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