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