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/ib_umem.h> 53 #include <rdma/rw.h> 54 #include <rdma/lag.h> 55 56 #include "rdma_core.h" 57 #include "core_priv.h" 58 #include <trace/events/rdma_core.h> 59 60 static int ib_resolve_eth_dmac(struct ib_device *device, 61 struct rdma_ah_attr *ah_attr); 62 63 static const char * const ib_events[] = { 64 [IB_EVENT_CQ_ERR] = "CQ error", 65 [IB_EVENT_QP_FATAL] = "QP fatal error", 66 [IB_EVENT_QP_REQ_ERR] = "QP request error", 67 [IB_EVENT_QP_ACCESS_ERR] = "QP access error", 68 [IB_EVENT_COMM_EST] = "communication established", 69 [IB_EVENT_SQ_DRAINED] = "send queue drained", 70 [IB_EVENT_PATH_MIG] = "path migration successful", 71 [IB_EVENT_PATH_MIG_ERR] = "path migration error", 72 [IB_EVENT_DEVICE_FATAL] = "device fatal error", 73 [IB_EVENT_PORT_ACTIVE] = "port active", 74 [IB_EVENT_PORT_ERR] = "port error", 75 [IB_EVENT_LID_CHANGE] = "LID change", 76 [IB_EVENT_PKEY_CHANGE] = "P_key change", 77 [IB_EVENT_SM_CHANGE] = "SM change", 78 [IB_EVENT_SRQ_ERR] = "SRQ error", 79 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached", 80 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached", 81 [IB_EVENT_CLIENT_REREGISTER] = "client reregister", 82 [IB_EVENT_GID_CHANGE] = "GID changed", 83 [IB_EVENT_DEVICE_SPEED_CHANGE] = "device speed change" 84 }; 85 86 const char *__attribute_const__ ib_event_msg(enum ib_event_type event) 87 { 88 size_t index = event; 89 90 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ? 91 ib_events[index] : "unrecognized event"; 92 } 93 EXPORT_SYMBOL(ib_event_msg); 94 95 static const char * const wc_statuses[] = { 96 [IB_WC_SUCCESS] = "success", 97 [IB_WC_LOC_LEN_ERR] = "local length error", 98 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error", 99 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error", 100 [IB_WC_LOC_PROT_ERR] = "local protection error", 101 [IB_WC_WR_FLUSH_ERR] = "WR flushed", 102 [IB_WC_MW_BIND_ERR] = "memory bind operation error", 103 [IB_WC_BAD_RESP_ERR] = "bad response error", 104 [IB_WC_LOC_ACCESS_ERR] = "local access error", 105 [IB_WC_REM_INV_REQ_ERR] = "remote invalid request error", 106 [IB_WC_REM_ACCESS_ERR] = "remote access error", 107 [IB_WC_REM_OP_ERR] = "remote operation error", 108 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded", 109 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded", 110 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error", 111 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request", 112 [IB_WC_REM_ABORT_ERR] = "operation aborted", 113 [IB_WC_INV_EECN_ERR] = "invalid EE context number", 114 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state", 115 [IB_WC_FATAL_ERR] = "fatal error", 116 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error", 117 [IB_WC_GENERAL_ERR] = "general error", 118 }; 119 120 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status) 121 { 122 size_t index = status; 123 124 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ? 125 wc_statuses[index] : "unrecognized status"; 126 } 127 EXPORT_SYMBOL(ib_wc_status_msg); 128 129 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate) 130 { 131 switch (rate) { 132 case IB_RATE_2_5_GBPS: return 1; 133 case IB_RATE_5_GBPS: return 2; 134 case IB_RATE_10_GBPS: return 4; 135 case IB_RATE_20_GBPS: return 8; 136 case IB_RATE_30_GBPS: return 12; 137 case IB_RATE_40_GBPS: return 16; 138 case IB_RATE_60_GBPS: return 24; 139 case IB_RATE_80_GBPS: return 32; 140 case IB_RATE_120_GBPS: return 48; 141 case IB_RATE_14_GBPS: return 6; 142 case IB_RATE_56_GBPS: return 22; 143 case IB_RATE_112_GBPS: return 45; 144 case IB_RATE_168_GBPS: return 67; 145 case IB_RATE_25_GBPS: return 10; 146 case IB_RATE_100_GBPS: return 40; 147 case IB_RATE_200_GBPS: return 80; 148 case IB_RATE_300_GBPS: return 120; 149 case IB_RATE_28_GBPS: return 11; 150 case IB_RATE_50_GBPS: return 20; 151 case IB_RATE_400_GBPS: return 160; 152 case IB_RATE_600_GBPS: return 240; 153 case IB_RATE_800_GBPS: return 320; 154 case IB_RATE_1600_GBPS: return 640; 155 default: return -1; 156 } 157 } 158 EXPORT_SYMBOL(ib_rate_to_mult); 159 160 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult) 161 { 162 switch (mult) { 163 case 1: return IB_RATE_2_5_GBPS; 164 case 2: return IB_RATE_5_GBPS; 165 case 4: return IB_RATE_10_GBPS; 166 case 8: return IB_RATE_20_GBPS; 167 case 12: return IB_RATE_30_GBPS; 168 case 16: return IB_RATE_40_GBPS; 169 case 24: return IB_RATE_60_GBPS; 170 case 32: return IB_RATE_80_GBPS; 171 case 48: return IB_RATE_120_GBPS; 172 case 6: return IB_RATE_14_GBPS; 173 case 22: return IB_RATE_56_GBPS; 174 case 45: return IB_RATE_112_GBPS; 175 case 67: return IB_RATE_168_GBPS; 176 case 10: return IB_RATE_25_GBPS; 177 case 40: return IB_RATE_100_GBPS; 178 case 80: return IB_RATE_200_GBPS; 179 case 120: return IB_RATE_300_GBPS; 180 case 11: return IB_RATE_28_GBPS; 181 case 20: return IB_RATE_50_GBPS; 182 case 160: return IB_RATE_400_GBPS; 183 case 240: return IB_RATE_600_GBPS; 184 case 320: return IB_RATE_800_GBPS; 185 case 640: return IB_RATE_1600_GBPS; 186 default: return IB_RATE_PORT_CURRENT; 187 } 188 } 189 EXPORT_SYMBOL(mult_to_ib_rate); 190 191 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate) 192 { 193 switch (rate) { 194 case IB_RATE_2_5_GBPS: return 2500; 195 case IB_RATE_5_GBPS: return 5000; 196 case IB_RATE_10_GBPS: return 10000; 197 case IB_RATE_20_GBPS: return 20000; 198 case IB_RATE_30_GBPS: return 30000; 199 case IB_RATE_40_GBPS: return 40000; 200 case IB_RATE_60_GBPS: return 60000; 201 case IB_RATE_80_GBPS: return 80000; 202 case IB_RATE_120_GBPS: return 120000; 203 case IB_RATE_14_GBPS: return 14062; 204 case IB_RATE_56_GBPS: return 56250; 205 case IB_RATE_112_GBPS: return 112500; 206 case IB_RATE_168_GBPS: return 168750; 207 case IB_RATE_25_GBPS: return 25781; 208 case IB_RATE_100_GBPS: return 103125; 209 case IB_RATE_200_GBPS: return 206250; 210 case IB_RATE_300_GBPS: return 309375; 211 case IB_RATE_28_GBPS: return 28125; 212 case IB_RATE_50_GBPS: return 53125; 213 case IB_RATE_400_GBPS: return 425000; 214 case IB_RATE_600_GBPS: return 637500; 215 case IB_RATE_800_GBPS: return 850000; 216 case IB_RATE_1600_GBPS: return 1700000; 217 default: return -1; 218 } 219 } 220 EXPORT_SYMBOL(ib_rate_to_mbps); 221 222 struct ib_speed_attr { 223 const char *str; 224 int speed; 225 }; 226 227 #define IB_SPEED_ATTR(speed_type, _str, _speed) \ 228 [speed_type] = {.str = _str, .speed = _speed} 229 230 static const struct ib_speed_attr ib_speed_attrs[] = { 231 IB_SPEED_ATTR(IB_SPEED_SDR, " SDR", 25), 232 IB_SPEED_ATTR(IB_SPEED_DDR, " DDR", 50), 233 IB_SPEED_ATTR(IB_SPEED_QDR, " QDR", 100), 234 IB_SPEED_ATTR(IB_SPEED_FDR10, " FDR10", 100), 235 IB_SPEED_ATTR(IB_SPEED_FDR, " FDR", 140), 236 IB_SPEED_ATTR(IB_SPEED_EDR, " EDR", 250), 237 IB_SPEED_ATTR(IB_SPEED_HDR, " HDR", 500), 238 IB_SPEED_ATTR(IB_SPEED_NDR, " NDR", 1000), 239 IB_SPEED_ATTR(IB_SPEED_XDR, " XDR", 2000), 240 }; 241 242 int ib_port_attr_to_speed_info(struct ib_port_attr *attr, 243 struct ib_port_speed_info *speed_info) 244 { 245 int speed_idx = attr->active_speed; 246 247 switch (attr->active_speed) { 248 case IB_SPEED_DDR: 249 case IB_SPEED_QDR: 250 case IB_SPEED_FDR10: 251 case IB_SPEED_FDR: 252 case IB_SPEED_EDR: 253 case IB_SPEED_HDR: 254 case IB_SPEED_NDR: 255 case IB_SPEED_XDR: 256 case IB_SPEED_SDR: 257 break; 258 default: 259 speed_idx = IB_SPEED_SDR; /* Default to SDR for invalid rates */ 260 break; 261 } 262 263 speed_info->str = ib_speed_attrs[speed_idx].str; 264 speed_info->rate = ib_speed_attrs[speed_idx].speed; 265 speed_info->rate *= ib_width_enum_to_int(attr->active_width); 266 if (speed_info->rate < 0) 267 return -EINVAL; 268 269 return 0; 270 } 271 EXPORT_SYMBOL(ib_port_attr_to_speed_info); 272 273 __attribute_const__ enum rdma_transport_type 274 rdma_node_get_transport(unsigned int node_type) 275 { 276 277 if (node_type == RDMA_NODE_USNIC) 278 return RDMA_TRANSPORT_USNIC; 279 if (node_type == RDMA_NODE_USNIC_UDP) 280 return RDMA_TRANSPORT_USNIC_UDP; 281 if (node_type == RDMA_NODE_RNIC) 282 return RDMA_TRANSPORT_IWARP; 283 if (node_type == RDMA_NODE_UNSPECIFIED) 284 return RDMA_TRANSPORT_UNSPECIFIED; 285 286 return RDMA_TRANSPORT_IB; 287 } 288 EXPORT_SYMBOL(rdma_node_get_transport); 289 290 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, 291 u32 port_num) 292 { 293 enum rdma_transport_type lt; 294 if (device->ops.get_link_layer) 295 return device->ops.get_link_layer(device, port_num); 296 297 lt = rdma_node_get_transport(device->node_type); 298 if (lt == RDMA_TRANSPORT_IB) 299 return IB_LINK_LAYER_INFINIBAND; 300 301 return IB_LINK_LAYER_ETHERNET; 302 } 303 EXPORT_SYMBOL(rdma_port_get_link_layer); 304 305 /* Protection domains */ 306 307 /** 308 * __ib_alloc_pd - Allocates an unused protection domain. 309 * @device: The device on which to allocate the protection domain. 310 * @flags: protection domain flags 311 * @caller: caller's build-time module name 312 * 313 * A protection domain object provides an association between QPs, shared 314 * receive queues, address handles, memory regions, and memory windows. 315 * 316 * Every PD has a local_dma_lkey which can be used as the lkey value for local 317 * memory operations. 318 */ 319 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags, 320 const char *caller) 321 { 322 struct ib_pd *pd; 323 int mr_access_flags = 0; 324 int ret; 325 326 pd = rdma_zalloc_drv_obj(device, ib_pd); 327 if (!pd) 328 return ERR_PTR(-ENOMEM); 329 330 pd->device = device; 331 pd->flags = flags; 332 333 rdma_restrack_new(&pd->res, RDMA_RESTRACK_PD); 334 rdma_restrack_set_name(&pd->res, caller); 335 336 ret = device->ops.alloc_pd(pd, NULL); 337 if (ret) { 338 rdma_restrack_put(&pd->res); 339 kfree(pd); 340 return ERR_PTR(ret); 341 } 342 rdma_restrack_add(&pd->res); 343 344 if (device->attrs.kernel_cap_flags & IBK_LOCAL_DMA_LKEY) 345 pd->local_dma_lkey = device->local_dma_lkey; 346 else 347 mr_access_flags |= IB_ACCESS_LOCAL_WRITE; 348 349 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) { 350 pr_warn("%s: enabling unsafe global rkey\n", caller); 351 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE; 352 } 353 354 if (mr_access_flags) { 355 struct ib_mr *mr; 356 357 mr = pd->device->ops.get_dma_mr(pd, mr_access_flags); 358 if (IS_ERR(mr)) { 359 ib_dealloc_pd(pd); 360 return ERR_CAST(mr); 361 } 362 363 mr->device = pd->device; 364 mr->pd = pd; 365 mr->type = IB_MR_TYPE_DMA; 366 mr->uobject = NULL; 367 mr->need_inval = false; 368 369 pd->__internal_mr = mr; 370 371 if (!(device->attrs.kernel_cap_flags & IBK_LOCAL_DMA_LKEY)) 372 pd->local_dma_lkey = pd->__internal_mr->lkey; 373 374 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) 375 pd->unsafe_global_rkey = pd->__internal_mr->rkey; 376 } 377 378 return pd; 379 } 380 EXPORT_SYMBOL(__ib_alloc_pd); 381 382 /** 383 * ib_dealloc_pd_user - Deallocates a protection domain. 384 * @pd: The protection domain to deallocate. 385 * @udata: Valid user data or NULL for kernel object 386 * 387 * It is an error to call this function while any resources in the pd still 388 * exist. The caller is responsible to synchronously destroy them and 389 * guarantee no new allocations will happen. 390 */ 391 int ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata) 392 { 393 int ret; 394 395 if (pd->__internal_mr) { 396 ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL); 397 WARN_ON(ret); 398 pd->__internal_mr = NULL; 399 } 400 401 ret = pd->device->ops.dealloc_pd(pd, udata); 402 if (ret) 403 return ret; 404 405 rdma_restrack_del(&pd->res); 406 kfree(pd); 407 return ret; 408 } 409 EXPORT_SYMBOL(ib_dealloc_pd_user); 410 411 /* Address handles */ 412 413 /** 414 * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination. 415 * @dest: Pointer to destination ah_attr. Contents of the destination 416 * pointer is assumed to be invalid and attribute are overwritten. 417 * @src: Pointer to source ah_attr. 418 */ 419 void rdma_copy_ah_attr(struct rdma_ah_attr *dest, 420 const struct rdma_ah_attr *src) 421 { 422 *dest = *src; 423 if (dest->grh.sgid_attr) 424 rdma_hold_gid_attr(dest->grh.sgid_attr); 425 } 426 EXPORT_SYMBOL(rdma_copy_ah_attr); 427 428 /** 429 * rdma_replace_ah_attr - Replace valid ah_attr with new one. 430 * @old: Pointer to existing ah_attr which needs to be replaced. 431 * old is assumed to be valid or zero'd 432 * @new: Pointer to the new ah_attr. 433 * 434 * rdma_replace_ah_attr() first releases any reference in the old ah_attr if 435 * old the ah_attr is valid; after that it copies the new attribute and holds 436 * the reference to the replaced ah_attr. 437 */ 438 void rdma_replace_ah_attr(struct rdma_ah_attr *old, 439 const struct rdma_ah_attr *new) 440 { 441 rdma_destroy_ah_attr(old); 442 *old = *new; 443 if (old->grh.sgid_attr) 444 rdma_hold_gid_attr(old->grh.sgid_attr); 445 } 446 EXPORT_SYMBOL(rdma_replace_ah_attr); 447 448 /** 449 * rdma_move_ah_attr - Move ah_attr pointed by source to destination. 450 * @dest: Pointer to destination ah_attr to copy to. 451 * dest is assumed to be valid or zero'd 452 * @src: Pointer to the new ah_attr. 453 * 454 * rdma_move_ah_attr() first releases any reference in the destination ah_attr 455 * if it is valid. This also transfers ownership of internal references from 456 * src to dest, making src invalid in the process. No new reference of the src 457 * ah_attr is taken. 458 */ 459 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src) 460 { 461 rdma_destroy_ah_attr(dest); 462 *dest = *src; 463 src->grh.sgid_attr = NULL; 464 } 465 EXPORT_SYMBOL(rdma_move_ah_attr); 466 467 /* 468 * Validate that the rdma_ah_attr is valid for the device before passing it 469 * off to the driver. 470 */ 471 static int rdma_check_ah_attr(struct ib_device *device, 472 struct rdma_ah_attr *ah_attr) 473 { 474 if (!rdma_is_port_valid(device, ah_attr->port_num)) 475 return -EINVAL; 476 477 if ((rdma_is_grh_required(device, ah_attr->port_num) || 478 ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) && 479 !(ah_attr->ah_flags & IB_AH_GRH)) 480 return -EINVAL; 481 482 if (ah_attr->grh.sgid_attr) { 483 /* 484 * Make sure the passed sgid_attr is consistent with the 485 * parameters 486 */ 487 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index || 488 ah_attr->grh.sgid_attr->port_num != ah_attr->port_num) 489 return -EINVAL; 490 } 491 return 0; 492 } 493 494 /* 495 * If the ah requires a GRH then ensure that sgid_attr pointer is filled in. 496 * On success the caller is responsible to call rdma_unfill_sgid_attr(). 497 */ 498 static int rdma_fill_sgid_attr(struct ib_device *device, 499 struct rdma_ah_attr *ah_attr, 500 const struct ib_gid_attr **old_sgid_attr) 501 { 502 const struct ib_gid_attr *sgid_attr; 503 struct ib_global_route *grh; 504 int ret; 505 506 *old_sgid_attr = ah_attr->grh.sgid_attr; 507 508 ret = rdma_check_ah_attr(device, ah_attr); 509 if (ret) 510 return ret; 511 512 if (!(ah_attr->ah_flags & IB_AH_GRH)) 513 return 0; 514 515 grh = rdma_ah_retrieve_grh(ah_attr); 516 if (grh->sgid_attr) 517 return 0; 518 519 sgid_attr = 520 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index); 521 if (IS_ERR(sgid_attr)) 522 return PTR_ERR(sgid_attr); 523 524 /* Move ownerhip of the kref into the ah_attr */ 525 grh->sgid_attr = sgid_attr; 526 return 0; 527 } 528 529 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr, 530 const struct ib_gid_attr *old_sgid_attr) 531 { 532 /* 533 * Fill didn't change anything, the caller retains ownership of 534 * whatever it passed 535 */ 536 if (ah_attr->grh.sgid_attr == old_sgid_attr) 537 return; 538 539 /* 540 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller 541 * doesn't see any change in the rdma_ah_attr. If we get here 542 * old_sgid_attr is NULL. 543 */ 544 rdma_destroy_ah_attr(ah_attr); 545 } 546 547 static const struct ib_gid_attr * 548 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr, 549 const struct ib_gid_attr *old_attr) 550 { 551 if (old_attr) 552 rdma_put_gid_attr(old_attr); 553 if (ah_attr->ah_flags & IB_AH_GRH) { 554 rdma_hold_gid_attr(ah_attr->grh.sgid_attr); 555 return ah_attr->grh.sgid_attr; 556 } 557 return NULL; 558 } 559 560 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd, 561 struct rdma_ah_attr *ah_attr, 562 u32 flags, 563 struct ib_udata *udata, 564 struct net_device *xmit_slave) 565 { 566 struct rdma_ah_init_attr init_attr = {}; 567 struct ib_device *device = pd->device; 568 struct ib_ah *ah; 569 int ret; 570 571 might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE); 572 573 if (!udata && !device->ops.create_ah) 574 return ERR_PTR(-EOPNOTSUPP); 575 576 ah = rdma_zalloc_drv_obj_gfp( 577 device, ib_ah, 578 (flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC); 579 if (!ah) 580 return ERR_PTR(-ENOMEM); 581 582 ah->device = device; 583 ah->pd = pd; 584 ah->type = ah_attr->type; 585 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL); 586 init_attr.ah_attr = ah_attr; 587 init_attr.flags = flags; 588 init_attr.xmit_slave = xmit_slave; 589 590 if (udata) 591 ret = device->ops.create_user_ah(ah, &init_attr, udata); 592 else 593 ret = device->ops.create_ah(ah, &init_attr, NULL); 594 if (ret) { 595 if (ah->sgid_attr) 596 rdma_put_gid_attr(ah->sgid_attr); 597 kfree(ah); 598 return ERR_PTR(ret); 599 } 600 601 atomic_inc(&pd->usecnt); 602 return ah; 603 } 604 605 /** 606 * rdma_create_ah - Creates an address handle for the 607 * given address vector. 608 * @pd: The protection domain associated with the address handle. 609 * @ah_attr: The attributes of the address vector. 610 * @flags: Create address handle flags (see enum rdma_create_ah_flags). 611 * 612 * It returns 0 on success and returns appropriate error code on error. 613 * The address handle is used to reference a local or global destination 614 * in all UD QP post sends. 615 */ 616 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr, 617 u32 flags) 618 { 619 const struct ib_gid_attr *old_sgid_attr; 620 struct net_device *slave; 621 struct ib_ah *ah; 622 int ret; 623 624 ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr); 625 if (ret) 626 return ERR_PTR(ret); 627 slave = rdma_lag_get_ah_roce_slave(pd->device, ah_attr, 628 (flags & RDMA_CREATE_AH_SLEEPABLE) ? 629 GFP_KERNEL : GFP_ATOMIC); 630 if (IS_ERR(slave)) { 631 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 632 return ERR_CAST(slave); 633 } 634 ah = _rdma_create_ah(pd, ah_attr, flags, NULL, slave); 635 rdma_lag_put_ah_roce_slave(slave); 636 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 637 return ah; 638 } 639 EXPORT_SYMBOL(rdma_create_ah); 640 641 /** 642 * rdma_create_user_ah - Creates an address handle for the 643 * given address vector. 644 * It resolves destination mac address for ah attribute of RoCE type. 645 * @pd: The protection domain associated with the address handle. 646 * @ah_attr: The attributes of the address vector. 647 * @udata: pointer to user's input output buffer information need by 648 * provider driver. 649 * 650 * It returns 0 on success and returns appropriate error code on error. 651 * The address handle is used to reference a local or global destination 652 * in all UD QP post sends. 653 */ 654 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd, 655 struct rdma_ah_attr *ah_attr, 656 struct ib_udata *udata) 657 { 658 const struct ib_gid_attr *old_sgid_attr; 659 struct ib_ah *ah; 660 int err; 661 662 err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr); 663 if (err) 664 return ERR_PTR(err); 665 666 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) { 667 err = ib_resolve_eth_dmac(pd->device, ah_attr); 668 if (err) { 669 ah = ERR_PTR(err); 670 goto out; 671 } 672 } 673 674 ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE, 675 udata, NULL); 676 677 out: 678 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 679 return ah; 680 } 681 EXPORT_SYMBOL(rdma_create_user_ah); 682 683 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr) 684 { 685 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh; 686 struct iphdr ip4h_checked; 687 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh; 688 689 /* If it's IPv6, the version must be 6, otherwise, the first 690 * 20 bytes (before the IPv4 header) are garbled. 691 */ 692 if (ip6h->version != 6) 693 return (ip4h->version == 4) ? 4 : 0; 694 /* version may be 6 or 4 because the first 20 bytes could be garbled */ 695 696 /* RoCE v2 requires no options, thus header length 697 * must be 5 words 698 */ 699 if (ip4h->ihl != 5) 700 return 6; 701 702 /* Verify checksum. 703 * We can't write on scattered buffers so we need to copy to 704 * temp buffer. 705 */ 706 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked)); 707 ip4h_checked.check = 0; 708 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5); 709 /* if IPv4 header checksum is OK, believe it */ 710 if (ip4h->check == ip4h_checked.check) 711 return 4; 712 return 6; 713 } 714 EXPORT_SYMBOL(ib_get_rdma_header_version); 715 716 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device, 717 u32 port_num, 718 const struct ib_grh *grh) 719 { 720 int grh_version; 721 722 if (rdma_protocol_ib(device, port_num)) 723 return RDMA_NETWORK_IB; 724 725 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh); 726 727 if (grh_version == 4) 728 return RDMA_NETWORK_IPV4; 729 730 if (grh->next_hdr == IPPROTO_UDP) 731 return RDMA_NETWORK_IPV6; 732 733 return RDMA_NETWORK_ROCE_V1; 734 } 735 736 struct find_gid_index_context { 737 u16 vlan_id; 738 enum ib_gid_type gid_type; 739 }; 740 741 static bool find_gid_index(const union ib_gid *gid, 742 const struct ib_gid_attr *gid_attr, 743 void *context) 744 { 745 struct find_gid_index_context *ctx = context; 746 u16 vlan_id = 0xffff; 747 int ret; 748 749 if (ctx->gid_type != gid_attr->gid_type) 750 return false; 751 752 ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL); 753 if (ret) 754 return false; 755 756 return ctx->vlan_id == vlan_id; 757 } 758 759 static const struct ib_gid_attr * 760 get_sgid_attr_from_eth(struct ib_device *device, u32 port_num, 761 u16 vlan_id, const union ib_gid *sgid, 762 enum ib_gid_type gid_type) 763 { 764 struct find_gid_index_context context = {.vlan_id = vlan_id, 765 .gid_type = gid_type}; 766 767 return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index, 768 &context); 769 } 770 771 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr, 772 enum rdma_network_type net_type, 773 union ib_gid *sgid, union ib_gid *dgid) 774 { 775 struct sockaddr_in src_in; 776 struct sockaddr_in dst_in; 777 __be32 src_saddr, dst_saddr; 778 779 if (!sgid || !dgid) 780 return -EINVAL; 781 782 if (net_type == RDMA_NETWORK_IPV4) { 783 memcpy(&src_in.sin_addr.s_addr, 784 &hdr->roce4grh.saddr, 4); 785 memcpy(&dst_in.sin_addr.s_addr, 786 &hdr->roce4grh.daddr, 4); 787 src_saddr = src_in.sin_addr.s_addr; 788 dst_saddr = dst_in.sin_addr.s_addr; 789 ipv6_addr_set_v4mapped(src_saddr, 790 (struct in6_addr *)sgid); 791 ipv6_addr_set_v4mapped(dst_saddr, 792 (struct in6_addr *)dgid); 793 return 0; 794 } else if (net_type == RDMA_NETWORK_IPV6 || 795 net_type == RDMA_NETWORK_IB || net_type == RDMA_NETWORK_ROCE_V1) { 796 *dgid = hdr->ibgrh.dgid; 797 *sgid = hdr->ibgrh.sgid; 798 return 0; 799 } else { 800 return -EINVAL; 801 } 802 } 803 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr); 804 805 /* Resolve destination mac address and hop limit for unicast destination 806 * GID entry, considering the source GID entry as well. 807 * ah_attribute must have valid port_num, sgid_index. 808 */ 809 static int ib_resolve_unicast_gid_dmac(struct ib_device *device, 810 struct rdma_ah_attr *ah_attr) 811 { 812 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr); 813 const struct ib_gid_attr *sgid_attr = grh->sgid_attr; 814 int hop_limit = 0xff; 815 int ret = 0; 816 817 /* If destination is link local and source GID is RoCEv1, 818 * IP stack is not used. 819 */ 820 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) && 821 sgid_attr->gid_type == IB_GID_TYPE_ROCE) { 822 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw, 823 ah_attr->roce.dmac); 824 return ret; 825 } 826 827 ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid, 828 ah_attr->roce.dmac, 829 sgid_attr, &hop_limit); 830 831 grh->hop_limit = hop_limit; 832 return ret; 833 } 834 835 /* 836 * This function initializes address handle attributes from the incoming packet. 837 * Incoming packet has dgid of the receiver node on which this code is 838 * getting executed and, sgid contains the GID of the sender. 839 * 840 * When resolving mac address of destination, the arrived dgid is used 841 * as sgid and, sgid is used as dgid because sgid contains destinations 842 * GID whom to respond to. 843 * 844 * On success the caller is responsible to call rdma_destroy_ah_attr on the 845 * attr. 846 */ 847 int ib_init_ah_attr_from_wc(struct ib_device *device, u32 port_num, 848 const struct ib_wc *wc, const struct ib_grh *grh, 849 struct rdma_ah_attr *ah_attr) 850 { 851 u32 flow_class; 852 int ret; 853 enum rdma_network_type net_type = RDMA_NETWORK_IB; 854 enum ib_gid_type gid_type = IB_GID_TYPE_IB; 855 const struct ib_gid_attr *sgid_attr; 856 int hoplimit = 0xff; 857 union ib_gid dgid; 858 union ib_gid sgid; 859 860 might_sleep(); 861 862 memset(ah_attr, 0, sizeof *ah_attr); 863 ah_attr->type = rdma_ah_find_type(device, port_num); 864 if (rdma_cap_eth_ah(device, port_num)) { 865 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE) 866 net_type = wc->network_hdr_type; 867 else 868 net_type = ib_get_net_type_by_grh(device, port_num, grh); 869 gid_type = ib_network_to_gid_type(net_type); 870 } 871 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type, 872 &sgid, &dgid); 873 if (ret) 874 return ret; 875 876 rdma_ah_set_sl(ah_attr, wc->sl); 877 rdma_ah_set_port_num(ah_attr, port_num); 878 879 if (rdma_protocol_roce(device, port_num)) { 880 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ? 881 wc->vlan_id : 0xffff; 882 883 if (!(wc->wc_flags & IB_WC_GRH)) 884 return -EPROTOTYPE; 885 886 sgid_attr = get_sgid_attr_from_eth(device, port_num, 887 vlan_id, &dgid, 888 gid_type); 889 if (IS_ERR(sgid_attr)) 890 return PTR_ERR(sgid_attr); 891 892 flow_class = be32_to_cpu(grh->version_tclass_flow); 893 rdma_move_grh_sgid_attr(ah_attr, 894 &sgid, 895 flow_class & 0xFFFFF, 896 hoplimit, 897 (flow_class >> 20) & 0xFF, 898 sgid_attr); 899 900 ret = ib_resolve_unicast_gid_dmac(device, ah_attr); 901 if (ret) 902 rdma_destroy_ah_attr(ah_attr); 903 904 return ret; 905 } else { 906 rdma_ah_set_dlid(ah_attr, wc->slid); 907 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits); 908 909 if ((wc->wc_flags & IB_WC_GRH) == 0) 910 return 0; 911 912 if (dgid.global.interface_id != 913 cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) { 914 sgid_attr = rdma_find_gid_by_port( 915 device, &dgid, IB_GID_TYPE_IB, port_num, NULL); 916 } else 917 sgid_attr = rdma_get_gid_attr(device, port_num, 0); 918 919 if (IS_ERR(sgid_attr)) 920 return PTR_ERR(sgid_attr); 921 flow_class = be32_to_cpu(grh->version_tclass_flow); 922 rdma_move_grh_sgid_attr(ah_attr, 923 &sgid, 924 flow_class & 0xFFFFF, 925 hoplimit, 926 (flow_class >> 20) & 0xFF, 927 sgid_attr); 928 929 return 0; 930 } 931 } 932 EXPORT_SYMBOL(ib_init_ah_attr_from_wc); 933 934 /** 935 * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership 936 * of the reference 937 * 938 * @attr: Pointer to AH attribute structure 939 * @dgid: Destination GID 940 * @flow_label: Flow label 941 * @hop_limit: Hop limit 942 * @traffic_class: traffic class 943 * @sgid_attr: Pointer to SGID attribute 944 * 945 * This takes ownership of the sgid_attr reference. The caller must ensure 946 * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after 947 * calling this function. 948 */ 949 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid, 950 u32 flow_label, u8 hop_limit, u8 traffic_class, 951 const struct ib_gid_attr *sgid_attr) 952 { 953 rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit, 954 traffic_class); 955 attr->grh.sgid_attr = sgid_attr; 956 } 957 EXPORT_SYMBOL(rdma_move_grh_sgid_attr); 958 959 /** 960 * rdma_destroy_ah_attr - Release reference to SGID attribute of 961 * ah attribute. 962 * @ah_attr: Pointer to ah attribute 963 * 964 * Release reference to the SGID attribute of the ah attribute if it is 965 * non NULL. It is safe to call this multiple times, and safe to call it on 966 * a zero initialized ah_attr. 967 */ 968 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr) 969 { 970 if (ah_attr->grh.sgid_attr) { 971 rdma_put_gid_attr(ah_attr->grh.sgid_attr); 972 ah_attr->grh.sgid_attr = NULL; 973 } 974 } 975 EXPORT_SYMBOL(rdma_destroy_ah_attr); 976 977 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc, 978 const struct ib_grh *grh, u32 port_num) 979 { 980 struct rdma_ah_attr ah_attr; 981 struct ib_ah *ah; 982 int ret; 983 984 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr); 985 if (ret) 986 return ERR_PTR(ret); 987 988 ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE); 989 990 rdma_destroy_ah_attr(&ah_attr); 991 return ah; 992 } 993 EXPORT_SYMBOL(ib_create_ah_from_wc); 994 995 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr) 996 { 997 const struct ib_gid_attr *old_sgid_attr; 998 int ret; 999 1000 if (ah->type != ah_attr->type) 1001 return -EINVAL; 1002 1003 ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr); 1004 if (ret) 1005 return ret; 1006 1007 ret = ah->device->ops.modify_ah ? 1008 ah->device->ops.modify_ah(ah, ah_attr) : 1009 -EOPNOTSUPP; 1010 1011 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr); 1012 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr); 1013 return ret; 1014 } 1015 EXPORT_SYMBOL(rdma_modify_ah); 1016 1017 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr) 1018 { 1019 ah_attr->grh.sgid_attr = NULL; 1020 1021 return ah->device->ops.query_ah ? 1022 ah->device->ops.query_ah(ah, ah_attr) : 1023 -EOPNOTSUPP; 1024 } 1025 EXPORT_SYMBOL(rdma_query_ah); 1026 1027 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata) 1028 { 1029 const struct ib_gid_attr *sgid_attr = ah->sgid_attr; 1030 struct ib_pd *pd; 1031 int ret; 1032 1033 might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE); 1034 1035 pd = ah->pd; 1036 1037 ret = ah->device->ops.destroy_ah(ah, flags); 1038 if (ret) 1039 return ret; 1040 1041 atomic_dec(&pd->usecnt); 1042 if (sgid_attr) 1043 rdma_put_gid_attr(sgid_attr); 1044 1045 kfree(ah); 1046 return ret; 1047 } 1048 EXPORT_SYMBOL(rdma_destroy_ah_user); 1049 1050 /* Shared receive queues */ 1051 1052 /** 1053 * ib_create_srq_user - Creates a SRQ associated with the specified protection 1054 * domain. 1055 * @pd: The protection domain associated with the SRQ. 1056 * @srq_init_attr: A list of initial attributes required to create the 1057 * SRQ. If SRQ creation succeeds, then the attributes are updated to 1058 * the actual capabilities of the created SRQ. 1059 * @uobject: uobject pointer if this is not a kernel SRQ 1060 * @udata: udata pointer if this is not a kernel SRQ 1061 * 1062 * srq_attr->max_wr and srq_attr->max_sge are read the determine the 1063 * requested size of the SRQ, and set to the actual values allocated 1064 * on return. If ib_create_srq() succeeds, then max_wr and max_sge 1065 * will always be at least as large as the requested values. 1066 */ 1067 struct ib_srq *ib_create_srq_user(struct ib_pd *pd, 1068 struct ib_srq_init_attr *srq_init_attr, 1069 struct ib_usrq_object *uobject, 1070 struct ib_udata *udata) 1071 { 1072 struct ib_srq *srq; 1073 int ret; 1074 1075 srq = rdma_zalloc_drv_obj(pd->device, ib_srq); 1076 if (!srq) 1077 return ERR_PTR(-ENOMEM); 1078 1079 srq->device = pd->device; 1080 srq->pd = pd; 1081 srq->event_handler = srq_init_attr->event_handler; 1082 srq->srq_context = srq_init_attr->srq_context; 1083 srq->srq_type = srq_init_attr->srq_type; 1084 srq->uobject = uobject; 1085 1086 if (ib_srq_has_cq(srq->srq_type)) { 1087 srq->ext.cq = srq_init_attr->ext.cq; 1088 atomic_inc(&srq->ext.cq->usecnt); 1089 } 1090 if (srq->srq_type == IB_SRQT_XRC) { 1091 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd; 1092 if (srq->ext.xrc.xrcd) 1093 atomic_inc(&srq->ext.xrc.xrcd->usecnt); 1094 } 1095 atomic_inc(&pd->usecnt); 1096 1097 rdma_restrack_new(&srq->res, RDMA_RESTRACK_SRQ); 1098 rdma_restrack_parent_name(&srq->res, &pd->res); 1099 1100 ret = pd->device->ops.create_srq(srq, srq_init_attr, udata); 1101 if (ret) { 1102 rdma_restrack_put(&srq->res); 1103 atomic_dec(&pd->usecnt); 1104 if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd) 1105 atomic_dec(&srq->ext.xrc.xrcd->usecnt); 1106 if (ib_srq_has_cq(srq->srq_type)) 1107 atomic_dec(&srq->ext.cq->usecnt); 1108 kfree(srq); 1109 return ERR_PTR(ret); 1110 } 1111 1112 rdma_restrack_add(&srq->res); 1113 1114 return srq; 1115 } 1116 EXPORT_SYMBOL(ib_create_srq_user); 1117 1118 int ib_modify_srq(struct ib_srq *srq, 1119 struct ib_srq_attr *srq_attr, 1120 enum ib_srq_attr_mask srq_attr_mask) 1121 { 1122 return srq->device->ops.modify_srq ? 1123 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask, 1124 NULL) : -EOPNOTSUPP; 1125 } 1126 EXPORT_SYMBOL(ib_modify_srq); 1127 1128 int ib_query_srq(struct ib_srq *srq, 1129 struct ib_srq_attr *srq_attr) 1130 { 1131 return srq->device->ops.query_srq ? 1132 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP; 1133 } 1134 EXPORT_SYMBOL(ib_query_srq); 1135 1136 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata) 1137 { 1138 int ret; 1139 1140 if (atomic_read(&srq->usecnt)) 1141 return -EBUSY; 1142 1143 ret = srq->device->ops.destroy_srq(srq, udata); 1144 if (ret) 1145 return ret; 1146 1147 atomic_dec(&srq->pd->usecnt); 1148 if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd) 1149 atomic_dec(&srq->ext.xrc.xrcd->usecnt); 1150 if (ib_srq_has_cq(srq->srq_type)) 1151 atomic_dec(&srq->ext.cq->usecnt); 1152 rdma_restrack_del(&srq->res); 1153 kfree(srq); 1154 1155 return ret; 1156 } 1157 EXPORT_SYMBOL(ib_destroy_srq_user); 1158 1159 /* Queue pairs */ 1160 1161 static void __ib_qp_event_handler(struct ib_event *event, void *context) 1162 { 1163 struct ib_qp *qp = event->element.qp; 1164 1165 if (event->event == IB_EVENT_QP_LAST_WQE_REACHED) 1166 complete(&qp->srq_completion); 1167 if (qp->registered_event_handler) 1168 qp->registered_event_handler(event, qp->qp_context); 1169 } 1170 1171 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context) 1172 { 1173 struct ib_qp *qp = context; 1174 unsigned long flags; 1175 1176 spin_lock_irqsave(&qp->device->qp_open_list_lock, flags); 1177 list_for_each_entry(event->element.qp, &qp->open_list, open_list) 1178 if (event->element.qp->event_handler) 1179 event->element.qp->event_handler(event, event->element.qp->qp_context); 1180 spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags); 1181 } 1182 1183 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp, 1184 void (*event_handler)(struct ib_event *, void *), 1185 void *qp_context) 1186 { 1187 struct ib_qp *qp; 1188 unsigned long flags; 1189 int err; 1190 1191 qp = kzalloc_obj(*qp); 1192 if (!qp) 1193 return ERR_PTR(-ENOMEM); 1194 1195 qp->real_qp = real_qp; 1196 err = ib_open_shared_qp_security(qp, real_qp->device); 1197 if (err) { 1198 kfree(qp); 1199 return ERR_PTR(err); 1200 } 1201 1202 qp->real_qp = real_qp; 1203 atomic_inc(&real_qp->usecnt); 1204 qp->device = real_qp->device; 1205 qp->event_handler = event_handler; 1206 qp->qp_context = qp_context; 1207 qp->qp_num = real_qp->qp_num; 1208 qp->qp_type = real_qp->qp_type; 1209 1210 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags); 1211 list_add(&qp->open_list, &real_qp->open_list); 1212 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags); 1213 1214 return qp; 1215 } 1216 1217 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd, 1218 struct ib_qp_open_attr *qp_open_attr) 1219 { 1220 struct ib_qp *qp, *real_qp; 1221 1222 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT) 1223 return ERR_PTR(-EINVAL); 1224 1225 down_read(&xrcd->tgt_qps_rwsem); 1226 real_qp = xa_load(&xrcd->tgt_qps, qp_open_attr->qp_num); 1227 if (!real_qp) { 1228 up_read(&xrcd->tgt_qps_rwsem); 1229 return ERR_PTR(-EINVAL); 1230 } 1231 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler, 1232 qp_open_attr->qp_context); 1233 up_read(&xrcd->tgt_qps_rwsem); 1234 return qp; 1235 } 1236 EXPORT_SYMBOL(ib_open_qp); 1237 1238 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp, 1239 struct ib_qp_init_attr *qp_init_attr) 1240 { 1241 struct ib_qp *real_qp = qp; 1242 int err; 1243 1244 qp->event_handler = __ib_shared_qp_event_handler; 1245 qp->qp_context = qp; 1246 qp->pd = NULL; 1247 qp->send_cq = qp->recv_cq = NULL; 1248 qp->srq = NULL; 1249 qp->xrcd = qp_init_attr->xrcd; 1250 atomic_inc(&qp_init_attr->xrcd->usecnt); 1251 INIT_LIST_HEAD(&qp->open_list); 1252 1253 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler, 1254 qp_init_attr->qp_context); 1255 if (IS_ERR(qp)) 1256 return qp; 1257 1258 err = xa_err(xa_store(&qp_init_attr->xrcd->tgt_qps, real_qp->qp_num, 1259 real_qp, GFP_KERNEL)); 1260 if (err) { 1261 ib_close_qp(qp); 1262 return ERR_PTR(err); 1263 } 1264 return qp; 1265 } 1266 1267 static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd, 1268 struct ib_qp_init_attr *attr, 1269 struct uverbs_attr_bundle *uattrs, 1270 struct ib_uqp_object *uobj, const char *caller) 1271 { 1272 struct ib_qp *qp; 1273 int ret; 1274 1275 if (!dev->ops.create_qp) 1276 return ERR_PTR(-EOPNOTSUPP); 1277 1278 qp = rdma_zalloc_drv_obj_numa(dev, ib_qp); 1279 if (!qp) 1280 return ERR_PTR(-ENOMEM); 1281 1282 qp->device = dev; 1283 qp->pd = pd; 1284 qp->uobject = uobj; 1285 qp->real_qp = qp; 1286 1287 qp->qp_type = attr->qp_type; 1288 qp->rwq_ind_tbl = attr->rwq_ind_tbl; 1289 qp->srq = attr->srq; 1290 qp->event_handler = __ib_qp_event_handler; 1291 qp->registered_event_handler = attr->event_handler; 1292 qp->port = attr->port_num; 1293 qp->qp_context = attr->qp_context; 1294 1295 spin_lock_init(&qp->mr_lock); 1296 INIT_LIST_HEAD(&qp->rdma_mrs); 1297 INIT_LIST_HEAD(&qp->sig_mrs); 1298 init_completion(&qp->srq_completion); 1299 1300 qp->send_cq = attr->send_cq; 1301 qp->recv_cq = attr->recv_cq; 1302 1303 rdma_restrack_new(&qp->res, RDMA_RESTRACK_QP); 1304 WARN_ONCE(!uattrs && !caller, "Missing kernel QP owner"); 1305 rdma_restrack_set_name(&qp->res, uattrs ? NULL : caller); 1306 ret = dev->ops.create_qp(qp, attr, 1307 uattrs ? &uattrs->driver_udata : NULL); 1308 if (ret) 1309 goto err_create; 1310 1311 /* 1312 * TODO: The mlx4 internally overwrites send_cq and recv_cq. 1313 * Unfortunately, it is not an easy task to fix that driver. 1314 */ 1315 qp->send_cq = attr->send_cq; 1316 qp->recv_cq = attr->recv_cq; 1317 1318 ret = ib_create_qp_security(qp, dev); 1319 if (ret) 1320 goto err_security; 1321 1322 rdma_restrack_add(&qp->res); 1323 return qp; 1324 1325 err_security: 1326 qp->device->ops.destroy_qp( 1327 qp, uattrs ? uverbs_get_cleared_udata(uattrs) : NULL); 1328 err_create: 1329 rdma_restrack_put(&qp->res); 1330 kfree(qp); 1331 return ERR_PTR(ret); 1332 1333 } 1334 1335 /** 1336 * ib_create_qp_user - Creates a QP associated with the specified protection 1337 * domain. 1338 * @dev: IB device 1339 * @pd: The protection domain associated with the QP. 1340 * @attr: A list of initial attributes required to create the 1341 * QP. If QP creation succeeds, then the attributes are updated to 1342 * the actual capabilities of the created QP. 1343 * @uattrs: User ioctl attributes and udata 1344 * @uobj: uverbs obect 1345 * @caller: caller's build-time module name 1346 */ 1347 struct ib_qp *ib_create_qp_user(struct ib_device *dev, struct ib_pd *pd, 1348 struct ib_qp_init_attr *attr, 1349 struct uverbs_attr_bundle *uattrs, 1350 struct ib_uqp_object *uobj, const char *caller) 1351 { 1352 struct ib_qp *qp, *xrc_qp; 1353 1354 if (attr->qp_type == IB_QPT_XRC_TGT) 1355 qp = create_qp(dev, pd, attr, NULL, NULL, caller); 1356 else 1357 qp = create_qp(dev, pd, attr, uattrs, uobj, NULL); 1358 if (attr->qp_type != IB_QPT_XRC_TGT || IS_ERR(qp)) 1359 return qp; 1360 1361 xrc_qp = create_xrc_qp_user(qp, attr); 1362 if (IS_ERR(xrc_qp)) { 1363 ib_destroy_qp(qp); 1364 return xrc_qp; 1365 } 1366 1367 xrc_qp->uobject = uobj; 1368 return xrc_qp; 1369 } 1370 EXPORT_SYMBOL(ib_create_qp_user); 1371 1372 void ib_qp_usecnt_inc(struct ib_qp *qp) 1373 { 1374 if (qp->pd) 1375 atomic_inc(&qp->pd->usecnt); 1376 if (qp->send_cq) 1377 atomic_inc(&qp->send_cq->usecnt); 1378 if (qp->recv_cq) 1379 atomic_inc(&qp->recv_cq->usecnt); 1380 if (qp->srq) 1381 atomic_inc(&qp->srq->usecnt); 1382 if (qp->rwq_ind_tbl) 1383 atomic_inc(&qp->rwq_ind_tbl->usecnt); 1384 } 1385 EXPORT_SYMBOL(ib_qp_usecnt_inc); 1386 1387 void ib_qp_usecnt_dec(struct ib_qp *qp) 1388 { 1389 if (qp->rwq_ind_tbl) 1390 atomic_dec(&qp->rwq_ind_tbl->usecnt); 1391 if (qp->srq) 1392 atomic_dec(&qp->srq->usecnt); 1393 if (qp->recv_cq) 1394 atomic_dec(&qp->recv_cq->usecnt); 1395 if (qp->send_cq) 1396 atomic_dec(&qp->send_cq->usecnt); 1397 if (qp->pd) 1398 atomic_dec(&qp->pd->usecnt); 1399 } 1400 EXPORT_SYMBOL(ib_qp_usecnt_dec); 1401 1402 struct ib_qp *ib_create_qp_kernel(struct ib_pd *pd, 1403 struct ib_qp_init_attr *qp_init_attr, 1404 const char *caller) 1405 { 1406 struct ib_device *device = pd->device; 1407 struct ib_qp *qp; 1408 int ret; 1409 1410 /* 1411 * If the callers is using the RDMA API calculate the resources 1412 * needed for the RDMA READ/WRITE operations. 1413 * 1414 * Note that these callers need to pass in a port number. 1415 */ 1416 if (qp_init_attr->cap.max_rdma_ctxs) 1417 rdma_rw_init_qp(device, qp_init_attr); 1418 1419 qp = create_qp(device, pd, qp_init_attr, NULL, NULL, caller); 1420 if (IS_ERR(qp)) 1421 return qp; 1422 1423 ib_qp_usecnt_inc(qp); 1424 1425 if (qp_init_attr->cap.max_rdma_ctxs) { 1426 ret = rdma_rw_init_mrs(qp, qp_init_attr); 1427 if (ret) 1428 goto err; 1429 } 1430 1431 /* 1432 * Note: all hw drivers guarantee that max_send_sge is lower than 1433 * the device RDMA WRITE SGE limit but not all hw drivers ensure that 1434 * max_send_sge <= max_sge_rd. 1435 */ 1436 qp->max_write_sge = qp_init_attr->cap.max_send_sge; 1437 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge, 1438 device->attrs.max_sge_rd); 1439 if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) 1440 qp->integrity_en = true; 1441 1442 return qp; 1443 1444 err: 1445 ib_destroy_qp(qp); 1446 return ERR_PTR(ret); 1447 1448 } 1449 EXPORT_SYMBOL(ib_create_qp_kernel); 1450 1451 static const struct { 1452 int valid; 1453 enum ib_qp_attr_mask req_param[IB_QPT_MAX]; 1454 enum ib_qp_attr_mask opt_param[IB_QPT_MAX]; 1455 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = { 1456 [IB_QPS_RESET] = { 1457 [IB_QPS_RESET] = { .valid = 1 }, 1458 [IB_QPS_INIT] = { 1459 .valid = 1, 1460 .req_param = { 1461 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1462 IB_QP_PORT | 1463 IB_QP_QKEY), 1464 [IB_QPT_RAW_PACKET] = IB_QP_PORT, 1465 [IB_QPT_UC] = (IB_QP_PKEY_INDEX | 1466 IB_QP_PORT | 1467 IB_QP_ACCESS_FLAGS), 1468 [IB_QPT_RC] = (IB_QP_PKEY_INDEX | 1469 IB_QP_PORT | 1470 IB_QP_ACCESS_FLAGS), 1471 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX | 1472 IB_QP_PORT | 1473 IB_QP_ACCESS_FLAGS), 1474 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX | 1475 IB_QP_PORT | 1476 IB_QP_ACCESS_FLAGS), 1477 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1478 IB_QP_QKEY), 1479 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1480 IB_QP_QKEY), 1481 } 1482 }, 1483 }, 1484 [IB_QPS_INIT] = { 1485 [IB_QPS_RESET] = { .valid = 1 }, 1486 [IB_QPS_ERR] = { .valid = 1 }, 1487 [IB_QPS_INIT] = { 1488 .valid = 1, 1489 .opt_param = { 1490 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1491 IB_QP_PORT | 1492 IB_QP_QKEY), 1493 [IB_QPT_UC] = (IB_QP_PKEY_INDEX | 1494 IB_QP_PORT | 1495 IB_QP_ACCESS_FLAGS), 1496 [IB_QPT_RC] = (IB_QP_PKEY_INDEX | 1497 IB_QP_PORT | 1498 IB_QP_ACCESS_FLAGS), 1499 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX | 1500 IB_QP_PORT | 1501 IB_QP_ACCESS_FLAGS), 1502 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX | 1503 IB_QP_PORT | 1504 IB_QP_ACCESS_FLAGS), 1505 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1506 IB_QP_QKEY), 1507 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1508 IB_QP_QKEY), 1509 } 1510 }, 1511 [IB_QPS_RTR] = { 1512 .valid = 1, 1513 .req_param = { 1514 [IB_QPT_UC] = (IB_QP_AV | 1515 IB_QP_PATH_MTU | 1516 IB_QP_DEST_QPN | 1517 IB_QP_RQ_PSN), 1518 [IB_QPT_RC] = (IB_QP_AV | 1519 IB_QP_PATH_MTU | 1520 IB_QP_DEST_QPN | 1521 IB_QP_RQ_PSN | 1522 IB_QP_MAX_DEST_RD_ATOMIC | 1523 IB_QP_MIN_RNR_TIMER), 1524 [IB_QPT_XRC_INI] = (IB_QP_AV | 1525 IB_QP_PATH_MTU | 1526 IB_QP_DEST_QPN | 1527 IB_QP_RQ_PSN), 1528 [IB_QPT_XRC_TGT] = (IB_QP_AV | 1529 IB_QP_PATH_MTU | 1530 IB_QP_DEST_QPN | 1531 IB_QP_RQ_PSN | 1532 IB_QP_MAX_DEST_RD_ATOMIC | 1533 IB_QP_MIN_RNR_TIMER), 1534 }, 1535 .opt_param = { 1536 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1537 IB_QP_QKEY), 1538 [IB_QPT_UC] = (IB_QP_ALT_PATH | 1539 IB_QP_ACCESS_FLAGS | 1540 IB_QP_PKEY_INDEX), 1541 [IB_QPT_RC] = (IB_QP_ALT_PATH | 1542 IB_QP_ACCESS_FLAGS | 1543 IB_QP_PKEY_INDEX), 1544 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH | 1545 IB_QP_ACCESS_FLAGS | 1546 IB_QP_PKEY_INDEX), 1547 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH | 1548 IB_QP_ACCESS_FLAGS | 1549 IB_QP_PKEY_INDEX), 1550 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1551 IB_QP_QKEY), 1552 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1553 IB_QP_QKEY), 1554 }, 1555 }, 1556 }, 1557 [IB_QPS_RTR] = { 1558 [IB_QPS_RESET] = { .valid = 1 }, 1559 [IB_QPS_ERR] = { .valid = 1 }, 1560 [IB_QPS_RTS] = { 1561 .valid = 1, 1562 .req_param = { 1563 [IB_QPT_UD] = IB_QP_SQ_PSN, 1564 [IB_QPT_UC] = IB_QP_SQ_PSN, 1565 [IB_QPT_RC] = (IB_QP_TIMEOUT | 1566 IB_QP_RETRY_CNT | 1567 IB_QP_RNR_RETRY | 1568 IB_QP_SQ_PSN | 1569 IB_QP_MAX_QP_RD_ATOMIC), 1570 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT | 1571 IB_QP_RETRY_CNT | 1572 IB_QP_RNR_RETRY | 1573 IB_QP_SQ_PSN | 1574 IB_QP_MAX_QP_RD_ATOMIC), 1575 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT | 1576 IB_QP_SQ_PSN), 1577 [IB_QPT_SMI] = IB_QP_SQ_PSN, 1578 [IB_QPT_GSI] = IB_QP_SQ_PSN, 1579 }, 1580 .opt_param = { 1581 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1582 IB_QP_QKEY), 1583 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1584 IB_QP_ALT_PATH | 1585 IB_QP_ACCESS_FLAGS | 1586 IB_QP_PATH_MIG_STATE), 1587 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1588 IB_QP_ALT_PATH | 1589 IB_QP_ACCESS_FLAGS | 1590 IB_QP_MIN_RNR_TIMER | 1591 IB_QP_PATH_MIG_STATE), 1592 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1593 IB_QP_ALT_PATH | 1594 IB_QP_ACCESS_FLAGS | 1595 IB_QP_PATH_MIG_STATE), 1596 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1597 IB_QP_ALT_PATH | 1598 IB_QP_ACCESS_FLAGS | 1599 IB_QP_MIN_RNR_TIMER | 1600 IB_QP_PATH_MIG_STATE), 1601 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1602 IB_QP_QKEY), 1603 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1604 IB_QP_QKEY), 1605 } 1606 } 1607 }, 1608 [IB_QPS_RTS] = { 1609 [IB_QPS_RESET] = { .valid = 1 }, 1610 [IB_QPS_ERR] = { .valid = 1 }, 1611 [IB_QPS_RTS] = { 1612 .valid = 1, 1613 .opt_param = { 1614 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1615 IB_QP_QKEY), 1616 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1617 IB_QP_ACCESS_FLAGS | 1618 IB_QP_ALT_PATH | 1619 IB_QP_PATH_MIG_STATE), 1620 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1621 IB_QP_ACCESS_FLAGS | 1622 IB_QP_ALT_PATH | 1623 IB_QP_PATH_MIG_STATE | 1624 IB_QP_MIN_RNR_TIMER), 1625 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1626 IB_QP_ACCESS_FLAGS | 1627 IB_QP_ALT_PATH | 1628 IB_QP_PATH_MIG_STATE), 1629 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1630 IB_QP_ACCESS_FLAGS | 1631 IB_QP_ALT_PATH | 1632 IB_QP_PATH_MIG_STATE | 1633 IB_QP_MIN_RNR_TIMER), 1634 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1635 IB_QP_QKEY), 1636 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1637 IB_QP_QKEY), 1638 } 1639 }, 1640 [IB_QPS_SQD] = { 1641 .valid = 1, 1642 .opt_param = { 1643 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1644 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1645 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1646 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1647 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */ 1648 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1649 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY 1650 } 1651 }, 1652 }, 1653 [IB_QPS_SQD] = { 1654 [IB_QPS_RESET] = { .valid = 1 }, 1655 [IB_QPS_ERR] = { .valid = 1 }, 1656 [IB_QPS_RTS] = { 1657 .valid = 1, 1658 .opt_param = { 1659 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1660 IB_QP_QKEY), 1661 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1662 IB_QP_ALT_PATH | 1663 IB_QP_ACCESS_FLAGS | 1664 IB_QP_PATH_MIG_STATE), 1665 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1666 IB_QP_ALT_PATH | 1667 IB_QP_ACCESS_FLAGS | 1668 IB_QP_MIN_RNR_TIMER | 1669 IB_QP_PATH_MIG_STATE), 1670 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1671 IB_QP_ALT_PATH | 1672 IB_QP_ACCESS_FLAGS | 1673 IB_QP_PATH_MIG_STATE), 1674 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1675 IB_QP_ALT_PATH | 1676 IB_QP_ACCESS_FLAGS | 1677 IB_QP_MIN_RNR_TIMER | 1678 IB_QP_PATH_MIG_STATE), 1679 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1680 IB_QP_QKEY), 1681 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1682 IB_QP_QKEY), 1683 } 1684 }, 1685 [IB_QPS_SQD] = { 1686 .valid = 1, 1687 .opt_param = { 1688 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1689 IB_QP_QKEY), 1690 [IB_QPT_UC] = (IB_QP_AV | 1691 IB_QP_ALT_PATH | 1692 IB_QP_ACCESS_FLAGS | 1693 IB_QP_PKEY_INDEX | 1694 IB_QP_PATH_MIG_STATE), 1695 [IB_QPT_RC] = (IB_QP_PORT | 1696 IB_QP_AV | 1697 IB_QP_TIMEOUT | 1698 IB_QP_RETRY_CNT | 1699 IB_QP_RNR_RETRY | 1700 IB_QP_MAX_QP_RD_ATOMIC | 1701 IB_QP_MAX_DEST_RD_ATOMIC | 1702 IB_QP_ALT_PATH | 1703 IB_QP_ACCESS_FLAGS | 1704 IB_QP_PKEY_INDEX | 1705 IB_QP_MIN_RNR_TIMER | 1706 IB_QP_PATH_MIG_STATE), 1707 [IB_QPT_XRC_INI] = (IB_QP_PORT | 1708 IB_QP_AV | 1709 IB_QP_TIMEOUT | 1710 IB_QP_RETRY_CNT | 1711 IB_QP_RNR_RETRY | 1712 IB_QP_MAX_QP_RD_ATOMIC | 1713 IB_QP_ALT_PATH | 1714 IB_QP_ACCESS_FLAGS | 1715 IB_QP_PKEY_INDEX | 1716 IB_QP_PATH_MIG_STATE), 1717 [IB_QPT_XRC_TGT] = (IB_QP_PORT | 1718 IB_QP_AV | 1719 IB_QP_TIMEOUT | 1720 IB_QP_MAX_DEST_RD_ATOMIC | 1721 IB_QP_ALT_PATH | 1722 IB_QP_ACCESS_FLAGS | 1723 IB_QP_PKEY_INDEX | 1724 IB_QP_MIN_RNR_TIMER | 1725 IB_QP_PATH_MIG_STATE), 1726 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1727 IB_QP_QKEY), 1728 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1729 IB_QP_QKEY), 1730 } 1731 } 1732 }, 1733 [IB_QPS_SQE] = { 1734 [IB_QPS_RESET] = { .valid = 1 }, 1735 [IB_QPS_ERR] = { .valid = 1 }, 1736 [IB_QPS_RTS] = { 1737 .valid = 1, 1738 .opt_param = { 1739 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1740 IB_QP_QKEY), 1741 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1742 IB_QP_ACCESS_FLAGS), 1743 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1744 IB_QP_QKEY), 1745 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1746 IB_QP_QKEY), 1747 } 1748 } 1749 }, 1750 [IB_QPS_ERR] = { 1751 [IB_QPS_RESET] = { .valid = 1 }, 1752 [IB_QPS_ERR] = { .valid = 1 } 1753 } 1754 }; 1755 1756 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state, 1757 enum ib_qp_type type, enum ib_qp_attr_mask mask) 1758 { 1759 enum ib_qp_attr_mask req_param, opt_param; 1760 1761 if (mask & IB_QP_CUR_STATE && 1762 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS && 1763 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE) 1764 return false; 1765 1766 if (!qp_state_table[cur_state][next_state].valid) 1767 return false; 1768 1769 req_param = qp_state_table[cur_state][next_state].req_param[type]; 1770 opt_param = qp_state_table[cur_state][next_state].opt_param[type]; 1771 1772 if ((mask & req_param) != req_param) 1773 return false; 1774 1775 if (mask & ~(req_param | opt_param | IB_QP_STATE | IB_QP_RATE_LIMIT)) 1776 return false; 1777 1778 return true; 1779 } 1780 EXPORT_SYMBOL(ib_modify_qp_is_ok); 1781 1782 /** 1783 * ib_resolve_eth_dmac - Resolve destination mac address 1784 * @device: Device to consider 1785 * @ah_attr: address handle attribute which describes the 1786 * source and destination parameters 1787 * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It 1788 * returns 0 on success or appropriate error code. It initializes the 1789 * necessary ah_attr fields when call is successful. 1790 */ 1791 static int ib_resolve_eth_dmac(struct ib_device *device, 1792 struct rdma_ah_attr *ah_attr) 1793 { 1794 int ret = 0; 1795 1796 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) { 1797 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) { 1798 __be32 addr = 0; 1799 1800 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4); 1801 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac); 1802 } else { 1803 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw, 1804 (char *)ah_attr->roce.dmac); 1805 } 1806 } else { 1807 ret = ib_resolve_unicast_gid_dmac(device, ah_attr); 1808 } 1809 return ret; 1810 } 1811 1812 static bool is_qp_type_connected(const struct ib_qp *qp) 1813 { 1814 return (qp->qp_type == IB_QPT_UC || 1815 qp->qp_type == IB_QPT_RC || 1816 qp->qp_type == IB_QPT_XRC_INI || 1817 qp->qp_type == IB_QPT_XRC_TGT); 1818 } 1819 1820 /* 1821 * IB core internal function to perform QP attributes modification. 1822 */ 1823 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr, 1824 int attr_mask, struct ib_udata *udata) 1825 { 1826 u32 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port; 1827 const struct ib_gid_attr *old_sgid_attr_av; 1828 const struct ib_gid_attr *old_sgid_attr_alt_av; 1829 int ret; 1830 1831 attr->xmit_slave = NULL; 1832 if (attr_mask & IB_QP_AV) { 1833 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr, 1834 &old_sgid_attr_av); 1835 if (ret) 1836 return ret; 1837 1838 if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE && 1839 is_qp_type_connected(qp)) { 1840 struct net_device *slave; 1841 1842 /* 1843 * If the user provided the qp_attr then we have to 1844 * resolve it. Kerne users have to provide already 1845 * resolved rdma_ah_attr's. 1846 */ 1847 if (udata) { 1848 ret = ib_resolve_eth_dmac(qp->device, 1849 &attr->ah_attr); 1850 if (ret) 1851 goto out_av; 1852 } 1853 slave = rdma_lag_get_ah_roce_slave(qp->device, 1854 &attr->ah_attr, 1855 GFP_KERNEL); 1856 if (IS_ERR(slave)) { 1857 ret = PTR_ERR(slave); 1858 goto out_av; 1859 } 1860 attr->xmit_slave = slave; 1861 } 1862 } 1863 if (attr_mask & IB_QP_ALT_PATH) { 1864 /* 1865 * FIXME: This does not track the migration state, so if the 1866 * user loads a new alternate path after the HW has migrated 1867 * from primary->alternate we will keep the wrong 1868 * references. This is OK for IB because the reference 1869 * counting does not serve any functional purpose. 1870 */ 1871 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr, 1872 &old_sgid_attr_alt_av); 1873 if (ret) 1874 goto out_av; 1875 1876 /* 1877 * Today the core code can only handle alternate paths and APM 1878 * for IB. Ban them in roce mode. 1879 */ 1880 if (!(rdma_protocol_ib(qp->device, 1881 attr->alt_ah_attr.port_num) && 1882 rdma_protocol_ib(qp->device, port))) { 1883 ret = -EINVAL; 1884 goto out; 1885 } 1886 } 1887 1888 if (rdma_ib_or_roce(qp->device, port)) { 1889 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) { 1890 dev_warn(&qp->device->dev, 1891 "%s rq_psn overflow, masking to 24 bits\n", 1892 __func__); 1893 attr->rq_psn &= 0xffffff; 1894 } 1895 1896 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) { 1897 dev_warn(&qp->device->dev, 1898 " %s sq_psn overflow, masking to 24 bits\n", 1899 __func__); 1900 attr->sq_psn &= 0xffffff; 1901 } 1902 } 1903 1904 /* 1905 * Bind this qp to a counter automatically based on the rdma counter 1906 * rules. This only set in RST2INIT with port specified 1907 */ 1908 if (!qp->counter && (attr_mask & IB_QP_PORT) && 1909 ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT)) 1910 rdma_counter_bind_qp_auto(qp, attr->port_num); 1911 1912 ret = ib_security_modify_qp(qp, attr, attr_mask, udata); 1913 if (ret) 1914 goto out; 1915 1916 if (attr_mask & IB_QP_PORT) 1917 qp->port = attr->port_num; 1918 if (attr_mask & IB_QP_AV) 1919 qp->av_sgid_attr = 1920 rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr); 1921 if (attr_mask & IB_QP_ALT_PATH) 1922 qp->alt_path_sgid_attr = rdma_update_sgid_attr( 1923 &attr->alt_ah_attr, qp->alt_path_sgid_attr); 1924 1925 out: 1926 if (attr_mask & IB_QP_ALT_PATH) 1927 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av); 1928 out_av: 1929 if (attr_mask & IB_QP_AV) { 1930 rdma_lag_put_ah_roce_slave(attr->xmit_slave); 1931 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av); 1932 } 1933 return ret; 1934 } 1935 1936 /** 1937 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP. 1938 * @ib_qp: The QP to modify. 1939 * @attr: On input, specifies the QP attributes to modify. On output, 1940 * the current values of selected QP attributes are returned. 1941 * @attr_mask: A bit-mask used to specify which attributes of the QP 1942 * are being modified. 1943 * @udata: pointer to user's input output buffer information 1944 * are being modified. 1945 * It returns 0 on success and returns appropriate error code on error. 1946 */ 1947 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr, 1948 int attr_mask, struct ib_udata *udata) 1949 { 1950 return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata); 1951 } 1952 EXPORT_SYMBOL(ib_modify_qp_with_udata); 1953 1954 static void ib_get_width_and_speed(u32 netdev_speed, u32 lanes, 1955 u16 *speed, u8 *width) 1956 { 1957 if (!lanes) { 1958 if (netdev_speed <= SPEED_1000) { 1959 *width = IB_WIDTH_1X; 1960 *speed = IB_SPEED_SDR; 1961 } else if (netdev_speed <= SPEED_10000) { 1962 *width = IB_WIDTH_1X; 1963 *speed = IB_SPEED_FDR10; 1964 } else if (netdev_speed <= SPEED_20000) { 1965 *width = IB_WIDTH_4X; 1966 *speed = IB_SPEED_DDR; 1967 } else if (netdev_speed <= SPEED_25000) { 1968 *width = IB_WIDTH_1X; 1969 *speed = IB_SPEED_EDR; 1970 } else if (netdev_speed <= SPEED_40000) { 1971 *width = IB_WIDTH_4X; 1972 *speed = IB_SPEED_FDR10; 1973 } else if (netdev_speed <= SPEED_50000) { 1974 *width = IB_WIDTH_2X; 1975 *speed = IB_SPEED_EDR; 1976 } else if (netdev_speed <= SPEED_100000) { 1977 *width = IB_WIDTH_4X; 1978 *speed = IB_SPEED_EDR; 1979 } else if (netdev_speed <= SPEED_200000) { 1980 *width = IB_WIDTH_4X; 1981 *speed = IB_SPEED_HDR; 1982 } else { 1983 *width = IB_WIDTH_4X; 1984 *speed = IB_SPEED_NDR; 1985 } 1986 1987 return; 1988 } 1989 1990 switch (lanes) { 1991 case 1: 1992 *width = IB_WIDTH_1X; 1993 break; 1994 case 2: 1995 *width = IB_WIDTH_2X; 1996 break; 1997 case 4: 1998 *width = IB_WIDTH_4X; 1999 break; 2000 case 8: 2001 *width = IB_WIDTH_8X; 2002 break; 2003 case 12: 2004 *width = IB_WIDTH_12X; 2005 break; 2006 default: 2007 *width = IB_WIDTH_1X; 2008 } 2009 2010 switch (netdev_speed / lanes) { 2011 case SPEED_2500: 2012 *speed = IB_SPEED_SDR; 2013 break; 2014 case SPEED_5000: 2015 *speed = IB_SPEED_DDR; 2016 break; 2017 case SPEED_10000: 2018 *speed = IB_SPEED_FDR10; 2019 break; 2020 case SPEED_14000: 2021 *speed = IB_SPEED_FDR; 2022 break; 2023 case SPEED_25000: 2024 *speed = IB_SPEED_EDR; 2025 break; 2026 case SPEED_50000: 2027 *speed = IB_SPEED_HDR; 2028 break; 2029 case SPEED_100000: 2030 *speed = IB_SPEED_NDR; 2031 break; 2032 default: 2033 *speed = IB_SPEED_SDR; 2034 } 2035 } 2036 2037 int ib_get_eth_speed(struct ib_device *dev, u32 port_num, u16 *speed, u8 *width) 2038 { 2039 int rc; 2040 u32 netdev_speed; 2041 struct net_device *netdev; 2042 struct ethtool_link_ksettings lksettings = {}; 2043 2044 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET) 2045 return -EINVAL; 2046 2047 netdev = ib_device_get_netdev(dev, port_num); 2048 if (!netdev) 2049 return -ENODEV; 2050 2051 rtnl_lock(); 2052 rc = __ethtool_get_link_ksettings(netdev, &lksettings); 2053 rtnl_unlock(); 2054 2055 dev_put(netdev); 2056 2057 if (!rc && lksettings.base.speed != (u32)SPEED_UNKNOWN) { 2058 netdev_speed = lksettings.base.speed; 2059 } else { 2060 netdev_speed = SPEED_1000; 2061 if (rc) 2062 pr_warn("%s speed is unknown, defaulting to %u\n", 2063 netdev->name, netdev_speed); 2064 } 2065 2066 ib_get_width_and_speed(netdev_speed, lksettings.lanes, 2067 speed, width); 2068 2069 return 0; 2070 } 2071 EXPORT_SYMBOL(ib_get_eth_speed); 2072 2073 int ib_modify_qp(struct ib_qp *qp, 2074 struct ib_qp_attr *qp_attr, 2075 int qp_attr_mask) 2076 { 2077 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL); 2078 } 2079 EXPORT_SYMBOL(ib_modify_qp); 2080 2081 int ib_query_qp(struct ib_qp *qp, 2082 struct ib_qp_attr *qp_attr, 2083 int qp_attr_mask, 2084 struct ib_qp_init_attr *qp_init_attr) 2085 { 2086 qp_attr->ah_attr.grh.sgid_attr = NULL; 2087 qp_attr->alt_ah_attr.grh.sgid_attr = NULL; 2088 2089 return qp->device->ops.query_qp ? 2090 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask, 2091 qp_init_attr) : -EOPNOTSUPP; 2092 } 2093 EXPORT_SYMBOL(ib_query_qp); 2094 2095 int ib_close_qp(struct ib_qp *qp) 2096 { 2097 struct ib_qp *real_qp; 2098 unsigned long flags; 2099 2100 real_qp = qp->real_qp; 2101 if (real_qp == qp) 2102 return -EINVAL; 2103 2104 spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags); 2105 list_del(&qp->open_list); 2106 spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags); 2107 2108 atomic_dec(&real_qp->usecnt); 2109 if (qp->qp_sec) 2110 ib_close_shared_qp_security(qp->qp_sec); 2111 kfree(qp); 2112 2113 return 0; 2114 } 2115 EXPORT_SYMBOL(ib_close_qp); 2116 2117 static int __ib_destroy_shared_qp(struct ib_qp *qp) 2118 { 2119 struct ib_xrcd *xrcd; 2120 struct ib_qp *real_qp; 2121 int ret; 2122 2123 real_qp = qp->real_qp; 2124 xrcd = real_qp->xrcd; 2125 down_write(&xrcd->tgt_qps_rwsem); 2126 ib_close_qp(qp); 2127 if (atomic_read(&real_qp->usecnt) == 0) 2128 xa_erase(&xrcd->tgt_qps, real_qp->qp_num); 2129 else 2130 real_qp = NULL; 2131 up_write(&xrcd->tgt_qps_rwsem); 2132 2133 if (real_qp) { 2134 ret = ib_destroy_qp(real_qp); 2135 if (!ret) 2136 atomic_dec(&xrcd->usecnt); 2137 } 2138 2139 return 0; 2140 } 2141 2142 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata) 2143 { 2144 const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr; 2145 const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr; 2146 struct ib_qp_security *sec; 2147 int ret; 2148 2149 WARN_ON_ONCE(qp->mrs_used > 0); 2150 2151 if (atomic_read(&qp->usecnt)) 2152 return -EBUSY; 2153 2154 if (qp->real_qp != qp) 2155 return __ib_destroy_shared_qp(qp); 2156 2157 sec = qp->qp_sec; 2158 if (sec) 2159 ib_destroy_qp_security_begin(sec); 2160 2161 if (!qp->uobject) 2162 rdma_rw_cleanup_mrs(qp); 2163 2164 rdma_counter_unbind_qp(qp, qp->port, true); 2165 ret = qp->device->ops.destroy_qp(qp, udata); 2166 if (ret) { 2167 if (sec) 2168 ib_destroy_qp_security_abort(sec); 2169 return ret; 2170 } 2171 2172 if (alt_path_sgid_attr) 2173 rdma_put_gid_attr(alt_path_sgid_attr); 2174 if (av_sgid_attr) 2175 rdma_put_gid_attr(av_sgid_attr); 2176 2177 ib_qp_usecnt_dec(qp); 2178 if (sec) 2179 ib_destroy_qp_security_end(sec); 2180 2181 rdma_restrack_del(&qp->res); 2182 kfree(qp); 2183 return ret; 2184 } 2185 EXPORT_SYMBOL(ib_destroy_qp_user); 2186 2187 /* Completion queues */ 2188 2189 struct ib_cq *__ib_create_cq(struct ib_device *device, 2190 ib_comp_handler comp_handler, 2191 void (*event_handler)(struct ib_event *, void *), 2192 void *cq_context, 2193 const struct ib_cq_init_attr *cq_attr, 2194 const char *caller) 2195 { 2196 struct ib_cq *cq; 2197 int ret; 2198 2199 cq = rdma_zalloc_drv_obj(device, ib_cq); 2200 if (!cq) 2201 return ERR_PTR(-ENOMEM); 2202 2203 if (WARN_ON_ONCE(!cq_attr->cqe)) 2204 return ERR_PTR(-EINVAL); 2205 2206 cq->device = device; 2207 cq->comp_handler = comp_handler; 2208 cq->event_handler = event_handler; 2209 cq->cq_context = cq_context; 2210 atomic_set(&cq->usecnt, 0); 2211 2212 rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ); 2213 rdma_restrack_set_name(&cq->res, caller); 2214 2215 ret = device->ops.create_cq(cq, cq_attr, NULL); 2216 if (ret) { 2217 rdma_restrack_put(&cq->res); 2218 kfree(cq); 2219 return ERR_PTR(ret); 2220 } 2221 rdma_restrack_add(&cq->res); 2222 return cq; 2223 } 2224 EXPORT_SYMBOL(__ib_create_cq); 2225 2226 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period) 2227 { 2228 if (cq->shared) 2229 return -EOPNOTSUPP; 2230 2231 return cq->device->ops.modify_cq ? 2232 cq->device->ops.modify_cq(cq, cq_count, 2233 cq_period) : -EOPNOTSUPP; 2234 } 2235 EXPORT_SYMBOL(rdma_set_cq_moderation); 2236 2237 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata) 2238 { 2239 int ret; 2240 2241 if (WARN_ON_ONCE(cq->shared)) 2242 return -EOPNOTSUPP; 2243 2244 if (atomic_read(&cq->usecnt)) 2245 return -EBUSY; 2246 2247 ret = cq->device->ops.destroy_cq(cq, udata); 2248 if (ret) 2249 return ret; 2250 2251 rdma_restrack_del(&cq->res); 2252 kfree(cq); 2253 return ret; 2254 } 2255 EXPORT_SYMBOL(ib_destroy_cq_user); 2256 2257 /* Memory regions */ 2258 2259 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, 2260 u64 virt_addr, int access_flags) 2261 { 2262 struct ib_mr *mr; 2263 2264 if (access_flags & IB_ACCESS_ON_DEMAND) { 2265 if (!(pd->device->attrs.kernel_cap_flags & 2266 IBK_ON_DEMAND_PAGING)) { 2267 pr_debug("ODP support not available\n"); 2268 return ERR_PTR(-EINVAL); 2269 } 2270 } 2271 2272 mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr, 2273 access_flags, NULL, NULL); 2274 2275 if (IS_ERR(mr)) 2276 return mr; 2277 2278 mr->device = pd->device; 2279 mr->type = IB_MR_TYPE_USER; 2280 mr->pd = pd; 2281 mr->dm = NULL; 2282 atomic_inc(&pd->usecnt); 2283 mr->iova = virt_addr; 2284 mr->length = length; 2285 2286 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR); 2287 rdma_restrack_parent_name(&mr->res, &pd->res); 2288 rdma_restrack_add(&mr->res); 2289 2290 return mr; 2291 } 2292 EXPORT_SYMBOL(ib_reg_user_mr); 2293 2294 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice, 2295 u32 flags, struct ib_sge *sg_list, u32 num_sge) 2296 { 2297 if (!pd->device->ops.advise_mr) 2298 return -EOPNOTSUPP; 2299 2300 if (!num_sge) 2301 return 0; 2302 2303 return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge, 2304 NULL); 2305 } 2306 EXPORT_SYMBOL(ib_advise_mr); 2307 2308 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata) 2309 { 2310 struct ib_pd *pd = mr->pd; 2311 struct ib_dm *dm = mr->dm; 2312 struct ib_dmah *dmah = mr->dmah; 2313 struct ib_sig_attrs *sig_attrs = mr->sig_attrs; 2314 int ret; 2315 2316 trace_mr_dereg(mr); 2317 rdma_restrack_del(&mr->res); 2318 ret = mr->device->ops.dereg_mr(mr, udata); 2319 if (!ret) { 2320 atomic_dec(&pd->usecnt); 2321 if (dm) 2322 atomic_dec(&dm->usecnt); 2323 if (dmah) 2324 atomic_dec(&dmah->usecnt); 2325 kfree(sig_attrs); 2326 } 2327 2328 return ret; 2329 } 2330 EXPORT_SYMBOL(ib_dereg_mr_user); 2331 2332 /** 2333 * ib_alloc_mr() - Allocates a memory region 2334 * @pd: protection domain associated with the region 2335 * @mr_type: memory region type 2336 * @max_num_sg: maximum sg entries available for registration. 2337 * 2338 * Notes: 2339 * Memory registeration page/sg lists must not exceed max_num_sg. 2340 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed 2341 * max_num_sg * used_page_size. 2342 * 2343 */ 2344 struct ib_mr *ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type, 2345 u32 max_num_sg) 2346 { 2347 struct ib_mr *mr; 2348 2349 if (!pd->device->ops.alloc_mr) { 2350 mr = ERR_PTR(-EOPNOTSUPP); 2351 goto out; 2352 } 2353 2354 if (mr_type == IB_MR_TYPE_INTEGRITY) { 2355 WARN_ON_ONCE(1); 2356 mr = ERR_PTR(-EINVAL); 2357 goto out; 2358 } 2359 2360 mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg); 2361 if (IS_ERR(mr)) 2362 goto out; 2363 2364 mr->device = pd->device; 2365 mr->pd = pd; 2366 mr->dm = NULL; 2367 mr->uobject = NULL; 2368 atomic_inc(&pd->usecnt); 2369 mr->need_inval = false; 2370 mr->type = mr_type; 2371 mr->sig_attrs = NULL; 2372 2373 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR); 2374 rdma_restrack_parent_name(&mr->res, &pd->res); 2375 rdma_restrack_add(&mr->res); 2376 out: 2377 trace_mr_alloc(pd, mr_type, max_num_sg, mr); 2378 return mr; 2379 } 2380 EXPORT_SYMBOL(ib_alloc_mr); 2381 2382 /** 2383 * ib_alloc_mr_integrity() - Allocates an integrity memory region 2384 * @pd: protection domain associated with the region 2385 * @max_num_data_sg: maximum data sg entries available for registration 2386 * @max_num_meta_sg: maximum metadata sg entries available for 2387 * registration 2388 * 2389 * Notes: 2390 * Memory registration page/sg lists must not exceed max_num_sg, 2391 * also the integrity page/sg lists must not exceed max_num_meta_sg. 2392 * 2393 */ 2394 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd, 2395 u32 max_num_data_sg, 2396 u32 max_num_meta_sg) 2397 { 2398 struct ib_mr *mr; 2399 struct ib_sig_attrs *sig_attrs; 2400 2401 if (!pd->device->ops.alloc_mr_integrity || 2402 !pd->device->ops.map_mr_sg_pi) { 2403 mr = ERR_PTR(-EOPNOTSUPP); 2404 goto out; 2405 } 2406 2407 if (!max_num_meta_sg) { 2408 mr = ERR_PTR(-EINVAL); 2409 goto out; 2410 } 2411 2412 sig_attrs = kzalloc_obj(struct ib_sig_attrs); 2413 if (!sig_attrs) { 2414 mr = ERR_PTR(-ENOMEM); 2415 goto out; 2416 } 2417 2418 mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg, 2419 max_num_meta_sg); 2420 if (IS_ERR(mr)) { 2421 kfree(sig_attrs); 2422 goto out; 2423 } 2424 2425 mr->device = pd->device; 2426 mr->pd = pd; 2427 mr->dm = NULL; 2428 mr->uobject = NULL; 2429 atomic_inc(&pd->usecnt); 2430 mr->need_inval = false; 2431 mr->type = IB_MR_TYPE_INTEGRITY; 2432 mr->sig_attrs = sig_attrs; 2433 2434 rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR); 2435 rdma_restrack_parent_name(&mr->res, &pd->res); 2436 rdma_restrack_add(&mr->res); 2437 out: 2438 trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr); 2439 return mr; 2440 } 2441 EXPORT_SYMBOL(ib_alloc_mr_integrity); 2442 2443 /* Multicast groups */ 2444 2445 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid) 2446 { 2447 struct ib_qp_init_attr init_attr = {}; 2448 struct ib_qp_attr attr = {}; 2449 int num_eth_ports = 0; 2450 unsigned int port; 2451 2452 /* If QP state >= init, it is assigned to a port and we can check this 2453 * port only. 2454 */ 2455 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) { 2456 if (attr.qp_state >= IB_QPS_INIT) { 2457 if (rdma_port_get_link_layer(qp->device, attr.port_num) != 2458 IB_LINK_LAYER_INFINIBAND) 2459 return true; 2460 goto lid_check; 2461 } 2462 } 2463 2464 /* Can't get a quick answer, iterate over all ports */ 2465 rdma_for_each_port(qp->device, port) 2466 if (rdma_port_get_link_layer(qp->device, port) != 2467 IB_LINK_LAYER_INFINIBAND) 2468 num_eth_ports++; 2469 2470 /* If we have at lease one Ethernet port, RoCE annex declares that 2471 * multicast LID should be ignored. We can't tell at this step if the 2472 * QP belongs to an IB or Ethernet port. 2473 */ 2474 if (num_eth_ports) 2475 return true; 2476 2477 /* If all the ports are IB, we can check according to IB spec. */ 2478 lid_check: 2479 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) || 2480 lid == be16_to_cpu(IB_LID_PERMISSIVE)); 2481 } 2482 2483 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid) 2484 { 2485 int ret; 2486 2487 if (!qp->device->ops.attach_mcast) 2488 return -EOPNOTSUPP; 2489 2490 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) || 2491 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid)) 2492 return -EINVAL; 2493 2494 ret = qp->device->ops.attach_mcast(qp, gid, lid); 2495 if (!ret) 2496 atomic_inc(&qp->usecnt); 2497 return ret; 2498 } 2499 EXPORT_SYMBOL(ib_attach_mcast); 2500 2501 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid) 2502 { 2503 int ret; 2504 2505 if (!qp->device->ops.detach_mcast) 2506 return -EOPNOTSUPP; 2507 2508 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) || 2509 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid)) 2510 return -EINVAL; 2511 2512 ret = qp->device->ops.detach_mcast(qp, gid, lid); 2513 if (!ret) 2514 atomic_dec(&qp->usecnt); 2515 return ret; 2516 } 2517 EXPORT_SYMBOL(ib_detach_mcast); 2518 2519 /** 2520 * ib_alloc_xrcd_user - Allocates an XRC domain. 2521 * @device: The device on which to allocate the XRC domain. 2522 * @inode: inode to connect XRCD 2523 * @udata: Valid user data or NULL for kernel object 2524 */ 2525 struct ib_xrcd *ib_alloc_xrcd_user(struct ib_device *device, 2526 struct inode *inode, struct ib_udata *udata) 2527 { 2528 struct ib_xrcd *xrcd; 2529 int ret; 2530 2531 if (!device->ops.alloc_xrcd) 2532 return ERR_PTR(-EOPNOTSUPP); 2533 2534 xrcd = rdma_zalloc_drv_obj(device, ib_xrcd); 2535 if (!xrcd) 2536 return ERR_PTR(-ENOMEM); 2537 2538 xrcd->device = device; 2539 xrcd->inode = inode; 2540 atomic_set(&xrcd->usecnt, 0); 2541 init_rwsem(&xrcd->tgt_qps_rwsem); 2542 xa_init(&xrcd->tgt_qps); 2543 2544 ret = device->ops.alloc_xrcd(xrcd, udata); 2545 if (ret) 2546 goto err; 2547 return xrcd; 2548 err: 2549 kfree(xrcd); 2550 return ERR_PTR(ret); 2551 } 2552 EXPORT_SYMBOL(ib_alloc_xrcd_user); 2553 2554 /** 2555 * ib_dealloc_xrcd_user - Deallocates an XRC domain. 2556 * @xrcd: The XRC domain to deallocate. 2557 * @udata: Valid user data or NULL for kernel object 2558 */ 2559 int ib_dealloc_xrcd_user(struct ib_xrcd *xrcd, struct ib_udata *udata) 2560 { 2561 int ret; 2562 2563 if (atomic_read(&xrcd->usecnt)) 2564 return -EBUSY; 2565 2566 WARN_ON(!xa_empty(&xrcd->tgt_qps)); 2567 ret = xrcd->device->ops.dealloc_xrcd(xrcd, udata); 2568 if (ret) 2569 return ret; 2570 kfree(xrcd); 2571 return ret; 2572 } 2573 EXPORT_SYMBOL(ib_dealloc_xrcd_user); 2574 2575 /** 2576 * ib_create_wq - Creates a WQ associated with the specified protection 2577 * domain. 2578 * @pd: The protection domain associated with the WQ. 2579 * @wq_attr: A list of initial attributes required to create the 2580 * WQ. If WQ creation succeeds, then the attributes are updated to 2581 * the actual capabilities of the created WQ. 2582 * 2583 * wq_attr->max_wr and wq_attr->max_sge determine 2584 * the requested size of the WQ, and set to the actual values allocated 2585 * on return. 2586 * If ib_create_wq() succeeds, then max_wr and max_sge will always be 2587 * at least as large as the requested values. 2588 */ 2589 struct ib_wq *ib_create_wq(struct ib_pd *pd, 2590 struct ib_wq_init_attr *wq_attr) 2591 { 2592 struct ib_wq *wq; 2593 2594 if (!pd->device->ops.create_wq) 2595 return ERR_PTR(-EOPNOTSUPP); 2596 2597 wq = pd->device->ops.create_wq(pd, wq_attr, NULL); 2598 if (!IS_ERR(wq)) { 2599 wq->event_handler = wq_attr->event_handler; 2600 wq->wq_context = wq_attr->wq_context; 2601 wq->wq_type = wq_attr->wq_type; 2602 wq->cq = wq_attr->cq; 2603 wq->device = pd->device; 2604 wq->pd = pd; 2605 wq->uobject = NULL; 2606 atomic_inc(&pd->usecnt); 2607 atomic_inc(&wq_attr->cq->usecnt); 2608 atomic_set(&wq->usecnt, 0); 2609 } 2610 return wq; 2611 } 2612 EXPORT_SYMBOL(ib_create_wq); 2613 2614 /** 2615 * ib_destroy_wq_user - Destroys the specified user WQ. 2616 * @wq: The WQ to destroy. 2617 * @udata: Valid user data 2618 */ 2619 int ib_destroy_wq_user(struct ib_wq *wq, struct ib_udata *udata) 2620 { 2621 struct ib_cq *cq = wq->cq; 2622 struct ib_pd *pd = wq->pd; 2623 int ret; 2624 2625 if (atomic_read(&wq->usecnt)) 2626 return -EBUSY; 2627 2628 ret = wq->device->ops.destroy_wq(wq, udata); 2629 if (ret) 2630 return ret; 2631 2632 atomic_dec(&pd->usecnt); 2633 atomic_dec(&cq->usecnt); 2634 return ret; 2635 } 2636 EXPORT_SYMBOL(ib_destroy_wq_user); 2637 2638 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask, 2639 struct ib_mr_status *mr_status) 2640 { 2641 if (!mr->device->ops.check_mr_status) 2642 return -EOPNOTSUPP; 2643 2644 return mr->device->ops.check_mr_status(mr, check_mask, mr_status); 2645 } 2646 EXPORT_SYMBOL(ib_check_mr_status); 2647 2648 int ib_set_vf_link_state(struct ib_device *device, int vf, u32 port, 2649 int state) 2650 { 2651 if (!device->ops.set_vf_link_state) 2652 return -EOPNOTSUPP; 2653 2654 return device->ops.set_vf_link_state(device, vf, port, state); 2655 } 2656 EXPORT_SYMBOL(ib_set_vf_link_state); 2657 2658 int ib_get_vf_config(struct ib_device *device, int vf, u32 port, 2659 struct ifla_vf_info *info) 2660 { 2661 if (!device->ops.get_vf_config) 2662 return -EOPNOTSUPP; 2663 2664 return device->ops.get_vf_config(device, vf, port, info); 2665 } 2666 EXPORT_SYMBOL(ib_get_vf_config); 2667 2668 int ib_get_vf_stats(struct ib_device *device, int vf, u32 port, 2669 struct ifla_vf_stats *stats) 2670 { 2671 if (!device->ops.get_vf_stats) 2672 return -EOPNOTSUPP; 2673 2674 return device->ops.get_vf_stats(device, vf, port, stats); 2675 } 2676 EXPORT_SYMBOL(ib_get_vf_stats); 2677 2678 int ib_set_vf_guid(struct ib_device *device, int vf, u32 port, u64 guid, 2679 int type) 2680 { 2681 if (!device->ops.set_vf_guid) 2682 return -EOPNOTSUPP; 2683 2684 return device->ops.set_vf_guid(device, vf, port, guid, type); 2685 } 2686 EXPORT_SYMBOL(ib_set_vf_guid); 2687 2688 int ib_get_vf_guid(struct ib_device *device, int vf, u32 port, 2689 struct ifla_vf_guid *node_guid, 2690 struct ifla_vf_guid *port_guid) 2691 { 2692 if (!device->ops.get_vf_guid) 2693 return -EOPNOTSUPP; 2694 2695 return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid); 2696 } 2697 EXPORT_SYMBOL(ib_get_vf_guid); 2698 /** 2699 * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection 2700 * information) and set an appropriate memory region for registration. 2701 * @mr: memory region 2702 * @data_sg: dma mapped scatterlist for data 2703 * @data_sg_nents: number of entries in data_sg 2704 * @data_sg_offset: offset in bytes into data_sg 2705 * @meta_sg: dma mapped scatterlist for metadata 2706 * @meta_sg_nents: number of entries in meta_sg 2707 * @meta_sg_offset: offset in bytes into meta_sg 2708 * @page_size: page vector desired page size 2709 * 2710 * Constraints: 2711 * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY. 2712 * 2713 * Return: 0 on success. 2714 * 2715 * After this completes successfully, the memory region 2716 * is ready for registration. 2717 */ 2718 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg, 2719 int data_sg_nents, unsigned int *data_sg_offset, 2720 struct scatterlist *meta_sg, int meta_sg_nents, 2721 unsigned int *meta_sg_offset, unsigned int page_size) 2722 { 2723 if (unlikely(!mr->device->ops.map_mr_sg_pi || 2724 WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY))) 2725 return -EOPNOTSUPP; 2726 2727 mr->page_size = page_size; 2728 2729 return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents, 2730 data_sg_offset, meta_sg, 2731 meta_sg_nents, meta_sg_offset); 2732 } 2733 EXPORT_SYMBOL(ib_map_mr_sg_pi); 2734 2735 /** 2736 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list 2737 * and set it the memory region. 2738 * @mr: memory region 2739 * @sg: dma mapped scatterlist 2740 * @sg_nents: number of entries in sg 2741 * @sg_offset: offset in bytes into sg 2742 * @page_size: page vector desired page size 2743 * 2744 * Constraints: 2745 * 2746 * - The first sg element is allowed to have an offset. 2747 * - Each sg element must either be aligned to page_size or virtually 2748 * contiguous to the previous element. In case an sg element has a 2749 * non-contiguous offset, the mapping prefix will not include it. 2750 * - The last sg element is allowed to have length less than page_size. 2751 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size 2752 * then only max_num_sg entries will be mapped. 2753 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these 2754 * constraints holds and the page_size argument is ignored. 2755 * 2756 * Returns the number of sg elements that were mapped to the memory region. 2757 * 2758 * After this completes successfully, the memory region 2759 * is ready for registration. 2760 */ 2761 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents, 2762 unsigned int *sg_offset, unsigned int page_size) 2763 { 2764 if (unlikely(!mr->device->ops.map_mr_sg)) 2765 return -EOPNOTSUPP; 2766 2767 mr->page_size = page_size; 2768 2769 return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset); 2770 } 2771 EXPORT_SYMBOL(ib_map_mr_sg); 2772 2773 /** 2774 * ib_sg_to_pages() - Convert the largest prefix of a sg list 2775 * to a page vector 2776 * @mr: memory region 2777 * @sgl: dma mapped scatterlist 2778 * @sg_nents: number of entries in sg 2779 * @sg_offset_p: ==== ======================================================= 2780 * IN start offset in bytes into sg 2781 * OUT offset in bytes for element n of the sg of the first 2782 * byte that has not been processed where n is the return 2783 * value of this function. 2784 * ==== ======================================================= 2785 * @set_page: driver page assignment function pointer 2786 * 2787 * Core service helper for drivers to convert the largest 2788 * prefix of given sg list to a page vector. The sg list 2789 * prefix converted is the prefix that meet the requirements 2790 * of ib_map_mr_sg. 2791 * 2792 * Returns the number of sg elements that were assigned to 2793 * a page vector. 2794 */ 2795 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents, 2796 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64)) 2797 { 2798 struct scatterlist *sg; 2799 u64 last_end_dma_addr = 0; 2800 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0; 2801 unsigned int last_page_off = 0; 2802 u64 page_mask = ~((u64)mr->page_size - 1); 2803 int i, ret; 2804 2805 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0]))) 2806 return -EINVAL; 2807 2808 mr->iova = sg_dma_address(&sgl[0]) + sg_offset; 2809 mr->length = 0; 2810 2811 for_each_sg(sgl, sg, sg_nents, i) { 2812 u64 dma_addr = sg_dma_address(sg) + sg_offset; 2813 u64 prev_addr = dma_addr; 2814 unsigned int dma_len = sg_dma_len(sg) - sg_offset; 2815 u64 end_dma_addr = dma_addr + dma_len; 2816 u64 page_addr = dma_addr & page_mask; 2817 2818 /* 2819 * For the second and later elements, check whether either the 2820 * end of element i-1 or the start of element i is not aligned 2821 * on a page boundary. 2822 */ 2823 if (i && (last_page_off != 0 || page_addr != dma_addr)) { 2824 /* Stop mapping if there is a gap. */ 2825 if (last_end_dma_addr != dma_addr) 2826 break; 2827 2828 /* 2829 * Coalesce this element with the last. If it is small 2830 * enough just update mr->length. Otherwise start 2831 * mapping from the next page. 2832 */ 2833 goto next_page; 2834 } 2835 2836 do { 2837 ret = set_page(mr, page_addr); 2838 if (unlikely(ret < 0)) { 2839 sg_offset = prev_addr - sg_dma_address(sg); 2840 mr->length += prev_addr - dma_addr; 2841 if (sg_offset_p) 2842 *sg_offset_p = sg_offset; 2843 return i || sg_offset ? i : ret; 2844 } 2845 prev_addr = page_addr; 2846 next_page: 2847 page_addr += mr->page_size; 2848 } while (page_addr < end_dma_addr); 2849 2850 mr->length += dma_len; 2851 last_end_dma_addr = end_dma_addr; 2852 last_page_off = end_dma_addr & ~page_mask; 2853 2854 sg_offset = 0; 2855 } 2856 2857 if (sg_offset_p) 2858 *sg_offset_p = 0; 2859 return i; 2860 } 2861 EXPORT_SYMBOL(ib_sg_to_pages); 2862 2863 struct ib_drain_cqe { 2864 struct ib_cqe cqe; 2865 struct completion done; 2866 }; 2867 2868 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc) 2869 { 2870 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe, 2871 cqe); 2872 2873 complete(&cqe->done); 2874 } 2875 2876 /* 2877 * Post a WR and block until its completion is reaped for the SQ. 2878 */ 2879 static void __ib_drain_sq(struct ib_qp *qp) 2880 { 2881 struct ib_cq *cq = qp->send_cq; 2882 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 2883 struct ib_drain_cqe sdrain; 2884 struct ib_rdma_wr swr = { 2885 .wr = { 2886 .next = NULL, 2887 { .wr_cqe = &sdrain.cqe, }, 2888 .opcode = IB_WR_RDMA_WRITE, 2889 }, 2890 }; 2891 int ret; 2892 2893 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 2894 if (ret) { 2895 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 2896 return; 2897 } 2898 2899 sdrain.cqe.done = ib_drain_qp_done; 2900 init_completion(&sdrain.done); 2901 2902 ret = ib_post_send(qp, &swr.wr, NULL); 2903 if (ret) { 2904 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 2905 return; 2906 } 2907 2908 if (cq->poll_ctx == IB_POLL_DIRECT) 2909 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0) 2910 ib_process_cq_direct(cq, -1); 2911 else 2912 wait_for_completion(&sdrain.done); 2913 } 2914 2915 /* 2916 * Post a WR and block until its completion is reaped for the RQ. 2917 */ 2918 static void __ib_drain_rq(struct ib_qp *qp) 2919 { 2920 struct ib_cq *cq = qp->recv_cq; 2921 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 2922 struct ib_drain_cqe rdrain; 2923 struct ib_recv_wr rwr = {}; 2924 int ret; 2925 2926 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 2927 if (ret) { 2928 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 2929 return; 2930 } 2931 2932 rwr.wr_cqe = &rdrain.cqe; 2933 rdrain.cqe.done = ib_drain_qp_done; 2934 init_completion(&rdrain.done); 2935 2936 ret = ib_post_recv(qp, &rwr, NULL); 2937 if (ret) { 2938 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 2939 return; 2940 } 2941 2942 if (cq->poll_ctx == IB_POLL_DIRECT) 2943 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0) 2944 ib_process_cq_direct(cq, -1); 2945 else 2946 wait_for_completion(&rdrain.done); 2947 } 2948 2949 /* 2950 * __ib_drain_srq() - Block until Last WQE Reached event arrives, or timeout 2951 * expires. 2952 * @qp: queue pair associated with SRQ to drain 2953 * 2954 * Quoting 10.3.1 Queue Pair and EE Context States: 2955 * 2956 * Note, for QPs that are associated with an SRQ, the Consumer should take the 2957 * QP through the Error State before invoking a Destroy QP or a Modify QP to the 2958 * Reset State. The Consumer may invoke the Destroy QP without first performing 2959 * a Modify QP to the Error State and waiting for the Affiliated Asynchronous 2960 * Last WQE Reached Event. However, if the Consumer does not wait for the 2961 * Affiliated Asynchronous Last WQE Reached Event, then WQE and Data Segment 2962 * leakage may occur. Therefore, it is good programming practice to tear down a 2963 * QP that is associated with an SRQ by using the following process: 2964 * 2965 * - Put the QP in the Error State 2966 * - Wait for the Affiliated Asynchronous Last WQE Reached Event; 2967 * - either: 2968 * drain the CQ by invoking the Poll CQ verb and either wait for CQ 2969 * to be empty or the number of Poll CQ operations has exceeded 2970 * CQ capacity size; 2971 * - or 2972 * post another WR that completes on the same CQ and wait for this 2973 * WR to return as a WC; 2974 * - and then invoke a Destroy QP or Reset QP. 2975 * 2976 * We use the first option. 2977 */ 2978 static void __ib_drain_srq(struct ib_qp *qp) 2979 { 2980 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 2981 struct ib_cq *cq; 2982 int n, polled = 0; 2983 int ret; 2984 2985 if (!qp->srq) { 2986 WARN_ONCE(1, "QP 0x%p is not associated with SRQ\n", qp); 2987 return; 2988 } 2989 2990 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 2991 if (ret) { 2992 WARN_ONCE(ret, "failed to drain shared recv queue: %d\n", ret); 2993 return; 2994 } 2995 2996 if (ib_srq_has_cq(qp->srq->srq_type)) { 2997 cq = qp->srq->ext.cq; 2998 } else if (qp->recv_cq) { 2999 cq = qp->recv_cq; 3000 } else { 3001 WARN_ONCE(1, "QP 0x%p has no CQ associated with SRQ\n", qp); 3002 return; 3003 } 3004 3005 if (wait_for_completion_timeout(&qp->srq_completion, 60 * HZ) > 0) { 3006 while (polled != cq->cqe) { 3007 n = ib_process_cq_direct(cq, cq->cqe - polled); 3008 if (!n) 3009 return; 3010 polled += n; 3011 } 3012 } 3013 } 3014 3015 /** 3016 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the 3017 * application. 3018 * @qp: queue pair to drain 3019 * 3020 * If the device has a provider-specific drain function, then 3021 * call that. Otherwise call the generic drain function 3022 * __ib_drain_sq(). 3023 * 3024 * The caller must: 3025 * 3026 * ensure there is room in the CQ and SQ for the drain work request and 3027 * completion. 3028 * 3029 * allocate the CQ using ib_alloc_cq(). 3030 * 3031 * ensure that there are no other contexts that are posting WRs concurrently. 3032 * Otherwise the drain is not guaranteed. 3033 */ 3034 void ib_drain_sq(struct ib_qp *qp) 3035 { 3036 if (qp->device->ops.drain_sq) 3037 qp->device->ops.drain_sq(qp); 3038 else 3039 __ib_drain_sq(qp); 3040 trace_cq_drain_complete(qp->send_cq); 3041 } 3042 EXPORT_SYMBOL(ib_drain_sq); 3043 3044 /** 3045 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the 3046 * application. 3047 * @qp: queue pair to drain 3048 * 3049 * If the device has a provider-specific drain function, then 3050 * call that. Otherwise call the generic drain function 3051 * __ib_drain_rq(). 3052 * 3053 * The caller must: 3054 * 3055 * ensure there is room in the CQ and RQ for the drain work request and 3056 * completion. 3057 * 3058 * allocate the CQ using ib_alloc_cq(). 3059 * 3060 * ensure that there are no other contexts that are posting WRs concurrently. 3061 * Otherwise the drain is not guaranteed. 3062 */ 3063 void ib_drain_rq(struct ib_qp *qp) 3064 { 3065 if (qp->device->ops.drain_rq) 3066 qp->device->ops.drain_rq(qp); 3067 else 3068 __ib_drain_rq(qp); 3069 trace_cq_drain_complete(qp->recv_cq); 3070 } 3071 EXPORT_SYMBOL(ib_drain_rq); 3072 3073 /** 3074 * ib_drain_qp() - Block until all CQEs have been consumed by the 3075 * application on both the RQ and SQ. 3076 * @qp: queue pair to drain 3077 * 3078 * The caller must: 3079 * 3080 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests 3081 * and completions. 3082 * 3083 * allocate the CQs using ib_alloc_cq(). 3084 * 3085 * ensure that there are no other contexts that are posting WRs concurrently. 3086 * Otherwise the drain is not guaranteed. 3087 */ 3088 void ib_drain_qp(struct ib_qp *qp) 3089 { 3090 ib_drain_sq(qp); 3091 if (!qp->srq) 3092 ib_drain_rq(qp); 3093 else 3094 __ib_drain_srq(qp); 3095 } 3096 EXPORT_SYMBOL(ib_drain_qp); 3097 3098 struct net_device *rdma_alloc_netdev(struct ib_device *device, u32 port_num, 3099 enum rdma_netdev_t type, const char *name, 3100 unsigned char name_assign_type, 3101 void (*setup)(struct net_device *)) 3102 { 3103 struct rdma_netdev_alloc_params params; 3104 struct net_device *netdev; 3105 int rc; 3106 3107 if (!device->ops.rdma_netdev_get_params) 3108 return ERR_PTR(-EOPNOTSUPP); 3109 3110 rc = device->ops.rdma_netdev_get_params(device, port_num, type, 3111 ¶ms); 3112 if (rc) 3113 return ERR_PTR(rc); 3114 3115 netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type, 3116 setup, params.txqs, params.rxqs); 3117 if (!netdev) 3118 return ERR_PTR(-ENOMEM); 3119 3120 return netdev; 3121 } 3122 EXPORT_SYMBOL(rdma_alloc_netdev); 3123 3124 int rdma_init_netdev(struct ib_device *device, u32 port_num, 3125 enum rdma_netdev_t type, const char *name, 3126 unsigned char name_assign_type, 3127 void (*setup)(struct net_device *), 3128 struct net_device *netdev) 3129 { 3130 struct rdma_netdev_alloc_params params; 3131 int rc; 3132 3133 if (!device->ops.rdma_netdev_get_params) 3134 return -EOPNOTSUPP; 3135 3136 rc = device->ops.rdma_netdev_get_params(device, port_num, type, 3137 ¶ms); 3138 if (rc) 3139 return rc; 3140 3141 return params.initialize_rdma_netdev(device, port_num, 3142 netdev, params.param); 3143 } 3144 EXPORT_SYMBOL(rdma_init_netdev); 3145 3146 /** 3147 * rdma_alloc_hw_stats_struct - Helper function to allocate dynamic struct 3148 * for the drivers. 3149 * @descs: array of static descriptors 3150 * @num_counters: number of elements in array 3151 * @lifespan: milliseconds between updates 3152 */ 3153 struct rdma_hw_stats *rdma_alloc_hw_stats_struct( 3154 const struct rdma_stat_desc *descs, int num_counters, 3155 unsigned long lifespan) 3156 { 3157 struct rdma_hw_stats *stats; 3158 3159 stats = kzalloc_flex(*stats, value, num_counters); 3160 if (!stats) 3161 return NULL; 3162 3163 stats->is_disabled = kcalloc(BITS_TO_LONGS(num_counters), 3164 sizeof(*stats->is_disabled), GFP_KERNEL); 3165 if (!stats->is_disabled) 3166 goto err; 3167 3168 stats->descs = descs; 3169 stats->num_counters = num_counters; 3170 stats->lifespan = msecs_to_jiffies(lifespan); 3171 mutex_init(&stats->lock); 3172 3173 return stats; 3174 3175 err: 3176 kfree(stats); 3177 return NULL; 3178 } 3179 EXPORT_SYMBOL(rdma_alloc_hw_stats_struct); 3180 3181 /** 3182 * rdma_free_hw_stats_struct - Helper function to release rdma_hw_stats 3183 * @stats: statistics to release 3184 */ 3185 void rdma_free_hw_stats_struct(struct rdma_hw_stats *stats) 3186 { 3187 if (!stats) 3188 return; 3189 3190 kfree(stats->is_disabled); 3191 kfree(stats); 3192 } 3193 EXPORT_SYMBOL(rdma_free_hw_stats_struct); 3194