1 /* 2 * Copyright (c) 2004 Mellanox Technologies Ltd. All rights reserved. 3 * Copyright (c) 2004 Infinicon Corporation. All rights reserved. 4 * Copyright (c) 2004 Intel Corporation. All rights reserved. 5 * Copyright (c) 2004 Topspin Corporation. All rights reserved. 6 * Copyright (c) 2004 Voltaire Corporation. All rights reserved. 7 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved. 8 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved. 9 * 10 * This software is available to you under a choice of one of two 11 * licenses. You may choose to be licensed under the terms of the GNU 12 * General Public License (GPL) Version 2, available from the file 13 * COPYING in the main directory of this source tree, or the 14 * OpenIB.org BSD license below: 15 * 16 * Redistribution and use in source and binary forms, with or 17 * without modification, are permitted provided that the following 18 * conditions are met: 19 * 20 * - Redistributions of source code must retain the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer. 23 * 24 * - Redistributions in binary form must reproduce the above 25 * copyright notice, this list of conditions and the following 26 * disclaimer in the documentation and/or other materials 27 * provided with the distribution. 28 * 29 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 30 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 31 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 32 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 33 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 34 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 35 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 36 * SOFTWARE. 37 */ 38 39 #include <linux/errno.h> 40 #include <linux/err.h> 41 #include <linux/export.h> 42 #include <linux/string.h> 43 #include <linux/slab.h> 44 #include <linux/in.h> 45 #include <linux/in6.h> 46 #include <net/addrconf.h> 47 #include <linux/security.h> 48 49 #include <rdma/ib_verbs.h> 50 #include <rdma/ib_cache.h> 51 #include <rdma/ib_addr.h> 52 #include <rdma/rw.h> 53 54 #include "core_priv.h" 55 56 static int ib_resolve_eth_dmac(struct ib_device *device, 57 struct rdma_ah_attr *ah_attr); 58 59 static const char * const ib_events[] = { 60 [IB_EVENT_CQ_ERR] = "CQ error", 61 [IB_EVENT_QP_FATAL] = "QP fatal error", 62 [IB_EVENT_QP_REQ_ERR] = "QP request error", 63 [IB_EVENT_QP_ACCESS_ERR] = "QP access error", 64 [IB_EVENT_COMM_EST] = "communication established", 65 [IB_EVENT_SQ_DRAINED] = "send queue drained", 66 [IB_EVENT_PATH_MIG] = "path migration successful", 67 [IB_EVENT_PATH_MIG_ERR] = "path migration error", 68 [IB_EVENT_DEVICE_FATAL] = "device fatal error", 69 [IB_EVENT_PORT_ACTIVE] = "port active", 70 [IB_EVENT_PORT_ERR] = "port error", 71 [IB_EVENT_LID_CHANGE] = "LID change", 72 [IB_EVENT_PKEY_CHANGE] = "P_key change", 73 [IB_EVENT_SM_CHANGE] = "SM change", 74 [IB_EVENT_SRQ_ERR] = "SRQ error", 75 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached", 76 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached", 77 [IB_EVENT_CLIENT_REREGISTER] = "client reregister", 78 [IB_EVENT_GID_CHANGE] = "GID changed", 79 }; 80 81 const char *__attribute_const__ ib_event_msg(enum ib_event_type event) 82 { 83 size_t index = event; 84 85 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ? 86 ib_events[index] : "unrecognized event"; 87 } 88 EXPORT_SYMBOL(ib_event_msg); 89 90 static const char * const wc_statuses[] = { 91 [IB_WC_SUCCESS] = "success", 92 [IB_WC_LOC_LEN_ERR] = "local length error", 93 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error", 94 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error", 95 [IB_WC_LOC_PROT_ERR] = "local protection error", 96 [IB_WC_WR_FLUSH_ERR] = "WR flushed", 97 [IB_WC_MW_BIND_ERR] = "memory management operation error", 98 [IB_WC_BAD_RESP_ERR] = "bad response error", 99 [IB_WC_LOC_ACCESS_ERR] = "local access error", 100 [IB_WC_REM_INV_REQ_ERR] = "invalid request error", 101 [IB_WC_REM_ACCESS_ERR] = "remote access error", 102 [IB_WC_REM_OP_ERR] = "remote operation error", 103 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded", 104 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded", 105 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error", 106 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request", 107 [IB_WC_REM_ABORT_ERR] = "operation aborted", 108 [IB_WC_INV_EECN_ERR] = "invalid EE context number", 109 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state", 110 [IB_WC_FATAL_ERR] = "fatal error", 111 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error", 112 [IB_WC_GENERAL_ERR] = "general error", 113 }; 114 115 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status) 116 { 117 size_t index = status; 118 119 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ? 120 wc_statuses[index] : "unrecognized status"; 121 } 122 EXPORT_SYMBOL(ib_wc_status_msg); 123 124 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate) 125 { 126 switch (rate) { 127 case IB_RATE_2_5_GBPS: return 1; 128 case IB_RATE_5_GBPS: return 2; 129 case IB_RATE_10_GBPS: return 4; 130 case IB_RATE_20_GBPS: return 8; 131 case IB_RATE_30_GBPS: return 12; 132 case IB_RATE_40_GBPS: return 16; 133 case IB_RATE_60_GBPS: return 24; 134 case IB_RATE_80_GBPS: return 32; 135 case IB_RATE_120_GBPS: return 48; 136 case IB_RATE_14_GBPS: return 6; 137 case IB_RATE_56_GBPS: return 22; 138 case IB_RATE_112_GBPS: return 45; 139 case IB_RATE_168_GBPS: return 67; 140 case IB_RATE_25_GBPS: return 10; 141 case IB_RATE_100_GBPS: return 40; 142 case IB_RATE_200_GBPS: return 80; 143 case IB_RATE_300_GBPS: return 120; 144 default: return -1; 145 } 146 } 147 EXPORT_SYMBOL(ib_rate_to_mult); 148 149 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult) 150 { 151 switch (mult) { 152 case 1: return IB_RATE_2_5_GBPS; 153 case 2: return IB_RATE_5_GBPS; 154 case 4: return IB_RATE_10_GBPS; 155 case 8: return IB_RATE_20_GBPS; 156 case 12: return IB_RATE_30_GBPS; 157 case 16: return IB_RATE_40_GBPS; 158 case 24: return IB_RATE_60_GBPS; 159 case 32: return IB_RATE_80_GBPS; 160 case 48: return IB_RATE_120_GBPS; 161 case 6: return IB_RATE_14_GBPS; 162 case 22: return IB_RATE_56_GBPS; 163 case 45: return IB_RATE_112_GBPS; 164 case 67: return IB_RATE_168_GBPS; 165 case 10: return IB_RATE_25_GBPS; 166 case 40: return IB_RATE_100_GBPS; 167 case 80: return IB_RATE_200_GBPS; 168 case 120: return IB_RATE_300_GBPS; 169 default: return IB_RATE_PORT_CURRENT; 170 } 171 } 172 EXPORT_SYMBOL(mult_to_ib_rate); 173 174 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate) 175 { 176 switch (rate) { 177 case IB_RATE_2_5_GBPS: return 2500; 178 case IB_RATE_5_GBPS: return 5000; 179 case IB_RATE_10_GBPS: return 10000; 180 case IB_RATE_20_GBPS: return 20000; 181 case IB_RATE_30_GBPS: return 30000; 182 case IB_RATE_40_GBPS: return 40000; 183 case IB_RATE_60_GBPS: return 60000; 184 case IB_RATE_80_GBPS: return 80000; 185 case IB_RATE_120_GBPS: return 120000; 186 case IB_RATE_14_GBPS: return 14062; 187 case IB_RATE_56_GBPS: return 56250; 188 case IB_RATE_112_GBPS: return 112500; 189 case IB_RATE_168_GBPS: return 168750; 190 case IB_RATE_25_GBPS: return 25781; 191 case IB_RATE_100_GBPS: return 103125; 192 case IB_RATE_200_GBPS: return 206250; 193 case IB_RATE_300_GBPS: return 309375; 194 default: return -1; 195 } 196 } 197 EXPORT_SYMBOL(ib_rate_to_mbps); 198 199 __attribute_const__ enum rdma_transport_type 200 rdma_node_get_transport(enum rdma_node_type node_type) 201 { 202 203 if (node_type == RDMA_NODE_USNIC) 204 return RDMA_TRANSPORT_USNIC; 205 if (node_type == RDMA_NODE_USNIC_UDP) 206 return RDMA_TRANSPORT_USNIC_UDP; 207 if (node_type == RDMA_NODE_RNIC) 208 return RDMA_TRANSPORT_IWARP; 209 210 return RDMA_TRANSPORT_IB; 211 } 212 EXPORT_SYMBOL(rdma_node_get_transport); 213 214 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num) 215 { 216 enum rdma_transport_type lt; 217 if (device->get_link_layer) 218 return device->get_link_layer(device, port_num); 219 220 lt = rdma_node_get_transport(device->node_type); 221 if (lt == RDMA_TRANSPORT_IB) 222 return IB_LINK_LAYER_INFINIBAND; 223 224 return IB_LINK_LAYER_ETHERNET; 225 } 226 EXPORT_SYMBOL(rdma_port_get_link_layer); 227 228 /* Protection domains */ 229 230 /** 231 * ib_alloc_pd - Allocates an unused protection domain. 232 * @device: The device on which to allocate the protection domain. 233 * 234 * A protection domain object provides an association between QPs, shared 235 * receive queues, address handles, memory regions, and memory windows. 236 * 237 * Every PD has a local_dma_lkey which can be used as the lkey value for local 238 * memory operations. 239 */ 240 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags, 241 const char *caller) 242 { 243 struct ib_pd *pd; 244 int mr_access_flags = 0; 245 246 pd = device->alloc_pd(device, NULL, NULL); 247 if (IS_ERR(pd)) 248 return pd; 249 250 pd->device = device; 251 pd->uobject = NULL; 252 pd->__internal_mr = NULL; 253 atomic_set(&pd->usecnt, 0); 254 pd->flags = flags; 255 256 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY) 257 pd->local_dma_lkey = device->local_dma_lkey; 258 else 259 mr_access_flags |= IB_ACCESS_LOCAL_WRITE; 260 261 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) { 262 pr_warn("%s: enabling unsafe global rkey\n", caller); 263 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE; 264 } 265 266 pd->res.type = RDMA_RESTRACK_PD; 267 pd->res.kern_name = caller; 268 rdma_restrack_add(&pd->res); 269 270 if (mr_access_flags) { 271 struct ib_mr *mr; 272 273 mr = pd->device->get_dma_mr(pd, mr_access_flags); 274 if (IS_ERR(mr)) { 275 ib_dealloc_pd(pd); 276 return ERR_CAST(mr); 277 } 278 279 mr->device = pd->device; 280 mr->pd = pd; 281 mr->uobject = NULL; 282 mr->need_inval = false; 283 284 pd->__internal_mr = mr; 285 286 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) 287 pd->local_dma_lkey = pd->__internal_mr->lkey; 288 289 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) 290 pd->unsafe_global_rkey = pd->__internal_mr->rkey; 291 } 292 293 return pd; 294 } 295 EXPORT_SYMBOL(__ib_alloc_pd); 296 297 /** 298 * ib_dealloc_pd - Deallocates a protection domain. 299 * @pd: The protection domain to deallocate. 300 * 301 * It is an error to call this function while any resources in the pd still 302 * exist. The caller is responsible to synchronously destroy them and 303 * guarantee no new allocations will happen. 304 */ 305 void ib_dealloc_pd(struct ib_pd *pd) 306 { 307 int ret; 308 309 if (pd->__internal_mr) { 310 ret = pd->device->dereg_mr(pd->__internal_mr); 311 WARN_ON(ret); 312 pd->__internal_mr = NULL; 313 } 314 315 /* uverbs manipulates usecnt with proper locking, while the kabi 316 requires the caller to guarantee we can't race here. */ 317 WARN_ON(atomic_read(&pd->usecnt)); 318 319 rdma_restrack_del(&pd->res); 320 /* Making delalloc_pd a void return is a WIP, no driver should return 321 an error here. */ 322 ret = pd->device->dealloc_pd(pd); 323 WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd"); 324 } 325 EXPORT_SYMBOL(ib_dealloc_pd); 326 327 /* Address handles */ 328 329 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd, 330 struct rdma_ah_attr *ah_attr, 331 struct ib_udata *udata) 332 { 333 struct ib_ah *ah; 334 335 ah = pd->device->create_ah(pd, ah_attr, udata); 336 337 if (!IS_ERR(ah)) { 338 ah->device = pd->device; 339 ah->pd = pd; 340 ah->uobject = NULL; 341 ah->type = ah_attr->type; 342 atomic_inc(&pd->usecnt); 343 } 344 345 return ah; 346 } 347 348 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr) 349 { 350 return _rdma_create_ah(pd, ah_attr, NULL); 351 } 352 EXPORT_SYMBOL(rdma_create_ah); 353 354 /** 355 * rdma_create_user_ah - Creates an address handle for the 356 * given address vector. 357 * It resolves destination mac address for ah attribute of RoCE type. 358 * @pd: The protection domain associated with the address handle. 359 * @ah_attr: The attributes of the address vector. 360 * @udata: pointer to user's input output buffer information need by 361 * provider driver. 362 * 363 * It returns 0 on success and returns appropriate error code on error. 364 * The address handle is used to reference a local or global destination 365 * in all UD QP post sends. 366 */ 367 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd, 368 struct rdma_ah_attr *ah_attr, 369 struct ib_udata *udata) 370 { 371 int err; 372 373 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) { 374 err = ib_resolve_eth_dmac(pd->device, ah_attr); 375 if (err) 376 return ERR_PTR(err); 377 } 378 379 return _rdma_create_ah(pd, ah_attr, udata); 380 } 381 EXPORT_SYMBOL(rdma_create_user_ah); 382 383 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr) 384 { 385 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh; 386 struct iphdr ip4h_checked; 387 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh; 388 389 /* If it's IPv6, the version must be 6, otherwise, the first 390 * 20 bytes (before the IPv4 header) are garbled. 391 */ 392 if (ip6h->version != 6) 393 return (ip4h->version == 4) ? 4 : 0; 394 /* version may be 6 or 4 because the first 20 bytes could be garbled */ 395 396 /* RoCE v2 requires no options, thus header length 397 * must be 5 words 398 */ 399 if (ip4h->ihl != 5) 400 return 6; 401 402 /* Verify checksum. 403 * We can't write on scattered buffers so we need to copy to 404 * temp buffer. 405 */ 406 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked)); 407 ip4h_checked.check = 0; 408 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5); 409 /* if IPv4 header checksum is OK, believe it */ 410 if (ip4h->check == ip4h_checked.check) 411 return 4; 412 return 6; 413 } 414 EXPORT_SYMBOL(ib_get_rdma_header_version); 415 416 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device, 417 u8 port_num, 418 const struct ib_grh *grh) 419 { 420 int grh_version; 421 422 if (rdma_protocol_ib(device, port_num)) 423 return RDMA_NETWORK_IB; 424 425 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh); 426 427 if (grh_version == 4) 428 return RDMA_NETWORK_IPV4; 429 430 if (grh->next_hdr == IPPROTO_UDP) 431 return RDMA_NETWORK_IPV6; 432 433 return RDMA_NETWORK_ROCE_V1; 434 } 435 436 struct find_gid_index_context { 437 u16 vlan_id; 438 enum ib_gid_type gid_type; 439 }; 440 441 static bool find_gid_index(const union ib_gid *gid, 442 const struct ib_gid_attr *gid_attr, 443 void *context) 444 { 445 struct find_gid_index_context *ctx = context; 446 447 if (ctx->gid_type != gid_attr->gid_type) 448 return false; 449 450 if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) || 451 (is_vlan_dev(gid_attr->ndev) && 452 vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id)) 453 return false; 454 455 return true; 456 } 457 458 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num, 459 u16 vlan_id, const union ib_gid *sgid, 460 enum ib_gid_type gid_type, 461 u16 *gid_index) 462 { 463 struct find_gid_index_context context = {.vlan_id = vlan_id, 464 .gid_type = gid_type}; 465 466 return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index, 467 &context, gid_index); 468 } 469 470 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr, 471 enum rdma_network_type net_type, 472 union ib_gid *sgid, union ib_gid *dgid) 473 { 474 struct sockaddr_in src_in; 475 struct sockaddr_in dst_in; 476 __be32 src_saddr, dst_saddr; 477 478 if (!sgid || !dgid) 479 return -EINVAL; 480 481 if (net_type == RDMA_NETWORK_IPV4) { 482 memcpy(&src_in.sin_addr.s_addr, 483 &hdr->roce4grh.saddr, 4); 484 memcpy(&dst_in.sin_addr.s_addr, 485 &hdr->roce4grh.daddr, 4); 486 src_saddr = src_in.sin_addr.s_addr; 487 dst_saddr = dst_in.sin_addr.s_addr; 488 ipv6_addr_set_v4mapped(src_saddr, 489 (struct in6_addr *)sgid); 490 ipv6_addr_set_v4mapped(dst_saddr, 491 (struct in6_addr *)dgid); 492 return 0; 493 } else if (net_type == RDMA_NETWORK_IPV6 || 494 net_type == RDMA_NETWORK_IB) { 495 *dgid = hdr->ibgrh.dgid; 496 *sgid = hdr->ibgrh.sgid; 497 return 0; 498 } else { 499 return -EINVAL; 500 } 501 } 502 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr); 503 504 /* Resolve destination mac address and hop limit for unicast destination 505 * GID entry, considering the source GID entry as well. 506 * ah_attribute must have have valid port_num, sgid_index. 507 */ 508 static int ib_resolve_unicast_gid_dmac(struct ib_device *device, 509 struct rdma_ah_attr *ah_attr) 510 { 511 struct ib_gid_attr sgid_attr; 512 struct ib_global_route *grh; 513 int hop_limit = 0xff; 514 union ib_gid sgid; 515 int ret; 516 517 grh = rdma_ah_retrieve_grh(ah_attr); 518 519 ret = ib_query_gid(device, 520 rdma_ah_get_port_num(ah_attr), 521 grh->sgid_index, 522 &sgid, &sgid_attr); 523 if (ret || !sgid_attr.ndev) { 524 if (!ret) 525 ret = -ENXIO; 526 return ret; 527 } 528 529 /* If destination is link local and source GID is RoCEv1, 530 * IP stack is not used. 531 */ 532 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) && 533 sgid_attr.gid_type == IB_GID_TYPE_ROCE) { 534 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw, 535 ah_attr->roce.dmac); 536 goto done; 537 } 538 539 ret = rdma_addr_find_l2_eth_by_grh(&sgid, &grh->dgid, 540 ah_attr->roce.dmac, 541 sgid_attr.ndev, &hop_limit); 542 done: 543 dev_put(sgid_attr.ndev); 544 545 grh->hop_limit = hop_limit; 546 return ret; 547 } 548 549 /* 550 * This function initializes address handle attributes from the incoming packet. 551 * Incoming packet has dgid of the receiver node on which this code is 552 * getting executed and, sgid contains the GID of the sender. 553 * 554 * When resolving mac address of destination, the arrived dgid is used 555 * as sgid and, sgid is used as dgid because sgid contains destinations 556 * GID whom to respond to. 557 * 558 */ 559 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num, 560 const struct ib_wc *wc, const struct ib_grh *grh, 561 struct rdma_ah_attr *ah_attr) 562 { 563 u32 flow_class; 564 u16 gid_index; 565 int ret; 566 enum rdma_network_type net_type = RDMA_NETWORK_IB; 567 enum ib_gid_type gid_type = IB_GID_TYPE_IB; 568 int hoplimit = 0xff; 569 union ib_gid dgid; 570 union ib_gid sgid; 571 572 might_sleep(); 573 574 memset(ah_attr, 0, sizeof *ah_attr); 575 ah_attr->type = rdma_ah_find_type(device, port_num); 576 if (rdma_cap_eth_ah(device, port_num)) { 577 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE) 578 net_type = wc->network_hdr_type; 579 else 580 net_type = ib_get_net_type_by_grh(device, port_num, grh); 581 gid_type = ib_network_to_gid_type(net_type); 582 } 583 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type, 584 &sgid, &dgid); 585 if (ret) 586 return ret; 587 588 rdma_ah_set_sl(ah_attr, wc->sl); 589 rdma_ah_set_port_num(ah_attr, port_num); 590 591 if (rdma_protocol_roce(device, port_num)) { 592 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ? 593 wc->vlan_id : 0xffff; 594 595 if (!(wc->wc_flags & IB_WC_GRH)) 596 return -EPROTOTYPE; 597 598 ret = get_sgid_index_from_eth(device, port_num, 599 vlan_id, &dgid, 600 gid_type, &gid_index); 601 if (ret) 602 return ret; 603 604 flow_class = be32_to_cpu(grh->version_tclass_flow); 605 rdma_ah_set_grh(ah_attr, &sgid, 606 flow_class & 0xFFFFF, 607 (u8)gid_index, hoplimit, 608 (flow_class >> 20) & 0xFF); 609 return ib_resolve_unicast_gid_dmac(device, ah_attr); 610 } else { 611 rdma_ah_set_dlid(ah_attr, wc->slid); 612 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits); 613 614 if (wc->wc_flags & IB_WC_GRH) { 615 if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) { 616 ret = ib_find_cached_gid_by_port(device, &dgid, 617 IB_GID_TYPE_IB, 618 port_num, NULL, 619 &gid_index); 620 if (ret) 621 return ret; 622 } else { 623 gid_index = 0; 624 } 625 626 flow_class = be32_to_cpu(grh->version_tclass_flow); 627 rdma_ah_set_grh(ah_attr, &sgid, 628 flow_class & 0xFFFFF, 629 (u8)gid_index, hoplimit, 630 (flow_class >> 20) & 0xFF); 631 } 632 return 0; 633 } 634 } 635 EXPORT_SYMBOL(ib_init_ah_attr_from_wc); 636 637 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc, 638 const struct ib_grh *grh, u8 port_num) 639 { 640 struct rdma_ah_attr ah_attr; 641 int ret; 642 643 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr); 644 if (ret) 645 return ERR_PTR(ret); 646 647 return rdma_create_ah(pd, &ah_attr); 648 } 649 EXPORT_SYMBOL(ib_create_ah_from_wc); 650 651 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr) 652 { 653 if (ah->type != ah_attr->type) 654 return -EINVAL; 655 656 return ah->device->modify_ah ? 657 ah->device->modify_ah(ah, ah_attr) : 658 -ENOSYS; 659 } 660 EXPORT_SYMBOL(rdma_modify_ah); 661 662 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr) 663 { 664 return ah->device->query_ah ? 665 ah->device->query_ah(ah, ah_attr) : 666 -ENOSYS; 667 } 668 EXPORT_SYMBOL(rdma_query_ah); 669 670 int rdma_destroy_ah(struct ib_ah *ah) 671 { 672 struct ib_pd *pd; 673 int ret; 674 675 pd = ah->pd; 676 ret = ah->device->destroy_ah(ah); 677 if (!ret) 678 atomic_dec(&pd->usecnt); 679 680 return ret; 681 } 682 EXPORT_SYMBOL(rdma_destroy_ah); 683 684 /* Shared receive queues */ 685 686 struct ib_srq *ib_create_srq(struct ib_pd *pd, 687 struct ib_srq_init_attr *srq_init_attr) 688 { 689 struct ib_srq *srq; 690 691 if (!pd->device->create_srq) 692 return ERR_PTR(-ENOSYS); 693 694 srq = pd->device->create_srq(pd, srq_init_attr, NULL); 695 696 if (!IS_ERR(srq)) { 697 srq->device = pd->device; 698 srq->pd = pd; 699 srq->uobject = NULL; 700 srq->event_handler = srq_init_attr->event_handler; 701 srq->srq_context = srq_init_attr->srq_context; 702 srq->srq_type = srq_init_attr->srq_type; 703 if (ib_srq_has_cq(srq->srq_type)) { 704 srq->ext.cq = srq_init_attr->ext.cq; 705 atomic_inc(&srq->ext.cq->usecnt); 706 } 707 if (srq->srq_type == IB_SRQT_XRC) { 708 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd; 709 atomic_inc(&srq->ext.xrc.xrcd->usecnt); 710 } 711 atomic_inc(&pd->usecnt); 712 atomic_set(&srq->usecnt, 0); 713 } 714 715 return srq; 716 } 717 EXPORT_SYMBOL(ib_create_srq); 718 719 int ib_modify_srq(struct ib_srq *srq, 720 struct ib_srq_attr *srq_attr, 721 enum ib_srq_attr_mask srq_attr_mask) 722 { 723 return srq->device->modify_srq ? 724 srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) : 725 -ENOSYS; 726 } 727 EXPORT_SYMBOL(ib_modify_srq); 728 729 int ib_query_srq(struct ib_srq *srq, 730 struct ib_srq_attr *srq_attr) 731 { 732 return srq->device->query_srq ? 733 srq->device->query_srq(srq, srq_attr) : -ENOSYS; 734 } 735 EXPORT_SYMBOL(ib_query_srq); 736 737 int ib_destroy_srq(struct ib_srq *srq) 738 { 739 struct ib_pd *pd; 740 enum ib_srq_type srq_type; 741 struct ib_xrcd *uninitialized_var(xrcd); 742 struct ib_cq *uninitialized_var(cq); 743 int ret; 744 745 if (atomic_read(&srq->usecnt)) 746 return -EBUSY; 747 748 pd = srq->pd; 749 srq_type = srq->srq_type; 750 if (ib_srq_has_cq(srq_type)) 751 cq = srq->ext.cq; 752 if (srq_type == IB_SRQT_XRC) 753 xrcd = srq->ext.xrc.xrcd; 754 755 ret = srq->device->destroy_srq(srq); 756 if (!ret) { 757 atomic_dec(&pd->usecnt); 758 if (srq_type == IB_SRQT_XRC) 759 atomic_dec(&xrcd->usecnt); 760 if (ib_srq_has_cq(srq_type)) 761 atomic_dec(&cq->usecnt); 762 } 763 764 return ret; 765 } 766 EXPORT_SYMBOL(ib_destroy_srq); 767 768 /* Queue pairs */ 769 770 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context) 771 { 772 struct ib_qp *qp = context; 773 unsigned long flags; 774 775 spin_lock_irqsave(&qp->device->event_handler_lock, flags); 776 list_for_each_entry(event->element.qp, &qp->open_list, open_list) 777 if (event->element.qp->event_handler) 778 event->element.qp->event_handler(event, event->element.qp->qp_context); 779 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags); 780 } 781 782 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp) 783 { 784 mutex_lock(&xrcd->tgt_qp_mutex); 785 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list); 786 mutex_unlock(&xrcd->tgt_qp_mutex); 787 } 788 789 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp, 790 void (*event_handler)(struct ib_event *, void *), 791 void *qp_context) 792 { 793 struct ib_qp *qp; 794 unsigned long flags; 795 int err; 796 797 qp = kzalloc(sizeof *qp, GFP_KERNEL); 798 if (!qp) 799 return ERR_PTR(-ENOMEM); 800 801 qp->real_qp = real_qp; 802 err = ib_open_shared_qp_security(qp, real_qp->device); 803 if (err) { 804 kfree(qp); 805 return ERR_PTR(err); 806 } 807 808 qp->real_qp = real_qp; 809 atomic_inc(&real_qp->usecnt); 810 qp->device = real_qp->device; 811 qp->event_handler = event_handler; 812 qp->qp_context = qp_context; 813 qp->qp_num = real_qp->qp_num; 814 qp->qp_type = real_qp->qp_type; 815 816 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags); 817 list_add(&qp->open_list, &real_qp->open_list); 818 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags); 819 820 return qp; 821 } 822 823 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd, 824 struct ib_qp_open_attr *qp_open_attr) 825 { 826 struct ib_qp *qp, *real_qp; 827 828 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT) 829 return ERR_PTR(-EINVAL); 830 831 qp = ERR_PTR(-EINVAL); 832 mutex_lock(&xrcd->tgt_qp_mutex); 833 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) { 834 if (real_qp->qp_num == qp_open_attr->qp_num) { 835 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler, 836 qp_open_attr->qp_context); 837 break; 838 } 839 } 840 mutex_unlock(&xrcd->tgt_qp_mutex); 841 return qp; 842 } 843 EXPORT_SYMBOL(ib_open_qp); 844 845 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp, 846 struct ib_qp_init_attr *qp_init_attr) 847 { 848 struct ib_qp *real_qp = qp; 849 850 qp->event_handler = __ib_shared_qp_event_handler; 851 qp->qp_context = qp; 852 qp->pd = NULL; 853 qp->send_cq = qp->recv_cq = NULL; 854 qp->srq = NULL; 855 qp->xrcd = qp_init_attr->xrcd; 856 atomic_inc(&qp_init_attr->xrcd->usecnt); 857 INIT_LIST_HEAD(&qp->open_list); 858 859 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler, 860 qp_init_attr->qp_context); 861 if (!IS_ERR(qp)) 862 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp); 863 else 864 real_qp->device->destroy_qp(real_qp); 865 return qp; 866 } 867 868 struct ib_qp *ib_create_qp(struct ib_pd *pd, 869 struct ib_qp_init_attr *qp_init_attr) 870 { 871 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device; 872 struct ib_qp *qp; 873 int ret; 874 875 if (qp_init_attr->rwq_ind_tbl && 876 (qp_init_attr->recv_cq || 877 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr || 878 qp_init_attr->cap.max_recv_sge)) 879 return ERR_PTR(-EINVAL); 880 881 /* 882 * If the callers is using the RDMA API calculate the resources 883 * needed for the RDMA READ/WRITE operations. 884 * 885 * Note that these callers need to pass in a port number. 886 */ 887 if (qp_init_attr->cap.max_rdma_ctxs) 888 rdma_rw_init_qp(device, qp_init_attr); 889 890 qp = _ib_create_qp(device, pd, qp_init_attr, NULL); 891 if (IS_ERR(qp)) 892 return qp; 893 894 ret = ib_create_qp_security(qp, device); 895 if (ret) { 896 ib_destroy_qp(qp); 897 return ERR_PTR(ret); 898 } 899 900 qp->real_qp = qp; 901 qp->uobject = NULL; 902 qp->qp_type = qp_init_attr->qp_type; 903 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl; 904 905 atomic_set(&qp->usecnt, 0); 906 qp->mrs_used = 0; 907 spin_lock_init(&qp->mr_lock); 908 INIT_LIST_HEAD(&qp->rdma_mrs); 909 INIT_LIST_HEAD(&qp->sig_mrs); 910 qp->port = 0; 911 912 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) 913 return ib_create_xrc_qp(qp, qp_init_attr); 914 915 qp->event_handler = qp_init_attr->event_handler; 916 qp->qp_context = qp_init_attr->qp_context; 917 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) { 918 qp->recv_cq = NULL; 919 qp->srq = NULL; 920 } else { 921 qp->recv_cq = qp_init_attr->recv_cq; 922 if (qp_init_attr->recv_cq) 923 atomic_inc(&qp_init_attr->recv_cq->usecnt); 924 qp->srq = qp_init_attr->srq; 925 if (qp->srq) 926 atomic_inc(&qp_init_attr->srq->usecnt); 927 } 928 929 qp->send_cq = qp_init_attr->send_cq; 930 qp->xrcd = NULL; 931 932 atomic_inc(&pd->usecnt); 933 if (qp_init_attr->send_cq) 934 atomic_inc(&qp_init_attr->send_cq->usecnt); 935 if (qp_init_attr->rwq_ind_tbl) 936 atomic_inc(&qp->rwq_ind_tbl->usecnt); 937 938 if (qp_init_attr->cap.max_rdma_ctxs) { 939 ret = rdma_rw_init_mrs(qp, qp_init_attr); 940 if (ret) { 941 pr_err("failed to init MR pool ret= %d\n", ret); 942 ib_destroy_qp(qp); 943 return ERR_PTR(ret); 944 } 945 } 946 947 /* 948 * Note: all hw drivers guarantee that max_send_sge is lower than 949 * the device RDMA WRITE SGE limit but not all hw drivers ensure that 950 * max_send_sge <= max_sge_rd. 951 */ 952 qp->max_write_sge = qp_init_attr->cap.max_send_sge; 953 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge, 954 device->attrs.max_sge_rd); 955 956 return qp; 957 } 958 EXPORT_SYMBOL(ib_create_qp); 959 960 static const struct { 961 int valid; 962 enum ib_qp_attr_mask req_param[IB_QPT_MAX]; 963 enum ib_qp_attr_mask opt_param[IB_QPT_MAX]; 964 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = { 965 [IB_QPS_RESET] = { 966 [IB_QPS_RESET] = { .valid = 1 }, 967 [IB_QPS_INIT] = { 968 .valid = 1, 969 .req_param = { 970 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 971 IB_QP_PORT | 972 IB_QP_QKEY), 973 [IB_QPT_RAW_PACKET] = IB_QP_PORT, 974 [IB_QPT_UC] = (IB_QP_PKEY_INDEX | 975 IB_QP_PORT | 976 IB_QP_ACCESS_FLAGS), 977 [IB_QPT_RC] = (IB_QP_PKEY_INDEX | 978 IB_QP_PORT | 979 IB_QP_ACCESS_FLAGS), 980 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX | 981 IB_QP_PORT | 982 IB_QP_ACCESS_FLAGS), 983 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX | 984 IB_QP_PORT | 985 IB_QP_ACCESS_FLAGS), 986 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 987 IB_QP_QKEY), 988 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 989 IB_QP_QKEY), 990 } 991 }, 992 }, 993 [IB_QPS_INIT] = { 994 [IB_QPS_RESET] = { .valid = 1 }, 995 [IB_QPS_ERR] = { .valid = 1 }, 996 [IB_QPS_INIT] = { 997 .valid = 1, 998 .opt_param = { 999 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1000 IB_QP_PORT | 1001 IB_QP_QKEY), 1002 [IB_QPT_UC] = (IB_QP_PKEY_INDEX | 1003 IB_QP_PORT | 1004 IB_QP_ACCESS_FLAGS), 1005 [IB_QPT_RC] = (IB_QP_PKEY_INDEX | 1006 IB_QP_PORT | 1007 IB_QP_ACCESS_FLAGS), 1008 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX | 1009 IB_QP_PORT | 1010 IB_QP_ACCESS_FLAGS), 1011 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX | 1012 IB_QP_PORT | 1013 IB_QP_ACCESS_FLAGS), 1014 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1015 IB_QP_QKEY), 1016 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1017 IB_QP_QKEY), 1018 } 1019 }, 1020 [IB_QPS_RTR] = { 1021 .valid = 1, 1022 .req_param = { 1023 [IB_QPT_UC] = (IB_QP_AV | 1024 IB_QP_PATH_MTU | 1025 IB_QP_DEST_QPN | 1026 IB_QP_RQ_PSN), 1027 [IB_QPT_RC] = (IB_QP_AV | 1028 IB_QP_PATH_MTU | 1029 IB_QP_DEST_QPN | 1030 IB_QP_RQ_PSN | 1031 IB_QP_MAX_DEST_RD_ATOMIC | 1032 IB_QP_MIN_RNR_TIMER), 1033 [IB_QPT_XRC_INI] = (IB_QP_AV | 1034 IB_QP_PATH_MTU | 1035 IB_QP_DEST_QPN | 1036 IB_QP_RQ_PSN), 1037 [IB_QPT_XRC_TGT] = (IB_QP_AV | 1038 IB_QP_PATH_MTU | 1039 IB_QP_DEST_QPN | 1040 IB_QP_RQ_PSN | 1041 IB_QP_MAX_DEST_RD_ATOMIC | 1042 IB_QP_MIN_RNR_TIMER), 1043 }, 1044 .opt_param = { 1045 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1046 IB_QP_QKEY), 1047 [IB_QPT_UC] = (IB_QP_ALT_PATH | 1048 IB_QP_ACCESS_FLAGS | 1049 IB_QP_PKEY_INDEX), 1050 [IB_QPT_RC] = (IB_QP_ALT_PATH | 1051 IB_QP_ACCESS_FLAGS | 1052 IB_QP_PKEY_INDEX), 1053 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH | 1054 IB_QP_ACCESS_FLAGS | 1055 IB_QP_PKEY_INDEX), 1056 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH | 1057 IB_QP_ACCESS_FLAGS | 1058 IB_QP_PKEY_INDEX), 1059 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1060 IB_QP_QKEY), 1061 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1062 IB_QP_QKEY), 1063 }, 1064 }, 1065 }, 1066 [IB_QPS_RTR] = { 1067 [IB_QPS_RESET] = { .valid = 1 }, 1068 [IB_QPS_ERR] = { .valid = 1 }, 1069 [IB_QPS_RTS] = { 1070 .valid = 1, 1071 .req_param = { 1072 [IB_QPT_UD] = IB_QP_SQ_PSN, 1073 [IB_QPT_UC] = IB_QP_SQ_PSN, 1074 [IB_QPT_RC] = (IB_QP_TIMEOUT | 1075 IB_QP_RETRY_CNT | 1076 IB_QP_RNR_RETRY | 1077 IB_QP_SQ_PSN | 1078 IB_QP_MAX_QP_RD_ATOMIC), 1079 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT | 1080 IB_QP_RETRY_CNT | 1081 IB_QP_RNR_RETRY | 1082 IB_QP_SQ_PSN | 1083 IB_QP_MAX_QP_RD_ATOMIC), 1084 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT | 1085 IB_QP_SQ_PSN), 1086 [IB_QPT_SMI] = IB_QP_SQ_PSN, 1087 [IB_QPT_GSI] = IB_QP_SQ_PSN, 1088 }, 1089 .opt_param = { 1090 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1091 IB_QP_QKEY), 1092 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1093 IB_QP_ALT_PATH | 1094 IB_QP_ACCESS_FLAGS | 1095 IB_QP_PATH_MIG_STATE), 1096 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1097 IB_QP_ALT_PATH | 1098 IB_QP_ACCESS_FLAGS | 1099 IB_QP_MIN_RNR_TIMER | 1100 IB_QP_PATH_MIG_STATE), 1101 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1102 IB_QP_ALT_PATH | 1103 IB_QP_ACCESS_FLAGS | 1104 IB_QP_PATH_MIG_STATE), 1105 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1106 IB_QP_ALT_PATH | 1107 IB_QP_ACCESS_FLAGS | 1108 IB_QP_MIN_RNR_TIMER | 1109 IB_QP_PATH_MIG_STATE), 1110 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1111 IB_QP_QKEY), 1112 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1113 IB_QP_QKEY), 1114 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT, 1115 } 1116 } 1117 }, 1118 [IB_QPS_RTS] = { 1119 [IB_QPS_RESET] = { .valid = 1 }, 1120 [IB_QPS_ERR] = { .valid = 1 }, 1121 [IB_QPS_RTS] = { 1122 .valid = 1, 1123 .opt_param = { 1124 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1125 IB_QP_QKEY), 1126 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1127 IB_QP_ACCESS_FLAGS | 1128 IB_QP_ALT_PATH | 1129 IB_QP_PATH_MIG_STATE), 1130 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1131 IB_QP_ACCESS_FLAGS | 1132 IB_QP_ALT_PATH | 1133 IB_QP_PATH_MIG_STATE | 1134 IB_QP_MIN_RNR_TIMER), 1135 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1136 IB_QP_ACCESS_FLAGS | 1137 IB_QP_ALT_PATH | 1138 IB_QP_PATH_MIG_STATE), 1139 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1140 IB_QP_ACCESS_FLAGS | 1141 IB_QP_ALT_PATH | 1142 IB_QP_PATH_MIG_STATE | 1143 IB_QP_MIN_RNR_TIMER), 1144 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1145 IB_QP_QKEY), 1146 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1147 IB_QP_QKEY), 1148 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT, 1149 } 1150 }, 1151 [IB_QPS_SQD] = { 1152 .valid = 1, 1153 .opt_param = { 1154 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1155 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1156 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1157 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1158 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */ 1159 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1160 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY 1161 } 1162 }, 1163 }, 1164 [IB_QPS_SQD] = { 1165 [IB_QPS_RESET] = { .valid = 1 }, 1166 [IB_QPS_ERR] = { .valid = 1 }, 1167 [IB_QPS_RTS] = { 1168 .valid = 1, 1169 .opt_param = { 1170 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1171 IB_QP_QKEY), 1172 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1173 IB_QP_ALT_PATH | 1174 IB_QP_ACCESS_FLAGS | 1175 IB_QP_PATH_MIG_STATE), 1176 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1177 IB_QP_ALT_PATH | 1178 IB_QP_ACCESS_FLAGS | 1179 IB_QP_MIN_RNR_TIMER | 1180 IB_QP_PATH_MIG_STATE), 1181 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1182 IB_QP_ALT_PATH | 1183 IB_QP_ACCESS_FLAGS | 1184 IB_QP_PATH_MIG_STATE), 1185 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1186 IB_QP_ALT_PATH | 1187 IB_QP_ACCESS_FLAGS | 1188 IB_QP_MIN_RNR_TIMER | 1189 IB_QP_PATH_MIG_STATE), 1190 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1191 IB_QP_QKEY), 1192 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1193 IB_QP_QKEY), 1194 } 1195 }, 1196 [IB_QPS_SQD] = { 1197 .valid = 1, 1198 .opt_param = { 1199 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1200 IB_QP_QKEY), 1201 [IB_QPT_UC] = (IB_QP_AV | 1202 IB_QP_ALT_PATH | 1203 IB_QP_ACCESS_FLAGS | 1204 IB_QP_PKEY_INDEX | 1205 IB_QP_PATH_MIG_STATE), 1206 [IB_QPT_RC] = (IB_QP_PORT | 1207 IB_QP_AV | 1208 IB_QP_TIMEOUT | 1209 IB_QP_RETRY_CNT | 1210 IB_QP_RNR_RETRY | 1211 IB_QP_MAX_QP_RD_ATOMIC | 1212 IB_QP_MAX_DEST_RD_ATOMIC | 1213 IB_QP_ALT_PATH | 1214 IB_QP_ACCESS_FLAGS | 1215 IB_QP_PKEY_INDEX | 1216 IB_QP_MIN_RNR_TIMER | 1217 IB_QP_PATH_MIG_STATE), 1218 [IB_QPT_XRC_INI] = (IB_QP_PORT | 1219 IB_QP_AV | 1220 IB_QP_TIMEOUT | 1221 IB_QP_RETRY_CNT | 1222 IB_QP_RNR_RETRY | 1223 IB_QP_MAX_QP_RD_ATOMIC | 1224 IB_QP_ALT_PATH | 1225 IB_QP_ACCESS_FLAGS | 1226 IB_QP_PKEY_INDEX | 1227 IB_QP_PATH_MIG_STATE), 1228 [IB_QPT_XRC_TGT] = (IB_QP_PORT | 1229 IB_QP_AV | 1230 IB_QP_TIMEOUT | 1231 IB_QP_MAX_DEST_RD_ATOMIC | 1232 IB_QP_ALT_PATH | 1233 IB_QP_ACCESS_FLAGS | 1234 IB_QP_PKEY_INDEX | 1235 IB_QP_MIN_RNR_TIMER | 1236 IB_QP_PATH_MIG_STATE), 1237 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1238 IB_QP_QKEY), 1239 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1240 IB_QP_QKEY), 1241 } 1242 } 1243 }, 1244 [IB_QPS_SQE] = { 1245 [IB_QPS_RESET] = { .valid = 1 }, 1246 [IB_QPS_ERR] = { .valid = 1 }, 1247 [IB_QPS_RTS] = { 1248 .valid = 1, 1249 .opt_param = { 1250 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1251 IB_QP_QKEY), 1252 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1253 IB_QP_ACCESS_FLAGS), 1254 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1255 IB_QP_QKEY), 1256 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1257 IB_QP_QKEY), 1258 } 1259 } 1260 }, 1261 [IB_QPS_ERR] = { 1262 [IB_QPS_RESET] = { .valid = 1 }, 1263 [IB_QPS_ERR] = { .valid = 1 } 1264 } 1265 }; 1266 1267 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state, 1268 enum ib_qp_type type, enum ib_qp_attr_mask mask, 1269 enum rdma_link_layer ll) 1270 { 1271 enum ib_qp_attr_mask req_param, opt_param; 1272 1273 if (cur_state < 0 || cur_state > IB_QPS_ERR || 1274 next_state < 0 || next_state > IB_QPS_ERR) 1275 return 0; 1276 1277 if (mask & IB_QP_CUR_STATE && 1278 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS && 1279 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE) 1280 return 0; 1281 1282 if (!qp_state_table[cur_state][next_state].valid) 1283 return 0; 1284 1285 req_param = qp_state_table[cur_state][next_state].req_param[type]; 1286 opt_param = qp_state_table[cur_state][next_state].opt_param[type]; 1287 1288 if ((mask & req_param) != req_param) 1289 return 0; 1290 1291 if (mask & ~(req_param | opt_param | IB_QP_STATE)) 1292 return 0; 1293 1294 return 1; 1295 } 1296 EXPORT_SYMBOL(ib_modify_qp_is_ok); 1297 1298 static int ib_resolve_eth_dmac(struct ib_device *device, 1299 struct rdma_ah_attr *ah_attr) 1300 { 1301 int ret = 0; 1302 struct ib_global_route *grh; 1303 1304 if (!rdma_is_port_valid(device, rdma_ah_get_port_num(ah_attr))) 1305 return -EINVAL; 1306 1307 grh = rdma_ah_retrieve_grh(ah_attr); 1308 1309 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) { 1310 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) { 1311 __be32 addr = 0; 1312 1313 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4); 1314 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac); 1315 } else { 1316 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw, 1317 (char *)ah_attr->roce.dmac); 1318 } 1319 } else { 1320 ret = ib_resolve_unicast_gid_dmac(device, ah_attr); 1321 } 1322 return ret; 1323 } 1324 1325 /** 1326 * IB core internal function to perform QP attributes modification. 1327 */ 1328 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr, 1329 int attr_mask, struct ib_udata *udata) 1330 { 1331 u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port; 1332 int ret; 1333 1334 if (rdma_ib_or_roce(qp->device, port)) { 1335 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) { 1336 pr_warn("%s: %s rq_psn overflow, masking to 24 bits\n", 1337 __func__, qp->device->name); 1338 attr->rq_psn &= 0xffffff; 1339 } 1340 1341 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) { 1342 pr_warn("%s: %s sq_psn overflow, masking to 24 bits\n", 1343 __func__, qp->device->name); 1344 attr->sq_psn &= 0xffffff; 1345 } 1346 } 1347 1348 ret = ib_security_modify_qp(qp, attr, attr_mask, udata); 1349 if (!ret && (attr_mask & IB_QP_PORT)) 1350 qp->port = attr->port_num; 1351 1352 return ret; 1353 } 1354 1355 static bool is_qp_type_connected(const struct ib_qp *qp) 1356 { 1357 return (qp->qp_type == IB_QPT_UC || 1358 qp->qp_type == IB_QPT_RC || 1359 qp->qp_type == IB_QPT_XRC_INI || 1360 qp->qp_type == IB_QPT_XRC_TGT); 1361 } 1362 1363 /** 1364 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP. 1365 * @ib_qp: The QP to modify. 1366 * @attr: On input, specifies the QP attributes to modify. On output, 1367 * the current values of selected QP attributes are returned. 1368 * @attr_mask: A bit-mask used to specify which attributes of the QP 1369 * are being modified. 1370 * @udata: pointer to user's input output buffer information 1371 * are being modified. 1372 * It returns 0 on success and returns appropriate error code on error. 1373 */ 1374 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr, 1375 int attr_mask, struct ib_udata *udata) 1376 { 1377 struct ib_qp *qp = ib_qp->real_qp; 1378 int ret; 1379 1380 if (attr_mask & IB_QP_AV && 1381 attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE && 1382 is_qp_type_connected(qp)) { 1383 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr); 1384 if (ret) 1385 return ret; 1386 } 1387 return _ib_modify_qp(qp, attr, attr_mask, udata); 1388 } 1389 EXPORT_SYMBOL(ib_modify_qp_with_udata); 1390 1391 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width) 1392 { 1393 int rc; 1394 u32 netdev_speed; 1395 struct net_device *netdev; 1396 struct ethtool_link_ksettings lksettings; 1397 1398 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET) 1399 return -EINVAL; 1400 1401 if (!dev->get_netdev) 1402 return -EOPNOTSUPP; 1403 1404 netdev = dev->get_netdev(dev, port_num); 1405 if (!netdev) 1406 return -ENODEV; 1407 1408 rtnl_lock(); 1409 rc = __ethtool_get_link_ksettings(netdev, &lksettings); 1410 rtnl_unlock(); 1411 1412 dev_put(netdev); 1413 1414 if (!rc) { 1415 netdev_speed = lksettings.base.speed; 1416 } else { 1417 netdev_speed = SPEED_1000; 1418 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name, 1419 netdev_speed); 1420 } 1421 1422 if (netdev_speed <= SPEED_1000) { 1423 *width = IB_WIDTH_1X; 1424 *speed = IB_SPEED_SDR; 1425 } else if (netdev_speed <= SPEED_10000) { 1426 *width = IB_WIDTH_1X; 1427 *speed = IB_SPEED_FDR10; 1428 } else if (netdev_speed <= SPEED_20000) { 1429 *width = IB_WIDTH_4X; 1430 *speed = IB_SPEED_DDR; 1431 } else if (netdev_speed <= SPEED_25000) { 1432 *width = IB_WIDTH_1X; 1433 *speed = IB_SPEED_EDR; 1434 } else if (netdev_speed <= SPEED_40000) { 1435 *width = IB_WIDTH_4X; 1436 *speed = IB_SPEED_FDR10; 1437 } else { 1438 *width = IB_WIDTH_4X; 1439 *speed = IB_SPEED_EDR; 1440 } 1441 1442 return 0; 1443 } 1444 EXPORT_SYMBOL(ib_get_eth_speed); 1445 1446 int ib_modify_qp(struct ib_qp *qp, 1447 struct ib_qp_attr *qp_attr, 1448 int qp_attr_mask) 1449 { 1450 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL); 1451 } 1452 EXPORT_SYMBOL(ib_modify_qp); 1453 1454 int ib_query_qp(struct ib_qp *qp, 1455 struct ib_qp_attr *qp_attr, 1456 int qp_attr_mask, 1457 struct ib_qp_init_attr *qp_init_attr) 1458 { 1459 return qp->device->query_qp ? 1460 qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) : 1461 -ENOSYS; 1462 } 1463 EXPORT_SYMBOL(ib_query_qp); 1464 1465 int ib_close_qp(struct ib_qp *qp) 1466 { 1467 struct ib_qp *real_qp; 1468 unsigned long flags; 1469 1470 real_qp = qp->real_qp; 1471 if (real_qp == qp) 1472 return -EINVAL; 1473 1474 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags); 1475 list_del(&qp->open_list); 1476 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags); 1477 1478 atomic_dec(&real_qp->usecnt); 1479 if (qp->qp_sec) 1480 ib_close_shared_qp_security(qp->qp_sec); 1481 kfree(qp); 1482 1483 return 0; 1484 } 1485 EXPORT_SYMBOL(ib_close_qp); 1486 1487 static int __ib_destroy_shared_qp(struct ib_qp *qp) 1488 { 1489 struct ib_xrcd *xrcd; 1490 struct ib_qp *real_qp; 1491 int ret; 1492 1493 real_qp = qp->real_qp; 1494 xrcd = real_qp->xrcd; 1495 1496 mutex_lock(&xrcd->tgt_qp_mutex); 1497 ib_close_qp(qp); 1498 if (atomic_read(&real_qp->usecnt) == 0) 1499 list_del(&real_qp->xrcd_list); 1500 else 1501 real_qp = NULL; 1502 mutex_unlock(&xrcd->tgt_qp_mutex); 1503 1504 if (real_qp) { 1505 ret = ib_destroy_qp(real_qp); 1506 if (!ret) 1507 atomic_dec(&xrcd->usecnt); 1508 else 1509 __ib_insert_xrcd_qp(xrcd, real_qp); 1510 } 1511 1512 return 0; 1513 } 1514 1515 int ib_destroy_qp(struct ib_qp *qp) 1516 { 1517 struct ib_pd *pd; 1518 struct ib_cq *scq, *rcq; 1519 struct ib_srq *srq; 1520 struct ib_rwq_ind_table *ind_tbl; 1521 struct ib_qp_security *sec; 1522 int ret; 1523 1524 WARN_ON_ONCE(qp->mrs_used > 0); 1525 1526 if (atomic_read(&qp->usecnt)) 1527 return -EBUSY; 1528 1529 if (qp->real_qp != qp) 1530 return __ib_destroy_shared_qp(qp); 1531 1532 pd = qp->pd; 1533 scq = qp->send_cq; 1534 rcq = qp->recv_cq; 1535 srq = qp->srq; 1536 ind_tbl = qp->rwq_ind_tbl; 1537 sec = qp->qp_sec; 1538 if (sec) 1539 ib_destroy_qp_security_begin(sec); 1540 1541 if (!qp->uobject) 1542 rdma_rw_cleanup_mrs(qp); 1543 1544 rdma_restrack_del(&qp->res); 1545 ret = qp->device->destroy_qp(qp); 1546 if (!ret) { 1547 if (pd) 1548 atomic_dec(&pd->usecnt); 1549 if (scq) 1550 atomic_dec(&scq->usecnt); 1551 if (rcq) 1552 atomic_dec(&rcq->usecnt); 1553 if (srq) 1554 atomic_dec(&srq->usecnt); 1555 if (ind_tbl) 1556 atomic_dec(&ind_tbl->usecnt); 1557 if (sec) 1558 ib_destroy_qp_security_end(sec); 1559 } else { 1560 if (sec) 1561 ib_destroy_qp_security_abort(sec); 1562 } 1563 1564 return ret; 1565 } 1566 EXPORT_SYMBOL(ib_destroy_qp); 1567 1568 /* Completion queues */ 1569 1570 struct ib_cq *ib_create_cq(struct ib_device *device, 1571 ib_comp_handler comp_handler, 1572 void (*event_handler)(struct ib_event *, void *), 1573 void *cq_context, 1574 const struct ib_cq_init_attr *cq_attr) 1575 { 1576 struct ib_cq *cq; 1577 1578 cq = device->create_cq(device, cq_attr, NULL, NULL); 1579 1580 if (!IS_ERR(cq)) { 1581 cq->device = device; 1582 cq->uobject = NULL; 1583 cq->comp_handler = comp_handler; 1584 cq->event_handler = event_handler; 1585 cq->cq_context = cq_context; 1586 atomic_set(&cq->usecnt, 0); 1587 cq->res.type = RDMA_RESTRACK_CQ; 1588 rdma_restrack_add(&cq->res); 1589 } 1590 1591 return cq; 1592 } 1593 EXPORT_SYMBOL(ib_create_cq); 1594 1595 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period) 1596 { 1597 return cq->device->modify_cq ? 1598 cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS; 1599 } 1600 EXPORT_SYMBOL(rdma_set_cq_moderation); 1601 1602 int ib_destroy_cq(struct ib_cq *cq) 1603 { 1604 if (atomic_read(&cq->usecnt)) 1605 return -EBUSY; 1606 1607 rdma_restrack_del(&cq->res); 1608 return cq->device->destroy_cq(cq); 1609 } 1610 EXPORT_SYMBOL(ib_destroy_cq); 1611 1612 int ib_resize_cq(struct ib_cq *cq, int cqe) 1613 { 1614 return cq->device->resize_cq ? 1615 cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS; 1616 } 1617 EXPORT_SYMBOL(ib_resize_cq); 1618 1619 /* Memory regions */ 1620 1621 int ib_dereg_mr(struct ib_mr *mr) 1622 { 1623 struct ib_pd *pd = mr->pd; 1624 int ret; 1625 1626 ret = mr->device->dereg_mr(mr); 1627 if (!ret) 1628 atomic_dec(&pd->usecnt); 1629 1630 return ret; 1631 } 1632 EXPORT_SYMBOL(ib_dereg_mr); 1633 1634 /** 1635 * ib_alloc_mr() - Allocates a memory region 1636 * @pd: protection domain associated with the region 1637 * @mr_type: memory region type 1638 * @max_num_sg: maximum sg entries available for registration. 1639 * 1640 * Notes: 1641 * Memory registeration page/sg lists must not exceed max_num_sg. 1642 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed 1643 * max_num_sg * used_page_size. 1644 * 1645 */ 1646 struct ib_mr *ib_alloc_mr(struct ib_pd *pd, 1647 enum ib_mr_type mr_type, 1648 u32 max_num_sg) 1649 { 1650 struct ib_mr *mr; 1651 1652 if (!pd->device->alloc_mr) 1653 return ERR_PTR(-ENOSYS); 1654 1655 mr = pd->device->alloc_mr(pd, mr_type, max_num_sg); 1656 if (!IS_ERR(mr)) { 1657 mr->device = pd->device; 1658 mr->pd = pd; 1659 mr->uobject = NULL; 1660 atomic_inc(&pd->usecnt); 1661 mr->need_inval = false; 1662 } 1663 1664 return mr; 1665 } 1666 EXPORT_SYMBOL(ib_alloc_mr); 1667 1668 /* "Fast" memory regions */ 1669 1670 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd, 1671 int mr_access_flags, 1672 struct ib_fmr_attr *fmr_attr) 1673 { 1674 struct ib_fmr *fmr; 1675 1676 if (!pd->device->alloc_fmr) 1677 return ERR_PTR(-ENOSYS); 1678 1679 fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr); 1680 if (!IS_ERR(fmr)) { 1681 fmr->device = pd->device; 1682 fmr->pd = pd; 1683 atomic_inc(&pd->usecnt); 1684 } 1685 1686 return fmr; 1687 } 1688 EXPORT_SYMBOL(ib_alloc_fmr); 1689 1690 int ib_unmap_fmr(struct list_head *fmr_list) 1691 { 1692 struct ib_fmr *fmr; 1693 1694 if (list_empty(fmr_list)) 1695 return 0; 1696 1697 fmr = list_entry(fmr_list->next, struct ib_fmr, list); 1698 return fmr->device->unmap_fmr(fmr_list); 1699 } 1700 EXPORT_SYMBOL(ib_unmap_fmr); 1701 1702 int ib_dealloc_fmr(struct ib_fmr *fmr) 1703 { 1704 struct ib_pd *pd; 1705 int ret; 1706 1707 pd = fmr->pd; 1708 ret = fmr->device->dealloc_fmr(fmr); 1709 if (!ret) 1710 atomic_dec(&pd->usecnt); 1711 1712 return ret; 1713 } 1714 EXPORT_SYMBOL(ib_dealloc_fmr); 1715 1716 /* Multicast groups */ 1717 1718 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid) 1719 { 1720 struct ib_qp_init_attr init_attr = {}; 1721 struct ib_qp_attr attr = {}; 1722 int num_eth_ports = 0; 1723 int port; 1724 1725 /* If QP state >= init, it is assigned to a port and we can check this 1726 * port only. 1727 */ 1728 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) { 1729 if (attr.qp_state >= IB_QPS_INIT) { 1730 if (rdma_port_get_link_layer(qp->device, attr.port_num) != 1731 IB_LINK_LAYER_INFINIBAND) 1732 return true; 1733 goto lid_check; 1734 } 1735 } 1736 1737 /* Can't get a quick answer, iterate over all ports */ 1738 for (port = 0; port < qp->device->phys_port_cnt; port++) 1739 if (rdma_port_get_link_layer(qp->device, port) != 1740 IB_LINK_LAYER_INFINIBAND) 1741 num_eth_ports++; 1742 1743 /* If we have at lease one Ethernet port, RoCE annex declares that 1744 * multicast LID should be ignored. We can't tell at this step if the 1745 * QP belongs to an IB or Ethernet port. 1746 */ 1747 if (num_eth_ports) 1748 return true; 1749 1750 /* If all the ports are IB, we can check according to IB spec. */ 1751 lid_check: 1752 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) || 1753 lid == be16_to_cpu(IB_LID_PERMISSIVE)); 1754 } 1755 1756 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid) 1757 { 1758 int ret; 1759 1760 if (!qp->device->attach_mcast) 1761 return -ENOSYS; 1762 1763 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) || 1764 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid)) 1765 return -EINVAL; 1766 1767 ret = qp->device->attach_mcast(qp, gid, lid); 1768 if (!ret) 1769 atomic_inc(&qp->usecnt); 1770 return ret; 1771 } 1772 EXPORT_SYMBOL(ib_attach_mcast); 1773 1774 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid) 1775 { 1776 int ret; 1777 1778 if (!qp->device->detach_mcast) 1779 return -ENOSYS; 1780 1781 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) || 1782 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid)) 1783 return -EINVAL; 1784 1785 ret = qp->device->detach_mcast(qp, gid, lid); 1786 if (!ret) 1787 atomic_dec(&qp->usecnt); 1788 return ret; 1789 } 1790 EXPORT_SYMBOL(ib_detach_mcast); 1791 1792 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller) 1793 { 1794 struct ib_xrcd *xrcd; 1795 1796 if (!device->alloc_xrcd) 1797 return ERR_PTR(-ENOSYS); 1798 1799 xrcd = device->alloc_xrcd(device, NULL, NULL); 1800 if (!IS_ERR(xrcd)) { 1801 xrcd->device = device; 1802 xrcd->inode = NULL; 1803 atomic_set(&xrcd->usecnt, 0); 1804 mutex_init(&xrcd->tgt_qp_mutex); 1805 INIT_LIST_HEAD(&xrcd->tgt_qp_list); 1806 } 1807 1808 return xrcd; 1809 } 1810 EXPORT_SYMBOL(__ib_alloc_xrcd); 1811 1812 int ib_dealloc_xrcd(struct ib_xrcd *xrcd) 1813 { 1814 struct ib_qp *qp; 1815 int ret; 1816 1817 if (atomic_read(&xrcd->usecnt)) 1818 return -EBUSY; 1819 1820 while (!list_empty(&xrcd->tgt_qp_list)) { 1821 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list); 1822 ret = ib_destroy_qp(qp); 1823 if (ret) 1824 return ret; 1825 } 1826 1827 return xrcd->device->dealloc_xrcd(xrcd); 1828 } 1829 EXPORT_SYMBOL(ib_dealloc_xrcd); 1830 1831 /** 1832 * ib_create_wq - Creates a WQ associated with the specified protection 1833 * domain. 1834 * @pd: The protection domain associated with the WQ. 1835 * @wq_attr: A list of initial attributes required to create the 1836 * WQ. If WQ creation succeeds, then the attributes are updated to 1837 * the actual capabilities of the created WQ. 1838 * 1839 * wq_attr->max_wr and wq_attr->max_sge determine 1840 * the requested size of the WQ, and set to the actual values allocated 1841 * on return. 1842 * If ib_create_wq() succeeds, then max_wr and max_sge will always be 1843 * at least as large as the requested values. 1844 */ 1845 struct ib_wq *ib_create_wq(struct ib_pd *pd, 1846 struct ib_wq_init_attr *wq_attr) 1847 { 1848 struct ib_wq *wq; 1849 1850 if (!pd->device->create_wq) 1851 return ERR_PTR(-ENOSYS); 1852 1853 wq = pd->device->create_wq(pd, wq_attr, NULL); 1854 if (!IS_ERR(wq)) { 1855 wq->event_handler = wq_attr->event_handler; 1856 wq->wq_context = wq_attr->wq_context; 1857 wq->wq_type = wq_attr->wq_type; 1858 wq->cq = wq_attr->cq; 1859 wq->device = pd->device; 1860 wq->pd = pd; 1861 wq->uobject = NULL; 1862 atomic_inc(&pd->usecnt); 1863 atomic_inc(&wq_attr->cq->usecnt); 1864 atomic_set(&wq->usecnt, 0); 1865 } 1866 return wq; 1867 } 1868 EXPORT_SYMBOL(ib_create_wq); 1869 1870 /** 1871 * ib_destroy_wq - Destroys the specified WQ. 1872 * @wq: The WQ to destroy. 1873 */ 1874 int ib_destroy_wq(struct ib_wq *wq) 1875 { 1876 int err; 1877 struct ib_cq *cq = wq->cq; 1878 struct ib_pd *pd = wq->pd; 1879 1880 if (atomic_read(&wq->usecnt)) 1881 return -EBUSY; 1882 1883 err = wq->device->destroy_wq(wq); 1884 if (!err) { 1885 atomic_dec(&pd->usecnt); 1886 atomic_dec(&cq->usecnt); 1887 } 1888 return err; 1889 } 1890 EXPORT_SYMBOL(ib_destroy_wq); 1891 1892 /** 1893 * ib_modify_wq - Modifies the specified WQ. 1894 * @wq: The WQ to modify. 1895 * @wq_attr: On input, specifies the WQ attributes to modify. 1896 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ 1897 * are being modified. 1898 * On output, the current values of selected WQ attributes are returned. 1899 */ 1900 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr, 1901 u32 wq_attr_mask) 1902 { 1903 int err; 1904 1905 if (!wq->device->modify_wq) 1906 return -ENOSYS; 1907 1908 err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL); 1909 return err; 1910 } 1911 EXPORT_SYMBOL(ib_modify_wq); 1912 1913 /* 1914 * ib_create_rwq_ind_table - Creates a RQ Indirection Table. 1915 * @device: The device on which to create the rwq indirection table. 1916 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to 1917 * create the Indirection Table. 1918 * 1919 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less 1920 * than the created ib_rwq_ind_table object and the caller is responsible 1921 * for its memory allocation/free. 1922 */ 1923 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device, 1924 struct ib_rwq_ind_table_init_attr *init_attr) 1925 { 1926 struct ib_rwq_ind_table *rwq_ind_table; 1927 int i; 1928 u32 table_size; 1929 1930 if (!device->create_rwq_ind_table) 1931 return ERR_PTR(-ENOSYS); 1932 1933 table_size = (1 << init_attr->log_ind_tbl_size); 1934 rwq_ind_table = device->create_rwq_ind_table(device, 1935 init_attr, NULL); 1936 if (IS_ERR(rwq_ind_table)) 1937 return rwq_ind_table; 1938 1939 rwq_ind_table->ind_tbl = init_attr->ind_tbl; 1940 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size; 1941 rwq_ind_table->device = device; 1942 rwq_ind_table->uobject = NULL; 1943 atomic_set(&rwq_ind_table->usecnt, 0); 1944 1945 for (i = 0; i < table_size; i++) 1946 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt); 1947 1948 return rwq_ind_table; 1949 } 1950 EXPORT_SYMBOL(ib_create_rwq_ind_table); 1951 1952 /* 1953 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table. 1954 * @wq_ind_table: The Indirection Table to destroy. 1955 */ 1956 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table) 1957 { 1958 int err, i; 1959 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size); 1960 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl; 1961 1962 if (atomic_read(&rwq_ind_table->usecnt)) 1963 return -EBUSY; 1964 1965 err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table); 1966 if (!err) { 1967 for (i = 0; i < table_size; i++) 1968 atomic_dec(&ind_tbl[i]->usecnt); 1969 } 1970 1971 return err; 1972 } 1973 EXPORT_SYMBOL(ib_destroy_rwq_ind_table); 1974 1975 struct ib_flow *ib_create_flow(struct ib_qp *qp, 1976 struct ib_flow_attr *flow_attr, 1977 int domain) 1978 { 1979 struct ib_flow *flow_id; 1980 if (!qp->device->create_flow) 1981 return ERR_PTR(-ENOSYS); 1982 1983 flow_id = qp->device->create_flow(qp, flow_attr, domain); 1984 if (!IS_ERR(flow_id)) { 1985 atomic_inc(&qp->usecnt); 1986 flow_id->qp = qp; 1987 } 1988 return flow_id; 1989 } 1990 EXPORT_SYMBOL(ib_create_flow); 1991 1992 int ib_destroy_flow(struct ib_flow *flow_id) 1993 { 1994 int err; 1995 struct ib_qp *qp = flow_id->qp; 1996 1997 err = qp->device->destroy_flow(flow_id); 1998 if (!err) 1999 atomic_dec(&qp->usecnt); 2000 return err; 2001 } 2002 EXPORT_SYMBOL(ib_destroy_flow); 2003 2004 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask, 2005 struct ib_mr_status *mr_status) 2006 { 2007 return mr->device->check_mr_status ? 2008 mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS; 2009 } 2010 EXPORT_SYMBOL(ib_check_mr_status); 2011 2012 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port, 2013 int state) 2014 { 2015 if (!device->set_vf_link_state) 2016 return -ENOSYS; 2017 2018 return device->set_vf_link_state(device, vf, port, state); 2019 } 2020 EXPORT_SYMBOL(ib_set_vf_link_state); 2021 2022 int ib_get_vf_config(struct ib_device *device, int vf, u8 port, 2023 struct ifla_vf_info *info) 2024 { 2025 if (!device->get_vf_config) 2026 return -ENOSYS; 2027 2028 return device->get_vf_config(device, vf, port, info); 2029 } 2030 EXPORT_SYMBOL(ib_get_vf_config); 2031 2032 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port, 2033 struct ifla_vf_stats *stats) 2034 { 2035 if (!device->get_vf_stats) 2036 return -ENOSYS; 2037 2038 return device->get_vf_stats(device, vf, port, stats); 2039 } 2040 EXPORT_SYMBOL(ib_get_vf_stats); 2041 2042 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid, 2043 int type) 2044 { 2045 if (!device->set_vf_guid) 2046 return -ENOSYS; 2047 2048 return device->set_vf_guid(device, vf, port, guid, type); 2049 } 2050 EXPORT_SYMBOL(ib_set_vf_guid); 2051 2052 /** 2053 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list 2054 * and set it the memory region. 2055 * @mr: memory region 2056 * @sg: dma mapped scatterlist 2057 * @sg_nents: number of entries in sg 2058 * @sg_offset: offset in bytes into sg 2059 * @page_size: page vector desired page size 2060 * 2061 * Constraints: 2062 * - The first sg element is allowed to have an offset. 2063 * - Each sg element must either be aligned to page_size or virtually 2064 * contiguous to the previous element. In case an sg element has a 2065 * non-contiguous offset, the mapping prefix will not include it. 2066 * - The last sg element is allowed to have length less than page_size. 2067 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size 2068 * then only max_num_sg entries will be mapped. 2069 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these 2070 * constraints holds and the page_size argument is ignored. 2071 * 2072 * Returns the number of sg elements that were mapped to the memory region. 2073 * 2074 * After this completes successfully, the memory region 2075 * is ready for registration. 2076 */ 2077 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents, 2078 unsigned int *sg_offset, unsigned int page_size) 2079 { 2080 if (unlikely(!mr->device->map_mr_sg)) 2081 return -ENOSYS; 2082 2083 mr->page_size = page_size; 2084 2085 return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset); 2086 } 2087 EXPORT_SYMBOL(ib_map_mr_sg); 2088 2089 /** 2090 * ib_sg_to_pages() - Convert the largest prefix of a sg list 2091 * to a page vector 2092 * @mr: memory region 2093 * @sgl: dma mapped scatterlist 2094 * @sg_nents: number of entries in sg 2095 * @sg_offset_p: IN: start offset in bytes into sg 2096 * OUT: offset in bytes for element n of the sg of the first 2097 * byte that has not been processed where n is the return 2098 * value of this function. 2099 * @set_page: driver page assignment function pointer 2100 * 2101 * Core service helper for drivers to convert the largest 2102 * prefix of given sg list to a page vector. The sg list 2103 * prefix converted is the prefix that meet the requirements 2104 * of ib_map_mr_sg. 2105 * 2106 * Returns the number of sg elements that were assigned to 2107 * a page vector. 2108 */ 2109 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents, 2110 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64)) 2111 { 2112 struct scatterlist *sg; 2113 u64 last_end_dma_addr = 0; 2114 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0; 2115 unsigned int last_page_off = 0; 2116 u64 page_mask = ~((u64)mr->page_size - 1); 2117 int i, ret; 2118 2119 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0]))) 2120 return -EINVAL; 2121 2122 mr->iova = sg_dma_address(&sgl[0]) + sg_offset; 2123 mr->length = 0; 2124 2125 for_each_sg(sgl, sg, sg_nents, i) { 2126 u64 dma_addr = sg_dma_address(sg) + sg_offset; 2127 u64 prev_addr = dma_addr; 2128 unsigned int dma_len = sg_dma_len(sg) - sg_offset; 2129 u64 end_dma_addr = dma_addr + dma_len; 2130 u64 page_addr = dma_addr & page_mask; 2131 2132 /* 2133 * For the second and later elements, check whether either the 2134 * end of element i-1 or the start of element i is not aligned 2135 * on a page boundary. 2136 */ 2137 if (i && (last_page_off != 0 || page_addr != dma_addr)) { 2138 /* Stop mapping if there is a gap. */ 2139 if (last_end_dma_addr != dma_addr) 2140 break; 2141 2142 /* 2143 * Coalesce this element with the last. If it is small 2144 * enough just update mr->length. Otherwise start 2145 * mapping from the next page. 2146 */ 2147 goto next_page; 2148 } 2149 2150 do { 2151 ret = set_page(mr, page_addr); 2152 if (unlikely(ret < 0)) { 2153 sg_offset = prev_addr - sg_dma_address(sg); 2154 mr->length += prev_addr - dma_addr; 2155 if (sg_offset_p) 2156 *sg_offset_p = sg_offset; 2157 return i || sg_offset ? i : ret; 2158 } 2159 prev_addr = page_addr; 2160 next_page: 2161 page_addr += mr->page_size; 2162 } while (page_addr < end_dma_addr); 2163 2164 mr->length += dma_len; 2165 last_end_dma_addr = end_dma_addr; 2166 last_page_off = end_dma_addr & ~page_mask; 2167 2168 sg_offset = 0; 2169 } 2170 2171 if (sg_offset_p) 2172 *sg_offset_p = 0; 2173 return i; 2174 } 2175 EXPORT_SYMBOL(ib_sg_to_pages); 2176 2177 struct ib_drain_cqe { 2178 struct ib_cqe cqe; 2179 struct completion done; 2180 }; 2181 2182 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc) 2183 { 2184 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe, 2185 cqe); 2186 2187 complete(&cqe->done); 2188 } 2189 2190 /* 2191 * Post a WR and block until its completion is reaped for the SQ. 2192 */ 2193 static void __ib_drain_sq(struct ib_qp *qp) 2194 { 2195 struct ib_cq *cq = qp->send_cq; 2196 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 2197 struct ib_drain_cqe sdrain; 2198 struct ib_send_wr swr = {}, *bad_swr; 2199 int ret; 2200 2201 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 2202 if (ret) { 2203 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 2204 return; 2205 } 2206 2207 swr.wr_cqe = &sdrain.cqe; 2208 sdrain.cqe.done = ib_drain_qp_done; 2209 init_completion(&sdrain.done); 2210 2211 ret = ib_post_send(qp, &swr, &bad_swr); 2212 if (ret) { 2213 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 2214 return; 2215 } 2216 2217 if (cq->poll_ctx == IB_POLL_DIRECT) 2218 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0) 2219 ib_process_cq_direct(cq, -1); 2220 else 2221 wait_for_completion(&sdrain.done); 2222 } 2223 2224 /* 2225 * Post a WR and block until its completion is reaped for the RQ. 2226 */ 2227 static void __ib_drain_rq(struct ib_qp *qp) 2228 { 2229 struct ib_cq *cq = qp->recv_cq; 2230 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 2231 struct ib_drain_cqe rdrain; 2232 struct ib_recv_wr rwr = {}, *bad_rwr; 2233 int ret; 2234 2235 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 2236 if (ret) { 2237 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 2238 return; 2239 } 2240 2241 rwr.wr_cqe = &rdrain.cqe; 2242 rdrain.cqe.done = ib_drain_qp_done; 2243 init_completion(&rdrain.done); 2244 2245 ret = ib_post_recv(qp, &rwr, &bad_rwr); 2246 if (ret) { 2247 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 2248 return; 2249 } 2250 2251 if (cq->poll_ctx == IB_POLL_DIRECT) 2252 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0) 2253 ib_process_cq_direct(cq, -1); 2254 else 2255 wait_for_completion(&rdrain.done); 2256 } 2257 2258 /** 2259 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the 2260 * application. 2261 * @qp: queue pair to drain 2262 * 2263 * If the device has a provider-specific drain function, then 2264 * call that. Otherwise call the generic drain function 2265 * __ib_drain_sq(). 2266 * 2267 * The caller must: 2268 * 2269 * ensure there is room in the CQ and SQ for the drain work request and 2270 * completion. 2271 * 2272 * allocate the CQ using ib_alloc_cq(). 2273 * 2274 * ensure that there are no other contexts that are posting WRs concurrently. 2275 * Otherwise the drain is not guaranteed. 2276 */ 2277 void ib_drain_sq(struct ib_qp *qp) 2278 { 2279 if (qp->device->drain_sq) 2280 qp->device->drain_sq(qp); 2281 else 2282 __ib_drain_sq(qp); 2283 } 2284 EXPORT_SYMBOL(ib_drain_sq); 2285 2286 /** 2287 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the 2288 * application. 2289 * @qp: queue pair to drain 2290 * 2291 * If the device has a provider-specific drain function, then 2292 * call that. Otherwise call the generic drain function 2293 * __ib_drain_rq(). 2294 * 2295 * The caller must: 2296 * 2297 * ensure there is room in the CQ and RQ for the drain work request and 2298 * completion. 2299 * 2300 * allocate the CQ using ib_alloc_cq(). 2301 * 2302 * ensure that there are no other contexts that are posting WRs concurrently. 2303 * Otherwise the drain is not guaranteed. 2304 */ 2305 void ib_drain_rq(struct ib_qp *qp) 2306 { 2307 if (qp->device->drain_rq) 2308 qp->device->drain_rq(qp); 2309 else 2310 __ib_drain_rq(qp); 2311 } 2312 EXPORT_SYMBOL(ib_drain_rq); 2313 2314 /** 2315 * ib_drain_qp() - Block until all CQEs have been consumed by the 2316 * application on both the RQ and SQ. 2317 * @qp: queue pair to drain 2318 * 2319 * The caller must: 2320 * 2321 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests 2322 * and completions. 2323 * 2324 * allocate the CQs using ib_alloc_cq(). 2325 * 2326 * ensure that there are no other contexts that are posting WRs concurrently. 2327 * Otherwise the drain is not guaranteed. 2328 */ 2329 void ib_drain_qp(struct ib_qp *qp) 2330 { 2331 ib_drain_sq(qp); 2332 if (!qp->srq) 2333 ib_drain_rq(qp); 2334 } 2335 EXPORT_SYMBOL(ib_drain_qp); 2336