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