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 54 #include "core_priv.h" 55 #include <trace/events/rdma_core.h> 56 57 #include <trace/events/rdma_core.h> 58 59 static int ib_resolve_eth_dmac(struct ib_device *device, 60 struct rdma_ah_attr *ah_attr); 61 62 static const char * const ib_events[] = { 63 [IB_EVENT_CQ_ERR] = "CQ error", 64 [IB_EVENT_QP_FATAL] = "QP fatal error", 65 [IB_EVENT_QP_REQ_ERR] = "QP request error", 66 [IB_EVENT_QP_ACCESS_ERR] = "QP access error", 67 [IB_EVENT_COMM_EST] = "communication established", 68 [IB_EVENT_SQ_DRAINED] = "send queue drained", 69 [IB_EVENT_PATH_MIG] = "path migration successful", 70 [IB_EVENT_PATH_MIG_ERR] = "path migration error", 71 [IB_EVENT_DEVICE_FATAL] = "device fatal error", 72 [IB_EVENT_PORT_ACTIVE] = "port active", 73 [IB_EVENT_PORT_ERR] = "port error", 74 [IB_EVENT_LID_CHANGE] = "LID change", 75 [IB_EVENT_PKEY_CHANGE] = "P_key change", 76 [IB_EVENT_SM_CHANGE] = "SM change", 77 [IB_EVENT_SRQ_ERR] = "SRQ error", 78 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached", 79 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached", 80 [IB_EVENT_CLIENT_REREGISTER] = "client reregister", 81 [IB_EVENT_GID_CHANGE] = "GID changed", 82 }; 83 84 const char *__attribute_const__ ib_event_msg(enum ib_event_type event) 85 { 86 size_t index = event; 87 88 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ? 89 ib_events[index] : "unrecognized event"; 90 } 91 EXPORT_SYMBOL(ib_event_msg); 92 93 static const char * const wc_statuses[] = { 94 [IB_WC_SUCCESS] = "success", 95 [IB_WC_LOC_LEN_ERR] = "local length error", 96 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error", 97 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error", 98 [IB_WC_LOC_PROT_ERR] = "local protection error", 99 [IB_WC_WR_FLUSH_ERR] = "WR flushed", 100 [IB_WC_MW_BIND_ERR] = "memory management operation error", 101 [IB_WC_BAD_RESP_ERR] = "bad response error", 102 [IB_WC_LOC_ACCESS_ERR] = "local access error", 103 [IB_WC_REM_INV_REQ_ERR] = "invalid request error", 104 [IB_WC_REM_ACCESS_ERR] = "remote access error", 105 [IB_WC_REM_OP_ERR] = "remote operation error", 106 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded", 107 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded", 108 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error", 109 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request", 110 [IB_WC_REM_ABORT_ERR] = "operation aborted", 111 [IB_WC_INV_EECN_ERR] = "invalid EE context number", 112 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state", 113 [IB_WC_FATAL_ERR] = "fatal error", 114 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error", 115 [IB_WC_GENERAL_ERR] = "general error", 116 }; 117 118 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status) 119 { 120 size_t index = status; 121 122 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ? 123 wc_statuses[index] : "unrecognized status"; 124 } 125 EXPORT_SYMBOL(ib_wc_status_msg); 126 127 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate) 128 { 129 switch (rate) { 130 case IB_RATE_2_5_GBPS: return 1; 131 case IB_RATE_5_GBPS: return 2; 132 case IB_RATE_10_GBPS: return 4; 133 case IB_RATE_20_GBPS: return 8; 134 case IB_RATE_30_GBPS: return 12; 135 case IB_RATE_40_GBPS: return 16; 136 case IB_RATE_60_GBPS: return 24; 137 case IB_RATE_80_GBPS: return 32; 138 case IB_RATE_120_GBPS: return 48; 139 case IB_RATE_14_GBPS: return 6; 140 case IB_RATE_56_GBPS: return 22; 141 case IB_RATE_112_GBPS: return 45; 142 case IB_RATE_168_GBPS: return 67; 143 case IB_RATE_25_GBPS: return 10; 144 case IB_RATE_100_GBPS: return 40; 145 case IB_RATE_200_GBPS: return 80; 146 case IB_RATE_300_GBPS: return 120; 147 case IB_RATE_28_GBPS: return 11; 148 case IB_RATE_50_GBPS: return 20; 149 case IB_RATE_400_GBPS: return 160; 150 case IB_RATE_600_GBPS: return 240; 151 default: return -1; 152 } 153 } 154 EXPORT_SYMBOL(ib_rate_to_mult); 155 156 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult) 157 { 158 switch (mult) { 159 case 1: return IB_RATE_2_5_GBPS; 160 case 2: return IB_RATE_5_GBPS; 161 case 4: return IB_RATE_10_GBPS; 162 case 8: return IB_RATE_20_GBPS; 163 case 12: return IB_RATE_30_GBPS; 164 case 16: return IB_RATE_40_GBPS; 165 case 24: return IB_RATE_60_GBPS; 166 case 32: return IB_RATE_80_GBPS; 167 case 48: return IB_RATE_120_GBPS; 168 case 6: return IB_RATE_14_GBPS; 169 case 22: return IB_RATE_56_GBPS; 170 case 45: return IB_RATE_112_GBPS; 171 case 67: return IB_RATE_168_GBPS; 172 case 10: return IB_RATE_25_GBPS; 173 case 40: return IB_RATE_100_GBPS; 174 case 80: return IB_RATE_200_GBPS; 175 case 120: return IB_RATE_300_GBPS; 176 case 11: return IB_RATE_28_GBPS; 177 case 20: return IB_RATE_50_GBPS; 178 case 160: return IB_RATE_400_GBPS; 179 case 240: return IB_RATE_600_GBPS; 180 default: return IB_RATE_PORT_CURRENT; 181 } 182 } 183 EXPORT_SYMBOL(mult_to_ib_rate); 184 185 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate) 186 { 187 switch (rate) { 188 case IB_RATE_2_5_GBPS: return 2500; 189 case IB_RATE_5_GBPS: return 5000; 190 case IB_RATE_10_GBPS: return 10000; 191 case IB_RATE_20_GBPS: return 20000; 192 case IB_RATE_30_GBPS: return 30000; 193 case IB_RATE_40_GBPS: return 40000; 194 case IB_RATE_60_GBPS: return 60000; 195 case IB_RATE_80_GBPS: return 80000; 196 case IB_RATE_120_GBPS: return 120000; 197 case IB_RATE_14_GBPS: return 14062; 198 case IB_RATE_56_GBPS: return 56250; 199 case IB_RATE_112_GBPS: return 112500; 200 case IB_RATE_168_GBPS: return 168750; 201 case IB_RATE_25_GBPS: return 25781; 202 case IB_RATE_100_GBPS: return 103125; 203 case IB_RATE_200_GBPS: return 206250; 204 case IB_RATE_300_GBPS: return 309375; 205 case IB_RATE_28_GBPS: return 28125; 206 case IB_RATE_50_GBPS: return 53125; 207 case IB_RATE_400_GBPS: return 425000; 208 case IB_RATE_600_GBPS: return 637500; 209 default: return -1; 210 } 211 } 212 EXPORT_SYMBOL(ib_rate_to_mbps); 213 214 __attribute_const__ enum rdma_transport_type 215 rdma_node_get_transport(unsigned int node_type) 216 { 217 218 if (node_type == RDMA_NODE_USNIC) 219 return RDMA_TRANSPORT_USNIC; 220 if (node_type == RDMA_NODE_USNIC_UDP) 221 return RDMA_TRANSPORT_USNIC_UDP; 222 if (node_type == RDMA_NODE_RNIC) 223 return RDMA_TRANSPORT_IWARP; 224 if (node_type == RDMA_NODE_UNSPECIFIED) 225 return RDMA_TRANSPORT_UNSPECIFIED; 226 227 return RDMA_TRANSPORT_IB; 228 } 229 EXPORT_SYMBOL(rdma_node_get_transport); 230 231 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 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 pd->res.type = RDMA_RESTRACK_PD; 277 rdma_restrack_set_task(&pd->res, caller); 278 279 ret = device->ops.alloc_pd(pd, NULL); 280 if (ret) { 281 kfree(pd); 282 return ERR_PTR(ret); 283 } 284 rdma_restrack_kadd(&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 } 322 EXPORT_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 */ 333 void 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 rdma_restrack_del(&pd->res); 348 pd->device->ops.dealloc_pd(pd, udata); 349 kfree(pd); 350 } 351 EXPORT_SYMBOL(ib_dealloc_pd_user); 352 353 /* Address handles */ 354 355 /** 356 * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination. 357 * @dest: Pointer to destination ah_attr. Contents of the destination 358 * pointer is assumed to be invalid and attribute are overwritten. 359 * @src: Pointer to source ah_attr. 360 */ 361 void rdma_copy_ah_attr(struct rdma_ah_attr *dest, 362 const struct rdma_ah_attr *src) 363 { 364 *dest = *src; 365 if (dest->grh.sgid_attr) 366 rdma_hold_gid_attr(dest->grh.sgid_attr); 367 } 368 EXPORT_SYMBOL(rdma_copy_ah_attr); 369 370 /** 371 * rdma_replace_ah_attr - Replace valid ah_attr with new new one. 372 * @old: Pointer to existing ah_attr which needs to be replaced. 373 * old is assumed to be valid or zero'd 374 * @new: Pointer to the new ah_attr. 375 * 376 * rdma_replace_ah_attr() first releases any reference in the old ah_attr if 377 * old the ah_attr is valid; after that it copies the new attribute and holds 378 * the reference to the replaced ah_attr. 379 */ 380 void rdma_replace_ah_attr(struct rdma_ah_attr *old, 381 const struct rdma_ah_attr *new) 382 { 383 rdma_destroy_ah_attr(old); 384 *old = *new; 385 if (old->grh.sgid_attr) 386 rdma_hold_gid_attr(old->grh.sgid_attr); 387 } 388 EXPORT_SYMBOL(rdma_replace_ah_attr); 389 390 /** 391 * rdma_move_ah_attr - Move ah_attr pointed by source to destination. 392 * @dest: Pointer to destination ah_attr to copy to. 393 * dest is assumed to be valid or zero'd 394 * @src: Pointer to the new ah_attr. 395 * 396 * rdma_move_ah_attr() first releases any reference in the destination ah_attr 397 * if it is valid. This also transfers ownership of internal references from 398 * src to dest, making src invalid in the process. No new reference of the src 399 * ah_attr is taken. 400 */ 401 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src) 402 { 403 rdma_destroy_ah_attr(dest); 404 *dest = *src; 405 src->grh.sgid_attr = NULL; 406 } 407 EXPORT_SYMBOL(rdma_move_ah_attr); 408 409 /* 410 * Validate that the rdma_ah_attr is valid for the device before passing it 411 * off to the driver. 412 */ 413 static int rdma_check_ah_attr(struct ib_device *device, 414 struct rdma_ah_attr *ah_attr) 415 { 416 if (!rdma_is_port_valid(device, ah_attr->port_num)) 417 return -EINVAL; 418 419 if ((rdma_is_grh_required(device, ah_attr->port_num) || 420 ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) && 421 !(ah_attr->ah_flags & IB_AH_GRH)) 422 return -EINVAL; 423 424 if (ah_attr->grh.sgid_attr) { 425 /* 426 * Make sure the passed sgid_attr is consistent with the 427 * parameters 428 */ 429 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index || 430 ah_attr->grh.sgid_attr->port_num != ah_attr->port_num) 431 return -EINVAL; 432 } 433 return 0; 434 } 435 436 /* 437 * If the ah requires a GRH then ensure that sgid_attr pointer is filled in. 438 * On success the caller is responsible to call rdma_unfill_sgid_attr(). 439 */ 440 static int rdma_fill_sgid_attr(struct ib_device *device, 441 struct rdma_ah_attr *ah_attr, 442 const struct ib_gid_attr **old_sgid_attr) 443 { 444 const struct ib_gid_attr *sgid_attr; 445 struct ib_global_route *grh; 446 int ret; 447 448 *old_sgid_attr = ah_attr->grh.sgid_attr; 449 450 ret = rdma_check_ah_attr(device, ah_attr); 451 if (ret) 452 return ret; 453 454 if (!(ah_attr->ah_flags & IB_AH_GRH)) 455 return 0; 456 457 grh = rdma_ah_retrieve_grh(ah_attr); 458 if (grh->sgid_attr) 459 return 0; 460 461 sgid_attr = 462 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index); 463 if (IS_ERR(sgid_attr)) 464 return PTR_ERR(sgid_attr); 465 466 /* Move ownerhip of the kref into the ah_attr */ 467 grh->sgid_attr = sgid_attr; 468 return 0; 469 } 470 471 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr, 472 const struct ib_gid_attr *old_sgid_attr) 473 { 474 /* 475 * Fill didn't change anything, the caller retains ownership of 476 * whatever it passed 477 */ 478 if (ah_attr->grh.sgid_attr == old_sgid_attr) 479 return; 480 481 /* 482 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller 483 * doesn't see any change in the rdma_ah_attr. If we get here 484 * old_sgid_attr is NULL. 485 */ 486 rdma_destroy_ah_attr(ah_attr); 487 } 488 489 static const struct ib_gid_attr * 490 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr, 491 const struct ib_gid_attr *old_attr) 492 { 493 if (old_attr) 494 rdma_put_gid_attr(old_attr); 495 if (ah_attr->ah_flags & IB_AH_GRH) { 496 rdma_hold_gid_attr(ah_attr->grh.sgid_attr); 497 return ah_attr->grh.sgid_attr; 498 } 499 return NULL; 500 } 501 502 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd, 503 struct rdma_ah_attr *ah_attr, 504 u32 flags, 505 struct ib_udata *udata) 506 { 507 struct ib_device *device = pd->device; 508 struct ib_ah *ah; 509 int ret; 510 511 might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE); 512 513 if (!device->ops.create_ah) 514 return ERR_PTR(-EOPNOTSUPP); 515 516 ah = rdma_zalloc_drv_obj_gfp( 517 device, ib_ah, 518 (flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC); 519 if (!ah) 520 return ERR_PTR(-ENOMEM); 521 522 ah->device = device; 523 ah->pd = pd; 524 ah->type = ah_attr->type; 525 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL); 526 527 ret = device->ops.create_ah(ah, ah_attr, flags, udata); 528 if (ret) { 529 kfree(ah); 530 return ERR_PTR(ret); 531 } 532 533 atomic_inc(&pd->usecnt); 534 return ah; 535 } 536 537 /** 538 * rdma_create_ah - Creates an address handle for the 539 * given address vector. 540 * @pd: The protection domain associated with the address handle. 541 * @ah_attr: The attributes of the address vector. 542 * @flags: Create address handle flags (see enum rdma_create_ah_flags). 543 * 544 * It returns 0 on success and returns appropriate error code on error. 545 * The address handle is used to reference a local or global destination 546 * in all UD QP post sends. 547 */ 548 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr, 549 u32 flags) 550 { 551 const struct ib_gid_attr *old_sgid_attr; 552 struct ib_ah *ah; 553 int ret; 554 555 ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr); 556 if (ret) 557 return ERR_PTR(ret); 558 559 ah = _rdma_create_ah(pd, ah_attr, flags, NULL); 560 561 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 562 return ah; 563 } 564 EXPORT_SYMBOL(rdma_create_ah); 565 566 /** 567 * rdma_create_user_ah - Creates an address handle for the 568 * given address vector. 569 * It resolves destination mac address for ah attribute of RoCE type. 570 * @pd: The protection domain associated with the address handle. 571 * @ah_attr: The attributes of the address vector. 572 * @udata: pointer to user's input output buffer information need by 573 * provider driver. 574 * 575 * It returns 0 on success and returns appropriate error code on error. 576 * The address handle is used to reference a local or global destination 577 * in all UD QP post sends. 578 */ 579 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd, 580 struct rdma_ah_attr *ah_attr, 581 struct ib_udata *udata) 582 { 583 const struct ib_gid_attr *old_sgid_attr; 584 struct ib_ah *ah; 585 int err; 586 587 err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr); 588 if (err) 589 return ERR_PTR(err); 590 591 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) { 592 err = ib_resolve_eth_dmac(pd->device, ah_attr); 593 if (err) { 594 ah = ERR_PTR(err); 595 goto out; 596 } 597 } 598 599 ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE, udata); 600 601 out: 602 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 603 return ah; 604 } 605 EXPORT_SYMBOL(rdma_create_user_ah); 606 607 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr) 608 { 609 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh; 610 struct iphdr ip4h_checked; 611 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh; 612 613 /* If it's IPv6, the version must be 6, otherwise, the first 614 * 20 bytes (before the IPv4 header) are garbled. 615 */ 616 if (ip6h->version != 6) 617 return (ip4h->version == 4) ? 4 : 0; 618 /* version may be 6 or 4 because the first 20 bytes could be garbled */ 619 620 /* RoCE v2 requires no options, thus header length 621 * must be 5 words 622 */ 623 if (ip4h->ihl != 5) 624 return 6; 625 626 /* Verify checksum. 627 * We can't write on scattered buffers so we need to copy to 628 * temp buffer. 629 */ 630 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked)); 631 ip4h_checked.check = 0; 632 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5); 633 /* if IPv4 header checksum is OK, believe it */ 634 if (ip4h->check == ip4h_checked.check) 635 return 4; 636 return 6; 637 } 638 EXPORT_SYMBOL(ib_get_rdma_header_version); 639 640 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device, 641 u8 port_num, 642 const struct ib_grh *grh) 643 { 644 int grh_version; 645 646 if (rdma_protocol_ib(device, port_num)) 647 return RDMA_NETWORK_IB; 648 649 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh); 650 651 if (grh_version == 4) 652 return RDMA_NETWORK_IPV4; 653 654 if (grh->next_hdr == IPPROTO_UDP) 655 return RDMA_NETWORK_IPV6; 656 657 return RDMA_NETWORK_ROCE_V1; 658 } 659 660 struct find_gid_index_context { 661 u16 vlan_id; 662 enum ib_gid_type gid_type; 663 }; 664 665 static bool find_gid_index(const union ib_gid *gid, 666 const struct ib_gid_attr *gid_attr, 667 void *context) 668 { 669 struct find_gid_index_context *ctx = context; 670 u16 vlan_id = 0xffff; 671 int ret; 672 673 if (ctx->gid_type != gid_attr->gid_type) 674 return false; 675 676 ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL); 677 if (ret) 678 return false; 679 680 return ctx->vlan_id == vlan_id; 681 } 682 683 static const struct ib_gid_attr * 684 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num, 685 u16 vlan_id, const union ib_gid *sgid, 686 enum ib_gid_type gid_type) 687 { 688 struct find_gid_index_context context = {.vlan_id = vlan_id, 689 .gid_type = gid_type}; 690 691 return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index, 692 &context); 693 } 694 695 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr, 696 enum rdma_network_type net_type, 697 union ib_gid *sgid, union ib_gid *dgid) 698 { 699 struct sockaddr_in src_in; 700 struct sockaddr_in dst_in; 701 __be32 src_saddr, dst_saddr; 702 703 if (!sgid || !dgid) 704 return -EINVAL; 705 706 if (net_type == RDMA_NETWORK_IPV4) { 707 memcpy(&src_in.sin_addr.s_addr, 708 &hdr->roce4grh.saddr, 4); 709 memcpy(&dst_in.sin_addr.s_addr, 710 &hdr->roce4grh.daddr, 4); 711 src_saddr = src_in.sin_addr.s_addr; 712 dst_saddr = dst_in.sin_addr.s_addr; 713 ipv6_addr_set_v4mapped(src_saddr, 714 (struct in6_addr *)sgid); 715 ipv6_addr_set_v4mapped(dst_saddr, 716 (struct in6_addr *)dgid); 717 return 0; 718 } else if (net_type == RDMA_NETWORK_IPV6 || 719 net_type == RDMA_NETWORK_IB) { 720 *dgid = hdr->ibgrh.dgid; 721 *sgid = hdr->ibgrh.sgid; 722 return 0; 723 } else { 724 return -EINVAL; 725 } 726 } 727 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr); 728 729 /* Resolve destination mac address and hop limit for unicast destination 730 * GID entry, considering the source GID entry as well. 731 * ah_attribute must have have valid port_num, sgid_index. 732 */ 733 static int ib_resolve_unicast_gid_dmac(struct ib_device *device, 734 struct rdma_ah_attr *ah_attr) 735 { 736 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr); 737 const struct ib_gid_attr *sgid_attr = grh->sgid_attr; 738 int hop_limit = 0xff; 739 int ret = 0; 740 741 /* If destination is link local and source GID is RoCEv1, 742 * IP stack is not used. 743 */ 744 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) && 745 sgid_attr->gid_type == IB_GID_TYPE_ROCE) { 746 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw, 747 ah_attr->roce.dmac); 748 return ret; 749 } 750 751 ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid, 752 ah_attr->roce.dmac, 753 sgid_attr, &hop_limit); 754 755 grh->hop_limit = hop_limit; 756 return ret; 757 } 758 759 /* 760 * This function initializes address handle attributes from the incoming packet. 761 * Incoming packet has dgid of the receiver node on which this code is 762 * getting executed and, sgid contains the GID of the sender. 763 * 764 * When resolving mac address of destination, the arrived dgid is used 765 * as sgid and, sgid is used as dgid because sgid contains destinations 766 * GID whom to respond to. 767 * 768 * On success the caller is responsible to call rdma_destroy_ah_attr on the 769 * attr. 770 */ 771 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num, 772 const struct ib_wc *wc, const struct ib_grh *grh, 773 struct rdma_ah_attr *ah_attr) 774 { 775 u32 flow_class; 776 int ret; 777 enum rdma_network_type net_type = RDMA_NETWORK_IB; 778 enum ib_gid_type gid_type = IB_GID_TYPE_IB; 779 const struct ib_gid_attr *sgid_attr; 780 int hoplimit = 0xff; 781 union ib_gid dgid; 782 union ib_gid sgid; 783 784 might_sleep(); 785 786 memset(ah_attr, 0, sizeof *ah_attr); 787 ah_attr->type = rdma_ah_find_type(device, port_num); 788 if (rdma_cap_eth_ah(device, port_num)) { 789 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE) 790 net_type = wc->network_hdr_type; 791 else 792 net_type = ib_get_net_type_by_grh(device, port_num, grh); 793 gid_type = ib_network_to_gid_type(net_type); 794 } 795 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type, 796 &sgid, &dgid); 797 if (ret) 798 return ret; 799 800 rdma_ah_set_sl(ah_attr, wc->sl); 801 rdma_ah_set_port_num(ah_attr, port_num); 802 803 if (rdma_protocol_roce(device, port_num)) { 804 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ? 805 wc->vlan_id : 0xffff; 806 807 if (!(wc->wc_flags & IB_WC_GRH)) 808 return -EPROTOTYPE; 809 810 sgid_attr = get_sgid_attr_from_eth(device, port_num, 811 vlan_id, &dgid, 812 gid_type); 813 if (IS_ERR(sgid_attr)) 814 return PTR_ERR(sgid_attr); 815 816 flow_class = be32_to_cpu(grh->version_tclass_flow); 817 rdma_move_grh_sgid_attr(ah_attr, 818 &sgid, 819 flow_class & 0xFFFFF, 820 hoplimit, 821 (flow_class >> 20) & 0xFF, 822 sgid_attr); 823 824 ret = ib_resolve_unicast_gid_dmac(device, ah_attr); 825 if (ret) 826 rdma_destroy_ah_attr(ah_attr); 827 828 return ret; 829 } else { 830 rdma_ah_set_dlid(ah_attr, wc->slid); 831 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits); 832 833 if ((wc->wc_flags & IB_WC_GRH) == 0) 834 return 0; 835 836 if (dgid.global.interface_id != 837 cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) { 838 sgid_attr = rdma_find_gid_by_port( 839 device, &dgid, IB_GID_TYPE_IB, port_num, NULL); 840 } else 841 sgid_attr = rdma_get_gid_attr(device, port_num, 0); 842 843 if (IS_ERR(sgid_attr)) 844 return PTR_ERR(sgid_attr); 845 flow_class = be32_to_cpu(grh->version_tclass_flow); 846 rdma_move_grh_sgid_attr(ah_attr, 847 &sgid, 848 flow_class & 0xFFFFF, 849 hoplimit, 850 (flow_class >> 20) & 0xFF, 851 sgid_attr); 852 853 return 0; 854 } 855 } 856 EXPORT_SYMBOL(ib_init_ah_attr_from_wc); 857 858 /** 859 * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership 860 * of the reference 861 * 862 * @attr: Pointer to AH attribute structure 863 * @dgid: Destination GID 864 * @flow_label: Flow label 865 * @hop_limit: Hop limit 866 * @traffic_class: traffic class 867 * @sgid_attr: Pointer to SGID attribute 868 * 869 * This takes ownership of the sgid_attr reference. The caller must ensure 870 * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after 871 * calling this function. 872 */ 873 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid, 874 u32 flow_label, u8 hop_limit, u8 traffic_class, 875 const struct ib_gid_attr *sgid_attr) 876 { 877 rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit, 878 traffic_class); 879 attr->grh.sgid_attr = sgid_attr; 880 } 881 EXPORT_SYMBOL(rdma_move_grh_sgid_attr); 882 883 /** 884 * rdma_destroy_ah_attr - Release reference to SGID attribute of 885 * ah attribute. 886 * @ah_attr: Pointer to ah attribute 887 * 888 * Release reference to the SGID attribute of the ah attribute if it is 889 * non NULL. It is safe to call this multiple times, and safe to call it on 890 * a zero initialized ah_attr. 891 */ 892 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr) 893 { 894 if (ah_attr->grh.sgid_attr) { 895 rdma_put_gid_attr(ah_attr->grh.sgid_attr); 896 ah_attr->grh.sgid_attr = NULL; 897 } 898 } 899 EXPORT_SYMBOL(rdma_destroy_ah_attr); 900 901 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc, 902 const struct ib_grh *grh, u8 port_num) 903 { 904 struct rdma_ah_attr ah_attr; 905 struct ib_ah *ah; 906 int ret; 907 908 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr); 909 if (ret) 910 return ERR_PTR(ret); 911 912 ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE); 913 914 rdma_destroy_ah_attr(&ah_attr); 915 return ah; 916 } 917 EXPORT_SYMBOL(ib_create_ah_from_wc); 918 919 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr) 920 { 921 const struct ib_gid_attr *old_sgid_attr; 922 int ret; 923 924 if (ah->type != ah_attr->type) 925 return -EINVAL; 926 927 ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr); 928 if (ret) 929 return ret; 930 931 ret = ah->device->ops.modify_ah ? 932 ah->device->ops.modify_ah(ah, ah_attr) : 933 -EOPNOTSUPP; 934 935 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr); 936 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 937 return ret; 938 } 939 EXPORT_SYMBOL(rdma_modify_ah); 940 941 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr) 942 { 943 ah_attr->grh.sgid_attr = NULL; 944 945 return ah->device->ops.query_ah ? 946 ah->device->ops.query_ah(ah, ah_attr) : 947 -EOPNOTSUPP; 948 } 949 EXPORT_SYMBOL(rdma_query_ah); 950 951 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata) 952 { 953 const struct ib_gid_attr *sgid_attr = ah->sgid_attr; 954 struct ib_pd *pd; 955 956 might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE); 957 958 pd = ah->pd; 959 960 ah->device->ops.destroy_ah(ah, flags); 961 atomic_dec(&pd->usecnt); 962 if (sgid_attr) 963 rdma_put_gid_attr(sgid_attr); 964 965 kfree(ah); 966 return 0; 967 } 968 EXPORT_SYMBOL(rdma_destroy_ah_user); 969 970 /* Shared receive queues */ 971 972 struct ib_srq *ib_create_srq(struct ib_pd *pd, 973 struct ib_srq_init_attr *srq_init_attr) 974 { 975 struct ib_srq *srq; 976 int ret; 977 978 if (!pd->device->ops.create_srq) 979 return ERR_PTR(-EOPNOTSUPP); 980 981 srq = rdma_zalloc_drv_obj(pd->device, ib_srq); 982 if (!srq) 983 return ERR_PTR(-ENOMEM); 984 985 srq->device = pd->device; 986 srq->pd = pd; 987 srq->event_handler = srq_init_attr->event_handler; 988 srq->srq_context = srq_init_attr->srq_context; 989 srq->srq_type = srq_init_attr->srq_type; 990 991 if (ib_srq_has_cq(srq->srq_type)) { 992 srq->ext.cq = srq_init_attr->ext.cq; 993 atomic_inc(&srq->ext.cq->usecnt); 994 } 995 if (srq->srq_type == IB_SRQT_XRC) { 996 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd; 997 atomic_inc(&srq->ext.xrc.xrcd->usecnt); 998 } 999 atomic_inc(&pd->usecnt); 1000 1001 ret = pd->device->ops.create_srq(srq, srq_init_attr, NULL); 1002 if (ret) { 1003 atomic_dec(&srq->pd->usecnt); 1004 if (srq->srq_type == IB_SRQT_XRC) 1005 atomic_dec(&srq->ext.xrc.xrcd->usecnt); 1006 if (ib_srq_has_cq(srq->srq_type)) 1007 atomic_dec(&srq->ext.cq->usecnt); 1008 kfree(srq); 1009 return ERR_PTR(ret); 1010 } 1011 1012 return srq; 1013 } 1014 EXPORT_SYMBOL(ib_create_srq); 1015 1016 int ib_modify_srq(struct ib_srq *srq, 1017 struct ib_srq_attr *srq_attr, 1018 enum ib_srq_attr_mask srq_attr_mask) 1019 { 1020 return srq->device->ops.modify_srq ? 1021 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask, 1022 NULL) : -EOPNOTSUPP; 1023 } 1024 EXPORT_SYMBOL(ib_modify_srq); 1025 1026 int ib_query_srq(struct ib_srq *srq, 1027 struct ib_srq_attr *srq_attr) 1028 { 1029 return srq->device->ops.query_srq ? 1030 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP; 1031 } 1032 EXPORT_SYMBOL(ib_query_srq); 1033 1034 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata) 1035 { 1036 if (atomic_read(&srq->usecnt)) 1037 return -EBUSY; 1038 1039 srq->device->ops.destroy_srq(srq, udata); 1040 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 1048 return 0; 1049 } 1050 EXPORT_SYMBOL(ib_destroy_srq_user); 1051 1052 /* Queue pairs */ 1053 1054 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context) 1055 { 1056 struct ib_qp *qp = context; 1057 unsigned long flags; 1058 1059 spin_lock_irqsave(&qp->device->qp_open_list_lock, flags); 1060 list_for_each_entry(event->element.qp, &qp->open_list, open_list) 1061 if (event->element.qp->event_handler) 1062 event->element.qp->event_handler(event, event->element.qp->qp_context); 1063 spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags); 1064 } 1065 1066 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp) 1067 { 1068 mutex_lock(&xrcd->tgt_qp_mutex); 1069 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list); 1070 mutex_unlock(&xrcd->tgt_qp_mutex); 1071 } 1072 1073 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp, 1074 void (*event_handler)(struct ib_event *, void *), 1075 void *qp_context) 1076 { 1077 struct ib_qp *qp; 1078 unsigned long flags; 1079 int err; 1080 1081 qp = kzalloc(sizeof *qp, GFP_KERNEL); 1082 if (!qp) 1083 return ERR_PTR(-ENOMEM); 1084 1085 qp->real_qp = real_qp; 1086 err = ib_open_shared_qp_security(qp, real_qp->device); 1087 if (err) { 1088 kfree(qp); 1089 return ERR_PTR(err); 1090 } 1091 1092 qp->real_qp = real_qp; 1093 atomic_inc(&real_qp->usecnt); 1094 qp->device = real_qp->device; 1095 qp->event_handler = event_handler; 1096 qp->qp_context = qp_context; 1097 qp->qp_num = real_qp->qp_num; 1098 qp->qp_type = real_qp->qp_type; 1099 1100 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags); 1101 list_add(&qp->open_list, &real_qp->open_list); 1102 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags); 1103 1104 return qp; 1105 } 1106 1107 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd, 1108 struct ib_qp_open_attr *qp_open_attr) 1109 { 1110 struct ib_qp *qp, *real_qp; 1111 1112 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT) 1113 return ERR_PTR(-EINVAL); 1114 1115 qp = ERR_PTR(-EINVAL); 1116 mutex_lock(&xrcd->tgt_qp_mutex); 1117 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) { 1118 if (real_qp->qp_num == qp_open_attr->qp_num) { 1119 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler, 1120 qp_open_attr->qp_context); 1121 break; 1122 } 1123 } 1124 mutex_unlock(&xrcd->tgt_qp_mutex); 1125 return qp; 1126 } 1127 EXPORT_SYMBOL(ib_open_qp); 1128 1129 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp, 1130 struct ib_qp_init_attr *qp_init_attr, 1131 struct ib_udata *udata) 1132 { 1133 struct ib_qp *real_qp = qp; 1134 1135 qp->event_handler = __ib_shared_qp_event_handler; 1136 qp->qp_context = qp; 1137 qp->pd = NULL; 1138 qp->send_cq = qp->recv_cq = NULL; 1139 qp->srq = NULL; 1140 qp->xrcd = qp_init_attr->xrcd; 1141 atomic_inc(&qp_init_attr->xrcd->usecnt); 1142 INIT_LIST_HEAD(&qp->open_list); 1143 1144 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler, 1145 qp_init_attr->qp_context); 1146 if (IS_ERR(qp)) 1147 return qp; 1148 1149 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp); 1150 return qp; 1151 } 1152 1153 struct ib_qp *ib_create_qp_user(struct ib_pd *pd, 1154 struct ib_qp_init_attr *qp_init_attr, 1155 struct ib_udata *udata) 1156 { 1157 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device; 1158 struct ib_qp *qp; 1159 int ret; 1160 1161 if (qp_init_attr->rwq_ind_tbl && 1162 (qp_init_attr->recv_cq || 1163 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr || 1164 qp_init_attr->cap.max_recv_sge)) 1165 return ERR_PTR(-EINVAL); 1166 1167 if ((qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) && 1168 !(device->attrs.device_cap_flags & IB_DEVICE_INTEGRITY_HANDOVER)) 1169 return ERR_PTR(-EINVAL); 1170 1171 /* 1172 * If the callers is using the RDMA API calculate the resources 1173 * needed for the RDMA READ/WRITE operations. 1174 * 1175 * Note that these callers need to pass in a port number. 1176 */ 1177 if (qp_init_attr->cap.max_rdma_ctxs) 1178 rdma_rw_init_qp(device, qp_init_attr); 1179 1180 qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL); 1181 if (IS_ERR(qp)) 1182 return qp; 1183 1184 ret = ib_create_qp_security(qp, device); 1185 if (ret) 1186 goto err; 1187 1188 qp->qp_type = qp_init_attr->qp_type; 1189 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl; 1190 1191 atomic_set(&qp->usecnt, 0); 1192 qp->mrs_used = 0; 1193 spin_lock_init(&qp->mr_lock); 1194 INIT_LIST_HEAD(&qp->rdma_mrs); 1195 INIT_LIST_HEAD(&qp->sig_mrs); 1196 qp->port = 0; 1197 1198 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) { 1199 struct ib_qp *xrc_qp = 1200 create_xrc_qp_user(qp, qp_init_attr, udata); 1201 1202 if (IS_ERR(xrc_qp)) { 1203 ret = PTR_ERR(xrc_qp); 1204 goto err; 1205 } 1206 return xrc_qp; 1207 } 1208 1209 qp->event_handler = qp_init_attr->event_handler; 1210 qp->qp_context = qp_init_attr->qp_context; 1211 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) { 1212 qp->recv_cq = NULL; 1213 qp->srq = NULL; 1214 } else { 1215 qp->recv_cq = qp_init_attr->recv_cq; 1216 if (qp_init_attr->recv_cq) 1217 atomic_inc(&qp_init_attr->recv_cq->usecnt); 1218 qp->srq = qp_init_attr->srq; 1219 if (qp->srq) 1220 atomic_inc(&qp_init_attr->srq->usecnt); 1221 } 1222 1223 qp->send_cq = qp_init_attr->send_cq; 1224 qp->xrcd = NULL; 1225 1226 atomic_inc(&pd->usecnt); 1227 if (qp_init_attr->send_cq) 1228 atomic_inc(&qp_init_attr->send_cq->usecnt); 1229 if (qp_init_attr->rwq_ind_tbl) 1230 atomic_inc(&qp->rwq_ind_tbl->usecnt); 1231 1232 if (qp_init_attr->cap.max_rdma_ctxs) { 1233 ret = rdma_rw_init_mrs(qp, qp_init_attr); 1234 if (ret) 1235 goto err; 1236 } 1237 1238 /* 1239 * Note: all hw drivers guarantee that max_send_sge is lower than 1240 * the device RDMA WRITE SGE limit but not all hw drivers ensure that 1241 * max_send_sge <= max_sge_rd. 1242 */ 1243 qp->max_write_sge = qp_init_attr->cap.max_send_sge; 1244 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge, 1245 device->attrs.max_sge_rd); 1246 if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) 1247 qp->integrity_en = true; 1248 1249 return qp; 1250 1251 err: 1252 ib_destroy_qp(qp); 1253 return ERR_PTR(ret); 1254 1255 } 1256 EXPORT_SYMBOL(ib_create_qp_user); 1257 1258 static const struct { 1259 int valid; 1260 enum ib_qp_attr_mask req_param[IB_QPT_MAX]; 1261 enum ib_qp_attr_mask opt_param[IB_QPT_MAX]; 1262 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = { 1263 [IB_QPS_RESET] = { 1264 [IB_QPS_RESET] = { .valid = 1 }, 1265 [IB_QPS_INIT] = { 1266 .valid = 1, 1267 .req_param = { 1268 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1269 IB_QP_PORT | 1270 IB_QP_QKEY), 1271 [IB_QPT_RAW_PACKET] = IB_QP_PORT, 1272 [IB_QPT_UC] = (IB_QP_PKEY_INDEX | 1273 IB_QP_PORT | 1274 IB_QP_ACCESS_FLAGS), 1275 [IB_QPT_RC] = (IB_QP_PKEY_INDEX | 1276 IB_QP_PORT | 1277 IB_QP_ACCESS_FLAGS), 1278 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX | 1279 IB_QP_PORT | 1280 IB_QP_ACCESS_FLAGS), 1281 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX | 1282 IB_QP_PORT | 1283 IB_QP_ACCESS_FLAGS), 1284 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1285 IB_QP_QKEY), 1286 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1287 IB_QP_QKEY), 1288 } 1289 }, 1290 }, 1291 [IB_QPS_INIT] = { 1292 [IB_QPS_RESET] = { .valid = 1 }, 1293 [IB_QPS_ERR] = { .valid = 1 }, 1294 [IB_QPS_INIT] = { 1295 .valid = 1, 1296 .opt_param = { 1297 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1298 IB_QP_PORT | 1299 IB_QP_QKEY), 1300 [IB_QPT_UC] = (IB_QP_PKEY_INDEX | 1301 IB_QP_PORT | 1302 IB_QP_ACCESS_FLAGS), 1303 [IB_QPT_RC] = (IB_QP_PKEY_INDEX | 1304 IB_QP_PORT | 1305 IB_QP_ACCESS_FLAGS), 1306 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX | 1307 IB_QP_PORT | 1308 IB_QP_ACCESS_FLAGS), 1309 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX | 1310 IB_QP_PORT | 1311 IB_QP_ACCESS_FLAGS), 1312 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1313 IB_QP_QKEY), 1314 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1315 IB_QP_QKEY), 1316 } 1317 }, 1318 [IB_QPS_RTR] = { 1319 .valid = 1, 1320 .req_param = { 1321 [IB_QPT_UC] = (IB_QP_AV | 1322 IB_QP_PATH_MTU | 1323 IB_QP_DEST_QPN | 1324 IB_QP_RQ_PSN), 1325 [IB_QPT_RC] = (IB_QP_AV | 1326 IB_QP_PATH_MTU | 1327 IB_QP_DEST_QPN | 1328 IB_QP_RQ_PSN | 1329 IB_QP_MAX_DEST_RD_ATOMIC | 1330 IB_QP_MIN_RNR_TIMER), 1331 [IB_QPT_XRC_INI] = (IB_QP_AV | 1332 IB_QP_PATH_MTU | 1333 IB_QP_DEST_QPN | 1334 IB_QP_RQ_PSN), 1335 [IB_QPT_XRC_TGT] = (IB_QP_AV | 1336 IB_QP_PATH_MTU | 1337 IB_QP_DEST_QPN | 1338 IB_QP_RQ_PSN | 1339 IB_QP_MAX_DEST_RD_ATOMIC | 1340 IB_QP_MIN_RNR_TIMER), 1341 }, 1342 .opt_param = { 1343 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1344 IB_QP_QKEY), 1345 [IB_QPT_UC] = (IB_QP_ALT_PATH | 1346 IB_QP_ACCESS_FLAGS | 1347 IB_QP_PKEY_INDEX), 1348 [IB_QPT_RC] = (IB_QP_ALT_PATH | 1349 IB_QP_ACCESS_FLAGS | 1350 IB_QP_PKEY_INDEX), 1351 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH | 1352 IB_QP_ACCESS_FLAGS | 1353 IB_QP_PKEY_INDEX), 1354 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH | 1355 IB_QP_ACCESS_FLAGS | 1356 IB_QP_PKEY_INDEX), 1357 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1358 IB_QP_QKEY), 1359 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1360 IB_QP_QKEY), 1361 }, 1362 }, 1363 }, 1364 [IB_QPS_RTR] = { 1365 [IB_QPS_RESET] = { .valid = 1 }, 1366 [IB_QPS_ERR] = { .valid = 1 }, 1367 [IB_QPS_RTS] = { 1368 .valid = 1, 1369 .req_param = { 1370 [IB_QPT_UD] = IB_QP_SQ_PSN, 1371 [IB_QPT_UC] = IB_QP_SQ_PSN, 1372 [IB_QPT_RC] = (IB_QP_TIMEOUT | 1373 IB_QP_RETRY_CNT | 1374 IB_QP_RNR_RETRY | 1375 IB_QP_SQ_PSN | 1376 IB_QP_MAX_QP_RD_ATOMIC), 1377 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT | 1378 IB_QP_RETRY_CNT | 1379 IB_QP_RNR_RETRY | 1380 IB_QP_SQ_PSN | 1381 IB_QP_MAX_QP_RD_ATOMIC), 1382 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT | 1383 IB_QP_SQ_PSN), 1384 [IB_QPT_SMI] = IB_QP_SQ_PSN, 1385 [IB_QPT_GSI] = IB_QP_SQ_PSN, 1386 }, 1387 .opt_param = { 1388 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1389 IB_QP_QKEY), 1390 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1391 IB_QP_ALT_PATH | 1392 IB_QP_ACCESS_FLAGS | 1393 IB_QP_PATH_MIG_STATE), 1394 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1395 IB_QP_ALT_PATH | 1396 IB_QP_ACCESS_FLAGS | 1397 IB_QP_MIN_RNR_TIMER | 1398 IB_QP_PATH_MIG_STATE), 1399 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1400 IB_QP_ALT_PATH | 1401 IB_QP_ACCESS_FLAGS | 1402 IB_QP_PATH_MIG_STATE), 1403 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1404 IB_QP_ALT_PATH | 1405 IB_QP_ACCESS_FLAGS | 1406 IB_QP_MIN_RNR_TIMER | 1407 IB_QP_PATH_MIG_STATE), 1408 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1409 IB_QP_QKEY), 1410 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1411 IB_QP_QKEY), 1412 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT, 1413 } 1414 } 1415 }, 1416 [IB_QPS_RTS] = { 1417 [IB_QPS_RESET] = { .valid = 1 }, 1418 [IB_QPS_ERR] = { .valid = 1 }, 1419 [IB_QPS_RTS] = { 1420 .valid = 1, 1421 .opt_param = { 1422 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1423 IB_QP_QKEY), 1424 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1425 IB_QP_ACCESS_FLAGS | 1426 IB_QP_ALT_PATH | 1427 IB_QP_PATH_MIG_STATE), 1428 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1429 IB_QP_ACCESS_FLAGS | 1430 IB_QP_ALT_PATH | 1431 IB_QP_PATH_MIG_STATE | 1432 IB_QP_MIN_RNR_TIMER), 1433 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1434 IB_QP_ACCESS_FLAGS | 1435 IB_QP_ALT_PATH | 1436 IB_QP_PATH_MIG_STATE), 1437 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1438 IB_QP_ACCESS_FLAGS | 1439 IB_QP_ALT_PATH | 1440 IB_QP_PATH_MIG_STATE | 1441 IB_QP_MIN_RNR_TIMER), 1442 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1443 IB_QP_QKEY), 1444 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1445 IB_QP_QKEY), 1446 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT, 1447 } 1448 }, 1449 [IB_QPS_SQD] = { 1450 .valid = 1, 1451 .opt_param = { 1452 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1453 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1454 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1455 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1456 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */ 1457 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1458 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY 1459 } 1460 }, 1461 }, 1462 [IB_QPS_SQD] = { 1463 [IB_QPS_RESET] = { .valid = 1 }, 1464 [IB_QPS_ERR] = { .valid = 1 }, 1465 [IB_QPS_RTS] = { 1466 .valid = 1, 1467 .opt_param = { 1468 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1469 IB_QP_QKEY), 1470 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1471 IB_QP_ALT_PATH | 1472 IB_QP_ACCESS_FLAGS | 1473 IB_QP_PATH_MIG_STATE), 1474 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1475 IB_QP_ALT_PATH | 1476 IB_QP_ACCESS_FLAGS | 1477 IB_QP_MIN_RNR_TIMER | 1478 IB_QP_PATH_MIG_STATE), 1479 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1480 IB_QP_ALT_PATH | 1481 IB_QP_ACCESS_FLAGS | 1482 IB_QP_PATH_MIG_STATE), 1483 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1484 IB_QP_ALT_PATH | 1485 IB_QP_ACCESS_FLAGS | 1486 IB_QP_MIN_RNR_TIMER | 1487 IB_QP_PATH_MIG_STATE), 1488 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1489 IB_QP_QKEY), 1490 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1491 IB_QP_QKEY), 1492 } 1493 }, 1494 [IB_QPS_SQD] = { 1495 .valid = 1, 1496 .opt_param = { 1497 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1498 IB_QP_QKEY), 1499 [IB_QPT_UC] = (IB_QP_AV | 1500 IB_QP_ALT_PATH | 1501 IB_QP_ACCESS_FLAGS | 1502 IB_QP_PKEY_INDEX | 1503 IB_QP_PATH_MIG_STATE), 1504 [IB_QPT_RC] = (IB_QP_PORT | 1505 IB_QP_AV | 1506 IB_QP_TIMEOUT | 1507 IB_QP_RETRY_CNT | 1508 IB_QP_RNR_RETRY | 1509 IB_QP_MAX_QP_RD_ATOMIC | 1510 IB_QP_MAX_DEST_RD_ATOMIC | 1511 IB_QP_ALT_PATH | 1512 IB_QP_ACCESS_FLAGS | 1513 IB_QP_PKEY_INDEX | 1514 IB_QP_MIN_RNR_TIMER | 1515 IB_QP_PATH_MIG_STATE), 1516 [IB_QPT_XRC_INI] = (IB_QP_PORT | 1517 IB_QP_AV | 1518 IB_QP_TIMEOUT | 1519 IB_QP_RETRY_CNT | 1520 IB_QP_RNR_RETRY | 1521 IB_QP_MAX_QP_RD_ATOMIC | 1522 IB_QP_ALT_PATH | 1523 IB_QP_ACCESS_FLAGS | 1524 IB_QP_PKEY_INDEX | 1525 IB_QP_PATH_MIG_STATE), 1526 [IB_QPT_XRC_TGT] = (IB_QP_PORT | 1527 IB_QP_AV | 1528 IB_QP_TIMEOUT | 1529 IB_QP_MAX_DEST_RD_ATOMIC | 1530 IB_QP_ALT_PATH | 1531 IB_QP_ACCESS_FLAGS | 1532 IB_QP_PKEY_INDEX | 1533 IB_QP_MIN_RNR_TIMER | 1534 IB_QP_PATH_MIG_STATE), 1535 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1536 IB_QP_QKEY), 1537 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1538 IB_QP_QKEY), 1539 } 1540 } 1541 }, 1542 [IB_QPS_SQE] = { 1543 [IB_QPS_RESET] = { .valid = 1 }, 1544 [IB_QPS_ERR] = { .valid = 1 }, 1545 [IB_QPS_RTS] = { 1546 .valid = 1, 1547 .opt_param = { 1548 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1549 IB_QP_QKEY), 1550 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1551 IB_QP_ACCESS_FLAGS), 1552 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1553 IB_QP_QKEY), 1554 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1555 IB_QP_QKEY), 1556 } 1557 } 1558 }, 1559 [IB_QPS_ERR] = { 1560 [IB_QPS_RESET] = { .valid = 1 }, 1561 [IB_QPS_ERR] = { .valid = 1 } 1562 } 1563 }; 1564 1565 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state, 1566 enum ib_qp_type type, enum ib_qp_attr_mask mask) 1567 { 1568 enum ib_qp_attr_mask req_param, opt_param; 1569 1570 if (mask & IB_QP_CUR_STATE && 1571 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS && 1572 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE) 1573 return false; 1574 1575 if (!qp_state_table[cur_state][next_state].valid) 1576 return false; 1577 1578 req_param = qp_state_table[cur_state][next_state].req_param[type]; 1579 opt_param = qp_state_table[cur_state][next_state].opt_param[type]; 1580 1581 if ((mask & req_param) != req_param) 1582 return false; 1583 1584 if (mask & ~(req_param | opt_param | IB_QP_STATE)) 1585 return false; 1586 1587 return true; 1588 } 1589 EXPORT_SYMBOL(ib_modify_qp_is_ok); 1590 1591 /** 1592 * ib_resolve_eth_dmac - Resolve destination mac address 1593 * @device: Device to consider 1594 * @ah_attr: address handle attribute which describes the 1595 * source and destination parameters 1596 * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It 1597 * returns 0 on success or appropriate error code. It initializes the 1598 * necessary ah_attr fields when call is successful. 1599 */ 1600 static int ib_resolve_eth_dmac(struct ib_device *device, 1601 struct rdma_ah_attr *ah_attr) 1602 { 1603 int ret = 0; 1604 1605 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) { 1606 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) { 1607 __be32 addr = 0; 1608 1609 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4); 1610 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac); 1611 } else { 1612 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw, 1613 (char *)ah_attr->roce.dmac); 1614 } 1615 } else { 1616 ret = ib_resolve_unicast_gid_dmac(device, ah_attr); 1617 } 1618 return ret; 1619 } 1620 1621 static bool is_qp_type_connected(const struct ib_qp *qp) 1622 { 1623 return (qp->qp_type == IB_QPT_UC || 1624 qp->qp_type == IB_QPT_RC || 1625 qp->qp_type == IB_QPT_XRC_INI || 1626 qp->qp_type == IB_QPT_XRC_TGT); 1627 } 1628 1629 /** 1630 * IB core internal function to perform QP attributes modification. 1631 */ 1632 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr, 1633 int attr_mask, struct ib_udata *udata) 1634 { 1635 u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port; 1636 const struct ib_gid_attr *old_sgid_attr_av; 1637 const struct ib_gid_attr *old_sgid_attr_alt_av; 1638 int ret; 1639 1640 if (attr_mask & IB_QP_AV) { 1641 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr, 1642 &old_sgid_attr_av); 1643 if (ret) 1644 return ret; 1645 } 1646 if (attr_mask & IB_QP_ALT_PATH) { 1647 /* 1648 * FIXME: This does not track the migration state, so if the 1649 * user loads a new alternate path after the HW has migrated 1650 * from primary->alternate we will keep the wrong 1651 * references. This is OK for IB because the reference 1652 * counting does not serve any functional purpose. 1653 */ 1654 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr, 1655 &old_sgid_attr_alt_av); 1656 if (ret) 1657 goto out_av; 1658 1659 /* 1660 * Today the core code can only handle alternate paths and APM 1661 * for IB. Ban them in roce mode. 1662 */ 1663 if (!(rdma_protocol_ib(qp->device, 1664 attr->alt_ah_attr.port_num) && 1665 rdma_protocol_ib(qp->device, port))) { 1666 ret = EINVAL; 1667 goto out; 1668 } 1669 } 1670 1671 /* 1672 * If the user provided the qp_attr then we have to resolve it. Kernel 1673 * users have to provide already resolved rdma_ah_attr's 1674 */ 1675 if (udata && (attr_mask & IB_QP_AV) && 1676 attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE && 1677 is_qp_type_connected(qp)) { 1678 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr); 1679 if (ret) 1680 goto out; 1681 } 1682 1683 if (rdma_ib_or_roce(qp->device, port)) { 1684 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) { 1685 dev_warn(&qp->device->dev, 1686 "%s rq_psn overflow, masking to 24 bits\n", 1687 __func__); 1688 attr->rq_psn &= 0xffffff; 1689 } 1690 1691 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) { 1692 dev_warn(&qp->device->dev, 1693 " %s sq_psn overflow, masking to 24 bits\n", 1694 __func__); 1695 attr->sq_psn &= 0xffffff; 1696 } 1697 } 1698 1699 /* 1700 * Bind this qp to a counter automatically based on the rdma counter 1701 * rules. This only set in RST2INIT with port specified 1702 */ 1703 if (!qp->counter && (attr_mask & IB_QP_PORT) && 1704 ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT)) 1705 rdma_counter_bind_qp_auto(qp, attr->port_num); 1706 1707 ret = ib_security_modify_qp(qp, attr, attr_mask, udata); 1708 if (ret) 1709 goto out; 1710 1711 if (attr_mask & IB_QP_PORT) 1712 qp->port = attr->port_num; 1713 if (attr_mask & IB_QP_AV) 1714 qp->av_sgid_attr = 1715 rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr); 1716 if (attr_mask & IB_QP_ALT_PATH) 1717 qp->alt_path_sgid_attr = rdma_update_sgid_attr( 1718 &attr->alt_ah_attr, qp->alt_path_sgid_attr); 1719 1720 out: 1721 if (attr_mask & IB_QP_ALT_PATH) 1722 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av); 1723 out_av: 1724 if (attr_mask & IB_QP_AV) 1725 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av); 1726 return ret; 1727 } 1728 1729 /** 1730 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP. 1731 * @ib_qp: The QP to modify. 1732 * @attr: On input, specifies the QP attributes to modify. On output, 1733 * the current values of selected QP attributes are returned. 1734 * @attr_mask: A bit-mask used to specify which attributes of the QP 1735 * are being modified. 1736 * @udata: pointer to user's input output buffer information 1737 * are being modified. 1738 * It returns 0 on success and returns appropriate error code on error. 1739 */ 1740 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr, 1741 int attr_mask, struct ib_udata *udata) 1742 { 1743 return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata); 1744 } 1745 EXPORT_SYMBOL(ib_modify_qp_with_udata); 1746 1747 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width) 1748 { 1749 int rc; 1750 u32 netdev_speed; 1751 struct net_device *netdev; 1752 struct ethtool_link_ksettings lksettings; 1753 1754 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET) 1755 return -EINVAL; 1756 1757 netdev = ib_device_get_netdev(dev, port_num); 1758 if (!netdev) 1759 return -ENODEV; 1760 1761 rtnl_lock(); 1762 rc = __ethtool_get_link_ksettings(netdev, &lksettings); 1763 rtnl_unlock(); 1764 1765 dev_put(netdev); 1766 1767 if (!rc) { 1768 netdev_speed = lksettings.base.speed; 1769 } else { 1770 netdev_speed = SPEED_1000; 1771 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name, 1772 netdev_speed); 1773 } 1774 1775 if (netdev_speed <= SPEED_1000) { 1776 *width = IB_WIDTH_1X; 1777 *speed = IB_SPEED_SDR; 1778 } else if (netdev_speed <= SPEED_10000) { 1779 *width = IB_WIDTH_1X; 1780 *speed = IB_SPEED_FDR10; 1781 } else if (netdev_speed <= SPEED_20000) { 1782 *width = IB_WIDTH_4X; 1783 *speed = IB_SPEED_DDR; 1784 } else if (netdev_speed <= SPEED_25000) { 1785 *width = IB_WIDTH_1X; 1786 *speed = IB_SPEED_EDR; 1787 } else if (netdev_speed <= SPEED_40000) { 1788 *width = IB_WIDTH_4X; 1789 *speed = IB_SPEED_FDR10; 1790 } else { 1791 *width = IB_WIDTH_4X; 1792 *speed = IB_SPEED_EDR; 1793 } 1794 1795 return 0; 1796 } 1797 EXPORT_SYMBOL(ib_get_eth_speed); 1798 1799 int ib_modify_qp(struct ib_qp *qp, 1800 struct ib_qp_attr *qp_attr, 1801 int qp_attr_mask) 1802 { 1803 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL); 1804 } 1805 EXPORT_SYMBOL(ib_modify_qp); 1806 1807 int ib_query_qp(struct ib_qp *qp, 1808 struct ib_qp_attr *qp_attr, 1809 int qp_attr_mask, 1810 struct ib_qp_init_attr *qp_init_attr) 1811 { 1812 qp_attr->ah_attr.grh.sgid_attr = NULL; 1813 qp_attr->alt_ah_attr.grh.sgid_attr = NULL; 1814 1815 return qp->device->ops.query_qp ? 1816 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask, 1817 qp_init_attr) : -EOPNOTSUPP; 1818 } 1819 EXPORT_SYMBOL(ib_query_qp); 1820 1821 int ib_close_qp(struct ib_qp *qp) 1822 { 1823 struct ib_qp *real_qp; 1824 unsigned long flags; 1825 1826 real_qp = qp->real_qp; 1827 if (real_qp == qp) 1828 return -EINVAL; 1829 1830 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags); 1831 list_del(&qp->open_list); 1832 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags); 1833 1834 atomic_dec(&real_qp->usecnt); 1835 if (qp->qp_sec) 1836 ib_close_shared_qp_security(qp->qp_sec); 1837 kfree(qp); 1838 1839 return 0; 1840 } 1841 EXPORT_SYMBOL(ib_close_qp); 1842 1843 static int __ib_destroy_shared_qp(struct ib_qp *qp) 1844 { 1845 struct ib_xrcd *xrcd; 1846 struct ib_qp *real_qp; 1847 int ret; 1848 1849 real_qp = qp->real_qp; 1850 xrcd = real_qp->xrcd; 1851 1852 mutex_lock(&xrcd->tgt_qp_mutex); 1853 ib_close_qp(qp); 1854 if (atomic_read(&real_qp->usecnt) == 0) 1855 list_del(&real_qp->xrcd_list); 1856 else 1857 real_qp = NULL; 1858 mutex_unlock(&xrcd->tgt_qp_mutex); 1859 1860 if (real_qp) { 1861 ret = ib_destroy_qp(real_qp); 1862 if (!ret) 1863 atomic_dec(&xrcd->usecnt); 1864 else 1865 __ib_insert_xrcd_qp(xrcd, real_qp); 1866 } 1867 1868 return 0; 1869 } 1870 1871 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata) 1872 { 1873 const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr; 1874 const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr; 1875 struct ib_pd *pd; 1876 struct ib_cq *scq, *rcq; 1877 struct ib_srq *srq; 1878 struct ib_rwq_ind_table *ind_tbl; 1879 struct ib_qp_security *sec; 1880 int ret; 1881 1882 WARN_ON_ONCE(qp->mrs_used > 0); 1883 1884 if (atomic_read(&qp->usecnt)) 1885 return -EBUSY; 1886 1887 if (qp->real_qp != qp) 1888 return __ib_destroy_shared_qp(qp); 1889 1890 pd = qp->pd; 1891 scq = qp->send_cq; 1892 rcq = qp->recv_cq; 1893 srq = qp->srq; 1894 ind_tbl = qp->rwq_ind_tbl; 1895 sec = qp->qp_sec; 1896 if (sec) 1897 ib_destroy_qp_security_begin(sec); 1898 1899 if (!qp->uobject) 1900 rdma_rw_cleanup_mrs(qp); 1901 1902 rdma_counter_unbind_qp(qp, true); 1903 rdma_restrack_del(&qp->res); 1904 ret = qp->device->ops.destroy_qp(qp, udata); 1905 if (!ret) { 1906 if (alt_path_sgid_attr) 1907 rdma_put_gid_attr(alt_path_sgid_attr); 1908 if (av_sgid_attr) 1909 rdma_put_gid_attr(av_sgid_attr); 1910 if (pd) 1911 atomic_dec(&pd->usecnt); 1912 if (scq) 1913 atomic_dec(&scq->usecnt); 1914 if (rcq) 1915 atomic_dec(&rcq->usecnt); 1916 if (srq) 1917 atomic_dec(&srq->usecnt); 1918 if (ind_tbl) 1919 atomic_dec(&ind_tbl->usecnt); 1920 if (sec) 1921 ib_destroy_qp_security_end(sec); 1922 } else { 1923 if (sec) 1924 ib_destroy_qp_security_abort(sec); 1925 } 1926 1927 return ret; 1928 } 1929 EXPORT_SYMBOL(ib_destroy_qp_user); 1930 1931 /* Completion queues */ 1932 1933 struct ib_cq *__ib_create_cq(struct ib_device *device, 1934 ib_comp_handler comp_handler, 1935 void (*event_handler)(struct ib_event *, void *), 1936 void *cq_context, 1937 const struct ib_cq_init_attr *cq_attr, 1938 const char *caller) 1939 { 1940 struct ib_cq *cq; 1941 int ret; 1942 1943 cq = rdma_zalloc_drv_obj(device, ib_cq); 1944 if (!cq) 1945 return ERR_PTR(-ENOMEM); 1946 1947 cq->device = device; 1948 cq->uobject = NULL; 1949 cq->comp_handler = comp_handler; 1950 cq->event_handler = event_handler; 1951 cq->cq_context = cq_context; 1952 atomic_set(&cq->usecnt, 0); 1953 cq->res.type = RDMA_RESTRACK_CQ; 1954 rdma_restrack_set_task(&cq->res, caller); 1955 1956 ret = device->ops.create_cq(cq, cq_attr, NULL); 1957 if (ret) { 1958 kfree(cq); 1959 return ERR_PTR(ret); 1960 } 1961 1962 rdma_restrack_kadd(&cq->res); 1963 return cq; 1964 } 1965 EXPORT_SYMBOL(__ib_create_cq); 1966 1967 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period) 1968 { 1969 return cq->device->ops.modify_cq ? 1970 cq->device->ops.modify_cq(cq, cq_count, 1971 cq_period) : -EOPNOTSUPP; 1972 } 1973 EXPORT_SYMBOL(rdma_set_cq_moderation); 1974 1975 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata) 1976 { 1977 if (atomic_read(&cq->usecnt)) 1978 return -EBUSY; 1979 1980 rdma_restrack_del(&cq->res); 1981 cq->device->ops.destroy_cq(cq, udata); 1982 kfree(cq); 1983 return 0; 1984 } 1985 EXPORT_SYMBOL(ib_destroy_cq_user); 1986 1987 int ib_resize_cq(struct ib_cq *cq, int cqe) 1988 { 1989 return cq->device->ops.resize_cq ? 1990 cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP; 1991 } 1992 EXPORT_SYMBOL(ib_resize_cq); 1993 1994 /* Memory regions */ 1995 1996 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, 1997 u64 virt_addr, int access_flags) 1998 { 1999 struct ib_mr *mr; 2000 2001 if (access_flags & IB_ACCESS_ON_DEMAND) { 2002 if (!(pd->device->attrs.device_cap_flags & 2003 IB_DEVICE_ON_DEMAND_PAGING)) { 2004 pr_debug("ODP support not available\n"); 2005 return ERR_PTR(-EINVAL); 2006 } 2007 } 2008 2009 mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr, 2010 access_flags, NULL); 2011 2012 if (IS_ERR(mr)) 2013 return mr; 2014 2015 mr->device = pd->device; 2016 mr->pd = pd; 2017 mr->dm = NULL; 2018 atomic_inc(&pd->usecnt); 2019 mr->res.type = RDMA_RESTRACK_MR; 2020 rdma_restrack_kadd(&mr->res); 2021 2022 return mr; 2023 } 2024 EXPORT_SYMBOL(ib_reg_user_mr); 2025 2026 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice, 2027 u32 flags, struct ib_sge *sg_list, u32 num_sge) 2028 { 2029 if (!pd->device->ops.advise_mr) 2030 return -EOPNOTSUPP; 2031 2032 return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge, 2033 NULL); 2034 } 2035 EXPORT_SYMBOL(ib_advise_mr); 2036 2037 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata) 2038 { 2039 struct ib_pd *pd = mr->pd; 2040 struct ib_dm *dm = mr->dm; 2041 struct ib_sig_attrs *sig_attrs = mr->sig_attrs; 2042 int ret; 2043 2044 trace_mr_dereg(mr); 2045 rdma_restrack_del(&mr->res); 2046 ret = mr->device->ops.dereg_mr(mr, udata); 2047 if (!ret) { 2048 atomic_dec(&pd->usecnt); 2049 if (dm) 2050 atomic_dec(&dm->usecnt); 2051 kfree(sig_attrs); 2052 } 2053 2054 return ret; 2055 } 2056 EXPORT_SYMBOL(ib_dereg_mr_user); 2057 2058 /** 2059 * ib_alloc_mr_user() - Allocates a memory region 2060 * @pd: protection domain associated with the region 2061 * @mr_type: memory region type 2062 * @max_num_sg: maximum sg entries available for registration. 2063 * @udata: user data or null for kernel objects 2064 * 2065 * Notes: 2066 * Memory registeration page/sg lists must not exceed max_num_sg. 2067 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed 2068 * max_num_sg * used_page_size. 2069 * 2070 */ 2071 struct ib_mr *ib_alloc_mr_user(struct ib_pd *pd, enum ib_mr_type mr_type, 2072 u32 max_num_sg, struct ib_udata *udata) 2073 { 2074 struct ib_mr *mr; 2075 2076 if (!pd->device->ops.alloc_mr) { 2077 mr = ERR_PTR(-EOPNOTSUPP); 2078 goto out; 2079 } 2080 2081 if (mr_type == IB_MR_TYPE_INTEGRITY) { 2082 WARN_ON_ONCE(1); 2083 mr = ERR_PTR(-EINVAL); 2084 goto out; 2085 } 2086 2087 mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg, udata); 2088 if (!IS_ERR(mr)) { 2089 mr->device = pd->device; 2090 mr->pd = pd; 2091 mr->dm = NULL; 2092 mr->uobject = NULL; 2093 atomic_inc(&pd->usecnt); 2094 mr->need_inval = false; 2095 mr->res.type = RDMA_RESTRACK_MR; 2096 rdma_restrack_kadd(&mr->res); 2097 mr->type = mr_type; 2098 mr->sig_attrs = NULL; 2099 } 2100 2101 out: 2102 trace_mr_alloc(pd, mr_type, max_num_sg, mr); 2103 return mr; 2104 } 2105 EXPORT_SYMBOL(ib_alloc_mr_user); 2106 2107 /** 2108 * ib_alloc_mr_integrity() - Allocates an integrity memory region 2109 * @pd: protection domain associated with the region 2110 * @max_num_data_sg: maximum data sg entries available for registration 2111 * @max_num_meta_sg: maximum metadata sg entries available for 2112 * registration 2113 * 2114 * Notes: 2115 * Memory registration page/sg lists must not exceed max_num_sg, 2116 * also the integrity page/sg lists must not exceed max_num_meta_sg. 2117 * 2118 */ 2119 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd, 2120 u32 max_num_data_sg, 2121 u32 max_num_meta_sg) 2122 { 2123 struct ib_mr *mr; 2124 struct ib_sig_attrs *sig_attrs; 2125 2126 if (!pd->device->ops.alloc_mr_integrity || 2127 !pd->device->ops.map_mr_sg_pi) { 2128 mr = ERR_PTR(-EOPNOTSUPP); 2129 goto out; 2130 } 2131 2132 if (!max_num_meta_sg) { 2133 mr = ERR_PTR(-EINVAL); 2134 goto out; 2135 } 2136 2137 sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL); 2138 if (!sig_attrs) { 2139 mr = ERR_PTR(-ENOMEM); 2140 goto out; 2141 } 2142 2143 mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg, 2144 max_num_meta_sg); 2145 if (IS_ERR(mr)) { 2146 kfree(sig_attrs); 2147 goto out; 2148 } 2149 2150 mr->device = pd->device; 2151 mr->pd = pd; 2152 mr->dm = NULL; 2153 mr->uobject = NULL; 2154 atomic_inc(&pd->usecnt); 2155 mr->need_inval = false; 2156 mr->res.type = RDMA_RESTRACK_MR; 2157 rdma_restrack_kadd(&mr->res); 2158 mr->type = IB_MR_TYPE_INTEGRITY; 2159 mr->sig_attrs = sig_attrs; 2160 2161 out: 2162 trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr); 2163 return mr; 2164 } 2165 EXPORT_SYMBOL(ib_alloc_mr_integrity); 2166 2167 /* "Fast" memory regions */ 2168 2169 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd, 2170 int mr_access_flags, 2171 struct ib_fmr_attr *fmr_attr) 2172 { 2173 struct ib_fmr *fmr; 2174 2175 if (!pd->device->ops.alloc_fmr) 2176 return ERR_PTR(-EOPNOTSUPP); 2177 2178 fmr = pd->device->ops.alloc_fmr(pd, mr_access_flags, fmr_attr); 2179 if (!IS_ERR(fmr)) { 2180 fmr->device = pd->device; 2181 fmr->pd = pd; 2182 atomic_inc(&pd->usecnt); 2183 } 2184 2185 return fmr; 2186 } 2187 EXPORT_SYMBOL(ib_alloc_fmr); 2188 2189 int ib_unmap_fmr(struct list_head *fmr_list) 2190 { 2191 struct ib_fmr *fmr; 2192 2193 if (list_empty(fmr_list)) 2194 return 0; 2195 2196 fmr = list_entry(fmr_list->next, struct ib_fmr, list); 2197 return fmr->device->ops.unmap_fmr(fmr_list); 2198 } 2199 EXPORT_SYMBOL(ib_unmap_fmr); 2200 2201 int ib_dealloc_fmr(struct ib_fmr *fmr) 2202 { 2203 struct ib_pd *pd; 2204 int ret; 2205 2206 pd = fmr->pd; 2207 ret = fmr->device->ops.dealloc_fmr(fmr); 2208 if (!ret) 2209 atomic_dec(&pd->usecnt); 2210 2211 return ret; 2212 } 2213 EXPORT_SYMBOL(ib_dealloc_fmr); 2214 2215 /* Multicast groups */ 2216 2217 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid) 2218 { 2219 struct ib_qp_init_attr init_attr = {}; 2220 struct ib_qp_attr attr = {}; 2221 int num_eth_ports = 0; 2222 int port; 2223 2224 /* If QP state >= init, it is assigned to a port and we can check this 2225 * port only. 2226 */ 2227 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) { 2228 if (attr.qp_state >= IB_QPS_INIT) { 2229 if (rdma_port_get_link_layer(qp->device, attr.port_num) != 2230 IB_LINK_LAYER_INFINIBAND) 2231 return true; 2232 goto lid_check; 2233 } 2234 } 2235 2236 /* Can't get a quick answer, iterate over all ports */ 2237 for (port = 0; port < qp->device->phys_port_cnt; port++) 2238 if (rdma_port_get_link_layer(qp->device, port) != 2239 IB_LINK_LAYER_INFINIBAND) 2240 num_eth_ports++; 2241 2242 /* If we have at lease one Ethernet port, RoCE annex declares that 2243 * multicast LID should be ignored. We can't tell at this step if the 2244 * QP belongs to an IB or Ethernet port. 2245 */ 2246 if (num_eth_ports) 2247 return true; 2248 2249 /* If all the ports are IB, we can check according to IB spec. */ 2250 lid_check: 2251 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) || 2252 lid == be16_to_cpu(IB_LID_PERMISSIVE)); 2253 } 2254 2255 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid) 2256 { 2257 int ret; 2258 2259 if (!qp->device->ops.attach_mcast) 2260 return -EOPNOTSUPP; 2261 2262 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) || 2263 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid)) 2264 return -EINVAL; 2265 2266 ret = qp->device->ops.attach_mcast(qp, gid, lid); 2267 if (!ret) 2268 atomic_inc(&qp->usecnt); 2269 return ret; 2270 } 2271 EXPORT_SYMBOL(ib_attach_mcast); 2272 2273 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid) 2274 { 2275 int ret; 2276 2277 if (!qp->device->ops.detach_mcast) 2278 return -EOPNOTSUPP; 2279 2280 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) || 2281 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid)) 2282 return -EINVAL; 2283 2284 ret = qp->device->ops.detach_mcast(qp, gid, lid); 2285 if (!ret) 2286 atomic_dec(&qp->usecnt); 2287 return ret; 2288 } 2289 EXPORT_SYMBOL(ib_detach_mcast); 2290 2291 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller) 2292 { 2293 struct ib_xrcd *xrcd; 2294 2295 if (!device->ops.alloc_xrcd) 2296 return ERR_PTR(-EOPNOTSUPP); 2297 2298 xrcd = device->ops.alloc_xrcd(device, NULL); 2299 if (!IS_ERR(xrcd)) { 2300 xrcd->device = device; 2301 xrcd->inode = NULL; 2302 atomic_set(&xrcd->usecnt, 0); 2303 mutex_init(&xrcd->tgt_qp_mutex); 2304 INIT_LIST_HEAD(&xrcd->tgt_qp_list); 2305 } 2306 2307 return xrcd; 2308 } 2309 EXPORT_SYMBOL(__ib_alloc_xrcd); 2310 2311 int ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata) 2312 { 2313 struct ib_qp *qp; 2314 int ret; 2315 2316 if (atomic_read(&xrcd->usecnt)) 2317 return -EBUSY; 2318 2319 while (!list_empty(&xrcd->tgt_qp_list)) { 2320 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list); 2321 ret = ib_destroy_qp(qp); 2322 if (ret) 2323 return ret; 2324 } 2325 mutex_destroy(&xrcd->tgt_qp_mutex); 2326 2327 return xrcd->device->ops.dealloc_xrcd(xrcd, udata); 2328 } 2329 EXPORT_SYMBOL(ib_dealloc_xrcd); 2330 2331 /** 2332 * ib_create_wq - Creates a WQ associated with the specified protection 2333 * domain. 2334 * @pd: The protection domain associated with the WQ. 2335 * @wq_attr: A list of initial attributes required to create the 2336 * WQ. If WQ creation succeeds, then the attributes are updated to 2337 * the actual capabilities of the created WQ. 2338 * 2339 * wq_attr->max_wr and wq_attr->max_sge determine 2340 * the requested size of the WQ, and set to the actual values allocated 2341 * on return. 2342 * If ib_create_wq() succeeds, then max_wr and max_sge will always be 2343 * at least as large as the requested values. 2344 */ 2345 struct ib_wq *ib_create_wq(struct ib_pd *pd, 2346 struct ib_wq_init_attr *wq_attr) 2347 { 2348 struct ib_wq *wq; 2349 2350 if (!pd->device->ops.create_wq) 2351 return ERR_PTR(-EOPNOTSUPP); 2352 2353 wq = pd->device->ops.create_wq(pd, wq_attr, NULL); 2354 if (!IS_ERR(wq)) { 2355 wq->event_handler = wq_attr->event_handler; 2356 wq->wq_context = wq_attr->wq_context; 2357 wq->wq_type = wq_attr->wq_type; 2358 wq->cq = wq_attr->cq; 2359 wq->device = pd->device; 2360 wq->pd = pd; 2361 wq->uobject = NULL; 2362 atomic_inc(&pd->usecnt); 2363 atomic_inc(&wq_attr->cq->usecnt); 2364 atomic_set(&wq->usecnt, 0); 2365 } 2366 return wq; 2367 } 2368 EXPORT_SYMBOL(ib_create_wq); 2369 2370 /** 2371 * ib_destroy_wq - Destroys the specified user WQ. 2372 * @wq: The WQ to destroy. 2373 * @udata: Valid user data 2374 */ 2375 int ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata) 2376 { 2377 struct ib_cq *cq = wq->cq; 2378 struct ib_pd *pd = wq->pd; 2379 2380 if (atomic_read(&wq->usecnt)) 2381 return -EBUSY; 2382 2383 wq->device->ops.destroy_wq(wq, udata); 2384 atomic_dec(&pd->usecnt); 2385 atomic_dec(&cq->usecnt); 2386 2387 return 0; 2388 } 2389 EXPORT_SYMBOL(ib_destroy_wq); 2390 2391 /** 2392 * ib_modify_wq - Modifies the specified WQ. 2393 * @wq: The WQ to modify. 2394 * @wq_attr: On input, specifies the WQ attributes to modify. 2395 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ 2396 * are being modified. 2397 * On output, the current values of selected WQ attributes are returned. 2398 */ 2399 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr, 2400 u32 wq_attr_mask) 2401 { 2402 int err; 2403 2404 if (!wq->device->ops.modify_wq) 2405 return -EOPNOTSUPP; 2406 2407 err = wq->device->ops.modify_wq(wq, wq_attr, wq_attr_mask, NULL); 2408 return err; 2409 } 2410 EXPORT_SYMBOL(ib_modify_wq); 2411 2412 /* 2413 * ib_create_rwq_ind_table - Creates a RQ Indirection Table. 2414 * @device: The device on which to create the rwq indirection table. 2415 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to 2416 * create the Indirection Table. 2417 * 2418 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less 2419 * than the created ib_rwq_ind_table object and the caller is responsible 2420 * for its memory allocation/free. 2421 */ 2422 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device, 2423 struct ib_rwq_ind_table_init_attr *init_attr) 2424 { 2425 struct ib_rwq_ind_table *rwq_ind_table; 2426 int i; 2427 u32 table_size; 2428 2429 if (!device->ops.create_rwq_ind_table) 2430 return ERR_PTR(-EOPNOTSUPP); 2431 2432 table_size = (1 << init_attr->log_ind_tbl_size); 2433 rwq_ind_table = device->ops.create_rwq_ind_table(device, 2434 init_attr, NULL); 2435 if (IS_ERR(rwq_ind_table)) 2436 return rwq_ind_table; 2437 2438 rwq_ind_table->ind_tbl = init_attr->ind_tbl; 2439 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size; 2440 rwq_ind_table->device = device; 2441 rwq_ind_table->uobject = NULL; 2442 atomic_set(&rwq_ind_table->usecnt, 0); 2443 2444 for (i = 0; i < table_size; i++) 2445 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt); 2446 2447 return rwq_ind_table; 2448 } 2449 EXPORT_SYMBOL(ib_create_rwq_ind_table); 2450 2451 /* 2452 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table. 2453 * @wq_ind_table: The Indirection Table to destroy. 2454 */ 2455 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table) 2456 { 2457 int err, i; 2458 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size); 2459 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl; 2460 2461 if (atomic_read(&rwq_ind_table->usecnt)) 2462 return -EBUSY; 2463 2464 err = rwq_ind_table->device->ops.destroy_rwq_ind_table(rwq_ind_table); 2465 if (!err) { 2466 for (i = 0; i < table_size; i++) 2467 atomic_dec(&ind_tbl[i]->usecnt); 2468 } 2469 2470 return err; 2471 } 2472 EXPORT_SYMBOL(ib_destroy_rwq_ind_table); 2473 2474 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask, 2475 struct ib_mr_status *mr_status) 2476 { 2477 if (!mr->device->ops.check_mr_status) 2478 return -EOPNOTSUPP; 2479 2480 return mr->device->ops.check_mr_status(mr, check_mask, mr_status); 2481 } 2482 EXPORT_SYMBOL(ib_check_mr_status); 2483 2484 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port, 2485 int state) 2486 { 2487 if (!device->ops.set_vf_link_state) 2488 return -EOPNOTSUPP; 2489 2490 return device->ops.set_vf_link_state(device, vf, port, state); 2491 } 2492 EXPORT_SYMBOL(ib_set_vf_link_state); 2493 2494 int ib_get_vf_config(struct ib_device *device, int vf, u8 port, 2495 struct ifla_vf_info *info) 2496 { 2497 if (!device->ops.get_vf_config) 2498 return -EOPNOTSUPP; 2499 2500 return device->ops.get_vf_config(device, vf, port, info); 2501 } 2502 EXPORT_SYMBOL(ib_get_vf_config); 2503 2504 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port, 2505 struct ifla_vf_stats *stats) 2506 { 2507 if (!device->ops.get_vf_stats) 2508 return -EOPNOTSUPP; 2509 2510 return device->ops.get_vf_stats(device, vf, port, stats); 2511 } 2512 EXPORT_SYMBOL(ib_get_vf_stats); 2513 2514 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid, 2515 int type) 2516 { 2517 if (!device->ops.set_vf_guid) 2518 return -EOPNOTSUPP; 2519 2520 return device->ops.set_vf_guid(device, vf, port, guid, type); 2521 } 2522 EXPORT_SYMBOL(ib_set_vf_guid); 2523 2524 int ib_get_vf_guid(struct ib_device *device, int vf, u8 port, 2525 struct ifla_vf_guid *node_guid, 2526 struct ifla_vf_guid *port_guid) 2527 { 2528 if (!device->ops.get_vf_guid) 2529 return -EOPNOTSUPP; 2530 2531 return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid); 2532 } 2533 EXPORT_SYMBOL(ib_get_vf_guid); 2534 /** 2535 * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection 2536 * information) and set an appropriate memory region for registration. 2537 * @mr: memory region 2538 * @data_sg: dma mapped scatterlist for data 2539 * @data_sg_nents: number of entries in data_sg 2540 * @data_sg_offset: offset in bytes into data_sg 2541 * @meta_sg: dma mapped scatterlist for metadata 2542 * @meta_sg_nents: number of entries in meta_sg 2543 * @meta_sg_offset: offset in bytes into meta_sg 2544 * @page_size: page vector desired page size 2545 * 2546 * Constraints: 2547 * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY. 2548 * 2549 * Return: 0 on success. 2550 * 2551 * After this completes successfully, the memory region 2552 * is ready for registration. 2553 */ 2554 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg, 2555 int data_sg_nents, unsigned int *data_sg_offset, 2556 struct scatterlist *meta_sg, int meta_sg_nents, 2557 unsigned int *meta_sg_offset, unsigned int page_size) 2558 { 2559 if (unlikely(!mr->device->ops.map_mr_sg_pi || 2560 WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY))) 2561 return -EOPNOTSUPP; 2562 2563 mr->page_size = page_size; 2564 2565 return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents, 2566 data_sg_offset, meta_sg, 2567 meta_sg_nents, meta_sg_offset); 2568 } 2569 EXPORT_SYMBOL(ib_map_mr_sg_pi); 2570 2571 /** 2572 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list 2573 * and set it the memory region. 2574 * @mr: memory region 2575 * @sg: dma mapped scatterlist 2576 * @sg_nents: number of entries in sg 2577 * @sg_offset: offset in bytes into sg 2578 * @page_size: page vector desired page size 2579 * 2580 * Constraints: 2581 * - The first sg element is allowed to have an offset. 2582 * - Each sg element must either be aligned to page_size or virtually 2583 * contiguous to the previous element. In case an sg element has a 2584 * non-contiguous offset, the mapping prefix will not include it. 2585 * - The last sg element is allowed to have length less than page_size. 2586 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size 2587 * then only max_num_sg entries will be mapped. 2588 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these 2589 * constraints holds and the page_size argument is ignored. 2590 * 2591 * Returns the number of sg elements that were mapped to the memory region. 2592 * 2593 * After this completes successfully, the memory region 2594 * is ready for registration. 2595 */ 2596 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents, 2597 unsigned int *sg_offset, unsigned int page_size) 2598 { 2599 if (unlikely(!mr->device->ops.map_mr_sg)) 2600 return -EOPNOTSUPP; 2601 2602 mr->page_size = page_size; 2603 2604 return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset); 2605 } 2606 EXPORT_SYMBOL(ib_map_mr_sg); 2607 2608 /** 2609 * ib_sg_to_pages() - Convert the largest prefix of a sg list 2610 * to a page vector 2611 * @mr: memory region 2612 * @sgl: dma mapped scatterlist 2613 * @sg_nents: number of entries in sg 2614 * @sg_offset_p: IN: start offset in bytes into sg 2615 * OUT: offset in bytes for element n of the sg of the first 2616 * byte that has not been processed where n is the return 2617 * value of this function. 2618 * @set_page: driver page assignment function pointer 2619 * 2620 * Core service helper for drivers to convert the largest 2621 * prefix of given sg list to a page vector. The sg list 2622 * prefix converted is the prefix that meet the requirements 2623 * of ib_map_mr_sg. 2624 * 2625 * Returns the number of sg elements that were assigned to 2626 * a page vector. 2627 */ 2628 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents, 2629 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64)) 2630 { 2631 struct scatterlist *sg; 2632 u64 last_end_dma_addr = 0; 2633 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0; 2634 unsigned int last_page_off = 0; 2635 u64 page_mask = ~((u64)mr->page_size - 1); 2636 int i, ret; 2637 2638 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0]))) 2639 return -EINVAL; 2640 2641 mr->iova = sg_dma_address(&sgl[0]) + sg_offset; 2642 mr->length = 0; 2643 2644 for_each_sg(sgl, sg, sg_nents, i) { 2645 u64 dma_addr = sg_dma_address(sg) + sg_offset; 2646 u64 prev_addr = dma_addr; 2647 unsigned int dma_len = sg_dma_len(sg) - sg_offset; 2648 u64 end_dma_addr = dma_addr + dma_len; 2649 u64 page_addr = dma_addr & page_mask; 2650 2651 /* 2652 * For the second and later elements, check whether either the 2653 * end of element i-1 or the start of element i is not aligned 2654 * on a page boundary. 2655 */ 2656 if (i && (last_page_off != 0 || page_addr != dma_addr)) { 2657 /* Stop mapping if there is a gap. */ 2658 if (last_end_dma_addr != dma_addr) 2659 break; 2660 2661 /* 2662 * Coalesce this element with the last. If it is small 2663 * enough just update mr->length. Otherwise start 2664 * mapping from the next page. 2665 */ 2666 goto next_page; 2667 } 2668 2669 do { 2670 ret = set_page(mr, page_addr); 2671 if (unlikely(ret < 0)) { 2672 sg_offset = prev_addr - sg_dma_address(sg); 2673 mr->length += prev_addr - dma_addr; 2674 if (sg_offset_p) 2675 *sg_offset_p = sg_offset; 2676 return i || sg_offset ? i : ret; 2677 } 2678 prev_addr = page_addr; 2679 next_page: 2680 page_addr += mr->page_size; 2681 } while (page_addr < end_dma_addr); 2682 2683 mr->length += dma_len; 2684 last_end_dma_addr = end_dma_addr; 2685 last_page_off = end_dma_addr & ~page_mask; 2686 2687 sg_offset = 0; 2688 } 2689 2690 if (sg_offset_p) 2691 *sg_offset_p = 0; 2692 return i; 2693 } 2694 EXPORT_SYMBOL(ib_sg_to_pages); 2695 2696 struct ib_drain_cqe { 2697 struct ib_cqe cqe; 2698 struct completion done; 2699 }; 2700 2701 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc) 2702 { 2703 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe, 2704 cqe); 2705 2706 complete(&cqe->done); 2707 } 2708 2709 /* 2710 * Post a WR and block until its completion is reaped for the SQ. 2711 */ 2712 static void __ib_drain_sq(struct ib_qp *qp) 2713 { 2714 struct ib_cq *cq = qp->send_cq; 2715 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 2716 struct ib_drain_cqe sdrain; 2717 struct ib_rdma_wr swr = { 2718 .wr = { 2719 .next = NULL, 2720 { .wr_cqe = &sdrain.cqe, }, 2721 .opcode = IB_WR_RDMA_WRITE, 2722 }, 2723 }; 2724 int ret; 2725 2726 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 2727 if (ret) { 2728 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 2729 return; 2730 } 2731 2732 sdrain.cqe.done = ib_drain_qp_done; 2733 init_completion(&sdrain.done); 2734 2735 ret = ib_post_send(qp, &swr.wr, NULL); 2736 if (ret) { 2737 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 2738 return; 2739 } 2740 2741 if (cq->poll_ctx == IB_POLL_DIRECT) 2742 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0) 2743 ib_process_cq_direct(cq, -1); 2744 else 2745 wait_for_completion(&sdrain.done); 2746 } 2747 2748 /* 2749 * Post a WR and block until its completion is reaped for the RQ. 2750 */ 2751 static void __ib_drain_rq(struct ib_qp *qp) 2752 { 2753 struct ib_cq *cq = qp->recv_cq; 2754 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 2755 struct ib_drain_cqe rdrain; 2756 struct ib_recv_wr rwr = {}; 2757 int ret; 2758 2759 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 2760 if (ret) { 2761 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 2762 return; 2763 } 2764 2765 rwr.wr_cqe = &rdrain.cqe; 2766 rdrain.cqe.done = ib_drain_qp_done; 2767 init_completion(&rdrain.done); 2768 2769 ret = ib_post_recv(qp, &rwr, NULL); 2770 if (ret) { 2771 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 2772 return; 2773 } 2774 2775 if (cq->poll_ctx == IB_POLL_DIRECT) 2776 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0) 2777 ib_process_cq_direct(cq, -1); 2778 else 2779 wait_for_completion(&rdrain.done); 2780 } 2781 2782 /** 2783 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the 2784 * application. 2785 * @qp: queue pair to drain 2786 * 2787 * If the device has a provider-specific drain function, then 2788 * call that. Otherwise call the generic drain function 2789 * __ib_drain_sq(). 2790 * 2791 * The caller must: 2792 * 2793 * ensure there is room in the CQ and SQ for the drain work request and 2794 * completion. 2795 * 2796 * allocate the CQ using ib_alloc_cq(). 2797 * 2798 * ensure that there are no other contexts that are posting WRs concurrently. 2799 * Otherwise the drain is not guaranteed. 2800 */ 2801 void ib_drain_sq(struct ib_qp *qp) 2802 { 2803 if (qp->device->ops.drain_sq) 2804 qp->device->ops.drain_sq(qp); 2805 else 2806 __ib_drain_sq(qp); 2807 trace_cq_drain_complete(qp->send_cq); 2808 } 2809 EXPORT_SYMBOL(ib_drain_sq); 2810 2811 /** 2812 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the 2813 * application. 2814 * @qp: queue pair to drain 2815 * 2816 * If the device has a provider-specific drain function, then 2817 * call that. Otherwise call the generic drain function 2818 * __ib_drain_rq(). 2819 * 2820 * The caller must: 2821 * 2822 * ensure there is room in the CQ and RQ for the drain work request and 2823 * completion. 2824 * 2825 * allocate the CQ using ib_alloc_cq(). 2826 * 2827 * ensure that there are no other contexts that are posting WRs concurrently. 2828 * Otherwise the drain is not guaranteed. 2829 */ 2830 void ib_drain_rq(struct ib_qp *qp) 2831 { 2832 if (qp->device->ops.drain_rq) 2833 qp->device->ops.drain_rq(qp); 2834 else 2835 __ib_drain_rq(qp); 2836 trace_cq_drain_complete(qp->recv_cq); 2837 } 2838 EXPORT_SYMBOL(ib_drain_rq); 2839 2840 /** 2841 * ib_drain_qp() - Block until all CQEs have been consumed by the 2842 * application on both the RQ and SQ. 2843 * @qp: queue pair to drain 2844 * 2845 * The caller must: 2846 * 2847 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests 2848 * and completions. 2849 * 2850 * allocate the CQs using ib_alloc_cq(). 2851 * 2852 * ensure that there are no other contexts that are posting WRs concurrently. 2853 * Otherwise the drain is not guaranteed. 2854 */ 2855 void ib_drain_qp(struct ib_qp *qp) 2856 { 2857 ib_drain_sq(qp); 2858 if (!qp->srq) 2859 ib_drain_rq(qp); 2860 } 2861 EXPORT_SYMBOL(ib_drain_qp); 2862 2863 struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num, 2864 enum rdma_netdev_t type, const char *name, 2865 unsigned char name_assign_type, 2866 void (*setup)(struct net_device *)) 2867 { 2868 struct rdma_netdev_alloc_params params; 2869 struct net_device *netdev; 2870 int rc; 2871 2872 if (!device->ops.rdma_netdev_get_params) 2873 return ERR_PTR(-EOPNOTSUPP); 2874 2875 rc = device->ops.rdma_netdev_get_params(device, port_num, type, 2876 ¶ms); 2877 if (rc) 2878 return ERR_PTR(rc); 2879 2880 netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type, 2881 setup, params.txqs, params.rxqs); 2882 if (!netdev) 2883 return ERR_PTR(-ENOMEM); 2884 2885 return netdev; 2886 } 2887 EXPORT_SYMBOL(rdma_alloc_netdev); 2888 2889 int rdma_init_netdev(struct ib_device *device, u8 port_num, 2890 enum rdma_netdev_t type, const char *name, 2891 unsigned char name_assign_type, 2892 void (*setup)(struct net_device *), 2893 struct net_device *netdev) 2894 { 2895 struct rdma_netdev_alloc_params params; 2896 int rc; 2897 2898 if (!device->ops.rdma_netdev_get_params) 2899 return -EOPNOTSUPP; 2900 2901 rc = device->ops.rdma_netdev_get_params(device, port_num, type, 2902 ¶ms); 2903 if (rc) 2904 return rc; 2905 2906 return params.initialize_rdma_netdev(device, port_num, 2907 netdev, params.param); 2908 } 2909 EXPORT_SYMBOL(rdma_init_netdev); 2910 2911 void __rdma_block_iter_start(struct ib_block_iter *biter, 2912 struct scatterlist *sglist, unsigned int nents, 2913 unsigned long pgsz) 2914 { 2915 memset(biter, 0, sizeof(struct ib_block_iter)); 2916 biter->__sg = sglist; 2917 biter->__sg_nents = nents; 2918 2919 /* Driver provides best block size to use */ 2920 biter->__pg_bit = __fls(pgsz); 2921 } 2922 EXPORT_SYMBOL(__rdma_block_iter_start); 2923 2924 bool __rdma_block_iter_next(struct ib_block_iter *biter) 2925 { 2926 unsigned int block_offset; 2927 2928 if (!biter->__sg_nents || !biter->__sg) 2929 return false; 2930 2931 biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance; 2932 block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1); 2933 biter->__sg_advance += BIT_ULL(biter->__pg_bit) - block_offset; 2934 2935 if (biter->__sg_advance >= sg_dma_len(biter->__sg)) { 2936 biter->__sg_advance = 0; 2937 biter->__sg = sg_next(biter->__sg); 2938 biter->__sg_nents--; 2939 } 2940 2941 return true; 2942 } 2943 EXPORT_SYMBOL(__rdma_block_iter_next); 2944