1676 return err; 1677} 1678 1679static void deallocate_uars(struct mlx5_ib_dev *dev, 1680 struct mlx5_ib_ucontext *context) 1681{ 1682 struct mlx5_bfreg_info *bfregi; 1683 int i; 1684 1685 bfregi = &context->bfregi; 1686 for (i = 0; i < bfregi->num_sys_pages; i++) 1687 if (i < bfregi->num_static_sys_pages || 1688 bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX) 1689 mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i], 1690 context->devx_uid); 1691} 1692 1693int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp) 1694{ 1695 int err = 0; 1696 1697 mutex_lock(&dev->lb.mutex); 1698 if (td) 1699 dev->lb.user_td++; 1700 if (qp) 1701 dev->lb.qps++; 1702 1703 if (dev->lb.user_td == 2 || 1704 dev->lb.qps == 1) { 1705 if (!dev->lb.enabled) { 1706 err = mlx5_nic_vport_update_local_lb(dev->mdev, true); 1707 dev->lb.enabled = true; 1708 } 1709 } 1710 1711 mutex_unlock(&dev->lb.mutex); 1712 1713 return err; 1714} 1715 1716void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp) 1717{ 1718 mutex_lock(&dev->lb.mutex); 1719 if (td) 1720 dev->lb.user_td--; 1721 if (qp) 1722 dev->lb.qps--; 1723 1724 if (dev->lb.user_td == 1 && 1725 dev->lb.qps == 0) { 1726 if (dev->lb.enabled) { 1727 mlx5_nic_vport_update_local_lb(dev->mdev, false); 1728 dev->lb.enabled = false; 1729 } 1730 } 1731 1732 mutex_unlock(&dev->lb.mutex); 1733} 1734 1735static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn, 1736 u16 uid) 1737{ 1738 int err; 1739 1740 if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) 1741 return 0; 1742 1743 err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid); 1744 if (err) 1745 return err; 1746 1747 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) || 1748 (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) && 1749 !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc))) 1750 return err; 1751 1752 return mlx5_ib_enable_lb(dev, true, false); 1753} 1754 1755static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn, 1756 u16 uid) 1757{ 1758 if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) 1759 return; 1760 1761 mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid); 1762 1763 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) || 1764 (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) && 1765 !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc))) 1766 return; 1767 1768 mlx5_ib_disable_lb(dev, true, false); 1769} 1770 1771static int set_ucontext_resp(struct ib_ucontext *uctx, 1772 struct mlx5_ib_alloc_ucontext_resp *resp) 1773{ 1774 struct ib_device *ibdev = uctx->device; 1775 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1776 struct mlx5_ib_ucontext *context = to_mucontext(uctx); 1777 struct mlx5_bfreg_info *bfregi = &context->bfregi; 1778 1779 if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) { 1780 resp->dump_fill_mkey = dev->mkeys.dump_fill_mkey; 1781 resp->comp_mask |= 1782 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY; 1783 } 1784 1785 resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp); 1786 if (dev->wc_support) 1787 resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, 1788 log_bf_reg_size); 1789 resp->cache_line_size = cache_line_size(); 1790 resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq); 1791 resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq); 1792 resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz); 1793 resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz); 1794 resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz); 1795 resp->cqe_version = context->cqe_version; 1796 resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ? 1797 MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT; 1798 resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? 1799 MLX5_CAP_GEN(dev->mdev, 1800 num_of_uars_per_page) : 1; 1801 resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 : 1802 bfregi->total_num_bfregs - bfregi->num_dyn_bfregs; 1803 resp->num_ports = dev->num_ports; 1804 resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE | 1805 MLX5_USER_CMDS_SUPP_UHW_CREATE_AH; 1806 1807 if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) { 1808 mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline); 1809 resp->eth_min_inline++; 1810 } 1811 1812 if (dev->mdev->clock_info) 1813 resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1); 1814 1815 /* 1816 * We don't want to expose information from the PCI bar that is located 1817 * after 4096 bytes, so if the arch only supports larger pages, let's 1818 * pretend we don't support reading the HCA's core clock. This is also 1819 * forced by mmap function. 1820 */ 1821 if (PAGE_SIZE <= 4096) { 1822 resp->comp_mask |= 1823 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET; 1824 resp->hca_core_clock_offset = 1825 offsetof(struct mlx5_init_seg, 1826 internal_timer_h) % PAGE_SIZE; 1827 } 1828 1829 if (MLX5_CAP_GEN(dev->mdev, ece_support)) 1830 resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE; 1831 1832 if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) && 1833 rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) && 1834 rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format))) 1835 resp->comp_mask |= 1836 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS; 1837 1838 resp->num_dyn_bfregs = bfregi->num_dyn_bfregs; 1839 1840 if (MLX5_CAP_GEN(dev->mdev, drain_sigerr)) 1841 resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS; 1842 1843 resp->comp_mask |= 1844 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_MKEY_UPDATE_TAG; 1845 1846 return 0; 1847} 1848 1849static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx, 1850 struct ib_udata *udata) 1851{ 1852 struct ib_device *ibdev = uctx->device; 1853 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1854 struct mlx5_ib_alloc_ucontext_req_v2 req = {}; 1855 struct mlx5_ib_alloc_ucontext_resp resp = {}; 1856 struct mlx5_ib_ucontext *context = to_mucontext(uctx); 1857 struct mlx5_bfreg_info *bfregi; 1858 int ver; 1859 int err; 1860 size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2, 1861 max_cqe_version); 1862 bool lib_uar_4k; 1863 bool lib_uar_dyn; 1864 1865 if (!dev->ib_active) 1866 return -EAGAIN; 1867 1868 if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req)) 1869 ver = 0; 1870 else if (udata->inlen >= min_req_v2) 1871 ver = 2; 1872 else 1873 return -EINVAL; 1874 1875 err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req))); 1876 if (err) 1877 return err; 1878 1879 if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX) 1880 return -EOPNOTSUPP; 1881 1882 if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2) 1883 return -EOPNOTSUPP; 1884 1885 req.total_num_bfregs = ALIGN(req.total_num_bfregs, 1886 MLX5_NON_FP_BFREGS_PER_UAR); 1887 if (req.num_low_latency_bfregs > req.total_num_bfregs - 1) 1888 return -EINVAL; 1889 1890 if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) { 1891 err = mlx5_ib_devx_create(dev, true); 1892 if (err < 0) 1893 goto out_ctx; 1894 context->devx_uid = err; 1895 } 1896 1897 lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR; 1898 lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR; 1899 bfregi = &context->bfregi; 1900 1901 if (lib_uar_dyn) { 1902 bfregi->lib_uar_dyn = lib_uar_dyn; 1903 goto uar_done; 1904 } 1905 1906 /* updates req->total_num_bfregs */ 1907 err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi); 1908 if (err) 1909 goto out_devx; 1910 1911 mutex_init(&bfregi->lock); 1912 bfregi->lib_uar_4k = lib_uar_4k; 1913 bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count), 1914 GFP_KERNEL); 1915 if (!bfregi->count) { 1916 err = -ENOMEM; 1917 goto out_devx; 1918 } 1919 1920 bfregi->sys_pages = kcalloc(bfregi->num_sys_pages, 1921 sizeof(*bfregi->sys_pages), 1922 GFP_KERNEL); 1923 if (!bfregi->sys_pages) { 1924 err = -ENOMEM; 1925 goto out_count; 1926 } 1927 1928 err = allocate_uars(dev, context); 1929 if (err) 1930 goto out_sys_pages; 1931 1932uar_done: 1933 err = mlx5_ib_alloc_transport_domain(dev, &context->tdn, 1934 context->devx_uid); 1935 if (err) 1936 goto out_uars; 1937 1938 INIT_LIST_HEAD(&context->db_page_list); 1939 mutex_init(&context->db_page_mutex); 1940 1941 context->cqe_version = min_t(__u8, 1942 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version), 1943 req.max_cqe_version); 1944 1945 err = set_ucontext_resp(uctx, &resp); 1946 if (err) 1947 goto out_mdev; 1948 1949 resp.response_length = min(udata->outlen, sizeof(resp)); 1950 err = ib_copy_to_udata(udata, &resp, resp.response_length); 1951 if (err) 1952 goto out_mdev; 1953 1954 bfregi->ver = ver; 1955 bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs; 1956 context->lib_caps = req.lib_caps; 1957 print_lib_caps(dev, context->lib_caps); 1958 1959 if (mlx5_ib_lag_should_assign_affinity(dev)) { 1960 u32 port = mlx5_core_native_port_num(dev->mdev) - 1; 1961 1962 atomic_set(&context->tx_port_affinity, 1963 atomic_add_return( 1964 1, &dev->port[port].roce.tx_port_affinity)); 1965 } 1966 1967 return 0; 1968 1969out_mdev: 1970 mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid); 1971 1972out_uars: 1973 deallocate_uars(dev, context); 1974 1975out_sys_pages: 1976 kfree(bfregi->sys_pages); 1977 1978out_count: 1979 kfree(bfregi->count); 1980 1981out_devx: 1982 if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) 1983 mlx5_ib_devx_destroy(dev, context->devx_uid); 1984 1985out_ctx: 1986 return err; 1987} 1988 1989static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext, 1990 struct uverbs_attr_bundle *attrs) 1991{ 1992 struct mlx5_ib_alloc_ucontext_resp uctx_resp = {}; 1993 int ret; 1994 1995 ret = set_ucontext_resp(ibcontext, &uctx_resp); 1996 if (ret) 1997 return ret; 1998 1999 uctx_resp.response_length = 2000 min_t(size_t, 2001 uverbs_attr_get_len(attrs, 2002 MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX), 2003 sizeof(uctx_resp)); 2004 2005 ret = uverbs_copy_to_struct_or_zero(attrs, 2006 MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX, 2007 &uctx_resp, 2008 sizeof(uctx_resp)); 2009 return ret; 2010} 2011 2012static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext) 2013{ 2014 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext); 2015 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device); 2016 struct mlx5_bfreg_info *bfregi; 2017 2018 bfregi = &context->bfregi; 2019 mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid); 2020 2021 deallocate_uars(dev, context); 2022 kfree(bfregi->sys_pages); 2023 kfree(bfregi->count); 2024 2025 if (context->devx_uid) 2026 mlx5_ib_devx_destroy(dev, context->devx_uid); 2027} 2028 2029static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, 2030 int uar_idx) 2031{ 2032 int fw_uars_per_page; 2033 2034 fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1; 2035 2036 return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page; 2037} 2038 2039static u64 uar_index2paddress(struct mlx5_ib_dev *dev, 2040 int uar_idx) 2041{ 2042 unsigned int fw_uars_per_page; 2043 2044 fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? 2045 MLX5_UARS_IN_PAGE : 1; 2046 2047 return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE); 2048} 2049 2050static int get_command(unsigned long offset) 2051{ 2052 return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK; 2053} 2054 2055static int get_arg(unsigned long offset) 2056{ 2057 return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1); 2058} 2059 2060static int get_index(unsigned long offset) 2061{ 2062 return get_arg(offset); 2063} 2064 2065/* Index resides in an extra byte to enable larger values than 255 */ 2066static int get_extended_index(unsigned long offset) 2067{ 2068 return get_arg(offset) | ((offset >> 16) & 0xff) << 8; 2069} 2070 2071 2072static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext) 2073{ 2074} 2075 2076static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd) 2077{ 2078 switch (cmd) { 2079 case MLX5_IB_MMAP_WC_PAGE: 2080 return "WC"; 2081 case MLX5_IB_MMAP_REGULAR_PAGE: 2082 return "best effort WC"; 2083 case MLX5_IB_MMAP_NC_PAGE: 2084 return "NC"; 2085 case MLX5_IB_MMAP_DEVICE_MEM: 2086 return "Device Memory"; 2087 default: 2088 return "Unknown"; 2089 } 2090} 2091 2092static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev, 2093 struct vm_area_struct *vma, 2094 struct mlx5_ib_ucontext *context) 2095{ 2096 if ((vma->vm_end - vma->vm_start != PAGE_SIZE) || 2097 !(vma->vm_flags & VM_SHARED)) 2098 return -EINVAL; 2099 2100 if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1) 2101 return -EOPNOTSUPP; 2102 2103 if (vma->vm_flags & (VM_WRITE | VM_EXEC)) 2104 return -EPERM; 2105 vm_flags_clear(vma, VM_MAYWRITE); 2106 2107 if (!dev->mdev->clock_info) 2108 return -EOPNOTSUPP; 2109 2110 return vm_insert_page(vma, vma->vm_start, 2111 virt_to_page(dev->mdev->clock_info)); 2112} 2113 2114static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry) 2115{ 2116 struct mlx5_user_mmap_entry *mentry = to_mmmap(entry); 2117 struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device); 2118 struct mlx5_var_table *var_table = &dev->var_table; 2119 struct mlx5_ib_ucontext *context = to_mucontext(entry->ucontext); 2120 2121 switch (mentry->mmap_flag) { 2122 case MLX5_IB_MMAP_TYPE_MEMIC: 2123 case MLX5_IB_MMAP_TYPE_MEMIC_OP: 2124 mlx5_ib_dm_mmap_free(dev, mentry); 2125 break; 2126 case MLX5_IB_MMAP_TYPE_VAR: 2127 mutex_lock(&var_table->bitmap_lock); 2128 clear_bit(mentry->page_idx, var_table->bitmap); 2129 mutex_unlock(&var_table->bitmap_lock); 2130 kfree(mentry); 2131 break; 2132 case MLX5_IB_MMAP_TYPE_UAR_WC: 2133 case MLX5_IB_MMAP_TYPE_UAR_NC: 2134 mlx5_cmd_uar_dealloc(dev->mdev, mentry->page_idx, 2135 context->devx_uid); 2136 kfree(mentry); 2137 break; 2138 default: 2139 WARN_ON(true); 2140 } 2141} 2142 2143static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd, 2144 struct vm_area_struct *vma, 2145 struct mlx5_ib_ucontext *context) 2146{ 2147 struct mlx5_bfreg_info *bfregi = &context->bfregi; 2148 int err; 2149 unsigned long idx; 2150 phys_addr_t pfn; 2151 pgprot_t prot; 2152 u32 bfreg_dyn_idx = 0; 2153 u32 uar_index; 2154 int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC); 2155 int max_valid_idx = dyn_uar ? bfregi->num_sys_pages : 2156 bfregi->num_static_sys_pages; 2157 2158 if (bfregi->lib_uar_dyn) 2159 return -EINVAL; 2160 2161 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 2162 return -EINVAL; 2163 2164 if (dyn_uar) 2165 idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages; 2166 else 2167 idx = get_index(vma->vm_pgoff); 2168 2169 if (idx >= max_valid_idx) { 2170 mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n", 2171 idx, max_valid_idx); 2172 return -EINVAL; 2173 } 2174 2175 switch (cmd) { 2176 case MLX5_IB_MMAP_WC_PAGE: 2177 case MLX5_IB_MMAP_ALLOC_WC: 2178 case MLX5_IB_MMAP_REGULAR_PAGE: 2179 /* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */ 2180 prot = pgprot_writecombine(vma->vm_page_prot); 2181 break; 2182 case MLX5_IB_MMAP_NC_PAGE: 2183 prot = pgprot_noncached(vma->vm_page_prot); 2184 break; 2185 default: 2186 return -EINVAL; 2187 } 2188 2189 if (dyn_uar) { 2190 int uars_per_page; 2191 2192 uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k); 2193 bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR); 2194 if (bfreg_dyn_idx >= bfregi->total_num_bfregs) { 2195 mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n", 2196 bfreg_dyn_idx, bfregi->total_num_bfregs); 2197 return -EINVAL; 2198 } 2199 2200 mutex_lock(&bfregi->lock); 2201 /* Fail if uar already allocated, first bfreg index of each 2202 * page holds its count. 2203 */ 2204 if (bfregi->count[bfreg_dyn_idx]) { 2205 mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx); 2206 mutex_unlock(&bfregi->lock); 2207 return -EINVAL; 2208 } 2209 2210 bfregi->count[bfreg_dyn_idx]++; 2211 mutex_unlock(&bfregi->lock); 2212 2213 err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index, 2214 context->devx_uid); 2215 if (err) { 2216 mlx5_ib_warn(dev, "UAR alloc failed\n"); 2217 goto free_bfreg; 2218 } 2219 } else { 2220 uar_index = bfregi->sys_pages[idx]; 2221 } 2222 2223 pfn = uar_index2pfn(dev, uar_index); 2224 mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn); 2225 2226 err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE, 2227 prot, NULL); 2228 if (err) { 2229 mlx5_ib_err(dev, 2230 "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n", 2231 err, mmap_cmd2str(cmd)); 2232 goto err; 2233 } 2234 2235 if (dyn_uar) 2236 bfregi->sys_pages[idx] = uar_index; 2237 return 0; 2238 2239err: 2240 if (!dyn_uar) 2241 return err; 2242 2243 mlx5_cmd_uar_dealloc(dev->mdev, idx, context->devx_uid); 2244 2245free_bfreg: 2246 mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx); 2247 2248 return err; 2249} 2250 2251static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma) 2252{ 2253 unsigned long idx; 2254 u8 command; 2255 2256 command = get_command(vma->vm_pgoff); 2257 idx = get_extended_index(vma->vm_pgoff); 2258 2259 return (command << 16 | idx); 2260} 2261 2262static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev, 2263 struct vm_area_struct *vma, 2264 struct ib_ucontext *ucontext) 2265{ 2266 struct mlx5_user_mmap_entry *mentry; 2267 struct rdma_user_mmap_entry *entry; 2268 unsigned long pgoff; 2269 pgprot_t prot; 2270 phys_addr_t pfn; 2271 int ret; 2272 2273 pgoff = mlx5_vma_to_pgoff(vma); 2274 entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff); 2275 if (!entry) 2276 return -EINVAL; 2277 2278 mentry = to_mmmap(entry); 2279 pfn = (mentry->address >> PAGE_SHIFT); 2280 if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR || 2281 mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC) 2282 prot = pgprot_noncached(vma->vm_page_prot); 2283 else 2284 prot = pgprot_writecombine(vma->vm_page_prot); 2285 ret = rdma_user_mmap_io(ucontext, vma, pfn, 2286 entry->npages * PAGE_SIZE, 2287 prot, 2288 entry); 2289 rdma_user_mmap_entry_put(&mentry->rdma_entry); 2290 return ret; 2291} 2292 2293static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry) 2294{ 2295 u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF; 2296 u64 index = entry->rdma_entry.start_pgoff & 0xFFFF; 2297 2298 return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) | 2299 (index & 0xFF)) << PAGE_SHIFT; 2300} 2301 2302static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma) 2303{ 2304 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext); 2305 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device); 2306 unsigned long command; 2307 phys_addr_t pfn; 2308 2309 command = get_command(vma->vm_pgoff); 2310 switch (command) { 2311 case MLX5_IB_MMAP_WC_PAGE: 2312 case MLX5_IB_MMAP_ALLOC_WC: 2313 if (!dev->wc_support) 2314 return -EPERM; 2315 fallthrough; 2316 case MLX5_IB_MMAP_NC_PAGE: 2317 case MLX5_IB_MMAP_REGULAR_PAGE: 2318 return uar_mmap(dev, command, vma, context); 2319 2320 case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES: 2321 return -ENOSYS; 2322 2323 case MLX5_IB_MMAP_CORE_CLOCK: 2324 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 2325 return -EINVAL; 2326 2327 if (vma->vm_flags & VM_WRITE) 2328 return -EPERM; 2329 vm_flags_clear(vma, VM_MAYWRITE); 2330 2331 /* Don't expose to user-space information it shouldn't have */ 2332 if (PAGE_SIZE > 4096) 2333 return -EOPNOTSUPP; 2334 2335 pfn = (dev->mdev->iseg_base + 2336 offsetof(struct mlx5_init_seg, internal_timer_h)) >> 2337 PAGE_SHIFT; 2338 return rdma_user_mmap_io(&context->ibucontext, vma, pfn, 2339 PAGE_SIZE, 2340 pgprot_noncached(vma->vm_page_prot), 2341 NULL); 2342 case MLX5_IB_MMAP_CLOCK_INFO: 2343 return mlx5_ib_mmap_clock_info_page(dev, vma, context); 2344 2345 default: 2346 return mlx5_ib_mmap_offset(dev, vma, ibcontext); 2347 } 2348 2349 return 0; 2350} 2351 2352static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) 2353{ 2354 struct mlx5_ib_pd *pd = to_mpd(ibpd); 2355 struct ib_device *ibdev = ibpd->device; 2356 struct mlx5_ib_alloc_pd_resp resp; 2357 int err; 2358 u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {}; 2359 u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {}; 2360 u16 uid = 0; 2361 struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context( 2362 udata, struct mlx5_ib_ucontext, ibucontext); 2363 2364 uid = context ? context->devx_uid : 0; 2365 MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD); 2366 MLX5_SET(alloc_pd_in, in, uid, uid); 2367 err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out); 2368 if (err) 2369 return err; 2370 2371 pd->pdn = MLX5_GET(alloc_pd_out, out, pd); 2372 pd->uid = uid; 2373 if (udata) { 2374 resp.pdn = pd->pdn; 2375 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) { 2376 mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid); 2377 return -EFAULT; 2378 } 2379 } 2380 2381 return 0; 2382} 2383 2384static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata) 2385{ 2386 struct mlx5_ib_dev *mdev = to_mdev(pd->device); 2387 struct mlx5_ib_pd *mpd = to_mpd(pd); 2388 2389 return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid); 2390} 2391 2392static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid) 2393{ 2394 struct mlx5_ib_dev *dev = to_mdev(ibqp->device); 2395 struct mlx5_ib_qp *mqp = to_mqp(ibqp); 2396 int err; 2397 u16 uid; 2398 2399 uid = ibqp->pd ? 2400 to_mpd(ibqp->pd)->uid : 0; 2401 2402 if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) { 2403 mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n"); 2404 return -EOPNOTSUPP; 2405 } 2406 2407 err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid); 2408 if (err) 2409 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n", 2410 ibqp->qp_num, gid->raw); 2411 2412 return err; 2413} 2414 2415static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid) 2416{ 2417 struct mlx5_ib_dev *dev = to_mdev(ibqp->device); 2418 int err; 2419 u16 uid; 2420 2421 uid = ibqp->pd ? 2422 to_mpd(ibqp->pd)->uid : 0; 2423 err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid); 2424 if (err) 2425 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n", 2426 ibqp->qp_num, gid->raw); 2427 2428 return err; 2429} 2430 2431static int init_node_data(struct mlx5_ib_dev *dev) 2432{ 2433 int err; 2434 2435 err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc); 2436 if (err) 2437 return err; 2438 2439 dev->mdev->rev_id = dev->mdev->pdev->revision; 2440 2441 return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid); 2442} 2443 2444static ssize_t fw_pages_show(struct device *device, 2445 struct device_attribute *attr, char *buf) 2446{ 2447 struct mlx5_ib_dev *dev = 2448 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev); 2449 2450 return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages); 2451} 2452static DEVICE_ATTR_RO(fw_pages); 2453 2454static ssize_t reg_pages_show(struct device *device, 2455 struct device_attribute *attr, char *buf) 2456{ 2457 struct mlx5_ib_dev *dev = 2458 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev); 2459 2460 return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages)); 2461} 2462static DEVICE_ATTR_RO(reg_pages); 2463 2464static ssize_t hca_type_show(struct device *device, 2465 struct device_attribute *attr, char *buf) 2466{ 2467 struct mlx5_ib_dev *dev = 2468 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev); 2469 2470 return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device); 2471} 2472static DEVICE_ATTR_RO(hca_type); 2473 2474static ssize_t hw_rev_show(struct device *device, 2475 struct device_attribute *attr, char *buf) 2476{ 2477 struct mlx5_ib_dev *dev = 2478 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev); 2479 2480 return sysfs_emit(buf, "%x\n", dev->mdev->rev_id); 2481} 2482static DEVICE_ATTR_RO(hw_rev); 2483 2484static ssize_t board_id_show(struct device *device, 2485 struct device_attribute *attr, char *buf) 2486{ 2487 struct mlx5_ib_dev *dev = 2488 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev); 2489 2490 return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN, 2491 dev->mdev->board_id); 2492} 2493static DEVICE_ATTR_RO(board_id); 2494 2495static struct attribute *mlx5_class_attributes[] = { 2496 &dev_attr_hw_rev.attr, 2497 &dev_attr_hca_type.attr, 2498 &dev_attr_board_id.attr, 2499 &dev_attr_fw_pages.attr, 2500 &dev_attr_reg_pages.attr, 2501 NULL, 2502}; 2503 2504static const struct attribute_group mlx5_attr_group = { 2505 .attrs = mlx5_class_attributes, 2506}; 2507 2508static void pkey_change_handler(struct work_struct *work) 2509{ 2510 struct mlx5_ib_port_resources *ports = 2511 container_of(work, struct mlx5_ib_port_resources, 2512 pkey_change_work); 2513 2514 if (!ports->gsi) 2515 /* 2516 * We got this event before device was fully configured 2517 * and MAD registration code wasn't called/finished yet. 2518 */ 2519 return; 2520 2521 mlx5_ib_gsi_pkey_change(ports->gsi); 2522} 2523 2524static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev) 2525{ 2526 struct mlx5_ib_qp *mqp; 2527 struct mlx5_ib_cq *send_mcq, *recv_mcq; 2528 struct mlx5_core_cq *mcq; 2529 struct list_head cq_armed_list; 2530 unsigned long flags_qp; 2531 unsigned long flags_cq; 2532 unsigned long flags; 2533 2534 INIT_LIST_HEAD(&cq_armed_list); 2535 2536 /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/ 2537 spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags); 2538 list_for_each_entry(mqp, &ibdev->qp_list, qps_list) { 2539 spin_lock_irqsave(&mqp->sq.lock, flags_qp); 2540 if (mqp->sq.tail != mqp->sq.head) { 2541 send_mcq = to_mcq(mqp->ibqp.send_cq); 2542 spin_lock_irqsave(&send_mcq->lock, flags_cq); 2543 if (send_mcq->mcq.comp && 2544 mqp->ibqp.send_cq->comp_handler) { 2545 if (!send_mcq->mcq.reset_notify_added) { 2546 send_mcq->mcq.reset_notify_added = 1; 2547 list_add_tail(&send_mcq->mcq.reset_notify, 2548 &cq_armed_list); 2549 } 2550 } 2551 spin_unlock_irqrestore(&send_mcq->lock, flags_cq); 2552 } 2553 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp); 2554 spin_lock_irqsave(&mqp->rq.lock, flags_qp); 2555 /* no handling is needed for SRQ */ 2556 if (!mqp->ibqp.srq) { 2557 if (mqp->rq.tail != mqp->rq.head) { 2558 recv_mcq = to_mcq(mqp->ibqp.recv_cq); 2559 spin_lock_irqsave(&recv_mcq->lock, flags_cq); 2560 if (recv_mcq->mcq.comp && 2561 mqp->ibqp.recv_cq->comp_handler) { 2562 if (!recv_mcq->mcq.reset_notify_added) { 2563 recv_mcq->mcq.reset_notify_added = 1; 2564 list_add_tail(&recv_mcq->mcq.reset_notify, 2565 &cq_armed_list); 2566 } 2567 } 2568 spin_unlock_irqrestore(&recv_mcq->lock, 2569 flags_cq); 2570 } 2571 } 2572 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp); 2573 } 2574 /*At that point all inflight post send were put to be executed as of we 2575 * lock/unlock above locks Now need to arm all involved CQs. 2576 */ 2577 list_for_each_entry(mcq, &cq_armed_list, reset_notify) { 2578 mcq->comp(mcq, NULL); 2579 } 2580 spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags); 2581} 2582 2583static void delay_drop_handler(struct work_struct *work) 2584{ 2585 int err; 2586 struct mlx5_ib_delay_drop *delay_drop = 2587 container_of(work, struct mlx5_ib_delay_drop, 2588 delay_drop_work); 2589 2590 atomic_inc(&delay_drop->events_cnt); 2591 2592 mutex_lock(&delay_drop->lock); 2593 err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout); 2594 if (err) { 2595 mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n", 2596 delay_drop->timeout); 2597 delay_drop->activate = false; 2598 } 2599 mutex_unlock(&delay_drop->lock); 2600} 2601 2602static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe, 2603 struct ib_event *ibev) 2604{ 2605 u32 port = (eqe->data.port.port >> 4) & 0xf; 2606 2607 switch (eqe->sub_type) { 2608 case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT: 2609 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) == 2610 IB_LINK_LAYER_ETHERNET) 2611 schedule_work(&ibdev->delay_drop.delay_drop_work); 2612 break; 2613 default: /* do nothing */ 2614 return; 2615 } 2616} 2617 2618static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe, 2619 struct ib_event *ibev) 2620{ 2621 u32 port = (eqe->data.port.port >> 4) & 0xf; 2622 2623 ibev->element.port_num = port; 2624 2625 switch (eqe->sub_type) { 2626 case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE: 2627 case MLX5_PORT_CHANGE_SUBTYPE_DOWN: 2628 case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED: 2629 /* In RoCE, port up/down events are handled in 2630 * mlx5_netdev_event(). 2631 */ 2632 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) == 2633 IB_LINK_LAYER_ETHERNET) 2634 return -EINVAL; 2635 2636 ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ? 2637 IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR; 2638 break; 2639 2640 case MLX5_PORT_CHANGE_SUBTYPE_LID: 2641 ibev->event = IB_EVENT_LID_CHANGE; 2642 break; 2643 2644 case MLX5_PORT_CHANGE_SUBTYPE_PKEY: 2645 ibev->event = IB_EVENT_PKEY_CHANGE; 2646 schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work); 2647 break; 2648 2649 case MLX5_PORT_CHANGE_SUBTYPE_GUID: 2650 ibev->event = IB_EVENT_GID_CHANGE; 2651 break; 2652 2653 case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG: 2654 ibev->event = IB_EVENT_CLIENT_REREGISTER; 2655 break; 2656 default: 2657 return -EINVAL; 2658 } 2659 2660 return 0; 2661} 2662 2663static void mlx5_ib_handle_event(struct work_struct *_work) 2664{ 2665 struct mlx5_ib_event_work *work = 2666 container_of(_work, struct mlx5_ib_event_work, work); 2667 struct mlx5_ib_dev *ibdev; 2668 struct ib_event ibev; 2669 bool fatal = false; 2670 2671 if (work->is_slave) { 2672 ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi); 2673 if (!ibdev) 2674 goto out; 2675 } else { 2676 ibdev = work->dev; 2677 } 2678 2679 switch (work->event) { 2680 case MLX5_DEV_EVENT_SYS_ERROR: 2681 ibev.event = IB_EVENT_DEVICE_FATAL; 2682 mlx5_ib_handle_internal_error(ibdev); 2683 ibev.element.port_num = (u8)(unsigned long)work->param; 2684 fatal = true; 2685 break; 2686 case MLX5_EVENT_TYPE_PORT_CHANGE: 2687 if (handle_port_change(ibdev, work->param, &ibev)) 2688 goto out; 2689 break; 2690 case MLX5_EVENT_TYPE_GENERAL_EVENT: 2691 handle_general_event(ibdev, work->param, &ibev); 2692 fallthrough; 2693 default: 2694 goto out; 2695 } 2696 2697 ibev.device = &ibdev->ib_dev; 2698 2699 if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) { 2700 mlx5_ib_warn(ibdev, "warning: event on port %d\n", ibev.element.port_num); 2701 goto out; 2702 } 2703 2704 if (ibdev->ib_active) 2705 ib_dispatch_event(&ibev); 2706 2707 if (fatal) 2708 ibdev->ib_active = false; 2709out: 2710 kfree(work); 2711} 2712 2713static int mlx5_ib_event(struct notifier_block *nb, 2714 unsigned long event, void *param) 2715{ 2716 struct mlx5_ib_event_work *work; 2717 2718 work = kmalloc(sizeof(*work), GFP_ATOMIC); 2719 if (!work) 2720 return NOTIFY_DONE; 2721 2722 INIT_WORK(&work->work, mlx5_ib_handle_event); 2723 work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events); 2724 work->is_slave = false; 2725 work->param = param; 2726 work->event = event; 2727 2728 queue_work(mlx5_ib_event_wq, &work->work); 2729 2730 return NOTIFY_OK; 2731} 2732 2733static int mlx5_ib_event_slave_port(struct notifier_block *nb, 2734 unsigned long event, void *param) 2735{ 2736 struct mlx5_ib_event_work *work; 2737 2738 work = kmalloc(sizeof(*work), GFP_ATOMIC); 2739 if (!work) 2740 return NOTIFY_DONE; 2741 2742 INIT_WORK(&work->work, mlx5_ib_handle_event); 2743 work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events); 2744 work->is_slave = true; 2745 work->param = param; 2746 work->event = event; 2747 queue_work(mlx5_ib_event_wq, &work->work); 2748 2749 return NOTIFY_OK; 2750} 2751 2752static int set_has_smi_cap(struct mlx5_ib_dev *dev) 2753{ 2754 struct mlx5_hca_vport_context vport_ctx; 2755 int err; 2756 int port; 2757 2758 if (MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_IB) 2759 return 0; 2760 2761 for (port = 1; port <= dev->num_ports; port++) { 2762 if (!MLX5_CAP_GEN(dev->mdev, ib_virt)) { 2763 dev->port_caps[port - 1].has_smi = true; 2764 continue; 2765 } 2766 err = mlx5_query_hca_vport_context(dev->mdev, 0, port, 0, 2767 &vport_ctx); 2768 if (err) { 2769 mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n", 2770 port, err); 2771 return err; 2772 } 2773 dev->port_caps[port - 1].has_smi = vport_ctx.has_smi; 2774 } 2775 2776 return 0; 2777} 2778 2779static void get_ext_port_caps(struct mlx5_ib_dev *dev) 2780{ 2781 unsigned int port; 2782 2783 rdma_for_each_port (&dev->ib_dev, port) 2784 mlx5_query_ext_port_caps(dev, port); 2785} 2786 2787static u8 mlx5_get_umr_fence(u8 umr_fence_cap) 2788{ 2789 switch (umr_fence_cap) { 2790 case MLX5_CAP_UMR_FENCE_NONE: 2791 return MLX5_FENCE_MODE_NONE; 2792 case MLX5_CAP_UMR_FENCE_SMALL: 2793 return MLX5_FENCE_MODE_INITIATOR_SMALL; 2794 default: 2795 return MLX5_FENCE_MODE_STRONG_ORDERING; 2796 } 2797} 2798 2799static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev) 2800{ 2801 struct mlx5_ib_resources *devr = &dev->devr; 2802 struct ib_srq_init_attr attr; 2803 struct ib_device *ibdev; 2804 struct ib_cq_init_attr cq_attr = {.cqe = 1}; 2805 int port; 2806 int ret = 0; 2807 2808 ibdev = &dev->ib_dev; 2809 2810 if (!MLX5_CAP_GEN(dev->mdev, xrc)) 2811 return -EOPNOTSUPP; 2812 2813 devr->p0 = ib_alloc_pd(ibdev, 0); 2814 if (IS_ERR(devr->p0)) 2815 return PTR_ERR(devr->p0); 2816 2817 devr->c0 = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr); 2818 if (IS_ERR(devr->c0)) { 2819 ret = PTR_ERR(devr->c0); 2820 goto error1; 2821 } 2822 2823 ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0); 2824 if (ret) 2825 goto error2; 2826 2827 ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0); 2828 if (ret) 2829 goto error3; 2830 2831 memset(&attr, 0, sizeof(attr)); 2832 attr.attr.max_sge = 1; 2833 attr.attr.max_wr = 1; 2834 attr.srq_type = IB_SRQT_XRC; 2835 attr.ext.cq = devr->c0; 2836 2837 devr->s0 = ib_create_srq(devr->p0, &attr); 2838 if (IS_ERR(devr->s0)) { 2839 ret = PTR_ERR(devr->s0); 2840 goto err_create; 2841 } 2842 2843 memset(&attr, 0, sizeof(attr)); 2844 attr.attr.max_sge = 1; 2845 attr.attr.max_wr = 1; 2846 attr.srq_type = IB_SRQT_BASIC; 2847 2848 devr->s1 = ib_create_srq(devr->p0, &attr); 2849 if (IS_ERR(devr->s1)) { 2850 ret = PTR_ERR(devr->s1); 2851 goto error6; 2852 } 2853 2854 for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) 2855 INIT_WORK(&devr->ports[port].pkey_change_work, 2856 pkey_change_handler); 2857 2858 return 0; 2859 2860error6: 2861 ib_destroy_srq(devr->s0); 2862err_create: 2863 mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0); 2864error3: 2865 mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0); 2866error2: 2867 ib_destroy_cq(devr->c0); 2868error1: 2869 ib_dealloc_pd(devr->p0); 2870 return ret; 2871} 2872 2873static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev) 2874{ 2875 struct mlx5_ib_resources *devr = &dev->devr; 2876 int port; 2877 2878 /* 2879 * Make sure no change P_Key work items are still executing. 2880 * 2881 * At this stage, the mlx5_ib_event should be unregistered 2882 * and it ensures that no new works are added. 2883 */ 2884 for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) 2885 cancel_work_sync(&devr->ports[port].pkey_change_work); 2886 2887 ib_destroy_srq(devr->s1); 2888 ib_destroy_srq(devr->s0); 2889 mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0); 2890 mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0); 2891 ib_destroy_cq(devr->c0); 2892 ib_dealloc_pd(devr->p0); 2893} 2894 2895static u32 get_core_cap_flags(struct ib_device *ibdev, 2896 struct mlx5_hca_vport_context *rep) 2897{ 2898 struct mlx5_ib_dev *dev = to_mdev(ibdev); 2899 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1); 2900 u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type); 2901 u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version); 2902 bool raw_support = !mlx5_core_mp_enabled(dev->mdev); 2903 u32 ret = 0; 2904 2905 if (rep->grh_required) 2906 ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED; 2907 2908 if (ll == IB_LINK_LAYER_INFINIBAND) 2909 return ret | RDMA_CORE_PORT_IBA_IB; 2910 2911 if (raw_support) 2912 ret |= RDMA_CORE_PORT_RAW_PACKET; 2913 2914 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP)) 2915 return ret; 2916 2917 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP)) 2918 return ret; 2919 2920 if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP) 2921 ret |= RDMA_CORE_PORT_IBA_ROCE; 2922 2923 if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP) 2924 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP; 2925 2926 return ret; 2927} 2928 2929static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num, 2930 struct ib_port_immutable *immutable) 2931{ 2932 struct ib_port_attr attr; 2933 struct mlx5_ib_dev *dev = to_mdev(ibdev); 2934 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num); 2935 struct mlx5_hca_vport_context rep = {0}; 2936 int err; 2937 2938 err = ib_query_port(ibdev, port_num, &attr); 2939 if (err) 2940 return err; 2941 2942 if (ll == IB_LINK_LAYER_INFINIBAND) { 2943 err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0, 2944 &rep); 2945 if (err) 2946 return err; 2947 } 2948 2949 immutable->pkey_tbl_len = attr.pkey_tbl_len; 2950 immutable->gid_tbl_len = attr.gid_tbl_len; 2951 immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep); 2952 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 2953 2954 return 0; 2955} 2956 2957static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num, 2958 struct ib_port_immutable *immutable) 2959{ 2960 struct ib_port_attr attr; 2961 int err; 2962 2963 immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET; 2964 2965 err = ib_query_port(ibdev, port_num, &attr); 2966 if (err) 2967 return err; 2968 2969 immutable->pkey_tbl_len = attr.pkey_tbl_len; 2970 immutable->gid_tbl_len = attr.gid_tbl_len; 2971 immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET; 2972 2973 return 0; 2974} 2975 2976static void get_dev_fw_str(struct ib_device *ibdev, char *str) 2977{ 2978 struct mlx5_ib_dev *dev = 2979 container_of(ibdev, struct mlx5_ib_dev, ib_dev); 2980 snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d", 2981 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev), 2982 fw_rev_sub(dev->mdev)); 2983} 2984 2985static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev) 2986{ 2987 struct mlx5_core_dev *mdev = dev->mdev; 2988 struct mlx5_flow_namespace *ns = mlx5_get_flow_namespace(mdev, 2989 MLX5_FLOW_NAMESPACE_LAG); 2990 struct mlx5_flow_table *ft; 2991 int err; 2992 2993 if (!ns || !mlx5_lag_is_active(mdev)) 2994 return 0; 2995 2996 err = mlx5_cmd_create_vport_lag(mdev); 2997 if (err) 2998 return err; 2999 3000 ft = mlx5_create_lag_demux_flow_table(ns, 0, 0); 3001 if (IS_ERR(ft)) { 3002 err = PTR_ERR(ft); 3003 goto err_destroy_vport_lag; 3004 } 3005 3006 dev->flow_db->lag_demux_ft = ft; 3007 dev->lag_ports = mlx5_lag_get_num_ports(mdev); 3008 dev->lag_active = true; 3009 return 0; 3010 3011err_destroy_vport_lag: 3012 mlx5_cmd_destroy_vport_lag(mdev); 3013 return err; 3014} 3015 3016static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev) 3017{ 3018 struct mlx5_core_dev *mdev = dev->mdev; 3019 3020 if (dev->lag_active) { 3021 dev->lag_active = false; 3022 3023 mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft); 3024 dev->flow_db->lag_demux_ft = NULL; 3025 3026 mlx5_cmd_destroy_vport_lag(mdev); 3027 } 3028} 3029 3030static void mlx5_netdev_notifier_register(struct mlx5_roce *roce, 3031 struct net_device *netdev) 3032{ 3033 int err; 3034 3035 if (roce->tracking_netdev) 3036 return; 3037 roce->tracking_netdev = netdev; 3038 roce->nb.notifier_call = mlx5_netdev_event; 3039 err = register_netdevice_notifier_dev_net(netdev, &roce->nb, &roce->nn); 3040 WARN_ON(err); 3041} 3042 3043static void mlx5_netdev_notifier_unregister(struct mlx5_roce *roce) 3044{ 3045 if (!roce->tracking_netdev) 3046 return; 3047 unregister_netdevice_notifier_dev_net(roce->tracking_netdev, &roce->nb, 3048 &roce->nn); 3049 roce->tracking_netdev = NULL; 3050} 3051 3052static int mlx5e_mdev_notifier_event(struct notifier_block *nb, 3053 unsigned long event, void *data) 3054{ 3055 struct mlx5_roce *roce = container_of(nb, struct mlx5_roce, mdev_nb); 3056 struct net_device *netdev = data; 3057 3058 switch (event) { 3059 case MLX5_DRIVER_EVENT_UPLINK_NETDEV: 3060 if (netdev) 3061 mlx5_netdev_notifier_register(roce, netdev); 3062 else 3063 mlx5_netdev_notifier_unregister(roce); 3064 break; 3065 default: 3066 return NOTIFY_DONE; 3067 } 3068 3069 return NOTIFY_OK; 3070} 3071 3072static void mlx5_mdev_netdev_track(struct mlx5_ib_dev *dev, u32 port_num) 3073{ 3074 struct mlx5_roce *roce = &dev->port[port_num].roce; 3075 3076 roce->mdev_nb.notifier_call = mlx5e_mdev_notifier_event; 3077 mlx5_blocking_notifier_register(dev->mdev, &roce->mdev_nb); 3078 mlx5_core_uplink_netdev_event_replay(dev->mdev); 3079} 3080 3081static void mlx5_mdev_netdev_untrack(struct mlx5_ib_dev *dev, u32 port_num) 3082{ 3083 struct mlx5_roce *roce = &dev->port[port_num].roce; 3084 3085 mlx5_blocking_notifier_unregister(dev->mdev, &roce->mdev_nb); 3086 mlx5_netdev_notifier_unregister(roce); 3087} 3088 3089static int mlx5_enable_eth(struct mlx5_ib_dev *dev) 3090{ 3091 int err; 3092 3093 if (!dev->is_rep && dev->profile != &raw_eth_profile) { 3094 err = mlx5_nic_vport_enable_roce(dev->mdev); 3095 if (err) 3096 return err; 3097 } 3098 3099 err = mlx5_eth_lag_init(dev); 3100 if (err) 3101 goto err_disable_roce; 3102 3103 return 0; 3104 3105err_disable_roce: 3106 if (!dev->is_rep && dev->profile != &raw_eth_profile) 3107 mlx5_nic_vport_disable_roce(dev->mdev); 3108 3109 return err; 3110} 3111 3112static void mlx5_disable_eth(struct mlx5_ib_dev *dev) 3113{ 3114 mlx5_eth_lag_cleanup(dev); 3115 if (!dev->is_rep && dev->profile != &raw_eth_profile) 3116 mlx5_nic_vport_disable_roce(dev->mdev); 3117} 3118 3119static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num, 3120 enum rdma_netdev_t type, 3121 struct rdma_netdev_alloc_params *params) 3122{ 3123 if (type != RDMA_NETDEV_IPOIB) 3124 return -EOPNOTSUPP; 3125 3126 return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params); 3127} 3128 3129static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf, 3130 size_t count, loff_t *pos) 3131{ 3132 struct mlx5_ib_delay_drop *delay_drop = filp->private_data; 3133 char lbuf[20]; 3134 int len; 3135 3136 len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout); 3137 return simple_read_from_buffer(buf, count, pos, lbuf, len); 3138} 3139 3140static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf, 3141 size_t count, loff_t *pos) 3142{ 3143 struct mlx5_ib_delay_drop *delay_drop = filp->private_data; 3144 u32 timeout; 3145 u32 var; 3146 3147 if (kstrtouint_from_user(buf, count, 0, &var)) 3148 return -EFAULT; 3149 3150 timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 3151 1000); 3152 if (timeout != var) 3153 mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n", 3154 timeout); 3155 3156 delay_drop->timeout = timeout; 3157 3158 return count; 3159} 3160 3161static const struct file_operations fops_delay_drop_timeout = { 3162 .owner = THIS_MODULE, 3163 .open = simple_open, 3164 .write = delay_drop_timeout_write, 3165 .read = delay_drop_timeout_read, 3166}; 3167 3168static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev, 3169 struct mlx5_ib_multiport_info *mpi) 3170{ 3171 u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1; 3172 struct mlx5_ib_port *port = &ibdev->port[port_num]; 3173 int comps; 3174 int err; 3175 int i; 3176 3177 lockdep_assert_held(&mlx5_ib_multiport_mutex); 3178 3179 mlx5_core_mp_event_replay(ibdev->mdev, 3180 MLX5_DRIVER_EVENT_AFFILIATION_REMOVED, 3181 NULL); 3182 mlx5_core_mp_event_replay(mpi->mdev, 3183 MLX5_DRIVER_EVENT_AFFILIATION_REMOVED, 3184 NULL); 3185 3186 mlx5_ib_cleanup_cong_debugfs(ibdev, port_num); 3187 3188 spin_lock(&port->mp.mpi_lock); 3189 if (!mpi->ibdev) { 3190 spin_unlock(&port->mp.mpi_lock); 3191 return; 3192 } 3193 3194 mpi->ibdev = NULL; 3195 3196 spin_unlock(&port->mp.mpi_lock); 3197 if (mpi->mdev_events.notifier_call) 3198 mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events); 3199 mpi->mdev_events.notifier_call = NULL; 3200 mlx5_mdev_netdev_untrack(ibdev, port_num); 3201 spin_lock(&port->mp.mpi_lock); 3202 3203 comps = mpi->mdev_refcnt; 3204 if (comps) { 3205 mpi->unaffiliate = true; 3206 init_completion(&mpi->unref_comp); 3207 spin_unlock(&port->mp.mpi_lock); 3208 3209 for (i = 0; i < comps; i++) 3210 wait_for_completion(&mpi->unref_comp); 3211 3212 spin_lock(&port->mp.mpi_lock); 3213 mpi->unaffiliate = false; 3214 } 3215 3216 port->mp.mpi = NULL; 3217 3218 spin_unlock(&port->mp.mpi_lock); 3219 3220 err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev); 3221 3222 mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1); 3223 /* Log an error, still needed to cleanup the pointers and add 3224 * it back to the list. 3225 */ 3226 if (err) 3227 mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n", 3228 port_num + 1); 3229 3230 ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN; 3231} 3232 3233static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev, 3234 struct mlx5_ib_multiport_info *mpi) 3235{ 3236 u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1; 3237 u64 key; 3238 int err; 3239 3240 lockdep_assert_held(&mlx5_ib_multiport_mutex); 3241 3242 spin_lock(&ibdev->port[port_num].mp.mpi_lock); 3243 if (ibdev->port[port_num].mp.mpi) { 3244 mlx5_ib_dbg(ibdev, "port %u already affiliated.\n", 3245 port_num + 1); 3246 spin_unlock(&ibdev->port[port_num].mp.mpi_lock); 3247 return false; 3248 } 3249 3250 ibdev->port[port_num].mp.mpi = mpi; 3251 mpi->ibdev = ibdev; 3252 mpi->mdev_events.notifier_call = NULL; 3253 spin_unlock(&ibdev->port[port_num].mp.mpi_lock); 3254 3255 err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev); 3256 if (err) 3257 goto unbind; 3258 3259 mlx5_mdev_netdev_track(ibdev, port_num); 3260 3261 mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port; 3262 mlx5_notifier_register(mpi->mdev, &mpi->mdev_events); 3263 3264 mlx5_ib_init_cong_debugfs(ibdev, port_num); 3265 3266 key = mpi->mdev->priv.adev_idx; 3267 mlx5_core_mp_event_replay(mpi->mdev, 3268 MLX5_DRIVER_EVENT_AFFILIATION_DONE, 3269 &key); 3270 mlx5_core_mp_event_replay(ibdev->mdev, 3271 MLX5_DRIVER_EVENT_AFFILIATION_DONE, 3272 &key); 3273 3274 return true; 3275 3276unbind: 3277 mlx5_ib_unbind_slave_port(ibdev, mpi); 3278 return false; 3279} 3280 3281static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev) 3282{ 3283 u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1; 3284 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 3285 port_num + 1); 3286 struct mlx5_ib_multiport_info *mpi; 3287 int err; 3288 u32 i; 3289 3290 if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET) 3291 return 0; 3292 3293 err = mlx5_query_nic_vport_system_image_guid(dev->mdev, 3294 &dev->sys_image_guid); 3295 if (err) 3296 return err; 3297 3298 err = mlx5_nic_vport_enable_roce(dev->mdev); 3299 if (err) 3300 return err; 3301 3302 mutex_lock(&mlx5_ib_multiport_mutex); 3303 for (i = 0; i < dev->num_ports; i++) { 3304 bool bound = false; 3305 3306 /* build a stub multiport info struct for the native port. */ 3307 if (i == port_num) { 3308 mpi = kzalloc(sizeof(*mpi), GFP_KERNEL); 3309 if (!mpi) { 3310 mutex_unlock(&mlx5_ib_multiport_mutex); 3311 mlx5_nic_vport_disable_roce(dev->mdev); 3312 return -ENOMEM; 3313 } 3314 3315 mpi->is_master = true; 3316 mpi->mdev = dev->mdev; 3317 mpi->sys_image_guid = dev->sys_image_guid; 3318 dev->port[i].mp.mpi = mpi; 3319 mpi->ibdev = dev; 3320 mpi = NULL; 3321 continue; 3322 } 3323 3324 list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list, 3325 list) { 3326 if (dev->sys_image_guid == mpi->sys_image_guid && 3327 (mlx5_core_native_port_num(mpi->mdev) - 1) == i) { 3328 bound = mlx5_ib_bind_slave_port(dev, mpi); 3329 } 3330 3331 if (bound) { 3332 dev_dbg(mpi->mdev->device, 3333 "removing port from unaffiliated list.\n"); 3334 mlx5_ib_dbg(dev, "port %d bound\n", i + 1); 3335 list_del(&mpi->list); 3336 break; 3337 } 3338 } 3339 if (!bound) 3340 mlx5_ib_dbg(dev, "no free port found for port %d\n", 3341 i + 1); 3342 } 3343 3344 list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list); 3345 mutex_unlock(&mlx5_ib_multiport_mutex); 3346 return err; 3347} 3348 3349static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev) 3350{ 3351 u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1; 3352 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 3353 port_num + 1); 3354 u32 i; 3355 3356 if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET) 3357 return; 3358 3359 mutex_lock(&mlx5_ib_multiport_mutex); 3360 for (i = 0; i < dev->num_ports; i++) { 3361 if (dev->port[i].mp.mpi) { 3362 /* Destroy the native port stub */ 3363 if (i == port_num) { 3364 kfree(dev->port[i].mp.mpi); 3365 dev->port[i].mp.mpi = NULL; 3366 } else { 3367 mlx5_ib_dbg(dev, "unbinding port_num: %u\n", 3368 i + 1); 3369 list_add_tail(&dev->port[i].mp.mpi->list, 3370 &mlx5_ib_unaffiliated_port_list); 3371 mlx5_ib_unbind_slave_port(dev, 3372 dev->port[i].mp.mpi); 3373 } 3374 } 3375 } 3376 3377 mlx5_ib_dbg(dev, "removing from devlist\n"); 3378 list_del(&dev->ib_dev_list); 3379 mutex_unlock(&mlx5_ib_multiport_mutex); 3380 3381 mlx5_nic_vport_disable_roce(dev->mdev); 3382} 3383 3384static int mmap_obj_cleanup(struct ib_uobject *uobject, 3385 enum rdma_remove_reason why, 3386 struct uverbs_attr_bundle *attrs) 3387{ 3388 struct mlx5_user_mmap_entry *obj = uobject->object; 3389 3390 rdma_user_mmap_entry_remove(&obj->rdma_entry); 3391 return 0; 3392} 3393 3394static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c, 3395 struct mlx5_user_mmap_entry *entry, 3396 size_t length) 3397{ 3398 return rdma_user_mmap_entry_insert_range( 3399 &c->ibucontext, &entry->rdma_entry, length, 3400 (MLX5_IB_MMAP_OFFSET_START << 16), 3401 ((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1)); 3402} 3403 3404static struct mlx5_user_mmap_entry * 3405alloc_var_entry(struct mlx5_ib_ucontext *c) 3406{ 3407 struct mlx5_user_mmap_entry *entry; 3408 struct mlx5_var_table *var_table; 3409 u32 page_idx; 3410 int err; 3411 3412 var_table = &to_mdev(c->ibucontext.device)->var_table; 3413 entry = kzalloc(sizeof(*entry), GFP_KERNEL); 3414 if (!entry) 3415 return ERR_PTR(-ENOMEM); 3416 3417 mutex_lock(&var_table->bitmap_lock); 3418 page_idx = find_first_zero_bit(var_table->bitmap, 3419 var_table->num_var_hw_entries); 3420 if (page_idx >= var_table->num_var_hw_entries) { 3421 err = -ENOSPC; 3422 mutex_unlock(&var_table->bitmap_lock); 3423 goto end; 3424 } 3425 3426 set_bit(page_idx, var_table->bitmap); 3427 mutex_unlock(&var_table->bitmap_lock); 3428 3429 entry->address = var_table->hw_start_addr + 3430 (page_idx * var_table->stride_size); 3431 entry->page_idx = page_idx; 3432 entry->mmap_flag = MLX5_IB_MMAP_TYPE_VAR; 3433 3434 err = mlx5_rdma_user_mmap_entry_insert(c, entry, 3435 var_table->stride_size); 3436 if (err) 3437 goto err_insert; 3438 3439 return entry; 3440 3441err_insert: 3442 mutex_lock(&var_table->bitmap_lock); 3443 clear_bit(page_idx, var_table->bitmap); 3444 mutex_unlock(&var_table->bitmap_lock); 3445end: 3446 kfree(entry); 3447 return ERR_PTR(err); 3448} 3449 3450static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)( 3451 struct uverbs_attr_bundle *attrs) 3452{ 3453 struct ib_uobject *uobj = uverbs_attr_get_uobject( 3454 attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE); 3455 struct mlx5_ib_ucontext *c; 3456 struct mlx5_user_mmap_entry *entry; 3457 u64 mmap_offset; 3458 u32 length; 3459 int err; 3460 3461 c = to_mucontext(ib_uverbs_get_ucontext(attrs)); 3462 if (IS_ERR(c)) 3463 return PTR_ERR(c); 3464 3465 entry = alloc_var_entry(c); 3466 if (IS_ERR(entry)) 3467 return PTR_ERR(entry); 3468 3469 mmap_offset = mlx5_entry_to_mmap_offset(entry); 3470 length = entry->rdma_entry.npages * PAGE_SIZE; 3471 uobj->object = entry; 3472 uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE); 3473 3474 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET, 3475 &mmap_offset, sizeof(mmap_offset)); 3476 if (err) 3477 return err; 3478 3479 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID, 3480 &entry->page_idx, sizeof(entry->page_idx)); 3481 if (err) 3482 return err; 3483 3484 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH, 3485 &length, sizeof(length)); 3486 return err; 3487} 3488 3489DECLARE_UVERBS_NAMED_METHOD( 3490 MLX5_IB_METHOD_VAR_OBJ_ALLOC, 3491 UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE, 3492 MLX5_IB_OBJECT_VAR, 3493 UVERBS_ACCESS_NEW, 3494 UA_MANDATORY), 3495 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID, 3496 UVERBS_ATTR_TYPE(u32), 3497 UA_MANDATORY), 3498 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH, 3499 UVERBS_ATTR_TYPE(u32), 3500 UA_MANDATORY), 3501 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET, 3502 UVERBS_ATTR_TYPE(u64), 3503 UA_MANDATORY)); 3504 3505DECLARE_UVERBS_NAMED_METHOD_DESTROY( 3506 MLX5_IB_METHOD_VAR_OBJ_DESTROY, 3507 UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE, 3508 MLX5_IB_OBJECT_VAR, 3509 UVERBS_ACCESS_DESTROY, 3510 UA_MANDATORY)); 3511 3512DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR, 3513 UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup), 3514 &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC), 3515 &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY)); 3516 3517static bool var_is_supported(struct ib_device *device) 3518{ 3519 struct mlx5_ib_dev *dev = to_mdev(device); 3520 3521 return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) & 3522 MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q); 3523} 3524 3525static struct mlx5_user_mmap_entry * 3526alloc_uar_entry(struct mlx5_ib_ucontext *c, 3527 enum mlx5_ib_uapi_uar_alloc_type alloc_type) 3528{ 3529 struct mlx5_user_mmap_entry *entry; 3530 struct mlx5_ib_dev *dev; 3531 u32 uar_index; 3532 int err; 3533 3534 entry = kzalloc(sizeof(*entry), GFP_KERNEL); 3535 if (!entry) 3536 return ERR_PTR(-ENOMEM); 3537 3538 dev = to_mdev(c->ibucontext.device); 3539 err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index, c->devx_uid); 3540 if (err) 3541 goto end; 3542 3543 entry->page_idx = uar_index; 3544 entry->address = uar_index2paddress(dev, uar_index); 3545 if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF) 3546 entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC; 3547 else 3548 entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC; 3549 3550 err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE); 3551 if (err) 3552 goto err_insert; 3553 3554 return entry; 3555 3556err_insert: 3557 mlx5_cmd_uar_dealloc(dev->mdev, uar_index, c->devx_uid); 3558end: 3559 kfree(entry); 3560 return ERR_PTR(err); 3561} 3562 3563static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)( 3564 struct uverbs_attr_bundle *attrs) 3565{ 3566 struct ib_uobject *uobj = uverbs_attr_get_uobject( 3567 attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE); 3568 enum mlx5_ib_uapi_uar_alloc_type alloc_type; 3569 struct mlx5_ib_ucontext *c; 3570 struct mlx5_user_mmap_entry *entry; 3571 u64 mmap_offset; 3572 u32 length; 3573 int err; 3574 3575 c = to_mucontext(ib_uverbs_get_ucontext(attrs)); 3576 if (IS_ERR(c)) 3577 return PTR_ERR(c); 3578 3579 err = uverbs_get_const(&alloc_type, attrs, 3580 MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE); 3581 if (err) 3582 return err; 3583 3584 if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF && 3585 alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC) 3586 return -EOPNOTSUPP; 3587 3588 if (!to_mdev(c->ibucontext.device)->wc_support && 3589 alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF) 3590 return -EOPNOTSUPP; 3591 3592 entry = alloc_uar_entry(c, alloc_type); 3593 if (IS_ERR(entry)) 3594 return PTR_ERR(entry); 3595 3596 mmap_offset = mlx5_entry_to_mmap_offset(entry); 3597 length = entry->rdma_entry.npages * PAGE_SIZE; 3598 uobj->object = entry; 3599 uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE); 3600 3601 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET, 3602 &mmap_offset, sizeof(mmap_offset)); 3603 if (err) 3604 return err; 3605 3606 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID, 3607 &entry->page_idx, sizeof(entry->page_idx)); 3608 if (err) 3609 return err; 3610 3611 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH, 3612 &length, sizeof(length)); 3613 return err; 3614} 3615 3616DECLARE_UVERBS_NAMED_METHOD( 3617 MLX5_IB_METHOD_UAR_OBJ_ALLOC, 3618 UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE, 3619 MLX5_IB_OBJECT_UAR, 3620 UVERBS_ACCESS_NEW, 3621 UA_MANDATORY), 3622 UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE, 3623 enum mlx5_ib_uapi_uar_alloc_type, 3624 UA_MANDATORY), 3625 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID, 3626 UVERBS_ATTR_TYPE(u32), 3627 UA_MANDATORY), 3628 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH, 3629 UVERBS_ATTR_TYPE(u32), 3630 UA_MANDATORY), 3631 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET, 3632 UVERBS_ATTR_TYPE(u64), 3633 UA_MANDATORY)); 3634 3635DECLARE_UVERBS_NAMED_METHOD_DESTROY( 3636 MLX5_IB_METHOD_UAR_OBJ_DESTROY, 3637 UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE, 3638 MLX5_IB_OBJECT_UAR, 3639 UVERBS_ACCESS_DESTROY, 3640 UA_MANDATORY)); 3641 3642DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR, 3643 UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup), 3644 &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC), 3645 &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY)); 3646 3647ADD_UVERBS_ATTRIBUTES_SIMPLE( 3648 mlx5_ib_query_context, 3649 UVERBS_OBJECT_DEVICE, 3650 UVERBS_METHOD_QUERY_CONTEXT, 3651 UVERBS_ATTR_PTR_OUT( 3652 MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX, 3653 UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp, 3654 dump_fill_mkey), 3655 UA_MANDATORY)); 3656 3657static const struct uapi_definition mlx5_ib_defs[] = { 3658 UAPI_DEF_CHAIN(mlx5_ib_devx_defs), 3659 UAPI_DEF_CHAIN(mlx5_ib_flow_defs), 3660 UAPI_DEF_CHAIN(mlx5_ib_qos_defs), 3661 UAPI_DEF_CHAIN(mlx5_ib_std_types_defs), 3662 UAPI_DEF_CHAIN(mlx5_ib_dm_defs), 3663 3664 UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context), 3665 UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR, 3666 UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)), 3667 UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR), 3668 {} 3669}; 3670 3671static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev) 3672{ 3673 mlx5_ib_cleanup_multiport_master(dev); 3674 WARN_ON(!xa_empty(&dev->odp_mkeys)); 3675 mutex_destroy(&dev->cap_mask_mutex); 3676 WARN_ON(!xa_empty(&dev->sig_mrs)); 3677 WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES)); 3678 mlx5r_macsec_dealloc_gids(dev); 3679} 3680 3681static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev) 3682{ 3683 struct mlx5_core_dev *mdev = dev->mdev; 3684 int err, i; 3685 3686 dev->ib_dev.node_type = RDMA_NODE_IB_CA; 3687 dev->ib_dev.local_dma_lkey = 0 /* not supported for now */; 3688 dev->ib_dev.phys_port_cnt = dev->num_ports; 3689 dev->ib_dev.dev.parent = mdev->device; 3690 dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES; 3691 3692 for (i = 0; i < dev->num_ports; i++) { 3693 spin_lock_init(&dev->port[i].mp.mpi_lock); 3694 rwlock_init(&dev->port[i].roce.netdev_lock); 3695 dev->port[i].roce.dev = dev; 3696 dev->port[i].roce.native_port_num = i + 1; 3697 dev->port[i].roce.last_port_state = IB_PORT_DOWN; 3698 } 3699 3700 err = mlx5r_cmd_query_special_mkeys(dev); 3701 if (err) 3702 return err; 3703 3704 err = mlx5r_macsec_init_gids_and_devlist(dev); 3705 if (err) 3706 return err; 3707 3708 err = mlx5_ib_init_multiport_master(dev); 3709 if (err) 3710 goto err; 3711 3712 err = set_has_smi_cap(dev); 3713 if (err) 3714 goto err_mp; 3715 3716 err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len); 3717 if (err) 3718 goto err_mp; 3719 3720 if (mlx5_use_mad_ifc(dev)) 3721 get_ext_port_caps(dev); 3722 3723 dev->ib_dev.num_comp_vectors = mlx5_comp_vectors_max(mdev); 3724 3725 mutex_init(&dev->cap_mask_mutex); 3726 INIT_LIST_HEAD(&dev->qp_list); 3727 spin_lock_init(&dev->reset_flow_resource_lock); 3728 xa_init(&dev->odp_mkeys); 3729 xa_init(&dev->sig_mrs); 3730 atomic_set(&dev->mkey_var, 0); 3731 3732 spin_lock_init(&dev->dm.lock); 3733 dev->dm.dev = mdev; 3734 return 0; 3735err: 3736 mlx5r_macsec_dealloc_gids(dev); 3737err_mp: 3738 mlx5_ib_cleanup_multiport_master(dev); 3739 return err; 3740} 3741 3742static int mlx5_ib_enable_driver(struct ib_device *dev) 3743{ 3744 struct mlx5_ib_dev *mdev = to_mdev(dev); 3745 int ret; 3746 3747 ret = mlx5_ib_test_wc(mdev); 3748 mlx5_ib_dbg(mdev, "Write-Combining %s", 3749 mdev->wc_support ? "supported" : "not supported"); 3750 3751 return ret; 3752} 3753 3754static const struct ib_device_ops mlx5_ib_dev_ops = { 3755 .owner = THIS_MODULE, 3756 .driver_id = RDMA_DRIVER_MLX5, 3757 .uverbs_abi_ver = MLX5_IB_UVERBS_ABI_VERSION, 3758 3759 .add_gid = mlx5_ib_add_gid, 3760 .alloc_mr = mlx5_ib_alloc_mr, 3761 .alloc_mr_integrity = mlx5_ib_alloc_mr_integrity, 3762 .alloc_pd = mlx5_ib_alloc_pd, 3763 .alloc_ucontext = mlx5_ib_alloc_ucontext, 3764 .attach_mcast = mlx5_ib_mcg_attach, 3765 .check_mr_status = mlx5_ib_check_mr_status, 3766 .create_ah = mlx5_ib_create_ah, 3767 .create_cq = mlx5_ib_create_cq, 3768 .create_qp = mlx5_ib_create_qp, 3769 .create_srq = mlx5_ib_create_srq, 3770 .create_user_ah = mlx5_ib_create_ah, 3771 .dealloc_pd = mlx5_ib_dealloc_pd, 3772 .dealloc_ucontext = mlx5_ib_dealloc_ucontext, 3773 .del_gid = mlx5_ib_del_gid, 3774 .dereg_mr = mlx5_ib_dereg_mr, 3775 .destroy_ah = mlx5_ib_destroy_ah, 3776 .destroy_cq = mlx5_ib_destroy_cq, 3777 .destroy_qp = mlx5_ib_destroy_qp, 3778 .destroy_srq = mlx5_ib_destroy_srq, 3779 .detach_mcast = mlx5_ib_mcg_detach, 3780 .disassociate_ucontext = mlx5_ib_disassociate_ucontext, 3781 .drain_rq = mlx5_ib_drain_rq, 3782 .drain_sq = mlx5_ib_drain_sq, 3783 .device_group = &mlx5_attr_group, 3784 .enable_driver = mlx5_ib_enable_driver, 3785 .get_dev_fw_str = get_dev_fw_str, 3786 .get_dma_mr = mlx5_ib_get_dma_mr, 3787 .get_link_layer = mlx5_ib_port_link_layer, 3788 .map_mr_sg = mlx5_ib_map_mr_sg, 3789 .map_mr_sg_pi = mlx5_ib_map_mr_sg_pi, 3790 .mmap = mlx5_ib_mmap, 3791 .mmap_free = mlx5_ib_mmap_free, 3792 .modify_cq = mlx5_ib_modify_cq, 3793 .modify_device = mlx5_ib_modify_device, 3794 .modify_port = mlx5_ib_modify_port, 3795 .modify_qp = mlx5_ib_modify_qp, 3796 .modify_srq = mlx5_ib_modify_srq, 3797 .poll_cq = mlx5_ib_poll_cq, 3798 .post_recv = mlx5_ib_post_recv_nodrain, 3799 .post_send = mlx5_ib_post_send_nodrain, 3800 .post_srq_recv = mlx5_ib_post_srq_recv, 3801 .process_mad = mlx5_ib_process_mad, 3802 .query_ah = mlx5_ib_query_ah, 3803 .query_device = mlx5_ib_query_device, 3804 .query_gid = mlx5_ib_query_gid, 3805 .query_pkey = mlx5_ib_query_pkey, 3806 .query_qp = mlx5_ib_query_qp, 3807 .query_srq = mlx5_ib_query_srq, 3808 .query_ucontext = mlx5_ib_query_ucontext, 3809 .reg_user_mr = mlx5_ib_reg_user_mr, 3810 .reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf, 3811 .req_notify_cq = mlx5_ib_arm_cq, 3812 .rereg_user_mr = mlx5_ib_rereg_user_mr, 3813 .resize_cq = mlx5_ib_resize_cq, 3814 3815 INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah), 3816 INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs), 3817 INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq), 3818 INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd), 3819 INIT_RDMA_OBJ_SIZE(ib_qp, mlx5_ib_qp, ibqp), 3820 INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq), 3821 INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext), 3822}; 3823 3824static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = { 3825 .rdma_netdev_get_params = mlx5_ib_rn_get_params, 3826}; 3827 3828static const struct ib_device_ops mlx5_ib_dev_sriov_ops = { 3829 .get_vf_config = mlx5_ib_get_vf_config, 3830 .get_vf_guid = mlx5_ib_get_vf_guid, 3831 .get_vf_stats = mlx5_ib_get_vf_stats, 3832 .set_vf_guid = mlx5_ib_set_vf_guid, 3833 .set_vf_link_state = mlx5_ib_set_vf_link_state, 3834}; 3835 3836static const struct ib_device_ops mlx5_ib_dev_mw_ops = { 3837 .alloc_mw = mlx5_ib_alloc_mw, 3838 .dealloc_mw = mlx5_ib_dealloc_mw, 3839 3840 INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw), 3841}; 3842 3843static const struct ib_device_ops mlx5_ib_dev_xrc_ops = { 3844 .alloc_xrcd = mlx5_ib_alloc_xrcd, 3845 .dealloc_xrcd = mlx5_ib_dealloc_xrcd, 3846 3847 INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd), 3848}; 3849 3850static int mlx5_ib_init_var_table(struct mlx5_ib_dev *dev) 3851{ 3852 struct mlx5_core_dev *mdev = dev->mdev; 3853 struct mlx5_var_table *var_table = &dev->var_table; 3854 u8 log_doorbell_bar_size; 3855 u8 log_doorbell_stride; 3856 u64 bar_size; 3857 3858 log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev, 3859 log_doorbell_bar_size); 3860 log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev, 3861 log_doorbell_stride); 3862 var_table->hw_start_addr = dev->mdev->bar_addr + 3863 MLX5_CAP64_DEV_VDPA_EMULATION(mdev, 3864 doorbell_bar_offset); 3865 bar_size = (1ULL << log_doorbell_bar_size) * 4096; 3866 var_table->stride_size = 1ULL << log_doorbell_stride; 3867 var_table->num_var_hw_entries = div_u64(bar_size, 3868 var_table->stride_size); 3869 mutex_init(&var_table->bitmap_lock); 3870 var_table->bitmap = bitmap_zalloc(var_table->num_var_hw_entries, 3871 GFP_KERNEL); 3872 return (var_table->bitmap) ? 0 : -ENOMEM; 3873} 3874 3875static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev) 3876{ 3877 bitmap_free(dev->var_table.bitmap); 3878} 3879 3880static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev) 3881{ 3882 struct mlx5_core_dev *mdev = dev->mdev; 3883 int err; 3884 3885 if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) && 3886 IS_ENABLED(CONFIG_MLX5_CORE_IPOIB)) 3887 ib_set_device_ops(&dev->ib_dev, 3888 &mlx5_ib_dev_ipoib_enhanced_ops); 3889 3890 if (mlx5_core_is_pf(mdev)) 3891 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops); 3892 3893 dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence)); 3894 3895 if (MLX5_CAP_GEN(mdev, imaicl)) 3896 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops); 3897 3898 if (MLX5_CAP_GEN(mdev, xrc)) 3899 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops); 3900 3901 if (MLX5_CAP_DEV_MEM(mdev, memic) || 3902 MLX5_CAP_GEN_64(dev->mdev, general_obj_types) & 3903 MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM) 3904 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops); 3905 3906 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops); 3907 3908 if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS)) 3909 dev->ib_dev.driver_def = mlx5_ib_defs; 3910 3911 err = init_node_data(dev); 3912 if (err) 3913 return err; 3914 3915 if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) && 3916 (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) || 3917 MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc))) 3918 mutex_init(&dev->lb.mutex); 3919 3920 if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) & 3921 MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) { 3922 err = mlx5_ib_init_var_table(dev); 3923 if (err) 3924 return err; 3925 } 3926 3927 dev->ib_dev.use_cq_dim = true; 3928 3929 return 0; 3930} 3931 3932static const struct ib_device_ops mlx5_ib_dev_port_ops = { 3933 .get_port_immutable = mlx5_port_immutable, 3934 .query_port = mlx5_ib_query_port, 3935}; 3936 3937static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev) 3938{ 3939 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops); 3940 return 0; 3941} 3942 3943static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = { 3944 .get_port_immutable = mlx5_port_rep_immutable, 3945 .query_port = mlx5_ib_rep_query_port, 3946 .query_pkey = mlx5_ib_rep_query_pkey, 3947}; 3948 3949static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev) 3950{ 3951 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops); 3952 return 0; 3953} 3954 3955static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = { 3956 .create_rwq_ind_table = mlx5_ib_create_rwq_ind_table, 3957 .create_wq = mlx5_ib_create_wq, 3958 .destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table, 3959 .destroy_wq = mlx5_ib_destroy_wq, 3960 .get_netdev = mlx5_ib_get_netdev, 3961 .modify_wq = mlx5_ib_modify_wq, 3962 3963 INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table, 3964 ib_rwq_ind_tbl), 3965}; 3966 3967static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev) 3968{ 3969 struct mlx5_core_dev *mdev = dev->mdev; 3970 enum rdma_link_layer ll; 3971 int port_type_cap; 3972 u32 port_num = 0; 3973 int err; 3974 3975 port_type_cap = MLX5_CAP_GEN(mdev, port_type); 3976 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap); 3977 3978 if (ll == IB_LINK_LAYER_ETHERNET) { 3979 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops); 3980 3981 port_num = mlx5_core_native_port_num(dev->mdev) - 1; 3982 3983 /* Register only for native ports */ 3984 mlx5_mdev_netdev_track(dev, port_num); 3985 3986 err = mlx5_enable_eth(dev); 3987 if (err) 3988 goto cleanup; 3989 } 3990 3991 return 0; 3992cleanup: 3993 mlx5_mdev_netdev_untrack(dev, port_num); 3994 return err; 3995} 3996 3997static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev) 3998{ 3999 struct mlx5_core_dev *mdev = dev->mdev; 4000 enum rdma_link_layer ll; 4001 int port_type_cap; 4002 u32 port_num; 4003 4004 port_type_cap = MLX5_CAP_GEN(mdev, port_type); 4005 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap); 4006 4007 if (ll == IB_LINK_LAYER_ETHERNET) { 4008 mlx5_disable_eth(dev); 4009 4010 port_num = mlx5_core_native_port_num(dev->mdev) - 1; 4011 mlx5_mdev_netdev_untrack(dev, port_num); 4012 } 4013} 4014 4015static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev) 4016{ 4017 mlx5_ib_init_cong_debugfs(dev, 4018 mlx5_core_native_port_num(dev->mdev) - 1); 4019 return 0; 4020} 4021 4022static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev) 4023{ 4024 mlx5_ib_cleanup_cong_debugfs(dev, 4025 mlx5_core_native_port_num(dev->mdev) - 1); 4026} 4027 4028static int mlx5_ib_stage_uar_init(struct mlx5_ib_dev *dev) 4029{ 4030 dev->mdev->priv.uar = mlx5_get_uars_page(dev->mdev); 4031 return PTR_ERR_OR_ZERO(dev->mdev->priv.uar); 4032} 4033 4034static void mlx5_ib_stage_uar_cleanup(struct mlx5_ib_dev *dev) 4035{ 4036 mlx5_put_uars_page(dev->mdev, dev->mdev->priv.uar); 4037} 4038 4039static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev) 4040{ 4041 int err; 4042 4043 err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false); 4044 if (err) 4045 return err; 4046 4047 err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true); 4048 if (err) 4049 mlx5_free_bfreg(dev->mdev, &dev->bfreg); 4050 4051 return err; 4052} 4053 4054static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev) 4055{ 4056 mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg); 4057 mlx5_free_bfreg(dev->mdev, &dev->bfreg); 4058} 4059 4060static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev) 4061{ 4062 const char *name; 4063 4064 if (!mlx5_lag_is_active(dev->mdev)) 4065 name = "mlx5_%d"; 4066 else 4067 name = "mlx5_bond_%d"; 4068 return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev); 4069} 4070 4071static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev) 4072{ 4073 mlx5_mkey_cache_cleanup(dev); 4074 mlx5r_umr_resource_cleanup(dev); 4075} 4076 4077static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev) 4078{ 4079 ib_unregister_device(&dev->ib_dev); 4080} 4081 4082static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev) 4083{ 4084 int ret; 4085 4086 ret = mlx5r_umr_resource_init(dev); 4087 if (ret) 4088 return ret; 4089 4090 ret = mlx5_mkey_cache_init(dev); 4091 if (ret) 4092 mlx5_ib_warn(dev, "mr cache init failed %d\n", ret); 4093 return ret; 4094} 4095 4096static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev) 4097{ 4098 struct dentry *root; 4099 4100 if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP)) 4101 return 0; 4102 4103 mutex_init(&dev->delay_drop.lock); 4104 dev->delay_drop.dev = dev; 4105 dev->delay_drop.activate = false; 4106 dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000; 4107 INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler); 4108 atomic_set(&dev->delay_drop.rqs_cnt, 0); 4109 atomic_set(&dev->delay_drop.events_cnt, 0); 4110 4111 if (!mlx5_debugfs_root) 4112 return 0; 4113 4114 root = debugfs_create_dir("delay_drop", mlx5_debugfs_get_dev_root(dev->mdev)); 4115 dev->delay_drop.dir_debugfs = root; 4116 4117 debugfs_create_atomic_t("num_timeout_events", 0400, root, 4118 &dev->delay_drop.events_cnt); 4119 debugfs_create_atomic_t("num_rqs", 0400, root, 4120 &dev->delay_drop.rqs_cnt); 4121 debugfs_create_file("timeout", 0600, root, &dev->delay_drop, 4122 &fops_delay_drop_timeout); 4123 return 0; 4124} 4125 4126static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev) 4127{ 4128 if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP)) 4129 return; 4130 4131 cancel_work_sync(&dev->delay_drop.delay_drop_work); 4132 if (!dev->delay_drop.dir_debugfs) 4133 return; 4134 4135 debugfs_remove_recursive(dev->delay_drop.dir_debugfs); 4136 dev->delay_drop.dir_debugfs = NULL; 4137} 4138 4139static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev) 4140{ 4141 dev->mdev_events.notifier_call = mlx5_ib_event; 4142 mlx5_notifier_register(dev->mdev, &dev->mdev_events); 4143 4144 mlx5r_macsec_event_register(dev); 4145 4146 return 0; 4147} 4148 4149static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev) 4150{ 4151 mlx5r_macsec_event_unregister(dev); 4152 mlx5_notifier_unregister(dev->mdev, &dev->mdev_events); 4153} 4154 4155void __mlx5_ib_remove(struct mlx5_ib_dev *dev, 4156 const struct mlx5_ib_profile *profile, 4157 int stage) 4158{ 4159 dev->ib_active = false; 4160 4161 /* Number of stages to cleanup */ 4162 while (stage) { 4163 stage--; 4164 if (profile->stage[stage].cleanup) 4165 profile->stage[stage].cleanup(dev); 4166 } 4167 4168 kfree(dev->port); 4169 ib_dealloc_device(&dev->ib_dev); 4170} 4171 4172int __mlx5_ib_add(struct mlx5_ib_dev *dev, 4173 const struct mlx5_ib_profile *profile) 4174{ 4175 int err; 4176 int i; 4177 4178 dev->profile = profile; 4179 4180 for (i = 0; i < MLX5_IB_STAGE_MAX; i++) { 4181 if (profile->stage[i].init) { 4182 err = profile->stage[i].init(dev); 4183 if (err) 4184 goto err_out; 4185 } 4186 } 4187 4188 dev->ib_active = true; 4189 return 0; 4190 4191err_out: 4192 /* Clean up stages which were initialized */ 4193 while (i) { 4194 i--; 4195 if (profile->stage[i].cleanup) 4196 profile->stage[i].cleanup(dev); 4197 } 4198 return -ENOMEM; 4199} 4200 4201static const struct mlx5_ib_profile pf_profile = { 4202 STAGE_CREATE(MLX5_IB_STAGE_INIT, 4203 mlx5_ib_stage_init_init, 4204 mlx5_ib_stage_init_cleanup), 4205 STAGE_CREATE(MLX5_IB_STAGE_FS, 4206 mlx5_ib_fs_init, 4207 mlx5_ib_fs_cleanup), 4208 STAGE_CREATE(MLX5_IB_STAGE_CAPS, 4209 mlx5_ib_stage_caps_init, 4210 mlx5_ib_stage_caps_cleanup), 4211 STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB, 4212 mlx5_ib_stage_non_default_cb, 4213 NULL), 4214 STAGE_CREATE(MLX5_IB_STAGE_ROCE, 4215 mlx5_ib_roce_init, 4216 mlx5_ib_roce_cleanup), 4217 STAGE_CREATE(MLX5_IB_STAGE_QP, 4218 mlx5_init_qp_table, 4219 mlx5_cleanup_qp_table), 4220 STAGE_CREATE(MLX5_IB_STAGE_SRQ, 4221 mlx5_init_srq_table, 4222 mlx5_cleanup_srq_table), 4223 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES, 4224 mlx5_ib_dev_res_init, 4225 mlx5_ib_dev_res_cleanup), 4226 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER, 4227 mlx5_ib_stage_dev_notifier_init, 4228 mlx5_ib_stage_dev_notifier_cleanup), 4229 STAGE_CREATE(MLX5_IB_STAGE_ODP, 4230 mlx5_ib_odp_init_one, 4231 mlx5_ib_odp_cleanup_one), 4232 STAGE_CREATE(MLX5_IB_STAGE_COUNTERS, 4233 mlx5_ib_counters_init, 4234 mlx5_ib_counters_cleanup), 4235 STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS, 4236 mlx5_ib_stage_cong_debugfs_init, 4237 mlx5_ib_stage_cong_debugfs_cleanup), 4238 STAGE_CREATE(MLX5_IB_STAGE_UAR, 4239 mlx5_ib_stage_uar_init, 4240 mlx5_ib_stage_uar_cleanup), 4241 STAGE_CREATE(MLX5_IB_STAGE_BFREG, 4242 mlx5_ib_stage_bfrag_init, 4243 mlx5_ib_stage_bfrag_cleanup), 4244 STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR, 4245 NULL, 4246 mlx5_ib_stage_pre_ib_reg_umr_cleanup), 4247 STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID, 4248 mlx5_ib_devx_init, 4249 mlx5_ib_devx_cleanup), 4250 STAGE_CREATE(MLX5_IB_STAGE_IB_REG, 4251 mlx5_ib_stage_ib_reg_init, 4252 mlx5_ib_stage_ib_reg_cleanup), 4253 STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR, 4254 mlx5_ib_stage_post_ib_reg_umr_init, 4255 NULL), 4256 STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP, 4257 mlx5_ib_stage_delay_drop_init, 4258 mlx5_ib_stage_delay_drop_cleanup), 4259 STAGE_CREATE(MLX5_IB_STAGE_RESTRACK, 4260 mlx5_ib_restrack_init, 4261 NULL), 4262}; 4263 4264const struct mlx5_ib_profile raw_eth_profile = { 4265 STAGE_CREATE(MLX5_IB_STAGE_INIT, 4266 mlx5_ib_stage_init_init, 4267 mlx5_ib_stage_init_cleanup), 4268 STAGE_CREATE(MLX5_IB_STAGE_FS, 4269 mlx5_ib_fs_init, 4270 mlx5_ib_fs_cleanup), 4271 STAGE_CREATE(MLX5_IB_STAGE_CAPS, 4272 mlx5_ib_stage_caps_init, 4273 mlx5_ib_stage_caps_cleanup), 4274 STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB, 4275 mlx5_ib_stage_raw_eth_non_default_cb, 4276 NULL), 4277 STAGE_CREATE(MLX5_IB_STAGE_ROCE, 4278 mlx5_ib_roce_init, 4279 mlx5_ib_roce_cleanup), 4280 STAGE_CREATE(MLX5_IB_STAGE_QP, 4281 mlx5_init_qp_table, 4282 mlx5_cleanup_qp_table), 4283 STAGE_CREATE(MLX5_IB_STAGE_SRQ, 4284 mlx5_init_srq_table, 4285 mlx5_cleanup_srq_table), 4286 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES, 4287 mlx5_ib_dev_res_init, 4288 mlx5_ib_dev_res_cleanup), 4289 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER, 4290 mlx5_ib_stage_dev_notifier_init, 4291 mlx5_ib_stage_dev_notifier_cleanup), 4292 STAGE_CREATE(MLX5_IB_STAGE_COUNTERS, 4293 mlx5_ib_counters_init, 4294 mlx5_ib_counters_cleanup), 4295 STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS, 4296 mlx5_ib_stage_cong_debugfs_init, 4297 mlx5_ib_stage_cong_debugfs_cleanup), 4298 STAGE_CREATE(MLX5_IB_STAGE_UAR, 4299 mlx5_ib_stage_uar_init, 4300 mlx5_ib_stage_uar_cleanup), 4301 STAGE_CREATE(MLX5_IB_STAGE_BFREG, 4302 mlx5_ib_stage_bfrag_init, 4303 mlx5_ib_stage_bfrag_cleanup), 4304 STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR, 4305 NULL, 4306 mlx5_ib_stage_pre_ib_reg_umr_cleanup), 4307 STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID, 4308 mlx5_ib_devx_init, 4309 mlx5_ib_devx_cleanup), 4310 STAGE_CREATE(MLX5_IB_STAGE_IB_REG, 4311 mlx5_ib_stage_ib_reg_init, 4312 mlx5_ib_stage_ib_reg_cleanup), 4313 STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR, 4314 mlx5_ib_stage_post_ib_reg_umr_init, 4315 NULL), 4316 STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP, 4317 mlx5_ib_stage_delay_drop_init, 4318 mlx5_ib_stage_delay_drop_cleanup), 4319 STAGE_CREATE(MLX5_IB_STAGE_RESTRACK, 4320 mlx5_ib_restrack_init, 4321 NULL), 4322}; 4323 4324static int mlx5r_mp_probe(struct auxiliary_device *adev, 4325 const struct auxiliary_device_id *id) 4326{ 4327 struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev); 4328 struct mlx5_core_dev *mdev = idev->mdev; 4329 struct mlx5_ib_multiport_info *mpi; 4330 struct mlx5_ib_dev *dev; 4331 bool bound = false; 4332 int err; 4333 4334 mpi = kzalloc(sizeof(*mpi), GFP_KERNEL); 4335 if (!mpi) 4336 return -ENOMEM; 4337 4338 mpi->mdev = mdev; 4339 err = mlx5_query_nic_vport_system_image_guid(mdev, 4340 &mpi->sys_image_guid); 4341 if (err) { 4342 kfree(mpi); 4343 return err; 4344 } 4345 4346 mutex_lock(&mlx5_ib_multiport_mutex); 4347 list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) { 4348 if (dev->sys_image_guid == mpi->sys_image_guid) 4349 bound = mlx5_ib_bind_slave_port(dev, mpi); 4350 4351 if (bound) { 4352 rdma_roce_rescan_device(&dev->ib_dev); 4353 mpi->ibdev->ib_active = true; 4354 break; 4355 } 4356 } 4357 4358 if (!bound) { 4359 list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list); 4360 dev_dbg(mdev->device, 4361 "no suitable IB device found to bind to, added to unaffiliated list.\n"); 4362 } 4363 mutex_unlock(&mlx5_ib_multiport_mutex); 4364 4365 auxiliary_set_drvdata(adev, mpi); 4366 return 0; 4367} 4368 4369static void mlx5r_mp_remove(struct auxiliary_device *adev) 4370{ 4371 struct mlx5_ib_multiport_info *mpi; 4372 4373 mpi = auxiliary_get_drvdata(adev); 4374 mutex_lock(&mlx5_ib_multiport_mutex); 4375 if (mpi->ibdev) 4376 mlx5_ib_unbind_slave_port(mpi->ibdev, mpi); 4377 else 4378 list_del(&mpi->list); 4379 mutex_unlock(&mlx5_ib_multiport_mutex); 4380 kfree(mpi); 4381} 4382 4383static int mlx5r_probe(struct auxiliary_device *adev, 4384 const struct auxiliary_device_id *id) 4385{ 4386 struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev); 4387 struct mlx5_core_dev *mdev = idev->mdev; 4388 const struct mlx5_ib_profile *profile; 4389 int port_type_cap, num_ports, ret; 4390 enum rdma_link_layer ll; 4391 struct mlx5_ib_dev *dev; 4392 4393 port_type_cap = MLX5_CAP_GEN(mdev, port_type); 4394 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap); 4395 4396 num_ports = max(MLX5_CAP_GEN(mdev, num_ports), 4397 MLX5_CAP_GEN(mdev, num_vhca_ports)); 4398 dev = ib_alloc_device(mlx5_ib_dev, ib_dev); 4399 if (!dev) 4400 return -ENOMEM; 4401 dev->port = kcalloc(num_ports, sizeof(*dev->port), 4402 GFP_KERNEL); 4403 if (!dev->port) { 4404 ib_dealloc_device(&dev->ib_dev); 4405 return -ENOMEM; 4406 } 4407 4408 dev->mdev = mdev; 4409 dev->num_ports = num_ports; 4410 4411 if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_get_roce_state(mdev)) 4412 profile = &raw_eth_profile; 4413 else 4414 profile = &pf_profile; 4415 4416 ret = __mlx5_ib_add(dev, profile); 4417 if (ret) { 4418 kfree(dev->port); 4419 ib_dealloc_device(&dev->ib_dev); 4420 return ret; 4421 } 4422 4423 auxiliary_set_drvdata(adev, dev); 4424 return 0; 4425} 4426 4427static void mlx5r_remove(struct auxiliary_device *adev) 4428{ 4429 struct mlx5_ib_dev *dev; 4430 4431 dev = auxiliary_get_drvdata(adev); 4432 __mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX); 4433} 4434 4435static const struct auxiliary_device_id mlx5r_mp_id_table[] = { 4436 { .name = MLX5_ADEV_NAME ".multiport", }, 4437 {}, 4438}; 4439 4440static const struct auxiliary_device_id mlx5r_id_table[] = { 4441 { .name = MLX5_ADEV_NAME ".rdma", }, 4442 {}, 4443}; 4444 4445MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table); 4446MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table); 4447 4448static struct auxiliary_driver mlx5r_mp_driver = { 4449 .name = "multiport", 4450 .probe = mlx5r_mp_probe, 4451 .remove = mlx5r_mp_remove, 4452 .id_table = mlx5r_mp_id_table, 4453}; 4454 4455static struct auxiliary_driver mlx5r_driver = { 4456 .name = "rdma", 4457 .probe = mlx5r_probe, 4458 .remove = mlx5r_remove, 4459 .id_table = mlx5r_id_table, 4460}; 4461 4462static int __init mlx5_ib_init(void) 4463{ 4464 int ret; 4465 4466 xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL); 4467 if (!xlt_emergency_page) 4468 return -ENOMEM; 4469 4470 mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0); 4471 if (!mlx5_ib_event_wq) { 4472 free_page((unsigned long)xlt_emergency_page); 4473 return -ENOMEM; 4474 } 4475 4476 ret = mlx5_ib_qp_event_init(); 4477 if (ret) 4478 goto qp_event_err; 4479 4480 mlx5_ib_odp_init(); 4481 ret = mlx5r_rep_init(); 4482 if (ret) 4483 goto rep_err; 4484 ret = auxiliary_driver_register(&mlx5r_mp_driver); 4485 if (ret) 4486 goto mp_err; 4487 ret = auxiliary_driver_register(&mlx5r_driver); 4488 if (ret) 4489 goto drv_err; 4490 return 0; 4491 4492drv_err: 4493 auxiliary_driver_unregister(&mlx5r_mp_driver); 4494mp_err: 4495 mlx5r_rep_cleanup(); 4496rep_err: 4497 mlx5_ib_qp_event_cleanup(); 4498qp_event_err: 4499 destroy_workqueue(mlx5_ib_event_wq); 4500 free_page((unsigned long)xlt_emergency_page); 4501 return ret; 4502} 4503 4504static void __exit mlx5_ib_cleanup(void) 4505{ 4506 auxiliary_driver_unregister(&mlx5r_driver); 4507 auxiliary_driver_unregister(&mlx5r_mp_driver); 4508 mlx5r_rep_cleanup(); 4509 4510 mlx5_ib_qp_event_cleanup(); 4511 destroy_workqueue(mlx5_ib_event_wq); 4512 free_page((unsigned long)xlt_emergency_page); 4513} 4514 4515module_init(mlx5_ib_init); 4516module_exit(mlx5_ib_cleanup); 4517