1/* 2 * Copyright (c) 2004 Mellanox Technologies Ltd. All rights reserved. 3 * Copyright (c) 2004 Infinicon Corporation. All rights reserved. 4 * Copyright (c) 2004 Intel Corporation. All rights reserved. 5 * Copyright (c) 2004 Topspin Corporation. All rights reserved. 6 * Copyright (c) 2004 Voltaire Corporation. All rights reserved. 7 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved. 8 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved. 9 * 10 * This software is available to you under a choice of one of two 11 * licenses. You may choose to be licensed under the terms of the GNU 12 * General Public License (GPL) Version 2, available from the file 13 * COPYING in the main directory of this source tree, or the 14 * OpenIB.org BSD license below: 15 * 16 * Redistribution and use in source and binary forms, with or 17 * without modification, are permitted provided that the following 18 * conditions are met: 19 * 20 * - Redistributions of source code must retain the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer. 23 * 24 * - Redistributions in binary form must reproduce the above 25 * copyright notice, this list of conditions and the following 26 * disclaimer in the documentation and/or other materials 27 * provided with the distribution. 28 * 29 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 30 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 31 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 32 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 33 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 34 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 35 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 36 * SOFTWARE. 37 */ 38 39#include <linux/errno.h> 40#include <linux/err.h> 41#include <linux/export.h> 42#include <linux/string.h> 43#include <linux/slab.h> 44#include <linux/in.h> 45#include <linux/in6.h> 46#include <net/addrconf.h> 47#include <linux/security.h> 48 49#include <rdma/ib_verbs.h> 50#include <rdma/ib_cache.h> 51#include <rdma/ib_addr.h> 52#include <rdma/rw.h> 53#include <rdma/lag.h> 54 55#include "core_priv.h" 56#include <trace/events/rdma_core.h> 57 58static int ib_resolve_eth_dmac(struct ib_device *device, 59 struct rdma_ah_attr *ah_attr); 60 61static const char * const ib_events[] = { 62 [IB_EVENT_CQ_ERR] = "CQ error", 63 [IB_EVENT_QP_FATAL] = "QP fatal error", 64 [IB_EVENT_QP_REQ_ERR] = "QP request error", 65 [IB_EVENT_QP_ACCESS_ERR] = "QP access error", 66 [IB_EVENT_COMM_EST] = "communication established", 67 [IB_EVENT_SQ_DRAINED] = "send queue drained", 68 [IB_EVENT_PATH_MIG] = "path migration successful", 69 [IB_EVENT_PATH_MIG_ERR] = "path migration error", 70 [IB_EVENT_DEVICE_FATAL] = "device fatal error", 71 [IB_EVENT_PORT_ACTIVE] = "port active", 72 [IB_EVENT_PORT_ERR] = "port error", 73 [IB_EVENT_LID_CHANGE] = "LID change", 74 [IB_EVENT_PKEY_CHANGE] = "P_key change", 75 [IB_EVENT_SM_CHANGE] = "SM change", 76 [IB_EVENT_SRQ_ERR] = "SRQ error", 77 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached", 78 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached", 79 [IB_EVENT_CLIENT_REREGISTER] = "client reregister", 80 [IB_EVENT_GID_CHANGE] = "GID changed", 81}; 82 83const char *__attribute_const__ ib_event_msg(enum ib_event_type event) 84{ 85 size_t index = event; 86 87 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ? 88 ib_events[index] : "unrecognized event"; 89} 90EXPORT_SYMBOL(ib_event_msg); 91 92static const char * const wc_statuses[] = { 93 [IB_WC_SUCCESS] = "success", 94 [IB_WC_LOC_LEN_ERR] = "local length error", 95 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error", 96 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error", 97 [IB_WC_LOC_PROT_ERR] = "local protection error", 98 [IB_WC_WR_FLUSH_ERR] = "WR flushed", 99 [IB_WC_MW_BIND_ERR] = "memory management operation error", 100 [IB_WC_BAD_RESP_ERR] = "bad response error", 101 [IB_WC_LOC_ACCESS_ERR] = "local access error", 102 [IB_WC_REM_INV_REQ_ERR] = "invalid request error", 103 [IB_WC_REM_ACCESS_ERR] = "remote access error", 104 [IB_WC_REM_OP_ERR] = "remote operation error", 105 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded", 106 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded", 107 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error", 108 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request", 109 [IB_WC_REM_ABORT_ERR] = "operation aborted", 110 [IB_WC_INV_EECN_ERR] = "invalid EE context number", 111 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state", 112 [IB_WC_FATAL_ERR] = "fatal error", 113 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error", 114 [IB_WC_GENERAL_ERR] = "general error", 115}; 116 117const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status) 118{ 119 size_t index = status; 120 121 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ? 122 wc_statuses[index] : "unrecognized status"; 123} 124EXPORT_SYMBOL(ib_wc_status_msg); 125 126__attribute_const__ int ib_rate_to_mult(enum ib_rate rate) 127{ 128 switch (rate) { 129 case IB_RATE_2_5_GBPS: return 1; 130 case IB_RATE_5_GBPS: return 2; 131 case IB_RATE_10_GBPS: return 4; 132 case IB_RATE_20_GBPS: return 8; 133 case IB_RATE_30_GBPS: return 12; 134 case IB_RATE_40_GBPS: return 16; 135 case IB_RATE_60_GBPS: return 24; 136 case IB_RATE_80_GBPS: return 32; 137 case IB_RATE_120_GBPS: return 48; 138 case IB_RATE_14_GBPS: return 6; 139 case IB_RATE_56_GBPS: return 22; 140 case IB_RATE_112_GBPS: return 45; 141 case IB_RATE_168_GBPS: return 67; 142 case IB_RATE_25_GBPS: return 10; 143 case IB_RATE_100_GBPS: return 40; 144 case IB_RATE_200_GBPS: return 80; 145 case IB_RATE_300_GBPS: return 120; 146 case IB_RATE_28_GBPS: return 11; 147 case IB_RATE_50_GBPS: return 20; 148 case IB_RATE_400_GBPS: return 160; 149 case IB_RATE_600_GBPS: return 240; 150 default: return -1; 151 } 152} 153EXPORT_SYMBOL(ib_rate_to_mult); 154 155__attribute_const__ enum ib_rate mult_to_ib_rate(int mult) 156{ 157 switch (mult) { 158 case 1: return IB_RATE_2_5_GBPS; 159 case 2: return IB_RATE_5_GBPS; 160 case 4: return IB_RATE_10_GBPS; 161 case 8: return IB_RATE_20_GBPS; 162 case 12: return IB_RATE_30_GBPS; 163 case 16: return IB_RATE_40_GBPS; 164 case 24: return IB_RATE_60_GBPS; 165 case 32: return IB_RATE_80_GBPS; 166 case 48: return IB_RATE_120_GBPS; 167 case 6: return IB_RATE_14_GBPS; 168 case 22: return IB_RATE_56_GBPS; 169 case 45: return IB_RATE_112_GBPS; 170 case 67: return IB_RATE_168_GBPS; 171 case 10: return IB_RATE_25_GBPS; 172 case 40: return IB_RATE_100_GBPS; 173 case 80: return IB_RATE_200_GBPS; 174 case 120: return IB_RATE_300_GBPS; 175 case 11: return IB_RATE_28_GBPS; 176 case 20: return IB_RATE_50_GBPS; 177 case 160: return IB_RATE_400_GBPS; 178 case 240: return IB_RATE_600_GBPS; 179 default: return IB_RATE_PORT_CURRENT; 180 } 181} 182EXPORT_SYMBOL(mult_to_ib_rate); 183 184__attribute_const__ int ib_rate_to_mbps(enum ib_rate rate) 185{ 186 switch (rate) { 187 case IB_RATE_2_5_GBPS: return 2500; 188 case IB_RATE_5_GBPS: return 5000; 189 case IB_RATE_10_GBPS: return 10000; 190 case IB_RATE_20_GBPS: return 20000; 191 case IB_RATE_30_GBPS: return 30000; 192 case IB_RATE_40_GBPS: return 40000; 193 case IB_RATE_60_GBPS: return 60000; 194 case IB_RATE_80_GBPS: return 80000; 195 case IB_RATE_120_GBPS: return 120000; 196 case IB_RATE_14_GBPS: return 14062; 197 case IB_RATE_56_GBPS: return 56250; 198 case IB_RATE_112_GBPS: return 112500; 199 case IB_RATE_168_GBPS: return 168750; 200 case IB_RATE_25_GBPS: return 25781; 201 case IB_RATE_100_GBPS: return 103125; 202 case IB_RATE_200_GBPS: return 206250; 203 case IB_RATE_300_GBPS: return 309375; 204 case IB_RATE_28_GBPS: return 28125; 205 case IB_RATE_50_GBPS: return 53125; 206 case IB_RATE_400_GBPS: return 425000; 207 case IB_RATE_600_GBPS: return 637500; 208 default: return -1; 209 } 210} 211EXPORT_SYMBOL(ib_rate_to_mbps); 212 213__attribute_const__ enum rdma_transport_type 214rdma_node_get_transport(unsigned int node_type) 215{ 216 217 if (node_type == RDMA_NODE_USNIC) 218 return RDMA_TRANSPORT_USNIC; 219 if (node_type == RDMA_NODE_USNIC_UDP) 220 return RDMA_TRANSPORT_USNIC_UDP; 221 if (node_type == RDMA_NODE_RNIC) 222 return RDMA_TRANSPORT_IWARP; 223 if (node_type == RDMA_NODE_UNSPECIFIED) 224 return RDMA_TRANSPORT_UNSPECIFIED; 225 226 return RDMA_TRANSPORT_IB; 227} 228EXPORT_SYMBOL(rdma_node_get_transport); 229 230enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num) 231{ 232 enum rdma_transport_type lt; 233 if (device->ops.get_link_layer) 234 return device->ops.get_link_layer(device, port_num); 235 236 lt = rdma_node_get_transport(device->node_type); 237 if (lt == RDMA_TRANSPORT_IB) 238 return IB_LINK_LAYER_INFINIBAND; 239 240 return IB_LINK_LAYER_ETHERNET; 241} 242EXPORT_SYMBOL(rdma_port_get_link_layer); 243 244/* Protection domains */ 245 246/** 247 * ib_alloc_pd - Allocates an unused protection domain. 248 * @device: The device on which to allocate the protection domain. 249 * @flags: protection domain flags 250 * @caller: caller's build-time module name 251 * 252 * A protection domain object provides an association between QPs, shared 253 * receive queues, address handles, memory regions, and memory windows. 254 * 255 * Every PD has a local_dma_lkey which can be used as the lkey value for local 256 * memory operations. 257 */ 258struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags, 259 const char *caller) 260{ 261 struct ib_pd *pd; 262 int mr_access_flags = 0; 263 int ret; 264 265 pd = rdma_zalloc_drv_obj(device, ib_pd); 266 if (!pd) 267 return ERR_PTR(-ENOMEM); 268 269 pd->device = device; 270 pd->uobject = NULL; 271 pd->__internal_mr = NULL; 272 atomic_set(&pd->usecnt, 0); 273 pd->flags = flags; 274 275 rdma_restrack_new(&pd->res, RDMA_RESTRACK_PD); 276 rdma_restrack_set_name(&pd->res, caller); 277 278 ret = device->ops.alloc_pd(pd, NULL); 279 if (ret) { 280 rdma_restrack_put(&pd->res); 281 kfree(pd); 282 return ERR_PTR(ret); 283 } 284 rdma_restrack_add(&pd->res); 285 286 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY) 287 pd->local_dma_lkey = device->local_dma_lkey; 288 else 289 mr_access_flags |= IB_ACCESS_LOCAL_WRITE; 290 291 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) { 292 pr_warn("%s: enabling unsafe global rkey\n", caller); 293 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE; 294 } 295 296 if (mr_access_flags) { 297 struct ib_mr *mr; 298 299 mr = pd->device->ops.get_dma_mr(pd, mr_access_flags); 300 if (IS_ERR(mr)) { 301 ib_dealloc_pd(pd); 302 return ERR_CAST(mr); 303 } 304 305 mr->device = pd->device; 306 mr->pd = pd; 307 mr->type = IB_MR_TYPE_DMA; 308 mr->uobject = NULL; 309 mr->need_inval = false; 310 311 pd->__internal_mr = mr; 312 313 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) 314 pd->local_dma_lkey = pd->__internal_mr->lkey; 315 316 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) 317 pd->unsafe_global_rkey = pd->__internal_mr->rkey; 318 } 319 320 return pd; 321} 322EXPORT_SYMBOL(__ib_alloc_pd); 323 324/** 325 * ib_dealloc_pd_user - Deallocates a protection domain. 326 * @pd: The protection domain to deallocate. 327 * @udata: Valid user data or NULL for kernel object 328 * 329 * It is an error to call this function while any resources in the pd still 330 * exist. The caller is responsible to synchronously destroy them and 331 * guarantee no new allocations will happen. 332 */ 333int ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata) 334{ 335 int ret; 336 337 if (pd->__internal_mr) { 338 ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL); 339 WARN_ON(ret); 340 pd->__internal_mr = NULL; 341 } 342 343 /* uverbs manipulates usecnt with proper locking, while the kabi 344 requires the caller to guarantee we can't race here. */ 345 WARN_ON(atomic_read(&pd->usecnt)); 346 347 ret = pd->device->ops.dealloc_pd(pd, udata); 348 if (ret) 349 return ret; 350 351 rdma_restrack_del(&pd->res); 352 kfree(pd); 353 return ret; 354} 355EXPORT_SYMBOL(ib_dealloc_pd_user); 356 357/* Address handles */ 358 359/** 360 * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination. 361 * @dest: Pointer to destination ah_attr. Contents of the destination 362 * pointer is assumed to be invalid and attribute are overwritten. 363 * @src: Pointer to source ah_attr. 364 */ 365void rdma_copy_ah_attr(struct rdma_ah_attr *dest, 366 const struct rdma_ah_attr *src) 367{ 368 *dest = *src; 369 if (dest->grh.sgid_attr) 370 rdma_hold_gid_attr(dest->grh.sgid_attr); 371} 372EXPORT_SYMBOL(rdma_copy_ah_attr); 373 374/** 375 * rdma_replace_ah_attr - Replace valid ah_attr with new new one. 376 * @old: Pointer to existing ah_attr which needs to be replaced. 377 * old is assumed to be valid or zero'd 378 * @new: Pointer to the new ah_attr. 379 * 380 * rdma_replace_ah_attr() first releases any reference in the old ah_attr if 381 * old the ah_attr is valid; after that it copies the new attribute and holds 382 * the reference to the replaced ah_attr. 383 */ 384void rdma_replace_ah_attr(struct rdma_ah_attr *old, 385 const struct rdma_ah_attr *new) 386{ 387 rdma_destroy_ah_attr(old); 388 *old = *new; 389 if (old->grh.sgid_attr) 390 rdma_hold_gid_attr(old->grh.sgid_attr); 391} 392EXPORT_SYMBOL(rdma_replace_ah_attr); 393 394/** 395 * rdma_move_ah_attr - Move ah_attr pointed by source to destination. 396 * @dest: Pointer to destination ah_attr to copy to. 397 * dest is assumed to be valid or zero'd 398 * @src: Pointer to the new ah_attr. 399 * 400 * rdma_move_ah_attr() first releases any reference in the destination ah_attr 401 * if it is valid. This also transfers ownership of internal references from 402 * src to dest, making src invalid in the process. No new reference of the src 403 * ah_attr is taken. 404 */ 405void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src) 406{ 407 rdma_destroy_ah_attr(dest); 408 *dest = *src; 409 src->grh.sgid_attr = NULL; 410} 411EXPORT_SYMBOL(rdma_move_ah_attr); 412 413/* 414 * Validate that the rdma_ah_attr is valid for the device before passing it 415 * off to the driver. 416 */ 417static int rdma_check_ah_attr(struct ib_device *device, 418 struct rdma_ah_attr *ah_attr) 419{ 420 if (!rdma_is_port_valid(device, ah_attr->port_num)) 421 return -EINVAL; 422 423 if ((rdma_is_grh_required(device, ah_attr->port_num) || 424 ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) && 425 !(ah_attr->ah_flags & IB_AH_GRH)) 426 return -EINVAL; 427 428 if (ah_attr->grh.sgid_attr) { 429 /* 430 * Make sure the passed sgid_attr is consistent with the 431 * parameters 432 */ 433 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index || 434 ah_attr->grh.sgid_attr->port_num != ah_attr->port_num) 435 return -EINVAL; 436 } 437 return 0; 438} 439 440/* 441 * If the ah requires a GRH then ensure that sgid_attr pointer is filled in. 442 * On success the caller is responsible to call rdma_unfill_sgid_attr(). 443 */ 444static int rdma_fill_sgid_attr(struct ib_device *device, 445 struct rdma_ah_attr *ah_attr, 446 const struct ib_gid_attr **old_sgid_attr) 447{ 448 const struct ib_gid_attr *sgid_attr; 449 struct ib_global_route *grh; 450 int ret; 451 452 *old_sgid_attr = ah_attr->grh.sgid_attr; 453 454 ret = rdma_check_ah_attr(device, ah_attr); 455 if (ret) 456 return ret; 457 458 if (!(ah_attr->ah_flags & IB_AH_GRH)) 459 return 0; 460 461 grh = rdma_ah_retrieve_grh(ah_attr); 462 if (grh->sgid_attr) 463 return 0; 464 465 sgid_attr = 466 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index); 467 if (IS_ERR(sgid_attr)) 468 return PTR_ERR(sgid_attr); 469 470 /* Move ownerhip of the kref into the ah_attr */ 471 grh->sgid_attr = sgid_attr; 472 return 0; 473} 474 475static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr, 476 const struct ib_gid_attr *old_sgid_attr) 477{ 478 /* 479 * Fill didn't change anything, the caller retains ownership of 480 * whatever it passed 481 */ 482 if (ah_attr->grh.sgid_attr == old_sgid_attr) 483 return; 484 485 /* 486 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller 487 * doesn't see any change in the rdma_ah_attr. If we get here 488 * old_sgid_attr is NULL. 489 */ 490 rdma_destroy_ah_attr(ah_attr); 491} 492 493static const struct ib_gid_attr * 494rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr, 495 const struct ib_gid_attr *old_attr) 496{ 497 if (old_attr) 498 rdma_put_gid_attr(old_attr); 499 if (ah_attr->ah_flags & IB_AH_GRH) { 500 rdma_hold_gid_attr(ah_attr->grh.sgid_attr); 501 return ah_attr->grh.sgid_attr; 502 } 503 return NULL; 504} 505 506static struct ib_ah *_rdma_create_ah(struct ib_pd *pd, 507 struct rdma_ah_attr *ah_attr, 508 u32 flags, 509 struct ib_udata *udata, 510 struct net_device *xmit_slave) 511{ 512 struct rdma_ah_init_attr init_attr = {}; 513 struct ib_device *device = pd->device; 514 struct ib_ah *ah; 515 int ret; 516 517 might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE); 518 519 if (!device->ops.create_ah) 520 return ERR_PTR(-EOPNOTSUPP); 521 522 ah = rdma_zalloc_drv_obj_gfp( 523 device, ib_ah, 524 (flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC); 525 if (!ah) 526 return ERR_PTR(-ENOMEM); 527 528 ah->device = device; 529 ah->pd = pd; 530 ah->type = ah_attr->type; 531 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL); 532 init_attr.ah_attr = ah_attr; 533 init_attr.flags = flags; 534 init_attr.xmit_slave = xmit_slave; 535 536 ret = device->ops.create_ah(ah, &init_attr, udata); 537 if (ret) { 538 if (ah->sgid_attr) 539 rdma_put_gid_attr(ah->sgid_attr); 540 kfree(ah); 541 return ERR_PTR(ret); 542 } 543 544 atomic_inc(&pd->usecnt); 545 return ah; 546} 547 548/** 549 * rdma_create_ah - Creates an address handle for the 550 * given address vector. 551 * @pd: The protection domain associated with the address handle. 552 * @ah_attr: The attributes of the address vector. 553 * @flags: Create address handle flags (see enum rdma_create_ah_flags). 554 * 555 * It returns 0 on success and returns appropriate error code on error. 556 * The address handle is used to reference a local or global destination 557 * in all UD QP post sends. 558 */ 559struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr, 560 u32 flags) 561{ 562 const struct ib_gid_attr *old_sgid_attr; 563 struct net_device *slave; 564 struct ib_ah *ah; 565 int ret; 566 567 ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr); 568 if (ret) 569 return ERR_PTR(ret); 570 slave = rdma_lag_get_ah_roce_slave(pd->device, ah_attr, 571 (flags & RDMA_CREATE_AH_SLEEPABLE) ? 572 GFP_KERNEL : GFP_ATOMIC); 573 if (IS_ERR(slave)) { 574 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 575 return (void *)slave; 576 } 577 ah = _rdma_create_ah(pd, ah_attr, flags, NULL, slave); 578 rdma_lag_put_ah_roce_slave(slave); 579 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 580 return ah; 581} 582EXPORT_SYMBOL(rdma_create_ah); 583 584/** 585 * rdma_create_user_ah - Creates an address handle for the 586 * given address vector. 587 * It resolves destination mac address for ah attribute of RoCE type. 588 * @pd: The protection domain associated with the address handle. 589 * @ah_attr: The attributes of the address vector. 590 * @udata: pointer to user's input output buffer information need by 591 * provider driver. 592 * 593 * It returns 0 on success and returns appropriate error code on error. 594 * The address handle is used to reference a local or global destination 595 * in all UD QP post sends. 596 */ 597struct ib_ah *rdma_create_user_ah(struct ib_pd *pd, 598 struct rdma_ah_attr *ah_attr, 599 struct ib_udata *udata) 600{ 601 const struct ib_gid_attr *old_sgid_attr; 602 struct ib_ah *ah; 603 int err; 604 605 err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr); 606 if (err) 607 return ERR_PTR(err); 608 609 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) { 610 err = ib_resolve_eth_dmac(pd->device, ah_attr); 611 if (err) { 612 ah = ERR_PTR(err); 613 goto out; 614 } 615 } 616 617 ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE, 618 udata, NULL); 619 620out: 621 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 622 return ah; 623} 624EXPORT_SYMBOL(rdma_create_user_ah); 625 626int ib_get_rdma_header_version(const union rdma_network_hdr *hdr) 627{ 628 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh; 629 struct iphdr ip4h_checked; 630 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh; 631 632 /* If it's IPv6, the version must be 6, otherwise, the first 633 * 20 bytes (before the IPv4 header) are garbled. 634 */ 635 if (ip6h->version != 6) 636 return (ip4h->version == 4) ? 4 : 0; 637 /* version may be 6 or 4 because the first 20 bytes could be garbled */ 638 639 /* RoCE v2 requires no options, thus header length 640 * must be 5 words 641 */ 642 if (ip4h->ihl != 5) 643 return 6; 644 645 /* Verify checksum. 646 * We can't write on scattered buffers so we need to copy to 647 * temp buffer. 648 */ 649 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked)); 650 ip4h_checked.check = 0; 651 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5); 652 /* if IPv4 header checksum is OK, believe it */ 653 if (ip4h->check == ip4h_checked.check) 654 return 4; 655 return 6; 656} 657EXPORT_SYMBOL(ib_get_rdma_header_version); 658 659static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device, 660 u8 port_num, 661 const struct ib_grh *grh) 662{ 663 int grh_version; 664 665 if (rdma_protocol_ib(device, port_num)) 666 return RDMA_NETWORK_IB; 667 668 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh); 669 670 if (grh_version == 4) 671 return RDMA_NETWORK_IPV4; 672 673 if (grh->next_hdr == IPPROTO_UDP) 674 return RDMA_NETWORK_IPV6; 675 676 return RDMA_NETWORK_ROCE_V1; 677} 678 679struct find_gid_index_context { 680 u16 vlan_id; 681 enum ib_gid_type gid_type; 682}; 683 684static bool find_gid_index(const union ib_gid *gid, 685 const struct ib_gid_attr *gid_attr, 686 void *context) 687{ 688 struct find_gid_index_context *ctx = context; 689 u16 vlan_id = 0xffff; 690 int ret; 691 692 if (ctx->gid_type != gid_attr->gid_type) 693 return false; 694 695 ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL); 696 if (ret) 697 return false; 698 699 return ctx->vlan_id == vlan_id; 700} 701 702static const struct ib_gid_attr * 703get_sgid_attr_from_eth(struct ib_device *device, u8 port_num, 704 u16 vlan_id, const union ib_gid *sgid, 705 enum ib_gid_type gid_type) 706{ 707 struct find_gid_index_context context = {.vlan_id = vlan_id, 708 .gid_type = gid_type}; 709 710 return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index, 711 &context); 712} 713 714int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr, 715 enum rdma_network_type net_type, 716 union ib_gid *sgid, union ib_gid *dgid) 717{ 718 struct sockaddr_in src_in; 719 struct sockaddr_in dst_in; 720 __be32 src_saddr, dst_saddr; 721 722 if (!sgid || !dgid) 723 return -EINVAL; 724 725 if (net_type == RDMA_NETWORK_IPV4) { 726 memcpy(&src_in.sin_addr.s_addr, 727 &hdr->roce4grh.saddr, 4); 728 memcpy(&dst_in.sin_addr.s_addr, 729 &hdr->roce4grh.daddr, 4); 730 src_saddr = src_in.sin_addr.s_addr; 731 dst_saddr = dst_in.sin_addr.s_addr; 732 ipv6_addr_set_v4mapped(src_saddr, 733 (struct in6_addr *)sgid); 734 ipv6_addr_set_v4mapped(dst_saddr, 735 (struct in6_addr *)dgid); 736 return 0; 737 } else if (net_type == RDMA_NETWORK_IPV6 || 738 net_type == RDMA_NETWORK_IB || RDMA_NETWORK_ROCE_V1) { 739 *dgid = hdr->ibgrh.dgid; 740 *sgid = hdr->ibgrh.sgid; 741 return 0; 742 } else { 743 return -EINVAL; 744 } 745} 746EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr); 747 748/* Resolve destination mac address and hop limit for unicast destination 749 * GID entry, considering the source GID entry as well. 750 * ah_attribute must have have valid port_num, sgid_index. 751 */ 752static int ib_resolve_unicast_gid_dmac(struct ib_device *device, 753 struct rdma_ah_attr *ah_attr) 754{ 755 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr); 756 const struct ib_gid_attr *sgid_attr = grh->sgid_attr; 757 int hop_limit = 0xff; 758 int ret = 0; 759 760 /* If destination is link local and source GID is RoCEv1, 761 * IP stack is not used. 762 */ 763 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) && 764 sgid_attr->gid_type == IB_GID_TYPE_ROCE) { 765 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw, 766 ah_attr->roce.dmac); 767 return ret; 768 } 769 770 ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid, 771 ah_attr->roce.dmac, 772 sgid_attr, &hop_limit); 773 774 grh->hop_limit = hop_limit; 775 return ret; 776} 777 778/* 779 * This function initializes address handle attributes from the incoming packet. 780 * Incoming packet has dgid of the receiver node on which this code is 781 * getting executed and, sgid contains the GID of the sender. 782 * 783 * When resolving mac address of destination, the arrived dgid is used 784 * as sgid and, sgid is used as dgid because sgid contains destinations 785 * GID whom to respond to. 786 * 787 * On success the caller is responsible to call rdma_destroy_ah_attr on the 788 * attr. 789 */ 790int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num, 791 const struct ib_wc *wc, const struct ib_grh *grh, 792 struct rdma_ah_attr *ah_attr) 793{ 794 u32 flow_class; 795 int ret; 796 enum rdma_network_type net_type = RDMA_NETWORK_IB; 797 enum ib_gid_type gid_type = IB_GID_TYPE_IB; 798 const struct ib_gid_attr *sgid_attr; 799 int hoplimit = 0xff; 800 union ib_gid dgid; 801 union ib_gid sgid; 802 803 might_sleep(); 804 805 memset(ah_attr, 0, sizeof *ah_attr); 806 ah_attr->type = rdma_ah_find_type(device, port_num); 807 if (rdma_cap_eth_ah(device, port_num)) { 808 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE) 809 net_type = wc->network_hdr_type; 810 else 811 net_type = ib_get_net_type_by_grh(device, port_num, grh); 812 gid_type = ib_network_to_gid_type(net_type); 813 } 814 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type, 815 &sgid, &dgid); 816 if (ret) 817 return ret; 818 819 rdma_ah_set_sl(ah_attr, wc->sl); 820 rdma_ah_set_port_num(ah_attr, port_num); 821 822 if (rdma_protocol_roce(device, port_num)) { 823 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ? 824 wc->vlan_id : 0xffff; 825 826 if (!(wc->wc_flags & IB_WC_GRH)) 827 return -EPROTOTYPE; 828 829 sgid_attr = get_sgid_attr_from_eth(device, port_num, 830 vlan_id, &dgid, 831 gid_type); 832 if (IS_ERR(sgid_attr)) 833 return PTR_ERR(sgid_attr); 834 835 flow_class = be32_to_cpu(grh->version_tclass_flow); 836 rdma_move_grh_sgid_attr(ah_attr, 837 &sgid, 838 flow_class & 0xFFFFF, 839 hoplimit, 840 (flow_class >> 20) & 0xFF, 841 sgid_attr); 842 843 ret = ib_resolve_unicast_gid_dmac(device, ah_attr); 844 if (ret) 845 rdma_destroy_ah_attr(ah_attr); 846 847 return ret; 848 } else { 849 rdma_ah_set_dlid(ah_attr, wc->slid); 850 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits); 851 852 if ((wc->wc_flags & IB_WC_GRH) == 0) 853 return 0; 854 855 if (dgid.global.interface_id != 856 cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) { 857 sgid_attr = rdma_find_gid_by_port( 858 device, &dgid, IB_GID_TYPE_IB, port_num, NULL); 859 } else 860 sgid_attr = rdma_get_gid_attr(device, port_num, 0); 861 862 if (IS_ERR(sgid_attr)) 863 return PTR_ERR(sgid_attr); 864 flow_class = be32_to_cpu(grh->version_tclass_flow); 865 rdma_move_grh_sgid_attr(ah_attr, 866 &sgid, 867 flow_class & 0xFFFFF, 868 hoplimit, 869 (flow_class >> 20) & 0xFF, 870 sgid_attr); 871 872 return 0; 873 } 874} 875EXPORT_SYMBOL(ib_init_ah_attr_from_wc); 876 877/** 878 * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership 879 * of the reference 880 * 881 * @attr: Pointer to AH attribute structure 882 * @dgid: Destination GID 883 * @flow_label: Flow label 884 * @hop_limit: Hop limit 885 * @traffic_class: traffic class 886 * @sgid_attr: Pointer to SGID attribute 887 * 888 * This takes ownership of the sgid_attr reference. The caller must ensure 889 * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after 890 * calling this function. 891 */ 892void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid, 893 u32 flow_label, u8 hop_limit, u8 traffic_class, 894 const struct ib_gid_attr *sgid_attr) 895{ 896 rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit, 897 traffic_class); 898 attr->grh.sgid_attr = sgid_attr; 899} 900EXPORT_SYMBOL(rdma_move_grh_sgid_attr); 901 902/** 903 * rdma_destroy_ah_attr - Release reference to SGID attribute of 904 * ah attribute. 905 * @ah_attr: Pointer to ah attribute 906 * 907 * Release reference to the SGID attribute of the ah attribute if it is 908 * non NULL. It is safe to call this multiple times, and safe to call it on 909 * a zero initialized ah_attr. 910 */ 911void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr) 912{ 913 if (ah_attr->grh.sgid_attr) { 914 rdma_put_gid_attr(ah_attr->grh.sgid_attr); 915 ah_attr->grh.sgid_attr = NULL; 916 } 917} 918EXPORT_SYMBOL(rdma_destroy_ah_attr); 919 920struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc, 921 const struct ib_grh *grh, u8 port_num) 922{ 923 struct rdma_ah_attr ah_attr; 924 struct ib_ah *ah; 925 int ret; 926 927 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr); 928 if (ret) 929 return ERR_PTR(ret); 930 931 ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE); 932 933 rdma_destroy_ah_attr(&ah_attr); 934 return ah; 935} 936EXPORT_SYMBOL(ib_create_ah_from_wc); 937 938int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr) 939{ 940 const struct ib_gid_attr *old_sgid_attr; 941 int ret; 942 943 if (ah->type != ah_attr->type) 944 return -EINVAL; 945 946 ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr); 947 if (ret) 948 return ret; 949 950 ret = ah->device->ops.modify_ah ? 951 ah->device->ops.modify_ah(ah, ah_attr) : 952 -EOPNOTSUPP; 953 954 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr); 955 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 956 return ret; 957} 958EXPORT_SYMBOL(rdma_modify_ah); 959 960int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr) 961{ 962 ah_attr->grh.sgid_attr = NULL; 963 964 return ah->device->ops.query_ah ? 965 ah->device->ops.query_ah(ah, ah_attr) : 966 -EOPNOTSUPP; 967} 968EXPORT_SYMBOL(rdma_query_ah); 969 970int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata) 971{ 972 const struct ib_gid_attr *sgid_attr = ah->sgid_attr; 973 struct ib_pd *pd; 974 int ret; 975 976 might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE); 977 978 pd = ah->pd; 979 980 ret = ah->device->ops.destroy_ah(ah, flags); 981 if (ret) 982 return ret; 983 984 atomic_dec(&pd->usecnt); 985 if (sgid_attr) 986 rdma_put_gid_attr(sgid_attr); 987 988 kfree(ah); 989 return ret; 990} 991EXPORT_SYMBOL(rdma_destroy_ah_user); 992 993/* Shared receive queues */ 994 995/** 996 * ib_create_srq_user - Creates a SRQ associated with the specified protection 997 * domain. 998 * @pd: The protection domain associated with the SRQ. 999 * @srq_init_attr: A list of initial attributes required to create the 1000 * SRQ. If SRQ creation succeeds, then the attributes are updated to 1001 * the actual capabilities of the created SRQ. 1002 * @uobject: uobject pointer if this is not a kernel SRQ 1003 * @udata: udata pointer if this is not a kernel SRQ 1004 * 1005 * srq_attr->max_wr and srq_attr->max_sge are read the determine the 1006 * requested size of the SRQ, and set to the actual values allocated 1007 * on return. If ib_create_srq() succeeds, then max_wr and max_sge 1008 * will always be at least as large as the requested values. 1009 */ 1010struct ib_srq *ib_create_srq_user(struct ib_pd *pd, 1011 struct ib_srq_init_attr *srq_init_attr, 1012 struct ib_usrq_object *uobject, 1013 struct ib_udata *udata) 1014{ 1015 struct ib_srq *srq; 1016 int ret; 1017 1018 srq = rdma_zalloc_drv_obj(pd->device, ib_srq); 1019 if (!srq) 1020 return ERR_PTR(-ENOMEM); 1021 1022 srq->device = pd->device; 1023 srq->pd = pd; 1024 srq->event_handler = srq_init_attr->event_handler; 1025 srq->srq_context = srq_init_attr->srq_context; 1026 srq->srq_type = srq_init_attr->srq_type; 1027 srq->uobject = uobject; 1028 1029 if (ib_srq_has_cq(srq->srq_type)) { 1030 srq->ext.cq = srq_init_attr->ext.cq; 1031 atomic_inc(&srq->ext.cq->usecnt); 1032 } 1033 if (srq->srq_type == IB_SRQT_XRC) { 1034 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd; 1035 atomic_inc(&srq->ext.xrc.xrcd->usecnt); 1036 } 1037 atomic_inc(&pd->usecnt); 1038 1039 ret = pd->device->ops.create_srq(srq, srq_init_attr, udata); 1040 if (ret) { 1041 atomic_dec(&srq->pd->usecnt); 1042 if (srq->srq_type == IB_SRQT_XRC) 1043 atomic_dec(&srq->ext.xrc.xrcd->usecnt); 1044 if (ib_srq_has_cq(srq->srq_type)) 1045 atomic_dec(&srq->ext.cq->usecnt); 1046 kfree(srq); 1047 return ERR_PTR(ret); 1048 } 1049 1050 return srq; 1051} 1052EXPORT_SYMBOL(ib_create_srq_user); 1053 1054int ib_modify_srq(struct ib_srq *srq, 1055 struct ib_srq_attr *srq_attr, 1056 enum ib_srq_attr_mask srq_attr_mask) 1057{ 1058 return srq->device->ops.modify_srq ? 1059 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask, 1060 NULL) : -EOPNOTSUPP; 1061} 1062EXPORT_SYMBOL(ib_modify_srq); 1063 1064int ib_query_srq(struct ib_srq *srq, 1065 struct ib_srq_attr *srq_attr) 1066{ 1067 return srq->device->ops.query_srq ? 1068 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP; 1069} 1070EXPORT_SYMBOL(ib_query_srq); 1071 1072int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata) 1073{ 1074 int ret; 1075 1076 if (atomic_read(&srq->usecnt)) 1077 return -EBUSY; 1078 1079 ret = srq->device->ops.destroy_srq(srq, udata); 1080 if (ret) 1081 return ret; 1082 1083 atomic_dec(&srq->pd->usecnt); 1084 if (srq->srq_type == IB_SRQT_XRC) 1085 atomic_dec(&srq->ext.xrc.xrcd->usecnt); 1086 if (ib_srq_has_cq(srq->srq_type)) 1087 atomic_dec(&srq->ext.cq->usecnt); 1088 kfree(srq); 1089 1090 return ret; 1091} 1092EXPORT_SYMBOL(ib_destroy_srq_user); 1093 1094/* Queue pairs */ 1095 1096static void __ib_shared_qp_event_handler(struct ib_event *event, void *context) 1097{ 1098 struct ib_qp *qp = context; 1099 unsigned long flags; 1100 1101 spin_lock_irqsave(&qp->device->qp_open_list_lock, flags); 1102 list_for_each_entry(event->element.qp, &qp->open_list, open_list) 1103 if (event->element.qp->event_handler) 1104 event->element.qp->event_handler(event, event->element.qp->qp_context); 1105 spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags); 1106} 1107 1108static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp, 1109 void (*event_handler)(struct ib_event *, void *), 1110 void *qp_context) 1111{ 1112 struct ib_qp *qp; 1113 unsigned long flags; 1114 int err; 1115 1116 qp = kzalloc(sizeof *qp, GFP_KERNEL); 1117 if (!qp) 1118 return ERR_PTR(-ENOMEM); 1119 1120 qp->real_qp = real_qp; 1121 err = ib_open_shared_qp_security(qp, real_qp->device); 1122 if (err) { 1123 kfree(qp); 1124 return ERR_PTR(err); 1125 } 1126 1127 qp->real_qp = real_qp; 1128 atomic_inc(&real_qp->usecnt); 1129 qp->device = real_qp->device; 1130 qp->event_handler = event_handler; 1131 qp->qp_context = qp_context; 1132 qp->qp_num = real_qp->qp_num; 1133 qp->qp_type = real_qp->qp_type; 1134 1135 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags); 1136 list_add(&qp->open_list, &real_qp->open_list); 1137 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags); 1138 1139 return qp; 1140} 1141 1142struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd, 1143 struct ib_qp_open_attr *qp_open_attr) 1144{ 1145 struct ib_qp *qp, *real_qp; 1146 1147 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT) 1148 return ERR_PTR(-EINVAL); 1149 1150 down_read(&xrcd->tgt_qps_rwsem); 1151 real_qp = xa_load(&xrcd->tgt_qps, qp_open_attr->qp_num); 1152 if (!real_qp) { 1153 up_read(&xrcd->tgt_qps_rwsem); 1154 return ERR_PTR(-EINVAL); 1155 } 1156 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler, 1157 qp_open_attr->qp_context); 1158 up_read(&xrcd->tgt_qps_rwsem); 1159 return qp; 1160} 1161EXPORT_SYMBOL(ib_open_qp); 1162 1163static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp, 1164 struct ib_qp_init_attr *qp_init_attr) 1165{ 1166 struct ib_qp *real_qp = qp; 1167 int err; 1168 1169 qp->event_handler = __ib_shared_qp_event_handler; 1170 qp->qp_context = qp; 1171 qp->pd = NULL; 1172 qp->send_cq = qp->recv_cq = NULL; 1173 qp->srq = NULL; 1174 qp->xrcd = qp_init_attr->xrcd; 1175 atomic_inc(&qp_init_attr->xrcd->usecnt); 1176 INIT_LIST_HEAD(&qp->open_list); 1177 1178 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler, 1179 qp_init_attr->qp_context); 1180 if (IS_ERR(qp)) 1181 return qp; 1182 1183 err = xa_err(xa_store(&qp_init_attr->xrcd->tgt_qps, real_qp->qp_num, 1184 real_qp, GFP_KERNEL)); 1185 if (err) { 1186 ib_close_qp(qp); 1187 return ERR_PTR(err); 1188 } 1189 return qp; 1190} 1191 1192/** 1193 * ib_create_qp - Creates a kernel QP associated with the specified protection 1194 * domain. 1195 * @pd: The protection domain associated with the QP. 1196 * @qp_init_attr: A list of initial attributes required to create the 1197 * QP. If QP creation succeeds, then the attributes are updated to 1198 * the actual capabilities of the created QP. 1199 * 1200 * NOTE: for user qp use ib_create_qp_user with valid udata! 1201 */ 1202struct ib_qp *ib_create_qp(struct ib_pd *pd, 1203 struct ib_qp_init_attr *qp_init_attr) 1204{ 1205 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device; 1206 struct ib_qp *qp; 1207 int ret; 1208 1209 if (qp_init_attr->rwq_ind_tbl && 1210 (qp_init_attr->recv_cq || 1211 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr || 1212 qp_init_attr->cap.max_recv_sge)) 1213 return ERR_PTR(-EINVAL); 1214 1215 if ((qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) && 1216 !(device->attrs.device_cap_flags & IB_DEVICE_INTEGRITY_HANDOVER)) 1217 return ERR_PTR(-EINVAL); 1218 1219 /* 1220 * If the callers is using the RDMA API calculate the resources 1221 * needed for the RDMA READ/WRITE operations. 1222 * 1223 * Note that these callers need to pass in a port number. 1224 */ 1225 if (qp_init_attr->cap.max_rdma_ctxs) 1226 rdma_rw_init_qp(device, qp_init_attr); 1227 1228 qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL); 1229 if (IS_ERR(qp)) 1230 return qp; 1231 1232 ret = ib_create_qp_security(qp, device); 1233 if (ret) 1234 goto err; 1235 1236 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) { 1237 struct ib_qp *xrc_qp = 1238 create_xrc_qp_user(qp, qp_init_attr); 1239 1240 if (IS_ERR(xrc_qp)) { 1241 ret = PTR_ERR(xrc_qp); 1242 goto err; 1243 } 1244 return xrc_qp; 1245 } 1246 1247 qp->event_handler = qp_init_attr->event_handler; 1248 qp->qp_context = qp_init_attr->qp_context; 1249 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) { 1250 qp->recv_cq = NULL; 1251 qp->srq = NULL; 1252 } else { 1253 qp->recv_cq = qp_init_attr->recv_cq; 1254 if (qp_init_attr->recv_cq) 1255 atomic_inc(&qp_init_attr->recv_cq->usecnt); 1256 qp->srq = qp_init_attr->srq; 1257 if (qp->srq) 1258 atomic_inc(&qp_init_attr->srq->usecnt); 1259 } 1260 1261 qp->send_cq = qp_init_attr->send_cq; 1262 qp->xrcd = NULL; 1263 1264 atomic_inc(&pd->usecnt); 1265 if (qp_init_attr->send_cq) 1266 atomic_inc(&qp_init_attr->send_cq->usecnt); 1267 if (qp_init_attr->rwq_ind_tbl) 1268 atomic_inc(&qp->rwq_ind_tbl->usecnt); 1269 1270 if (qp_init_attr->cap.max_rdma_ctxs) { 1271 ret = rdma_rw_init_mrs(qp, qp_init_attr); 1272 if (ret) 1273 goto err; 1274 } 1275 1276 /* 1277 * Note: all hw drivers guarantee that max_send_sge is lower than 1278 * the device RDMA WRITE SGE limit but not all hw drivers ensure that 1279 * max_send_sge <= max_sge_rd. 1280 */ 1281 qp->max_write_sge = qp_init_attr->cap.max_send_sge; 1282 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge, 1283 device->attrs.max_sge_rd); 1284 if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) 1285 qp->integrity_en = true; 1286 1287 return qp; 1288 1289err: 1290 ib_destroy_qp(qp); 1291 return ERR_PTR(ret); 1292 1293} 1294EXPORT_SYMBOL(ib_create_qp); 1295 1296static const struct { 1297 int valid; 1298 enum ib_qp_attr_mask req_param[IB_QPT_MAX]; 1299 enum ib_qp_attr_mask opt_param[IB_QPT_MAX]; 1300} qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = { 1301 [IB_QPS_RESET] = { 1302 [IB_QPS_RESET] = { .valid = 1 }, 1303 [IB_QPS_INIT] = { 1304 .valid = 1, 1305 .req_param = { 1306 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1307 IB_QP_PORT | 1308 IB_QP_QKEY), 1309 [IB_QPT_RAW_PACKET] = IB_QP_PORT, 1310 [IB_QPT_UC] = (IB_QP_PKEY_INDEX | 1311 IB_QP_PORT | 1312 IB_QP_ACCESS_FLAGS), 1313 [IB_QPT_RC] = (IB_QP_PKEY_INDEX | 1314 IB_QP_PORT | 1315 IB_QP_ACCESS_FLAGS), 1316 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX | 1317 IB_QP_PORT | 1318 IB_QP_ACCESS_FLAGS), 1319 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX | 1320 IB_QP_PORT | 1321 IB_QP_ACCESS_FLAGS), 1322 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1323 IB_QP_QKEY), 1324 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1325 IB_QP_QKEY), 1326 } 1327 }, 1328 }, 1329 [IB_QPS_INIT] = { 1330 [IB_QPS_RESET] = { .valid = 1 }, 1331 [IB_QPS_ERR] = { .valid = 1 }, 1332 [IB_QPS_INIT] = { 1333 .valid = 1, 1334 .opt_param = { 1335 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1336 IB_QP_PORT | 1337 IB_QP_QKEY), 1338 [IB_QPT_UC] = (IB_QP_PKEY_INDEX | 1339 IB_QP_PORT | 1340 IB_QP_ACCESS_FLAGS), 1341 [IB_QPT_RC] = (IB_QP_PKEY_INDEX | 1342 IB_QP_PORT | 1343 IB_QP_ACCESS_FLAGS), 1344 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX | 1345 IB_QP_PORT | 1346 IB_QP_ACCESS_FLAGS), 1347 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX | 1348 IB_QP_PORT | 1349 IB_QP_ACCESS_FLAGS), 1350 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1351 IB_QP_QKEY), 1352 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1353 IB_QP_QKEY), 1354 } 1355 }, 1356 [IB_QPS_RTR] = { 1357 .valid = 1, 1358 .req_param = { 1359 [IB_QPT_UC] = (IB_QP_AV | 1360 IB_QP_PATH_MTU | 1361 IB_QP_DEST_QPN | 1362 IB_QP_RQ_PSN), 1363 [IB_QPT_RC] = (IB_QP_AV | 1364 IB_QP_PATH_MTU | 1365 IB_QP_DEST_QPN | 1366 IB_QP_RQ_PSN | 1367 IB_QP_MAX_DEST_RD_ATOMIC | 1368 IB_QP_MIN_RNR_TIMER), 1369 [IB_QPT_XRC_INI] = (IB_QP_AV | 1370 IB_QP_PATH_MTU | 1371 IB_QP_DEST_QPN | 1372 IB_QP_RQ_PSN), 1373 [IB_QPT_XRC_TGT] = (IB_QP_AV | 1374 IB_QP_PATH_MTU | 1375 IB_QP_DEST_QPN | 1376 IB_QP_RQ_PSN | 1377 IB_QP_MAX_DEST_RD_ATOMIC | 1378 IB_QP_MIN_RNR_TIMER), 1379 }, 1380 .opt_param = { 1381 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1382 IB_QP_QKEY), 1383 [IB_QPT_UC] = (IB_QP_ALT_PATH | 1384 IB_QP_ACCESS_FLAGS | 1385 IB_QP_PKEY_INDEX), 1386 [IB_QPT_RC] = (IB_QP_ALT_PATH | 1387 IB_QP_ACCESS_FLAGS | 1388 IB_QP_PKEY_INDEX), 1389 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH | 1390 IB_QP_ACCESS_FLAGS | 1391 IB_QP_PKEY_INDEX), 1392 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH | 1393 IB_QP_ACCESS_FLAGS | 1394 IB_QP_PKEY_INDEX), 1395 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1396 IB_QP_QKEY), 1397 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1398 IB_QP_QKEY), 1399 }, 1400 }, 1401 }, 1402 [IB_QPS_RTR] = { 1403 [IB_QPS_RESET] = { .valid = 1 }, 1404 [IB_QPS_ERR] = { .valid = 1 }, 1405 [IB_QPS_RTS] = { 1406 .valid = 1, 1407 .req_param = { 1408 [IB_QPT_UD] = IB_QP_SQ_PSN, 1409 [IB_QPT_UC] = IB_QP_SQ_PSN, 1410 [IB_QPT_RC] = (IB_QP_TIMEOUT | 1411 IB_QP_RETRY_CNT | 1412 IB_QP_RNR_RETRY | 1413 IB_QP_SQ_PSN | 1414 IB_QP_MAX_QP_RD_ATOMIC), 1415 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT | 1416 IB_QP_RETRY_CNT | 1417 IB_QP_RNR_RETRY | 1418 IB_QP_SQ_PSN | 1419 IB_QP_MAX_QP_RD_ATOMIC), 1420 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT | 1421 IB_QP_SQ_PSN), 1422 [IB_QPT_SMI] = IB_QP_SQ_PSN, 1423 [IB_QPT_GSI] = IB_QP_SQ_PSN, 1424 }, 1425 .opt_param = { 1426 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1427 IB_QP_QKEY), 1428 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1429 IB_QP_ALT_PATH | 1430 IB_QP_ACCESS_FLAGS | 1431 IB_QP_PATH_MIG_STATE), 1432 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1433 IB_QP_ALT_PATH | 1434 IB_QP_ACCESS_FLAGS | 1435 IB_QP_MIN_RNR_TIMER | 1436 IB_QP_PATH_MIG_STATE), 1437 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1438 IB_QP_ALT_PATH | 1439 IB_QP_ACCESS_FLAGS | 1440 IB_QP_PATH_MIG_STATE), 1441 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1442 IB_QP_ALT_PATH | 1443 IB_QP_ACCESS_FLAGS | 1444 IB_QP_MIN_RNR_TIMER | 1445 IB_QP_PATH_MIG_STATE), 1446 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1447 IB_QP_QKEY), 1448 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1449 IB_QP_QKEY), 1450 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT, 1451 } 1452 } 1453 }, 1454 [IB_QPS_RTS] = { 1455 [IB_QPS_RESET] = { .valid = 1 }, 1456 [IB_QPS_ERR] = { .valid = 1 }, 1457 [IB_QPS_RTS] = { 1458 .valid = 1, 1459 .opt_param = { 1460 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1461 IB_QP_QKEY), 1462 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1463 IB_QP_ACCESS_FLAGS | 1464 IB_QP_ALT_PATH | 1465 IB_QP_PATH_MIG_STATE), 1466 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1467 IB_QP_ACCESS_FLAGS | 1468 IB_QP_ALT_PATH | 1469 IB_QP_PATH_MIG_STATE | 1470 IB_QP_MIN_RNR_TIMER), 1471 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1472 IB_QP_ACCESS_FLAGS | 1473 IB_QP_ALT_PATH | 1474 IB_QP_PATH_MIG_STATE), 1475 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1476 IB_QP_ACCESS_FLAGS | 1477 IB_QP_ALT_PATH | 1478 IB_QP_PATH_MIG_STATE | 1479 IB_QP_MIN_RNR_TIMER), 1480 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1481 IB_QP_QKEY), 1482 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1483 IB_QP_QKEY), 1484 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT, 1485 } 1486 }, 1487 [IB_QPS_SQD] = { 1488 .valid = 1, 1489 .opt_param = { 1490 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1491 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1492 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1493 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1494 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */ 1495 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1496 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY 1497 } 1498 }, 1499 }, 1500 [IB_QPS_SQD] = { 1501 [IB_QPS_RESET] = { .valid = 1 }, 1502 [IB_QPS_ERR] = { .valid = 1 }, 1503 [IB_QPS_RTS] = { 1504 .valid = 1, 1505 .opt_param = { 1506 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1507 IB_QP_QKEY), 1508 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1509 IB_QP_ALT_PATH | 1510 IB_QP_ACCESS_FLAGS | 1511 IB_QP_PATH_MIG_STATE), 1512 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1513 IB_QP_ALT_PATH | 1514 IB_QP_ACCESS_FLAGS | 1515 IB_QP_MIN_RNR_TIMER | 1516 IB_QP_PATH_MIG_STATE), 1517 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1518 IB_QP_ALT_PATH | 1519 IB_QP_ACCESS_FLAGS | 1520 IB_QP_PATH_MIG_STATE), 1521 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1522 IB_QP_ALT_PATH | 1523 IB_QP_ACCESS_FLAGS | 1524 IB_QP_MIN_RNR_TIMER | 1525 IB_QP_PATH_MIG_STATE), 1526 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1527 IB_QP_QKEY), 1528 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1529 IB_QP_QKEY), 1530 } 1531 }, 1532 [IB_QPS_SQD] = { 1533 .valid = 1, 1534 .opt_param = { 1535 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1536 IB_QP_QKEY), 1537 [IB_QPT_UC] = (IB_QP_AV | 1538 IB_QP_ALT_PATH | 1539 IB_QP_ACCESS_FLAGS | 1540 IB_QP_PKEY_INDEX | 1541 IB_QP_PATH_MIG_STATE), 1542 [IB_QPT_RC] = (IB_QP_PORT | 1543 IB_QP_AV | 1544 IB_QP_TIMEOUT | 1545 IB_QP_RETRY_CNT | 1546 IB_QP_RNR_RETRY | 1547 IB_QP_MAX_QP_RD_ATOMIC | 1548 IB_QP_MAX_DEST_RD_ATOMIC | 1549 IB_QP_ALT_PATH | 1550 IB_QP_ACCESS_FLAGS | 1551 IB_QP_PKEY_INDEX | 1552 IB_QP_MIN_RNR_TIMER | 1553 IB_QP_PATH_MIG_STATE), 1554 [IB_QPT_XRC_INI] = (IB_QP_PORT | 1555 IB_QP_AV | 1556 IB_QP_TIMEOUT | 1557 IB_QP_RETRY_CNT | 1558 IB_QP_RNR_RETRY | 1559 IB_QP_MAX_QP_RD_ATOMIC | 1560 IB_QP_ALT_PATH | 1561 IB_QP_ACCESS_FLAGS | 1562 IB_QP_PKEY_INDEX | 1563 IB_QP_PATH_MIG_STATE), 1564 [IB_QPT_XRC_TGT] = (IB_QP_PORT | 1565 IB_QP_AV | 1566 IB_QP_TIMEOUT | 1567 IB_QP_MAX_DEST_RD_ATOMIC | 1568 IB_QP_ALT_PATH | 1569 IB_QP_ACCESS_FLAGS | 1570 IB_QP_PKEY_INDEX | 1571 IB_QP_MIN_RNR_TIMER | 1572 IB_QP_PATH_MIG_STATE), 1573 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1574 IB_QP_QKEY), 1575 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1576 IB_QP_QKEY), 1577 } 1578 } 1579 }, 1580 [IB_QPS_SQE] = { 1581 [IB_QPS_RESET] = { .valid = 1 }, 1582 [IB_QPS_ERR] = { .valid = 1 }, 1583 [IB_QPS_RTS] = { 1584 .valid = 1, 1585 .opt_param = { 1586 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1587 IB_QP_QKEY), 1588 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1589 IB_QP_ACCESS_FLAGS), 1590 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1591 IB_QP_QKEY), 1592 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1593 IB_QP_QKEY), 1594 } 1595 } 1596 }, 1597 [IB_QPS_ERR] = { 1598 [IB_QPS_RESET] = { .valid = 1 }, 1599 [IB_QPS_ERR] = { .valid = 1 } 1600 } 1601}; 1602 1603bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state, 1604 enum ib_qp_type type, enum ib_qp_attr_mask mask) 1605{ 1606 enum ib_qp_attr_mask req_param, opt_param; 1607 1608 if (mask & IB_QP_CUR_STATE && 1609 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS && 1610 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE) 1611 return false; 1612 1613 if (!qp_state_table[cur_state][next_state].valid) 1614 return false; 1615 1616 req_param = qp_state_table[cur_state][next_state].req_param[type]; 1617 opt_param = qp_state_table[cur_state][next_state].opt_param[type]; 1618 1619 if ((mask & req_param) != req_param) 1620 return false; 1621 1622 if (mask & ~(req_param | opt_param | IB_QP_STATE)) 1623 return false; 1624 1625 return true; 1626} 1627EXPORT_SYMBOL(ib_modify_qp_is_ok); 1628 1629/** 1630 * ib_resolve_eth_dmac - Resolve destination mac address 1631 * @device: Device to consider 1632 * @ah_attr: address handle attribute which describes the 1633 * source and destination parameters 1634 * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It 1635 * returns 0 on success or appropriate error code. It initializes the 1636 * necessary ah_attr fields when call is successful. 1637 */ 1638static int ib_resolve_eth_dmac(struct ib_device *device, 1639 struct rdma_ah_attr *ah_attr) 1640{ 1641 int ret = 0; 1642 1643 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) { 1644 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) { 1645 __be32 addr = 0; 1646 1647 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4); 1648 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac); 1649 } else { 1650 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw, 1651 (char *)ah_attr->roce.dmac); 1652 } 1653 } else { 1654 ret = ib_resolve_unicast_gid_dmac(device, ah_attr); 1655 } 1656 return ret; 1657} 1658 1659static bool is_qp_type_connected(const struct ib_qp *qp) 1660{ 1661 return (qp->qp_type == IB_QPT_UC || 1662 qp->qp_type == IB_QPT_RC || 1663 qp->qp_type == IB_QPT_XRC_INI || 1664 qp->qp_type == IB_QPT_XRC_TGT); 1665} 1666 1667/** 1668 * IB core internal function to perform QP attributes modification. 1669 */ 1670static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr, 1671 int attr_mask, struct ib_udata *udata) 1672{ 1673 u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port; 1674 const struct ib_gid_attr *old_sgid_attr_av; 1675 const struct ib_gid_attr *old_sgid_attr_alt_av; 1676 int ret; 1677 1678 attr->xmit_slave = NULL; 1679 if (attr_mask & IB_QP_AV) { 1680 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr, 1681 &old_sgid_attr_av); 1682 if (ret) 1683 return ret; 1684 1685 if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE && 1686 is_qp_type_connected(qp)) { 1687 struct net_device *slave; 1688 1689 /* 1690 * If the user provided the qp_attr then we have to 1691 * resolve it. Kerne users have to provide already 1692 * resolved rdma_ah_attr's. 1693 */ 1694 if (udata) { 1695 ret = ib_resolve_eth_dmac(qp->device, 1696 &attr->ah_attr); 1697 if (ret) 1698 goto out_av; 1699 } 1700 slave = rdma_lag_get_ah_roce_slave(qp->device, 1701 &attr->ah_attr, 1702 GFP_KERNEL); 1703 if (IS_ERR(slave)) { 1704 ret = PTR_ERR(slave); 1705 goto out_av; 1706 } 1707 attr->xmit_slave = slave; 1708 } 1709 } 1710 if (attr_mask & IB_QP_ALT_PATH) { 1711 /* 1712 * FIXME: This does not track the migration state, so if the 1713 * user loads a new alternate path after the HW has migrated 1714 * from primary->alternate we will keep the wrong 1715 * references. This is OK for IB because the reference 1716 * counting does not serve any functional purpose. 1717 */ 1718 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr, 1719 &old_sgid_attr_alt_av); 1720 if (ret) 1721 goto out_av; 1722 1723 /* 1724 * Today the core code can only handle alternate paths and APM 1725 * for IB. Ban them in roce mode. 1726 */ 1727 if (!(rdma_protocol_ib(qp->device, 1728 attr->alt_ah_attr.port_num) && 1729 rdma_protocol_ib(qp->device, port))) { 1730 ret = -EINVAL; 1731 goto out; 1732 } 1733 } 1734 1735 if (rdma_ib_or_roce(qp->device, port)) { 1736 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) { 1737 dev_warn(&qp->device->dev, 1738 "%s rq_psn overflow, masking to 24 bits\n", 1739 __func__); 1740 attr->rq_psn &= 0xffffff; 1741 } 1742 1743 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) { 1744 dev_warn(&qp->device->dev, 1745 " %s sq_psn overflow, masking to 24 bits\n", 1746 __func__); 1747 attr->sq_psn &= 0xffffff; 1748 } 1749 } 1750 1751 /* 1752 * Bind this qp to a counter automatically based on the rdma counter 1753 * rules. This only set in RST2INIT with port specified 1754 */ 1755 if (!qp->counter && (attr_mask & IB_QP_PORT) && 1756 ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT)) 1757 rdma_counter_bind_qp_auto(qp, attr->port_num); 1758 1759 ret = ib_security_modify_qp(qp, attr, attr_mask, udata); 1760 if (ret) 1761 goto out; 1762 1763 if (attr_mask & IB_QP_PORT) 1764 qp->port = attr->port_num; 1765 if (attr_mask & IB_QP_AV) 1766 qp->av_sgid_attr = 1767 rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr); 1768 if (attr_mask & IB_QP_ALT_PATH) 1769 qp->alt_path_sgid_attr = rdma_update_sgid_attr( 1770 &attr->alt_ah_attr, qp->alt_path_sgid_attr); 1771 1772out: 1773 if (attr_mask & IB_QP_ALT_PATH) 1774 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av); 1775out_av: 1776 if (attr_mask & IB_QP_AV) { 1777 rdma_lag_put_ah_roce_slave(attr->xmit_slave); 1778 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av); 1779 } 1780 return ret; 1781} 1782 1783/** 1784 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP. 1785 * @ib_qp: The QP to modify. 1786 * @attr: On input, specifies the QP attributes to modify. On output, 1787 * the current values of selected QP attributes are returned. 1788 * @attr_mask: A bit-mask used to specify which attributes of the QP 1789 * are being modified. 1790 * @udata: pointer to user's input output buffer information 1791 * are being modified. 1792 * It returns 0 on success and returns appropriate error code on error. 1793 */ 1794int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr, 1795 int attr_mask, struct ib_udata *udata) 1796{ 1797 return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata); 1798} 1799EXPORT_SYMBOL(ib_modify_qp_with_udata); 1800 1801int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u16 *speed, u8 *width) 1802{ 1803 int rc; 1804 u32 netdev_speed; 1805 struct net_device *netdev; 1806 struct ethtool_link_ksettings lksettings; 1807 1808 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET) 1809 return -EINVAL; 1810 1811 netdev = ib_device_get_netdev(dev, port_num); 1812 if (!netdev) 1813 return -ENODEV; 1814 1815 rtnl_lock(); 1816 rc = __ethtool_get_link_ksettings(netdev, &lksettings); 1817 rtnl_unlock(); 1818 1819 dev_put(netdev); 1820 1821 if (!rc && lksettings.base.speed != (u32)SPEED_UNKNOWN) { 1822 netdev_speed = lksettings.base.speed; 1823 } else { 1824 netdev_speed = SPEED_1000; 1825 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name, 1826 netdev_speed); 1827 } 1828 1829 if (netdev_speed <= SPEED_1000) { 1830 *width = IB_WIDTH_1X; 1831 *speed = IB_SPEED_SDR; 1832 } else if (netdev_speed <= SPEED_10000) { 1833 *width = IB_WIDTH_1X; 1834 *speed = IB_SPEED_FDR10; 1835 } else if (netdev_speed <= SPEED_20000) { 1836 *width = IB_WIDTH_4X; 1837 *speed = IB_SPEED_DDR; 1838 } else if (netdev_speed <= SPEED_25000) { 1839 *width = IB_WIDTH_1X; 1840 *speed = IB_SPEED_EDR; 1841 } else if (netdev_speed <= SPEED_40000) { 1842 *width = IB_WIDTH_4X; 1843 *speed = IB_SPEED_FDR10; 1844 } else { 1845 *width = IB_WIDTH_4X; 1846 *speed = IB_SPEED_EDR; 1847 } 1848 1849 return 0; 1850} 1851EXPORT_SYMBOL(ib_get_eth_speed); 1852 1853int ib_modify_qp(struct ib_qp *qp, 1854 struct ib_qp_attr *qp_attr, 1855 int qp_attr_mask) 1856{ 1857 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL); 1858} 1859EXPORT_SYMBOL(ib_modify_qp); 1860 1861int ib_query_qp(struct ib_qp *qp, 1862 struct ib_qp_attr *qp_attr, 1863 int qp_attr_mask, 1864 struct ib_qp_init_attr *qp_init_attr) 1865{ 1866 qp_attr->ah_attr.grh.sgid_attr = NULL; 1867 qp_attr->alt_ah_attr.grh.sgid_attr = NULL; 1868 1869 return qp->device->ops.query_qp ? 1870 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask, 1871 qp_init_attr) : -EOPNOTSUPP; 1872} 1873EXPORT_SYMBOL(ib_query_qp); 1874 1875int ib_close_qp(struct ib_qp *qp) 1876{ 1877 struct ib_qp *real_qp; 1878 unsigned long flags; 1879 1880 real_qp = qp->real_qp; 1881 if (real_qp == qp) 1882 return -EINVAL; 1883 1884 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags); 1885 list_del(&qp->open_list); 1886 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags); 1887 1888 atomic_dec(&real_qp->usecnt); 1889 if (qp->qp_sec) 1890 ib_close_shared_qp_security(qp->qp_sec); 1891 kfree(qp); 1892 1893 return 0; 1894} 1895EXPORT_SYMBOL(ib_close_qp); 1896 1897static int __ib_destroy_shared_qp(struct ib_qp *qp) 1898{ 1899 struct ib_xrcd *xrcd; 1900 struct ib_qp *real_qp; 1901 int ret; 1902 1903 real_qp = qp->real_qp; 1904 xrcd = real_qp->xrcd; 1905 down_write(&xrcd->tgt_qps_rwsem); 1906 ib_close_qp(qp); 1907 if (atomic_read(&real_qp->usecnt) == 0) 1908 xa_erase(&xrcd->tgt_qps, real_qp->qp_num); 1909 else 1910 real_qp = NULL; 1911 up_write(&xrcd->tgt_qps_rwsem); 1912 1913 if (real_qp) { 1914 ret = ib_destroy_qp(real_qp); 1915 if (!ret) 1916 atomic_dec(&xrcd->usecnt); 1917 } 1918 1919 return 0; 1920} 1921 1922int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata) 1923{ 1924 const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr; 1925 const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr; 1926 struct ib_pd *pd; 1927 struct ib_cq *scq, *rcq; 1928 struct ib_srq *srq; 1929 struct ib_rwq_ind_table *ind_tbl; 1930 struct ib_qp_security *sec; 1931 int ret; 1932 1933 WARN_ON_ONCE(qp->mrs_used > 0); 1934 1935 if (atomic_read(&qp->usecnt)) 1936 return -EBUSY; 1937 1938 if (qp->real_qp != qp) 1939 return __ib_destroy_shared_qp(qp); 1940 1941 pd = qp->pd; 1942 scq = qp->send_cq; 1943 rcq = qp->recv_cq; 1944 srq = qp->srq; 1945 ind_tbl = qp->rwq_ind_tbl; 1946 sec = qp->qp_sec; 1947 if (sec) 1948 ib_destroy_qp_security_begin(sec); 1949 1950 if (!qp->uobject) 1951 rdma_rw_cleanup_mrs(qp); 1952 1953 rdma_counter_unbind_qp(qp, true); 1954 rdma_restrack_del(&qp->res); 1955 ret = qp->device->ops.destroy_qp(qp, udata); 1956 if (!ret) { 1957 if (alt_path_sgid_attr) 1958 rdma_put_gid_attr(alt_path_sgid_attr); 1959 if (av_sgid_attr) 1960 rdma_put_gid_attr(av_sgid_attr); 1961 if (pd) 1962 atomic_dec(&pd->usecnt); 1963 if (scq) 1964 atomic_dec(&scq->usecnt); 1965 if (rcq) 1966 atomic_dec(&rcq->usecnt); 1967 if (srq) 1968 atomic_dec(&srq->usecnt); 1969 if (ind_tbl) 1970 atomic_dec(&ind_tbl->usecnt); 1971 if (sec) 1972 ib_destroy_qp_security_end(sec); 1973 } else { 1974 if (sec) 1975 ib_destroy_qp_security_abort(sec); 1976 } 1977 1978 return ret; 1979} 1980EXPORT_SYMBOL(ib_destroy_qp_user); 1981 1982/* Completion queues */ 1983 1984struct ib_cq *__ib_create_cq(struct ib_device *device, 1985 ib_comp_handler comp_handler, 1986 void (*event_handler)(struct ib_event *, void *), 1987 void *cq_context, 1988 const struct ib_cq_init_attr *cq_attr, 1989 const char *caller) 1990{ 1991 struct ib_cq *cq; 1992 int ret; 1993 1994 cq = rdma_zalloc_drv_obj(device, ib_cq); 1995 if (!cq) 1996 return ERR_PTR(-ENOMEM); 1997 1998 cq->device = device; 1999 cq->uobject = NULL; 2000 cq->comp_handler = comp_handler; 2001 cq->event_handler = event_handler; 2002 cq->cq_context = cq_context; 2003 atomic_set(&cq->usecnt, 0); 2004 2005 rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ); 2006 rdma_restrack_set_name(&cq->res, caller); 2007 2008 ret = device->ops.create_cq(cq, cq_attr, NULL); 2009 if (ret) { 2010 rdma_restrack_put(&cq->res); 2011 kfree(cq); 2012 return ERR_PTR(ret); 2013 } 2014 2015 rdma_restrack_add(&cq->res); 2016 return cq; 2017} 2018EXPORT_SYMBOL(__ib_create_cq); 2019 2020int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period) 2021{ 2022 if (cq->shared) 2023 return -EOPNOTSUPP; 2024 2025 return cq->device->ops.modify_cq ? 2026 cq->device->ops.modify_cq(cq, cq_count, 2027 cq_period) : -EOPNOTSUPP; 2028} 2029EXPORT_SYMBOL(rdma_set_cq_moderation); 2030 2031int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata) 2032{ 2033 int ret; 2034 2035 if (WARN_ON_ONCE(cq->shared)) 2036 return -EOPNOTSUPP; 2037 2038 if (atomic_read(&cq->usecnt)) 2039 return -EBUSY; 2040 2041 ret = cq->device->ops.destroy_cq(cq, udata); 2042 if (ret) 2043 return ret; 2044 2045 rdma_restrack_del(&cq->res); 2046 kfree(cq); 2047 return ret; 2048} 2049EXPORT_SYMBOL(ib_destroy_cq_user); 2050 2051int ib_resize_cq(struct ib_cq *cq, int cqe) 2052{ 2053 if (cq->shared) 2054 return -EOPNOTSUPP; 2055 2056 return cq->device->ops.resize_cq ? 2057 cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP; 2058} 2059EXPORT_SYMBOL(ib_resize_cq); 2060 2061/* Memory regions */ 2062 2063struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, 2064 u64 virt_addr, int access_flags) 2065{ 2066 struct ib_mr *mr; 2067 2068 if (access_flags & IB_ACCESS_ON_DEMAND) { 2069 if (!(pd->device->attrs.device_cap_flags & 2070 IB_DEVICE_ON_DEMAND_PAGING)) { 2071 pr_debug("ODP support not available\n"); 2072 return ERR_PTR(-EINVAL); 2073 } 2074 } 2075 2076 mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr, 2077 access_flags, NULL); 2078 2079 if (IS_ERR(mr)) 2080 return mr; 2081 2082 mr->device = pd->device; 2083 mr->type = IB_MR_TYPE_USER; 2084 mr->pd = pd; 2085 mr->dm = NULL; 2086 atomic_inc(&pd->usecnt); 2087 mr->iova = virt_addr; 2088 mr->length = length; 2089 2090 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR); 2091 rdma_restrack_parent_name(&mr->res, &pd->res); 2092 rdma_restrack_add(&mr->res); 2093 2094 return mr; 2095} 2096EXPORT_SYMBOL(ib_reg_user_mr); 2097 2098int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice, 2099 u32 flags, struct ib_sge *sg_list, u32 num_sge) 2100{ 2101 if (!pd->device->ops.advise_mr) 2102 return -EOPNOTSUPP; 2103 2104 if (!num_sge) 2105 return 0; 2106 2107 return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge, 2108 NULL); 2109} 2110EXPORT_SYMBOL(ib_advise_mr); 2111 2112int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata) 2113{ 2114 struct ib_pd *pd = mr->pd; 2115 struct ib_dm *dm = mr->dm; 2116 struct ib_sig_attrs *sig_attrs = mr->sig_attrs; 2117 int ret; 2118 2119 trace_mr_dereg(mr); 2120 rdma_restrack_del(&mr->res); 2121 ret = mr->device->ops.dereg_mr(mr, udata); 2122 if (!ret) { 2123 atomic_dec(&pd->usecnt); 2124 if (dm) 2125 atomic_dec(&dm->usecnt); 2126 kfree(sig_attrs); 2127 } 2128 2129 return ret; 2130} 2131EXPORT_SYMBOL(ib_dereg_mr_user); 2132 2133/** 2134 * ib_alloc_mr() - Allocates a memory region 2135 * @pd: protection domain associated with the region 2136 * @mr_type: memory region type 2137 * @max_num_sg: maximum sg entries available for registration. 2138 * 2139 * Notes: 2140 * Memory registeration page/sg lists must not exceed max_num_sg. 2141 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed 2142 * max_num_sg * used_page_size. 2143 * 2144 */ 2145struct ib_mr *ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type, 2146 u32 max_num_sg) 2147{ 2148 struct ib_mr *mr; 2149 2150 if (!pd->device->ops.alloc_mr) { 2151 mr = ERR_PTR(-EOPNOTSUPP); 2152 goto out; 2153 } 2154 2155 if (mr_type == IB_MR_TYPE_INTEGRITY) { 2156 WARN_ON_ONCE(1); 2157 mr = ERR_PTR(-EINVAL); 2158 goto out; 2159 } 2160 2161 mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg); 2162 if (IS_ERR(mr)) 2163 goto out; 2164 2165 mr->device = pd->device; 2166 mr->pd = pd; 2167 mr->dm = NULL; 2168 mr->uobject = NULL; 2169 atomic_inc(&pd->usecnt); 2170 mr->need_inval = false; 2171 mr->type = mr_type; 2172 mr->sig_attrs = NULL; 2173 2174 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR); 2175 rdma_restrack_parent_name(&mr->res, &pd->res); 2176 rdma_restrack_add(&mr->res); 2177out: 2178 trace_mr_alloc(pd, mr_type, max_num_sg, mr); 2179 return mr; 2180} 2181EXPORT_SYMBOL(ib_alloc_mr); 2182 2183/** 2184 * ib_alloc_mr_integrity() - Allocates an integrity memory region 2185 * @pd: protection domain associated with the region 2186 * @max_num_data_sg: maximum data sg entries available for registration 2187 * @max_num_meta_sg: maximum metadata sg entries available for 2188 * registration 2189 * 2190 * Notes: 2191 * Memory registration page/sg lists must not exceed max_num_sg, 2192 * also the integrity page/sg lists must not exceed max_num_meta_sg. 2193 * 2194 */ 2195struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd, 2196 u32 max_num_data_sg, 2197 u32 max_num_meta_sg) 2198{ 2199 struct ib_mr *mr; 2200 struct ib_sig_attrs *sig_attrs; 2201 2202 if (!pd->device->ops.alloc_mr_integrity || 2203 !pd->device->ops.map_mr_sg_pi) { 2204 mr = ERR_PTR(-EOPNOTSUPP); 2205 goto out; 2206 } 2207 2208 if (!max_num_meta_sg) { 2209 mr = ERR_PTR(-EINVAL); 2210 goto out; 2211 } 2212 2213 sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL); 2214 if (!sig_attrs) { 2215 mr = ERR_PTR(-ENOMEM); 2216 goto out; 2217 } 2218 2219 mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg, 2220 max_num_meta_sg); 2221 if (IS_ERR(mr)) { 2222 kfree(sig_attrs); 2223 goto out; 2224 } 2225 2226 mr->device = pd->device; 2227 mr->pd = pd; 2228 mr->dm = NULL; 2229 mr->uobject = NULL; 2230 atomic_inc(&pd->usecnt); 2231 mr->need_inval = false; 2232 mr->type = IB_MR_TYPE_INTEGRITY; 2233 mr->sig_attrs = sig_attrs; 2234 2235 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR); 2236 rdma_restrack_parent_name(&mr->res, &pd->res); 2237 rdma_restrack_add(&mr->res); 2238out: 2239 trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr); 2240 return mr; 2241} 2242EXPORT_SYMBOL(ib_alloc_mr_integrity); 2243 2244/* Multicast groups */ 2245 2246static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid) 2247{ 2248 struct ib_qp_init_attr init_attr = {}; 2249 struct ib_qp_attr attr = {}; 2250 int num_eth_ports = 0; 2251 int port; 2252 2253 /* If QP state >= init, it is assigned to a port and we can check this 2254 * port only. 2255 */ 2256 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) { 2257 if (attr.qp_state >= IB_QPS_INIT) { 2258 if (rdma_port_get_link_layer(qp->device, attr.port_num) != 2259 IB_LINK_LAYER_INFINIBAND) 2260 return true; 2261 goto lid_check; 2262 } 2263 } 2264 2265 /* Can't get a quick answer, iterate over all ports */ 2266 for (port = 0; port < qp->device->phys_port_cnt; port++) 2267 if (rdma_port_get_link_layer(qp->device, port) != 2268 IB_LINK_LAYER_INFINIBAND) 2269 num_eth_ports++; 2270 2271 /* If we have at lease one Ethernet port, RoCE annex declares that 2272 * multicast LID should be ignored. We can't tell at this step if the 2273 * QP belongs to an IB or Ethernet port. 2274 */ 2275 if (num_eth_ports) 2276 return true; 2277 2278 /* If all the ports are IB, we can check according to IB spec. */ 2279lid_check: 2280 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) || 2281 lid == be16_to_cpu(IB_LID_PERMISSIVE)); 2282} 2283 2284int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid) 2285{ 2286 int ret; 2287 2288 if (!qp->device->ops.attach_mcast) 2289 return -EOPNOTSUPP; 2290 2291 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) || 2292 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid)) 2293 return -EINVAL; 2294 2295 ret = qp->device->ops.attach_mcast(qp, gid, lid); 2296 if (!ret) 2297 atomic_inc(&qp->usecnt); 2298 return ret; 2299} 2300EXPORT_SYMBOL(ib_attach_mcast); 2301 2302int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid) 2303{ 2304 int ret; 2305 2306 if (!qp->device->ops.detach_mcast) 2307 return -EOPNOTSUPP; 2308 2309 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) || 2310 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid)) 2311 return -EINVAL; 2312 2313 ret = qp->device->ops.detach_mcast(qp, gid, lid); 2314 if (!ret) 2315 atomic_dec(&qp->usecnt); 2316 return ret; 2317} 2318EXPORT_SYMBOL(ib_detach_mcast); 2319 2320/** 2321 * ib_alloc_xrcd_user - Allocates an XRC domain. 2322 * @device: The device on which to allocate the XRC domain. 2323 * @inode: inode to connect XRCD 2324 * @udata: Valid user data or NULL for kernel object 2325 */ 2326struct ib_xrcd *ib_alloc_xrcd_user(struct ib_device *device, 2327 struct inode *inode, struct ib_udata *udata) 2328{ 2329 struct ib_xrcd *xrcd; 2330 int ret; 2331 2332 if (!device->ops.alloc_xrcd) 2333 return ERR_PTR(-EOPNOTSUPP); 2334 2335 xrcd = rdma_zalloc_drv_obj(device, ib_xrcd); 2336 if (!xrcd) 2337 return ERR_PTR(-ENOMEM); 2338 2339 xrcd->device = device; 2340 xrcd->inode = inode; 2341 atomic_set(&xrcd->usecnt, 0); 2342 init_rwsem(&xrcd->tgt_qps_rwsem); 2343 xa_init(&xrcd->tgt_qps); 2344 2345 ret = device->ops.alloc_xrcd(xrcd, udata); 2346 if (ret) 2347 goto err; 2348 return xrcd; 2349err: 2350 kfree(xrcd); 2351 return ERR_PTR(ret); 2352} 2353EXPORT_SYMBOL(ib_alloc_xrcd_user); 2354 2355/** 2356 * ib_dealloc_xrcd_user - Deallocates an XRC domain. 2357 * @xrcd: The XRC domain to deallocate. 2358 * @udata: Valid user data or NULL for kernel object 2359 */ 2360int ib_dealloc_xrcd_user(struct ib_xrcd *xrcd, struct ib_udata *udata) 2361{ 2362 int ret; 2363 2364 if (atomic_read(&xrcd->usecnt)) 2365 return -EBUSY; 2366 2367 WARN_ON(!xa_empty(&xrcd->tgt_qps)); 2368 ret = xrcd->device->ops.dealloc_xrcd(xrcd, udata); 2369 if (ret) 2370 return ret; 2371 kfree(xrcd); 2372 return ret; 2373} 2374EXPORT_SYMBOL(ib_dealloc_xrcd_user); 2375 2376/** 2377 * ib_create_wq - Creates a WQ associated with the specified protection 2378 * domain. 2379 * @pd: The protection domain associated with the WQ. 2380 * @wq_attr: A list of initial attributes required to create the 2381 * WQ. If WQ creation succeeds, then the attributes are updated to 2382 * the actual capabilities of the created WQ. 2383 * 2384 * wq_attr->max_wr and wq_attr->max_sge determine 2385 * the requested size of the WQ, and set to the actual values allocated 2386 * on return. 2387 * If ib_create_wq() succeeds, then max_wr and max_sge will always be 2388 * at least as large as the requested values. 2389 */ 2390struct ib_wq *ib_create_wq(struct ib_pd *pd, 2391 struct ib_wq_init_attr *wq_attr) 2392{ 2393 struct ib_wq *wq; 2394 2395 if (!pd->device->ops.create_wq) 2396 return ERR_PTR(-EOPNOTSUPP); 2397 2398 wq = pd->device->ops.create_wq(pd, wq_attr, NULL); 2399 if (!IS_ERR(wq)) { 2400 wq->event_handler = wq_attr->event_handler; 2401 wq->wq_context = wq_attr->wq_context; 2402 wq->wq_type = wq_attr->wq_type; 2403 wq->cq = wq_attr->cq; 2404 wq->device = pd->device; 2405 wq->pd = pd; 2406 wq->uobject = NULL; 2407 atomic_inc(&pd->usecnt); 2408 atomic_inc(&wq_attr->cq->usecnt); 2409 atomic_set(&wq->usecnt, 0); 2410 } 2411 return wq; 2412} 2413EXPORT_SYMBOL(ib_create_wq); 2414 2415/** 2416 * ib_destroy_wq_user - Destroys the specified user WQ. 2417 * @wq: The WQ to destroy. 2418 * @udata: Valid user data 2419 */ 2420int ib_destroy_wq_user(struct ib_wq *wq, struct ib_udata *udata) 2421{ 2422 struct ib_cq *cq = wq->cq; 2423 struct ib_pd *pd = wq->pd; 2424 int ret; 2425 2426 if (atomic_read(&wq->usecnt)) 2427 return -EBUSY; 2428 2429 ret = wq->device->ops.destroy_wq(wq, udata); 2430 if (ret) 2431 return ret; 2432 2433 atomic_dec(&pd->usecnt); 2434 atomic_dec(&cq->usecnt); 2435 return ret; 2436} 2437EXPORT_SYMBOL(ib_destroy_wq_user); 2438 2439/** 2440 * ib_modify_wq - Modifies the specified WQ. 2441 * @wq: The WQ to modify. 2442 * @wq_attr: On input, specifies the WQ attributes to modify. 2443 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ 2444 * are being modified. 2445 * On output, the current values of selected WQ attributes are returned. 2446 */ 2447int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr, 2448 u32 wq_attr_mask) 2449{ 2450 int err; 2451 2452 if (!wq->device->ops.modify_wq) 2453 return -EOPNOTSUPP; 2454 2455 err = wq->device->ops.modify_wq(wq, wq_attr, wq_attr_mask, NULL); 2456 return err; 2457} 2458EXPORT_SYMBOL(ib_modify_wq); 2459 2460int ib_check_mr_status(struct ib_mr *mr, u32 check_mask, 2461 struct ib_mr_status *mr_status) 2462{ 2463 if (!mr->device->ops.check_mr_status) 2464 return -EOPNOTSUPP; 2465 2466 return mr->device->ops.check_mr_status(mr, check_mask, mr_status); 2467} 2468EXPORT_SYMBOL(ib_check_mr_status); 2469 2470int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port, 2471 int state) 2472{ 2473 if (!device->ops.set_vf_link_state) 2474 return -EOPNOTSUPP; 2475 2476 return device->ops.set_vf_link_state(device, vf, port, state); 2477} 2478EXPORT_SYMBOL(ib_set_vf_link_state); 2479 2480int ib_get_vf_config(struct ib_device *device, int vf, u8 port, 2481 struct ifla_vf_info *info) 2482{ 2483 if (!device->ops.get_vf_config) 2484 return -EOPNOTSUPP; 2485 2486 return device->ops.get_vf_config(device, vf, port, info); 2487} 2488EXPORT_SYMBOL(ib_get_vf_config); 2489 2490int ib_get_vf_stats(struct ib_device *device, int vf, u8 port, 2491 struct ifla_vf_stats *stats) 2492{ 2493 if (!device->ops.get_vf_stats) 2494 return -EOPNOTSUPP; 2495 2496 return device->ops.get_vf_stats(device, vf, port, stats); 2497} 2498EXPORT_SYMBOL(ib_get_vf_stats); 2499 2500int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid, 2501 int type) 2502{ 2503 if (!device->ops.set_vf_guid) 2504 return -EOPNOTSUPP; 2505 2506 return device->ops.set_vf_guid(device, vf, port, guid, type); 2507} 2508EXPORT_SYMBOL(ib_set_vf_guid); 2509 2510int ib_get_vf_guid(struct ib_device *device, int vf, u8 port, 2511 struct ifla_vf_guid *node_guid, 2512 struct ifla_vf_guid *port_guid) 2513{ 2514 if (!device->ops.get_vf_guid) 2515 return -EOPNOTSUPP; 2516 2517 return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid); 2518} 2519EXPORT_SYMBOL(ib_get_vf_guid); 2520/** 2521 * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection 2522 * information) and set an appropriate memory region for registration. 2523 * @mr: memory region 2524 * @data_sg: dma mapped scatterlist for data 2525 * @data_sg_nents: number of entries in data_sg 2526 * @data_sg_offset: offset in bytes into data_sg 2527 * @meta_sg: dma mapped scatterlist for metadata 2528 * @meta_sg_nents: number of entries in meta_sg 2529 * @meta_sg_offset: offset in bytes into meta_sg 2530 * @page_size: page vector desired page size 2531 * 2532 * Constraints: 2533 * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY. 2534 * 2535 * Return: 0 on success. 2536 * 2537 * After this completes successfully, the memory region 2538 * is ready for registration. 2539 */ 2540int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg, 2541 int data_sg_nents, unsigned int *data_sg_offset, 2542 struct scatterlist *meta_sg, int meta_sg_nents, 2543 unsigned int *meta_sg_offset, unsigned int page_size) 2544{ 2545 if (unlikely(!mr->device->ops.map_mr_sg_pi || 2546 WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY))) 2547 return -EOPNOTSUPP; 2548 2549 mr->page_size = page_size; 2550 2551 return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents, 2552 data_sg_offset, meta_sg, 2553 meta_sg_nents, meta_sg_offset); 2554} 2555EXPORT_SYMBOL(ib_map_mr_sg_pi); 2556 2557/** 2558 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list 2559 * and set it the memory region. 2560 * @mr: memory region 2561 * @sg: dma mapped scatterlist 2562 * @sg_nents: number of entries in sg 2563 * @sg_offset: offset in bytes into sg 2564 * @page_size: page vector desired page size 2565 * 2566 * Constraints: 2567 * 2568 * - The first sg element is allowed to have an offset. 2569 * - Each sg element must either be aligned to page_size or virtually 2570 * contiguous to the previous element. In case an sg element has a 2571 * non-contiguous offset, the mapping prefix will not include it. 2572 * - The last sg element is allowed to have length less than page_size. 2573 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size 2574 * then only max_num_sg entries will be mapped. 2575 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these 2576 * constraints holds and the page_size argument is ignored. 2577 * 2578 * Returns the number of sg elements that were mapped to the memory region. 2579 * 2580 * After this completes successfully, the memory region 2581 * is ready for registration. 2582 */ 2583int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents, 2584 unsigned int *sg_offset, unsigned int page_size) 2585{ 2586 if (unlikely(!mr->device->ops.map_mr_sg)) 2587 return -EOPNOTSUPP; 2588 2589 mr->page_size = page_size; 2590 2591 return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset); 2592} 2593EXPORT_SYMBOL(ib_map_mr_sg); 2594 2595/** 2596 * ib_sg_to_pages() - Convert the largest prefix of a sg list 2597 * to a page vector 2598 * @mr: memory region 2599 * @sgl: dma mapped scatterlist 2600 * @sg_nents: number of entries in sg 2601 * @sg_offset_p: ==== ======================================================= 2602 * IN start offset in bytes into sg 2603 * OUT offset in bytes for element n of the sg of the first 2604 * byte that has not been processed where n is the return 2605 * value of this function. 2606 * ==== ======================================================= 2607 * @set_page: driver page assignment function pointer 2608 * 2609 * Core service helper for drivers to convert the largest 2610 * prefix of given sg list to a page vector. The sg list 2611 * prefix converted is the prefix that meet the requirements 2612 * of ib_map_mr_sg. 2613 * 2614 * Returns the number of sg elements that were assigned to 2615 * a page vector. 2616 */ 2617int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents, 2618 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64)) 2619{ 2620 struct scatterlist *sg; 2621 u64 last_end_dma_addr = 0; 2622 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0; 2623 unsigned int last_page_off = 0; 2624 u64 page_mask = ~((u64)mr->page_size - 1); 2625 int i, ret; 2626 2627 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0]))) 2628 return -EINVAL; 2629 2630 mr->iova = sg_dma_address(&sgl[0]) + sg_offset; 2631 mr->length = 0; 2632 2633 for_each_sg(sgl, sg, sg_nents, i) { 2634 u64 dma_addr = sg_dma_address(sg) + sg_offset; 2635 u64 prev_addr = dma_addr; 2636 unsigned int dma_len = sg_dma_len(sg) - sg_offset; 2637 u64 end_dma_addr = dma_addr + dma_len; 2638 u64 page_addr = dma_addr & page_mask; 2639 2640 /* 2641 * For the second and later elements, check whether either the 2642 * end of element i-1 or the start of element i is not aligned 2643 * on a page boundary. 2644 */ 2645 if (i && (last_page_off != 0 || page_addr != dma_addr)) { 2646 /* Stop mapping if there is a gap. */ 2647 if (last_end_dma_addr != dma_addr) 2648 break; 2649 2650 /* 2651 * Coalesce this element with the last. If it is small 2652 * enough just update mr->length. Otherwise start 2653 * mapping from the next page. 2654 */ 2655 goto next_page; 2656 } 2657 2658 do { 2659 ret = set_page(mr, page_addr); 2660 if (unlikely(ret < 0)) { 2661 sg_offset = prev_addr - sg_dma_address(sg); 2662 mr->length += prev_addr - dma_addr; 2663 if (sg_offset_p) 2664 *sg_offset_p = sg_offset; 2665 return i || sg_offset ? i : ret; 2666 } 2667 prev_addr = page_addr; 2668next_page: 2669 page_addr += mr->page_size; 2670 } while (page_addr < end_dma_addr); 2671 2672 mr->length += dma_len; 2673 last_end_dma_addr = end_dma_addr; 2674 last_page_off = end_dma_addr & ~page_mask; 2675 2676 sg_offset = 0; 2677 } 2678 2679 if (sg_offset_p) 2680 *sg_offset_p = 0; 2681 return i; 2682} 2683EXPORT_SYMBOL(ib_sg_to_pages); 2684 2685struct ib_drain_cqe { 2686 struct ib_cqe cqe; 2687 struct completion done; 2688}; 2689 2690static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc) 2691{ 2692 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe, 2693 cqe); 2694 2695 complete(&cqe->done); 2696} 2697 2698/* 2699 * Post a WR and block until its completion is reaped for the SQ. 2700 */ 2701static void __ib_drain_sq(struct ib_qp *qp) 2702{ 2703 struct ib_cq *cq = qp->send_cq; 2704 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 2705 struct ib_drain_cqe sdrain; 2706 struct ib_rdma_wr swr = { 2707 .wr = { 2708 .next = NULL, 2709 { .wr_cqe = &sdrain.cqe, }, 2710 .opcode = IB_WR_RDMA_WRITE, 2711 }, 2712 }; 2713 int ret; 2714 2715 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 2716 if (ret) { 2717 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 2718 return; 2719 } 2720 2721 sdrain.cqe.done = ib_drain_qp_done; 2722 init_completion(&sdrain.done); 2723 2724 ret = ib_post_send(qp, &swr.wr, NULL); 2725 if (ret) { 2726 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 2727 return; 2728 } 2729 2730 if (cq->poll_ctx == IB_POLL_DIRECT) 2731 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0) 2732 ib_process_cq_direct(cq, -1); 2733 else 2734 wait_for_completion(&sdrain.done); 2735} 2736 2737/* 2738 * Post a WR and block until its completion is reaped for the RQ. 2739 */ 2740static void __ib_drain_rq(struct ib_qp *qp) 2741{ 2742 struct ib_cq *cq = qp->recv_cq; 2743 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 2744 struct ib_drain_cqe rdrain; 2745 struct ib_recv_wr rwr = {}; 2746 int ret; 2747 2748 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 2749 if (ret) { 2750 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 2751 return; 2752 } 2753 2754 rwr.wr_cqe = &rdrain.cqe; 2755 rdrain.cqe.done = ib_drain_qp_done; 2756 init_completion(&rdrain.done); 2757 2758 ret = ib_post_recv(qp, &rwr, NULL); 2759 if (ret) { 2760 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 2761 return; 2762 } 2763 2764 if (cq->poll_ctx == IB_POLL_DIRECT) 2765 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0) 2766 ib_process_cq_direct(cq, -1); 2767 else 2768 wait_for_completion(&rdrain.done); 2769} 2770 2771/** 2772 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the 2773 * application. 2774 * @qp: queue pair to drain 2775 * 2776 * If the device has a provider-specific drain function, then 2777 * call that. Otherwise call the generic drain function 2778 * __ib_drain_sq(). 2779 * 2780 * The caller must: 2781 * 2782 * ensure there is room in the CQ and SQ for the drain work request and 2783 * completion. 2784 * 2785 * allocate the CQ using ib_alloc_cq(). 2786 * 2787 * ensure that there are no other contexts that are posting WRs concurrently. 2788 * Otherwise the drain is not guaranteed. 2789 */ 2790void ib_drain_sq(struct ib_qp *qp) 2791{ 2792 if (qp->device->ops.drain_sq) 2793 qp->device->ops.drain_sq(qp); 2794 else 2795 __ib_drain_sq(qp); 2796 trace_cq_drain_complete(qp->send_cq); 2797} 2798EXPORT_SYMBOL(ib_drain_sq); 2799 2800/** 2801 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the 2802 * application. 2803 * @qp: queue pair to drain 2804 * 2805 * If the device has a provider-specific drain function, then 2806 * call that. Otherwise call the generic drain function 2807 * __ib_drain_rq(). 2808 * 2809 * The caller must: 2810 * 2811 * ensure there is room in the CQ and RQ for the drain work request and 2812 * completion. 2813 * 2814 * allocate the CQ using ib_alloc_cq(). 2815 * 2816 * ensure that there are no other contexts that are posting WRs concurrently. 2817 * Otherwise the drain is not guaranteed. 2818 */ 2819void ib_drain_rq(struct ib_qp *qp) 2820{ 2821 if (qp->device->ops.drain_rq) 2822 qp->device->ops.drain_rq(qp); 2823 else 2824 __ib_drain_rq(qp); 2825 trace_cq_drain_complete(qp->recv_cq); 2826} 2827EXPORT_SYMBOL(ib_drain_rq); 2828 2829/** 2830 * ib_drain_qp() - Block until all CQEs have been consumed by the 2831 * application on both the RQ and SQ. 2832 * @qp: queue pair to drain 2833 * 2834 * The caller must: 2835 * 2836 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests 2837 * and completions. 2838 * 2839 * allocate the CQs using ib_alloc_cq(). 2840 * 2841 * ensure that there are no other contexts that are posting WRs concurrently. 2842 * Otherwise the drain is not guaranteed. 2843 */ 2844void ib_drain_qp(struct ib_qp *qp) 2845{ 2846 ib_drain_sq(qp); 2847 if (!qp->srq) 2848 ib_drain_rq(qp); 2849} 2850EXPORT_SYMBOL(ib_drain_qp); 2851 2852struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num, 2853 enum rdma_netdev_t type, const char *name, 2854 unsigned char name_assign_type, 2855 void (*setup)(struct net_device *)) 2856{ 2857 struct rdma_netdev_alloc_params params; 2858 struct net_device *netdev; 2859 int rc; 2860 2861 if (!device->ops.rdma_netdev_get_params) 2862 return ERR_PTR(-EOPNOTSUPP); 2863 2864 rc = device->ops.rdma_netdev_get_params(device, port_num, type, 2865 ¶ms); 2866 if (rc) 2867 return ERR_PTR(rc); 2868 2869 netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type, 2870 setup, params.txqs, params.rxqs); 2871 if (!netdev) 2872 return ERR_PTR(-ENOMEM); 2873 2874 return netdev; 2875} 2876EXPORT_SYMBOL(rdma_alloc_netdev); 2877 2878int rdma_init_netdev(struct ib_device *device, u8 port_num, 2879 enum rdma_netdev_t type, const char *name, 2880 unsigned char name_assign_type, 2881 void (*setup)(struct net_device *), 2882 struct net_device *netdev) 2883{ 2884 struct rdma_netdev_alloc_params params; 2885 int rc; 2886 2887 if (!device->ops.rdma_netdev_get_params) 2888 return -EOPNOTSUPP; 2889 2890 rc = device->ops.rdma_netdev_get_params(device, port_num, type, 2891 ¶ms); 2892 if (rc) 2893 return rc; 2894 2895 return params.initialize_rdma_netdev(device, port_num, 2896 netdev, params.param); 2897} 2898EXPORT_SYMBOL(rdma_init_netdev); 2899 2900void __rdma_block_iter_start(struct ib_block_iter *biter, 2901 struct scatterlist *sglist, unsigned int nents, 2902 unsigned long pgsz) 2903{ 2904 memset(biter, 0, sizeof(struct ib_block_iter)); 2905 biter->__sg = sglist; 2906 biter->__sg_nents = nents; 2907 2908 /* Driver provides best block size to use */ 2909 biter->__pg_bit = __fls(pgsz); 2910} 2911EXPORT_SYMBOL(__rdma_block_iter_start); 2912 2913bool __rdma_block_iter_next(struct ib_block_iter *biter) 2914{ 2915 unsigned int block_offset; 2916 unsigned int sg_delta; 2917 2918 if (!biter->__sg_nents || !biter->__sg) 2919 return false; 2920 2921 biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance; 2922 block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1); 2923 sg_delta = BIT_ULL(biter->__pg_bit) - block_offset; 2924 2925 if (sg_dma_len(biter->__sg) - biter->__sg_advance > sg_delta) { 2926 biter->__sg_advance += sg_delta; 2927 } else { 2928 biter->__sg_advance = 0; 2929 biter->__sg = sg_next(biter->__sg); 2930 biter->__sg_nents--; 2931 } 2932 2933 return true; 2934} 2935EXPORT_SYMBOL(__rdma_block_iter_next); 2936