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/log2.h> 17 #include <linux/sched.h> 18 #include <linux/sched/mm.h> 19 #include <linux/sched/task.h> 20 #include <linux/delay.h> 21 #include <rdma/ib_user_verbs.h> 22 #include <rdma/ib_addr.h> 23 #include <rdma/ib_cache.h> 24 #include <linux/mlx5/port.h> 25 #include <linux/mlx5/vport.h> 26 #include <linux/mlx5/fs.h> 27 #include <linux/mlx5/eswitch.h> 28 #include <linux/mlx5/driver.h> 29 #include <linux/mlx5/lag.h> 30 #include <linux/list.h> 31 #include <rdma/ib_smi.h> 32 #include <rdma/ib_umem_odp.h> 33 #include <rdma/lag.h> 34 #include <linux/in.h> 35 #include <linux/etherdevice.h> 36 #include "mlx5_ib.h" 37 #include "ib_rep.h" 38 #include "cmd.h" 39 #include "devx.h" 40 #include "dm.h" 41 #include "fs.h" 42 #include "srq.h" 43 #include "qp.h" 44 #include "wr.h" 45 #include "restrack.h" 46 #include "counters.h" 47 #include "umr.h" 48 #include <rdma/uverbs_std_types.h> 49 #include <rdma/uverbs_ioctl.h> 50 #include <rdma/mlx5_user_ioctl_verbs.h> 51 #include <rdma/mlx5_user_ioctl_cmds.h> 52 #include <rdma/ib_ucaps.h> 53 #include "macsec.h" 54 #include "data_direct.h" 55 #include "dmah.h" 56 57 #define UVERBS_MODULE_NAME mlx5_ib 58 #include <rdma/uverbs_named_ioctl.h> 59 60 MODULE_AUTHOR("Eli Cohen <eli@mellanox.com>"); 61 MODULE_DESCRIPTION("Mellanox 5th generation network adapters (ConnectX series) IB driver"); 62 MODULE_LICENSE("Dual BSD/GPL"); 63 64 struct mlx5_ib_event_work { 65 struct work_struct work; 66 union { 67 struct mlx5_ib_dev *dev; 68 struct mlx5_ib_multiport_info *mpi; 69 }; 70 bool is_slave; 71 unsigned int event; 72 void *param; 73 }; 74 75 enum { 76 MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3, 77 }; 78 79 static struct workqueue_struct *mlx5_ib_event_wq; 80 static LIST_HEAD(mlx5_ib_unaffiliated_port_list); 81 static LIST_HEAD(mlx5_ib_dev_list); 82 /* 83 * This mutex should be held when accessing either of the above lists 84 */ 85 static DEFINE_MUTEX(mlx5_ib_multiport_mutex); 86 87 struct mlx5_ib_dev *mlx5_ib_get_ibdev_from_mpi(struct mlx5_ib_multiport_info *mpi) 88 { 89 struct mlx5_ib_dev *dev; 90 91 mutex_lock(&mlx5_ib_multiport_mutex); 92 dev = mpi->ibdev; 93 mutex_unlock(&mlx5_ib_multiport_mutex); 94 return dev; 95 } 96 97 static enum rdma_link_layer 98 mlx5_port_type_cap_to_rdma_ll(int port_type_cap) 99 { 100 switch (port_type_cap) { 101 case MLX5_CAP_PORT_TYPE_IB: 102 return IB_LINK_LAYER_INFINIBAND; 103 case MLX5_CAP_PORT_TYPE_ETH: 104 return IB_LINK_LAYER_ETHERNET; 105 default: 106 return IB_LINK_LAYER_UNSPECIFIED; 107 } 108 } 109 110 static enum rdma_link_layer 111 mlx5_ib_port_link_layer(struct ib_device *device, u32 port_num) 112 { 113 struct mlx5_ib_dev *dev = to_mdev(device); 114 int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type); 115 116 return mlx5_port_type_cap_to_rdma_ll(port_type_cap); 117 } 118 119 static int get_port_state(struct ib_device *ibdev, 120 u32 port_num, 121 enum ib_port_state *state) 122 { 123 struct ib_port_attr attr; 124 int ret; 125 126 memset(&attr, 0, sizeof(attr)); 127 ret = ibdev->ops.query_port(ibdev, port_num, &attr); 128 if (!ret) 129 *state = attr.state; 130 return ret; 131 } 132 133 static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev, 134 struct net_device *ndev, 135 struct net_device *upper, 136 u32 *port_num) 137 { 138 struct net_device *rep_ndev; 139 struct mlx5_ib_port *port; 140 int i; 141 142 for (i = 0; i < dev->num_ports; i++) { 143 port = &dev->port[i]; 144 if (!port->rep) 145 continue; 146 147 if (upper == ndev && port->rep->vport == MLX5_VPORT_UPLINK) { 148 *port_num = i + 1; 149 return &port->roce; 150 } 151 152 if (upper && port->rep->vport == MLX5_VPORT_UPLINK) 153 continue; 154 rep_ndev = ib_device_get_netdev(&dev->ib_dev, i + 1); 155 if (rep_ndev && rep_ndev == ndev) { 156 dev_put(rep_ndev); 157 *port_num = i + 1; 158 return &port->roce; 159 } 160 161 dev_put(rep_ndev); 162 } 163 164 return NULL; 165 } 166 167 static bool mlx5_netdev_send_event(struct mlx5_ib_dev *dev, 168 struct net_device *ndev, 169 struct net_device *upper, 170 struct net_device *ib_ndev) 171 { 172 if (!dev->ib_active) 173 return false; 174 175 /* Event is about our upper device */ 176 if (upper == ndev) 177 return true; 178 179 /* RDMA device is not in lag and not in switchdev */ 180 if (!dev->is_rep && !upper && ndev == ib_ndev) 181 return true; 182 183 /* RDMA devie is in switchdev */ 184 if (dev->is_rep && ndev == ib_ndev) 185 return true; 186 187 return false; 188 } 189 190 static struct net_device *mlx5_ib_get_rep_uplink_netdev(struct mlx5_ib_dev *ibdev) 191 { 192 struct mlx5_ib_port *port; 193 int i; 194 195 for (i = 0; i < ibdev->num_ports; i++) { 196 port = &ibdev->port[i]; 197 if (port->rep && port->rep->vport == MLX5_VPORT_UPLINK) { 198 return ib_device_get_netdev(&ibdev->ib_dev, i + 1); 199 } 200 } 201 202 return NULL; 203 } 204 205 static int mlx5_netdev_event(struct notifier_block *this, 206 unsigned long event, void *ptr) 207 { 208 struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb); 209 struct net_device *ndev = netdev_notifier_info_to_dev(ptr); 210 u32 port_num = roce->native_port_num; 211 struct net_device *ib_ndev = NULL; 212 struct mlx5_core_dev *mdev; 213 struct mlx5_ib_dev *ibdev; 214 215 ibdev = roce->dev; 216 mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL); 217 if (!mdev) 218 return NOTIFY_DONE; 219 220 switch (event) { 221 case NETDEV_REGISTER: 222 /* Should already be registered during the load */ 223 if (ibdev->is_rep) 224 break; 225 226 ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num); 227 /* Exit if already registered */ 228 if (ib_ndev) 229 goto put_ndev; 230 231 if (ndev->dev.parent == mdev->device) 232 ib_device_set_netdev(&ibdev->ib_dev, ndev, port_num); 233 break; 234 235 case NETDEV_UNREGISTER: 236 /* In case of reps, ib device goes away before the netdevs */ 237 if (ibdev->is_rep) 238 break; 239 ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num); 240 if (ib_ndev == ndev) 241 ib_device_set_netdev(&ibdev->ib_dev, NULL, port_num); 242 goto put_ndev; 243 244 case NETDEV_CHANGE: 245 case NETDEV_UP: 246 case NETDEV_DOWN: { 247 struct net_device *upper = NULL; 248 249 if (!netif_is_lag_master(ndev) && !netif_is_lag_port(ndev) && 250 !mlx5_core_mp_enabled(mdev)) 251 return NOTIFY_DONE; 252 253 if (mlx5_lag_is_roce(mdev) || mlx5_lag_is_sriov(mdev)) { 254 struct net_device *lag_ndev; 255 256 if(mlx5_lag_is_roce(mdev)) 257 lag_ndev = ib_device_get_netdev(&ibdev->ib_dev, 1); 258 else /* sriov lag */ 259 lag_ndev = mlx5_ib_get_rep_uplink_netdev(ibdev); 260 261 if (lag_ndev) { 262 upper = netdev_master_upper_dev_get(lag_ndev); 263 dev_put(lag_ndev); 264 } else { 265 goto done; 266 } 267 } 268 269 if (ibdev->is_rep) 270 roce = mlx5_get_rep_roce(ibdev, ndev, upper, &port_num); 271 if (!roce) 272 return NOTIFY_DONE; 273 274 ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num); 275 276 if (mlx5_netdev_send_event(ibdev, ndev, upper, ib_ndev)) { 277 struct ib_event ibev = { }; 278 enum ib_port_state port_state; 279 280 if (get_port_state(&ibdev->ib_dev, port_num, 281 &port_state)) 282 goto put_ndev; 283 284 if (roce->last_port_state == port_state) 285 goto put_ndev; 286 287 roce->last_port_state = port_state; 288 ibev.device = &ibdev->ib_dev; 289 if (port_state == IB_PORT_DOWN) 290 ibev.event = IB_EVENT_PORT_ERR; 291 else if (port_state == IB_PORT_ACTIVE) 292 ibev.event = IB_EVENT_PORT_ACTIVE; 293 else 294 goto put_ndev; 295 296 ibev.element.port_num = port_num; 297 ib_dispatch_event(&ibev); 298 } 299 break; 300 } 301 302 default: 303 break; 304 } 305 put_ndev: 306 dev_put(ib_ndev); 307 done: 308 mlx5_ib_put_native_port_mdev(ibdev, port_num); 309 return NOTIFY_DONE; 310 } 311 312 struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev, 313 u32 ib_port_num, 314 u32 *native_port_num) 315 { 316 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev, 317 ib_port_num); 318 struct mlx5_core_dev *mdev = NULL; 319 struct mlx5_ib_multiport_info *mpi; 320 struct mlx5_ib_port *port; 321 322 if (ibdev->ib_dev.type == RDMA_DEVICE_TYPE_SMI) { 323 if (native_port_num) 324 *native_port_num = smi_to_native_portnum(ibdev, 325 ib_port_num); 326 return ibdev->mdev; 327 328 } 329 330 if (!mlx5_core_mp_enabled(ibdev->mdev) || 331 ll != IB_LINK_LAYER_ETHERNET) { 332 if (native_port_num) 333 *native_port_num = ib_port_num; 334 return ibdev->mdev; 335 } 336 337 if (native_port_num) 338 *native_port_num = 1; 339 340 port = &ibdev->port[ib_port_num - 1]; 341 spin_lock(&port->mp.mpi_lock); 342 mpi = ibdev->port[ib_port_num - 1].mp.mpi; 343 if (mpi && !mpi->unaffiliate) { 344 mdev = mpi->mdev; 345 /* If it's the master no need to refcount, it'll exist 346 * as long as the ib_dev exists. 347 */ 348 if (!mpi->is_master) 349 mpi->mdev_refcnt++; 350 } 351 spin_unlock(&port->mp.mpi_lock); 352 353 return mdev; 354 } 355 356 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 port_num) 357 { 358 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&ibdev->ib_dev, 359 port_num); 360 struct mlx5_ib_multiport_info *mpi; 361 struct mlx5_ib_port *port; 362 363 if (!mlx5_core_mp_enabled(ibdev->mdev) || ll != IB_LINK_LAYER_ETHERNET) 364 return; 365 366 port = &ibdev->port[port_num - 1]; 367 368 spin_lock(&port->mp.mpi_lock); 369 mpi = ibdev->port[port_num - 1].mp.mpi; 370 if (mpi->is_master) 371 goto out; 372 373 mpi->mdev_refcnt--; 374 if (mpi->unaffiliate) 375 complete(&mpi->unref_comp); 376 out: 377 spin_unlock(&port->mp.mpi_lock); 378 } 379 380 static int translate_eth_legacy_proto_oper(u32 eth_proto_oper, 381 u16 *active_speed, u8 *active_width) 382 { 383 switch (eth_proto_oper) { 384 case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII): 385 case MLX5E_PROT_MASK(MLX5E_1000BASE_KX): 386 case MLX5E_PROT_MASK(MLX5E_100BASE_TX): 387 case MLX5E_PROT_MASK(MLX5E_1000BASE_T): 388 *active_width = IB_WIDTH_1X; 389 *active_speed = IB_SPEED_SDR; 390 break; 391 case MLX5E_PROT_MASK(MLX5E_10GBASE_T): 392 case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4): 393 case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4): 394 case MLX5E_PROT_MASK(MLX5E_10GBASE_KR): 395 case MLX5E_PROT_MASK(MLX5E_10GBASE_CR): 396 case MLX5E_PROT_MASK(MLX5E_10GBASE_SR): 397 case MLX5E_PROT_MASK(MLX5E_10GBASE_ER): 398 *active_width = IB_WIDTH_1X; 399 *active_speed = IB_SPEED_QDR; 400 break; 401 case MLX5E_PROT_MASK(MLX5E_25GBASE_CR): 402 case MLX5E_PROT_MASK(MLX5E_25GBASE_KR): 403 case MLX5E_PROT_MASK(MLX5E_25GBASE_SR): 404 *active_width = IB_WIDTH_1X; 405 *active_speed = IB_SPEED_EDR; 406 break; 407 case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4): 408 case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4): 409 case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4): 410 case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4): 411 *active_width = IB_WIDTH_4X; 412 *active_speed = IB_SPEED_QDR; 413 break; 414 case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2): 415 case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2): 416 case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2): 417 *active_width = IB_WIDTH_1X; 418 *active_speed = IB_SPEED_HDR; 419 break; 420 case MLX5E_PROT_MASK(MLX5E_56GBASE_R4): 421 *active_width = IB_WIDTH_4X; 422 *active_speed = IB_SPEED_FDR; 423 break; 424 case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4): 425 case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4): 426 case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4): 427 case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4): 428 *active_width = IB_WIDTH_4X; 429 *active_speed = IB_SPEED_EDR; 430 break; 431 default: 432 return -EINVAL; 433 } 434 435 return 0; 436 } 437 438 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u16 *active_speed, 439 u8 *active_width) 440 { 441 switch (eth_proto_oper) { 442 case MLX5E_PROT_MASK(MLX5E_SGMII_100M): 443 case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII): 444 *active_width = IB_WIDTH_1X; 445 *active_speed = IB_SPEED_SDR; 446 break; 447 case MLX5E_PROT_MASK(MLX5E_5GBASE_R): 448 *active_width = IB_WIDTH_1X; 449 *active_speed = IB_SPEED_DDR; 450 break; 451 case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1): 452 *active_width = IB_WIDTH_1X; 453 *active_speed = IB_SPEED_QDR; 454 break; 455 case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4): 456 *active_width = IB_WIDTH_4X; 457 *active_speed = IB_SPEED_QDR; 458 break; 459 case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR): 460 *active_width = IB_WIDTH_1X; 461 *active_speed = IB_SPEED_EDR; 462 break; 463 case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2): 464 *active_width = IB_WIDTH_2X; 465 *active_speed = IB_SPEED_EDR; 466 break; 467 case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR): 468 *active_width = IB_WIDTH_1X; 469 *active_speed = IB_SPEED_HDR; 470 break; 471 case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4): 472 *active_width = IB_WIDTH_4X; 473 *active_speed = IB_SPEED_EDR; 474 break; 475 case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2): 476 *active_width = IB_WIDTH_2X; 477 *active_speed = IB_SPEED_HDR; 478 break; 479 case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR): 480 *active_width = IB_WIDTH_1X; 481 *active_speed = IB_SPEED_NDR; 482 break; 483 case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4): 484 *active_width = IB_WIDTH_4X; 485 *active_speed = IB_SPEED_HDR; 486 break; 487 case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2): 488 *active_width = IB_WIDTH_2X; 489 *active_speed = IB_SPEED_NDR; 490 break; 491 case MLX5E_PROT_MASK(MLX5E_200GAUI_1_200GBASE_CR1_KR1): 492 *active_width = IB_WIDTH_1X; 493 *active_speed = IB_SPEED_XDR; 494 break; 495 case MLX5E_PROT_MASK(MLX5E_400GAUI_8_400GBASE_CR8): 496 *active_width = IB_WIDTH_8X; 497 *active_speed = IB_SPEED_HDR; 498 break; 499 case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4): 500 *active_width = IB_WIDTH_4X; 501 *active_speed = IB_SPEED_NDR; 502 break; 503 case MLX5E_PROT_MASK(MLX5E_400GAUI_2_400GBASE_CR2_KR2): 504 *active_width = IB_WIDTH_2X; 505 *active_speed = IB_SPEED_XDR; 506 break; 507 case MLX5E_PROT_MASK(MLX5E_800GAUI_8_800GBASE_CR8_KR8): 508 *active_width = IB_WIDTH_8X; 509 *active_speed = IB_SPEED_NDR; 510 break; 511 case MLX5E_PROT_MASK(MLX5E_800GAUI_4_800GBASE_CR4_KR4): 512 *active_width = IB_WIDTH_4X; 513 *active_speed = IB_SPEED_XDR; 514 break; 515 case MLX5E_PROT_MASK(MLX5E_1600GAUI_8_1600GBASE_CR8_KR8): 516 *active_width = IB_WIDTH_8X; 517 *active_speed = IB_SPEED_XDR; 518 break; 519 default: 520 return -EINVAL; 521 } 522 523 return 0; 524 } 525 526 static int translate_eth_proto_oper(u32 eth_proto_oper, u16 *active_speed, 527 u8 *active_width, bool ext) 528 { 529 return ext ? 530 translate_eth_ext_proto_oper(eth_proto_oper, active_speed, 531 active_width) : 532 translate_eth_legacy_proto_oper(eth_proto_oper, active_speed, 533 active_width); 534 } 535 536 static int mlx5_query_port_roce(struct ib_device *device, u32 port_num, 537 struct ib_port_attr *props) 538 { 539 struct mlx5_ib_dev *dev = to_mdev(device); 540 u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {0}; 541 struct mlx5_core_dev *mdev; 542 struct net_device *ndev, *upper; 543 enum ib_mtu ndev_ib_mtu; 544 bool put_mdev = true; 545 u32 eth_prot_oper; 546 u32 mdev_port_num; 547 bool ext; 548 int err; 549 550 mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num); 551 if (!mdev) { 552 /* This means the port isn't affiliated yet. Get the 553 * info for the master port instead. 554 */ 555 put_mdev = false; 556 mdev = dev->mdev; 557 mdev_port_num = 1; 558 port_num = 1; 559 } 560 561 /* Possible bad flows are checked before filling out props so in case 562 * of an error it will still be zeroed out. 563 * Use native port in case of reps 564 */ 565 if (dev->is_rep) { 566 struct mlx5_eswitch_rep *rep; 567 568 rep = dev->port[port_num - 1].rep; 569 if (rep) { 570 mdev = mlx5_eswitch_get_core_dev(rep->esw); 571 WARN_ON(!mdev); 572 } 573 mdev_port_num = 1; 574 } 575 576 err = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN, 577 mdev_port_num, 0); 578 579 if (err) 580 goto out; 581 ext = !!MLX5_GET_ETH_PROTO(ptys_reg, out, true, eth_proto_capability); 582 eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper); 583 584 props->active_width = IB_WIDTH_4X; 585 props->active_speed = IB_SPEED_QDR; 586 587 translate_eth_proto_oper(eth_prot_oper, &props->active_speed, 588 &props->active_width, ext); 589 590 if (!dev->is_rep && dev->mdev->roce.roce_en) { 591 u16 qkey_viol_cntr; 592 593 props->port_cap_flags |= IB_PORT_CM_SUP; 594 props->ip_gids = true; 595 props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev, 596 roce_address_table_size); 597 mlx5_query_nic_vport_qkey_viol_cntr(mdev, &qkey_viol_cntr); 598 props->qkey_viol_cntr = qkey_viol_cntr; 599 } 600 props->max_mtu = IB_MTU_4096; 601 props->max_msg_sz = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg); 602 props->pkey_tbl_len = 1; 603 props->state = IB_PORT_DOWN; 604 props->phys_state = IB_PORT_PHYS_STATE_DISABLED; 605 606 /* If this is a stub query for an unaffiliated port stop here */ 607 if (!put_mdev) 608 goto out; 609 610 ndev = ib_device_get_netdev(device, port_num); 611 if (!ndev) 612 goto out; 613 614 if (mlx5_lag_is_roce(mdev) || mlx5_lag_is_sriov(mdev)) { 615 rcu_read_lock(); 616 upper = netdev_master_upper_dev_get_rcu(ndev); 617 if (upper) { 618 dev_put(ndev); 619 ndev = upper; 620 dev_hold(ndev); 621 } 622 rcu_read_unlock(); 623 } 624 625 if (netif_running(ndev) && netif_carrier_ok(ndev)) { 626 props->state = IB_PORT_ACTIVE; 627 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP; 628 } 629 630 ndev_ib_mtu = iboe_get_mtu(ndev->mtu); 631 632 dev_put(ndev); 633 634 props->active_mtu = min(props->max_mtu, ndev_ib_mtu); 635 out: 636 if (put_mdev) 637 mlx5_ib_put_native_port_mdev(dev, port_num); 638 return err; 639 } 640 641 int set_roce_addr(struct mlx5_ib_dev *dev, u32 port_num, 642 unsigned int index, const union ib_gid *gid, 643 const struct ib_gid_attr *attr) 644 { 645 enum ib_gid_type gid_type; 646 u16 vlan_id = 0xffff; 647 u8 roce_version = 0; 648 u8 roce_l3_type = 0; 649 u8 mac[ETH_ALEN]; 650 int ret; 651 652 gid_type = attr->gid_type; 653 if (gid) { 654 ret = rdma_read_gid_l2_fields(attr, &vlan_id, &mac[0]); 655 if (ret) 656 return ret; 657 } 658 659 switch (gid_type) { 660 case IB_GID_TYPE_ROCE: 661 roce_version = MLX5_ROCE_VERSION_1; 662 break; 663 case IB_GID_TYPE_ROCE_UDP_ENCAP: 664 roce_version = MLX5_ROCE_VERSION_2; 665 if (gid && ipv6_addr_v4mapped((void *)gid)) 666 roce_l3_type = MLX5_ROCE_L3_TYPE_IPV4; 667 else 668 roce_l3_type = MLX5_ROCE_L3_TYPE_IPV6; 669 break; 670 671 default: 672 mlx5_ib_warn(dev, "Unexpected GID type %u\n", gid_type); 673 } 674 675 return mlx5_core_roce_gid_set(dev->mdev, index, roce_version, 676 roce_l3_type, gid->raw, mac, 677 vlan_id < VLAN_CFI_MASK, vlan_id, 678 port_num); 679 } 680 681 static int mlx5_ib_add_gid(const struct ib_gid_attr *attr, 682 __always_unused void **context) 683 { 684 int ret; 685 686 ret = mlx5r_add_gid_macsec_operations(attr); 687 if (ret) 688 return ret; 689 690 return set_roce_addr(to_mdev(attr->device), attr->port_num, 691 attr->index, &attr->gid, attr); 692 } 693 694 static int mlx5_ib_del_gid(const struct ib_gid_attr *attr, 695 __always_unused void **context) 696 { 697 int ret; 698 699 ret = set_roce_addr(to_mdev(attr->device), attr->port_num, 700 attr->index, NULL, attr); 701 if (ret) 702 return ret; 703 704 mlx5r_del_gid_macsec_operations(attr); 705 return 0; 706 } 707 708 __be16 mlx5_get_roce_udp_sport_min(const struct mlx5_ib_dev *dev, 709 const struct ib_gid_attr *attr) 710 { 711 if (attr->gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP) 712 return 0; 713 714 return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port)); 715 } 716 717 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev) 718 { 719 if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB) 720 return !MLX5_CAP_GEN(dev->mdev, ib_virt); 721 return 0; 722 } 723 724 enum { 725 MLX5_VPORT_ACCESS_METHOD_MAD, 726 MLX5_VPORT_ACCESS_METHOD_HCA, 727 MLX5_VPORT_ACCESS_METHOD_NIC, 728 }; 729 730 static int mlx5_get_vport_access_method(struct ib_device *ibdev) 731 { 732 if (mlx5_use_mad_ifc(to_mdev(ibdev))) 733 return MLX5_VPORT_ACCESS_METHOD_MAD; 734 735 if (mlx5_ib_port_link_layer(ibdev, 1) == 736 IB_LINK_LAYER_ETHERNET) 737 return MLX5_VPORT_ACCESS_METHOD_NIC; 738 739 return MLX5_VPORT_ACCESS_METHOD_HCA; 740 } 741 742 static void get_atomic_caps(struct mlx5_ib_dev *dev, 743 u8 atomic_size_qp, 744 struct ib_device_attr *props) 745 { 746 u8 tmp; 747 u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations); 748 u8 atomic_req_8B_endianness_mode = 749 MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianness_mode); 750 751 /* Check if HW supports 8 bytes standard atomic operations and capable 752 * of host endianness respond 753 */ 754 tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD; 755 if (((atomic_operations & tmp) == tmp) && 756 (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) && 757 (atomic_req_8B_endianness_mode)) { 758 props->atomic_cap = IB_ATOMIC_HCA; 759 } else { 760 props->atomic_cap = IB_ATOMIC_NONE; 761 } 762 } 763 764 static void get_atomic_caps_qp(struct mlx5_ib_dev *dev, 765 struct ib_device_attr *props) 766 { 767 u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp); 768 769 get_atomic_caps(dev, atomic_size_qp, props); 770 } 771 772 static int mlx5_query_system_image_guid(struct ib_device *ibdev, 773 __be64 *sys_image_guid) 774 { 775 struct mlx5_ib_dev *dev = to_mdev(ibdev); 776 struct mlx5_core_dev *mdev = dev->mdev; 777 u64 tmp; 778 int err; 779 780 switch (mlx5_get_vport_access_method(ibdev)) { 781 case MLX5_VPORT_ACCESS_METHOD_MAD: 782 return mlx5_query_mad_ifc_system_image_guid(ibdev, 783 sys_image_guid); 784 785 case MLX5_VPORT_ACCESS_METHOD_HCA: 786 err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp); 787 break; 788 789 case MLX5_VPORT_ACCESS_METHOD_NIC: 790 err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp); 791 break; 792 793 default: 794 return -EINVAL; 795 } 796 797 if (!err) 798 *sys_image_guid = cpu_to_be64(tmp); 799 800 return err; 801 802 } 803 804 static int mlx5_query_max_pkeys(struct ib_device *ibdev, 805 u16 *max_pkeys) 806 { 807 struct mlx5_ib_dev *dev = to_mdev(ibdev); 808 struct mlx5_core_dev *mdev = dev->mdev; 809 810 switch (mlx5_get_vport_access_method(ibdev)) { 811 case MLX5_VPORT_ACCESS_METHOD_MAD: 812 return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys); 813 814 case MLX5_VPORT_ACCESS_METHOD_HCA: 815 case MLX5_VPORT_ACCESS_METHOD_NIC: 816 *max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, 817 pkey_table_size)); 818 return 0; 819 820 default: 821 return -EINVAL; 822 } 823 } 824 825 static int mlx5_query_vendor_id(struct ib_device *ibdev, 826 u32 *vendor_id) 827 { 828 struct mlx5_ib_dev *dev = to_mdev(ibdev); 829 830 switch (mlx5_get_vport_access_method(ibdev)) { 831 case MLX5_VPORT_ACCESS_METHOD_MAD: 832 return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id); 833 834 case MLX5_VPORT_ACCESS_METHOD_HCA: 835 case MLX5_VPORT_ACCESS_METHOD_NIC: 836 return mlx5_core_query_vendor_id(dev->mdev, vendor_id); 837 838 default: 839 return -EINVAL; 840 } 841 } 842 843 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev, 844 __be64 *node_guid) 845 { 846 u64 tmp; 847 int err; 848 849 switch (mlx5_get_vport_access_method(&dev->ib_dev)) { 850 case MLX5_VPORT_ACCESS_METHOD_MAD: 851 return mlx5_query_mad_ifc_node_guid(dev, node_guid); 852 853 case MLX5_VPORT_ACCESS_METHOD_HCA: 854 err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp); 855 break; 856 857 case MLX5_VPORT_ACCESS_METHOD_NIC: 858 err = mlx5_query_nic_vport_node_guid(dev->mdev, 0, false, &tmp); 859 break; 860 861 default: 862 return -EINVAL; 863 } 864 865 if (!err) 866 *node_guid = cpu_to_be64(tmp); 867 868 return err; 869 } 870 871 struct mlx5_reg_node_desc { 872 u8 desc[IB_DEVICE_NODE_DESC_MAX]; 873 }; 874 875 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc) 876 { 877 struct mlx5_reg_node_desc in; 878 879 if (mlx5_use_mad_ifc(dev)) 880 return mlx5_query_mad_ifc_node_desc(dev, node_desc); 881 882 memset(&in, 0, sizeof(in)); 883 884 return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc, 885 sizeof(struct mlx5_reg_node_desc), 886 MLX5_REG_NODE_DESC, 0, 0); 887 } 888 889 static void fill_esw_mgr_reg_c0(struct mlx5_core_dev *mdev, 890 struct mlx5_ib_query_device_resp *resp) 891 { 892 struct mlx5_eswitch *esw = mdev->priv.eswitch; 893 u16 vport = mlx5_eswitch_manager_vport(mdev); 894 895 resp->reg_c0.value = mlx5_eswitch_get_vport_metadata_for_match(esw, 896 vport); 897 resp->reg_c0.mask = mlx5_eswitch_get_vport_metadata_mask(); 898 } 899 900 /* 901 * Calculate maximum SQ overhead across all QP types. 902 * Other QP types (REG_UMR, UC, RC, UD/SMI/GSI, XRC_TGT) 903 * have smaller overhead than the types calculated below, 904 * so they are implicitly included. 905 */ 906 static u32 mlx5_ib_calc_max_sq_overhead(void) 907 { 908 u32 max_overhead_xrc, overhead_ud_lso, a, b; 909 910 /* XRC_INI */ 911 max_overhead_xrc = sizeof(struct mlx5_wqe_xrc_seg); 912 max_overhead_xrc += sizeof(struct mlx5_wqe_ctrl_seg); 913 a = sizeof(struct mlx5_wqe_atomic_seg) + 914 sizeof(struct mlx5_wqe_raddr_seg); 915 b = sizeof(struct mlx5_wqe_umr_ctrl_seg) + 916 sizeof(struct mlx5_mkey_seg) + 917 MLX5_IB_SQ_UMR_INLINE_THRESHOLD / MLX5_IB_UMR_OCTOWORD; 918 max_overhead_xrc += max(a, b); 919 920 /* UD with LSO */ 921 overhead_ud_lso = sizeof(struct mlx5_wqe_ctrl_seg); 922 overhead_ud_lso += sizeof(struct mlx5_wqe_eth_pad); 923 overhead_ud_lso += sizeof(struct mlx5_wqe_eth_seg); 924 overhead_ud_lso += sizeof(struct mlx5_wqe_datagram_seg); 925 926 return max(max_overhead_xrc, overhead_ud_lso); 927 } 928 929 static u32 mlx5_ib_calc_max_qp_wr(struct mlx5_ib_dev *dev) 930 { 931 struct mlx5_core_dev *mdev = dev->mdev; 932 u32 max_wqe_bb_units = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz); 933 u32 max_wqe_size; 934 /* max QP overhead + 1 SGE, no inline, no special features */ 935 max_wqe_size = mlx5_ib_calc_max_sq_overhead() + 936 sizeof(struct mlx5_wqe_data_seg); 937 938 max_wqe_size = roundup_pow_of_two(max_wqe_size); 939 940 max_wqe_size = ALIGN(max_wqe_size, MLX5_SEND_WQE_BB); 941 942 return (max_wqe_bb_units * MLX5_SEND_WQE_BB) / max_wqe_size; 943 } 944 945 static int mlx5_ib_query_device(struct ib_device *ibdev, 946 struct ib_device_attr *props, 947 struct ib_udata *uhw) 948 { 949 size_t uhw_outlen = (uhw) ? uhw->outlen : 0; 950 struct mlx5_ib_dev *dev = to_mdev(ibdev); 951 struct mlx5_core_dev *mdev = dev->mdev; 952 int err = -ENOMEM; 953 int max_sq_desc; 954 int max_rq_sg; 955 int max_sq_sg; 956 u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz); 957 bool raw_support = !mlx5_core_mp_enabled(mdev); 958 struct mlx5_ib_query_device_resp resp = {}; 959 size_t resp_len; 960 u64 max_tso; 961 962 resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length); 963 if (uhw_outlen && uhw_outlen < resp_len) 964 return -EINVAL; 965 966 resp.response_length = resp_len; 967 968 if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen)) 969 return -EINVAL; 970 971 memset(props, 0, sizeof(*props)); 972 err = mlx5_query_system_image_guid(ibdev, 973 &props->sys_image_guid); 974 if (err) 975 return err; 976 977 props->max_pkeys = dev->pkey_table_len; 978 979 err = mlx5_query_vendor_id(ibdev, &props->vendor_id); 980 if (err) 981 return err; 982 983 props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) | 984 (fw_rev_min(dev->mdev) << 16) | 985 fw_rev_sub(dev->mdev); 986 props->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT | 987 IB_DEVICE_PORT_ACTIVE_EVENT | 988 IB_DEVICE_SYS_IMAGE_GUID | 989 IB_DEVICE_RC_RNR_NAK_GEN; 990 991 if (MLX5_CAP_GEN(mdev, pkv)) 992 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR; 993 if (MLX5_CAP_GEN(mdev, qkv)) 994 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR; 995 if (MLX5_CAP_GEN(mdev, apm)) 996 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG; 997 if (MLX5_CAP_GEN(mdev, xrc)) 998 props->device_cap_flags |= IB_DEVICE_XRC; 999 if (MLX5_CAP_GEN(mdev, imaicl)) { 1000 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW | 1001 IB_DEVICE_MEM_WINDOW_TYPE_2B; 1002 props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey); 1003 /* We support 'Gappy' memory registration too */ 1004 props->kernel_cap_flags |= IBK_SG_GAPS_REG; 1005 } 1006 /* IB_WR_REG_MR always requires changing the entity size with UMR */ 1007 if (!MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled)) 1008 props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS; 1009 if (MLX5_CAP_GEN(mdev, sho)) { 1010 props->kernel_cap_flags |= IBK_INTEGRITY_HANDOVER; 1011 /* At this stage no support for signature handover */ 1012 props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 | 1013 IB_PROT_T10DIF_TYPE_2 | 1014 IB_PROT_T10DIF_TYPE_3; 1015 props->sig_guard_cap = IB_GUARD_T10DIF_CRC | 1016 IB_GUARD_T10DIF_CSUM; 1017 } 1018 if (MLX5_CAP_GEN(mdev, block_lb_mc)) 1019 props->kernel_cap_flags |= IBK_BLOCK_MULTICAST_LOOPBACK; 1020 1021 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && raw_support) { 1022 if (MLX5_CAP_ETH(mdev, csum_cap)) { 1023 /* Legacy bit to support old userspace libraries */ 1024 props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM; 1025 props->raw_packet_caps |= IB_RAW_PACKET_CAP_IP_CSUM; 1026 } 1027 1028 if (MLX5_CAP_ETH(dev->mdev, vlan_cap)) 1029 props->raw_packet_caps |= 1030 IB_RAW_PACKET_CAP_CVLAN_STRIPPING; 1031 1032 if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) { 1033 max_tso = MLX5_CAP_ETH(mdev, max_lso_cap); 1034 if (max_tso) { 1035 resp.tso_caps.max_tso = 1 << max_tso; 1036 resp.tso_caps.supported_qpts |= 1037 1 << IB_QPT_RAW_PACKET; 1038 resp.response_length += sizeof(resp.tso_caps); 1039 } 1040 } 1041 1042 if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) { 1043 resp.rss_caps.rx_hash_function = 1044 MLX5_RX_HASH_FUNC_TOEPLITZ; 1045 resp.rss_caps.rx_hash_fields_mask = 1046 MLX5_RX_HASH_SRC_IPV4 | 1047 MLX5_RX_HASH_DST_IPV4 | 1048 MLX5_RX_HASH_SRC_IPV6 | 1049 MLX5_RX_HASH_DST_IPV6 | 1050 MLX5_RX_HASH_SRC_PORT_TCP | 1051 MLX5_RX_HASH_DST_PORT_TCP | 1052 MLX5_RX_HASH_SRC_PORT_UDP | 1053 MLX5_RX_HASH_DST_PORT_UDP | 1054 MLX5_RX_HASH_INNER; 1055 resp.response_length += sizeof(resp.rss_caps); 1056 } 1057 } else { 1058 if (offsetofend(typeof(resp), tso_caps) <= uhw_outlen) 1059 resp.response_length += sizeof(resp.tso_caps); 1060 if (offsetofend(typeof(resp), rss_caps) <= uhw_outlen) 1061 resp.response_length += sizeof(resp.rss_caps); 1062 } 1063 1064 if (MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) { 1065 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM; 1066 props->kernel_cap_flags |= IBK_UD_TSO; 1067 } 1068 1069 if (MLX5_CAP_GEN(dev->mdev, rq_delay_drop) && 1070 MLX5_CAP_GEN(dev->mdev, general_notification_event) && 1071 raw_support) 1072 props->raw_packet_caps |= IB_RAW_PACKET_CAP_DELAY_DROP; 1073 1074 if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) && 1075 MLX5_CAP_IPOIB_ENHANCED(mdev, csum_cap)) 1076 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM; 1077 1078 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && 1079 MLX5_CAP_ETH(dev->mdev, scatter_fcs) && 1080 raw_support) { 1081 /* Legacy bit to support old userspace libraries */ 1082 props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS; 1083 props->raw_packet_caps |= IB_RAW_PACKET_CAP_SCATTER_FCS; 1084 } 1085 1086 if (MLX5_CAP_DEV_MEM(mdev, memic)) { 1087 props->max_dm_size = 1088 MLX5_CAP_DEV_MEM(mdev, max_memic_size); 1089 } 1090 1091 if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS)) 1092 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING; 1093 1094 if (MLX5_CAP_GEN(mdev, end_pad)) 1095 props->device_cap_flags |= IB_DEVICE_PCI_WRITE_END_PADDING; 1096 1097 props->vendor_part_id = mdev->pdev->device; 1098 props->hw_ver = mdev->pdev->revision; 1099 1100 props->max_mr_size = ~0ull; 1101 props->page_size_cap = ~(min_page_size - 1); 1102 props->max_qp = 1 << MLX5_CAP_GEN(mdev, log_max_qp); 1103 props->max_qp_wr = mlx5_ib_calc_max_qp_wr(dev); 1104 max_rq_sg = MLX5_CAP_GEN(mdev, max_wqe_sz_rq) / 1105 sizeof(struct mlx5_wqe_data_seg); 1106 max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512); 1107 max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) - 1108 sizeof(struct mlx5_wqe_raddr_seg)) / 1109 sizeof(struct mlx5_wqe_data_seg); 1110 props->max_send_sge = max_sq_sg; 1111 props->max_recv_sge = max_rq_sg; 1112 props->max_sge_rd = MLX5_MAX_SGE_RD; 1113 props->max_cq = 1 << MLX5_CAP_GEN(mdev, log_max_cq); 1114 props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1; 1115 props->max_mr = 1 << MLX5_CAP_GEN(mdev, log_max_mkey); 1116 props->max_pd = 1 << MLX5_CAP_GEN(mdev, log_max_pd); 1117 props->max_qp_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp); 1118 props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp); 1119 props->max_srq = 1 << MLX5_CAP_GEN(mdev, log_max_srq); 1120 props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1; 1121 props->local_ca_ack_delay = MLX5_CAP_GEN(mdev, local_ca_ack_delay); 1122 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp; 1123 props->max_srq_sge = max_rq_sg - 1; 1124 props->max_fast_reg_page_list_len = 1125 1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size); 1126 props->max_pi_fast_reg_page_list_len = 1127 props->max_fast_reg_page_list_len / 2; 1128 props->max_sgl_rd = 1129 MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance); 1130 get_atomic_caps_qp(dev, props); 1131 props->masked_atomic_cap = IB_ATOMIC_NONE; 1132 props->max_mcast_grp = 1 << MLX5_CAP_GEN(mdev, log_max_mcg); 1133 props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg); 1134 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach * 1135 props->max_mcast_grp; 1136 props->max_ah = INT_MAX; 1137 props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz); 1138 props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL; 1139 1140 if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) { 1141 if (dev->odp_caps.general_caps & IB_ODP_SUPPORT) 1142 props->kernel_cap_flags |= IBK_ON_DEMAND_PAGING; 1143 props->odp_caps = dev->odp_caps; 1144 if (!uhw) { 1145 /* ODP for kernel QPs is not implemented for receive 1146 * WQEs and SRQ WQEs 1147 */ 1148 props->odp_caps.per_transport_caps.rc_odp_caps &= 1149 ~(IB_ODP_SUPPORT_READ | 1150 IB_ODP_SUPPORT_SRQ_RECV); 1151 props->odp_caps.per_transport_caps.uc_odp_caps &= 1152 ~(IB_ODP_SUPPORT_READ | 1153 IB_ODP_SUPPORT_SRQ_RECV); 1154 props->odp_caps.per_transport_caps.ud_odp_caps &= 1155 ~(IB_ODP_SUPPORT_READ | 1156 IB_ODP_SUPPORT_SRQ_RECV); 1157 props->odp_caps.per_transport_caps.xrc_odp_caps &= 1158 ~(IB_ODP_SUPPORT_READ | 1159 IB_ODP_SUPPORT_SRQ_RECV); 1160 } 1161 } 1162 1163 if (mlx5_core_is_vf(mdev)) 1164 props->kernel_cap_flags |= IBK_VIRTUAL_FUNCTION; 1165 1166 if (mlx5_ib_port_link_layer(ibdev, 1) == 1167 IB_LINK_LAYER_ETHERNET && raw_support) { 1168 props->rss_caps.max_rwq_indirection_tables = 1169 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt); 1170 props->rss_caps.max_rwq_indirection_table_size = 1171 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size); 1172 props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET; 1173 props->max_wq_type_rq = 1174 1 << MLX5_CAP_GEN(dev->mdev, log_max_rq); 1175 } 1176 1177 if (MLX5_CAP_GEN(mdev, tag_matching)) { 1178 props->tm_caps.max_num_tags = 1179 (1 << MLX5_CAP_GEN(mdev, log_tag_matching_list_sz)) - 1; 1180 props->tm_caps.max_ops = 1181 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz); 1182 props->tm_caps.max_sge = MLX5_TM_MAX_SGE; 1183 } 1184 1185 if (MLX5_CAP_GEN(mdev, tag_matching) && 1186 MLX5_CAP_GEN(mdev, rndv_offload_rc)) { 1187 props->tm_caps.flags = IB_TM_CAP_RNDV_RC; 1188 props->tm_caps.max_rndv_hdr_size = MLX5_TM_MAX_RNDV_MSG_SIZE; 1189 } 1190 1191 if (MLX5_CAP_GEN(dev->mdev, cq_moderation)) { 1192 props->cq_caps.max_cq_moderation_count = 1193 MLX5_MAX_CQ_COUNT; 1194 props->cq_caps.max_cq_moderation_period = 1195 MLX5_MAX_CQ_PERIOD; 1196 } 1197 1198 if (offsetofend(typeof(resp), cqe_comp_caps) <= uhw_outlen) { 1199 resp.response_length += sizeof(resp.cqe_comp_caps); 1200 1201 if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) { 1202 resp.cqe_comp_caps.max_num = 1203 MLX5_CAP_GEN(dev->mdev, 1204 cqe_compression_max_num); 1205 1206 resp.cqe_comp_caps.supported_format = 1207 MLX5_IB_CQE_RES_FORMAT_HASH | 1208 MLX5_IB_CQE_RES_FORMAT_CSUM; 1209 1210 if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index)) 1211 resp.cqe_comp_caps.supported_format |= 1212 MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX; 1213 } 1214 } 1215 1216 if (offsetofend(typeof(resp), packet_pacing_caps) <= uhw_outlen && 1217 raw_support) { 1218 if (MLX5_CAP_QOS(mdev, packet_pacing) && 1219 MLX5_CAP_GEN(mdev, qos)) { 1220 resp.packet_pacing_caps.qp_rate_limit_max = 1221 MLX5_CAP_QOS(mdev, packet_pacing_max_rate); 1222 resp.packet_pacing_caps.qp_rate_limit_min = 1223 MLX5_CAP_QOS(mdev, packet_pacing_min_rate); 1224 resp.packet_pacing_caps.supported_qpts |= 1225 1 << IB_QPT_RAW_PACKET; 1226 if (MLX5_CAP_QOS(mdev, packet_pacing_burst_bound) && 1227 MLX5_CAP_QOS(mdev, packet_pacing_typical_size)) 1228 resp.packet_pacing_caps.cap_flags |= 1229 MLX5_IB_PP_SUPPORT_BURST; 1230 } 1231 resp.response_length += sizeof(resp.packet_pacing_caps); 1232 } 1233 1234 if (offsetofend(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes) <= 1235 uhw_outlen) { 1236 if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe)) 1237 resp.mlx5_ib_support_multi_pkt_send_wqes = 1238 MLX5_IB_ALLOW_MPW; 1239 1240 if (MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe)) 1241 resp.mlx5_ib_support_multi_pkt_send_wqes |= 1242 MLX5_IB_SUPPORT_EMPW; 1243 1244 resp.response_length += 1245 sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes); 1246 } 1247 1248 if (offsetofend(typeof(resp), flags) <= uhw_outlen) { 1249 resp.response_length += sizeof(resp.flags); 1250 1251 if (MLX5_CAP_GEN(mdev, cqe_compression_128)) 1252 resp.flags |= 1253 MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_COMP; 1254 1255 if (MLX5_CAP_GEN(mdev, cqe_128_always)) 1256 resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_CQE_128B_PAD; 1257 if (MLX5_CAP_GEN(mdev, qp_packet_based)) 1258 resp.flags |= 1259 MLX5_IB_QUERY_DEV_RESP_PACKET_BASED_CREDIT_MODE; 1260 1261 resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT; 1262 1263 if (MLX5_CAP_GEN_2(mdev, dp_ordering_force) && 1264 (MLX5_CAP_GEN(mdev, dp_ordering_ooo_all_xrc) || 1265 MLX5_CAP_GEN(mdev, dp_ordering_ooo_all_dc) || 1266 MLX5_CAP_GEN(mdev, dp_ordering_ooo_all_rc) || 1267 MLX5_CAP_GEN(mdev, dp_ordering_ooo_all_ud) || 1268 MLX5_CAP_GEN(mdev, dp_ordering_ooo_all_uc))) 1269 resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_OOO_DP; 1270 } 1271 1272 if (offsetofend(typeof(resp), sw_parsing_caps) <= uhw_outlen) { 1273 resp.response_length += sizeof(resp.sw_parsing_caps); 1274 if (MLX5_CAP_ETH(mdev, swp)) { 1275 resp.sw_parsing_caps.sw_parsing_offloads |= 1276 MLX5_IB_SW_PARSING; 1277 1278 if (MLX5_CAP_ETH(mdev, swp_csum)) 1279 resp.sw_parsing_caps.sw_parsing_offloads |= 1280 MLX5_IB_SW_PARSING_CSUM; 1281 1282 if (MLX5_CAP_ETH(mdev, swp_lso)) 1283 resp.sw_parsing_caps.sw_parsing_offloads |= 1284 MLX5_IB_SW_PARSING_LSO; 1285 1286 if (resp.sw_parsing_caps.sw_parsing_offloads) 1287 resp.sw_parsing_caps.supported_qpts = 1288 BIT(IB_QPT_RAW_PACKET); 1289 } 1290 } 1291 1292 if (offsetofend(typeof(resp), striding_rq_caps) <= uhw_outlen && 1293 raw_support) { 1294 resp.response_length += sizeof(resp.striding_rq_caps); 1295 if (MLX5_CAP_GEN(mdev, striding_rq)) { 1296 resp.striding_rq_caps.min_single_stride_log_num_of_bytes = 1297 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES; 1298 resp.striding_rq_caps.max_single_stride_log_num_of_bytes = 1299 MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES; 1300 if (MLX5_CAP_GEN(dev->mdev, ext_stride_num_range)) 1301 resp.striding_rq_caps 1302 .min_single_wqe_log_num_of_strides = 1303 MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES; 1304 else 1305 resp.striding_rq_caps 1306 .min_single_wqe_log_num_of_strides = 1307 MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES; 1308 resp.striding_rq_caps.max_single_wqe_log_num_of_strides = 1309 MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES; 1310 resp.striding_rq_caps.supported_qpts = 1311 BIT(IB_QPT_RAW_PACKET); 1312 } 1313 } 1314 1315 if (offsetofend(typeof(resp), tunnel_offloads_caps) <= uhw_outlen) { 1316 resp.response_length += sizeof(resp.tunnel_offloads_caps); 1317 if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan)) 1318 resp.tunnel_offloads_caps |= 1319 MLX5_IB_TUNNELED_OFFLOADS_VXLAN; 1320 if (MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx)) 1321 resp.tunnel_offloads_caps |= 1322 MLX5_IB_TUNNELED_OFFLOADS_GENEVE; 1323 if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre)) 1324 resp.tunnel_offloads_caps |= 1325 MLX5_IB_TUNNELED_OFFLOADS_GRE; 1326 if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre)) 1327 resp.tunnel_offloads_caps |= 1328 MLX5_IB_TUNNELED_OFFLOADS_MPLS_GRE; 1329 if (MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_udp)) 1330 resp.tunnel_offloads_caps |= 1331 MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP; 1332 } 1333 1334 if (offsetofend(typeof(resp), dci_streams_caps) <= uhw_outlen) { 1335 resp.response_length += sizeof(resp.dci_streams_caps); 1336 1337 resp.dci_streams_caps.max_log_num_concurent = 1338 MLX5_CAP_GEN(mdev, log_max_dci_stream_channels); 1339 1340 resp.dci_streams_caps.max_log_num_errored = 1341 MLX5_CAP_GEN(mdev, log_max_dci_errored_streams); 1342 } 1343 1344 if (offsetofend(typeof(resp), reserved) <= uhw_outlen) 1345 resp.response_length += sizeof(resp.reserved); 1346 1347 if (offsetofend(typeof(resp), reg_c0) <= uhw_outlen) { 1348 struct mlx5_eswitch *esw = mdev->priv.eswitch; 1349 1350 resp.response_length += sizeof(resp.reg_c0); 1351 1352 if (mlx5_eswitch_mode(mdev) == MLX5_ESWITCH_OFFLOADS && 1353 mlx5_eswitch_vport_match_metadata_enabled(esw)) 1354 fill_esw_mgr_reg_c0(mdev, &resp); 1355 } 1356 1357 if (uhw_outlen) { 1358 err = ib_copy_to_udata(uhw, &resp, resp.response_length); 1359 1360 if (err) 1361 return err; 1362 } 1363 1364 return 0; 1365 } 1366 1367 static void translate_active_width(struct ib_device *ibdev, u16 active_width, 1368 u8 *ib_width) 1369 { 1370 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1371 1372 if (active_width & MLX5_PTYS_WIDTH_1X) 1373 *ib_width = IB_WIDTH_1X; 1374 else if (active_width & MLX5_PTYS_WIDTH_2X) 1375 *ib_width = IB_WIDTH_2X; 1376 else if (active_width & MLX5_PTYS_WIDTH_4X) 1377 *ib_width = IB_WIDTH_4X; 1378 else if (active_width & MLX5_PTYS_WIDTH_8X) 1379 *ib_width = IB_WIDTH_8X; 1380 else if (active_width & MLX5_PTYS_WIDTH_12X) 1381 *ib_width = IB_WIDTH_12X; 1382 else { 1383 mlx5_ib_dbg(dev, "Invalid active_width %d, setting width to default value: 4x\n", 1384 active_width); 1385 *ib_width = IB_WIDTH_4X; 1386 } 1387 1388 return; 1389 } 1390 1391 static int mlx5_mtu_to_ib_mtu(int mtu) 1392 { 1393 switch (mtu) { 1394 case 256: return 1; 1395 case 512: return 2; 1396 case 1024: return 3; 1397 case 2048: return 4; 1398 case 4096: return 5; 1399 default: 1400 pr_warn("invalid mtu\n"); 1401 return -1; 1402 } 1403 } 1404 1405 enum ib_max_vl_num { 1406 __IB_MAX_VL_0 = 1, 1407 __IB_MAX_VL_0_1 = 2, 1408 __IB_MAX_VL_0_3 = 3, 1409 __IB_MAX_VL_0_7 = 4, 1410 __IB_MAX_VL_0_14 = 5, 1411 }; 1412 1413 enum mlx5_vl_hw_cap { 1414 MLX5_VL_HW_0 = 1, 1415 MLX5_VL_HW_0_1 = 2, 1416 MLX5_VL_HW_0_2 = 3, 1417 MLX5_VL_HW_0_3 = 4, 1418 MLX5_VL_HW_0_4 = 5, 1419 MLX5_VL_HW_0_5 = 6, 1420 MLX5_VL_HW_0_6 = 7, 1421 MLX5_VL_HW_0_7 = 8, 1422 MLX5_VL_HW_0_14 = 15 1423 }; 1424 1425 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap, 1426 u8 *max_vl_num) 1427 { 1428 switch (vl_hw_cap) { 1429 case MLX5_VL_HW_0: 1430 *max_vl_num = __IB_MAX_VL_0; 1431 break; 1432 case MLX5_VL_HW_0_1: 1433 *max_vl_num = __IB_MAX_VL_0_1; 1434 break; 1435 case MLX5_VL_HW_0_3: 1436 *max_vl_num = __IB_MAX_VL_0_3; 1437 break; 1438 case MLX5_VL_HW_0_7: 1439 *max_vl_num = __IB_MAX_VL_0_7; 1440 break; 1441 case MLX5_VL_HW_0_14: 1442 *max_vl_num = __IB_MAX_VL_0_14; 1443 break; 1444 1445 default: 1446 return -EINVAL; 1447 } 1448 1449 return 0; 1450 } 1451 1452 static int mlx5_query_hca_port(struct ib_device *ibdev, u32 port, 1453 struct ib_port_attr *props) 1454 { 1455 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1456 struct mlx5_core_dev *mdev = dev->mdev; 1457 struct mlx5_hca_vport_context *rep; 1458 u8 vl_hw_cap, plane_index = 0; 1459 u16 max_mtu; 1460 u16 oper_mtu; 1461 int err; 1462 u16 ib_link_width_oper; 1463 1464 rep = kzalloc_obj(*rep); 1465 if (!rep) { 1466 err = -ENOMEM; 1467 goto out; 1468 } 1469 1470 /* props being zeroed by the caller, avoid zeroing it here */ 1471 1472 if (ibdev->type == RDMA_DEVICE_TYPE_SMI) { 1473 plane_index = port; 1474 port = smi_to_native_portnum(dev, port); 1475 } 1476 1477 err = mlx5_query_hca_vport_context(mdev, 0, port, 0, rep); 1478 if (err) 1479 goto out; 1480 1481 props->lid = rep->lid; 1482 props->lmc = rep->lmc; 1483 props->sm_lid = rep->sm_lid; 1484 props->sm_sl = rep->sm_sl; 1485 props->state = rep->vport_state; 1486 props->phys_state = rep->port_physical_state; 1487 1488 props->port_cap_flags = rep->cap_mask1; 1489 if (dev->num_plane) { 1490 props->port_cap_flags |= IB_PORT_SM_DISABLED; 1491 props->port_cap_flags &= ~IB_PORT_SM; 1492 } else if (ibdev->type == RDMA_DEVICE_TYPE_SMI) 1493 props->port_cap_flags &= ~IB_PORT_CM_SUP; 1494 1495 props->gid_tbl_len = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size)); 1496 props->max_msg_sz = 1 << MLX5_CAP_GEN(mdev, log_max_msg); 1497 props->pkey_tbl_len = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size)); 1498 props->bad_pkey_cntr = rep->pkey_violation_counter; 1499 props->qkey_viol_cntr = rep->qkey_violation_counter; 1500 props->subnet_timeout = rep->subnet_timeout; 1501 props->init_type_reply = rep->init_type_reply; 1502 1503 if (props->port_cap_flags & IB_PORT_CAP_MASK2_SUP) 1504 props->port_cap_flags2 = rep->cap_mask2; 1505 1506 err = mlx5_query_ib_port_oper(mdev, &ib_link_width_oper, 1507 &props->active_speed, port, plane_index); 1508 if (err) 1509 goto out; 1510 1511 translate_active_width(ibdev, ib_link_width_oper, &props->active_width); 1512 1513 mlx5_query_port_max_mtu(mdev, &max_mtu, port); 1514 1515 props->max_mtu = mlx5_mtu_to_ib_mtu(max_mtu); 1516 1517 mlx5_query_port_oper_mtu(mdev, &oper_mtu, port); 1518 1519 props->active_mtu = mlx5_mtu_to_ib_mtu(oper_mtu); 1520 1521 err = mlx5_query_port_vl_hw_cap(mdev, &vl_hw_cap, port); 1522 if (err) 1523 goto out; 1524 1525 err = translate_max_vl_num(ibdev, vl_hw_cap, 1526 &props->max_vl_num); 1527 out: 1528 kfree(rep); 1529 return err; 1530 } 1531 1532 int mlx5_ib_query_port(struct ib_device *ibdev, u32 port, 1533 struct ib_port_attr *props) 1534 { 1535 unsigned int count; 1536 int ret; 1537 1538 switch (mlx5_get_vport_access_method(ibdev)) { 1539 case MLX5_VPORT_ACCESS_METHOD_MAD: 1540 ret = mlx5_query_mad_ifc_port(ibdev, port, props); 1541 break; 1542 1543 case MLX5_VPORT_ACCESS_METHOD_HCA: 1544 ret = mlx5_query_hca_port(ibdev, port, props); 1545 break; 1546 1547 case MLX5_VPORT_ACCESS_METHOD_NIC: 1548 ret = mlx5_query_port_roce(ibdev, port, props); 1549 break; 1550 1551 default: 1552 ret = -EINVAL; 1553 } 1554 1555 if (!ret && props) { 1556 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1557 struct mlx5_core_dev *mdev; 1558 bool put_mdev = true; 1559 1560 mdev = mlx5_ib_get_native_port_mdev(dev, port, NULL); 1561 if (!mdev) { 1562 /* If the port isn't affiliated yet query the master. 1563 * The master and slave will have the same values. 1564 */ 1565 mdev = dev->mdev; 1566 port = 1; 1567 put_mdev = false; 1568 } 1569 count = mlx5_core_reserved_gids_count(mdev); 1570 if (put_mdev) 1571 mlx5_ib_put_native_port_mdev(dev, port); 1572 props->gid_tbl_len -= count; 1573 } 1574 return ret; 1575 } 1576 1577 static int mlx5_ib_rep_query_port(struct ib_device *ibdev, u32 port, 1578 struct ib_port_attr *props) 1579 { 1580 return mlx5_query_port_roce(ibdev, port, props); 1581 } 1582 1583 static int mlx5_ib_rep_query_pkey(struct ib_device *ibdev, u32 port, u16 index, 1584 u16 *pkey) 1585 { 1586 /* Default special Pkey for representor device port as per the 1587 * IB specification 1.3 section 10.9.1.2. 1588 */ 1589 *pkey = 0xffff; 1590 return 0; 1591 } 1592 1593 static int mlx5_ib_query_port_speed_from_port(struct mlx5_ib_dev *dev, 1594 u32 port_num, u64 *speed) 1595 { 1596 struct ib_port_speed_info speed_info; 1597 struct ib_port_attr attr = {}; 1598 int err; 1599 1600 err = mlx5_ib_query_port(&dev->ib_dev, port_num, &attr); 1601 if (err) 1602 return err; 1603 1604 if (attr.state == IB_PORT_DOWN) { 1605 *speed = 0; 1606 return 0; 1607 } 1608 1609 err = ib_port_attr_to_speed_info(&attr, &speed_info); 1610 if (err) 1611 return err; 1612 1613 *speed = speed_info.rate; 1614 return 0; 1615 } 1616 1617 static int mlx5_ib_query_port_speed_from_vport(struct mlx5_core_dev *mdev, 1618 u8 op_mod, u16 vport, 1619 u8 other_vport, u64 *speed, 1620 struct mlx5_ib_dev *dev, 1621 u32 port_num) 1622 { 1623 u32 max_tx_speed; 1624 int err; 1625 1626 err = mlx5_query_vport_max_tx_speed(mdev, op_mod, vport, other_vport, 1627 &max_tx_speed); 1628 if (err) 1629 return err; 1630 1631 if (max_tx_speed == 0) 1632 /* Value 0 indicates field not supported, fallback */ 1633 return mlx5_ib_query_port_speed_from_port(dev, port_num, 1634 speed); 1635 1636 *speed = max_tx_speed; 1637 return 0; 1638 } 1639 1640 static int mlx5_ib_query_port_speed_from_bond(struct mlx5_ib_dev *dev, 1641 u32 port_num, u64 *speed) 1642 { 1643 struct mlx5_core_dev *mdev = dev->mdev; 1644 u32 bond_speed; 1645 int err; 1646 1647 err = mlx5_lag_query_bond_speed(mdev, &bond_speed); 1648 if (err) 1649 return err; 1650 1651 *speed = bond_speed / MLX5_MAX_TX_SPEED_UNIT; 1652 1653 return 0; 1654 } 1655 1656 static int mlx5_ib_query_port_speed_non_rep(struct mlx5_ib_dev *dev, 1657 u32 port_num, u64 *speed) 1658 { 1659 u16 op_mod = MLX5_VPORT_STATE_OP_MOD_VNIC_VPORT; 1660 1661 if (mlx5_lag_is_roce(dev->mdev)) 1662 return mlx5_ib_query_port_speed_from_bond(dev, port_num, 1663 speed); 1664 1665 return mlx5_ib_query_port_speed_from_vport(dev->mdev, op_mod, 0, false, 1666 speed, dev, port_num); 1667 } 1668 1669 static int mlx5_ib_query_port_speed_rep(struct mlx5_ib_dev *dev, u32 port_num, 1670 u64 *speed) 1671 { 1672 struct mlx5_eswitch_rep *rep; 1673 struct mlx5_core_dev *mdev; 1674 u16 op_mod; 1675 1676 if (!dev->port[port_num - 1].rep) { 1677 mlx5_ib_warn(dev, "Representor doesn't exist for port %u\n", 1678 port_num); 1679 return -EINVAL; 1680 } 1681 1682 rep = dev->port[port_num - 1].rep; 1683 mdev = mlx5_eswitch_get_core_dev(rep->esw); 1684 if (!mdev) 1685 return -ENODEV; 1686 1687 if (rep->vport == MLX5_VPORT_UPLINK) { 1688 if (mlx5_lag_is_sriov(mdev)) 1689 return mlx5_ib_query_port_speed_from_bond(dev, 1690 port_num, 1691 speed); 1692 1693 return mlx5_ib_query_port_speed_from_port(dev, port_num, 1694 speed); 1695 } 1696 1697 op_mod = MLX5_VPORT_STATE_OP_MOD_ESW_VPORT; 1698 return mlx5_ib_query_port_speed_from_vport(dev->mdev, op_mod, 1699 rep->vport, true, speed, dev, 1700 port_num); 1701 } 1702 1703 int mlx5_ib_query_port_speed(struct ib_device *ibdev, u32 port_num, u64 *speed) 1704 { 1705 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1706 1707 if (mlx5_ib_port_link_layer(ibdev, port_num) == 1708 IB_LINK_LAYER_INFINIBAND || mlx5_core_mp_enabled(dev->mdev)) 1709 return mlx5_ib_query_port_speed_from_port(dev, port_num, speed); 1710 else if (!dev->is_rep) 1711 return mlx5_ib_query_port_speed_non_rep(dev, port_num, speed); 1712 else 1713 return mlx5_ib_query_port_speed_rep(dev, port_num, speed); 1714 } 1715 1716 static int mlx5_ib_query_gid(struct ib_device *ibdev, u32 port, int index, 1717 union ib_gid *gid) 1718 { 1719 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1720 struct mlx5_core_dev *mdev = dev->mdev; 1721 1722 switch (mlx5_get_vport_access_method(ibdev)) { 1723 case MLX5_VPORT_ACCESS_METHOD_MAD: 1724 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid); 1725 1726 case MLX5_VPORT_ACCESS_METHOD_HCA: 1727 return mlx5_query_hca_vport_gid(mdev, 0, port, 0, index, gid); 1728 1729 default: 1730 return -EINVAL; 1731 } 1732 1733 } 1734 1735 static int mlx5_query_hca_nic_pkey(struct ib_device *ibdev, u32 port, 1736 u16 index, u16 *pkey) 1737 { 1738 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1739 struct mlx5_core_dev *mdev; 1740 bool put_mdev = true; 1741 u32 mdev_port_num; 1742 int err; 1743 1744 mdev = mlx5_ib_get_native_port_mdev(dev, port, &mdev_port_num); 1745 if (!mdev) { 1746 /* The port isn't affiliated yet, get the PKey from the master 1747 * port. For RoCE the PKey tables will be the same. 1748 */ 1749 put_mdev = false; 1750 mdev = dev->mdev; 1751 mdev_port_num = 1; 1752 } 1753 1754 err = mlx5_query_hca_vport_pkey(mdev, 0, mdev_port_num, 0, 1755 index, pkey); 1756 if (put_mdev) 1757 mlx5_ib_put_native_port_mdev(dev, port); 1758 1759 return err; 1760 } 1761 1762 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index, 1763 u16 *pkey) 1764 { 1765 switch (mlx5_get_vport_access_method(ibdev)) { 1766 case MLX5_VPORT_ACCESS_METHOD_MAD: 1767 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey); 1768 1769 case MLX5_VPORT_ACCESS_METHOD_HCA: 1770 case MLX5_VPORT_ACCESS_METHOD_NIC: 1771 return mlx5_query_hca_nic_pkey(ibdev, port, index, pkey); 1772 default: 1773 return -EINVAL; 1774 } 1775 } 1776 1777 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask, 1778 struct ib_device_modify *props) 1779 { 1780 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1781 struct mlx5_reg_node_desc in; 1782 struct mlx5_reg_node_desc out; 1783 int err; 1784 1785 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC) 1786 return -EOPNOTSUPP; 1787 1788 if (!(mask & IB_DEVICE_MODIFY_NODE_DESC)) 1789 return 0; 1790 1791 /* 1792 * If possible, pass node desc to FW, so it can generate 1793 * a 144 trap. If cmd fails, just ignore. 1794 */ 1795 memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX); 1796 err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out, 1797 sizeof(out), MLX5_REG_NODE_DESC, 0, 1); 1798 if (err) 1799 return err; 1800 1801 memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX); 1802 1803 return err; 1804 } 1805 1806 static int set_port_caps_atomic(struct mlx5_ib_dev *dev, u32 port_num, u32 mask, 1807 u32 value) 1808 { 1809 struct mlx5_hca_vport_context ctx = {}; 1810 struct mlx5_core_dev *mdev; 1811 u32 mdev_port_num; 1812 int err; 1813 1814 mdev = mlx5_ib_get_native_port_mdev(dev, port_num, &mdev_port_num); 1815 if (!mdev) 1816 return -ENODEV; 1817 1818 err = mlx5_query_hca_vport_context(mdev, 0, mdev_port_num, 0, &ctx); 1819 if (err) 1820 goto out; 1821 1822 if (~ctx.cap_mask1_perm & mask) { 1823 mlx5_ib_warn(dev, "trying to change bitmask 0x%X but change supported 0x%X\n", 1824 mask, ctx.cap_mask1_perm); 1825 err = -EINVAL; 1826 goto out; 1827 } 1828 1829 ctx.cap_mask1 = value; 1830 ctx.cap_mask1_perm = mask; 1831 err = mlx5_core_modify_hca_vport_context(mdev, 0, mdev_port_num, 1832 0, &ctx); 1833 1834 out: 1835 mlx5_ib_put_native_port_mdev(dev, port_num); 1836 1837 return err; 1838 } 1839 1840 static int mlx5_ib_modify_port(struct ib_device *ibdev, u32 port, int mask, 1841 struct ib_port_modify *props) 1842 { 1843 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1844 struct ib_port_attr attr; 1845 u32 tmp; 1846 int err; 1847 u32 change_mask; 1848 u32 value; 1849 bool is_ib = (mlx5_ib_port_link_layer(ibdev, port) == 1850 IB_LINK_LAYER_INFINIBAND); 1851 1852 /* CM layer calls ib_modify_port() regardless of the link layer. For 1853 * Ethernet ports, qkey violation and Port capabilities are meaningless. 1854 */ 1855 if (!is_ib) 1856 return 0; 1857 1858 if (MLX5_CAP_GEN(dev->mdev, ib_virt) && is_ib) { 1859 change_mask = props->clr_port_cap_mask | props->set_port_cap_mask; 1860 value = ~props->clr_port_cap_mask | props->set_port_cap_mask; 1861 return set_port_caps_atomic(dev, port, change_mask, value); 1862 } 1863 1864 mutex_lock(&dev->cap_mask_mutex); 1865 1866 err = ib_query_port(ibdev, port, &attr); 1867 if (err) 1868 goto out; 1869 1870 tmp = (attr.port_cap_flags | props->set_port_cap_mask) & 1871 ~props->clr_port_cap_mask; 1872 1873 err = mlx5_set_port_caps(dev->mdev, port, tmp); 1874 1875 out: 1876 mutex_unlock(&dev->cap_mask_mutex); 1877 return err; 1878 } 1879 1880 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps) 1881 { 1882 mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n", 1883 caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n"); 1884 } 1885 1886 static u16 calc_dynamic_bfregs(int uars_per_sys_page) 1887 { 1888 /* Large page with non 4k uar support might limit the dynamic size */ 1889 if (uars_per_sys_page == 1 && PAGE_SIZE > 4096) 1890 return MLX5_MIN_DYN_BFREGS; 1891 1892 return MLX5_MAX_DYN_BFREGS; 1893 } 1894 1895 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k, 1896 struct mlx5_ib_alloc_ucontext_req_v2 *req, 1897 struct mlx5_bfreg_info *bfregi) 1898 { 1899 int uars_per_sys_page; 1900 int bfregs_per_sys_page; 1901 int ref_bfregs = req->total_num_bfregs; 1902 1903 if (req->total_num_bfregs == 0) 1904 return -EINVAL; 1905 1906 BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE); 1907 BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE); 1908 1909 if (req->total_num_bfregs > MLX5_MAX_BFREGS) 1910 return -ENOMEM; 1911 1912 uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k); 1913 bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR; 1914 /* This holds the required static allocation asked by the user */ 1915 req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page); 1916 if (req->num_low_latency_bfregs > req->total_num_bfregs - 1) 1917 return -EINVAL; 1918 1919 bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page; 1920 bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page); 1921 bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs; 1922 bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page; 1923 1924 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", 1925 MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no", 1926 lib_uar_4k ? "yes" : "no", ref_bfregs, 1927 req->total_num_bfregs, bfregi->total_num_bfregs, 1928 bfregi->num_sys_pages); 1929 1930 return 0; 1931 } 1932 1933 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context) 1934 { 1935 struct mlx5_bfreg_info *bfregi; 1936 int err; 1937 int i; 1938 1939 bfregi = &context->bfregi; 1940 for (i = 0; i < bfregi->num_static_sys_pages; i++) { 1941 err = mlx5_cmd_uar_alloc(dev->mdev, &bfregi->sys_pages[i], 1942 context->devx_uid); 1943 if (err) 1944 goto error; 1945 1946 mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]); 1947 } 1948 1949 for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++) 1950 bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX; 1951 1952 return 0; 1953 1954 error: 1955 for (--i; i >= 0; i--) 1956 if (mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i], 1957 context->devx_uid)) 1958 mlx5_ib_warn(dev, "failed to free uar %d\n", i); 1959 1960 return err; 1961 } 1962 1963 static void deallocate_uars(struct mlx5_ib_dev *dev, 1964 struct mlx5_ib_ucontext *context) 1965 { 1966 struct mlx5_bfreg_info *bfregi; 1967 int i; 1968 1969 bfregi = &context->bfregi; 1970 for (i = 0; i < bfregi->num_sys_pages; i++) 1971 if (i < bfregi->num_static_sys_pages || 1972 bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX) 1973 mlx5_cmd_uar_dealloc(dev->mdev, bfregi->sys_pages[i], 1974 context->devx_uid); 1975 } 1976 1977 static int mlx5_ib_enable_lb_mp(struct mlx5_core_dev *master, 1978 struct mlx5_core_dev *slave, 1979 struct mlx5_ib_lb_state *lb_state) 1980 { 1981 int err; 1982 1983 err = mlx5_nic_vport_update_local_lb(master, true); 1984 if (err) 1985 return err; 1986 1987 err = mlx5_nic_vport_update_local_lb(slave, true); 1988 if (err) 1989 goto out; 1990 1991 lb_state->force_enable = true; 1992 return 0; 1993 1994 out: 1995 mlx5_nic_vport_update_local_lb(master, false); 1996 return err; 1997 } 1998 1999 static void mlx5_ib_disable_lb_mp(struct mlx5_core_dev *master, 2000 struct mlx5_core_dev *slave, 2001 struct mlx5_ib_lb_state *lb_state) 2002 { 2003 mlx5_nic_vport_update_local_lb(slave, false); 2004 mlx5_nic_vport_update_local_lb(master, false); 2005 2006 lb_state->force_enable = false; 2007 } 2008 2009 int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp) 2010 { 2011 int err = 0; 2012 2013 if (dev->lb.force_enable) 2014 return 0; 2015 2016 mutex_lock(&dev->lb.mutex); 2017 if (td) 2018 dev->lb.user_td++; 2019 if (qp) 2020 dev->lb.qps++; 2021 2022 if (dev->lb.user_td == 2 || 2023 dev->lb.qps == 1) { 2024 if (!dev->lb.enabled) { 2025 err = mlx5_nic_vport_update_local_lb(dev->mdev, true); 2026 dev->lb.enabled = true; 2027 } 2028 } 2029 2030 mutex_unlock(&dev->lb.mutex); 2031 2032 return err; 2033 } 2034 2035 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp) 2036 { 2037 if (dev->lb.force_enable) 2038 return; 2039 2040 mutex_lock(&dev->lb.mutex); 2041 if (td) 2042 dev->lb.user_td--; 2043 if (qp) 2044 dev->lb.qps--; 2045 2046 if (dev->lb.user_td == 1 && 2047 dev->lb.qps == 0) { 2048 if (dev->lb.enabled) { 2049 mlx5_nic_vport_update_local_lb(dev->mdev, false); 2050 dev->lb.enabled = false; 2051 } 2052 } 2053 2054 mutex_unlock(&dev->lb.mutex); 2055 } 2056 2057 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn, 2058 u16 uid) 2059 { 2060 int err; 2061 2062 if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) 2063 return 0; 2064 2065 err = mlx5_cmd_alloc_transport_domain(dev->mdev, tdn, uid); 2066 if (err) 2067 return err; 2068 2069 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) || 2070 (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) && 2071 !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc))) 2072 return err; 2073 2074 return mlx5_ib_enable_lb(dev, true, false); 2075 } 2076 2077 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn, 2078 u16 uid) 2079 { 2080 if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) 2081 return; 2082 2083 mlx5_cmd_dealloc_transport_domain(dev->mdev, tdn, uid); 2084 2085 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) || 2086 (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) && 2087 !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc))) 2088 return; 2089 2090 mlx5_ib_disable_lb(dev, true, false); 2091 } 2092 2093 static int set_ucontext_resp(struct ib_ucontext *uctx, 2094 struct mlx5_ib_alloc_ucontext_resp *resp) 2095 { 2096 struct ib_device *ibdev = uctx->device; 2097 struct mlx5_ib_dev *dev = to_mdev(ibdev); 2098 struct mlx5_ib_ucontext *context = to_mucontext(uctx); 2099 struct mlx5_bfreg_info *bfregi = &context->bfregi; 2100 2101 if (MLX5_CAP_GEN(dev->mdev, dump_fill_mkey)) { 2102 resp->dump_fill_mkey = dev->mkeys.dump_fill_mkey; 2103 resp->comp_mask |= 2104 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY; 2105 } 2106 2107 resp->qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp); 2108 if (mlx5_wc_support_get(dev->mdev)) 2109 resp->bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, 2110 log_bf_reg_size); 2111 resp->cache_line_size = cache_line_size(); 2112 resp->max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq); 2113 resp->max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq); 2114 resp->max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz); 2115 resp->max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz); 2116 resp->max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz); 2117 resp->cqe_version = context->cqe_version; 2118 resp->log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ? 2119 MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT; 2120 resp->num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? 2121 MLX5_CAP_GEN(dev->mdev, 2122 num_of_uars_per_page) : 1; 2123 resp->tot_bfregs = bfregi->lib_uar_dyn ? 0 : 2124 bfregi->total_num_bfregs - bfregi->num_dyn_bfregs; 2125 resp->num_ports = dev->num_ports; 2126 resp->cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE | 2127 MLX5_USER_CMDS_SUPP_UHW_CREATE_AH; 2128 2129 if (mlx5_ib_port_link_layer(ibdev, 1) == IB_LINK_LAYER_ETHERNET) { 2130 mlx5_query_min_inline(dev->mdev, &resp->eth_min_inline); 2131 resp->eth_min_inline++; 2132 } 2133 2134 if (dev->mdev->clock_info) 2135 resp->clock_info_versions = BIT(MLX5_IB_CLOCK_INFO_V1); 2136 2137 /* 2138 * We don't want to expose information from the PCI bar that is located 2139 * after 4096 bytes, so if the arch only supports larger pages, let's 2140 * pretend we don't support reading the HCA's core clock. This is also 2141 * forced by mmap function. 2142 */ 2143 if (PAGE_SIZE <= 4096) { 2144 resp->comp_mask |= 2145 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET; 2146 resp->hca_core_clock_offset = 2147 offsetof(struct mlx5_init_seg, 2148 internal_timer_h) % PAGE_SIZE; 2149 } 2150 2151 if (MLX5_CAP_GEN(dev->mdev, ece_support)) 2152 resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE; 2153 2154 if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) && 2155 rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) && 2156 rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format))) 2157 resp->comp_mask |= 2158 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS; 2159 2160 resp->num_dyn_bfregs = bfregi->num_dyn_bfregs; 2161 2162 if (MLX5_CAP_GEN(dev->mdev, drain_sigerr)) 2163 resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_SQD2RTS; 2164 2165 resp->comp_mask |= 2166 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_MKEY_UPDATE_TAG; 2167 2168 return 0; 2169 } 2170 2171 static bool uctx_rdma_ctrl_is_enabled(u64 enabled_caps) 2172 { 2173 return UCAP_ENABLED(enabled_caps, RDMA_UCAP_MLX5_CTRL_LOCAL) || 2174 UCAP_ENABLED(enabled_caps, RDMA_UCAP_MLX5_CTRL_OTHER_VHCA); 2175 } 2176 2177 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx, 2178 struct ib_udata *udata) 2179 { 2180 struct ib_device *ibdev = uctx->device; 2181 struct mlx5_ib_dev *dev = to_mdev(ibdev); 2182 struct mlx5_ib_alloc_ucontext_req_v2 req; 2183 struct mlx5_ib_alloc_ucontext_resp resp = {}; 2184 struct mlx5_ib_ucontext *context = to_mucontext(uctx); 2185 struct mlx5_bfreg_info *bfregi; 2186 int ver; 2187 int err; 2188 size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2, 2189 max_cqe_version); 2190 bool lib_uar_4k; 2191 bool lib_uar_dyn; 2192 2193 if (!dev->ib_active) 2194 return -EAGAIN; 2195 2196 if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req)) 2197 ver = 0; 2198 else if (udata->inlen >= min_req_v2) 2199 ver = 2; 2200 else 2201 return -EINVAL; 2202 2203 err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req))); 2204 if (err) 2205 return err; 2206 2207 if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX) 2208 return -EOPNOTSUPP; 2209 2210 if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2) 2211 return -EOPNOTSUPP; 2212 2213 req.total_num_bfregs = ALIGN(req.total_num_bfregs, 2214 MLX5_NON_FP_BFREGS_PER_UAR); 2215 if (req.num_low_latency_bfregs > req.total_num_bfregs - 1) 2216 return -EINVAL; 2217 2218 if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) { 2219 err = mlx5_ib_devx_create(dev, true, uctx->enabled_caps); 2220 if (err < 0) 2221 goto out_ctx; 2222 context->devx_uid = err; 2223 2224 if (uctx_rdma_ctrl_is_enabled(uctx->enabled_caps)) { 2225 err = mlx5_cmd_add_privileged_uid(dev->mdev, 2226 context->devx_uid); 2227 if (err) 2228 goto out_devx; 2229 } 2230 } 2231 2232 lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR; 2233 lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR; 2234 bfregi = &context->bfregi; 2235 2236 if (lib_uar_dyn) { 2237 bfregi->lib_uar_dyn = lib_uar_dyn; 2238 goto uar_done; 2239 } 2240 2241 /* updates req->total_num_bfregs */ 2242 err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi); 2243 if (err) 2244 goto out_ucap; 2245 2246 mutex_init(&bfregi->lock); 2247 bfregi->lib_uar_4k = lib_uar_4k; 2248 bfregi->count = kzalloc_objs(*bfregi->count, bfregi->total_num_bfregs); 2249 if (!bfregi->count) { 2250 err = -ENOMEM; 2251 goto out_ucap; 2252 } 2253 2254 bfregi->sys_pages = 2255 kzalloc_objs(*bfregi->sys_pages, bfregi->num_sys_pages); 2256 if (!bfregi->sys_pages) { 2257 err = -ENOMEM; 2258 goto out_count; 2259 } 2260 2261 err = allocate_uars(dev, context); 2262 if (err) 2263 goto out_sys_pages; 2264 2265 uar_done: 2266 err = mlx5_ib_alloc_transport_domain(dev, &context->tdn, 2267 context->devx_uid); 2268 if (err) 2269 goto out_uars; 2270 2271 INIT_LIST_HEAD(&context->db_page_list); 2272 mutex_init(&context->db_page_mutex); 2273 2274 context->cqe_version = min_t(__u8, 2275 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version), 2276 req.max_cqe_version); 2277 2278 err = set_ucontext_resp(uctx, &resp); 2279 if (err) 2280 goto out_mdev; 2281 2282 resp.response_length = min(udata->outlen, sizeof(resp)); 2283 err = ib_copy_to_udata(udata, &resp, resp.response_length); 2284 if (err) 2285 goto out_mdev; 2286 2287 bfregi->ver = ver; 2288 bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs; 2289 context->lib_caps = req.lib_caps; 2290 print_lib_caps(dev, context->lib_caps); 2291 2292 if (mlx5_ib_lag_should_assign_affinity(dev)) { 2293 u32 port = mlx5_core_native_port_num(dev->mdev) - 1; 2294 2295 atomic_set(&context->tx_port_affinity, 2296 atomic_add_return( 2297 1, &dev->port[port].roce.tx_port_affinity)); 2298 } 2299 2300 return 0; 2301 2302 out_mdev: 2303 mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid); 2304 2305 out_uars: 2306 deallocate_uars(dev, context); 2307 2308 out_sys_pages: 2309 kfree(bfregi->sys_pages); 2310 2311 out_count: 2312 kfree(bfregi->count); 2313 2314 out_ucap: 2315 if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX && 2316 uctx_rdma_ctrl_is_enabled(uctx->enabled_caps)) 2317 mlx5_cmd_remove_privileged_uid(dev->mdev, context->devx_uid); 2318 2319 out_devx: 2320 if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) 2321 mlx5_ib_devx_destroy(dev, context->devx_uid); 2322 2323 out_ctx: 2324 return err; 2325 } 2326 2327 static int mlx5_ib_query_ucontext(struct ib_ucontext *ibcontext, 2328 struct uverbs_attr_bundle *attrs) 2329 { 2330 struct mlx5_ib_alloc_ucontext_resp uctx_resp = {}; 2331 int ret; 2332 2333 ret = set_ucontext_resp(ibcontext, &uctx_resp); 2334 if (ret) 2335 return ret; 2336 2337 uctx_resp.response_length = 2338 min_t(size_t, 2339 uverbs_attr_get_len(attrs, 2340 MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX), 2341 sizeof(uctx_resp)); 2342 2343 ret = uverbs_copy_to_struct_or_zero(attrs, 2344 MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX, 2345 &uctx_resp, 2346 sizeof(uctx_resp)); 2347 return ret; 2348 } 2349 2350 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext) 2351 { 2352 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext); 2353 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device); 2354 struct mlx5_bfreg_info *bfregi; 2355 2356 bfregi = &context->bfregi; 2357 mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid); 2358 2359 deallocate_uars(dev, context); 2360 kfree(bfregi->sys_pages); 2361 kfree(bfregi->count); 2362 2363 if (context->devx_uid) { 2364 if (uctx_rdma_ctrl_is_enabled(ibcontext->enabled_caps)) 2365 mlx5_cmd_remove_privileged_uid(dev->mdev, 2366 context->devx_uid); 2367 mlx5_ib_devx_destroy(dev, context->devx_uid); 2368 } 2369 } 2370 2371 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, 2372 int uar_idx) 2373 { 2374 int fw_uars_per_page; 2375 2376 fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1; 2377 2378 return (dev->mdev->bar_addr >> PAGE_SHIFT) + uar_idx / fw_uars_per_page; 2379 } 2380 2381 static u64 uar_index2paddress(struct mlx5_ib_dev *dev, 2382 int uar_idx) 2383 { 2384 unsigned int fw_uars_per_page; 2385 2386 fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? 2387 MLX5_UARS_IN_PAGE : 1; 2388 2389 return (dev->mdev->bar_addr + (uar_idx / fw_uars_per_page) * PAGE_SIZE); 2390 } 2391 2392 static int get_command(unsigned long offset) 2393 { 2394 return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK; 2395 } 2396 2397 static int get_arg(unsigned long offset) 2398 { 2399 return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1); 2400 } 2401 2402 static int get_index(unsigned long offset) 2403 { 2404 return get_arg(offset); 2405 } 2406 2407 /* Index resides in an extra byte to enable larger values than 255 */ 2408 static int get_extended_index(unsigned long offset) 2409 { 2410 return get_arg(offset) | ((offset >> 16) & 0xff) << 8; 2411 } 2412 2413 2414 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext) 2415 { 2416 } 2417 2418 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd) 2419 { 2420 switch (cmd) { 2421 case MLX5_IB_MMAP_WC_PAGE: 2422 return "WC"; 2423 case MLX5_IB_MMAP_REGULAR_PAGE: 2424 return "best effort WC"; 2425 case MLX5_IB_MMAP_NC_PAGE: 2426 return "NC"; 2427 case MLX5_IB_MMAP_DEVICE_MEM: 2428 return "Device Memory"; 2429 default: 2430 return "Unknown"; 2431 } 2432 } 2433 2434 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev, 2435 struct vm_area_struct *vma, 2436 struct mlx5_ib_ucontext *context) 2437 { 2438 if ((vma->vm_end - vma->vm_start != PAGE_SIZE) || 2439 !(vma->vm_flags & VM_SHARED)) 2440 return -EINVAL; 2441 2442 if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1) 2443 return -EOPNOTSUPP; 2444 2445 if (vma->vm_flags & (VM_WRITE | VM_EXEC)) 2446 return -EPERM; 2447 vm_flags_clear(vma, VM_MAYWRITE); 2448 2449 if (!dev->mdev->clock_info) 2450 return -EOPNOTSUPP; 2451 2452 return vm_insert_page(vma, vma->vm_start, 2453 virt_to_page(dev->mdev->clock_info)); 2454 } 2455 2456 static int phys_addr_to_bar(struct pci_dev *pdev, phys_addr_t pa) 2457 { 2458 resource_size_t start, end; 2459 int bar; 2460 2461 for (bar = 0; bar < PCI_STD_NUM_BARS; bar++) { 2462 /* Skip BARs not present or not memory-mapped */ 2463 if (!(pci_resource_flags(pdev, bar) & IORESOURCE_MEM)) 2464 continue; 2465 2466 start = pci_resource_start(pdev, bar); 2467 end = pci_resource_end(pdev, bar); 2468 2469 if (!start || !end) 2470 continue; 2471 2472 if (pa >= start && pa <= end) 2473 return bar; 2474 } 2475 2476 return -1; 2477 } 2478 2479 static int mlx5_ib_mmap_get_pfns(struct rdma_user_mmap_entry *entry, 2480 struct phys_vec *phys_vec, 2481 struct p2pdma_provider **provider) 2482 { 2483 struct mlx5_user_mmap_entry *mentry = to_mmmap(entry); 2484 struct pci_dev *pdev = to_mdev(entry->ucontext->device)->mdev->pdev; 2485 int bar; 2486 2487 phys_vec->paddr = mentry->address; 2488 phys_vec->len = entry->npages * PAGE_SIZE; 2489 2490 bar = phys_addr_to_bar(pdev, phys_vec->paddr); 2491 if (bar < 0) 2492 return -EINVAL; 2493 2494 *provider = pcim_p2pdma_provider(pdev, bar); 2495 /* If the kernel was not compiled with CONFIG_PCI_P2PDMA the 2496 * functionality is not supported. 2497 */ 2498 if (!*provider) 2499 return -EOPNOTSUPP; 2500 2501 return 0; 2502 } 2503 2504 static struct rdma_user_mmap_entry * 2505 mlx5_ib_pgoff_to_mmap_entry(struct ib_ucontext *ucontext, off_t pg_off) 2506 { 2507 unsigned long entry_pgoff; 2508 unsigned long idx; 2509 u8 command; 2510 2511 pg_off = pg_off >> PAGE_SHIFT; 2512 command = get_command(pg_off); 2513 idx = get_extended_index(pg_off); 2514 2515 entry_pgoff = command << 16 | idx; 2516 2517 return rdma_user_mmap_entry_get_pgoff(ucontext, entry_pgoff); 2518 } 2519 2520 static void mlx5_ib_free_var_mmap_entry(struct mlx5_user_mmap_entry *mentry, 2521 struct mlx5_var_region *var_region) 2522 { 2523 mutex_lock(&var_region->bitmap_lock); 2524 clear_bit(mentry->page_idx, var_region->bitmap); 2525 mutex_unlock(&var_region->bitmap_lock); 2526 kfree(mentry); 2527 } 2528 2529 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry) 2530 { 2531 struct mlx5_user_mmap_entry *mentry = to_mmmap(entry); 2532 struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device); 2533 struct mlx5_var_table *var_table = &dev->var_table; 2534 struct mlx5_ib_ucontext *context = to_mucontext(entry->ucontext); 2535 struct mlx5_var_region *var_region; 2536 2537 switch (mentry->mmap_flag) { 2538 case MLX5_IB_MMAP_TYPE_MEMIC: 2539 case MLX5_IB_MMAP_TYPE_MEMIC_OP: 2540 mlx5_ib_dm_mmap_free(dev, mentry); 2541 break; 2542 case MLX5_IB_MMAP_TYPE_VAR: 2543 var_region = &var_table->var_region; 2544 mlx5_ib_free_var_mmap_entry(mentry, var_region); 2545 break; 2546 case MLX5_IB_MMAP_TYPE_TLP_VAR: 2547 var_region = &var_table->tlp_var_region; 2548 mlx5_ib_free_var_mmap_entry(mentry, var_region); 2549 break; 2550 case MLX5_IB_MMAP_TYPE_UAR_WC: 2551 case MLX5_IB_MMAP_TYPE_UAR_NC: 2552 mlx5_cmd_uar_dealloc(dev->mdev, mentry->page_idx, 2553 context->devx_uid); 2554 kfree(mentry); 2555 break; 2556 default: 2557 WARN_ON(true); 2558 } 2559 } 2560 2561 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd, 2562 struct vm_area_struct *vma, 2563 struct mlx5_ib_ucontext *context) 2564 { 2565 struct mlx5_bfreg_info *bfregi = &context->bfregi; 2566 int err; 2567 unsigned long idx; 2568 phys_addr_t pfn; 2569 pgprot_t prot; 2570 u32 bfreg_dyn_idx = 0; 2571 u32 uar_index; 2572 int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC); 2573 int max_valid_idx = dyn_uar ? bfregi->num_sys_pages : 2574 bfregi->num_static_sys_pages; 2575 2576 if (bfregi->lib_uar_dyn) 2577 return -EINVAL; 2578 2579 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 2580 return -EINVAL; 2581 2582 if (dyn_uar) 2583 idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages; 2584 else 2585 idx = get_index(vma->vm_pgoff); 2586 2587 if (idx >= max_valid_idx) { 2588 mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n", 2589 idx, max_valid_idx); 2590 return -EINVAL; 2591 } 2592 2593 switch (cmd) { 2594 case MLX5_IB_MMAP_WC_PAGE: 2595 case MLX5_IB_MMAP_ALLOC_WC: 2596 case MLX5_IB_MMAP_REGULAR_PAGE: 2597 /* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */ 2598 prot = pgprot_writecombine(vma->vm_page_prot); 2599 break; 2600 case MLX5_IB_MMAP_NC_PAGE: 2601 prot = pgprot_noncached(vma->vm_page_prot); 2602 break; 2603 default: 2604 return -EINVAL; 2605 } 2606 2607 if (dyn_uar) { 2608 int uars_per_page; 2609 2610 uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k); 2611 bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR); 2612 if (bfreg_dyn_idx >= bfregi->total_num_bfregs) { 2613 mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n", 2614 bfreg_dyn_idx, bfregi->total_num_bfregs); 2615 return -EINVAL; 2616 } 2617 2618 mutex_lock(&bfregi->lock); 2619 /* Fail if uar already allocated, first bfreg index of each 2620 * page holds its count. 2621 */ 2622 if (bfregi->count[bfreg_dyn_idx]) { 2623 mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx); 2624 mutex_unlock(&bfregi->lock); 2625 return -EINVAL; 2626 } 2627 2628 bfregi->count[bfreg_dyn_idx]++; 2629 mutex_unlock(&bfregi->lock); 2630 2631 err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index, 2632 context->devx_uid); 2633 if (err) { 2634 mlx5_ib_warn(dev, "UAR alloc failed\n"); 2635 goto free_bfreg; 2636 } 2637 } else { 2638 uar_index = bfregi->sys_pages[idx]; 2639 } 2640 2641 pfn = uar_index2pfn(dev, uar_index); 2642 mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn); 2643 2644 err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE, 2645 prot, NULL); 2646 if (err) { 2647 mlx5_ib_err(dev, 2648 "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n", 2649 err, mmap_cmd2str(cmd)); 2650 goto err; 2651 } 2652 2653 if (dyn_uar) 2654 bfregi->sys_pages[idx] = uar_index; 2655 return 0; 2656 2657 err: 2658 if (!dyn_uar) 2659 return err; 2660 2661 mlx5_cmd_uar_dealloc(dev->mdev, idx, context->devx_uid); 2662 2663 free_bfreg: 2664 mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx); 2665 2666 return err; 2667 } 2668 2669 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma) 2670 { 2671 unsigned long idx; 2672 u8 command; 2673 2674 command = get_command(vma->vm_pgoff); 2675 idx = get_extended_index(vma->vm_pgoff); 2676 2677 return (command << 16 | idx); 2678 } 2679 2680 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev, 2681 struct vm_area_struct *vma, 2682 struct ib_ucontext *ucontext) 2683 { 2684 struct mlx5_user_mmap_entry *mentry; 2685 struct rdma_user_mmap_entry *entry; 2686 unsigned long pgoff; 2687 pgprot_t prot; 2688 phys_addr_t pfn; 2689 int ret; 2690 2691 pgoff = mlx5_vma_to_pgoff(vma); 2692 entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff); 2693 if (!entry) 2694 return -EINVAL; 2695 2696 mentry = to_mmmap(entry); 2697 pfn = (mentry->address >> PAGE_SHIFT); 2698 if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR || 2699 mentry->mmap_flag == MLX5_IB_MMAP_TYPE_TLP_VAR || 2700 mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC) 2701 prot = pgprot_noncached(vma->vm_page_prot); 2702 else 2703 prot = pgprot_writecombine(vma->vm_page_prot); 2704 ret = rdma_user_mmap_io(ucontext, vma, pfn, 2705 entry->npages * PAGE_SIZE, 2706 prot, 2707 entry); 2708 rdma_user_mmap_entry_put(&mentry->rdma_entry); 2709 return ret; 2710 } 2711 2712 static u64 mlx5_entry_to_mmap_offset(struct mlx5_user_mmap_entry *entry) 2713 { 2714 u64 cmd = (entry->rdma_entry.start_pgoff >> 16) & 0xFFFF; 2715 u64 index = entry->rdma_entry.start_pgoff & 0xFFFF; 2716 2717 return (((index >> 8) << 16) | (cmd << MLX5_IB_MMAP_CMD_SHIFT) | 2718 (index & 0xFF)) << PAGE_SHIFT; 2719 } 2720 2721 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma) 2722 { 2723 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext); 2724 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device); 2725 unsigned long command; 2726 phys_addr_t pfn; 2727 2728 command = get_command(vma->vm_pgoff); 2729 switch (command) { 2730 case MLX5_IB_MMAP_WC_PAGE: 2731 case MLX5_IB_MMAP_ALLOC_WC: 2732 if (!mlx5_wc_support_get(dev->mdev)) 2733 return -EPERM; 2734 fallthrough; 2735 case MLX5_IB_MMAP_NC_PAGE: 2736 case MLX5_IB_MMAP_REGULAR_PAGE: 2737 return uar_mmap(dev, command, vma, context); 2738 2739 case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES: 2740 return -ENOSYS; 2741 2742 case MLX5_IB_MMAP_CORE_CLOCK: 2743 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 2744 return -EINVAL; 2745 2746 if (vma->vm_flags & VM_WRITE) 2747 return -EPERM; 2748 vm_flags_clear(vma, VM_MAYWRITE); 2749 2750 /* Don't expose to user-space information it shouldn't have */ 2751 if (PAGE_SIZE > 4096) 2752 return -EOPNOTSUPP; 2753 2754 pfn = (dev->mdev->bar_addr + 2755 offsetof(struct mlx5_init_seg, internal_timer_h)) >> 2756 PAGE_SHIFT; 2757 return rdma_user_mmap_io(&context->ibucontext, vma, pfn, 2758 PAGE_SIZE, 2759 pgprot_noncached(vma->vm_page_prot), 2760 NULL); 2761 case MLX5_IB_MMAP_CLOCK_INFO: 2762 return mlx5_ib_mmap_clock_info_page(dev, vma, context); 2763 2764 default: 2765 return mlx5_ib_mmap_offset(dev, vma, ibcontext); 2766 } 2767 2768 return 0; 2769 } 2770 2771 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) 2772 { 2773 struct mlx5_ib_pd *pd = to_mpd(ibpd); 2774 struct ib_device *ibdev = ibpd->device; 2775 struct mlx5_ib_alloc_pd_resp resp; 2776 int err; 2777 u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {}; 2778 u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {}; 2779 u16 uid = 0; 2780 struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context( 2781 udata, struct mlx5_ib_ucontext, ibucontext); 2782 2783 uid = context ? context->devx_uid : 0; 2784 MLX5_SET(alloc_pd_in, in, opcode, MLX5_CMD_OP_ALLOC_PD); 2785 MLX5_SET(alloc_pd_in, in, uid, uid); 2786 err = mlx5_cmd_exec_inout(to_mdev(ibdev)->mdev, alloc_pd, in, out); 2787 if (err) 2788 return err; 2789 2790 pd->pdn = MLX5_GET(alloc_pd_out, out, pd); 2791 pd->uid = uid; 2792 if (udata) { 2793 resp.pdn = pd->pdn; 2794 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) { 2795 mlx5_cmd_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid); 2796 return -EFAULT; 2797 } 2798 } 2799 2800 return 0; 2801 } 2802 2803 static int mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata) 2804 { 2805 struct mlx5_ib_dev *mdev = to_mdev(pd->device); 2806 struct mlx5_ib_pd *mpd = to_mpd(pd); 2807 2808 return mlx5_cmd_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid); 2809 } 2810 2811 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid) 2812 { 2813 struct mlx5_ib_dev *dev = to_mdev(ibqp->device); 2814 struct mlx5_ib_qp *mqp = to_mqp(ibqp); 2815 int err; 2816 u16 uid; 2817 2818 uid = ibqp->pd ? 2819 to_mpd(ibqp->pd)->uid : 0; 2820 2821 if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) { 2822 mlx5_ib_dbg(dev, "Attaching a multi cast group to underlay QP is not supported\n"); 2823 return -EOPNOTSUPP; 2824 } 2825 2826 err = mlx5_cmd_attach_mcg(dev->mdev, gid, ibqp->qp_num, uid); 2827 if (err) 2828 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n", 2829 ibqp->qp_num, gid->raw); 2830 2831 return err; 2832 } 2833 2834 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid) 2835 { 2836 struct mlx5_ib_dev *dev = to_mdev(ibqp->device); 2837 int err; 2838 u16 uid; 2839 2840 uid = ibqp->pd ? 2841 to_mpd(ibqp->pd)->uid : 0; 2842 err = mlx5_cmd_detach_mcg(dev->mdev, gid, ibqp->qp_num, uid); 2843 if (err) 2844 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n", 2845 ibqp->qp_num, gid->raw); 2846 2847 return err; 2848 } 2849 2850 static int init_node_data(struct mlx5_ib_dev *dev) 2851 { 2852 int err; 2853 2854 err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc); 2855 if (err) 2856 return err; 2857 2858 dev->mdev->rev_id = dev->mdev->pdev->revision; 2859 2860 return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid); 2861 } 2862 2863 static ssize_t fw_pages_show(struct device *device, 2864 struct device_attribute *attr, char *buf) 2865 { 2866 struct mlx5_ib_dev *dev = 2867 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev); 2868 2869 return sysfs_emit(buf, "%d\n", dev->mdev->priv.fw_pages); 2870 } 2871 static DEVICE_ATTR_RO(fw_pages); 2872 2873 static ssize_t reg_pages_show(struct device *device, 2874 struct device_attribute *attr, char *buf) 2875 { 2876 struct mlx5_ib_dev *dev = 2877 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev); 2878 2879 return sysfs_emit(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages)); 2880 } 2881 static DEVICE_ATTR_RO(reg_pages); 2882 2883 static ssize_t hca_type_show(struct device *device, 2884 struct device_attribute *attr, char *buf) 2885 { 2886 struct mlx5_ib_dev *dev = 2887 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev); 2888 2889 return sysfs_emit(buf, "MT%d\n", dev->mdev->pdev->device); 2890 } 2891 static DEVICE_ATTR_RO(hca_type); 2892 2893 static ssize_t hw_rev_show(struct device *device, 2894 struct device_attribute *attr, char *buf) 2895 { 2896 struct mlx5_ib_dev *dev = 2897 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev); 2898 2899 return sysfs_emit(buf, "%x\n", dev->mdev->rev_id); 2900 } 2901 static DEVICE_ATTR_RO(hw_rev); 2902 2903 static ssize_t board_id_show(struct device *device, 2904 struct device_attribute *attr, char *buf) 2905 { 2906 struct mlx5_ib_dev *dev = 2907 rdma_device_to_drv_device(device, struct mlx5_ib_dev, ib_dev); 2908 2909 return sysfs_emit(buf, "%.*s\n", MLX5_BOARD_ID_LEN, 2910 dev->mdev->board_id); 2911 } 2912 static DEVICE_ATTR_RO(board_id); 2913 2914 static struct attribute *mlx5_class_attributes[] = { 2915 &dev_attr_hw_rev.attr, 2916 &dev_attr_hca_type.attr, 2917 &dev_attr_board_id.attr, 2918 &dev_attr_fw_pages.attr, 2919 &dev_attr_reg_pages.attr, 2920 NULL, 2921 }; 2922 2923 static const struct attribute_group mlx5_attr_group = { 2924 .attrs = mlx5_class_attributes, 2925 }; 2926 2927 static void pkey_change_handler(struct work_struct *work) 2928 { 2929 struct mlx5_ib_port_resources *ports = 2930 container_of(work, struct mlx5_ib_port_resources, 2931 pkey_change_work); 2932 2933 if (!ports->gsi) 2934 /* 2935 * We got this event before device was fully configured 2936 * and MAD registration code wasn't called/finished yet. 2937 */ 2938 return; 2939 2940 mlx5_ib_gsi_pkey_change(ports->gsi); 2941 } 2942 2943 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev) 2944 { 2945 struct mlx5_ib_qp *mqp; 2946 struct mlx5_ib_cq *send_mcq, *recv_mcq; 2947 struct mlx5_core_cq *mcq; 2948 struct list_head cq_armed_list; 2949 unsigned long flags_qp; 2950 unsigned long flags_cq; 2951 unsigned long flags; 2952 2953 INIT_LIST_HEAD(&cq_armed_list); 2954 2955 /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/ 2956 spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags); 2957 list_for_each_entry(mqp, &ibdev->qp_list, qps_list) { 2958 spin_lock_irqsave(&mqp->sq.lock, flags_qp); 2959 if (mqp->sq.tail != mqp->sq.head) { 2960 send_mcq = to_mcq(mqp->ibqp.send_cq); 2961 spin_lock_irqsave(&send_mcq->lock, flags_cq); 2962 if (send_mcq->mcq.comp && 2963 mqp->ibqp.send_cq->comp_handler) { 2964 if (!send_mcq->mcq.reset_notify_added) { 2965 send_mcq->mcq.reset_notify_added = 1; 2966 list_add_tail(&send_mcq->mcq.reset_notify, 2967 &cq_armed_list); 2968 } 2969 } 2970 spin_unlock_irqrestore(&send_mcq->lock, flags_cq); 2971 } 2972 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp); 2973 spin_lock_irqsave(&mqp->rq.lock, flags_qp); 2974 /* no handling is needed for SRQ */ 2975 if (!mqp->ibqp.srq) { 2976 if (mqp->rq.tail != mqp->rq.head) { 2977 recv_mcq = to_mcq(mqp->ibqp.recv_cq); 2978 spin_lock_irqsave(&recv_mcq->lock, flags_cq); 2979 if (recv_mcq->mcq.comp && 2980 mqp->ibqp.recv_cq->comp_handler) { 2981 if (!recv_mcq->mcq.reset_notify_added) { 2982 recv_mcq->mcq.reset_notify_added = 1; 2983 list_add_tail(&recv_mcq->mcq.reset_notify, 2984 &cq_armed_list); 2985 } 2986 } 2987 spin_unlock_irqrestore(&recv_mcq->lock, 2988 flags_cq); 2989 } 2990 } 2991 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp); 2992 } 2993 /*At that point all inflight post send were put to be executed as of we 2994 * lock/unlock above locks Now need to arm all involved CQs. 2995 */ 2996 list_for_each_entry(mcq, &cq_armed_list, reset_notify) { 2997 mcq->comp(mcq, NULL); 2998 } 2999 spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags); 3000 } 3001 3002 static void delay_drop_handler(struct work_struct *work) 3003 { 3004 int err; 3005 struct mlx5_ib_delay_drop *delay_drop = 3006 container_of(work, struct mlx5_ib_delay_drop, 3007 delay_drop_work); 3008 3009 atomic_inc(&delay_drop->events_cnt); 3010 3011 mutex_lock(&delay_drop->lock); 3012 err = mlx5_core_set_delay_drop(delay_drop->dev, delay_drop->timeout); 3013 if (err) { 3014 mlx5_ib_warn(delay_drop->dev, "Failed to set delay drop, timeout=%u\n", 3015 delay_drop->timeout); 3016 delay_drop->activate = false; 3017 } 3018 mutex_unlock(&delay_drop->lock); 3019 } 3020 3021 static void handle_general_event(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe, 3022 struct ib_event *ibev) 3023 { 3024 u32 port = (eqe->data.port.port >> 4) & 0xf; 3025 3026 switch (eqe->sub_type) { 3027 case MLX5_GENERAL_SUBTYPE_DELAY_DROP_TIMEOUT: 3028 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) == 3029 IB_LINK_LAYER_ETHERNET) 3030 schedule_work(&ibdev->delay_drop.delay_drop_work); 3031 break; 3032 default: /* do nothing */ 3033 return; 3034 } 3035 } 3036 3037 static int handle_port_change(struct mlx5_ib_dev *ibdev, struct mlx5_eqe *eqe, 3038 struct ib_event *ibev) 3039 { 3040 u32 port = (eqe->data.port.port >> 4) & 0xf; 3041 3042 ibev->element.port_num = port; 3043 3044 switch (eqe->sub_type) { 3045 case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE: 3046 case MLX5_PORT_CHANGE_SUBTYPE_DOWN: 3047 case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED: 3048 if (ibdev->ib_active) { 3049 struct ib_event speed_event = {}; 3050 3051 speed_event.device = &ibdev->ib_dev; 3052 speed_event.event = IB_EVENT_DEVICE_SPEED_CHANGE; 3053 ib_dispatch_event(&speed_event); 3054 } 3055 3056 /* In RoCE, port up/down events are handled in 3057 * mlx5_netdev_event(). 3058 */ 3059 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) == 3060 IB_LINK_LAYER_ETHERNET) 3061 return -EINVAL; 3062 3063 ibev->event = (eqe->sub_type == MLX5_PORT_CHANGE_SUBTYPE_ACTIVE) ? 3064 IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR; 3065 break; 3066 3067 case MLX5_PORT_CHANGE_SUBTYPE_LID: 3068 ibev->event = IB_EVENT_LID_CHANGE; 3069 break; 3070 3071 case MLX5_PORT_CHANGE_SUBTYPE_PKEY: 3072 ibev->event = IB_EVENT_PKEY_CHANGE; 3073 schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work); 3074 break; 3075 3076 case MLX5_PORT_CHANGE_SUBTYPE_GUID: 3077 ibev->event = IB_EVENT_GID_CHANGE; 3078 break; 3079 3080 case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG: 3081 ibev->event = IB_EVENT_CLIENT_REREGISTER; 3082 break; 3083 default: 3084 return -EINVAL; 3085 } 3086 3087 return 0; 3088 } 3089 3090 static void mlx5_ib_handle_event(struct work_struct *_work) 3091 { 3092 struct mlx5_ib_event_work *work = 3093 container_of(_work, struct mlx5_ib_event_work, work); 3094 struct mlx5_ib_dev *ibdev; 3095 struct ib_event ibev; 3096 3097 if (work->is_slave) { 3098 ibdev = mlx5_ib_get_ibdev_from_mpi(work->mpi); 3099 if (!ibdev) 3100 goto out; 3101 } else { 3102 ibdev = work->dev; 3103 } 3104 3105 switch (work->event) { 3106 case MLX5_EVENT_TYPE_PORT_CHANGE: 3107 if (handle_port_change(ibdev, work->param, &ibev)) 3108 goto out; 3109 break; 3110 case MLX5_EVENT_TYPE_GENERAL_EVENT: 3111 handle_general_event(ibdev, work->param, &ibev); 3112 fallthrough; 3113 default: 3114 goto out; 3115 } 3116 3117 ibev.device = &ibdev->ib_dev; 3118 3119 if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) { 3120 mlx5_ib_warn(ibdev, "warning: event on port %d\n", ibev.element.port_num); 3121 goto out; 3122 } 3123 3124 if (ibdev->ib_active) 3125 ib_dispatch_event(&ibev); 3126 3127 out: 3128 kfree(work); 3129 } 3130 3131 static int mlx5_ib_event(struct notifier_block *nb, 3132 unsigned long event, void *param) 3133 { 3134 struct mlx5_ib_event_work *work; 3135 3136 work = kmalloc_obj(*work, GFP_ATOMIC); 3137 if (!work) 3138 return NOTIFY_DONE; 3139 3140 INIT_WORK(&work->work, mlx5_ib_handle_event); 3141 work->dev = container_of(nb, struct mlx5_ib_dev, mdev_events); 3142 work->is_slave = false; 3143 work->param = param; 3144 work->event = event; 3145 3146 queue_work(mlx5_ib_event_wq, &work->work); 3147 3148 return NOTIFY_OK; 3149 } 3150 3151 static int mlx5_ib_event_slave_port(struct notifier_block *nb, 3152 unsigned long event, void *param) 3153 { 3154 struct mlx5_ib_event_work *work; 3155 3156 work = kmalloc_obj(*work, GFP_ATOMIC); 3157 if (!work) 3158 return NOTIFY_DONE; 3159 3160 INIT_WORK(&work->work, mlx5_ib_handle_event); 3161 work->mpi = container_of(nb, struct mlx5_ib_multiport_info, mdev_events); 3162 work->is_slave = true; 3163 work->param = param; 3164 work->event = event; 3165 queue_work(mlx5_ib_event_wq, &work->work); 3166 3167 return NOTIFY_OK; 3168 } 3169 3170 static void mlx5_ib_handle_sys_error_event(struct work_struct *_work) 3171 { 3172 struct mlx5_ib_event_work *work = 3173 container_of(_work, struct mlx5_ib_event_work, work); 3174 struct mlx5_ib_dev *ibdev = work->dev; 3175 struct ib_event ibev; 3176 3177 ibev.event = IB_EVENT_DEVICE_FATAL; 3178 mlx5_ib_handle_internal_error(ibdev); 3179 ibev.element.port_num = (u8)(unsigned long)work->param; 3180 ibev.device = &ibdev->ib_dev; 3181 3182 if (!rdma_is_port_valid(&ibdev->ib_dev, ibev.element.port_num)) { 3183 mlx5_ib_warn(ibdev, "warning: event on port %d\n", ibev.element.port_num); 3184 goto out; 3185 } 3186 3187 if (ibdev->ib_active) 3188 ib_dispatch_event(&ibev); 3189 3190 ibdev->ib_active = false; 3191 out: 3192 kfree(work); 3193 } 3194 3195 static int mlx5_ib_sys_error_event(struct notifier_block *nb, 3196 unsigned long event, void *param) 3197 { 3198 struct mlx5_ib_event_work *work; 3199 3200 if (event != MLX5_DEV_EVENT_SYS_ERROR) 3201 return NOTIFY_DONE; 3202 3203 work = kmalloc_obj(*work, GFP_ATOMIC); 3204 if (!work) 3205 return NOTIFY_DONE; 3206 3207 INIT_WORK(&work->work, mlx5_ib_handle_sys_error_event); 3208 work->dev = container_of(nb, struct mlx5_ib_dev, sys_error_events); 3209 work->is_slave = false; 3210 work->param = param; 3211 work->event = event; 3212 3213 queue_work(mlx5_ib_event_wq, &work->work); 3214 3215 return NOTIFY_OK; 3216 } 3217 3218 static int mlx5_ib_stage_sys_error_notifier_init(struct mlx5_ib_dev *dev) 3219 { 3220 dev->sys_error_events.notifier_call = mlx5_ib_sys_error_event; 3221 mlx5_notifier_register(dev->mdev, &dev->sys_error_events); 3222 return 0; 3223 } 3224 3225 static void mlx5_ib_stage_sys_error_notifier_cleanup(struct mlx5_ib_dev *dev) 3226 { 3227 mlx5_notifier_unregister(dev->mdev, &dev->sys_error_events); 3228 } 3229 3230 static int mlx5_ib_get_plane_num(struct mlx5_core_dev *mdev, u8 *num_plane) 3231 { 3232 struct mlx5_hca_vport_context vport_ctx; 3233 int err; 3234 3235 *num_plane = 0; 3236 if (!MLX5_CAP_GEN(mdev, ib_virt) || !MLX5_CAP_GEN_2(mdev, multiplane)) 3237 return 0; 3238 3239 err = mlx5_query_hca_vport_context(mdev, 0, 1, 0, &vport_ctx); 3240 if (err) 3241 return err; 3242 3243 *num_plane = vport_ctx.num_plane; 3244 return 0; 3245 } 3246 3247 static int set_has_smi_cap(struct mlx5_ib_dev *dev) 3248 { 3249 struct mlx5_hca_vport_context vport_ctx; 3250 int err; 3251 int port; 3252 3253 if (MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_IB) 3254 return 0; 3255 3256 for (port = 1; port <= dev->num_ports; port++) { 3257 if (dev->num_plane) { 3258 dev->port_caps[port - 1].has_smi = false; 3259 continue; 3260 } else if (!MLX5_CAP_GEN(dev->mdev, ib_virt) || 3261 dev->ib_dev.type == RDMA_DEVICE_TYPE_SMI) { 3262 dev->port_caps[port - 1].has_smi = true; 3263 continue; 3264 } 3265 3266 err = mlx5_query_hca_vport_context(dev->mdev, 0, port, 0, 3267 &vport_ctx); 3268 if (err) { 3269 mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n", 3270 port, err); 3271 return err; 3272 } 3273 dev->port_caps[port - 1].has_smi = vport_ctx.has_smi; 3274 } 3275 3276 return 0; 3277 } 3278 3279 static void get_ext_port_caps(struct mlx5_ib_dev *dev) 3280 { 3281 unsigned int port; 3282 3283 rdma_for_each_port (&dev->ib_dev, port) 3284 mlx5_query_ext_port_caps(dev, port); 3285 } 3286 3287 static u8 mlx5_get_umr_fence(u8 umr_fence_cap) 3288 { 3289 switch (umr_fence_cap) { 3290 case MLX5_CAP_UMR_FENCE_NONE: 3291 return MLX5_FENCE_MODE_NONE; 3292 case MLX5_CAP_UMR_FENCE_SMALL: 3293 return MLX5_FENCE_MODE_INITIATOR_SMALL; 3294 default: 3295 return MLX5_FENCE_MODE_STRONG_ORDERING; 3296 } 3297 } 3298 3299 int mlx5_ib_dev_res_cq_init(struct mlx5_ib_dev *dev) 3300 { 3301 struct mlx5_ib_resources *devr = &dev->devr; 3302 struct ib_cq_init_attr cq_attr = {.cqe = 1}; 3303 struct ib_device *ibdev; 3304 struct ib_pd *pd; 3305 struct ib_cq *cq; 3306 int ret = 0; 3307 3308 3309 /* 3310 * devr->c0 is set once, never changed until device unload. 3311 * Avoid taking the mutex if initialization is already done. 3312 */ 3313 if (devr->c0) 3314 return 0; 3315 3316 mutex_lock(&devr->cq_lock); 3317 if (devr->c0) 3318 goto unlock; 3319 3320 ibdev = &dev->ib_dev; 3321 pd = ib_alloc_pd(ibdev, 0); 3322 if (IS_ERR(pd)) { 3323 ret = PTR_ERR(pd); 3324 mlx5_ib_err(dev, "Couldn't allocate PD for res init, err=%pe\n", 3325 pd); 3326 goto unlock; 3327 } 3328 3329 cq = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr); 3330 if (IS_ERR(cq)) { 3331 ret = PTR_ERR(cq); 3332 mlx5_ib_err(dev, "Couldn't create CQ for res init, err=%pe\n", 3333 cq); 3334 ib_dealloc_pd(pd); 3335 goto unlock; 3336 } 3337 3338 devr->p0 = pd; 3339 devr->c0 = cq; 3340 3341 unlock: 3342 mutex_unlock(&devr->cq_lock); 3343 return ret; 3344 } 3345 3346 int mlx5_ib_dev_res_srq_init(struct mlx5_ib_dev *dev) 3347 { 3348 struct mlx5_ib_resources *devr = &dev->devr; 3349 struct ib_srq_init_attr attr; 3350 struct ib_srq *s0, *s1; 3351 int ret = 0; 3352 3353 /* 3354 * devr->s1 is set once, never changed until device unload. 3355 * Avoid taking the mutex if initialization is already done. 3356 */ 3357 if (devr->s1) 3358 return 0; 3359 3360 mutex_lock(&devr->srq_lock); 3361 if (devr->s1) 3362 goto unlock; 3363 3364 ret = mlx5_ib_dev_res_cq_init(dev); 3365 if (ret) 3366 goto unlock; 3367 3368 memset(&attr, 0, sizeof(attr)); 3369 attr.attr.max_sge = 1; 3370 attr.attr.max_wr = 1; 3371 attr.srq_type = IB_SRQT_XRC; 3372 attr.ext.cq = devr->c0; 3373 3374 s0 = ib_create_srq(devr->p0, &attr); 3375 if (IS_ERR(s0)) { 3376 ret = PTR_ERR(s0); 3377 mlx5_ib_err(dev, 3378 "Couldn't create SRQ 0 for res init, err=%pe\n", 3379 s0); 3380 goto unlock; 3381 } 3382 3383 memset(&attr, 0, sizeof(attr)); 3384 attr.attr.max_sge = 1; 3385 attr.attr.max_wr = 1; 3386 attr.srq_type = IB_SRQT_BASIC; 3387 3388 s1 = ib_create_srq(devr->p0, &attr); 3389 if (IS_ERR(s1)) { 3390 ret = PTR_ERR(s1); 3391 mlx5_ib_err(dev, 3392 "Couldn't create SRQ 1 for res init, err=%pe\n", 3393 s1); 3394 ib_destroy_srq(s0); 3395 } 3396 3397 devr->s0 = s0; 3398 devr->s1 = s1; 3399 3400 unlock: 3401 mutex_unlock(&devr->srq_lock); 3402 return ret; 3403 } 3404 3405 static int mlx5_ib_dev_res_init(struct mlx5_ib_dev *dev) 3406 { 3407 struct mlx5_ib_resources *devr = &dev->devr; 3408 int ret; 3409 3410 if (!MLX5_CAP_GEN(dev->mdev, xrc)) 3411 return -EOPNOTSUPP; 3412 3413 ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn0, 0); 3414 if (ret) 3415 return ret; 3416 3417 ret = mlx5_cmd_xrcd_alloc(dev->mdev, &devr->xrcdn1, 0); 3418 if (ret) { 3419 mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0); 3420 return ret; 3421 } 3422 3423 mutex_init(&devr->cq_lock); 3424 mutex_init(&devr->srq_lock); 3425 3426 return 0; 3427 } 3428 3429 static void mlx5_ib_dev_res_cleanup(struct mlx5_ib_dev *dev) 3430 { 3431 struct mlx5_ib_resources *devr = &dev->devr; 3432 3433 /* After s0/s1 init, they are not unset during the device lifetime. */ 3434 if (devr->s1) { 3435 ib_destroy_srq(devr->s1); 3436 ib_destroy_srq(devr->s0); 3437 } 3438 mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn1, 0); 3439 mlx5_cmd_xrcd_dealloc(dev->mdev, devr->xrcdn0, 0); 3440 /* After p0/c0 init, they are not unset during the device lifetime. */ 3441 if (devr->c0) { 3442 ib_destroy_cq(devr->c0); 3443 ib_dealloc_pd(devr->p0); 3444 } 3445 mutex_destroy(&devr->cq_lock); 3446 mutex_destroy(&devr->srq_lock); 3447 } 3448 3449 static int 3450 mlx5_ib_create_data_direct_resources(struct mlx5_ib_dev *dev) 3451 { 3452 int inlen = MLX5_ST_SZ_BYTES(create_mkey_in); 3453 struct mlx5_core_dev *mdev = dev->mdev; 3454 bool ro_supp = false; 3455 void *mkc; 3456 u32 mkey; 3457 u32 pdn; 3458 u32 *in; 3459 int err; 3460 3461 err = mlx5_core_alloc_pd(mdev, &pdn); 3462 if (err) 3463 return err; 3464 3465 in = kvzalloc(inlen, GFP_KERNEL); 3466 if (!in) { 3467 err = -ENOMEM; 3468 goto err; 3469 } 3470 3471 MLX5_SET(create_mkey_in, in, data_direct, 1); 3472 mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry); 3473 MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_PA); 3474 MLX5_SET(mkc, mkc, lw, 1); 3475 MLX5_SET(mkc, mkc, lr, 1); 3476 MLX5_SET(mkc, mkc, rw, 1); 3477 MLX5_SET(mkc, mkc, rr, 1); 3478 MLX5_SET(mkc, mkc, a, 1); 3479 MLX5_SET(mkc, mkc, pd, pdn); 3480 MLX5_SET(mkc, mkc, length64, 1); 3481 MLX5_SET(mkc, mkc, qpn, 0xffffff); 3482 err = mlx5_core_create_mkey(mdev, &mkey, in, inlen); 3483 if (err) 3484 goto err_mkey; 3485 3486 dev->ddr.mkey = mkey; 3487 dev->ddr.pdn = pdn; 3488 3489 /* create another mkey with RO support */ 3490 if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write)) { 3491 MLX5_SET(mkc, mkc, relaxed_ordering_write, 1); 3492 ro_supp = true; 3493 } 3494 3495 if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read)) { 3496 MLX5_SET(mkc, mkc, relaxed_ordering_read, 1); 3497 ro_supp = true; 3498 } 3499 3500 if (ro_supp) { 3501 err = mlx5_core_create_mkey(mdev, &mkey, in, inlen); 3502 /* RO is defined as best effort */ 3503 if (!err) { 3504 dev->ddr.mkey_ro = mkey; 3505 dev->ddr.mkey_ro_valid = true; 3506 } 3507 } 3508 3509 kvfree(in); 3510 return 0; 3511 3512 err_mkey: 3513 kvfree(in); 3514 err: 3515 mlx5_core_dealloc_pd(mdev, pdn); 3516 return err; 3517 } 3518 3519 static void 3520 mlx5_ib_free_data_direct_resources(struct mlx5_ib_dev *dev) 3521 { 3522 3523 if (dev->ddr.mkey_ro_valid) 3524 mlx5_core_destroy_mkey(dev->mdev, dev->ddr.mkey_ro); 3525 3526 mlx5_core_destroy_mkey(dev->mdev, dev->ddr.mkey); 3527 mlx5_core_dealloc_pd(dev->mdev, dev->ddr.pdn); 3528 } 3529 3530 static u32 get_core_cap_flags(struct ib_device *ibdev, 3531 struct mlx5_hca_vport_context *rep) 3532 { 3533 struct mlx5_ib_dev *dev = to_mdev(ibdev); 3534 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1); 3535 u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type); 3536 u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version); 3537 bool raw_support = !mlx5_core_mp_enabled(dev->mdev); 3538 u32 ret = 0; 3539 3540 if (rep->grh_required) 3541 ret |= RDMA_CORE_CAP_IB_GRH_REQUIRED; 3542 3543 if (dev->num_plane) 3544 return ret | RDMA_CORE_CAP_PROT_IB | RDMA_CORE_CAP_IB_MAD | 3545 RDMA_CORE_CAP_IB_CM | RDMA_CORE_CAP_IB_SA | 3546 RDMA_CORE_CAP_AF_IB; 3547 else if (ibdev->type == RDMA_DEVICE_TYPE_SMI) 3548 return ret | RDMA_CORE_CAP_IB_MAD | RDMA_CORE_CAP_IB_SMI; 3549 3550 if (ll == IB_LINK_LAYER_INFINIBAND) 3551 return ret | RDMA_CORE_PORT_IBA_IB; 3552 3553 if (raw_support) 3554 ret |= RDMA_CORE_PORT_RAW_PACKET; 3555 3556 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP)) 3557 return ret; 3558 3559 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP)) 3560 return ret; 3561 3562 if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP) 3563 ret |= RDMA_CORE_PORT_IBA_ROCE; 3564 3565 if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP) 3566 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP; 3567 3568 return ret; 3569 } 3570 3571 static int mlx5_port_immutable(struct ib_device *ibdev, u32 port_num, 3572 struct ib_port_immutable *immutable) 3573 { 3574 struct ib_port_attr attr; 3575 struct mlx5_ib_dev *dev = to_mdev(ibdev); 3576 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num); 3577 struct mlx5_hca_vport_context rep = {0}; 3578 int err; 3579 3580 err = ib_query_port(ibdev, port_num, &attr); 3581 if (err) 3582 return err; 3583 3584 if (ll == IB_LINK_LAYER_INFINIBAND) { 3585 if (ibdev->type == RDMA_DEVICE_TYPE_SMI) 3586 port_num = smi_to_native_portnum(dev, port_num); 3587 3588 err = mlx5_query_hca_vport_context(dev->mdev, 0, port_num, 0, 3589 &rep); 3590 if (err) 3591 return err; 3592 } 3593 3594 immutable->pkey_tbl_len = attr.pkey_tbl_len; 3595 immutable->gid_tbl_len = attr.gid_tbl_len; 3596 immutable->core_cap_flags = get_core_cap_flags(ibdev, &rep); 3597 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 3598 3599 return 0; 3600 } 3601 3602 static int mlx5_port_rep_immutable(struct ib_device *ibdev, u32 port_num, 3603 struct ib_port_immutable *immutable) 3604 { 3605 struct ib_port_attr attr; 3606 int err; 3607 3608 immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET; 3609 3610 err = ib_query_port(ibdev, port_num, &attr); 3611 if (err) 3612 return err; 3613 3614 immutable->pkey_tbl_len = attr.pkey_tbl_len; 3615 immutable->gid_tbl_len = attr.gid_tbl_len; 3616 immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET; 3617 3618 return 0; 3619 } 3620 3621 static void get_dev_fw_str(struct ib_device *ibdev, char *str) 3622 { 3623 struct mlx5_ib_dev *dev = 3624 container_of(ibdev, struct mlx5_ib_dev, ib_dev); 3625 snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%04d", 3626 fw_rev_maj(dev->mdev), fw_rev_min(dev->mdev), 3627 fw_rev_sub(dev->mdev)); 3628 } 3629 3630 static int lag_event(struct notifier_block *nb, unsigned long event, void *data) 3631 { 3632 struct mlx5_ib_dev *dev = container_of(nb, struct mlx5_ib_dev, 3633 lag_events); 3634 struct mlx5_core_dev *mdev = dev->mdev; 3635 struct ib_device *ibdev = &dev->ib_dev; 3636 struct net_device *old_ndev = NULL; 3637 struct mlx5_ib_port *port; 3638 struct net_device *ndev; 3639 u32 portnum = 0; 3640 int ret = 0; 3641 int i; 3642 3643 switch (event) { 3644 case MLX5_DRIVER_EVENT_ACTIVE_BACKUP_LAG_CHANGE_LOWERSTATE: 3645 ndev = data; 3646 if (ndev) { 3647 if (!mlx5_lag_is_roce(mdev)) { 3648 // sriov lag 3649 for (i = 0; i < dev->num_ports; i++) { 3650 port = &dev->port[i]; 3651 if (port->rep && port->rep->vport == 3652 MLX5_VPORT_UPLINK) { 3653 portnum = i; 3654 break; 3655 } 3656 } 3657 } 3658 old_ndev = ib_device_get_netdev(ibdev, portnum + 1); 3659 ret = ib_device_set_netdev(ibdev, ndev, portnum + 1); 3660 if (ret) 3661 goto out; 3662 3663 if (old_ndev) 3664 roce_del_all_netdev_gids(ibdev, portnum + 1, 3665 old_ndev); 3666 rdma_roce_rescan_port(ibdev, portnum + 1); 3667 } 3668 break; 3669 default: 3670 return NOTIFY_DONE; 3671 } 3672 3673 out: 3674 dev_put(old_ndev); 3675 return notifier_from_errno(ret); 3676 } 3677 3678 static void mlx5e_lag_event_register(struct mlx5_ib_dev *dev) 3679 { 3680 dev->lag_events.notifier_call = lag_event; 3681 blocking_notifier_chain_register(&dev->mdev->priv.lag_nh, 3682 &dev->lag_events); 3683 } 3684 3685 static void mlx5e_lag_event_unregister(struct mlx5_ib_dev *dev) 3686 { 3687 blocking_notifier_chain_unregister(&dev->mdev->priv.lag_nh, 3688 &dev->lag_events); 3689 } 3690 3691 static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev) 3692 { 3693 struct mlx5_flow_table_attr ft_attr = {}; 3694 struct mlx5_core_dev *mdev = dev->mdev; 3695 struct mlx5_flow_namespace *ns; 3696 int err; 3697 3698 ns = mlx5_get_flow_namespace(mdev, MLX5_FLOW_NAMESPACE_LAG); 3699 if (!ns || !mlx5_lag_is_active(mdev)) 3700 return 0; 3701 3702 err = mlx5_cmd_create_vport_lag(mdev); 3703 if (err) 3704 return err; 3705 3706 ft_attr.level = 0; 3707 ft_attr.prio = 0; 3708 ft_attr.max_fte = dev->num_ports; 3709 3710 err = mlx5_lag_demux_init(mdev, &ft_attr); 3711 if (err) 3712 goto err_destroy_vport_lag; 3713 3714 mlx5e_lag_event_register(dev); 3715 dev->lag_ports = mlx5_lag_get_num_ports(mdev); 3716 dev->lag_active = true; 3717 return 0; 3718 3719 err_destroy_vport_lag: 3720 mlx5_cmd_destroy_vport_lag(mdev); 3721 return err; 3722 } 3723 3724 static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev) 3725 { 3726 struct mlx5_core_dev *mdev = dev->mdev; 3727 3728 if (dev->lag_active) { 3729 dev->lag_active = false; 3730 3731 mlx5e_lag_event_unregister(dev); 3732 mlx5_lag_demux_cleanup(mdev); 3733 3734 mlx5_cmd_destroy_vport_lag(mdev); 3735 } 3736 } 3737 3738 static void mlx5_netdev_notifier_register(struct mlx5_roce *roce, 3739 struct net_device *netdev) 3740 { 3741 int err; 3742 3743 if (roce->tracking_netdev) 3744 return; 3745 roce->tracking_netdev = netdev; 3746 roce->nb.notifier_call = mlx5_netdev_event; 3747 err = register_netdevice_notifier_dev_net(netdev, &roce->nb, &roce->nn); 3748 WARN_ON(err); 3749 } 3750 3751 static void mlx5_netdev_notifier_unregister(struct mlx5_roce *roce) 3752 { 3753 if (!roce->tracking_netdev) 3754 return; 3755 unregister_netdevice_notifier_dev_net(roce->tracking_netdev, &roce->nb, 3756 &roce->nn); 3757 roce->tracking_netdev = NULL; 3758 } 3759 3760 static int mlx5e_mdev_notifier_event(struct notifier_block *nb, 3761 unsigned long event, void *data) 3762 { 3763 struct mlx5_roce *roce = container_of(nb, struct mlx5_roce, mdev_nb); 3764 struct net_device *netdev = data; 3765 3766 switch (event) { 3767 case MLX5_DRIVER_EVENT_UPLINK_NETDEV: 3768 if (netdev) 3769 mlx5_netdev_notifier_register(roce, netdev); 3770 else 3771 mlx5_netdev_notifier_unregister(roce); 3772 break; 3773 default: 3774 return NOTIFY_DONE; 3775 } 3776 3777 return NOTIFY_OK; 3778 } 3779 3780 static void mlx5_mdev_netdev_track(struct mlx5_ib_dev *dev, u32 port_num) 3781 { 3782 struct mlx5_roce *roce = &dev->port[port_num].roce; 3783 3784 roce->mdev_nb.notifier_call = mlx5e_mdev_notifier_event; 3785 mlx5_blocking_notifier_register(dev->mdev, &roce->mdev_nb); 3786 mlx5_core_uplink_netdev_event_replay(dev->mdev); 3787 } 3788 3789 static void mlx5_mdev_netdev_untrack(struct mlx5_ib_dev *dev, u32 port_num) 3790 { 3791 struct mlx5_roce *roce = &dev->port[port_num].roce; 3792 3793 mlx5_blocking_notifier_unregister(dev->mdev, &roce->mdev_nb); 3794 mlx5_netdev_notifier_unregister(roce); 3795 } 3796 3797 static int mlx5_enable_eth(struct mlx5_ib_dev *dev) 3798 { 3799 int err; 3800 3801 if (!dev->is_rep && dev->profile != &raw_eth_profile) { 3802 err = mlx5_nic_vport_enable_roce(dev->mdev); 3803 if (err) 3804 return err; 3805 } 3806 3807 err = mlx5_eth_lag_init(dev); 3808 if (err) 3809 goto err_disable_roce; 3810 3811 return 0; 3812 3813 err_disable_roce: 3814 if (!dev->is_rep && dev->profile != &raw_eth_profile) 3815 mlx5_nic_vport_disable_roce(dev->mdev); 3816 3817 return err; 3818 } 3819 3820 static void mlx5_disable_eth(struct mlx5_ib_dev *dev) 3821 { 3822 mlx5_eth_lag_cleanup(dev); 3823 if (!dev->is_rep && dev->profile != &raw_eth_profile) 3824 mlx5_nic_vport_disable_roce(dev->mdev); 3825 } 3826 3827 static int mlx5_ib_rn_get_params(struct ib_device *device, u32 port_num, 3828 enum rdma_netdev_t type, 3829 struct rdma_netdev_alloc_params *params) 3830 { 3831 if (type != RDMA_NETDEV_IPOIB) 3832 return -EOPNOTSUPP; 3833 3834 return mlx5_rdma_rn_get_params(to_mdev(device)->mdev, device, params); 3835 } 3836 3837 static ssize_t delay_drop_timeout_read(struct file *filp, char __user *buf, 3838 size_t count, loff_t *pos) 3839 { 3840 struct mlx5_ib_delay_drop *delay_drop = filp->private_data; 3841 char lbuf[20]; 3842 int len; 3843 3844 len = snprintf(lbuf, sizeof(lbuf), "%u\n", delay_drop->timeout); 3845 return simple_read_from_buffer(buf, count, pos, lbuf, len); 3846 } 3847 3848 static ssize_t delay_drop_timeout_write(struct file *filp, const char __user *buf, 3849 size_t count, loff_t *pos) 3850 { 3851 struct mlx5_ib_delay_drop *delay_drop = filp->private_data; 3852 u32 timeout; 3853 u32 var; 3854 3855 if (kstrtouint_from_user(buf, count, 0, &var)) 3856 return -EFAULT; 3857 3858 timeout = min_t(u32, roundup(var, 100), MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 3859 1000); 3860 if (timeout != var) 3861 mlx5_ib_dbg(delay_drop->dev, "Round delay drop timeout to %u usec\n", 3862 timeout); 3863 3864 delay_drop->timeout = timeout; 3865 3866 return count; 3867 } 3868 3869 static const struct file_operations fops_delay_drop_timeout = { 3870 .owner = THIS_MODULE, 3871 .open = simple_open, 3872 .write = delay_drop_timeout_write, 3873 .read = delay_drop_timeout_read, 3874 }; 3875 3876 static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev, 3877 struct mlx5_ib_multiport_info *mpi) 3878 { 3879 u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1; 3880 struct mlx5_ib_port *port = &ibdev->port[port_num]; 3881 int comps; 3882 int err; 3883 int i; 3884 3885 lockdep_assert_held(&mlx5_ib_multiport_mutex); 3886 3887 mlx5_ib_disable_lb_mp(ibdev->mdev, mpi->mdev, &ibdev->lb); 3888 3889 mlx5_core_mp_event_replay(ibdev->mdev, 3890 MLX5_DRIVER_EVENT_AFFILIATION_REMOVED, 3891 NULL); 3892 mlx5_core_mp_event_replay(mpi->mdev, 3893 MLX5_DRIVER_EVENT_AFFILIATION_REMOVED, 3894 NULL); 3895 3896 mlx5_ib_cleanup_cong_debugfs(ibdev, port_num); 3897 3898 spin_lock(&port->mp.mpi_lock); 3899 if (!mpi->ibdev) { 3900 spin_unlock(&port->mp.mpi_lock); 3901 return; 3902 } 3903 3904 mpi->ibdev = NULL; 3905 3906 spin_unlock(&port->mp.mpi_lock); 3907 if (mpi->mdev_events.notifier_call) 3908 mlx5_notifier_unregister(mpi->mdev, &mpi->mdev_events); 3909 mpi->mdev_events.notifier_call = NULL; 3910 mlx5_mdev_netdev_untrack(ibdev, port_num); 3911 spin_lock(&port->mp.mpi_lock); 3912 3913 comps = mpi->mdev_refcnt; 3914 if (comps) { 3915 mpi->unaffiliate = true; 3916 init_completion(&mpi->unref_comp); 3917 spin_unlock(&port->mp.mpi_lock); 3918 3919 for (i = 0; i < comps; i++) 3920 wait_for_completion(&mpi->unref_comp); 3921 3922 spin_lock(&port->mp.mpi_lock); 3923 mpi->unaffiliate = false; 3924 } 3925 3926 port->mp.mpi = NULL; 3927 3928 spin_unlock(&port->mp.mpi_lock); 3929 3930 err = mlx5_nic_vport_unaffiliate_multiport(mpi->mdev); 3931 3932 mlx5_ib_dbg(ibdev, "unaffiliated port %u\n", port_num + 1); 3933 /* Log an error, still needed to cleanup the pointers and add 3934 * it back to the list. 3935 */ 3936 if (err) 3937 mlx5_ib_err(ibdev, "Failed to unaffiliate port %u\n", 3938 port_num + 1); 3939 3940 ibdev->port[port_num].roce.last_port_state = IB_PORT_DOWN; 3941 } 3942 3943 static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev, 3944 struct mlx5_ib_multiport_info *mpi) 3945 { 3946 u32 port_num = mlx5_core_native_port_num(mpi->mdev) - 1; 3947 u64 key; 3948 int err; 3949 3950 lockdep_assert_held(&mlx5_ib_multiport_mutex); 3951 3952 spin_lock(&ibdev->port[port_num].mp.mpi_lock); 3953 if (ibdev->port[port_num].mp.mpi) { 3954 mlx5_ib_dbg(ibdev, "port %u already affiliated.\n", 3955 port_num + 1); 3956 spin_unlock(&ibdev->port[port_num].mp.mpi_lock); 3957 return false; 3958 } 3959 3960 ibdev->port[port_num].mp.mpi = mpi; 3961 mpi->ibdev = ibdev; 3962 mpi->mdev_events.notifier_call = NULL; 3963 spin_unlock(&ibdev->port[port_num].mp.mpi_lock); 3964 3965 err = mlx5_nic_vport_affiliate_multiport(ibdev->mdev, mpi->mdev); 3966 if (err) 3967 goto unbind; 3968 3969 mlx5_mdev_netdev_track(ibdev, port_num); 3970 3971 mpi->mdev_events.notifier_call = mlx5_ib_event_slave_port; 3972 mlx5_notifier_register(mpi->mdev, &mpi->mdev_events); 3973 3974 mlx5_ib_init_cong_debugfs(ibdev, port_num); 3975 3976 key = mpi->mdev->priv.adev_idx; 3977 mlx5_core_mp_event_replay(mpi->mdev, 3978 MLX5_DRIVER_EVENT_AFFILIATION_DONE, 3979 &key); 3980 mlx5_core_mp_event_replay(ibdev->mdev, 3981 MLX5_DRIVER_EVENT_AFFILIATION_DONE, 3982 &key); 3983 3984 err = mlx5_ib_enable_lb_mp(ibdev->mdev, mpi->mdev, &ibdev->lb); 3985 if (err) 3986 goto unbind; 3987 3988 return true; 3989 3990 unbind: 3991 mlx5_ib_unbind_slave_port(ibdev, mpi); 3992 return false; 3993 } 3994 3995 static int mlx5_ib_data_direct_init(struct mlx5_ib_dev *dev) 3996 { 3997 char vuid[MLX5_ST_SZ_BYTES(array1024_auto) + 1] = {}; 3998 int ret; 3999 4000 if (!MLX5_CAP_GEN(dev->mdev, data_direct) || 4001 !MLX5_CAP_GEN_2(dev->mdev, query_vuid)) 4002 return 0; 4003 4004 ret = mlx5_cmd_query_vuid(dev->mdev, true, vuid); 4005 if (ret) 4006 return ret; 4007 4008 ret = mlx5_ib_create_data_direct_resources(dev); 4009 if (ret) 4010 return ret; 4011 4012 INIT_LIST_HEAD(&dev->data_direct_mr_list); 4013 ret = mlx5_data_direct_ib_reg(dev, vuid); 4014 if (ret) 4015 mlx5_ib_free_data_direct_resources(dev); 4016 4017 return ret; 4018 } 4019 4020 static void mlx5_ib_data_direct_cleanup(struct mlx5_ib_dev *dev) 4021 { 4022 if (!MLX5_CAP_GEN(dev->mdev, data_direct) || 4023 !MLX5_CAP_GEN_2(dev->mdev, query_vuid)) 4024 return; 4025 4026 mlx5_data_direct_ib_unreg(dev); 4027 mlx5_ib_free_data_direct_resources(dev); 4028 } 4029 4030 static int mlx5_ib_init_multiport_master(struct mlx5_ib_dev *dev) 4031 { 4032 u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1; 4033 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 4034 port_num + 1); 4035 struct mlx5_ib_multiport_info *mpi; 4036 int err; 4037 u32 i; 4038 4039 if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET) 4040 return 0; 4041 4042 err = mlx5_query_nic_vport_system_image_guid(dev->mdev, 4043 &dev->sys_image_guid); 4044 if (err) 4045 return err; 4046 4047 err = mlx5_nic_vport_enable_roce(dev->mdev); 4048 if (err) 4049 return err; 4050 4051 mutex_lock(&mlx5_ib_multiport_mutex); 4052 for (i = 0; i < dev->num_ports; i++) { 4053 bool bound = false; 4054 4055 /* build a stub multiport info struct for the native port. */ 4056 if (i == port_num) { 4057 mpi = kzalloc_obj(*mpi); 4058 if (!mpi) { 4059 mutex_unlock(&mlx5_ib_multiport_mutex); 4060 mlx5_nic_vport_disable_roce(dev->mdev); 4061 return -ENOMEM; 4062 } 4063 4064 mpi->is_master = true; 4065 mpi->mdev = dev->mdev; 4066 mpi->sys_image_guid = dev->sys_image_guid; 4067 dev->port[i].mp.mpi = mpi; 4068 mpi->ibdev = dev; 4069 mpi = NULL; 4070 continue; 4071 } 4072 4073 list_for_each_entry(mpi, &mlx5_ib_unaffiliated_port_list, 4074 list) { 4075 if (dev->sys_image_guid == mpi->sys_image_guid && 4076 (mlx5_core_native_port_num(mpi->mdev) - 1) == i && 4077 mlx5_core_same_coredev_type(dev->mdev, mpi->mdev)) { 4078 bound = mlx5_ib_bind_slave_port(dev, mpi); 4079 } 4080 4081 if (bound) { 4082 dev_dbg(mpi->mdev->device, 4083 "removing port from unaffiliated list.\n"); 4084 mlx5_ib_dbg(dev, "port %d bound\n", i + 1); 4085 list_del(&mpi->list); 4086 break; 4087 } 4088 } 4089 if (!bound) 4090 mlx5_ib_dbg(dev, "no free port found for port %d\n", 4091 i + 1); 4092 } 4093 4094 list_add_tail(&dev->ib_dev_list, &mlx5_ib_dev_list); 4095 mutex_unlock(&mlx5_ib_multiport_mutex); 4096 return err; 4097 } 4098 4099 static void mlx5_ib_cleanup_multiport_master(struct mlx5_ib_dev *dev) 4100 { 4101 u32 port_num = mlx5_core_native_port_num(dev->mdev) - 1; 4102 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 4103 port_num + 1); 4104 u32 i; 4105 4106 if (!mlx5_core_is_mp_master(dev->mdev) || ll != IB_LINK_LAYER_ETHERNET) 4107 return; 4108 4109 mutex_lock(&mlx5_ib_multiport_mutex); 4110 for (i = 0; i < dev->num_ports; i++) { 4111 if (dev->port[i].mp.mpi) { 4112 /* Destroy the native port stub */ 4113 if (i == port_num) { 4114 kfree(dev->port[i].mp.mpi); 4115 dev->port[i].mp.mpi = NULL; 4116 } else { 4117 mlx5_ib_dbg(dev, "unbinding port_num: %u\n", 4118 i + 1); 4119 list_add_tail(&dev->port[i].mp.mpi->list, 4120 &mlx5_ib_unaffiliated_port_list); 4121 mlx5_ib_unbind_slave_port(dev, 4122 dev->port[i].mp.mpi); 4123 } 4124 } 4125 } 4126 4127 mlx5_ib_dbg(dev, "removing from devlist\n"); 4128 list_del(&dev->ib_dev_list); 4129 mutex_unlock(&mlx5_ib_multiport_mutex); 4130 4131 mlx5_nic_vport_disable_roce(dev->mdev); 4132 } 4133 4134 static int mmap_obj_cleanup(struct ib_uobject *uobject, 4135 enum rdma_remove_reason why, 4136 struct uverbs_attr_bundle *attrs) 4137 { 4138 struct mlx5_user_mmap_entry *obj = uobject->object; 4139 4140 rdma_user_mmap_entry_remove(&obj->rdma_entry); 4141 return 0; 4142 } 4143 4144 static int mlx5_rdma_user_mmap_entry_insert(struct mlx5_ib_ucontext *c, 4145 struct mlx5_user_mmap_entry *entry, 4146 size_t length) 4147 { 4148 return rdma_user_mmap_entry_insert_range( 4149 &c->ibucontext, &entry->rdma_entry, length, 4150 (MLX5_IB_MMAP_OFFSET_START << 16), 4151 ((MLX5_IB_MMAP_OFFSET_END << 16) + (1UL << 16) - 1)); 4152 } 4153 4154 static struct mlx5_user_mmap_entry * 4155 alloc_var_entry(struct mlx5_ib_ucontext *c, u32 flags) 4156 { 4157 struct mlx5_user_mmap_entry *entry; 4158 struct mlx5_var_region *var_region; 4159 struct mlx5_var_table *var_table; 4160 u32 page_idx; 4161 int err; 4162 4163 var_table = &to_mdev(c->ibucontext.device)->var_table; 4164 if (flags & MLX5_IB_UAPI_VAR_ALLOC_FLAG_TLP) 4165 var_region = &var_table->tlp_var_region; 4166 else 4167 var_region = &var_table->var_region; 4168 4169 entry = kzalloc_obj(*entry); 4170 if (!entry) 4171 return ERR_PTR(-ENOMEM); 4172 4173 mutex_lock(&var_region->bitmap_lock); 4174 page_idx = find_first_zero_bit(var_region->bitmap, 4175 var_region->num_var_hw_entries); 4176 if (page_idx >= var_region->num_var_hw_entries) { 4177 err = -ENOSPC; 4178 mutex_unlock(&var_region->bitmap_lock); 4179 goto end; 4180 } 4181 4182 set_bit(page_idx, var_region->bitmap); 4183 mutex_unlock(&var_region->bitmap_lock); 4184 4185 entry->address = var_region->hw_start_addr + 4186 (page_idx * var_region->stride_size); 4187 entry->page_idx = page_idx; 4188 entry->mmap_flag = flags & MLX5_IB_UAPI_VAR_ALLOC_FLAG_TLP ? 4189 MLX5_IB_MMAP_TYPE_TLP_VAR : 4190 MLX5_IB_MMAP_TYPE_VAR; 4191 4192 err = mlx5_rdma_user_mmap_entry_insert(c, entry, 4193 var_region->stride_size); 4194 if (err) 4195 goto err_insert; 4196 4197 return entry; 4198 4199 err_insert: 4200 mutex_lock(&var_region->bitmap_lock); 4201 clear_bit(page_idx, var_region->bitmap); 4202 mutex_unlock(&var_region->bitmap_lock); 4203 end: 4204 kfree(entry); 4205 return ERR_PTR(err); 4206 } 4207 4208 static int UVERBS_HANDLER(MLX5_IB_METHOD_VAR_OBJ_ALLOC)( 4209 struct uverbs_attr_bundle *attrs) 4210 { 4211 struct ib_uobject *uobj = uverbs_attr_get_uobject( 4212 attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE); 4213 struct mlx5_user_mmap_entry *entry; 4214 struct mlx5_ib_ucontext *c; 4215 u64 mmap_offset; 4216 u32 flags = 0; 4217 u32 length; 4218 int err; 4219 4220 c = to_mucontext(ib_uverbs_get_ucontext(attrs)); 4221 if (IS_ERR(c)) 4222 return PTR_ERR(c); 4223 4224 err = uverbs_get_flags32(&flags, attrs, 4225 MLX5_IB_ATTR_VAR_OBJ_ALLOC_FLAGS, 4226 MLX5_IB_UAPI_VAR_ALLOC_FLAG_TLP); 4227 if (err) 4228 return err; 4229 4230 if (flags & MLX5_IB_UAPI_VAR_ALLOC_FLAG_TLP) { 4231 if (!MLX5_CAP_GEN(to_mdev(c->ibucontext.device)->mdev, 4232 tlp_device_emulation_manager)) 4233 return -EOPNOTSUPP; 4234 } else { 4235 if (!(MLX5_CAP_GEN_64(to_mdev(c->ibucontext.device)->mdev, 4236 general_obj_types) & 4237 MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q)) 4238 return -EOPNOTSUPP; 4239 } 4240 4241 entry = alloc_var_entry(c, flags); 4242 if (IS_ERR(entry)) 4243 return PTR_ERR(entry); 4244 4245 mmap_offset = mlx5_entry_to_mmap_offset(entry); 4246 length = entry->rdma_entry.npages * PAGE_SIZE; 4247 uobj->object = entry; 4248 uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE); 4249 4250 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET, 4251 &mmap_offset, sizeof(mmap_offset)); 4252 if (err) 4253 return err; 4254 4255 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID, 4256 &entry->page_idx, sizeof(entry->page_idx)); 4257 if (err) 4258 return err; 4259 4260 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH, 4261 &length, sizeof(length)); 4262 return err; 4263 } 4264 4265 DECLARE_UVERBS_NAMED_METHOD( 4266 MLX5_IB_METHOD_VAR_OBJ_ALLOC, 4267 UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_ALLOC_HANDLE, 4268 MLX5_IB_OBJECT_VAR, 4269 UVERBS_ACCESS_NEW, 4270 UA_MANDATORY), 4271 UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_VAR_OBJ_ALLOC_FLAGS, 4272 enum mlx5_ib_uapi_var_alloc_flags, 4273 UA_OPTIONAL), 4274 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_PAGE_ID, 4275 UVERBS_ATTR_TYPE(u32), 4276 UA_MANDATORY), 4277 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_LENGTH, 4278 UVERBS_ATTR_TYPE(u32), 4279 UA_MANDATORY), 4280 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_VAR_OBJ_ALLOC_MMAP_OFFSET, 4281 UVERBS_ATTR_TYPE(u64), 4282 UA_MANDATORY)); 4283 4284 DECLARE_UVERBS_NAMED_METHOD_DESTROY( 4285 MLX5_IB_METHOD_VAR_OBJ_DESTROY, 4286 UVERBS_ATTR_IDR(MLX5_IB_ATTR_VAR_OBJ_DESTROY_HANDLE, 4287 MLX5_IB_OBJECT_VAR, 4288 UVERBS_ACCESS_DESTROY, 4289 UA_MANDATORY)); 4290 4291 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_VAR, 4292 UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup), 4293 &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_ALLOC), 4294 &UVERBS_METHOD(MLX5_IB_METHOD_VAR_OBJ_DESTROY)); 4295 4296 static bool var_is_supported(struct ib_device *device) 4297 { 4298 struct mlx5_ib_dev *dev = to_mdev(device); 4299 4300 return (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) & 4301 MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) || 4302 MLX5_CAP_GEN(dev->mdev, tlp_device_emulation_manager); 4303 } 4304 4305 static struct mlx5_user_mmap_entry * 4306 alloc_uar_entry(struct mlx5_ib_ucontext *c, 4307 enum mlx5_ib_uapi_uar_alloc_type alloc_type) 4308 { 4309 struct mlx5_user_mmap_entry *entry; 4310 struct mlx5_ib_dev *dev; 4311 u32 uar_index; 4312 int err; 4313 4314 entry = kzalloc_obj(*entry); 4315 if (!entry) 4316 return ERR_PTR(-ENOMEM); 4317 4318 dev = to_mdev(c->ibucontext.device); 4319 err = mlx5_cmd_uar_alloc(dev->mdev, &uar_index, c->devx_uid); 4320 if (err) 4321 goto end; 4322 4323 entry->page_idx = uar_index; 4324 entry->address = uar_index2paddress(dev, uar_index); 4325 if (alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF) 4326 entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_WC; 4327 else 4328 entry->mmap_flag = MLX5_IB_MMAP_TYPE_UAR_NC; 4329 4330 err = mlx5_rdma_user_mmap_entry_insert(c, entry, PAGE_SIZE); 4331 if (err) 4332 goto err_insert; 4333 4334 return entry; 4335 4336 err_insert: 4337 mlx5_cmd_uar_dealloc(dev->mdev, uar_index, c->devx_uid); 4338 end: 4339 kfree(entry); 4340 return ERR_PTR(err); 4341 } 4342 4343 static int UVERBS_HANDLER(MLX5_IB_METHOD_UAR_OBJ_ALLOC)( 4344 struct uverbs_attr_bundle *attrs) 4345 { 4346 struct ib_uobject *uobj = uverbs_attr_get_uobject( 4347 attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE); 4348 enum mlx5_ib_uapi_uar_alloc_type alloc_type; 4349 struct mlx5_ib_ucontext *c; 4350 struct mlx5_user_mmap_entry *entry; 4351 u64 mmap_offset; 4352 u32 length; 4353 int err; 4354 4355 c = to_mucontext(ib_uverbs_get_ucontext(attrs)); 4356 if (IS_ERR(c)) 4357 return PTR_ERR(c); 4358 4359 err = uverbs_get_const(&alloc_type, attrs, 4360 MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE); 4361 if (err) 4362 return err; 4363 4364 if (alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF && 4365 alloc_type != MLX5_IB_UAPI_UAR_ALLOC_TYPE_NC) 4366 return -EOPNOTSUPP; 4367 4368 if (!mlx5_wc_support_get(to_mdev(c->ibucontext.device)->mdev) && 4369 alloc_type == MLX5_IB_UAPI_UAR_ALLOC_TYPE_BF) 4370 return -EOPNOTSUPP; 4371 4372 entry = alloc_uar_entry(c, alloc_type); 4373 if (IS_ERR(entry)) 4374 return PTR_ERR(entry); 4375 4376 mmap_offset = mlx5_entry_to_mmap_offset(entry); 4377 length = entry->rdma_entry.npages * PAGE_SIZE; 4378 uobj->object = entry; 4379 uverbs_finalize_uobj_create(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE); 4380 4381 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET, 4382 &mmap_offset, sizeof(mmap_offset)); 4383 if (err) 4384 return err; 4385 4386 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID, 4387 &entry->page_idx, sizeof(entry->page_idx)); 4388 if (err) 4389 return err; 4390 4391 err = uverbs_copy_to(attrs, MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH, 4392 &length, sizeof(length)); 4393 return err; 4394 } 4395 4396 DECLARE_UVERBS_NAMED_METHOD( 4397 MLX5_IB_METHOD_UAR_OBJ_ALLOC, 4398 UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_ALLOC_HANDLE, 4399 MLX5_IB_OBJECT_UAR, 4400 UVERBS_ACCESS_NEW, 4401 UA_MANDATORY), 4402 UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_UAR_OBJ_ALLOC_TYPE, 4403 enum mlx5_ib_uapi_uar_alloc_type, 4404 UA_MANDATORY), 4405 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_PAGE_ID, 4406 UVERBS_ATTR_TYPE(u32), 4407 UA_MANDATORY), 4408 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_LENGTH, 4409 UVERBS_ATTR_TYPE(u32), 4410 UA_MANDATORY), 4411 UVERBS_ATTR_PTR_OUT(MLX5_IB_ATTR_UAR_OBJ_ALLOC_MMAP_OFFSET, 4412 UVERBS_ATTR_TYPE(u64), 4413 UA_MANDATORY)); 4414 4415 DECLARE_UVERBS_NAMED_METHOD_DESTROY( 4416 MLX5_IB_METHOD_UAR_OBJ_DESTROY, 4417 UVERBS_ATTR_IDR(MLX5_IB_ATTR_UAR_OBJ_DESTROY_HANDLE, 4418 MLX5_IB_OBJECT_UAR, 4419 UVERBS_ACCESS_DESTROY, 4420 UA_MANDATORY)); 4421 4422 DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_UAR, 4423 UVERBS_TYPE_ALLOC_IDR(mmap_obj_cleanup), 4424 &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_ALLOC), 4425 &UVERBS_METHOD(MLX5_IB_METHOD_UAR_OBJ_DESTROY)); 4426 4427 ADD_UVERBS_ATTRIBUTES_SIMPLE( 4428 mlx5_ib_query_context, 4429 UVERBS_OBJECT_DEVICE, 4430 UVERBS_METHOD_QUERY_CONTEXT, 4431 UVERBS_ATTR_PTR_OUT( 4432 MLX5_IB_ATTR_QUERY_CONTEXT_RESP_UCTX, 4433 UVERBS_ATTR_STRUCT(struct mlx5_ib_alloc_ucontext_resp, 4434 dump_fill_mkey), 4435 UA_MANDATORY)); 4436 4437 ADD_UVERBS_ATTRIBUTES_SIMPLE( 4438 mlx5_ib_reg_dmabuf_mr, 4439 UVERBS_OBJECT_MR, 4440 UVERBS_METHOD_REG_DMABUF_MR, 4441 UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_REG_DMABUF_MR_ACCESS_FLAGS, 4442 enum mlx5_ib_uapi_reg_dmabuf_flags, 4443 UA_OPTIONAL)); 4444 4445 static const struct uapi_definition mlx5_ib_defs[] = { 4446 UAPI_DEF_CHAIN(mlx5_ib_devx_defs), 4447 UAPI_DEF_CHAIN(mlx5_ib_flow_defs), 4448 UAPI_DEF_CHAIN(mlx5_ib_qos_defs), 4449 UAPI_DEF_CHAIN(mlx5_ib_std_types_defs), 4450 UAPI_DEF_CHAIN(mlx5_ib_dm_defs), 4451 UAPI_DEF_CHAIN(mlx5_ib_create_cq_defs), 4452 4453 UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context), 4454 UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_MR, &mlx5_ib_reg_dmabuf_mr), 4455 UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR, 4456 UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)), 4457 UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR), 4458 {} 4459 }; 4460 4461 static void mlx5_ib_stage_init_cleanup(struct mlx5_ib_dev *dev) 4462 { 4463 mlx5_cmd_cleanup_async_ctx(&dev->async_ctx); 4464 mlx5_ib_data_direct_cleanup(dev); 4465 mlx5_ib_cleanup_multiport_master(dev); 4466 WARN_ON(!xa_empty(&dev->odp_mkeys)); 4467 mutex_destroy(&dev->cap_mask_mutex); 4468 WARN_ON(!xa_empty(&dev->sig_mrs)); 4469 WARN_ON(!bitmap_empty(dev->dm.memic_alloc_pages, MLX5_MAX_MEMIC_PAGES)); 4470 mlx5r_macsec_dealloc_gids(dev); 4471 } 4472 4473 static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev) 4474 { 4475 struct mlx5_core_dev *mdev = dev->mdev; 4476 int err, i; 4477 4478 dev->ib_dev.node_type = RDMA_NODE_IB_CA; 4479 dev->ib_dev.local_dma_lkey = 0 /* not supported for now */; 4480 dev->ib_dev.dev.parent = mdev->device; 4481 dev->ib_dev.lag_flags = RDMA_LAG_FLAGS_HASH_ALL_SLAVES; 4482 4483 for (i = 0; i < dev->num_ports; i++) { 4484 spin_lock_init(&dev->port[i].mp.mpi_lock); 4485 dev->port[i].roce.dev = dev; 4486 dev->port[i].roce.native_port_num = i + 1; 4487 dev->port[i].roce.last_port_state = IB_PORT_DOWN; 4488 } 4489 4490 err = mlx5r_cmd_query_special_mkeys(dev); 4491 if (err) 4492 return err; 4493 4494 err = mlx5r_macsec_init_gids_and_devlist(dev); 4495 if (err) 4496 return err; 4497 4498 err = mlx5_ib_init_multiport_master(dev); 4499 if (err) 4500 goto err; 4501 4502 err = set_has_smi_cap(dev); 4503 if (err) 4504 goto err_mp; 4505 4506 err = mlx5_query_max_pkeys(&dev->ib_dev, &dev->pkey_table_len); 4507 if (err) 4508 goto err_mp; 4509 4510 if (mlx5_use_mad_ifc(dev)) 4511 get_ext_port_caps(dev); 4512 4513 dev->ib_dev.num_comp_vectors = mlx5_comp_vectors_max(mdev); 4514 4515 mutex_init(&dev->cap_mask_mutex); 4516 mutex_init(&dev->data_direct_lock); 4517 INIT_LIST_HEAD(&dev->qp_list); 4518 spin_lock_init(&dev->reset_flow_resource_lock); 4519 xa_init(&dev->odp_mkeys); 4520 xa_init(&dev->sig_mrs); 4521 atomic_set(&dev->mkey_var, 0); 4522 4523 spin_lock_init(&dev->dm.lock); 4524 dev->dm.dev = mdev; 4525 err = mlx5_ib_data_direct_init(dev); 4526 if (err) 4527 goto err_mp; 4528 4529 err = pcim_p2pdma_init(mdev->pdev); 4530 if (err && err != -EOPNOTSUPP) 4531 goto err_dd; 4532 4533 mlx5_cmd_init_async_ctx(mdev, &dev->async_ctx); 4534 4535 return 0; 4536 err_dd: 4537 mlx5_ib_data_direct_cleanup(dev); 4538 err_mp: 4539 mlx5_ib_cleanup_multiport_master(dev); 4540 err: 4541 mlx5r_macsec_dealloc_gids(dev); 4542 return err; 4543 } 4544 4545 static struct ib_device *mlx5_ib_add_sub_dev(struct ib_device *parent, 4546 enum rdma_nl_dev_type type, 4547 const char *name); 4548 static void mlx5_ib_del_sub_dev(struct ib_device *sub_dev); 4549 4550 static const struct ib_device_ops mlx5_ib_dev_ops = { 4551 .owner = THIS_MODULE, 4552 .driver_id = RDMA_DRIVER_MLX5, 4553 .uverbs_abi_ver = MLX5_IB_UVERBS_ABI_VERSION, 4554 4555 .add_gid = mlx5_ib_add_gid, 4556 .add_sub_dev = mlx5_ib_add_sub_dev, 4557 .alloc_mr = mlx5_ib_alloc_mr, 4558 .alloc_mr_integrity = mlx5_ib_alloc_mr_integrity, 4559 .alloc_pd = mlx5_ib_alloc_pd, 4560 .alloc_ucontext = mlx5_ib_alloc_ucontext, 4561 .attach_mcast = mlx5_ib_mcg_attach, 4562 .check_mr_status = mlx5_ib_check_mr_status, 4563 .create_ah = mlx5_ib_create_ah, 4564 .create_cq = mlx5_ib_create_cq, 4565 .create_user_cq = mlx5_ib_create_user_cq, 4566 .create_qp = mlx5_ib_create_qp, 4567 .create_srq = mlx5_ib_create_srq, 4568 .create_user_ah = mlx5_ib_create_ah, 4569 .dealloc_pd = mlx5_ib_dealloc_pd, 4570 .dealloc_ucontext = mlx5_ib_dealloc_ucontext, 4571 .del_gid = mlx5_ib_del_gid, 4572 .del_sub_dev = mlx5_ib_del_sub_dev, 4573 .dereg_mr = mlx5_ib_dereg_mr, 4574 .destroy_ah = mlx5_ib_destroy_ah, 4575 .destroy_cq = mlx5_ib_destroy_cq, 4576 .destroy_qp = mlx5_ib_destroy_qp, 4577 .destroy_srq = mlx5_ib_destroy_srq, 4578 .detach_mcast = mlx5_ib_mcg_detach, 4579 .disassociate_ucontext = mlx5_ib_disassociate_ucontext, 4580 .drain_rq = mlx5_ib_drain_rq, 4581 .drain_sq = mlx5_ib_drain_sq, 4582 .device_group = &mlx5_attr_group, 4583 .get_dev_fw_str = get_dev_fw_str, 4584 .get_dma_mr = mlx5_ib_get_dma_mr, 4585 .get_link_layer = mlx5_ib_port_link_layer, 4586 .map_mr_sg = mlx5_ib_map_mr_sg, 4587 .map_mr_sg_pi = mlx5_ib_map_mr_sg_pi, 4588 .mmap = mlx5_ib_mmap, 4589 .mmap_free = mlx5_ib_mmap_free, 4590 .mmap_get_pfns = mlx5_ib_mmap_get_pfns, 4591 .modify_cq = mlx5_ib_modify_cq, 4592 .modify_device = mlx5_ib_modify_device, 4593 .modify_port = mlx5_ib_modify_port, 4594 .modify_qp = mlx5_ib_modify_qp, 4595 .modify_srq = mlx5_ib_modify_srq, 4596 .pgoff_to_mmap_entry = mlx5_ib_pgoff_to_mmap_entry, 4597 .pre_destroy_cq = mlx5_ib_pre_destroy_cq, 4598 .poll_cq = mlx5_ib_poll_cq, 4599 .post_destroy_cq = mlx5_ib_post_destroy_cq, 4600 .post_recv = mlx5_ib_post_recv_nodrain, 4601 .post_send = mlx5_ib_post_send_nodrain, 4602 .post_srq_recv = mlx5_ib_post_srq_recv, 4603 .process_mad = mlx5_ib_process_mad, 4604 .query_ah = mlx5_ib_query_ah, 4605 .query_device = mlx5_ib_query_device, 4606 .query_gid = mlx5_ib_query_gid, 4607 .query_pkey = mlx5_ib_query_pkey, 4608 .query_port_speed = mlx5_ib_query_port_speed, 4609 .query_qp = mlx5_ib_query_qp, 4610 .query_srq = mlx5_ib_query_srq, 4611 .query_ucontext = mlx5_ib_query_ucontext, 4612 .reg_user_mr = mlx5_ib_reg_user_mr, 4613 .reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf, 4614 .req_notify_cq = mlx5_ib_arm_cq, 4615 .rereg_user_mr = mlx5_ib_rereg_user_mr, 4616 .resize_user_cq = mlx5_ib_resize_cq, 4617 .ufile_hw_cleanup = mlx5_ib_ufile_hw_cleanup, 4618 4619 INIT_RDMA_OBJ_SIZE(ib_ah, mlx5_ib_ah, ibah), 4620 INIT_RDMA_OBJ_SIZE(ib_counters, mlx5_ib_mcounters, ibcntrs), 4621 INIT_RDMA_OBJ_SIZE(ib_cq, mlx5_ib_cq, ibcq), 4622 INIT_RDMA_OBJ_SIZE(ib_dmah, mlx5_ib_dmah, ibdmah), 4623 INIT_RDMA_OBJ_SIZE(ib_pd, mlx5_ib_pd, ibpd), 4624 INIT_RDMA_OBJ_SIZE(ib_qp, mlx5_ib_qp, ibqp), 4625 INIT_RDMA_OBJ_SIZE(ib_srq, mlx5_ib_srq, ibsrq), 4626 INIT_RDMA_OBJ_SIZE(ib_ucontext, mlx5_ib_ucontext, ibucontext), 4627 }; 4628 4629 static const struct ib_device_ops mlx5_ib_dev_ipoib_enhanced_ops = { 4630 .rdma_netdev_get_params = mlx5_ib_rn_get_params, 4631 }; 4632 4633 static const struct ib_device_ops mlx5_ib_dev_sriov_ops = { 4634 .get_vf_config = mlx5_ib_get_vf_config, 4635 .get_vf_guid = mlx5_ib_get_vf_guid, 4636 .get_vf_stats = mlx5_ib_get_vf_stats, 4637 .set_vf_guid = mlx5_ib_set_vf_guid, 4638 .set_vf_link_state = mlx5_ib_set_vf_link_state, 4639 }; 4640 4641 static const struct ib_device_ops mlx5_ib_dev_mw_ops = { 4642 .alloc_mw = mlx5_ib_alloc_mw, 4643 .dealloc_mw = mlx5_ib_dealloc_mw, 4644 4645 INIT_RDMA_OBJ_SIZE(ib_mw, mlx5_ib_mw, ibmw), 4646 }; 4647 4648 static const struct ib_device_ops mlx5_ib_dev_xrc_ops = { 4649 .alloc_xrcd = mlx5_ib_alloc_xrcd, 4650 .dealloc_xrcd = mlx5_ib_dealloc_xrcd, 4651 4652 INIT_RDMA_OBJ_SIZE(ib_xrcd, mlx5_ib_xrcd, ibxrcd), 4653 }; 4654 4655 static int mlx5_ib_init_var_region(struct mlx5_ib_dev *dev) 4656 { 4657 struct mlx5_var_region *var_region = &dev->var_table.var_region; 4658 struct mlx5_core_dev *mdev = dev->mdev; 4659 u8 log_doorbell_bar_size; 4660 u8 log_doorbell_stride; 4661 u64 bar_size; 4662 4663 log_doorbell_bar_size = MLX5_CAP_DEV_VDPA_EMULATION(mdev, 4664 log_doorbell_bar_size); 4665 log_doorbell_stride = MLX5_CAP_DEV_VDPA_EMULATION(mdev, 4666 log_doorbell_stride); 4667 var_region->hw_start_addr = dev->mdev->bar_addr + 4668 MLX5_CAP64_DEV_VDPA_EMULATION(mdev, 4669 doorbell_bar_offset); 4670 bar_size = (1ULL << log_doorbell_bar_size) * 4096; 4671 var_region->stride_size = 1ULL << log_doorbell_stride; 4672 var_region->num_var_hw_entries = div_u64(bar_size, 4673 var_region->stride_size); 4674 mutex_init(&var_region->bitmap_lock); 4675 var_region->bitmap = bitmap_zalloc(var_region->num_var_hw_entries, 4676 GFP_KERNEL); 4677 return (var_region->bitmap) ? 0 : -ENOMEM; 4678 } 4679 4680 static int mlx5_ib_init_tlp_var_region(struct mlx5_ib_dev *dev) 4681 { 4682 struct mlx5_var_region *var_region = &dev->var_table.tlp_var_region; 4683 struct mlx5_core_dev *mdev = dev->mdev; 4684 u8 log_tlp_var_stride; 4685 4686 log_tlp_var_stride = 4687 MLX5_CAP_DEV_TLP_EMULATION(mdev, log_tlp_rsp_gw_page_stride); 4688 var_region->hw_start_addr = 4689 dev->mdev->bar_addr + 4690 MLX5_CAP64_DEV_TLP_EMULATION(mdev, tlp_rsp_gw_pages_bar_offset); 4691 4692 var_region->stride_size = (1ULL << log_tlp_var_stride) * 4096; 4693 var_region->num_var_hw_entries = 4694 MLX5_CAP_DEV_TLP_EMULATION(mdev, tlp_rsp_gw_num_pages); 4695 4696 mutex_init(&var_region->bitmap_lock); 4697 var_region->bitmap = bitmap_zalloc(var_region->num_var_hw_entries, 4698 GFP_KERNEL); 4699 return (var_region->bitmap) ? 0 : -ENOMEM; 4700 } 4701 4702 static void mlx5_ib_cleanup_ucaps(struct mlx5_ib_dev *dev) 4703 { 4704 if (MLX5_CAP_GEN(dev->mdev, uctx_cap) & MLX5_UCTX_CAP_RDMA_CTRL) 4705 ib_remove_ucap(RDMA_UCAP_MLX5_CTRL_LOCAL); 4706 4707 if (MLX5_CAP_GEN(dev->mdev, uctx_cap) & 4708 MLX5_UCTX_CAP_RDMA_CTRL_OTHER_VHCA) 4709 ib_remove_ucap(RDMA_UCAP_MLX5_CTRL_OTHER_VHCA); 4710 } 4711 4712 static int mlx5_ib_init_ucaps(struct mlx5_ib_dev *dev) 4713 { 4714 int ret; 4715 4716 if (MLX5_CAP_GEN(dev->mdev, uctx_cap) & MLX5_UCTX_CAP_RDMA_CTRL) { 4717 ret = ib_create_ucap(RDMA_UCAP_MLX5_CTRL_LOCAL); 4718 if (ret) 4719 return ret; 4720 } 4721 4722 if (MLX5_CAP_GEN(dev->mdev, uctx_cap) & 4723 MLX5_UCTX_CAP_RDMA_CTRL_OTHER_VHCA) { 4724 ret = ib_create_ucap(RDMA_UCAP_MLX5_CTRL_OTHER_VHCA); 4725 if (ret) 4726 goto remove_local; 4727 } 4728 4729 return 0; 4730 4731 remove_local: 4732 if (MLX5_CAP_GEN(dev->mdev, uctx_cap) & MLX5_UCTX_CAP_RDMA_CTRL) 4733 ib_remove_ucap(RDMA_UCAP_MLX5_CTRL_LOCAL); 4734 return ret; 4735 } 4736 4737 static void mlx5_ib_cleanup_var_table(struct mlx5_ib_dev *dev) 4738 { 4739 bitmap_free(dev->var_table.var_region.bitmap); 4740 bitmap_free(dev->var_table.tlp_var_region.bitmap); 4741 } 4742 4743 static void mlx5_ib_stage_caps_cleanup(struct mlx5_ib_dev *dev) 4744 { 4745 if (MLX5_CAP_GEN_2_64(dev->mdev, general_obj_types_127_64) & 4746 MLX5_HCA_CAP_2_GENERAL_OBJECT_TYPES_RDMA_CTRL) 4747 mlx5_ib_cleanup_ucaps(dev); 4748 4749 mlx5_ib_cleanup_var_table(dev); 4750 } 4751 4752 static int mlx5_ib_stage_caps_init(struct mlx5_ib_dev *dev) 4753 { 4754 struct mlx5_core_dev *mdev = dev->mdev; 4755 int err; 4756 4757 if (MLX5_CAP_GEN(mdev, ipoib_enhanced_offloads) && 4758 IS_ENABLED(CONFIG_MLX5_CORE_IPOIB)) 4759 ib_set_device_ops(&dev->ib_dev, 4760 &mlx5_ib_dev_ipoib_enhanced_ops); 4761 4762 if (mlx5_core_is_pf(mdev)) 4763 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_sriov_ops); 4764 4765 dev->umr_fence = mlx5_get_umr_fence(MLX5_CAP_GEN(mdev, umr_fence)); 4766 4767 if (MLX5_CAP_GEN(mdev, imaicl)) 4768 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_mw_ops); 4769 4770 if (MLX5_CAP_GEN(mdev, xrc)) 4771 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_xrc_ops); 4772 4773 if (MLX5_CAP_DEV_MEM(mdev, memic) || 4774 MLX5_CAP_GEN_64(dev->mdev, general_obj_types) & 4775 MLX5_GENERAL_OBJ_TYPES_CAP_SW_ICM) 4776 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dm_ops); 4777 4778 if (mdev->st) 4779 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_dmah_ops); 4780 4781 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_ops); 4782 4783 if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS)) 4784 dev->ib_dev.driver_def = mlx5_ib_defs; 4785 4786 err = init_node_data(dev); 4787 if (err) 4788 return err; 4789 4790 if ((MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_ETH) && 4791 (MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) || 4792 MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc))) 4793 mutex_init(&dev->lb.mutex); 4794 4795 if (MLX5_CAP_GEN_64(dev->mdev, general_obj_types) & 4796 MLX5_GENERAL_OBJ_TYPES_CAP_VIRTIO_NET_Q) { 4797 err = mlx5_ib_init_var_region(dev); 4798 if (err) 4799 return err; 4800 } 4801 4802 if (MLX5_CAP_GEN_2_64(dev->mdev, general_obj_types_127_64) & 4803 MLX5_HCA_CAP_2_GENERAL_OBJECT_TYPES_RDMA_CTRL) { 4804 err = mlx5_ib_init_ucaps(dev); 4805 if (err) 4806 goto err_ucaps; 4807 } 4808 4809 if (MLX5_CAP_GEN(dev->mdev, tlp_device_emulation_manager)) { 4810 err = mlx5_ib_init_tlp_var_region(dev); 4811 if (err) 4812 goto err_tlp_var; 4813 } 4814 4815 dev->ib_dev.use_cq_dim = true; 4816 4817 return 0; 4818 4819 err_tlp_var: 4820 mlx5_ib_cleanup_ucaps(dev); 4821 err_ucaps: 4822 bitmap_free(dev->var_table.var_region.bitmap); 4823 return err; 4824 } 4825 4826 static const struct ib_device_ops mlx5_ib_dev_port_ops = { 4827 .get_port_immutable = mlx5_port_immutable, 4828 .query_port = mlx5_ib_query_port, 4829 }; 4830 4831 static int mlx5_ib_stage_non_default_cb(struct mlx5_ib_dev *dev) 4832 { 4833 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_ops); 4834 return 0; 4835 } 4836 4837 static const struct ib_device_ops mlx5_ib_dev_port_rep_ops = { 4838 .get_port_immutable = mlx5_port_rep_immutable, 4839 .query_port = mlx5_ib_rep_query_port, 4840 .query_pkey = mlx5_ib_rep_query_pkey, 4841 }; 4842 4843 static int mlx5_ib_stage_raw_eth_non_default_cb(struct mlx5_ib_dev *dev) 4844 { 4845 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_port_rep_ops); 4846 return 0; 4847 } 4848 4849 static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = { 4850 .create_rwq_ind_table = mlx5_ib_create_rwq_ind_table, 4851 .create_wq = mlx5_ib_create_wq, 4852 .destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table, 4853 .destroy_wq = mlx5_ib_destroy_wq, 4854 .modify_wq = mlx5_ib_modify_wq, 4855 4856 INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table, 4857 ib_rwq_ind_tbl), 4858 }; 4859 4860 static int mlx5_ib_roce_init(struct mlx5_ib_dev *dev) 4861 { 4862 struct mlx5_core_dev *mdev = dev->mdev; 4863 enum rdma_link_layer ll; 4864 int port_type_cap; 4865 u32 port_num = 0; 4866 int err; 4867 4868 port_type_cap = MLX5_CAP_GEN(mdev, port_type); 4869 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap); 4870 4871 if (ll == IB_LINK_LAYER_ETHERNET) { 4872 ib_set_device_ops(&dev->ib_dev, &mlx5_ib_dev_common_roce_ops); 4873 4874 port_num = mlx5_core_native_port_num(dev->mdev) - 1; 4875 4876 /* Register only for native ports */ 4877 mlx5_mdev_netdev_track(dev, port_num); 4878 4879 err = mlx5_enable_eth(dev); 4880 if (err) 4881 goto cleanup; 4882 } 4883 4884 return 0; 4885 cleanup: 4886 mlx5_mdev_netdev_untrack(dev, port_num); 4887 return err; 4888 } 4889 4890 static void mlx5_ib_roce_cleanup(struct mlx5_ib_dev *dev) 4891 { 4892 struct mlx5_core_dev *mdev = dev->mdev; 4893 enum rdma_link_layer ll; 4894 int port_type_cap; 4895 u32 port_num; 4896 4897 port_type_cap = MLX5_CAP_GEN(mdev, port_type); 4898 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap); 4899 4900 if (ll == IB_LINK_LAYER_ETHERNET) { 4901 mlx5_disable_eth(dev); 4902 4903 port_num = mlx5_core_native_port_num(dev->mdev) - 1; 4904 mlx5_mdev_netdev_untrack(dev, port_num); 4905 } 4906 } 4907 4908 static int mlx5_ib_stage_cong_debugfs_init(struct mlx5_ib_dev *dev) 4909 { 4910 mlx5_ib_init_cong_debugfs(dev, 4911 mlx5_core_native_port_num(dev->mdev) - 1); 4912 return 0; 4913 } 4914 4915 static void mlx5_ib_stage_cong_debugfs_cleanup(struct mlx5_ib_dev *dev) 4916 { 4917 mlx5_ib_cleanup_cong_debugfs(dev, 4918 mlx5_core_native_port_num(dev->mdev) - 1); 4919 } 4920 4921 static int mlx5_ib_stage_bfrag_init(struct mlx5_ib_dev *dev) 4922 { 4923 int err; 4924 4925 err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false); 4926 if (err) 4927 return err; 4928 4929 err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true); 4930 if (err) 4931 mlx5_free_bfreg(dev->mdev, &dev->bfreg); 4932 4933 return err; 4934 } 4935 4936 static void mlx5_ib_stage_bfrag_cleanup(struct mlx5_ib_dev *dev) 4937 { 4938 mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg); 4939 mlx5_free_bfreg(dev->mdev, &dev->bfreg); 4940 } 4941 4942 static int mlx5_ib_stage_ib_reg_init(struct mlx5_ib_dev *dev) 4943 { 4944 const char *name; 4945 4946 if (dev->sub_dev_name) { 4947 name = dev->sub_dev_name; 4948 ib_mark_name_assigned_by_user(&dev->ib_dev); 4949 } else if (!mlx5_lag_is_active(dev->mdev)) 4950 name = "mlx5_%d"; 4951 else 4952 name = "mlx5_bond_%d"; 4953 return ib_register_device(&dev->ib_dev, name, &dev->mdev->pdev->dev); 4954 } 4955 4956 static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev) 4957 { 4958 mlx5r_frmr_pools_cleanup(&dev->ib_dev); 4959 mlx5r_umr_resource_cleanup(dev); 4960 mlx5r_umr_cleanup(dev); 4961 } 4962 4963 static void mlx5_ib_stage_ib_reg_cleanup(struct mlx5_ib_dev *dev) 4964 { 4965 ib_unregister_device(&dev->ib_dev); 4966 } 4967 4968 static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev) 4969 { 4970 int ret; 4971 4972 ret = mlx5r_umr_init(dev); 4973 if (ret) 4974 return ret; 4975 4976 ret = mlx5r_frmr_pools_init(&dev->ib_dev); 4977 if (ret) 4978 mlx5_ib_warn(dev, "frmr pools init failed %d\n", ret); 4979 4980 return ret; 4981 } 4982 4983 static int mlx5_ib_stage_delay_drop_init(struct mlx5_ib_dev *dev) 4984 { 4985 struct dentry *root; 4986 4987 if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP)) 4988 return 0; 4989 4990 mutex_init(&dev->delay_drop.lock); 4991 dev->delay_drop.dev = dev; 4992 dev->delay_drop.activate = false; 4993 dev->delay_drop.timeout = MLX5_MAX_DELAY_DROP_TIMEOUT_MS * 1000; 4994 INIT_WORK(&dev->delay_drop.delay_drop_work, delay_drop_handler); 4995 atomic_set(&dev->delay_drop.rqs_cnt, 0); 4996 atomic_set(&dev->delay_drop.events_cnt, 0); 4997 4998 if (!mlx5_debugfs_root) 4999 return 0; 5000 5001 root = debugfs_create_dir("delay_drop", mlx5_debugfs_get_dev_root(dev->mdev)); 5002 dev->delay_drop.dir_debugfs = root; 5003 5004 debugfs_create_atomic_t("num_timeout_events", 0400, root, 5005 &dev->delay_drop.events_cnt); 5006 debugfs_create_atomic_t("num_rqs", 0400, root, 5007 &dev->delay_drop.rqs_cnt); 5008 debugfs_create_file("timeout", 0600, root, &dev->delay_drop, 5009 &fops_delay_drop_timeout); 5010 return 0; 5011 } 5012 5013 static void mlx5_ib_stage_delay_drop_cleanup(struct mlx5_ib_dev *dev) 5014 { 5015 if (!(dev->ib_dev.attrs.raw_packet_caps & IB_RAW_PACKET_CAP_DELAY_DROP)) 5016 return; 5017 5018 cancel_work_sync(&dev->delay_drop.delay_drop_work); 5019 if (!dev->delay_drop.dir_debugfs) 5020 return; 5021 5022 debugfs_remove_recursive(dev->delay_drop.dir_debugfs); 5023 dev->delay_drop.dir_debugfs = NULL; 5024 } 5025 5026 static int mlx5_ib_stage_dev_notifier_init(struct mlx5_ib_dev *dev) 5027 { 5028 struct mlx5_ib_resources *devr = &dev->devr; 5029 int port; 5030 5031 for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) 5032 INIT_WORK(&devr->ports[port].pkey_change_work, 5033 pkey_change_handler); 5034 5035 dev->mdev_events.notifier_call = mlx5_ib_event; 5036 mlx5_notifier_register(dev->mdev, &dev->mdev_events); 5037 5038 mlx5r_macsec_event_register(dev); 5039 5040 return 0; 5041 } 5042 5043 static void mlx5_ib_stage_dev_notifier_cleanup(struct mlx5_ib_dev *dev) 5044 { 5045 struct mlx5_ib_resources *devr = &dev->devr; 5046 int port; 5047 5048 mlx5r_macsec_event_unregister(dev); 5049 mlx5_notifier_unregister(dev->mdev, &dev->mdev_events); 5050 5051 for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) 5052 cancel_work_sync(&devr->ports[port].pkey_change_work); 5053 } 5054 5055 void mlx5_ib_data_direct_bind(struct mlx5_ib_dev *ibdev, 5056 struct mlx5_data_direct_dev *dev) 5057 { 5058 mutex_lock(&ibdev->data_direct_lock); 5059 ibdev->data_direct_dev = dev; 5060 mutex_unlock(&ibdev->data_direct_lock); 5061 } 5062 5063 void mlx5_ib_data_direct_unbind(struct mlx5_ib_dev *ibdev) 5064 { 5065 mutex_lock(&ibdev->data_direct_lock); 5066 mlx5_ib_revoke_data_direct_mrs(ibdev); 5067 ibdev->data_direct_dev = NULL; 5068 mutex_unlock(&ibdev->data_direct_lock); 5069 } 5070 5071 void __mlx5_ib_remove(struct mlx5_ib_dev *dev, 5072 const struct mlx5_ib_profile *profile, 5073 int stage) 5074 { 5075 dev->ib_active = false; 5076 5077 /* Number of stages to cleanup */ 5078 while (stage) { 5079 stage--; 5080 if (profile->stage[stage].cleanup) 5081 profile->stage[stage].cleanup(dev); 5082 } 5083 5084 kfree(dev->port); 5085 ib_dealloc_device(&dev->ib_dev); 5086 } 5087 5088 int __mlx5_ib_add(struct mlx5_ib_dev *dev, 5089 const struct mlx5_ib_profile *profile) 5090 { 5091 int err; 5092 int i; 5093 5094 dev->profile = profile; 5095 5096 for (i = 0; i < MLX5_IB_STAGE_MAX; i++) { 5097 if (profile->stage[i].init) { 5098 err = profile->stage[i].init(dev); 5099 if (err) 5100 goto err_out; 5101 } 5102 } 5103 5104 dev->ib_active = true; 5105 return 0; 5106 5107 err_out: 5108 /* Clean up stages which were initialized */ 5109 while (i) { 5110 i--; 5111 if (profile->stage[i].cleanup) 5112 profile->stage[i].cleanup(dev); 5113 } 5114 return -ENOMEM; 5115 } 5116 5117 static const struct mlx5_ib_profile pf_profile = { 5118 STAGE_CREATE(MLX5_IB_STAGE_INIT, 5119 mlx5_ib_stage_init_init, 5120 mlx5_ib_stage_init_cleanup), 5121 STAGE_CREATE(MLX5_IB_STAGE_FS, 5122 mlx5_ib_fs_init, 5123 mlx5_ib_fs_cleanup), 5124 STAGE_CREATE(MLX5_IB_STAGE_CAPS, 5125 mlx5_ib_stage_caps_init, 5126 mlx5_ib_stage_caps_cleanup), 5127 STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB, 5128 mlx5_ib_stage_non_default_cb, 5129 NULL), 5130 STAGE_CREATE(MLX5_IB_STAGE_ROCE, 5131 mlx5_ib_roce_init, 5132 mlx5_ib_roce_cleanup), 5133 STAGE_CREATE(MLX5_IB_STAGE_QP, 5134 mlx5_init_qp_table, 5135 mlx5_cleanup_qp_table), 5136 STAGE_CREATE(MLX5_IB_STAGE_SRQ, 5137 mlx5_init_srq_table, 5138 mlx5_cleanup_srq_table), 5139 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES, 5140 mlx5_ib_dev_res_init, 5141 mlx5_ib_dev_res_cleanup), 5142 STAGE_CREATE(MLX5_IB_STAGE_ODP, 5143 mlx5_ib_odp_init_one, 5144 mlx5_ib_odp_cleanup_one), 5145 STAGE_CREATE(MLX5_IB_STAGE_COUNTERS, 5146 mlx5_ib_counters_init, 5147 mlx5_ib_counters_cleanup), 5148 STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS, 5149 mlx5_ib_stage_cong_debugfs_init, 5150 mlx5_ib_stage_cong_debugfs_cleanup), 5151 STAGE_CREATE(MLX5_IB_STAGE_BFREG, 5152 mlx5_ib_stage_bfrag_init, 5153 mlx5_ib_stage_bfrag_cleanup), 5154 STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR, 5155 NULL, 5156 mlx5_ib_stage_pre_ib_reg_umr_cleanup), 5157 STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID, 5158 mlx5_ib_devx_init, 5159 mlx5_ib_devx_cleanup), 5160 STAGE_CREATE(MLX5_IB_STAGE_SYS_ERROR_NOTIFIER, 5161 mlx5_ib_stage_sys_error_notifier_init, 5162 mlx5_ib_stage_sys_error_notifier_cleanup), 5163 STAGE_CREATE(MLX5_IB_STAGE_IB_REG, 5164 mlx5_ib_stage_ib_reg_init, 5165 mlx5_ib_stage_ib_reg_cleanup), 5166 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER, 5167 mlx5_ib_stage_dev_notifier_init, 5168 mlx5_ib_stage_dev_notifier_cleanup), 5169 STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR, 5170 mlx5_ib_stage_post_ib_reg_umr_init, 5171 NULL), 5172 STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP, 5173 mlx5_ib_stage_delay_drop_init, 5174 mlx5_ib_stage_delay_drop_cleanup), 5175 STAGE_CREATE(MLX5_IB_STAGE_RESTRACK, 5176 mlx5_ib_restrack_init, 5177 NULL), 5178 }; 5179 5180 const struct mlx5_ib_profile raw_eth_profile = { 5181 STAGE_CREATE(MLX5_IB_STAGE_INIT, 5182 mlx5_ib_stage_init_init, 5183 mlx5_ib_stage_init_cleanup), 5184 STAGE_CREATE(MLX5_IB_STAGE_FS, 5185 mlx5_ib_fs_init, 5186 mlx5_ib_fs_cleanup), 5187 STAGE_CREATE(MLX5_IB_STAGE_CAPS, 5188 mlx5_ib_stage_caps_init, 5189 mlx5_ib_stage_caps_cleanup), 5190 STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB, 5191 mlx5_ib_stage_raw_eth_non_default_cb, 5192 NULL), 5193 STAGE_CREATE(MLX5_IB_STAGE_ROCE, 5194 mlx5_ib_roce_init, 5195 mlx5_ib_roce_cleanup), 5196 STAGE_CREATE(MLX5_IB_STAGE_QP, 5197 mlx5_init_qp_table, 5198 mlx5_cleanup_qp_table), 5199 STAGE_CREATE(MLX5_IB_STAGE_SRQ, 5200 mlx5_init_srq_table, 5201 mlx5_cleanup_srq_table), 5202 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES, 5203 mlx5_ib_dev_res_init, 5204 mlx5_ib_dev_res_cleanup), 5205 STAGE_CREATE(MLX5_IB_STAGE_COUNTERS, 5206 mlx5_ib_counters_init, 5207 mlx5_ib_counters_cleanup), 5208 STAGE_CREATE(MLX5_IB_STAGE_CONG_DEBUGFS, 5209 mlx5_ib_stage_cong_debugfs_init, 5210 mlx5_ib_stage_cong_debugfs_cleanup), 5211 STAGE_CREATE(MLX5_IB_STAGE_BFREG, 5212 mlx5_ib_stage_bfrag_init, 5213 mlx5_ib_stage_bfrag_cleanup), 5214 STAGE_CREATE(MLX5_IB_STAGE_PRE_IB_REG_UMR, 5215 NULL, 5216 mlx5_ib_stage_pre_ib_reg_umr_cleanup), 5217 STAGE_CREATE(MLX5_IB_STAGE_WHITELIST_UID, 5218 mlx5_ib_devx_init, 5219 mlx5_ib_devx_cleanup), 5220 STAGE_CREATE(MLX5_IB_STAGE_SYS_ERROR_NOTIFIER, 5221 mlx5_ib_stage_sys_error_notifier_init, 5222 mlx5_ib_stage_sys_error_notifier_cleanup), 5223 STAGE_CREATE(MLX5_IB_STAGE_IB_REG, 5224 mlx5_ib_stage_ib_reg_init, 5225 mlx5_ib_stage_ib_reg_cleanup), 5226 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_NOTIFIER, 5227 mlx5_ib_stage_dev_notifier_init, 5228 mlx5_ib_stage_dev_notifier_cleanup), 5229 STAGE_CREATE(MLX5_IB_STAGE_POST_IB_REG_UMR, 5230 mlx5_ib_stage_post_ib_reg_umr_init, 5231 NULL), 5232 STAGE_CREATE(MLX5_IB_STAGE_DELAY_DROP, 5233 mlx5_ib_stage_delay_drop_init, 5234 mlx5_ib_stage_delay_drop_cleanup), 5235 STAGE_CREATE(MLX5_IB_STAGE_RESTRACK, 5236 mlx5_ib_restrack_init, 5237 NULL), 5238 }; 5239 5240 static const struct mlx5_ib_profile plane_profile = { 5241 STAGE_CREATE(MLX5_IB_STAGE_INIT, 5242 mlx5_ib_stage_init_init, 5243 mlx5_ib_stage_init_cleanup), 5244 STAGE_CREATE(MLX5_IB_STAGE_CAPS, 5245 mlx5_ib_stage_caps_init, 5246 mlx5_ib_stage_caps_cleanup), 5247 STAGE_CREATE(MLX5_IB_STAGE_NON_DEFAULT_CB, 5248 mlx5_ib_stage_non_default_cb, 5249 NULL), 5250 STAGE_CREATE(MLX5_IB_STAGE_QP, 5251 mlx5_init_qp_table, 5252 mlx5_cleanup_qp_table), 5253 STAGE_CREATE(MLX5_IB_STAGE_SRQ, 5254 mlx5_init_srq_table, 5255 mlx5_cleanup_srq_table), 5256 STAGE_CREATE(MLX5_IB_STAGE_DEVICE_RESOURCES, 5257 mlx5_ib_dev_res_init, 5258 mlx5_ib_dev_res_cleanup), 5259 STAGE_CREATE(MLX5_IB_STAGE_BFREG, 5260 mlx5_ib_stage_bfrag_init, 5261 mlx5_ib_stage_bfrag_cleanup), 5262 STAGE_CREATE(MLX5_IB_STAGE_IB_REG, 5263 mlx5_ib_stage_ib_reg_init, 5264 mlx5_ib_stage_ib_reg_cleanup), 5265 }; 5266 5267 static struct ib_device *mlx5_ib_add_sub_dev(struct ib_device *parent, 5268 enum rdma_nl_dev_type type, 5269 const char *name) 5270 { 5271 struct mlx5_ib_dev *mparent = to_mdev(parent), *mplane; 5272 enum rdma_link_layer ll; 5273 int ret; 5274 5275 if (mparent->smi_dev) 5276 return ERR_PTR(-EEXIST); 5277 5278 ll = mlx5_port_type_cap_to_rdma_ll(MLX5_CAP_GEN(mparent->mdev, 5279 port_type)); 5280 if (type != RDMA_DEVICE_TYPE_SMI || !mparent->num_plane || 5281 ll != IB_LINK_LAYER_INFINIBAND || 5282 !MLX5_CAP_GEN_2(mparent->mdev, multiplane_qp_ud)) 5283 return ERR_PTR(-EOPNOTSUPP); 5284 5285 mplane = ib_alloc_device_with_net(mlx5_ib_dev, ib_dev, 5286 mlx5_core_net(mparent->mdev)); 5287 if (!mplane) 5288 return ERR_PTR(-ENOMEM); 5289 5290 mplane->port = kzalloc_objs(*mplane->port, 5291 mparent->num_plane * mparent->num_ports); 5292 if (!mplane->port) { 5293 ret = -ENOMEM; 5294 goto fail_kcalloc; 5295 } 5296 5297 mplane->ib_dev.type = type; 5298 mplane->mdev = mparent->mdev; 5299 mplane->num_ports = mparent->num_plane; 5300 mplane->sub_dev_name = name; 5301 mplane->ib_dev.phys_port_cnt = mplane->num_ports; 5302 5303 ret = __mlx5_ib_add(mplane, &plane_profile); 5304 if (ret) 5305 goto fail_ib_add; 5306 5307 mparent->smi_dev = mplane; 5308 return &mplane->ib_dev; 5309 5310 fail_ib_add: 5311 kfree(mplane->port); 5312 fail_kcalloc: 5313 ib_dealloc_device(&mplane->ib_dev); 5314 return ERR_PTR(ret); 5315 } 5316 5317 static void mlx5_ib_del_sub_dev(struct ib_device *sub_dev) 5318 { 5319 struct mlx5_ib_dev *mdev = to_mdev(sub_dev); 5320 5321 to_mdev(sub_dev->parent)->smi_dev = NULL; 5322 __mlx5_ib_remove(mdev, mdev->profile, MLX5_IB_STAGE_MAX); 5323 } 5324 5325 static int mlx5r_mp_probe(struct auxiliary_device *adev, 5326 const struct auxiliary_device_id *id) 5327 { 5328 struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev); 5329 struct mlx5_core_dev *mdev = idev->mdev; 5330 struct mlx5_ib_multiport_info *mpi; 5331 struct mlx5_ib_dev *dev; 5332 bool bound = false; 5333 int err; 5334 5335 mpi = kzalloc_obj(*mpi); 5336 if (!mpi) 5337 return -ENOMEM; 5338 5339 mpi->mdev = mdev; 5340 err = mlx5_query_nic_vport_system_image_guid(mdev, 5341 &mpi->sys_image_guid); 5342 if (err) { 5343 kfree(mpi); 5344 return err; 5345 } 5346 5347 mutex_lock(&mlx5_ib_multiport_mutex); 5348 list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) { 5349 if (dev->sys_image_guid == mpi->sys_image_guid && 5350 mlx5_core_same_coredev_type(dev->mdev, mpi->mdev)) 5351 bound = mlx5_ib_bind_slave_port(dev, mpi); 5352 5353 if (bound) { 5354 rdma_roce_rescan_device(&dev->ib_dev); 5355 mpi->ibdev->ib_active = true; 5356 break; 5357 } 5358 } 5359 5360 if (!bound) { 5361 list_add_tail(&mpi->list, &mlx5_ib_unaffiliated_port_list); 5362 dev_dbg(mdev->device, 5363 "no suitable IB device found to bind to, added to unaffiliated list.\n"); 5364 } 5365 mutex_unlock(&mlx5_ib_multiport_mutex); 5366 5367 auxiliary_set_drvdata(adev, mpi); 5368 return 0; 5369 } 5370 5371 static void mlx5r_mp_remove(struct auxiliary_device *adev) 5372 { 5373 struct mlx5_ib_multiport_info *mpi; 5374 5375 mpi = auxiliary_get_drvdata(adev); 5376 mutex_lock(&mlx5_ib_multiport_mutex); 5377 if (mpi->ibdev) 5378 mlx5_ib_unbind_slave_port(mpi->ibdev, mpi); 5379 else 5380 list_del(&mpi->list); 5381 mutex_unlock(&mlx5_ib_multiport_mutex); 5382 kfree(mpi); 5383 } 5384 5385 static int mlx5r_probe(struct auxiliary_device *adev, 5386 const struct auxiliary_device_id *id) 5387 { 5388 struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev); 5389 struct mlx5_core_dev *mdev = idev->mdev; 5390 const struct mlx5_ib_profile *profile; 5391 int port_type_cap, num_ports, ret; 5392 enum rdma_link_layer ll; 5393 struct mlx5_ib_dev *dev; 5394 5395 port_type_cap = MLX5_CAP_GEN(mdev, port_type); 5396 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap); 5397 5398 num_ports = max(MLX5_CAP_GEN(mdev, num_ports), 5399 MLX5_CAP_GEN(mdev, num_vhca_ports)); 5400 dev = ib_alloc_device_with_net(mlx5_ib_dev, ib_dev, 5401 mlx5_core_net(mdev)); 5402 if (!dev) 5403 return -ENOMEM; 5404 5405 if (ll == IB_LINK_LAYER_INFINIBAND) { 5406 ret = mlx5_ib_get_plane_num(mdev, &dev->num_plane); 5407 if (ret) 5408 goto fail; 5409 } 5410 5411 dev->port = kzalloc_objs(*dev->port, num_ports); 5412 if (!dev->port) { 5413 ret = -ENOMEM; 5414 goto fail; 5415 } 5416 5417 dev->mdev = mdev; 5418 dev->num_ports = num_ports; 5419 dev->ib_dev.phys_port_cnt = num_ports; 5420 5421 if (ll == IB_LINK_LAYER_ETHERNET && !mlx5_get_roce_state(mdev)) 5422 profile = &raw_eth_profile; 5423 else 5424 profile = &pf_profile; 5425 5426 ret = __mlx5_ib_add(dev, profile); 5427 if (ret) 5428 goto fail_ib_add; 5429 5430 auxiliary_set_drvdata(adev, dev); 5431 return 0; 5432 5433 fail_ib_add: 5434 kfree(dev->port); 5435 fail: 5436 ib_dealloc_device(&dev->ib_dev); 5437 return ret; 5438 } 5439 5440 static void mlx5r_remove(struct auxiliary_device *adev) 5441 { 5442 struct mlx5_ib_dev *dev; 5443 5444 dev = auxiliary_get_drvdata(adev); 5445 __mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX); 5446 } 5447 5448 static const struct auxiliary_device_id mlx5r_mp_id_table[] = { 5449 { .name = MLX5_ADEV_NAME ".multiport", }, 5450 {}, 5451 }; 5452 5453 static const struct auxiliary_device_id mlx5r_id_table[] = { 5454 { .name = MLX5_ADEV_NAME ".rdma", }, 5455 {}, 5456 }; 5457 5458 MODULE_DEVICE_TABLE(auxiliary, mlx5r_mp_id_table); 5459 MODULE_DEVICE_TABLE(auxiliary, mlx5r_id_table); 5460 5461 static struct auxiliary_driver mlx5r_mp_driver = { 5462 .name = "multiport", 5463 .probe = mlx5r_mp_probe, 5464 .remove = mlx5r_mp_remove, 5465 .id_table = mlx5r_mp_id_table, 5466 }; 5467 5468 static struct auxiliary_driver mlx5r_driver = { 5469 .name = "rdma", 5470 .probe = mlx5r_probe, 5471 .remove = mlx5r_remove, 5472 .id_table = mlx5r_id_table, 5473 }; 5474 5475 static int __init mlx5_ib_init(void) 5476 { 5477 int ret; 5478 5479 xlt_emergency_page = (void *)__get_free_page(GFP_KERNEL); 5480 if (!xlt_emergency_page) 5481 return -ENOMEM; 5482 5483 mlx5_ib_event_wq = alloc_ordered_workqueue("mlx5_ib_event_wq", 0); 5484 if (!mlx5_ib_event_wq) { 5485 free_page((unsigned long)xlt_emergency_page); 5486 return -ENOMEM; 5487 } 5488 5489 ret = mlx5_ib_qp_event_init(); 5490 if (ret) 5491 goto qp_event_err; 5492 5493 mlx5_ib_odp_init(); 5494 ret = mlx5r_rep_init(); 5495 if (ret) 5496 goto rep_err; 5497 ret = mlx5_data_direct_driver_register(); 5498 if (ret) 5499 goto dd_err; 5500 ret = auxiliary_driver_register(&mlx5r_mp_driver); 5501 if (ret) 5502 goto mp_err; 5503 ret = auxiliary_driver_register(&mlx5r_driver); 5504 if (ret) 5505 goto drv_err; 5506 5507 return 0; 5508 5509 drv_err: 5510 auxiliary_driver_unregister(&mlx5r_mp_driver); 5511 mp_err: 5512 mlx5_data_direct_driver_unregister(); 5513 dd_err: 5514 mlx5r_rep_cleanup(); 5515 rep_err: 5516 mlx5_ib_qp_event_cleanup(); 5517 qp_event_err: 5518 destroy_workqueue(mlx5_ib_event_wq); 5519 free_page((unsigned long)xlt_emergency_page); 5520 return ret; 5521 } 5522 5523 static void __exit mlx5_ib_cleanup(void) 5524 { 5525 mlx5_data_direct_driver_unregister(); 5526 auxiliary_driver_unregister(&mlx5r_driver); 5527 auxiliary_driver_unregister(&mlx5r_mp_driver); 5528 mlx5r_rep_cleanup(); 5529 5530 mlx5_ib_qp_event_cleanup(); 5531 destroy_workqueue(mlx5_ib_event_wq); 5532 free_page((unsigned long)xlt_emergency_page); 5533 } 5534 5535 module_init(mlx5_ib_init); 5536 module_exit(mlx5_ib_cleanup); 5537