1 /*- 2 * Copyright (c) 2013-2021, Mellanox Technologies, Ltd. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 */ 25 26 #include "opt_rss.h" 27 #include "opt_ratelimit.h" 28 29 #include <linux/module.h> 30 #include <linux/errno.h> 31 #include <linux/pci.h> 32 #include <linux/dma-mapping.h> 33 #include <linux/slab.h> 34 #if defined(CONFIG_X86) 35 #include <asm/pat.h> 36 #endif 37 #include <linux/sched.h> 38 #include <linux/delay.h> 39 #include <linux/fs.h> 40 #undef inode 41 #include <rdma/ib_user_verbs.h> 42 #include <rdma/ib_addr.h> 43 #include <rdma/ib_cache.h> 44 #include <dev/mlx5/port.h> 45 #include <dev/mlx5/vport.h> 46 #include <linux/list.h> 47 #include <rdma/ib_smi.h> 48 #include <rdma/ib_umem.h> 49 #include <rdma/uverbs_ioctl.h> 50 #include <linux/in.h> 51 #include <linux/etherdevice.h> 52 #include <dev/mlx5/fs.h> 53 #include <dev/mlx5/mlx5_ib/mlx5_ib.h> 54 55 MODULE_DESCRIPTION("Mellanox Connect-IB HCA IB driver"); 56 MODULE_LICENSE("Dual BSD/GPL"); 57 MODULE_DEPEND(mlx5ib, linuxkpi, 1, 1, 1); 58 MODULE_DEPEND(mlx5ib, mlx5, 1, 1, 1); 59 MODULE_DEPEND(mlx5ib, ibcore, 1, 1, 1); 60 MODULE_VERSION(mlx5ib, 1); 61 62 enum { 63 MLX5_ATOMIC_SIZE_QP_8BYTES = 1 << 3, 64 }; 65 66 static enum rdma_link_layer 67 mlx5_port_type_cap_to_rdma_ll(int port_type_cap) 68 { 69 switch (port_type_cap) { 70 case MLX5_CAP_PORT_TYPE_IB: 71 return IB_LINK_LAYER_INFINIBAND; 72 case MLX5_CAP_PORT_TYPE_ETH: 73 return IB_LINK_LAYER_ETHERNET; 74 default: 75 return IB_LINK_LAYER_UNSPECIFIED; 76 } 77 } 78 79 static enum rdma_link_layer 80 mlx5_ib_port_link_layer(struct ib_device *device, u8 port_num) 81 { 82 struct mlx5_ib_dev *dev = to_mdev(device); 83 int port_type_cap = MLX5_CAP_GEN(dev->mdev, port_type); 84 85 return mlx5_port_type_cap_to_rdma_ll(port_type_cap); 86 } 87 88 static bool mlx5_netdev_match(if_t ndev, 89 struct mlx5_core_dev *mdev, 90 const char *dname) 91 { 92 return if_gettype(ndev) == IFT_ETHER && 93 if_getdname(ndev) != NULL && 94 strcmp(if_getdname(ndev), dname) == 0 && 95 if_getsoftc(ndev) != NULL && 96 *(struct mlx5_core_dev **)if_getsoftc(ndev) == mdev; 97 } 98 99 static int mlx5_netdev_event(struct notifier_block *this, 100 unsigned long event, void *ptr) 101 { 102 if_t ndev = netdev_notifier_info_to_ifp(ptr); 103 struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev, 104 roce.nb); 105 106 switch (event) { 107 case NETDEV_REGISTER: 108 case NETDEV_UNREGISTER: 109 write_lock(&ibdev->roce.netdev_lock); 110 /* check if network interface belongs to mlx5en */ 111 if (mlx5_netdev_match(ndev, ibdev->mdev, "mce")) 112 ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ? 113 NULL : ndev; 114 write_unlock(&ibdev->roce.netdev_lock); 115 break; 116 117 case NETDEV_UP: 118 case NETDEV_DOWN: { 119 if_t upper = NULL; 120 121 if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev)) 122 && ibdev->ib_active) { 123 struct ib_event ibev = {0}; 124 125 ibev.device = &ibdev->ib_dev; 126 ibev.event = (event == NETDEV_UP) ? 127 IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR; 128 ibev.element.port_num = 1; 129 ib_dispatch_event(&ibev); 130 } 131 break; 132 } 133 134 default: 135 break; 136 } 137 138 return NOTIFY_DONE; 139 } 140 141 static if_t mlx5_ib_get_netdev(struct ib_device *device, 142 u8 port_num) 143 { 144 struct mlx5_ib_dev *ibdev = to_mdev(device); 145 if_t ndev; 146 147 /* Ensure ndev does not disappear before we invoke if_ref() 148 */ 149 read_lock(&ibdev->roce.netdev_lock); 150 ndev = ibdev->roce.netdev; 151 if (ndev) 152 if_ref(ndev); 153 read_unlock(&ibdev->roce.netdev_lock); 154 155 return ndev; 156 } 157 158 static int translate_eth_proto_oper(u32 eth_proto_oper, u8 *active_speed, 159 u8 *active_width) 160 { 161 switch (eth_proto_oper) { 162 case MLX5E_PROT_MASK(MLX5E_1000BASE_CX_SGMII): 163 case MLX5E_PROT_MASK(MLX5E_1000BASE_KX): 164 case MLX5E_PROT_MASK(MLX5E_100BASE_TX): 165 case MLX5E_PROT_MASK(MLX5E_1000BASE_T): 166 *active_width = IB_WIDTH_1X; 167 *active_speed = IB_SPEED_SDR; 168 break; 169 case MLX5E_PROT_MASK(MLX5E_10GBASE_T): 170 case MLX5E_PROT_MASK(MLX5E_10GBASE_CX4): 171 case MLX5E_PROT_MASK(MLX5E_10GBASE_KX4): 172 case MLX5E_PROT_MASK(MLX5E_10GBASE_KR): 173 case MLX5E_PROT_MASK(MLX5E_10GBASE_CR): 174 case MLX5E_PROT_MASK(MLX5E_10GBASE_SR): 175 case MLX5E_PROT_MASK(MLX5E_10GBASE_ER_LR): 176 *active_width = IB_WIDTH_1X; 177 *active_speed = IB_SPEED_QDR; 178 break; 179 case MLX5E_PROT_MASK(MLX5E_25GBASE_CR): 180 case MLX5E_PROT_MASK(MLX5E_25GBASE_KR): 181 case MLX5E_PROT_MASK(MLX5E_25GBASE_SR): 182 *active_width = IB_WIDTH_1X; 183 *active_speed = IB_SPEED_EDR; 184 break; 185 case MLX5E_PROT_MASK(MLX5E_40GBASE_CR4): 186 case MLX5E_PROT_MASK(MLX5E_40GBASE_KR4): 187 case MLX5E_PROT_MASK(MLX5E_40GBASE_SR4): 188 case MLX5E_PROT_MASK(MLX5E_40GBASE_LR4_ER4): 189 *active_width = IB_WIDTH_4X; 190 *active_speed = IB_SPEED_QDR; 191 break; 192 case MLX5E_PROT_MASK(MLX5E_50GBASE_CR2): 193 case MLX5E_PROT_MASK(MLX5E_50GBASE_KR2): 194 case MLX5E_PROT_MASK(MLX5E_50GBASE_KR4): 195 case MLX5E_PROT_MASK(MLX5E_50GBASE_SR2): 196 *active_width = IB_WIDTH_1X; 197 *active_speed = IB_SPEED_HDR; 198 break; 199 case MLX5E_PROT_MASK(MLX5E_56GBASE_R4): 200 *active_width = IB_WIDTH_4X; 201 *active_speed = IB_SPEED_FDR; 202 break; 203 case MLX5E_PROT_MASK(MLX5E_100GBASE_CR4): 204 case MLX5E_PROT_MASK(MLX5E_100GBASE_SR4): 205 case MLX5E_PROT_MASK(MLX5E_100GBASE_KR4): 206 case MLX5E_PROT_MASK(MLX5E_100GBASE_LR4): 207 *active_width = IB_WIDTH_4X; 208 *active_speed = IB_SPEED_EDR; 209 break; 210 default: 211 *active_width = IB_WIDTH_4X; 212 *active_speed = IB_SPEED_QDR; 213 return -EINVAL; 214 } 215 216 return 0; 217 } 218 219 static int translate_eth_ext_proto_oper(u32 eth_proto_oper, u8 *active_speed, 220 u8 *active_width) 221 { 222 switch (eth_proto_oper) { 223 case MLX5E_PROT_MASK(MLX5E_SGMII_100M): 224 case MLX5E_PROT_MASK(MLX5E_1000BASE_X_SGMII): 225 *active_width = IB_WIDTH_1X; 226 *active_speed = IB_SPEED_SDR; 227 break; 228 case MLX5E_PROT_MASK(MLX5E_5GBASE_R): 229 *active_width = IB_WIDTH_1X; 230 *active_speed = IB_SPEED_DDR; 231 break; 232 case MLX5E_PROT_MASK(MLX5E_10GBASE_XFI_XAUI_1): 233 *active_width = IB_WIDTH_1X; 234 *active_speed = IB_SPEED_QDR; 235 break; 236 case MLX5E_PROT_MASK(MLX5E_40GBASE_XLAUI_4_XLPPI_4): 237 *active_width = IB_WIDTH_4X; 238 *active_speed = IB_SPEED_QDR; 239 break; 240 case MLX5E_PROT_MASK(MLX5E_25GAUI_1_25GBASE_CR_KR): 241 *active_width = IB_WIDTH_1X; 242 *active_speed = IB_SPEED_EDR; 243 break; 244 case MLX5E_PROT_MASK(MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2): 245 *active_width = IB_WIDTH_2X; 246 *active_speed = IB_SPEED_EDR; 247 break; 248 case MLX5E_PROT_MASK(MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR): 249 *active_width = IB_WIDTH_1X; 250 *active_speed = IB_SPEED_HDR; 251 break; 252 case MLX5E_PROT_MASK(MLX5E_CAUI_4_100GBASE_CR4_KR4): 253 *active_width = IB_WIDTH_4X; 254 *active_speed = IB_SPEED_EDR; 255 break; 256 case MLX5E_PROT_MASK(MLX5E_100GAUI_2_100GBASE_CR2_KR2): 257 *active_width = IB_WIDTH_2X; 258 *active_speed = IB_SPEED_HDR; 259 break; 260 case MLX5E_PROT_MASK(MLX5E_100GAUI_1_100GBASE_CR_KR): 261 *active_width = IB_WIDTH_1X; 262 *active_speed = IB_SPEED_NDR; 263 break; 264 case MLX5E_PROT_MASK(MLX5E_200GAUI_4_200GBASE_CR4_KR4): 265 *active_width = IB_WIDTH_4X; 266 *active_speed = IB_SPEED_HDR; 267 break; 268 case MLX5E_PROT_MASK(MLX5E_200GAUI_2_200GBASE_CR2_KR2): 269 *active_width = IB_WIDTH_2X; 270 *active_speed = IB_SPEED_NDR; 271 break; 272 case MLX5E_PROT_MASK(MLX5E_400GAUI_4_400GBASE_CR4_KR4): 273 *active_width = IB_WIDTH_4X; 274 *active_speed = IB_SPEED_NDR; 275 break; 276 default: 277 *active_width = IB_WIDTH_4X; 278 *active_speed = IB_SPEED_QDR; 279 return -EINVAL; 280 } 281 282 return 0; 283 } 284 285 static int mlx5_query_port_roce(struct ib_device *device, u8 port_num, 286 struct ib_port_attr *props) 287 { 288 struct mlx5_ib_dev *dev = to_mdev(device); 289 u32 out[MLX5_ST_SZ_DW(ptys_reg)] = {}; 290 if_t ndev; 291 enum ib_mtu ndev_ib_mtu; 292 u16 qkey_viol_cntr; 293 u32 eth_prot_oper; 294 bool ext; 295 int err; 296 297 memset(props, 0, sizeof(*props)); 298 299 /* Possible bad flows are checked before filling out props so in case 300 * of an error it will still be zeroed out. 301 */ 302 err = mlx5_query_port_ptys(dev->mdev, out, sizeof(out), MLX5_PTYS_EN, 303 port_num); 304 if (err) 305 return err; 306 307 ext = MLX5_CAP_PCAM_FEATURE(dev->mdev, ptys_extended_ethernet); 308 eth_prot_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext, eth_proto_oper); 309 310 if (ext) 311 translate_eth_ext_proto_oper(eth_prot_oper, &props->active_speed, 312 &props->active_width); 313 else 314 translate_eth_proto_oper(eth_prot_oper, &props->active_speed, 315 &props->active_width); 316 317 props->port_cap_flags |= IB_PORT_CM_SUP; 318 props->port_cap_flags |= IB_PORT_IP_BASED_GIDS; 319 320 props->gid_tbl_len = MLX5_CAP_ROCE(dev->mdev, 321 roce_address_table_size); 322 props->max_mtu = IB_MTU_4096; 323 props->max_msg_sz = 1 << MLX5_CAP_GEN(dev->mdev, log_max_msg); 324 props->pkey_tbl_len = 1; 325 props->state = IB_PORT_DOWN; 326 props->phys_state = IB_PORT_PHYS_STATE_DISABLED; 327 328 mlx5_query_nic_vport_qkey_viol_cntr(dev->mdev, &qkey_viol_cntr); 329 props->qkey_viol_cntr = qkey_viol_cntr; 330 331 ndev = mlx5_ib_get_netdev(device, port_num); 332 if (!ndev) 333 return 0; 334 335 if (if_getdrvflags(ndev) & IFF_DRV_RUNNING && 336 if_getlinkstate(ndev) == LINK_STATE_UP) { 337 props->state = IB_PORT_ACTIVE; 338 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP; 339 } 340 341 ndev_ib_mtu = iboe_get_mtu(if_getmtu(ndev)); 342 343 if_rele(ndev); 344 345 props->active_mtu = min(props->max_mtu, ndev_ib_mtu); 346 return 0; 347 } 348 349 static void ib_gid_to_mlx5_roce_addr(const union ib_gid *gid, 350 const struct ib_gid_attr *attr, 351 void *mlx5_addr) 352 { 353 #define MLX5_SET_RA(p, f, v) MLX5_SET(roce_addr_layout, p, f, v) 354 char *mlx5_addr_l3_addr = MLX5_ADDR_OF(roce_addr_layout, mlx5_addr, 355 source_l3_address); 356 void *mlx5_addr_mac = MLX5_ADDR_OF(roce_addr_layout, mlx5_addr, 357 source_mac_47_32); 358 u16 vlan_id; 359 360 if (!gid) 361 return; 362 ether_addr_copy(mlx5_addr_mac, if_getlladdr(attr->ndev)); 363 364 vlan_id = rdma_vlan_dev_vlan_id(attr->ndev); 365 if (vlan_id != 0xffff) { 366 MLX5_SET_RA(mlx5_addr, vlan_valid, 1); 367 MLX5_SET_RA(mlx5_addr, vlan_id, vlan_id); 368 } 369 370 switch (attr->gid_type) { 371 case IB_GID_TYPE_IB: 372 MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_1); 373 break; 374 case IB_GID_TYPE_ROCE_UDP_ENCAP: 375 MLX5_SET_RA(mlx5_addr, roce_version, MLX5_ROCE_VERSION_2); 376 break; 377 378 default: 379 WARN_ON(true); 380 } 381 382 if (attr->gid_type != IB_GID_TYPE_IB) { 383 if (ipv6_addr_v4mapped((void *)gid)) 384 MLX5_SET_RA(mlx5_addr, roce_l3_type, 385 MLX5_ROCE_L3_TYPE_IPV4); 386 else 387 MLX5_SET_RA(mlx5_addr, roce_l3_type, 388 MLX5_ROCE_L3_TYPE_IPV6); 389 } 390 391 if ((attr->gid_type == IB_GID_TYPE_IB) || 392 !ipv6_addr_v4mapped((void *)gid)) 393 memcpy(mlx5_addr_l3_addr, gid, sizeof(*gid)); 394 else 395 memcpy(&mlx5_addr_l3_addr[12], &gid->raw[12], 4); 396 } 397 398 static int set_roce_addr(struct ib_device *device, u8 port_num, 399 unsigned int index, 400 const union ib_gid *gid, 401 const struct ib_gid_attr *attr) 402 { 403 struct mlx5_ib_dev *dev = to_mdev(device); 404 u32 in[MLX5_ST_SZ_DW(set_roce_address_in)] = {0}; 405 u32 out[MLX5_ST_SZ_DW(set_roce_address_out)] = {0}; 406 void *in_addr = MLX5_ADDR_OF(set_roce_address_in, in, roce_address); 407 enum rdma_link_layer ll = mlx5_ib_port_link_layer(device, port_num); 408 409 if (ll != IB_LINK_LAYER_ETHERNET) 410 return -EINVAL; 411 412 ib_gid_to_mlx5_roce_addr(gid, attr, in_addr); 413 414 MLX5_SET(set_roce_address_in, in, roce_address_index, index); 415 MLX5_SET(set_roce_address_in, in, opcode, MLX5_CMD_OP_SET_ROCE_ADDRESS); 416 return mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out)); 417 } 418 419 static int mlx5_ib_add_gid(struct ib_device *device, u8 port_num, 420 unsigned int index, const union ib_gid *gid, 421 const struct ib_gid_attr *attr, 422 __always_unused void **context) 423 { 424 return set_roce_addr(device, port_num, index, gid, attr); 425 } 426 427 static int mlx5_ib_del_gid(struct ib_device *device, u8 port_num, 428 unsigned int index, __always_unused void **context) 429 { 430 return set_roce_addr(device, port_num, index, NULL, NULL); 431 } 432 433 __be16 mlx5_get_roce_udp_sport(struct mlx5_ib_dev *dev, u8 port_num, 434 int index) 435 { 436 struct ib_gid_attr attr; 437 union ib_gid gid; 438 439 if (ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr)) 440 return 0; 441 442 if (!attr.ndev) 443 return 0; 444 445 if_rele(attr.ndev); 446 447 if (attr.gid_type != IB_GID_TYPE_ROCE_UDP_ENCAP) 448 return 0; 449 450 return cpu_to_be16(MLX5_CAP_ROCE(dev->mdev, r_roce_min_src_udp_port)); 451 } 452 453 int mlx5_get_roce_gid_type(struct mlx5_ib_dev *dev, u8 port_num, 454 int index, enum ib_gid_type *gid_type) 455 { 456 struct ib_gid_attr attr; 457 union ib_gid gid; 458 int ret; 459 460 ret = ib_get_cached_gid(&dev->ib_dev, port_num, index, &gid, &attr); 461 if (ret) 462 return ret; 463 464 if (!attr.ndev) 465 return -ENODEV; 466 467 if_rele(attr.ndev); 468 469 *gid_type = attr.gid_type; 470 471 return 0; 472 } 473 474 static int mlx5_use_mad_ifc(struct mlx5_ib_dev *dev) 475 { 476 if (MLX5_CAP_GEN(dev->mdev, port_type) == MLX5_CAP_PORT_TYPE_IB) 477 return !MLX5_CAP_GEN(dev->mdev, ib_virt); 478 return 0; 479 } 480 481 enum { 482 MLX5_VPORT_ACCESS_METHOD_MAD, 483 MLX5_VPORT_ACCESS_METHOD_HCA, 484 MLX5_VPORT_ACCESS_METHOD_NIC, 485 }; 486 487 static int mlx5_get_vport_access_method(struct ib_device *ibdev) 488 { 489 if (mlx5_use_mad_ifc(to_mdev(ibdev))) 490 return MLX5_VPORT_ACCESS_METHOD_MAD; 491 492 if (mlx5_ib_port_link_layer(ibdev, 1) == 493 IB_LINK_LAYER_ETHERNET) 494 return MLX5_VPORT_ACCESS_METHOD_NIC; 495 496 return MLX5_VPORT_ACCESS_METHOD_HCA; 497 } 498 499 static void get_atomic_caps(struct mlx5_ib_dev *dev, 500 struct ib_device_attr *props) 501 { 502 u8 tmp; 503 u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations); 504 u8 atomic_size_qp = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp); 505 u8 atomic_req_8B_endianness_mode = 506 MLX5_CAP_ATOMIC(dev->mdev, atomic_req_8B_endianess_mode); 507 508 /* Check if HW supports 8 bytes standard atomic operations and capable 509 * of host endianness respond 510 */ 511 tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD; 512 if (((atomic_operations & tmp) == tmp) && 513 (atomic_size_qp & MLX5_ATOMIC_SIZE_QP_8BYTES) && 514 (atomic_req_8B_endianness_mode)) { 515 props->atomic_cap = IB_ATOMIC_HCA; 516 } else { 517 props->atomic_cap = IB_ATOMIC_NONE; 518 } 519 } 520 521 static int mlx5_query_system_image_guid(struct ib_device *ibdev, 522 __be64 *sys_image_guid) 523 { 524 struct mlx5_ib_dev *dev = to_mdev(ibdev); 525 struct mlx5_core_dev *mdev = dev->mdev; 526 u64 tmp; 527 int err; 528 529 switch (mlx5_get_vport_access_method(ibdev)) { 530 case MLX5_VPORT_ACCESS_METHOD_MAD: 531 return mlx5_query_mad_ifc_system_image_guid(ibdev, 532 sys_image_guid); 533 534 case MLX5_VPORT_ACCESS_METHOD_HCA: 535 err = mlx5_query_hca_vport_system_image_guid(mdev, &tmp); 536 break; 537 538 case MLX5_VPORT_ACCESS_METHOD_NIC: 539 err = mlx5_query_nic_vport_system_image_guid(mdev, &tmp); 540 break; 541 542 default: 543 return -EINVAL; 544 } 545 546 if (!err) 547 *sys_image_guid = cpu_to_be64(tmp); 548 549 return err; 550 551 } 552 553 static int mlx5_query_max_pkeys(struct ib_device *ibdev, 554 u16 *max_pkeys) 555 { 556 struct mlx5_ib_dev *dev = to_mdev(ibdev); 557 struct mlx5_core_dev *mdev = dev->mdev; 558 559 switch (mlx5_get_vport_access_method(ibdev)) { 560 case MLX5_VPORT_ACCESS_METHOD_MAD: 561 return mlx5_query_mad_ifc_max_pkeys(ibdev, max_pkeys); 562 563 case MLX5_VPORT_ACCESS_METHOD_HCA: 564 case MLX5_VPORT_ACCESS_METHOD_NIC: 565 *max_pkeys = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, 566 pkey_table_size)); 567 return 0; 568 569 default: 570 return -EINVAL; 571 } 572 } 573 574 static int mlx5_query_vendor_id(struct ib_device *ibdev, 575 u32 *vendor_id) 576 { 577 struct mlx5_ib_dev *dev = to_mdev(ibdev); 578 579 switch (mlx5_get_vport_access_method(ibdev)) { 580 case MLX5_VPORT_ACCESS_METHOD_MAD: 581 return mlx5_query_mad_ifc_vendor_id(ibdev, vendor_id); 582 583 case MLX5_VPORT_ACCESS_METHOD_HCA: 584 case MLX5_VPORT_ACCESS_METHOD_NIC: 585 return mlx5_core_query_vendor_id(dev->mdev, vendor_id); 586 587 default: 588 return -EINVAL; 589 } 590 } 591 592 static int mlx5_query_node_guid(struct mlx5_ib_dev *dev, 593 __be64 *node_guid) 594 { 595 u64 tmp; 596 int err; 597 598 switch (mlx5_get_vport_access_method(&dev->ib_dev)) { 599 case MLX5_VPORT_ACCESS_METHOD_MAD: 600 return mlx5_query_mad_ifc_node_guid(dev, node_guid); 601 602 case MLX5_VPORT_ACCESS_METHOD_HCA: 603 err = mlx5_query_hca_vport_node_guid(dev->mdev, &tmp); 604 break; 605 606 case MLX5_VPORT_ACCESS_METHOD_NIC: 607 err = mlx5_query_nic_vport_node_guid(dev->mdev, &tmp); 608 break; 609 610 default: 611 return -EINVAL; 612 } 613 614 if (!err) 615 *node_guid = cpu_to_be64(tmp); 616 617 return err; 618 } 619 620 struct mlx5_reg_node_desc { 621 u8 desc[IB_DEVICE_NODE_DESC_MAX]; 622 }; 623 624 static int mlx5_query_node_desc(struct mlx5_ib_dev *dev, char *node_desc) 625 { 626 struct mlx5_reg_node_desc in; 627 628 if (mlx5_use_mad_ifc(dev)) 629 return mlx5_query_mad_ifc_node_desc(dev, node_desc); 630 631 memset(&in, 0, sizeof(in)); 632 633 return mlx5_core_access_reg(dev->mdev, &in, sizeof(in), node_desc, 634 sizeof(struct mlx5_reg_node_desc), 635 MLX5_REG_NODE_DESC, 0, 0); 636 } 637 638 static int mlx5_ib_query_device(struct ib_device *ibdev, 639 struct ib_device_attr *props, 640 struct ib_udata *uhw) 641 { 642 struct mlx5_ib_dev *dev = to_mdev(ibdev); 643 struct mlx5_core_dev *mdev = dev->mdev; 644 int err = -ENOMEM; 645 int max_sq_desc; 646 int max_rq_sg; 647 int max_sq_sg; 648 u64 min_page_size = 1ull << MLX5_CAP_GEN(mdev, log_pg_sz); 649 struct mlx5_ib_query_device_resp resp = {}; 650 size_t resp_len; 651 u64 max_tso; 652 653 resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length); 654 if (uhw->outlen && uhw->outlen < resp_len) 655 return -EINVAL; 656 else 657 resp.response_length = resp_len; 658 659 if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen)) 660 return -EINVAL; 661 662 memset(props, 0, sizeof(*props)); 663 err = mlx5_query_system_image_guid(ibdev, 664 &props->sys_image_guid); 665 if (err) 666 return err; 667 668 err = mlx5_query_max_pkeys(ibdev, &props->max_pkeys); 669 if (err) 670 return err; 671 672 err = mlx5_query_vendor_id(ibdev, &props->vendor_id); 673 if (err) 674 return err; 675 676 props->fw_ver = ((u64)fw_rev_maj(dev->mdev) << 32) | 677 ((u32)fw_rev_min(dev->mdev) << 16) | 678 fw_rev_sub(dev->mdev); 679 props->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT | 680 IB_DEVICE_PORT_ACTIVE_EVENT | 681 IB_DEVICE_SYS_IMAGE_GUID | 682 IB_DEVICE_RC_RNR_NAK_GEN | 683 IB_DEVICE_KNOWSEPOCH; 684 685 if (MLX5_CAP_GEN(mdev, pkv)) 686 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR; 687 if (MLX5_CAP_GEN(mdev, qkv)) 688 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR; 689 if (MLX5_CAP_GEN(mdev, apm)) 690 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG; 691 if (MLX5_CAP_GEN(mdev, xrc)) 692 props->device_cap_flags |= IB_DEVICE_XRC; 693 if (MLX5_CAP_GEN(mdev, imaicl)) { 694 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW | 695 IB_DEVICE_MEM_WINDOW_TYPE_2B; 696 props->max_mw = 1 << MLX5_CAP_GEN(mdev, log_max_mkey); 697 /* We support 'Gappy' memory registration too */ 698 props->device_cap_flags |= IB_DEVICE_SG_GAPS_REG; 699 } 700 props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS; 701 if (MLX5_CAP_GEN(mdev, sho)) { 702 props->device_cap_flags |= IB_DEVICE_SIGNATURE_HANDOVER; 703 /* At this stage no support for signature handover */ 704 props->sig_prot_cap = IB_PROT_T10DIF_TYPE_1 | 705 IB_PROT_T10DIF_TYPE_2 | 706 IB_PROT_T10DIF_TYPE_3; 707 props->sig_guard_cap = IB_GUARD_T10DIF_CRC | 708 IB_GUARD_T10DIF_CSUM; 709 } 710 if (MLX5_CAP_GEN(mdev, block_lb_mc)) 711 props->device_cap_flags |= IB_DEVICE_BLOCK_MULTICAST_LOOPBACK; 712 713 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads)) { 714 if (MLX5_CAP_ETH(mdev, csum_cap)) 715 props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM; 716 717 if (field_avail(typeof(resp), tso_caps, uhw->outlen)) { 718 max_tso = MLX5_CAP_ETH(mdev, max_lso_cap); 719 if (max_tso) { 720 resp.tso_caps.max_tso = 1 << max_tso; 721 resp.tso_caps.supported_qpts |= 722 1 << IB_QPT_RAW_PACKET; 723 resp.response_length += sizeof(resp.tso_caps); 724 } 725 } 726 727 if (field_avail(typeof(resp), rss_caps, uhw->outlen)) { 728 resp.rss_caps.rx_hash_function = 729 MLX5_RX_HASH_FUNC_TOEPLITZ; 730 resp.rss_caps.rx_hash_fields_mask = 731 MLX5_RX_HASH_SRC_IPV4 | 732 MLX5_RX_HASH_DST_IPV4 | 733 MLX5_RX_HASH_SRC_IPV6 | 734 MLX5_RX_HASH_DST_IPV6 | 735 MLX5_RX_HASH_SRC_PORT_TCP | 736 MLX5_RX_HASH_DST_PORT_TCP | 737 MLX5_RX_HASH_SRC_PORT_UDP | 738 MLX5_RX_HASH_DST_PORT_UDP; 739 resp.response_length += sizeof(resp.rss_caps); 740 } 741 } else { 742 if (field_avail(typeof(resp), tso_caps, uhw->outlen)) 743 resp.response_length += sizeof(resp.tso_caps); 744 if (field_avail(typeof(resp), rss_caps, uhw->outlen)) 745 resp.response_length += sizeof(resp.rss_caps); 746 } 747 748 if (MLX5_CAP_GEN(mdev, ipoib_ipoib_offloads)) { 749 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM; 750 props->device_cap_flags |= IB_DEVICE_UD_TSO; 751 } 752 753 if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) && 754 MLX5_CAP_ETH(dev->mdev, scatter_fcs)) 755 props->device_cap_flags |= IB_DEVICE_RAW_SCATTER_FCS; 756 757 if (mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS)) 758 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING; 759 760 props->vendor_part_id = mdev->pdev->device; 761 props->hw_ver = mdev->pdev->revision; 762 763 props->max_mr_size = ~0ull; 764 props->page_size_cap = ~(min_page_size - 1); 765 props->max_qp = 1 << MLX5_CAP_GEN(mdev, log_max_qp); 766 props->max_qp_wr = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz); 767 max_rq_sg = MLX5_CAP_GEN(mdev, max_wqe_sz_rq) / 768 sizeof(struct mlx5_wqe_data_seg); 769 max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512); 770 max_sq_sg = (max_sq_desc - sizeof(struct mlx5_wqe_ctrl_seg) - 771 sizeof(struct mlx5_wqe_raddr_seg)) / 772 sizeof(struct mlx5_wqe_data_seg); 773 props->max_sge = min(max_rq_sg, max_sq_sg); 774 props->max_sge_rd = MLX5_MAX_SGE_RD; 775 props->max_cq = 1 << MLX5_CAP_GEN(mdev, log_max_cq); 776 props->max_cqe = (1 << MLX5_CAP_GEN(mdev, log_max_cq_sz)) - 1; 777 props->max_mr = 1 << MLX5_CAP_GEN(mdev, log_max_mkey); 778 props->max_pd = 1 << MLX5_CAP_GEN(mdev, log_max_pd); 779 props->max_qp_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_req_qp); 780 props->max_qp_init_rd_atom = 1 << MLX5_CAP_GEN(mdev, log_max_ra_res_qp); 781 props->max_srq = 1 << MLX5_CAP_GEN(mdev, log_max_srq); 782 props->max_srq_wr = (1 << MLX5_CAP_GEN(mdev, log_max_srq_sz)) - 1; 783 props->local_ca_ack_delay = MLX5_CAP_GEN(mdev, local_ca_ack_delay); 784 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp; 785 props->max_srq_sge = max_rq_sg - 1; 786 props->max_fast_reg_page_list_len = 787 1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size); 788 get_atomic_caps(dev, props); 789 props->masked_atomic_cap = IB_ATOMIC_NONE; 790 props->max_mcast_grp = 1 << MLX5_CAP_GEN(mdev, log_max_mcg); 791 props->max_mcast_qp_attach = MLX5_CAP_GEN(mdev, max_qp_mcg); 792 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach * 793 props->max_mcast_grp; 794 props->max_map_per_fmr = INT_MAX; /* no limit in ConnectIB */ 795 props->hca_core_clock = MLX5_CAP_GEN(mdev, device_frequency_khz); 796 props->timestamp_mask = 0x7FFFFFFFFFFFFFFFULL; 797 798 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING 799 if (MLX5_CAP_GEN(mdev, pg)) 800 props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING; 801 props->odp_caps = dev->odp_caps; 802 #endif 803 804 if (MLX5_CAP_GEN(mdev, cd)) 805 props->device_cap_flags |= IB_DEVICE_CROSS_CHANNEL; 806 807 if (!mlx5_core_is_pf(mdev)) 808 props->device_cap_flags |= IB_DEVICE_VIRTUAL_FUNCTION; 809 810 if (mlx5_ib_port_link_layer(ibdev, 1) == 811 IB_LINK_LAYER_ETHERNET) { 812 props->rss_caps.max_rwq_indirection_tables = 813 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt); 814 props->rss_caps.max_rwq_indirection_table_size = 815 1 << MLX5_CAP_GEN(dev->mdev, log_max_rqt_size); 816 props->rss_caps.supported_qpts = 1 << IB_QPT_RAW_PACKET; 817 props->max_wq_type_rq = 818 1 << MLX5_CAP_GEN(dev->mdev, log_max_rq); 819 } 820 821 if (uhw->outlen) { 822 err = ib_copy_to_udata(uhw, &resp, resp.response_length); 823 824 if (err) 825 return err; 826 } 827 828 return 0; 829 } 830 831 enum mlx5_ib_width { 832 MLX5_IB_WIDTH_1X = 1 << 0, 833 MLX5_IB_WIDTH_2X = 1 << 1, 834 MLX5_IB_WIDTH_4X = 1 << 2, 835 MLX5_IB_WIDTH_8X = 1 << 3, 836 MLX5_IB_WIDTH_12X = 1 << 4 837 }; 838 839 static int translate_active_width(struct ib_device *ibdev, u8 active_width, 840 u8 *ib_width) 841 { 842 struct mlx5_ib_dev *dev = to_mdev(ibdev); 843 int err = 0; 844 845 if (active_width & MLX5_IB_WIDTH_1X) { 846 *ib_width = IB_WIDTH_1X; 847 } else if (active_width & MLX5_IB_WIDTH_2X) { 848 *ib_width = IB_WIDTH_2X; 849 } else if (active_width & MLX5_IB_WIDTH_4X) { 850 *ib_width = IB_WIDTH_4X; 851 } else if (active_width & MLX5_IB_WIDTH_8X) { 852 *ib_width = IB_WIDTH_8X; 853 } else if (active_width & MLX5_IB_WIDTH_12X) { 854 *ib_width = IB_WIDTH_12X; 855 } else { 856 mlx5_ib_dbg(dev, "Invalid active_width %d\n", 857 (int)active_width); 858 err = -EINVAL; 859 } 860 861 return err; 862 } 863 864 enum ib_max_vl_num { 865 __IB_MAX_VL_0 = 1, 866 __IB_MAX_VL_0_1 = 2, 867 __IB_MAX_VL_0_3 = 3, 868 __IB_MAX_VL_0_7 = 4, 869 __IB_MAX_VL_0_14 = 5, 870 }; 871 872 enum mlx5_vl_hw_cap { 873 MLX5_VL_HW_0 = 1, 874 MLX5_VL_HW_0_1 = 2, 875 MLX5_VL_HW_0_2 = 3, 876 MLX5_VL_HW_0_3 = 4, 877 MLX5_VL_HW_0_4 = 5, 878 MLX5_VL_HW_0_5 = 6, 879 MLX5_VL_HW_0_6 = 7, 880 MLX5_VL_HW_0_7 = 8, 881 MLX5_VL_HW_0_14 = 15 882 }; 883 884 static int translate_max_vl_num(struct ib_device *ibdev, u8 vl_hw_cap, 885 u8 *max_vl_num) 886 { 887 switch (vl_hw_cap) { 888 case MLX5_VL_HW_0: 889 *max_vl_num = __IB_MAX_VL_0; 890 break; 891 case MLX5_VL_HW_0_1: 892 *max_vl_num = __IB_MAX_VL_0_1; 893 break; 894 case MLX5_VL_HW_0_3: 895 *max_vl_num = __IB_MAX_VL_0_3; 896 break; 897 case MLX5_VL_HW_0_7: 898 *max_vl_num = __IB_MAX_VL_0_7; 899 break; 900 case MLX5_VL_HW_0_14: 901 *max_vl_num = __IB_MAX_VL_0_14; 902 break; 903 904 default: 905 return -EINVAL; 906 } 907 908 return 0; 909 } 910 911 static int mlx5_query_hca_port(struct ib_device *ibdev, u8 port, 912 struct ib_port_attr *props) 913 { 914 struct mlx5_ib_dev *dev = to_mdev(ibdev); 915 struct mlx5_core_dev *mdev = dev->mdev; 916 u32 *rep; 917 int replen = MLX5_ST_SZ_BYTES(query_hca_vport_context_out); 918 struct mlx5_ptys_reg *ptys; 919 struct mlx5_pmtu_reg *pmtu; 920 struct mlx5_pvlc_reg pvlc; 921 void *ctx; 922 int err; 923 924 rep = mlx5_vzalloc(replen); 925 ptys = kzalloc(sizeof(*ptys), GFP_KERNEL); 926 pmtu = kzalloc(sizeof(*pmtu), GFP_KERNEL); 927 if (!rep || !ptys || !pmtu) { 928 err = -ENOMEM; 929 goto out; 930 } 931 932 memset(props, 0, sizeof(*props)); 933 934 err = mlx5_query_hca_vport_context(mdev, port, 0, rep, replen); 935 if (err) 936 goto out; 937 938 ctx = MLX5_ADDR_OF(query_hca_vport_context_out, rep, hca_vport_context); 939 940 props->lid = MLX5_GET(hca_vport_context, ctx, lid); 941 props->lmc = MLX5_GET(hca_vport_context, ctx, lmc); 942 props->sm_lid = MLX5_GET(hca_vport_context, ctx, sm_lid); 943 props->sm_sl = MLX5_GET(hca_vport_context, ctx, sm_sl); 944 props->state = MLX5_GET(hca_vport_context, ctx, vport_state); 945 props->phys_state = MLX5_GET(hca_vport_context, ctx, 946 port_physical_state); 947 props->port_cap_flags = MLX5_GET(hca_vport_context, ctx, cap_mask1); 948 props->gid_tbl_len = mlx5_get_gid_table_len(MLX5_CAP_GEN(mdev, gid_table_size)); 949 props->max_msg_sz = 1 << MLX5_CAP_GEN(mdev, log_max_msg); 950 props->pkey_tbl_len = mlx5_to_sw_pkey_sz(MLX5_CAP_GEN(mdev, pkey_table_size)); 951 props->bad_pkey_cntr = MLX5_GET(hca_vport_context, ctx, 952 pkey_violation_counter); 953 props->qkey_viol_cntr = MLX5_GET(hca_vport_context, ctx, 954 qkey_violation_counter); 955 props->subnet_timeout = MLX5_GET(hca_vport_context, ctx, 956 subnet_timeout); 957 props->init_type_reply = MLX5_GET(hca_vport_context, ctx, 958 init_type_reply); 959 props->grh_required = MLX5_GET(hca_vport_context, ctx, grh_required); 960 961 ptys->proto_mask |= MLX5_PTYS_IB; 962 ptys->local_port = port; 963 err = mlx5_core_access_ptys(mdev, ptys, 0); 964 if (err) 965 goto out; 966 967 err = translate_active_width(ibdev, ptys->ib_link_width_oper, 968 &props->active_width); 969 if (err) 970 goto out; 971 972 props->active_speed = (u8)ptys->ib_proto_oper; 973 974 pmtu->local_port = port; 975 err = mlx5_core_access_pmtu(mdev, pmtu, 0); 976 if (err) 977 goto out; 978 979 props->max_mtu = pmtu->max_mtu; 980 props->active_mtu = pmtu->oper_mtu; 981 982 memset(&pvlc, 0, sizeof(pvlc)); 983 pvlc.local_port = port; 984 err = mlx5_core_access_pvlc(mdev, &pvlc, 0); 985 if (err) 986 goto out; 987 988 err = translate_max_vl_num(ibdev, pvlc.vl_hw_cap, 989 &props->max_vl_num); 990 out: 991 kvfree(rep); 992 kfree(ptys); 993 kfree(pmtu); 994 return err; 995 } 996 997 int mlx5_ib_query_port(struct ib_device *ibdev, u8 port, 998 struct ib_port_attr *props) 999 { 1000 switch (mlx5_get_vport_access_method(ibdev)) { 1001 case MLX5_VPORT_ACCESS_METHOD_MAD: 1002 return mlx5_query_mad_ifc_port(ibdev, port, props); 1003 1004 case MLX5_VPORT_ACCESS_METHOD_HCA: 1005 return mlx5_query_hca_port(ibdev, port, props); 1006 1007 case MLX5_VPORT_ACCESS_METHOD_NIC: 1008 return mlx5_query_port_roce(ibdev, port, props); 1009 1010 default: 1011 return -EINVAL; 1012 } 1013 } 1014 1015 static int mlx5_ib_query_gid(struct ib_device *ibdev, u8 port, int index, 1016 union ib_gid *gid) 1017 { 1018 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1019 struct mlx5_core_dev *mdev = dev->mdev; 1020 1021 switch (mlx5_get_vport_access_method(ibdev)) { 1022 case MLX5_VPORT_ACCESS_METHOD_MAD: 1023 return mlx5_query_mad_ifc_gids(ibdev, port, index, gid); 1024 1025 case MLX5_VPORT_ACCESS_METHOD_HCA: 1026 return mlx5_query_hca_vport_gid(mdev, port, 0, index, gid); 1027 1028 default: 1029 return -EINVAL; 1030 } 1031 1032 } 1033 1034 static int mlx5_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index, 1035 u16 *pkey) 1036 { 1037 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1038 struct mlx5_core_dev *mdev = dev->mdev; 1039 1040 switch (mlx5_get_vport_access_method(ibdev)) { 1041 case MLX5_VPORT_ACCESS_METHOD_MAD: 1042 return mlx5_query_mad_ifc_pkey(ibdev, port, index, pkey); 1043 1044 case MLX5_VPORT_ACCESS_METHOD_HCA: 1045 case MLX5_VPORT_ACCESS_METHOD_NIC: 1046 return mlx5_query_hca_vport_pkey(mdev, 0, port, 0, index, 1047 pkey); 1048 default: 1049 return -EINVAL; 1050 } 1051 } 1052 1053 static int mlx5_ib_modify_device(struct ib_device *ibdev, int mask, 1054 struct ib_device_modify *props) 1055 { 1056 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1057 struct mlx5_reg_node_desc in; 1058 struct mlx5_reg_node_desc out; 1059 int err; 1060 1061 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC) 1062 return -EOPNOTSUPP; 1063 1064 if (!(mask & IB_DEVICE_MODIFY_NODE_DESC)) 1065 return 0; 1066 1067 /* 1068 * If possible, pass node desc to FW, so it can generate 1069 * a 144 trap. If cmd fails, just ignore. 1070 */ 1071 memcpy(&in, props->node_desc, IB_DEVICE_NODE_DESC_MAX); 1072 err = mlx5_core_access_reg(dev->mdev, &in, sizeof(in), &out, 1073 sizeof(out), MLX5_REG_NODE_DESC, 0, 1); 1074 if (err) 1075 return err; 1076 1077 memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX); 1078 1079 return err; 1080 } 1081 1082 static int mlx5_ib_modify_port(struct ib_device *ibdev, u8 port, int mask, 1083 struct ib_port_modify *props) 1084 { 1085 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1086 struct ib_port_attr attr; 1087 u32 tmp; 1088 int err; 1089 1090 /* 1091 * CM layer calls ib_modify_port() regardless of the link 1092 * layer. For Ethernet ports, qkey violation and Port 1093 * capabilities are meaningless. 1094 */ 1095 if (mlx5_ib_port_link_layer(ibdev, port) == IB_LINK_LAYER_ETHERNET) 1096 return 0; 1097 1098 mutex_lock(&dev->cap_mask_mutex); 1099 1100 err = mlx5_ib_query_port(ibdev, port, &attr); 1101 if (err) 1102 goto out; 1103 1104 tmp = (attr.port_cap_flags | props->set_port_cap_mask) & 1105 ~props->clr_port_cap_mask; 1106 1107 err = mlx5_set_port_caps(dev->mdev, port, tmp); 1108 1109 out: 1110 mutex_unlock(&dev->cap_mask_mutex); 1111 return err; 1112 } 1113 1114 static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps) 1115 { 1116 mlx5_ib_dbg(dev, "MLX5_LIB_CAP_4K_UAR = %s\n", 1117 caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n"); 1118 } 1119 1120 static u16 calc_dynamic_bfregs(int uars_per_sys_page) 1121 { 1122 /* Large page with non 4k uar support might limit the dynamic size */ 1123 if (uars_per_sys_page == 1 && PAGE_SIZE > 4096) 1124 return MLX5_MIN_DYN_BFREGS; 1125 1126 return MLX5_MAX_DYN_BFREGS; 1127 } 1128 1129 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k, 1130 struct mlx5_ib_alloc_ucontext_req_v2 *req, 1131 struct mlx5_bfreg_info *bfregi) 1132 { 1133 int uars_per_sys_page; 1134 int bfregs_per_sys_page; 1135 int ref_bfregs = req->total_num_bfregs; 1136 1137 if (req->total_num_bfregs == 0) 1138 return -EINVAL; 1139 1140 BUILD_BUG_ON(MLX5_MAX_BFREGS % MLX5_NON_FP_BFREGS_IN_PAGE); 1141 BUILD_BUG_ON(MLX5_MAX_BFREGS < MLX5_NON_FP_BFREGS_IN_PAGE); 1142 1143 if (req->total_num_bfregs > MLX5_MAX_BFREGS) 1144 return -ENOMEM; 1145 1146 uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k); 1147 bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR; 1148 /* This holds the required static allocation asked by the user */ 1149 req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page); 1150 if (req->num_low_latency_bfregs > req->total_num_bfregs - 1) 1151 return -EINVAL; 1152 1153 bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page; 1154 bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page); 1155 bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs; 1156 bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page; 1157 1158 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", 1159 MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no", 1160 lib_uar_4k ? "yes" : "no", ref_bfregs, 1161 req->total_num_bfregs, bfregi->total_num_bfregs, 1162 bfregi->num_sys_pages); 1163 1164 return 0; 1165 } 1166 1167 static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *context) 1168 { 1169 struct mlx5_bfreg_info *bfregi; 1170 int err; 1171 int i; 1172 1173 bfregi = &context->bfregi; 1174 for (i = 0; i < bfregi->num_static_sys_pages; i++) { 1175 err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]); 1176 if (err) 1177 goto error; 1178 1179 mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]); 1180 } 1181 1182 for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++) 1183 bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX; 1184 1185 return 0; 1186 1187 error: 1188 for (--i; i >= 0; i--) 1189 if (mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i])) 1190 mlx5_ib_warn(dev, "failed to free uar %d\n", i); 1191 1192 return err; 1193 } 1194 1195 static void deallocate_uars(struct mlx5_ib_dev *dev, 1196 struct mlx5_ib_ucontext *context) 1197 { 1198 struct mlx5_bfreg_info *bfregi; 1199 int i; 1200 1201 bfregi = &context->bfregi; 1202 for (i = 0; i < bfregi->num_sys_pages; i++) 1203 if (i < bfregi->num_static_sys_pages || 1204 bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX) 1205 mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]); 1206 } 1207 1208 static int mlx5_ib_alloc_transport_domain(struct mlx5_ib_dev *dev, u32 *tdn, 1209 u16 uid) 1210 { 1211 int err; 1212 1213 if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) 1214 return 0; 1215 1216 err = mlx5_alloc_transport_domain(dev->mdev, tdn, uid); 1217 if (err) 1218 return err; 1219 1220 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) || 1221 (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) && 1222 !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc))) 1223 return 0; 1224 1225 mutex_lock(&dev->lb_mutex); 1226 dev->user_td++; 1227 1228 if (dev->user_td == 2) 1229 err = mlx5_nic_vport_update_local_lb(dev->mdev, true); 1230 1231 mutex_unlock(&dev->lb_mutex); 1232 1233 if (err != 0) 1234 mlx5_dealloc_transport_domain(dev->mdev, *tdn, uid); 1235 return err; 1236 } 1237 1238 static void mlx5_ib_dealloc_transport_domain(struct mlx5_ib_dev *dev, u32 tdn, 1239 u16 uid) 1240 { 1241 if (!MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) 1242 return; 1243 1244 mlx5_dealloc_transport_domain(dev->mdev, tdn, uid); 1245 1246 if ((MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH) || 1247 (!MLX5_CAP_GEN(dev->mdev, disable_local_lb_uc) && 1248 !MLX5_CAP_GEN(dev->mdev, disable_local_lb_mc))) 1249 return; 1250 1251 mutex_lock(&dev->lb_mutex); 1252 dev->user_td--; 1253 1254 if (dev->user_td < 2) 1255 mlx5_nic_vport_update_local_lb(dev->mdev, false); 1256 1257 mutex_unlock(&dev->lb_mutex); 1258 } 1259 1260 static int mlx5_ib_alloc_ucontext(struct ib_ucontext *uctx, 1261 struct ib_udata *udata) 1262 { 1263 struct ib_device *ibdev = uctx->device; 1264 struct mlx5_ib_dev *dev = to_mdev(ibdev); 1265 struct mlx5_ib_alloc_ucontext_req_v2 req = {}; 1266 struct mlx5_ib_alloc_ucontext_resp resp = {}; 1267 struct mlx5_ib_ucontext *context = to_mucontext(uctx); 1268 struct mlx5_bfreg_info *bfregi; 1269 int ver; 1270 int err; 1271 size_t min_req_v2 = offsetof(struct mlx5_ib_alloc_ucontext_req_v2, 1272 max_cqe_version); 1273 bool lib_uar_4k; 1274 bool lib_uar_dyn; 1275 1276 if (!dev->ib_active) 1277 return -EAGAIN; 1278 1279 if (udata->inlen == sizeof(struct mlx5_ib_alloc_ucontext_req)) 1280 ver = 0; 1281 else if (udata->inlen >= min_req_v2) 1282 ver = 2; 1283 else 1284 return -EINVAL; 1285 1286 err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req))); 1287 if (err) 1288 return err; 1289 1290 if (req.flags & ~MLX5_IB_ALLOC_UCTX_DEVX) 1291 return -EOPNOTSUPP; 1292 1293 if (req.comp_mask || req.reserved0 || req.reserved1 || req.reserved2) 1294 return -EOPNOTSUPP; 1295 1296 req.total_num_bfregs = ALIGN(req.total_num_bfregs, 1297 MLX5_NON_FP_BFREGS_PER_UAR); 1298 if (req.num_low_latency_bfregs > req.total_num_bfregs - 1) 1299 return -EINVAL; 1300 1301 resp.qp_tab_size = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp); 1302 if (mlx5_core_is_pf(dev->mdev) && MLX5_CAP_GEN(dev->mdev, bf)) 1303 resp.bf_reg_size = 1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size); 1304 resp.cache_line_size = cache_line_size(); 1305 resp.max_sq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq); 1306 resp.max_rq_desc_sz = MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq); 1307 resp.max_send_wqebb = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz); 1308 resp.max_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz); 1309 resp.max_srq_recv_wr = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz); 1310 resp.cqe_version = min_t(__u8, 1311 (__u8)MLX5_CAP_GEN(dev->mdev, cqe_version), 1312 req.max_cqe_version); 1313 resp.log_uar_size = MLX5_CAP_GEN(dev->mdev, uar_4k) ? 1314 MLX5_ADAPTER_PAGE_SHIFT : PAGE_SHIFT; 1315 resp.num_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? 1316 MLX5_CAP_GEN(dev->mdev, num_of_uars_per_page) : 1; 1317 resp.response_length = min(offsetof(typeof(resp), response_length) + 1318 sizeof(resp.response_length), udata->outlen); 1319 1320 lib_uar_4k = req.lib_caps & MLX5_LIB_CAP_4K_UAR; 1321 lib_uar_dyn = req.lib_caps & MLX5_LIB_CAP_DYN_UAR; 1322 bfregi = &context->bfregi; 1323 1324 if (lib_uar_dyn) { 1325 bfregi->lib_uar_dyn = lib_uar_dyn; 1326 goto uar_done; 1327 } 1328 1329 /* updates req->total_num_bfregs */ 1330 err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi); 1331 if (err) 1332 goto out_ctx; 1333 1334 mutex_init(&bfregi->lock); 1335 bfregi->lib_uar_4k = lib_uar_4k; 1336 bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count), 1337 GFP_KERNEL); 1338 if (!bfregi->count) { 1339 err = -ENOMEM; 1340 goto out_ctx; 1341 } 1342 1343 bfregi->sys_pages = kcalloc(bfregi->num_sys_pages, 1344 sizeof(*bfregi->sys_pages), 1345 GFP_KERNEL); 1346 if (!bfregi->sys_pages) { 1347 err = -ENOMEM; 1348 goto out_count; 1349 } 1350 1351 err = allocate_uars(dev, context); 1352 if (err) 1353 goto out_sys_pages; 1354 1355 uar_done: 1356 if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) { 1357 err = mlx5_ib_devx_create(dev, true); 1358 if (err < 0) 1359 goto out_uars; 1360 context->devx_uid = err; 1361 } 1362 1363 err = mlx5_ib_alloc_transport_domain(dev, &context->tdn, 1364 context->devx_uid); 1365 if (err) 1366 goto out_devx; 1367 1368 INIT_LIST_HEAD(&context->db_page_list); 1369 mutex_init(&context->db_page_mutex); 1370 1371 resp.tot_bfregs = lib_uar_dyn ? 0 : req.total_num_bfregs; 1372 resp.num_ports = MLX5_CAP_GEN(dev->mdev, num_ports); 1373 1374 if (field_avail(typeof(resp), cqe_version, udata->outlen)) 1375 resp.response_length += sizeof(resp.cqe_version); 1376 1377 if (field_avail(typeof(resp), cmds_supp_uhw, udata->outlen)) { 1378 resp.cmds_supp_uhw |= MLX5_USER_CMDS_SUPP_UHW_QUERY_DEVICE | 1379 MLX5_USER_CMDS_SUPP_UHW_CREATE_AH; 1380 resp.response_length += sizeof(resp.cmds_supp_uhw); 1381 } 1382 1383 /* 1384 * We don't want to expose information from the PCI bar that is located 1385 * after 4096 bytes, so if the arch only supports larger pages, let's 1386 * pretend we don't support reading the HCA's core clock. This is also 1387 * forced by mmap function. 1388 */ 1389 if (offsetofend(typeof(resp), hca_core_clock_offset) <= udata->outlen) { 1390 if (PAGE_SIZE <= 4096) { 1391 resp.comp_mask |= 1392 MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET; 1393 resp.hca_core_clock_offset = 1394 offsetof(struct mlx5_init_seg, internal_timer_h) % PAGE_SIZE; 1395 } 1396 resp.response_length += sizeof(resp.hca_core_clock_offset); 1397 } 1398 1399 if (offsetofend(typeof(resp), log_uar_size) <= udata->outlen) 1400 resp.response_length += sizeof(resp.log_uar_size); 1401 1402 if (offsetofend(typeof(resp), num_uars_per_page) <= udata->outlen) 1403 resp.response_length += sizeof(resp.num_uars_per_page); 1404 1405 if (offsetofend(typeof(resp), num_dyn_bfregs) <= udata->outlen) { 1406 resp.num_dyn_bfregs = bfregi->num_dyn_bfregs; 1407 resp.response_length += sizeof(resp.num_dyn_bfregs); 1408 } 1409 1410 err = ib_copy_to_udata(udata, &resp, resp.response_length); 1411 if (err) 1412 goto out_mdev; 1413 1414 bfregi->ver = ver; 1415 bfregi->num_low_latency_bfregs = req.num_low_latency_bfregs; 1416 context->cqe_version = resp.cqe_version; 1417 context->lib_caps = req.lib_caps; 1418 print_lib_caps(dev, context->lib_caps); 1419 1420 return 0; 1421 1422 out_mdev: 1423 mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid); 1424 out_devx: 1425 if (req.flags & MLX5_IB_ALLOC_UCTX_DEVX) 1426 mlx5_ib_devx_destroy(dev, context->devx_uid); 1427 1428 out_uars: 1429 deallocate_uars(dev, context); 1430 1431 out_sys_pages: 1432 kfree(bfregi->sys_pages); 1433 1434 out_count: 1435 kfree(bfregi->count); 1436 1437 out_ctx: 1438 return err; 1439 } 1440 1441 static void mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext) 1442 { 1443 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext); 1444 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device); 1445 struct mlx5_bfreg_info *bfregi; 1446 1447 bfregi = &context->bfregi; 1448 mlx5_ib_dealloc_transport_domain(dev, context->tdn, context->devx_uid); 1449 1450 if (context->devx_uid) 1451 mlx5_ib_devx_destroy(dev, context->devx_uid); 1452 1453 deallocate_uars(dev, context); 1454 kfree(bfregi->sys_pages); 1455 kfree(bfregi->count); 1456 } 1457 1458 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev, 1459 int uar_idx) 1460 { 1461 int fw_uars_per_page; 1462 1463 fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1; 1464 1465 return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + uar_idx / fw_uars_per_page; 1466 } 1467 1468 static int get_command(unsigned long offset) 1469 { 1470 return (offset >> MLX5_IB_MMAP_CMD_SHIFT) & MLX5_IB_MMAP_CMD_MASK; 1471 } 1472 1473 static int get_arg(unsigned long offset) 1474 { 1475 return offset & ((1 << MLX5_IB_MMAP_CMD_SHIFT) - 1); 1476 } 1477 1478 static int get_index(unsigned long offset) 1479 { 1480 return get_arg(offset); 1481 } 1482 1483 /* Index resides in an extra byte to enable larger values than 255 */ 1484 static int get_extended_index(unsigned long offset) 1485 { 1486 return get_arg(offset) | ((offset >> 16) & 0xff) << 8; 1487 } 1488 1489 1490 static void mlx5_ib_disassociate_ucontext(struct ib_ucontext *ibcontext) 1491 { 1492 } 1493 1494 static inline char *mmap_cmd2str(enum mlx5_ib_mmap_cmd cmd) 1495 { 1496 switch (cmd) { 1497 case MLX5_IB_MMAP_WC_PAGE: 1498 return "WC"; 1499 case MLX5_IB_MMAP_REGULAR_PAGE: 1500 return "best effort WC"; 1501 case MLX5_IB_MMAP_NC_PAGE: 1502 return "NC"; 1503 default: 1504 return NULL; 1505 } 1506 } 1507 1508 static int mlx5_ib_mmap_clock_info_page(struct mlx5_ib_dev *dev, 1509 struct vm_area_struct *vma, 1510 struct mlx5_ib_ucontext *context) 1511 { 1512 if ((vma->vm_end - vma->vm_start != PAGE_SIZE) || 1513 !(vma->vm_flags & VM_SHARED)) 1514 return -EINVAL; 1515 1516 if (get_index(vma->vm_pgoff) != MLX5_IB_CLOCK_INFO_V1) 1517 return -EOPNOTSUPP; 1518 1519 if (vma->vm_flags & (VM_WRITE | VM_EXEC)) 1520 return -EPERM; 1521 1522 return -EOPNOTSUPP; 1523 } 1524 1525 static void mlx5_ib_mmap_free(struct rdma_user_mmap_entry *entry) 1526 { 1527 struct mlx5_user_mmap_entry *mentry = to_mmmap(entry); 1528 struct mlx5_ib_dev *dev = to_mdev(entry->ucontext->device); 1529 1530 switch (mentry->mmap_flag) { 1531 case MLX5_IB_MMAP_TYPE_UAR_WC: 1532 case MLX5_IB_MMAP_TYPE_UAR_NC: 1533 mlx5_cmd_free_uar(dev->mdev, mentry->page_idx); 1534 kfree(mentry); 1535 break; 1536 default: 1537 WARN_ON(true); 1538 } 1539 } 1540 1541 static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd, 1542 struct vm_area_struct *vma, 1543 struct mlx5_ib_ucontext *context) 1544 { 1545 struct mlx5_bfreg_info *bfregi = &context->bfregi; 1546 int err; 1547 unsigned long idx; 1548 phys_addr_t pfn; 1549 pgprot_t prot; 1550 u32 bfreg_dyn_idx = 0; 1551 u32 uar_index; 1552 int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC); 1553 int max_valid_idx = dyn_uar ? bfregi->num_sys_pages : 1554 bfregi->num_static_sys_pages; 1555 1556 if (bfregi->lib_uar_dyn) 1557 return -EINVAL; 1558 1559 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 1560 return -EINVAL; 1561 1562 if (dyn_uar) 1563 idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages; 1564 else 1565 idx = get_index(vma->vm_pgoff); 1566 1567 if (idx >= max_valid_idx) { 1568 mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n", 1569 idx, max_valid_idx); 1570 return -EINVAL; 1571 } 1572 1573 switch (cmd) { 1574 case MLX5_IB_MMAP_WC_PAGE: 1575 case MLX5_IB_MMAP_ALLOC_WC: 1576 case MLX5_IB_MMAP_REGULAR_PAGE: 1577 /* For MLX5_IB_MMAP_REGULAR_PAGE do the best effort to get WC */ 1578 prot = pgprot_writecombine(vma->vm_page_prot); 1579 break; 1580 case MLX5_IB_MMAP_NC_PAGE: 1581 prot = pgprot_noncached(vma->vm_page_prot); 1582 break; 1583 default: 1584 return -EINVAL; 1585 } 1586 1587 if (dyn_uar) { 1588 int uars_per_page; 1589 1590 uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k); 1591 bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR); 1592 if (bfreg_dyn_idx >= bfregi->total_num_bfregs) { 1593 mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n", 1594 bfreg_dyn_idx, bfregi->total_num_bfregs); 1595 return -EINVAL; 1596 } 1597 1598 mutex_lock(&bfregi->lock); 1599 /* Fail if uar already allocated, first bfreg index of each 1600 * page holds its count. 1601 */ 1602 if (bfregi->count[bfreg_dyn_idx]) { 1603 mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx); 1604 mutex_unlock(&bfregi->lock); 1605 return -EINVAL; 1606 } 1607 1608 bfregi->count[bfreg_dyn_idx]++; 1609 mutex_unlock(&bfregi->lock); 1610 1611 err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index); 1612 if (err) { 1613 mlx5_ib_warn(dev, "UAR alloc failed\n"); 1614 goto free_bfreg; 1615 } 1616 } else { 1617 uar_index = bfregi->sys_pages[idx]; 1618 } 1619 1620 pfn = uar_index2pfn(dev, uar_index); 1621 mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn); 1622 1623 err = rdma_user_mmap_io(&context->ibucontext, vma, pfn, PAGE_SIZE, 1624 prot, NULL); 1625 if (err) { 1626 mlx5_ib_err(dev, 1627 "rdma_user_mmap_io failed with error=%d, mmap_cmd=%s\n", 1628 err, mmap_cmd2str(cmd)); 1629 goto err; 1630 } 1631 1632 if (dyn_uar) 1633 bfregi->sys_pages[idx] = uar_index; 1634 return 0; 1635 1636 err: 1637 if (!dyn_uar) 1638 return err; 1639 1640 mlx5_cmd_free_uar(dev->mdev, idx); 1641 1642 free_bfreg: 1643 mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx); 1644 1645 return err; 1646 } 1647 1648 static unsigned long mlx5_vma_to_pgoff(struct vm_area_struct *vma) 1649 { 1650 unsigned long idx; 1651 u8 command; 1652 1653 command = get_command(vma->vm_pgoff); 1654 idx = get_extended_index(vma->vm_pgoff); 1655 1656 return (command << 16 | idx); 1657 } 1658 1659 static int mlx5_ib_mmap_offset(struct mlx5_ib_dev *dev, 1660 struct vm_area_struct *vma, 1661 struct ib_ucontext *ucontext) 1662 { 1663 struct mlx5_user_mmap_entry *mentry; 1664 struct rdma_user_mmap_entry *entry; 1665 unsigned long pgoff; 1666 pgprot_t prot; 1667 phys_addr_t pfn; 1668 int ret; 1669 1670 pgoff = mlx5_vma_to_pgoff(vma); 1671 entry = rdma_user_mmap_entry_get_pgoff(ucontext, pgoff); 1672 if (!entry) 1673 return -EINVAL; 1674 1675 mentry = to_mmmap(entry); 1676 pfn = (mentry->address >> PAGE_SHIFT); 1677 if (mentry->mmap_flag == MLX5_IB_MMAP_TYPE_VAR || 1678 mentry->mmap_flag == MLX5_IB_MMAP_TYPE_UAR_NC) 1679 prot = pgprot_noncached(vma->vm_page_prot); 1680 else 1681 prot = pgprot_writecombine(vma->vm_page_prot); 1682 ret = rdma_user_mmap_io(ucontext, vma, pfn, 1683 entry->npages * PAGE_SIZE, 1684 prot, 1685 entry); 1686 rdma_user_mmap_entry_put(&mentry->rdma_entry); 1687 return ret; 1688 } 1689 1690 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma) 1691 { 1692 struct mlx5_ib_ucontext *context = to_mucontext(ibcontext); 1693 struct mlx5_ib_dev *dev = to_mdev(ibcontext->device); 1694 unsigned long command; 1695 phys_addr_t pfn; 1696 1697 command = get_command(vma->vm_pgoff); 1698 switch (command) { 1699 case MLX5_IB_MMAP_WC_PAGE: 1700 case MLX5_IB_MMAP_ALLOC_WC: 1701 if (!dev->wc_support) 1702 return -EPERM; 1703 /* FALLTHROUGH */ 1704 case MLX5_IB_MMAP_NC_PAGE: 1705 case MLX5_IB_MMAP_REGULAR_PAGE: 1706 return uar_mmap(dev, command, vma, context); 1707 1708 case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES: 1709 return -ENOSYS; 1710 1711 case MLX5_IB_MMAP_CORE_CLOCK: 1712 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 1713 return -EINVAL; 1714 1715 if (vma->vm_flags & VM_WRITE) 1716 return -EPERM; 1717 1718 /* Don't expose to user-space information it shouldn't have */ 1719 if (PAGE_SIZE > 4096) 1720 return -EOPNOTSUPP; 1721 1722 pfn = (dev->mdev->iseg_base + 1723 offsetof(struct mlx5_init_seg, internal_timer_h)) >> 1724 PAGE_SHIFT; 1725 return rdma_user_mmap_io(&context->ibucontext, vma, pfn, 1726 PAGE_SIZE, 1727 pgprot_noncached(vma->vm_page_prot), 1728 NULL); 1729 case MLX5_IB_MMAP_CLOCK_INFO: 1730 return mlx5_ib_mmap_clock_info_page(dev, vma, context); 1731 1732 default: 1733 return mlx5_ib_mmap_offset(dev, vma, ibcontext); 1734 } 1735 1736 return 0; 1737 } 1738 1739 static int mlx5_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) 1740 { 1741 struct mlx5_ib_pd *pd = to_mpd(ibpd); 1742 struct ib_device *ibdev = ibpd->device; 1743 struct mlx5_ib_alloc_pd_resp resp; 1744 int err; 1745 struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context( 1746 udata, struct mlx5_ib_ucontext, ibucontext); 1747 u16 uid = context ? context->devx_uid : 0; 1748 1749 err = mlx5_core_alloc_pd(to_mdev(ibdev)->mdev, &pd->pdn, uid); 1750 if (err) 1751 return (err); 1752 1753 pd->uid = uid; 1754 if (udata) { 1755 resp.pdn = pd->pdn; 1756 if (ib_copy_to_udata(udata, &resp, sizeof(resp))) { 1757 mlx5_core_dealloc_pd(to_mdev(ibdev)->mdev, pd->pdn, uid); 1758 return -EFAULT; 1759 } 1760 } 1761 1762 return 0; 1763 } 1764 1765 static void mlx5_ib_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata) 1766 { 1767 struct mlx5_ib_dev *mdev = to_mdev(pd->device); 1768 struct mlx5_ib_pd *mpd = to_mpd(pd); 1769 1770 mlx5_core_dealloc_pd(mdev->mdev, mpd->pdn, mpd->uid); 1771 } 1772 1773 enum { 1774 MATCH_CRITERIA_ENABLE_OUTER_BIT, 1775 MATCH_CRITERIA_ENABLE_MISC_BIT, 1776 MATCH_CRITERIA_ENABLE_INNER_BIT 1777 }; 1778 1779 #define HEADER_IS_ZERO(match_criteria, headers) \ 1780 !(memchr_inv(MLX5_ADDR_OF(fte_match_param, match_criteria, headers), \ 1781 0, MLX5_FLD_SZ_BYTES(fte_match_param, headers))) \ 1782 1783 static u8 get_match_criteria_enable(u32 *match_criteria) 1784 { 1785 u8 match_criteria_enable; 1786 1787 match_criteria_enable = 1788 (!HEADER_IS_ZERO(match_criteria, outer_headers)) << 1789 MATCH_CRITERIA_ENABLE_OUTER_BIT; 1790 match_criteria_enable |= 1791 (!HEADER_IS_ZERO(match_criteria, misc_parameters)) << 1792 MATCH_CRITERIA_ENABLE_MISC_BIT; 1793 match_criteria_enable |= 1794 (!HEADER_IS_ZERO(match_criteria, inner_headers)) << 1795 MATCH_CRITERIA_ENABLE_INNER_BIT; 1796 1797 return match_criteria_enable; 1798 } 1799 1800 static void set_proto(void *outer_c, void *outer_v, u8 mask, u8 val) 1801 { 1802 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_protocol, mask); 1803 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_protocol, val); 1804 } 1805 1806 static void set_tos(void *outer_c, void *outer_v, u8 mask, u8 val) 1807 { 1808 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_ecn, mask); 1809 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_ecn, val); 1810 MLX5_SET(fte_match_set_lyr_2_4, outer_c, ip_dscp, mask >> 2); 1811 MLX5_SET(fte_match_set_lyr_2_4, outer_v, ip_dscp, val >> 2); 1812 } 1813 1814 #define LAST_ETH_FIELD vlan_tag 1815 #define LAST_IB_FIELD sl 1816 #define LAST_IPV4_FIELD tos 1817 #define LAST_IPV6_FIELD traffic_class 1818 #define LAST_TCP_UDP_FIELD src_port 1819 1820 /* Field is the last supported field */ 1821 #define FIELDS_NOT_SUPPORTED(filter, field)\ 1822 memchr_inv((void *)&filter.field +\ 1823 sizeof(filter.field), 0,\ 1824 sizeof(filter) -\ 1825 offsetof(typeof(filter), field) -\ 1826 sizeof(filter.field)) 1827 1828 static int parse_flow_attr(u32 *match_c, u32 *match_v, 1829 const union ib_flow_spec *ib_spec) 1830 { 1831 void *outer_headers_c = MLX5_ADDR_OF(fte_match_param, match_c, 1832 outer_headers); 1833 void *outer_headers_v = MLX5_ADDR_OF(fte_match_param, match_v, 1834 outer_headers); 1835 void *misc_params_c = MLX5_ADDR_OF(fte_match_param, match_c, 1836 misc_parameters); 1837 void *misc_params_v = MLX5_ADDR_OF(fte_match_param, match_v, 1838 misc_parameters); 1839 1840 switch (ib_spec->type) { 1841 case IB_FLOW_SPEC_ETH: 1842 if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD)) 1843 return -ENOTSUPP; 1844 1845 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c, 1846 dmac_47_16), 1847 ib_spec->eth.mask.dst_mac); 1848 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v, 1849 dmac_47_16), 1850 ib_spec->eth.val.dst_mac); 1851 1852 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c, 1853 smac_47_16), 1854 ib_spec->eth.mask.src_mac); 1855 ether_addr_copy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v, 1856 smac_47_16), 1857 ib_spec->eth.val.src_mac); 1858 1859 if (ib_spec->eth.mask.vlan_tag) { 1860 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, 1861 cvlan_tag, 1); 1862 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, 1863 cvlan_tag, 1); 1864 1865 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, 1866 first_vid, ntohs(ib_spec->eth.mask.vlan_tag)); 1867 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, 1868 first_vid, ntohs(ib_spec->eth.val.vlan_tag)); 1869 1870 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, 1871 first_cfi, 1872 ntohs(ib_spec->eth.mask.vlan_tag) >> 12); 1873 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, 1874 first_cfi, 1875 ntohs(ib_spec->eth.val.vlan_tag) >> 12); 1876 1877 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, 1878 first_prio, 1879 ntohs(ib_spec->eth.mask.vlan_tag) >> 13); 1880 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, 1881 first_prio, 1882 ntohs(ib_spec->eth.val.vlan_tag) >> 13); 1883 } 1884 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, 1885 ethertype, ntohs(ib_spec->eth.mask.ether_type)); 1886 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, 1887 ethertype, ntohs(ib_spec->eth.val.ether_type)); 1888 break; 1889 case IB_FLOW_SPEC_IPV4: 1890 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD)) 1891 return -ENOTSUPP; 1892 1893 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, 1894 ethertype, 0xffff); 1895 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, 1896 ethertype, ETH_P_IP); 1897 1898 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c, 1899 src_ipv4_src_ipv6.ipv4_layout.ipv4), 1900 &ib_spec->ipv4.mask.src_ip, 1901 sizeof(ib_spec->ipv4.mask.src_ip)); 1902 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v, 1903 src_ipv4_src_ipv6.ipv4_layout.ipv4), 1904 &ib_spec->ipv4.val.src_ip, 1905 sizeof(ib_spec->ipv4.val.src_ip)); 1906 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c, 1907 dst_ipv4_dst_ipv6.ipv4_layout.ipv4), 1908 &ib_spec->ipv4.mask.dst_ip, 1909 sizeof(ib_spec->ipv4.mask.dst_ip)); 1910 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v, 1911 dst_ipv4_dst_ipv6.ipv4_layout.ipv4), 1912 &ib_spec->ipv4.val.dst_ip, 1913 sizeof(ib_spec->ipv4.val.dst_ip)); 1914 1915 set_tos(outer_headers_c, outer_headers_v, 1916 ib_spec->ipv4.mask.tos, ib_spec->ipv4.val.tos); 1917 1918 set_proto(outer_headers_c, outer_headers_v, 1919 ib_spec->ipv4.mask.proto, ib_spec->ipv4.val.proto); 1920 break; 1921 case IB_FLOW_SPEC_IPV6: 1922 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv6.mask, LAST_IPV6_FIELD)) 1923 return -ENOTSUPP; 1924 1925 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, 1926 ethertype, 0xffff); 1927 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, 1928 ethertype, ETH_P_IPV6); 1929 1930 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c, 1931 src_ipv4_src_ipv6.ipv6_layout.ipv6), 1932 &ib_spec->ipv6.mask.src_ip, 1933 sizeof(ib_spec->ipv6.mask.src_ip)); 1934 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v, 1935 src_ipv4_src_ipv6.ipv6_layout.ipv6), 1936 &ib_spec->ipv6.val.src_ip, 1937 sizeof(ib_spec->ipv6.val.src_ip)); 1938 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_c, 1939 dst_ipv4_dst_ipv6.ipv6_layout.ipv6), 1940 &ib_spec->ipv6.mask.dst_ip, 1941 sizeof(ib_spec->ipv6.mask.dst_ip)); 1942 memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, outer_headers_v, 1943 dst_ipv4_dst_ipv6.ipv6_layout.ipv6), 1944 &ib_spec->ipv6.val.dst_ip, 1945 sizeof(ib_spec->ipv6.val.dst_ip)); 1946 1947 set_tos(outer_headers_c, outer_headers_v, 1948 ib_spec->ipv6.mask.traffic_class, 1949 ib_spec->ipv6.val.traffic_class); 1950 1951 set_proto(outer_headers_c, outer_headers_v, 1952 ib_spec->ipv6.mask.next_hdr, 1953 ib_spec->ipv6.val.next_hdr); 1954 1955 MLX5_SET(fte_match_set_misc, misc_params_c, 1956 outer_ipv6_flow_label, 1957 ntohl(ib_spec->ipv6.mask.flow_label)); 1958 MLX5_SET(fte_match_set_misc, misc_params_v, 1959 outer_ipv6_flow_label, 1960 ntohl(ib_spec->ipv6.val.flow_label)); 1961 break; 1962 case IB_FLOW_SPEC_TCP: 1963 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask, 1964 LAST_TCP_UDP_FIELD)) 1965 return -ENOTSUPP; 1966 1967 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol, 1968 0xff); 1969 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol, 1970 IPPROTO_TCP); 1971 1972 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_sport, 1973 ntohs(ib_spec->tcp_udp.mask.src_port)); 1974 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_sport, 1975 ntohs(ib_spec->tcp_udp.val.src_port)); 1976 1977 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, tcp_dport, 1978 ntohs(ib_spec->tcp_udp.mask.dst_port)); 1979 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, tcp_dport, 1980 ntohs(ib_spec->tcp_udp.val.dst_port)); 1981 break; 1982 case IB_FLOW_SPEC_UDP: 1983 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask, 1984 LAST_TCP_UDP_FIELD)) 1985 return -ENOTSUPP; 1986 1987 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, ip_protocol, 1988 0xff); 1989 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, ip_protocol, 1990 IPPROTO_UDP); 1991 1992 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_sport, 1993 ntohs(ib_spec->tcp_udp.mask.src_port)); 1994 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_sport, 1995 ntohs(ib_spec->tcp_udp.val.src_port)); 1996 1997 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_c, udp_dport, 1998 ntohs(ib_spec->tcp_udp.mask.dst_port)); 1999 MLX5_SET(fte_match_set_lyr_2_4, outer_headers_v, udp_dport, 2000 ntohs(ib_spec->tcp_udp.val.dst_port)); 2001 break; 2002 default: 2003 return -EINVAL; 2004 } 2005 2006 return 0; 2007 } 2008 2009 /* If a flow could catch both multicast and unicast packets, 2010 * it won't fall into the multicast flow steering table and this rule 2011 * could steal other multicast packets. 2012 */ 2013 static bool flow_is_multicast_only(struct ib_flow_attr *ib_attr) 2014 { 2015 struct ib_flow_spec_eth *eth_spec; 2016 2017 if (ib_attr->type != IB_FLOW_ATTR_NORMAL || 2018 ib_attr->size < sizeof(struct ib_flow_attr) + 2019 sizeof(struct ib_flow_spec_eth) || 2020 ib_attr->num_of_specs < 1) 2021 return false; 2022 2023 eth_spec = (struct ib_flow_spec_eth *)(ib_attr + 1); 2024 if (eth_spec->type != IB_FLOW_SPEC_ETH || 2025 eth_spec->size != sizeof(*eth_spec)) 2026 return false; 2027 2028 return is_multicast_ether_addr(eth_spec->mask.dst_mac) && 2029 is_multicast_ether_addr(eth_spec->val.dst_mac); 2030 } 2031 2032 static bool is_valid_attr(const struct ib_flow_attr *flow_attr) 2033 { 2034 union ib_flow_spec *ib_spec = (union ib_flow_spec *)(flow_attr + 1); 2035 bool has_ipv4_spec = false; 2036 bool eth_type_ipv4 = true; 2037 unsigned int spec_index; 2038 2039 /* Validate that ethertype is correct */ 2040 for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) { 2041 if (ib_spec->type == IB_FLOW_SPEC_ETH && 2042 ib_spec->eth.mask.ether_type) { 2043 if (!((ib_spec->eth.mask.ether_type == htons(0xffff)) && 2044 ib_spec->eth.val.ether_type == htons(ETH_P_IP))) 2045 eth_type_ipv4 = false; 2046 } else if (ib_spec->type == IB_FLOW_SPEC_IPV4) { 2047 has_ipv4_spec = true; 2048 } 2049 ib_spec = (void *)ib_spec + ib_spec->size; 2050 } 2051 return !has_ipv4_spec || eth_type_ipv4; 2052 } 2053 2054 static void put_flow_table(struct mlx5_ib_dev *dev, 2055 struct mlx5_ib_flow_prio *prio, bool ft_added) 2056 { 2057 prio->refcount -= !!ft_added; 2058 if (!prio->refcount) { 2059 mlx5_destroy_flow_table(prio->flow_table); 2060 prio->flow_table = NULL; 2061 } 2062 } 2063 2064 static int mlx5_ib_destroy_flow(struct ib_flow *flow_id) 2065 { 2066 struct mlx5_ib_dev *dev = to_mdev(flow_id->qp->device); 2067 struct mlx5_ib_flow_handler *handler = container_of(flow_id, 2068 struct mlx5_ib_flow_handler, 2069 ibflow); 2070 struct mlx5_ib_flow_handler *iter, *tmp; 2071 2072 mutex_lock(&dev->flow_db.lock); 2073 2074 list_for_each_entry_safe(iter, tmp, &handler->list, list) { 2075 mlx5_del_flow_rules(&iter->rule); 2076 put_flow_table(dev, iter->prio, true); 2077 list_del(&iter->list); 2078 kfree(iter); 2079 } 2080 2081 mlx5_del_flow_rules(&handler->rule); 2082 put_flow_table(dev, handler->prio, true); 2083 mutex_unlock(&dev->flow_db.lock); 2084 2085 kfree(handler); 2086 2087 return 0; 2088 } 2089 2090 static int ib_prio_to_core_prio(unsigned int priority, bool dont_trap) 2091 { 2092 priority *= 2; 2093 if (!dont_trap) 2094 priority++; 2095 return priority; 2096 } 2097 2098 enum flow_table_type { 2099 MLX5_IB_FT_RX, 2100 MLX5_IB_FT_TX 2101 }; 2102 2103 #define MLX5_FS_MAX_TYPES 10 2104 #define MLX5_FS_MAX_ENTRIES 32000UL 2105 static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev, 2106 struct ib_flow_attr *flow_attr, 2107 enum flow_table_type ft_type) 2108 { 2109 bool dont_trap = flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP; 2110 struct mlx5_flow_table_attr ft_attr = {}; 2111 struct mlx5_flow_namespace *ns = NULL; 2112 struct mlx5_ib_flow_prio *prio; 2113 struct mlx5_flow_table *ft; 2114 int num_entries; 2115 int num_groups; 2116 int priority; 2117 int err = 0; 2118 2119 if (flow_attr->type == IB_FLOW_ATTR_NORMAL) { 2120 if (flow_is_multicast_only(flow_attr) && 2121 !dont_trap) 2122 priority = MLX5_IB_FLOW_MCAST_PRIO; 2123 else 2124 priority = ib_prio_to_core_prio(flow_attr->priority, 2125 dont_trap); 2126 ns = mlx5_get_flow_namespace(dev->mdev, 2127 MLX5_FLOW_NAMESPACE_BYPASS); 2128 num_entries = MLX5_FS_MAX_ENTRIES; 2129 num_groups = MLX5_FS_MAX_TYPES; 2130 prio = &dev->flow_db.prios[priority]; 2131 } else if (flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT || 2132 flow_attr->type == IB_FLOW_ATTR_MC_DEFAULT) { 2133 ns = mlx5_get_flow_namespace(dev->mdev, 2134 MLX5_FLOW_NAMESPACE_LEFTOVERS); 2135 build_leftovers_ft_param("bypass", &priority, 2136 &num_entries, 2137 &num_groups); 2138 prio = &dev->flow_db.prios[MLX5_IB_FLOW_LEFTOVERS_PRIO]; 2139 } else if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) { 2140 if (!MLX5_CAP_FLOWTABLE(dev->mdev, 2141 allow_sniffer_and_nic_rx_shared_tir)) 2142 return ERR_PTR(-ENOTSUPP); 2143 2144 ns = mlx5_get_flow_namespace(dev->mdev, ft_type == MLX5_IB_FT_RX ? 2145 MLX5_FLOW_NAMESPACE_SNIFFER_RX : 2146 MLX5_FLOW_NAMESPACE_SNIFFER_TX); 2147 2148 prio = &dev->flow_db.sniffer[ft_type]; 2149 priority = 0; 2150 num_entries = 1; 2151 num_groups = 1; 2152 } 2153 2154 if (!ns) 2155 return ERR_PTR(-ENOTSUPP); 2156 2157 ft = prio->flow_table; 2158 if (!ft) { 2159 ft_attr.prio = priority; 2160 ft_attr.max_fte = num_entries; 2161 ft_attr.autogroup.max_num_groups = num_groups; 2162 2163 ft = mlx5_create_auto_grouped_flow_table(ns, &ft_attr); 2164 2165 if (!IS_ERR(ft)) { 2166 prio->refcount = 0; 2167 prio->flow_table = ft; 2168 } else { 2169 err = PTR_ERR(ft); 2170 } 2171 } 2172 2173 return err ? ERR_PTR(err) : prio; 2174 } 2175 2176 static struct mlx5_ib_flow_handler *create_flow_rule(struct mlx5_ib_dev *dev, 2177 struct mlx5_ib_flow_prio *ft_prio, 2178 const struct ib_flow_attr *flow_attr, 2179 struct mlx5_flow_destination *dst) 2180 { 2181 struct mlx5_flow_table *ft = ft_prio->flow_table; 2182 struct mlx5_ib_flow_handler *handler; 2183 struct mlx5_flow_spec *spec; 2184 const void *ib_flow = (const void *)flow_attr + sizeof(*flow_attr); 2185 unsigned int spec_index; 2186 struct mlx5_flow_act flow_act = {}; 2187 2188 u32 action; 2189 int err = 0; 2190 2191 if (!is_valid_attr(flow_attr)) 2192 return ERR_PTR(-EINVAL); 2193 2194 spec = mlx5_vzalloc(sizeof(*spec)); 2195 handler = kzalloc(sizeof(*handler), GFP_KERNEL); 2196 if (!handler || !spec) { 2197 err = -ENOMEM; 2198 goto free; 2199 } 2200 2201 spec->flow_context.flags = FLOW_CONTEXT_HAS_TAG; 2202 spec->flow_context.flow_tag = MLX5_FS_DEFAULT_FLOW_TAG; 2203 2204 INIT_LIST_HEAD(&handler->list); 2205 2206 for (spec_index = 0; spec_index < flow_attr->num_of_specs; spec_index++) { 2207 err = parse_flow_attr(spec->match_criteria, 2208 spec->match_value, ib_flow); 2209 if (err < 0) 2210 goto free; 2211 2212 ib_flow += ((union ib_flow_spec *)ib_flow)->size; 2213 } 2214 2215 spec->match_criteria_enable = get_match_criteria_enable(spec->match_criteria); 2216 action = dst ? MLX5_FLOW_RULE_FWD_ACTION_DEST : 0; 2217 flow_act.action = action; 2218 handler->rule = mlx5_add_flow_rules(ft, spec, &flow_act, dst, 1); 2219 2220 if (IS_ERR(handler->rule)) { 2221 err = PTR_ERR(handler->rule); 2222 goto free; 2223 } 2224 2225 ft_prio->refcount++; 2226 handler->prio = ft_prio; 2227 2228 ft_prio->flow_table = ft; 2229 free: 2230 if (err) 2231 kfree(handler); 2232 kvfree(spec); 2233 return err ? ERR_PTR(err) : handler; 2234 } 2235 2236 static struct mlx5_ib_flow_handler *create_dont_trap_rule(struct mlx5_ib_dev *dev, 2237 struct mlx5_ib_flow_prio *ft_prio, 2238 struct ib_flow_attr *flow_attr, 2239 struct mlx5_flow_destination *dst) 2240 { 2241 struct mlx5_ib_flow_handler *handler_dst = NULL; 2242 struct mlx5_ib_flow_handler *handler = NULL; 2243 2244 handler = create_flow_rule(dev, ft_prio, flow_attr, NULL); 2245 if (!IS_ERR(handler)) { 2246 handler_dst = create_flow_rule(dev, ft_prio, 2247 flow_attr, dst); 2248 if (IS_ERR(handler_dst)) { 2249 mlx5_del_flow_rules(&handler->rule); 2250 ft_prio->refcount--; 2251 kfree(handler); 2252 handler = handler_dst; 2253 } else { 2254 list_add(&handler_dst->list, &handler->list); 2255 } 2256 } 2257 2258 return handler; 2259 } 2260 enum { 2261 LEFTOVERS_MC, 2262 LEFTOVERS_UC, 2263 }; 2264 2265 static struct mlx5_ib_flow_handler *create_leftovers_rule(struct mlx5_ib_dev *dev, 2266 struct mlx5_ib_flow_prio *ft_prio, 2267 struct ib_flow_attr *flow_attr, 2268 struct mlx5_flow_destination *dst) 2269 { 2270 struct mlx5_ib_flow_handler *handler_ucast = NULL; 2271 struct mlx5_ib_flow_handler *handler = NULL; 2272 2273 static struct { 2274 struct ib_flow_attr flow_attr; 2275 struct ib_flow_spec_eth eth_flow; 2276 } leftovers_specs[] = { 2277 [LEFTOVERS_MC] = { 2278 .flow_attr = { 2279 .num_of_specs = 1, 2280 .size = sizeof(leftovers_specs[0]) 2281 }, 2282 .eth_flow = { 2283 .type = IB_FLOW_SPEC_ETH, 2284 .size = sizeof(struct ib_flow_spec_eth), 2285 .mask = {.dst_mac = {0x1} }, 2286 .val = {.dst_mac = {0x1} } 2287 } 2288 }, 2289 [LEFTOVERS_UC] = { 2290 .flow_attr = { 2291 .num_of_specs = 1, 2292 .size = sizeof(leftovers_specs[0]) 2293 }, 2294 .eth_flow = { 2295 .type = IB_FLOW_SPEC_ETH, 2296 .size = sizeof(struct ib_flow_spec_eth), 2297 .mask = {.dst_mac = {0x1} }, 2298 .val = {.dst_mac = {} } 2299 } 2300 } 2301 }; 2302 2303 handler = create_flow_rule(dev, ft_prio, 2304 &leftovers_specs[LEFTOVERS_MC].flow_attr, 2305 dst); 2306 if (!IS_ERR(handler) && 2307 flow_attr->type == IB_FLOW_ATTR_ALL_DEFAULT) { 2308 handler_ucast = create_flow_rule(dev, ft_prio, 2309 &leftovers_specs[LEFTOVERS_UC].flow_attr, 2310 dst); 2311 if (IS_ERR(handler_ucast)) { 2312 mlx5_del_flow_rules(&handler->rule); 2313 ft_prio->refcount--; 2314 kfree(handler); 2315 handler = handler_ucast; 2316 } else { 2317 list_add(&handler_ucast->list, &handler->list); 2318 } 2319 } 2320 2321 return handler; 2322 } 2323 2324 static struct mlx5_ib_flow_handler *create_sniffer_rule(struct mlx5_ib_dev *dev, 2325 struct mlx5_ib_flow_prio *ft_rx, 2326 struct mlx5_ib_flow_prio *ft_tx, 2327 struct mlx5_flow_destination *dst) 2328 { 2329 struct mlx5_ib_flow_handler *handler_rx; 2330 struct mlx5_ib_flow_handler *handler_tx; 2331 int err; 2332 static const struct ib_flow_attr flow_attr = { 2333 .num_of_specs = 0, 2334 .type = IB_FLOW_ATTR_SNIFFER, 2335 .size = sizeof(flow_attr) 2336 }; 2337 2338 handler_rx = create_flow_rule(dev, ft_rx, &flow_attr, dst); 2339 if (IS_ERR(handler_rx)) { 2340 err = PTR_ERR(handler_rx); 2341 goto err; 2342 } 2343 2344 handler_tx = create_flow_rule(dev, ft_tx, &flow_attr, dst); 2345 if (IS_ERR(handler_tx)) { 2346 err = PTR_ERR(handler_tx); 2347 goto err_tx; 2348 } 2349 2350 list_add(&handler_tx->list, &handler_rx->list); 2351 2352 return handler_rx; 2353 2354 err_tx: 2355 mlx5_del_flow_rules(&handler_rx->rule); 2356 ft_rx->refcount--; 2357 kfree(handler_rx); 2358 err: 2359 return ERR_PTR(err); 2360 } 2361 2362 static struct ib_flow *mlx5_ib_create_flow(struct ib_qp *qp, 2363 struct ib_flow_attr *flow_attr, 2364 int domain, 2365 struct ib_udata *udata) 2366 { 2367 struct mlx5_ib_dev *dev = to_mdev(qp->device); 2368 struct mlx5_ib_qp *mqp = to_mqp(qp); 2369 struct mlx5_ib_flow_handler *handler = NULL; 2370 struct mlx5_flow_destination *dst = NULL; 2371 struct mlx5_ib_flow_prio *ft_prio_tx = NULL; 2372 struct mlx5_ib_flow_prio *ft_prio; 2373 struct mlx5_ib_create_flow *ucmd = NULL, ucmd_hdr; 2374 size_t min_ucmd_sz, required_ucmd_sz; 2375 int err; 2376 2377 if (udata && udata->inlen) { 2378 min_ucmd_sz = offsetofend(struct mlx5_ib_create_flow, reserved); 2379 if (udata->inlen < min_ucmd_sz) 2380 return ERR_PTR(-EOPNOTSUPP); 2381 2382 err = ib_copy_from_udata(&ucmd_hdr, udata, min_ucmd_sz); 2383 if (err) 2384 return ERR_PTR(err); 2385 2386 /* currently supports only one counters data */ 2387 if (ucmd_hdr.ncounters_data > 1) 2388 return ERR_PTR(-EINVAL); 2389 2390 required_ucmd_sz = min_ucmd_sz + 2391 sizeof(struct mlx5_ib_flow_counters_data) * 2392 ucmd_hdr.ncounters_data; 2393 if (udata->inlen > required_ucmd_sz && 2394 !ib_is_udata_cleared(udata, required_ucmd_sz, 2395 udata->inlen - required_ucmd_sz)) 2396 return ERR_PTR(-EOPNOTSUPP); 2397 2398 ucmd = kzalloc(required_ucmd_sz, GFP_KERNEL); 2399 if (!ucmd) 2400 return ERR_PTR(-ENOMEM); 2401 2402 err = ib_copy_from_udata(ucmd, udata, required_ucmd_sz); 2403 if (err) 2404 goto free_ucmd; 2405 } 2406 2407 if (flow_attr->priority > MLX5_IB_FLOW_LAST_PRIO) { 2408 err = -ENOMEM; 2409 goto free_ucmd; 2410 } 2411 2412 if (flow_attr->flags & ~IB_FLOW_ATTR_FLAGS_DONT_TRAP) { 2413 err = -EINVAL; 2414 goto free_ucmd; 2415 } 2416 2417 dst = kzalloc(sizeof(*dst), GFP_KERNEL); 2418 if (!dst) { 2419 err = -ENOMEM; 2420 goto free_ucmd; 2421 } 2422 2423 mutex_lock(&dev->flow_db.lock); 2424 2425 ft_prio = get_flow_table(dev, flow_attr, MLX5_IB_FT_RX); 2426 if (IS_ERR(ft_prio)) { 2427 err = PTR_ERR(ft_prio); 2428 goto unlock; 2429 } 2430 if (flow_attr->type == IB_FLOW_ATTR_SNIFFER) { 2431 ft_prio_tx = get_flow_table(dev, flow_attr, MLX5_IB_FT_TX); 2432 if (IS_ERR(ft_prio_tx)) { 2433 err = PTR_ERR(ft_prio_tx); 2434 ft_prio_tx = NULL; 2435 goto destroy_ft; 2436 } 2437 } 2438 2439 dst->type = MLX5_FLOW_DESTINATION_TYPE_TIR; 2440 if (mqp->flags & MLX5_IB_QP_RSS) 2441 dst->tir_num = mqp->rss_qp.tirn; 2442 else 2443 dst->tir_num = mqp->raw_packet_qp.rq.tirn; 2444 2445 switch (flow_attr->type) { 2446 case IB_FLOW_ATTR_NORMAL: 2447 if (mqp->flags & IB_QP_CREATE_SOURCE_QPN) { 2448 err = -EOPNOTSUPP; 2449 goto destroy_ft; 2450 } 2451 if (flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) { 2452 handler = create_dont_trap_rule(dev, ft_prio, flow_attr, dst); 2453 } else { 2454 handler = create_flow_rule(dev, ft_prio, flow_attr, dst); 2455 } 2456 break; 2457 case IB_FLOW_ATTR_ALL_DEFAULT: 2458 case IB_FLOW_ATTR_MC_DEFAULT: 2459 handler = create_leftovers_rule(dev, ft_prio, flow_attr, dst); 2460 break; 2461 case IB_FLOW_ATTR_SNIFFER: 2462 handler = create_sniffer_rule(dev, ft_prio, ft_prio_tx, dst); 2463 break; 2464 default: 2465 err = -EINVAL; 2466 goto destroy_ft; 2467 } 2468 2469 if (IS_ERR(handler)) { 2470 err = PTR_ERR(handler); 2471 handler = NULL; 2472 goto destroy_ft; 2473 } 2474 2475 mutex_unlock(&dev->flow_db.lock); 2476 kfree(dst); 2477 kfree(ucmd); 2478 2479 return &handler->ibflow; 2480 2481 destroy_ft: 2482 put_flow_table(dev, ft_prio, false); 2483 if (ft_prio_tx) 2484 put_flow_table(dev, ft_prio_tx, false); 2485 unlock: 2486 mutex_unlock(&dev->flow_db.lock); 2487 kfree(dst); 2488 free_ucmd: 2489 kfree(ucmd); 2490 return ERR_PTR(err); 2491 } 2492 2493 static int mlx5_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid) 2494 { 2495 struct mlx5_ib_dev *dev = to_mdev(ibqp->device); 2496 int err; 2497 2498 err = mlx5_core_attach_mcg(dev->mdev, gid, ibqp->qp_num); 2499 if (err) 2500 mlx5_ib_warn(dev, "failed attaching QPN 0x%x, MGID %pI6\n", 2501 ibqp->qp_num, gid->raw); 2502 2503 return err; 2504 } 2505 2506 static int mlx5_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid) 2507 { 2508 struct mlx5_ib_dev *dev = to_mdev(ibqp->device); 2509 int err; 2510 2511 err = mlx5_core_detach_mcg(dev->mdev, gid, ibqp->qp_num); 2512 if (err) 2513 mlx5_ib_warn(dev, "failed detaching QPN 0x%x, MGID %pI6\n", 2514 ibqp->qp_num, gid->raw); 2515 2516 return err; 2517 } 2518 2519 static int init_node_data(struct mlx5_ib_dev *dev) 2520 { 2521 int err; 2522 2523 err = mlx5_query_node_desc(dev, dev->ib_dev.node_desc); 2524 if (err) 2525 return err; 2526 2527 return mlx5_query_node_guid(dev, &dev->ib_dev.node_guid); 2528 } 2529 2530 static ssize_t show_fw_pages(struct device *device, struct device_attribute *attr, 2531 char *buf) 2532 { 2533 struct mlx5_ib_dev *dev = 2534 container_of(device, struct mlx5_ib_dev, ib_dev.dev); 2535 2536 return sprintf(buf, "%lld\n", (long long)dev->mdev->priv.fw_pages); 2537 } 2538 2539 static ssize_t show_reg_pages(struct device *device, 2540 struct device_attribute *attr, char *buf) 2541 { 2542 struct mlx5_ib_dev *dev = 2543 container_of(device, struct mlx5_ib_dev, ib_dev.dev); 2544 2545 return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages)); 2546 } 2547 2548 static ssize_t show_hca(struct device *device, struct device_attribute *attr, 2549 char *buf) 2550 { 2551 struct mlx5_ib_dev *dev = 2552 container_of(device, struct mlx5_ib_dev, ib_dev.dev); 2553 return sprintf(buf, "MT%d\n", dev->mdev->pdev->device); 2554 } 2555 2556 static ssize_t show_rev(struct device *device, struct device_attribute *attr, 2557 char *buf) 2558 { 2559 struct mlx5_ib_dev *dev = 2560 container_of(device, struct mlx5_ib_dev, ib_dev.dev); 2561 return sprintf(buf, "%x\n", dev->mdev->pdev->revision); 2562 } 2563 2564 static ssize_t show_board(struct device *device, struct device_attribute *attr, 2565 char *buf) 2566 { 2567 struct mlx5_ib_dev *dev = 2568 container_of(device, struct mlx5_ib_dev, ib_dev.dev); 2569 return sprintf(buf, "%.*s\n", MLX5_BOARD_ID_LEN, 2570 dev->mdev->board_id); 2571 } 2572 2573 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL); 2574 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL); 2575 static DEVICE_ATTR(board_id, S_IRUGO, show_board, NULL); 2576 static DEVICE_ATTR(fw_pages, S_IRUGO, show_fw_pages, NULL); 2577 static DEVICE_ATTR(reg_pages, S_IRUGO, show_reg_pages, NULL); 2578 2579 static struct device_attribute *mlx5_class_attributes[] = { 2580 &dev_attr_hw_rev, 2581 &dev_attr_hca_type, 2582 &dev_attr_board_id, 2583 &dev_attr_fw_pages, 2584 &dev_attr_reg_pages, 2585 }; 2586 2587 static void pkey_change_handler(struct work_struct *work) 2588 { 2589 struct mlx5_ib_port_resources *ports = 2590 container_of(work, struct mlx5_ib_port_resources, 2591 pkey_change_work); 2592 2593 mutex_lock(&ports->devr->mutex); 2594 mlx5_ib_gsi_pkey_change(ports->gsi); 2595 mutex_unlock(&ports->devr->mutex); 2596 } 2597 2598 static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev) 2599 { 2600 struct mlx5_ib_qp *mqp; 2601 struct mlx5_ib_cq *send_mcq, *recv_mcq; 2602 struct mlx5_core_cq *mcq; 2603 struct list_head cq_armed_list; 2604 unsigned long flags_qp; 2605 unsigned long flags_cq; 2606 unsigned long flags; 2607 2608 INIT_LIST_HEAD(&cq_armed_list); 2609 2610 /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/ 2611 spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags); 2612 list_for_each_entry(mqp, &ibdev->qp_list, qps_list) { 2613 spin_lock_irqsave(&mqp->sq.lock, flags_qp); 2614 if (mqp->sq.tail != mqp->sq.head) { 2615 send_mcq = to_mcq(mqp->ibqp.send_cq); 2616 spin_lock_irqsave(&send_mcq->lock, flags_cq); 2617 if (send_mcq->mcq.comp && 2618 mqp->ibqp.send_cq->comp_handler) { 2619 if (!send_mcq->mcq.reset_notify_added) { 2620 send_mcq->mcq.reset_notify_added = 1; 2621 list_add_tail(&send_mcq->mcq.reset_notify, 2622 &cq_armed_list); 2623 } 2624 } 2625 spin_unlock_irqrestore(&send_mcq->lock, flags_cq); 2626 } 2627 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp); 2628 spin_lock_irqsave(&mqp->rq.lock, flags_qp); 2629 /* no handling is needed for SRQ */ 2630 if (!mqp->ibqp.srq) { 2631 if (mqp->rq.tail != mqp->rq.head) { 2632 recv_mcq = to_mcq(mqp->ibqp.recv_cq); 2633 spin_lock_irqsave(&recv_mcq->lock, flags_cq); 2634 if (recv_mcq->mcq.comp && 2635 mqp->ibqp.recv_cq->comp_handler) { 2636 if (!recv_mcq->mcq.reset_notify_added) { 2637 recv_mcq->mcq.reset_notify_added = 1; 2638 list_add_tail(&recv_mcq->mcq.reset_notify, 2639 &cq_armed_list); 2640 } 2641 } 2642 spin_unlock_irqrestore(&recv_mcq->lock, 2643 flags_cq); 2644 } 2645 } 2646 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp); 2647 } 2648 /*At that point all inflight post send were put to be executed as of we 2649 * lock/unlock above locks Now need to arm all involved CQs. 2650 */ 2651 list_for_each_entry(mcq, &cq_armed_list, reset_notify) { 2652 mcq->comp(mcq, NULL); 2653 } 2654 spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags); 2655 } 2656 2657 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context, 2658 enum mlx5_dev_event event, unsigned long param) 2659 { 2660 struct mlx5_ib_dev *ibdev = (struct mlx5_ib_dev *)context; 2661 struct ib_event ibev; 2662 bool fatal = false; 2663 u8 port = (u8)param; 2664 2665 switch (event) { 2666 case MLX5_DEV_EVENT_SYS_ERROR: 2667 ibev.event = IB_EVENT_DEVICE_FATAL; 2668 mlx5_ib_handle_internal_error(ibdev); 2669 fatal = true; 2670 break; 2671 2672 case MLX5_DEV_EVENT_PORT_UP: 2673 case MLX5_DEV_EVENT_PORT_DOWN: 2674 case MLX5_DEV_EVENT_PORT_INITIALIZED: 2675 /* In RoCE, port up/down events are handled in 2676 * mlx5_netdev_event(). 2677 */ 2678 if (mlx5_ib_port_link_layer(&ibdev->ib_dev, port) == 2679 IB_LINK_LAYER_ETHERNET) 2680 return; 2681 2682 ibev.event = (event == MLX5_DEV_EVENT_PORT_UP) ? 2683 IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR; 2684 break; 2685 2686 case MLX5_DEV_EVENT_LID_CHANGE: 2687 ibev.event = IB_EVENT_LID_CHANGE; 2688 break; 2689 2690 case MLX5_DEV_EVENT_PKEY_CHANGE: 2691 ibev.event = IB_EVENT_PKEY_CHANGE; 2692 2693 schedule_work(&ibdev->devr.ports[port - 1].pkey_change_work); 2694 break; 2695 2696 case MLX5_DEV_EVENT_GUID_CHANGE: 2697 ibev.event = IB_EVENT_GID_CHANGE; 2698 break; 2699 2700 case MLX5_DEV_EVENT_CLIENT_REREG: 2701 ibev.event = IB_EVENT_CLIENT_REREGISTER; 2702 break; 2703 2704 default: 2705 /* unsupported event */ 2706 return; 2707 } 2708 2709 ibev.device = &ibdev->ib_dev; 2710 ibev.element.port_num = port; 2711 2712 if (!rdma_is_port_valid(&ibdev->ib_dev, port)) { 2713 mlx5_ib_warn(ibdev, "warning: event(%d) on port %d\n", event, port); 2714 return; 2715 } 2716 2717 if (ibdev->ib_active) 2718 ib_dispatch_event(&ibev); 2719 2720 if (fatal) 2721 ibdev->ib_active = false; 2722 } 2723 2724 static void get_ext_port_caps(struct mlx5_ib_dev *dev) 2725 { 2726 int port; 2727 2728 for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) 2729 mlx5_query_ext_port_caps(dev, port); 2730 } 2731 2732 static int get_port_caps(struct mlx5_ib_dev *dev) 2733 { 2734 struct ib_device_attr *dprops = NULL; 2735 struct ib_port_attr *pprops = NULL; 2736 int err = -ENOMEM; 2737 int port; 2738 struct ib_udata uhw = {.inlen = 0, .outlen = 0}; 2739 2740 pprops = kmalloc(sizeof(*pprops), GFP_KERNEL); 2741 if (!pprops) 2742 goto out; 2743 2744 dprops = kmalloc(sizeof(*dprops), GFP_KERNEL); 2745 if (!dprops) 2746 goto out; 2747 2748 err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw); 2749 if (err) { 2750 mlx5_ib_warn(dev, "query_device failed %d\n", err); 2751 goto out; 2752 } 2753 2754 for (port = 1; port <= MLX5_CAP_GEN(dev->mdev, num_ports); port++) { 2755 err = mlx5_ib_query_port(&dev->ib_dev, port, pprops); 2756 if (err) { 2757 mlx5_ib_warn(dev, "query_port %d failed %d\n", 2758 port, err); 2759 break; 2760 } 2761 dev->mdev->port_caps[port - 1].pkey_table_len = 2762 dprops->max_pkeys; 2763 dev->mdev->port_caps[port - 1].gid_table_len = 2764 pprops->gid_tbl_len; 2765 mlx5_ib_dbg(dev, "pkey_table_len %d, gid_table_len %d\n", 2766 dprops->max_pkeys, pprops->gid_tbl_len); 2767 } 2768 2769 out: 2770 kfree(pprops); 2771 kfree(dprops); 2772 2773 return err; 2774 } 2775 2776 static void destroy_umrc_res(struct mlx5_ib_dev *dev) 2777 { 2778 int err; 2779 2780 err = mlx5_mr_cache_cleanup(dev); 2781 if (err) 2782 mlx5_ib_warn(dev, "mr cache cleanup failed\n"); 2783 2784 if (dev->umrc.qp) 2785 mlx5_ib_destroy_qp(dev->umrc.qp, NULL); 2786 if (dev->umrc.cq) 2787 ib_free_cq(dev->umrc.cq); 2788 if (dev->umrc.pd) 2789 ib_dealloc_pd(dev->umrc.pd); 2790 } 2791 2792 enum { 2793 MAX_UMR_WR = 128, 2794 }; 2795 2796 static int create_umr_res(struct mlx5_ib_dev *dev) 2797 { 2798 struct ib_qp_init_attr *init_attr = NULL; 2799 struct ib_qp_attr *attr = NULL; 2800 struct ib_pd *pd; 2801 struct ib_cq *cq; 2802 struct ib_qp *qp; 2803 int ret; 2804 2805 attr = kzalloc(sizeof(*attr), GFP_KERNEL); 2806 init_attr = kzalloc(sizeof(*init_attr), GFP_KERNEL); 2807 if (!attr || !init_attr) { 2808 ret = -ENOMEM; 2809 goto error_0; 2810 } 2811 2812 pd = ib_alloc_pd(&dev->ib_dev, 0); 2813 if (IS_ERR(pd)) { 2814 mlx5_ib_dbg(dev, "Couldn't create PD for sync UMR QP\n"); 2815 ret = PTR_ERR(pd); 2816 goto error_0; 2817 } 2818 2819 cq = ib_alloc_cq(&dev->ib_dev, NULL, 128, 0, IB_POLL_SOFTIRQ); 2820 if (IS_ERR(cq)) { 2821 mlx5_ib_dbg(dev, "Couldn't create CQ for sync UMR QP\n"); 2822 ret = PTR_ERR(cq); 2823 goto error_2; 2824 } 2825 2826 init_attr->send_cq = cq; 2827 init_attr->recv_cq = cq; 2828 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR; 2829 init_attr->cap.max_send_wr = MAX_UMR_WR; 2830 init_attr->cap.max_send_sge = 1; 2831 init_attr->qp_type = MLX5_IB_QPT_REG_UMR; 2832 init_attr->port_num = 1; 2833 qp = mlx5_ib_create_qp(pd, init_attr, NULL); 2834 if (IS_ERR(qp)) { 2835 mlx5_ib_dbg(dev, "Couldn't create sync UMR QP\n"); 2836 ret = PTR_ERR(qp); 2837 goto error_3; 2838 } 2839 qp->device = &dev->ib_dev; 2840 qp->real_qp = qp; 2841 qp->uobject = NULL; 2842 qp->qp_type = MLX5_IB_QPT_REG_UMR; 2843 2844 attr->qp_state = IB_QPS_INIT; 2845 attr->port_num = 1; 2846 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_PKEY_INDEX | 2847 IB_QP_PORT, NULL); 2848 if (ret) { 2849 mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n"); 2850 goto error_4; 2851 } 2852 2853 memset(attr, 0, sizeof(*attr)); 2854 attr->qp_state = IB_QPS_RTR; 2855 attr->path_mtu = IB_MTU_256; 2856 2857 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL); 2858 if (ret) { 2859 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rtr\n"); 2860 goto error_4; 2861 } 2862 2863 memset(attr, 0, sizeof(*attr)); 2864 attr->qp_state = IB_QPS_RTS; 2865 ret = mlx5_ib_modify_qp(qp, attr, IB_QP_STATE, NULL); 2866 if (ret) { 2867 mlx5_ib_dbg(dev, "Couldn't modify umr QP to rts\n"); 2868 goto error_4; 2869 } 2870 2871 dev->umrc.qp = qp; 2872 dev->umrc.cq = cq; 2873 dev->umrc.pd = pd; 2874 2875 sema_init(&dev->umrc.sem, MAX_UMR_WR); 2876 ret = mlx5_mr_cache_init(dev); 2877 if (ret) { 2878 mlx5_ib_warn(dev, "mr cache init failed %d\n", ret); 2879 goto error_4; 2880 } 2881 2882 kfree(attr); 2883 kfree(init_attr); 2884 2885 return 0; 2886 2887 error_4: 2888 mlx5_ib_destroy_qp(qp, NULL); 2889 dev->umrc.qp = NULL; 2890 2891 error_3: 2892 ib_free_cq(cq); 2893 dev->umrc.cq = NULL; 2894 2895 error_2: 2896 ib_dealloc_pd(pd); 2897 dev->umrc.pd = NULL; 2898 2899 error_0: 2900 kfree(attr); 2901 kfree(init_attr); 2902 return ret; 2903 } 2904 2905 static int create_dev_resources(struct mlx5_ib_resources *devr) 2906 { 2907 struct ib_srq_init_attr attr; 2908 struct mlx5_ib_dev *dev; 2909 struct ib_device *ibdev; 2910 struct ib_cq_init_attr cq_attr = {.cqe = 1}; 2911 int port; 2912 int ret = 0; 2913 2914 dev = container_of(devr, struct mlx5_ib_dev, devr); 2915 ibdev = &dev->ib_dev; 2916 2917 mutex_init(&devr->mutex); 2918 2919 devr->p0 = rdma_zalloc_drv_obj(ibdev, ib_pd); 2920 if (!devr->p0) 2921 return -ENOMEM; 2922 2923 devr->p0->device = ibdev; 2924 devr->p0->uobject = NULL; 2925 atomic_set(&devr->p0->usecnt, 0); 2926 2927 ret = mlx5_ib_alloc_pd(devr->p0, NULL); 2928 if (ret) 2929 goto error0; 2930 2931 devr->c0 = rdma_zalloc_drv_obj(ibdev, ib_cq); 2932 if (!devr->c0) { 2933 ret = -ENOMEM; 2934 goto error1; 2935 } 2936 2937 devr->c0->device = &dev->ib_dev; 2938 atomic_set(&devr->c0->usecnt, 0); 2939 2940 ret = mlx5_ib_create_cq(devr->c0, &cq_attr, NULL); 2941 if (ret) 2942 goto err_create_cq; 2943 2944 devr->x0 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL); 2945 if (IS_ERR(devr->x0)) { 2946 ret = PTR_ERR(devr->x0); 2947 goto error2; 2948 } 2949 devr->x0->device = &dev->ib_dev; 2950 devr->x0->inode = NULL; 2951 atomic_set(&devr->x0->usecnt, 0); 2952 mutex_init(&devr->x0->tgt_qp_mutex); 2953 INIT_LIST_HEAD(&devr->x0->tgt_qp_list); 2954 2955 devr->x1 = mlx5_ib_alloc_xrcd(&dev->ib_dev, NULL); 2956 if (IS_ERR(devr->x1)) { 2957 ret = PTR_ERR(devr->x1); 2958 goto error3; 2959 } 2960 devr->x1->device = &dev->ib_dev; 2961 devr->x1->inode = NULL; 2962 atomic_set(&devr->x1->usecnt, 0); 2963 mutex_init(&devr->x1->tgt_qp_mutex); 2964 INIT_LIST_HEAD(&devr->x1->tgt_qp_list); 2965 2966 memset(&attr, 0, sizeof(attr)); 2967 attr.attr.max_sge = 1; 2968 attr.attr.max_wr = 1; 2969 attr.srq_type = IB_SRQT_XRC; 2970 attr.ext.cq = devr->c0; 2971 attr.ext.xrc.xrcd = devr->x0; 2972 2973 devr->s0 = rdma_zalloc_drv_obj(ibdev, ib_srq); 2974 if (!devr->s0) { 2975 ret = -ENOMEM; 2976 goto error4; 2977 } 2978 2979 devr->s0->device = &dev->ib_dev; 2980 devr->s0->pd = devr->p0; 2981 devr->s0->srq_type = IB_SRQT_XRC; 2982 devr->s0->ext.xrc.xrcd = devr->x0; 2983 devr->s0->ext.cq = devr->c0; 2984 ret = mlx5_ib_create_srq(devr->s0, &attr, NULL); 2985 if (ret) 2986 goto err_create; 2987 2988 atomic_inc(&devr->s0->ext.xrc.xrcd->usecnt); 2989 atomic_inc(&devr->s0->ext.cq->usecnt); 2990 atomic_inc(&devr->p0->usecnt); 2991 atomic_set(&devr->s0->usecnt, 0); 2992 2993 memset(&attr, 0, sizeof(attr)); 2994 attr.attr.max_sge = 1; 2995 attr.attr.max_wr = 1; 2996 attr.srq_type = IB_SRQT_BASIC; 2997 devr->s1 = rdma_zalloc_drv_obj(ibdev, ib_srq); 2998 if (!devr->s1) { 2999 ret = -ENOMEM; 3000 goto error5; 3001 } 3002 3003 devr->s1->device = &dev->ib_dev; 3004 devr->s1->pd = devr->p0; 3005 devr->s1->srq_type = IB_SRQT_BASIC; 3006 devr->s1->ext.cq = devr->c0; 3007 3008 ret = mlx5_ib_create_srq(devr->s1, &attr, NULL); 3009 if (ret) 3010 goto error6; 3011 3012 atomic_inc(&devr->p0->usecnt); 3013 atomic_set(&devr->s1->usecnt, 0); 3014 3015 for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) { 3016 INIT_WORK(&devr->ports[port].pkey_change_work, 3017 pkey_change_handler); 3018 devr->ports[port].devr = devr; 3019 } 3020 3021 return 0; 3022 3023 error6: 3024 kfree(devr->s1); 3025 error5: 3026 mlx5_ib_destroy_srq(devr->s0, NULL); 3027 err_create: 3028 kfree(devr->s0); 3029 error4: 3030 mlx5_ib_dealloc_xrcd(devr->x1, NULL); 3031 error3: 3032 mlx5_ib_dealloc_xrcd(devr->x0, NULL); 3033 error2: 3034 mlx5_ib_destroy_cq(devr->c0, NULL); 3035 err_create_cq: 3036 kfree(devr->c0); 3037 error1: 3038 mlx5_ib_dealloc_pd(devr->p0, NULL); 3039 error0: 3040 kfree(devr->p0); 3041 return ret; 3042 } 3043 3044 static void destroy_dev_resources(struct mlx5_ib_resources *devr) 3045 { 3046 int port; 3047 3048 mlx5_ib_destroy_srq(devr->s1, NULL); 3049 kfree(devr->s1); 3050 mlx5_ib_destroy_srq(devr->s0, NULL); 3051 kfree(devr->s0); 3052 mlx5_ib_dealloc_xrcd(devr->x0, NULL); 3053 mlx5_ib_dealloc_xrcd(devr->x1, NULL); 3054 mlx5_ib_destroy_cq(devr->c0, NULL); 3055 kfree(devr->c0); 3056 mlx5_ib_dealloc_pd(devr->p0, NULL); 3057 kfree(devr->p0); 3058 3059 /* Make sure no change P_Key work items are still executing */ 3060 for (port = 0; port < ARRAY_SIZE(devr->ports); ++port) 3061 cancel_work_sync(&devr->ports[port].pkey_change_work); 3062 } 3063 3064 static u32 get_core_cap_flags(struct ib_device *ibdev) 3065 { 3066 struct mlx5_ib_dev *dev = to_mdev(ibdev); 3067 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, 1); 3068 u8 l3_type_cap = MLX5_CAP_ROCE(dev->mdev, l3_type); 3069 u8 roce_version_cap = MLX5_CAP_ROCE(dev->mdev, roce_version); 3070 u32 ret = 0; 3071 3072 if (ll == IB_LINK_LAYER_INFINIBAND) 3073 return RDMA_CORE_PORT_IBA_IB; 3074 3075 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV4_CAP)) 3076 return 0; 3077 3078 if (!(l3_type_cap & MLX5_ROCE_L3_TYPE_IPV6_CAP)) 3079 return 0; 3080 3081 if (roce_version_cap & MLX5_ROCE_VERSION_1_CAP) 3082 ret |= RDMA_CORE_PORT_IBA_ROCE; 3083 3084 if (roce_version_cap & MLX5_ROCE_VERSION_2_CAP) 3085 ret |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP; 3086 3087 return ret; 3088 } 3089 3090 static int mlx5_port_immutable(struct ib_device *ibdev, u8 port_num, 3091 struct ib_port_immutable *immutable) 3092 { 3093 struct ib_port_attr attr; 3094 struct mlx5_ib_dev *dev = to_mdev(ibdev); 3095 enum rdma_link_layer ll = mlx5_ib_port_link_layer(ibdev, port_num); 3096 int err; 3097 3098 err = mlx5_ib_query_port(ibdev, port_num, &attr); 3099 if (err) 3100 return err; 3101 3102 immutable->pkey_tbl_len = attr.pkey_tbl_len; 3103 immutable->gid_tbl_len = attr.gid_tbl_len; 3104 immutable->core_cap_flags = get_core_cap_flags(ibdev); 3105 if ((ll == IB_LINK_LAYER_INFINIBAND) || MLX5_CAP_GEN(dev->mdev, roce)) 3106 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 3107 3108 return 0; 3109 } 3110 3111 static void get_dev_fw_str(struct ib_device *ibdev, char *str, 3112 size_t str_len) 3113 { 3114 struct mlx5_ib_dev *dev = 3115 container_of(ibdev, struct mlx5_ib_dev, ib_dev); 3116 snprintf(str, str_len, "%d.%d.%04d", fw_rev_maj(dev->mdev), 3117 fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev)); 3118 } 3119 3120 static int mlx5_roce_lag_init(struct mlx5_ib_dev *dev) 3121 { 3122 return 0; 3123 } 3124 3125 static void mlx5_roce_lag_cleanup(struct mlx5_ib_dev *dev) 3126 { 3127 } 3128 3129 static void mlx5_remove_roce_notifier(struct mlx5_ib_dev *dev) 3130 { 3131 if (dev->roce.nb.notifier_call) { 3132 unregister_netdevice_notifier(&dev->roce.nb); 3133 dev->roce.nb.notifier_call = NULL; 3134 } 3135 } 3136 3137 static int 3138 mlx5_enable_roce_if_cb(if_t ifp, void *arg) 3139 { 3140 struct mlx5_ib_dev *dev = arg; 3141 3142 /* check if network interface belongs to mlx5en */ 3143 if (!mlx5_netdev_match(ifp, dev->mdev, "mce")) 3144 return (0); 3145 3146 write_lock(&dev->roce.netdev_lock); 3147 dev->roce.netdev = ifp; 3148 write_unlock(&dev->roce.netdev_lock); 3149 3150 return (0); 3151 } 3152 3153 static int mlx5_enable_roce(struct mlx5_ib_dev *dev) 3154 { 3155 struct epoch_tracker et; 3156 VNET_ITERATOR_DECL(vnet_iter); 3157 int err; 3158 3159 /* Check if mlx5en net device already exists */ 3160 VNET_LIST_RLOCK(); 3161 NET_EPOCH_ENTER(et); 3162 VNET_FOREACH(vnet_iter) { 3163 CURVNET_SET_QUIET(vnet_iter); 3164 if_foreach(mlx5_enable_roce_if_cb, dev); 3165 CURVNET_RESTORE(); 3166 } 3167 NET_EPOCH_EXIT(et); 3168 VNET_LIST_RUNLOCK(); 3169 3170 dev->roce.nb.notifier_call = mlx5_netdev_event; 3171 err = register_netdevice_notifier(&dev->roce.nb); 3172 if (err) { 3173 dev->roce.nb.notifier_call = NULL; 3174 return err; 3175 } 3176 3177 if (MLX5_CAP_GEN(dev->mdev, roce)) { 3178 err = mlx5_nic_vport_enable_roce(dev->mdev); 3179 if (err) 3180 goto err_unregister_netdevice_notifier; 3181 } 3182 3183 err = mlx5_roce_lag_init(dev); 3184 if (err) 3185 goto err_disable_roce; 3186 3187 return 0; 3188 3189 err_disable_roce: 3190 if (MLX5_CAP_GEN(dev->mdev, roce)) 3191 mlx5_nic_vport_disable_roce(dev->mdev); 3192 3193 err_unregister_netdevice_notifier: 3194 mlx5_remove_roce_notifier(dev); 3195 return err; 3196 } 3197 3198 static void mlx5_disable_roce(struct mlx5_ib_dev *dev) 3199 { 3200 mlx5_roce_lag_cleanup(dev); 3201 if (MLX5_CAP_GEN(dev->mdev, roce)) 3202 mlx5_nic_vport_disable_roce(dev->mdev); 3203 } 3204 3205 static void mlx5_ib_dealloc_q_port_counter(struct mlx5_ib_dev *dev, u8 port_num) 3206 { 3207 mlx5_vport_dealloc_q_counter(dev->mdev, 3208 MLX5_INTERFACE_PROTOCOL_IB, 3209 dev->port[port_num].q_cnt_id); 3210 dev->port[port_num].q_cnt_id = 0; 3211 } 3212 3213 static void mlx5_ib_dealloc_q_counters(struct mlx5_ib_dev *dev) 3214 { 3215 unsigned int i; 3216 3217 for (i = 0; i < dev->num_ports; i++) 3218 mlx5_ib_dealloc_q_port_counter(dev, i); 3219 } 3220 3221 static int mlx5_ib_alloc_q_counters(struct mlx5_ib_dev *dev) 3222 { 3223 int i; 3224 int ret; 3225 3226 for (i = 0; i < dev->num_ports; i++) { 3227 ret = mlx5_vport_alloc_q_counter(dev->mdev, 3228 MLX5_INTERFACE_PROTOCOL_IB, 3229 &dev->port[i].q_cnt_id); 3230 if (ret) { 3231 mlx5_ib_warn(dev, 3232 "couldn't allocate queue counter for port %d, err %d\n", 3233 i + 1, ret); 3234 goto dealloc_counters; 3235 } 3236 } 3237 3238 return 0; 3239 3240 dealloc_counters: 3241 while (--i >= 0) 3242 mlx5_ib_dealloc_q_port_counter(dev, i); 3243 3244 return ret; 3245 } 3246 3247 static const char * const names[] = { 3248 "rx_write_requests", 3249 "rx_read_requests", 3250 "rx_atomic_requests", 3251 "out_of_buffer", 3252 "out_of_sequence", 3253 "duplicate_request", 3254 "rnr_nak_retry_err", 3255 "packet_seq_err", 3256 "implied_nak_seq_err", 3257 "local_ack_timeout_err", 3258 }; 3259 3260 static const size_t stats_offsets[] = { 3261 MLX5_BYTE_OFF(query_q_counter_out, rx_write_requests), 3262 MLX5_BYTE_OFF(query_q_counter_out, rx_read_requests), 3263 MLX5_BYTE_OFF(query_q_counter_out, rx_atomic_requests), 3264 MLX5_BYTE_OFF(query_q_counter_out, out_of_buffer), 3265 MLX5_BYTE_OFF(query_q_counter_out, out_of_sequence), 3266 MLX5_BYTE_OFF(query_q_counter_out, duplicate_request), 3267 MLX5_BYTE_OFF(query_q_counter_out, rnr_nak_retry_err), 3268 MLX5_BYTE_OFF(query_q_counter_out, packet_seq_err), 3269 MLX5_BYTE_OFF(query_q_counter_out, implied_nak_seq_err), 3270 MLX5_BYTE_OFF(query_q_counter_out, local_ack_timeout_err), 3271 }; 3272 3273 static struct rdma_hw_stats *mlx5_ib_alloc_hw_stats(struct ib_device *ibdev, 3274 u8 port_num) 3275 { 3276 BUILD_BUG_ON(ARRAY_SIZE(names) != ARRAY_SIZE(stats_offsets)); 3277 3278 /* We support only per port stats */ 3279 if (port_num == 0) 3280 return NULL; 3281 3282 return rdma_alloc_hw_stats_struct(names, ARRAY_SIZE(names), 3283 RDMA_HW_STATS_DEFAULT_LIFESPAN); 3284 } 3285 3286 static int mlx5_ib_get_hw_stats(struct ib_device *ibdev, 3287 struct rdma_hw_stats *stats, 3288 u8 port, int index) 3289 { 3290 struct mlx5_ib_dev *dev = to_mdev(ibdev); 3291 int outlen = MLX5_ST_SZ_BYTES(query_q_counter_out); 3292 void *out; 3293 __be32 val; 3294 int ret; 3295 int i; 3296 3297 if (!port || !stats) 3298 return -ENOSYS; 3299 3300 out = mlx5_vzalloc(outlen); 3301 if (!out) 3302 return -ENOMEM; 3303 3304 ret = mlx5_vport_query_q_counter(dev->mdev, 3305 dev->port[port - 1].q_cnt_id, 0, 3306 out, outlen); 3307 if (ret) 3308 goto free; 3309 3310 for (i = 0; i < ARRAY_SIZE(names); i++) { 3311 val = *(__be32 *)(out + stats_offsets[i]); 3312 stats->value[i] = (u64)be32_to_cpu(val); 3313 } 3314 free: 3315 kvfree(out); 3316 return ARRAY_SIZE(names); 3317 } 3318 3319 static int mlx5_ib_stage_bfreg_init(struct mlx5_ib_dev *dev) 3320 { 3321 int err; 3322 3323 err = mlx5_alloc_bfreg(dev->mdev, &dev->bfreg, false, false); 3324 if (err) 3325 return err; 3326 3327 err = mlx5_alloc_bfreg(dev->mdev, &dev->fp_bfreg, false, true); 3328 if (err) { 3329 mlx5_free_bfreg(dev->mdev, &dev->bfreg); 3330 return err; 3331 } 3332 3333 err = mlx5_alloc_bfreg(dev->mdev, &dev->wc_bfreg, true, false); 3334 if (err) { 3335 mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg); 3336 mlx5_free_bfreg(dev->mdev, &dev->bfreg); 3337 } 3338 3339 return err; 3340 } 3341 3342 static void mlx5_ib_stage_bfreg_cleanup(struct mlx5_ib_dev *dev) 3343 { 3344 mlx5_free_bfreg(dev->mdev, &dev->wc_bfreg); 3345 mlx5_free_bfreg(dev->mdev, &dev->fp_bfreg); 3346 mlx5_free_bfreg(dev->mdev, &dev->bfreg); 3347 } 3348 3349 static void *mlx5_ib_add(struct mlx5_core_dev *mdev) 3350 { 3351 struct mlx5_ib_dev *dev; 3352 enum rdma_link_layer ll; 3353 int port_type_cap; 3354 int err; 3355 int i; 3356 3357 port_type_cap = MLX5_CAP_GEN(mdev, port_type); 3358 ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap); 3359 3360 dev = (struct mlx5_ib_dev *)ib_alloc_device(sizeof(*dev)); 3361 if (!dev) 3362 return NULL; 3363 3364 dev->mdev = mdev; 3365 3366 dev->port = kcalloc(MLX5_CAP_GEN(mdev, num_ports), sizeof(*dev->port), 3367 GFP_KERNEL); 3368 if (!dev->port) 3369 goto err_dealloc; 3370 3371 rwlock_init(&dev->roce.netdev_lock); 3372 err = get_port_caps(dev); 3373 if (err) 3374 goto err_free_port; 3375 3376 if (mlx5_use_mad_ifc(dev)) 3377 get_ext_port_caps(dev); 3378 3379 MLX5_INIT_DOORBELL_LOCK(&dev->uar_lock); 3380 3381 mutex_init(&dev->lb_mutex); 3382 3383 INIT_IB_DEVICE_OPS(&dev->ib_dev.ops, mlx5, MLX5); 3384 snprintf(dev->ib_dev.name, IB_DEVICE_NAME_MAX, "mlx5_%d", device_get_unit(mdev->pdev->dev.bsddev)); 3385 dev->ib_dev.owner = THIS_MODULE; 3386 dev->ib_dev.node_type = RDMA_NODE_IB_CA; 3387 dev->ib_dev.local_dma_lkey = 0 /* not supported for now */; 3388 dev->num_ports = MLX5_CAP_GEN(mdev, num_ports); 3389 dev->ib_dev.phys_port_cnt = dev->num_ports; 3390 dev->ib_dev.num_comp_vectors = 3391 dev->mdev->priv.eq_table.num_comp_vectors; 3392 dev->ib_dev.dma_device = &mdev->pdev->dev; 3393 3394 dev->ib_dev.uverbs_abi_ver = MLX5_IB_UVERBS_ABI_VERSION; 3395 dev->ib_dev.uverbs_cmd_mask = 3396 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) | 3397 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) | 3398 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) | 3399 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) | 3400 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) | 3401 (1ull << IB_USER_VERBS_CMD_CREATE_AH) | 3402 (1ull << IB_USER_VERBS_CMD_DESTROY_AH) | 3403 (1ull << IB_USER_VERBS_CMD_REG_MR) | 3404 (1ull << IB_USER_VERBS_CMD_REREG_MR) | 3405 (1ull << IB_USER_VERBS_CMD_DEREG_MR) | 3406 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) | 3407 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) | 3408 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) | 3409 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) | 3410 (1ull << IB_USER_VERBS_CMD_CREATE_QP) | 3411 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) | 3412 (1ull << IB_USER_VERBS_CMD_QUERY_QP) | 3413 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) | 3414 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) | 3415 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) | 3416 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) | 3417 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) | 3418 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) | 3419 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) | 3420 (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ) | 3421 (1ull << IB_USER_VERBS_CMD_OPEN_QP); 3422 dev->ib_dev.uverbs_ex_cmd_mask = 3423 (1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE) | 3424 (1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ) | 3425 (1ull << IB_USER_VERBS_EX_CMD_CREATE_QP); 3426 3427 dev->ib_dev.query_device = mlx5_ib_query_device; 3428 dev->ib_dev.query_port = mlx5_ib_query_port; 3429 dev->ib_dev.get_link_layer = mlx5_ib_port_link_layer; 3430 if (ll == IB_LINK_LAYER_ETHERNET) 3431 dev->ib_dev.get_netdev = mlx5_ib_get_netdev; 3432 dev->ib_dev.query_gid = mlx5_ib_query_gid; 3433 dev->ib_dev.add_gid = mlx5_ib_add_gid; 3434 dev->ib_dev.del_gid = mlx5_ib_del_gid; 3435 dev->ib_dev.query_pkey = mlx5_ib_query_pkey; 3436 dev->ib_dev.modify_device = mlx5_ib_modify_device; 3437 dev->ib_dev.modify_port = mlx5_ib_modify_port; 3438 dev->ib_dev.alloc_ucontext = mlx5_ib_alloc_ucontext; 3439 dev->ib_dev.dealloc_ucontext = mlx5_ib_dealloc_ucontext; 3440 dev->ib_dev.mmap = mlx5_ib_mmap; 3441 dev->ib_dev.mmap_free = mlx5_ib_mmap_free; 3442 dev->ib_dev.alloc_pd = mlx5_ib_alloc_pd; 3443 dev->ib_dev.dealloc_pd = mlx5_ib_dealloc_pd; 3444 dev->ib_dev.create_ah = mlx5_ib_create_ah; 3445 dev->ib_dev.query_ah = mlx5_ib_query_ah; 3446 dev->ib_dev.destroy_ah = mlx5_ib_destroy_ah; 3447 dev->ib_dev.create_srq = mlx5_ib_create_srq; 3448 dev->ib_dev.modify_srq = mlx5_ib_modify_srq; 3449 dev->ib_dev.query_srq = mlx5_ib_query_srq; 3450 dev->ib_dev.destroy_srq = mlx5_ib_destroy_srq; 3451 dev->ib_dev.post_srq_recv = mlx5_ib_post_srq_recv; 3452 dev->ib_dev.create_qp = mlx5_ib_create_qp; 3453 dev->ib_dev.modify_qp = mlx5_ib_modify_qp; 3454 dev->ib_dev.query_qp = mlx5_ib_query_qp; 3455 dev->ib_dev.destroy_qp = mlx5_ib_destroy_qp; 3456 dev->ib_dev.post_send = mlx5_ib_post_send; 3457 dev->ib_dev.post_recv = mlx5_ib_post_recv; 3458 dev->ib_dev.create_cq = mlx5_ib_create_cq; 3459 dev->ib_dev.modify_cq = mlx5_ib_modify_cq; 3460 dev->ib_dev.resize_cq = mlx5_ib_resize_cq; 3461 dev->ib_dev.destroy_cq = mlx5_ib_destroy_cq; 3462 dev->ib_dev.poll_cq = mlx5_ib_poll_cq; 3463 dev->ib_dev.req_notify_cq = mlx5_ib_arm_cq; 3464 dev->ib_dev.get_dma_mr = mlx5_ib_get_dma_mr; 3465 dev->ib_dev.reg_user_mr = mlx5_ib_reg_user_mr; 3466 dev->ib_dev.rereg_user_mr = mlx5_ib_rereg_user_mr; 3467 dev->ib_dev.dereg_mr = mlx5_ib_dereg_mr; 3468 dev->ib_dev.attach_mcast = mlx5_ib_mcg_attach; 3469 dev->ib_dev.detach_mcast = mlx5_ib_mcg_detach; 3470 dev->ib_dev.process_mad = mlx5_ib_process_mad; 3471 dev->ib_dev.alloc_mr = mlx5_ib_alloc_mr; 3472 dev->ib_dev.map_mr_sg = mlx5_ib_map_mr_sg; 3473 dev->ib_dev.check_mr_status = mlx5_ib_check_mr_status; 3474 dev->ib_dev.get_port_immutable = mlx5_port_immutable; 3475 dev->ib_dev.get_dev_fw_str = get_dev_fw_str; 3476 if (mlx5_core_is_pf(mdev)) { 3477 dev->ib_dev.get_vf_config = mlx5_ib_get_vf_config; 3478 dev->ib_dev.set_vf_link_state = mlx5_ib_set_vf_link_state; 3479 dev->ib_dev.get_vf_stats = mlx5_ib_get_vf_stats; 3480 dev->ib_dev.set_vf_guid = mlx5_ib_set_vf_guid; 3481 } 3482 3483 dev->ib_dev.disassociate_ucontext = mlx5_ib_disassociate_ucontext; 3484 3485 mlx5_ib_internal_fill_odp_caps(dev); 3486 3487 if (MLX5_CAP_GEN(mdev, imaicl)) { 3488 dev->ib_dev.alloc_mw = mlx5_ib_alloc_mw; 3489 dev->ib_dev.dealloc_mw = mlx5_ib_dealloc_mw; 3490 dev->ib_dev.uverbs_cmd_mask |= 3491 (1ull << IB_USER_VERBS_CMD_ALLOC_MW) | 3492 (1ull << IB_USER_VERBS_CMD_DEALLOC_MW); 3493 } 3494 3495 if (MLX5_CAP_GEN(dev->mdev, out_of_seq_cnt) && 3496 MLX5_CAP_GEN(dev->mdev, retransmission_q_counters)) { 3497 dev->ib_dev.get_hw_stats = mlx5_ib_get_hw_stats; 3498 dev->ib_dev.alloc_hw_stats = mlx5_ib_alloc_hw_stats; 3499 } 3500 3501 if (MLX5_CAP_GEN(mdev, xrc)) { 3502 dev->ib_dev.alloc_xrcd = mlx5_ib_alloc_xrcd; 3503 dev->ib_dev.dealloc_xrcd = mlx5_ib_dealloc_xrcd; 3504 dev->ib_dev.uverbs_cmd_mask |= 3505 (1ull << IB_USER_VERBS_CMD_OPEN_XRCD) | 3506 (1ull << IB_USER_VERBS_CMD_CLOSE_XRCD); 3507 } 3508 3509 if (mlx5_ib_port_link_layer(&dev->ib_dev, 1) == 3510 IB_LINK_LAYER_ETHERNET) { 3511 dev->ib_dev.create_flow = mlx5_ib_create_flow; 3512 dev->ib_dev.destroy_flow = mlx5_ib_destroy_flow; 3513 dev->ib_dev.create_wq = mlx5_ib_create_wq; 3514 dev->ib_dev.modify_wq = mlx5_ib_modify_wq; 3515 dev->ib_dev.destroy_wq = mlx5_ib_destroy_wq; 3516 dev->ib_dev.create_rwq_ind_table = mlx5_ib_create_rwq_ind_table; 3517 dev->ib_dev.destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table; 3518 dev->ib_dev.uverbs_ex_cmd_mask |= 3519 (1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) | 3520 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW) | 3521 (1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) | 3522 (1ull << IB_USER_VERBS_EX_CMD_MODIFY_WQ) | 3523 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) | 3524 (1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) | 3525 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL); 3526 } 3527 err = init_node_data(dev); 3528 if (err) 3529 goto err_free_port; 3530 3531 mutex_init(&dev->flow_db.lock); 3532 mutex_init(&dev->cap_mask_mutex); 3533 INIT_LIST_HEAD(&dev->qp_list); 3534 spin_lock_init(&dev->reset_flow_resource_lock); 3535 3536 if (ll == IB_LINK_LAYER_ETHERNET) { 3537 err = mlx5_enable_roce(dev); 3538 if (err) 3539 goto err_free_port; 3540 } 3541 3542 err = create_dev_resources(&dev->devr); 3543 if (err) 3544 goto err_disable_roce; 3545 3546 err = mlx5_ib_odp_init_one(dev); 3547 if (err) 3548 goto err_rsrc; 3549 3550 err = mlx5_ib_alloc_q_counters(dev); 3551 if (err) 3552 goto err_odp; 3553 3554 err = mlx5_ib_stage_bfreg_init(dev); 3555 if (err) 3556 goto err_q_cnt; 3557 3558 err = ib_register_device(&dev->ib_dev, NULL); 3559 if (err) 3560 goto err_bfreg; 3561 3562 err = create_umr_res(dev); 3563 if (err) 3564 goto err_dev; 3565 3566 for (i = 0; i < ARRAY_SIZE(mlx5_class_attributes); i++) { 3567 err = device_create_file(&dev->ib_dev.dev, 3568 mlx5_class_attributes[i]); 3569 if (err) 3570 goto err_umrc; 3571 } 3572 3573 err = mlx5_ib_init_congestion(dev); 3574 if (err) 3575 goto err_umrc; 3576 3577 dev->ib_active = true; 3578 3579 return dev; 3580 3581 err_umrc: 3582 destroy_umrc_res(dev); 3583 3584 err_dev: 3585 ib_unregister_device(&dev->ib_dev); 3586 3587 err_bfreg: 3588 mlx5_ib_stage_bfreg_cleanup(dev); 3589 3590 err_q_cnt: 3591 mlx5_ib_dealloc_q_counters(dev); 3592 3593 err_odp: 3594 mlx5_ib_odp_remove_one(dev); 3595 3596 err_rsrc: 3597 destroy_dev_resources(&dev->devr); 3598 3599 err_disable_roce: 3600 if (ll == IB_LINK_LAYER_ETHERNET) { 3601 mlx5_disable_roce(dev); 3602 mlx5_remove_roce_notifier(dev); 3603 } 3604 3605 err_free_port: 3606 kfree(dev->port); 3607 3608 err_dealloc: 3609 ib_dealloc_device((struct ib_device *)dev); 3610 3611 return NULL; 3612 } 3613 3614 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context) 3615 { 3616 struct mlx5_ib_dev *dev = context; 3617 enum rdma_link_layer ll = mlx5_ib_port_link_layer(&dev->ib_dev, 1); 3618 3619 mlx5_ib_cleanup_congestion(dev); 3620 mlx5_remove_roce_notifier(dev); 3621 ib_unregister_device(&dev->ib_dev); 3622 mlx5_ib_stage_bfreg_cleanup(dev); 3623 mlx5_ib_dealloc_q_counters(dev); 3624 destroy_umrc_res(dev); 3625 mlx5_ib_odp_remove_one(dev); 3626 destroy_dev_resources(&dev->devr); 3627 if (ll == IB_LINK_LAYER_ETHERNET) 3628 mlx5_disable_roce(dev); 3629 kfree(dev->port); 3630 ib_dealloc_device(&dev->ib_dev); 3631 } 3632 3633 static struct mlx5_interface mlx5_ib_interface = { 3634 .add = mlx5_ib_add, 3635 .remove = mlx5_ib_remove, 3636 .event = mlx5_ib_event, 3637 .protocol = MLX5_INTERFACE_PROTOCOL_IB, 3638 }; 3639 3640 static int __init mlx5_ib_init(void) 3641 { 3642 int err; 3643 3644 err = mlx5_ib_odp_init(); 3645 if (err) 3646 return err; 3647 3648 err = mlx5_register_interface(&mlx5_ib_interface); 3649 if (err) 3650 goto clean_odp; 3651 3652 return err; 3653 3654 clean_odp: 3655 mlx5_ib_odp_cleanup(); 3656 return err; 3657 } 3658 3659 static void __exit mlx5_ib_cleanup(void) 3660 { 3661 mlx5_unregister_interface(&mlx5_ib_interface); 3662 mlx5_ib_odp_cleanup(); 3663 } 3664 3665 module_init_order(mlx5_ib_init, SI_ORDER_SEVENTH); 3666 module_exit_order(mlx5_ib_cleanup, SI_ORDER_SEVENTH); 3667