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/string.h> 42 #include <linux/slab.h> 43 #include <linux/in.h> 44 #include <linux/in6.h> 45 46 #include <rdma/ib_verbs.h> 47 #include <rdma/ib_cache.h> 48 #include <rdma/ib_addr.h> 49 50 #include <netinet/ip.h> 51 #include <netinet/ip6.h> 52 53 #include <machine/in_cksum.h> 54 55 #include "core_priv.h" 56 57 static const char * const ib_events[] = { 58 [IB_EVENT_CQ_ERR] = "CQ error", 59 [IB_EVENT_QP_FATAL] = "QP fatal error", 60 [IB_EVENT_QP_REQ_ERR] = "QP request error", 61 [IB_EVENT_QP_ACCESS_ERR] = "QP access error", 62 [IB_EVENT_COMM_EST] = "communication established", 63 [IB_EVENT_SQ_DRAINED] = "send queue drained", 64 [IB_EVENT_PATH_MIG] = "path migration successful", 65 [IB_EVENT_PATH_MIG_ERR] = "path migration error", 66 [IB_EVENT_DEVICE_FATAL] = "device fatal error", 67 [IB_EVENT_PORT_ACTIVE] = "port active", 68 [IB_EVENT_PORT_ERR] = "port error", 69 [IB_EVENT_LID_CHANGE] = "LID change", 70 [IB_EVENT_PKEY_CHANGE] = "P_key change", 71 [IB_EVENT_SM_CHANGE] = "SM change", 72 [IB_EVENT_SRQ_ERR] = "SRQ error", 73 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached", 74 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached", 75 [IB_EVENT_CLIENT_REREGISTER] = "client reregister", 76 [IB_EVENT_GID_CHANGE] = "GID changed", 77 }; 78 79 const char *__attribute_const__ ib_event_msg(enum ib_event_type event) 80 { 81 size_t index = event; 82 83 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ? 84 ib_events[index] : "unrecognized event"; 85 } 86 EXPORT_SYMBOL(ib_event_msg); 87 88 static const char * const wc_statuses[] = { 89 [IB_WC_SUCCESS] = "success", 90 [IB_WC_LOC_LEN_ERR] = "local length error", 91 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error", 92 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error", 93 [IB_WC_LOC_PROT_ERR] = "local protection error", 94 [IB_WC_WR_FLUSH_ERR] = "WR flushed", 95 [IB_WC_MW_BIND_ERR] = "memory management operation error", 96 [IB_WC_BAD_RESP_ERR] = "bad response error", 97 [IB_WC_LOC_ACCESS_ERR] = "local access error", 98 [IB_WC_REM_INV_REQ_ERR] = "invalid request error", 99 [IB_WC_REM_ACCESS_ERR] = "remote access error", 100 [IB_WC_REM_OP_ERR] = "remote operation error", 101 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded", 102 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded", 103 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error", 104 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request", 105 [IB_WC_REM_ABORT_ERR] = "operation aborted", 106 [IB_WC_INV_EECN_ERR] = "invalid EE context number", 107 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state", 108 [IB_WC_FATAL_ERR] = "fatal error", 109 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error", 110 [IB_WC_GENERAL_ERR] = "general error", 111 }; 112 113 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status) 114 { 115 size_t index = status; 116 117 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ? 118 wc_statuses[index] : "unrecognized status"; 119 } 120 EXPORT_SYMBOL(ib_wc_status_msg); 121 122 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate) 123 { 124 switch (rate) { 125 case IB_RATE_2_5_GBPS: return 1; 126 case IB_RATE_5_GBPS: return 2; 127 case IB_RATE_10_GBPS: return 4; 128 case IB_RATE_20_GBPS: return 8; 129 case IB_RATE_30_GBPS: return 12; 130 case IB_RATE_40_GBPS: return 16; 131 case IB_RATE_60_GBPS: return 24; 132 case IB_RATE_80_GBPS: return 32; 133 case IB_RATE_120_GBPS: return 48; 134 default: return -1; 135 } 136 } 137 EXPORT_SYMBOL(ib_rate_to_mult); 138 139 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult) 140 { 141 switch (mult) { 142 case 1: return IB_RATE_2_5_GBPS; 143 case 2: return IB_RATE_5_GBPS; 144 case 4: return IB_RATE_10_GBPS; 145 case 8: return IB_RATE_20_GBPS; 146 case 12: return IB_RATE_30_GBPS; 147 case 16: return IB_RATE_40_GBPS; 148 case 24: return IB_RATE_60_GBPS; 149 case 32: return IB_RATE_80_GBPS; 150 case 48: return IB_RATE_120_GBPS; 151 default: return IB_RATE_PORT_CURRENT; 152 } 153 } 154 EXPORT_SYMBOL(mult_to_ib_rate); 155 156 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate) 157 { 158 switch (rate) { 159 case IB_RATE_2_5_GBPS: return 2500; 160 case IB_RATE_5_GBPS: return 5000; 161 case IB_RATE_10_GBPS: return 10000; 162 case IB_RATE_20_GBPS: return 20000; 163 case IB_RATE_30_GBPS: return 30000; 164 case IB_RATE_40_GBPS: return 40000; 165 case IB_RATE_60_GBPS: return 60000; 166 case IB_RATE_80_GBPS: return 80000; 167 case IB_RATE_120_GBPS: return 120000; 168 case IB_RATE_14_GBPS: return 14062; 169 case IB_RATE_56_GBPS: return 56250; 170 case IB_RATE_112_GBPS: return 112500; 171 case IB_RATE_168_GBPS: return 168750; 172 case IB_RATE_25_GBPS: return 25781; 173 case IB_RATE_100_GBPS: return 103125; 174 case IB_RATE_200_GBPS: return 206250; 175 case IB_RATE_300_GBPS: return 309375; 176 default: return -1; 177 } 178 } 179 EXPORT_SYMBOL(ib_rate_to_mbps); 180 181 __attribute_const__ enum rdma_transport_type 182 rdma_node_get_transport(enum rdma_node_type node_type) 183 { 184 switch (node_type) { 185 case RDMA_NODE_IB_CA: 186 case RDMA_NODE_IB_SWITCH: 187 case RDMA_NODE_IB_ROUTER: 188 return RDMA_TRANSPORT_IB; 189 case RDMA_NODE_RNIC: 190 return RDMA_TRANSPORT_IWARP; 191 case RDMA_NODE_USNIC: 192 return RDMA_TRANSPORT_USNIC; 193 case RDMA_NODE_USNIC_UDP: 194 return RDMA_TRANSPORT_USNIC_UDP; 195 default: 196 BUG(); 197 return 0; 198 } 199 } 200 EXPORT_SYMBOL(rdma_node_get_transport); 201 202 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num) 203 { 204 if (device->get_link_layer) 205 return device->get_link_layer(device, port_num); 206 207 switch (rdma_node_get_transport(device->node_type)) { 208 case RDMA_TRANSPORT_IB: 209 return IB_LINK_LAYER_INFINIBAND; 210 case RDMA_TRANSPORT_IWARP: 211 case RDMA_TRANSPORT_USNIC: 212 case RDMA_TRANSPORT_USNIC_UDP: 213 return IB_LINK_LAYER_ETHERNET; 214 default: 215 return IB_LINK_LAYER_UNSPECIFIED; 216 } 217 } 218 EXPORT_SYMBOL(rdma_port_get_link_layer); 219 220 /* Protection domains */ 221 222 /** 223 * ib_alloc_pd - Allocates an unused protection domain. 224 * @device: The device on which to allocate the protection domain. 225 * 226 * A protection domain object provides an association between QPs, shared 227 * receive queues, address handles, memory regions, and memory windows. 228 * 229 * Every PD has a local_dma_lkey which can be used as the lkey value for local 230 * memory operations. 231 */ 232 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags, 233 const char *caller) 234 { 235 struct ib_pd *pd; 236 int mr_access_flags = 0; 237 238 pd = device->alloc_pd(device, NULL, NULL); 239 if (IS_ERR(pd)) 240 return pd; 241 242 pd->device = device; 243 pd->uobject = NULL; 244 pd->__internal_mr = NULL; 245 atomic_set(&pd->usecnt, 0); 246 pd->flags = flags; 247 248 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY) 249 pd->local_dma_lkey = device->local_dma_lkey; 250 else 251 mr_access_flags |= IB_ACCESS_LOCAL_WRITE; 252 253 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) { 254 pr_warn("%s: enabling unsafe global rkey\n", caller); 255 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE; 256 } 257 258 if (mr_access_flags) { 259 struct ib_mr *mr; 260 261 mr = pd->device->get_dma_mr(pd, mr_access_flags); 262 if (IS_ERR(mr)) { 263 ib_dealloc_pd(pd); 264 return ERR_CAST(mr); 265 } 266 267 mr->device = pd->device; 268 mr->pd = pd; 269 mr->uobject = NULL; 270 mr->need_inval = false; 271 272 pd->__internal_mr = mr; 273 274 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) 275 pd->local_dma_lkey = pd->__internal_mr->lkey; 276 277 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) 278 pd->unsafe_global_rkey = pd->__internal_mr->rkey; 279 } 280 281 return pd; 282 } 283 EXPORT_SYMBOL(__ib_alloc_pd); 284 285 /** 286 * ib_dealloc_pd - Deallocates a protection domain. 287 * @pd: The protection domain to deallocate. 288 * 289 * It is an error to call this function while any resources in the pd still 290 * exist. The caller is responsible to synchronously destroy them and 291 * guarantee no new allocations will happen. 292 */ 293 void ib_dealloc_pd(struct ib_pd *pd) 294 { 295 int ret; 296 297 if (pd->__internal_mr) { 298 ret = pd->device->dereg_mr(pd->__internal_mr); 299 WARN_ON(ret); 300 pd->__internal_mr = NULL; 301 } 302 303 /* uverbs manipulates usecnt with proper locking, while the kabi 304 requires the caller to guarantee we can't race here. */ 305 WARN_ON(atomic_read(&pd->usecnt)); 306 307 /* Making delalloc_pd a void return is a WIP, no driver should return 308 an error here. */ 309 ret = pd->device->dealloc_pd(pd); 310 WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd"); 311 } 312 EXPORT_SYMBOL(ib_dealloc_pd); 313 314 /* Address handles */ 315 316 struct ib_ah *ib_create_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr) 317 { 318 struct ib_ah *ah; 319 320 ah = pd->device->create_ah(pd, ah_attr); 321 322 if (!IS_ERR(ah)) { 323 ah->device = pd->device; 324 ah->pd = pd; 325 ah->uobject = NULL; 326 atomic_inc(&pd->usecnt); 327 } 328 329 return ah; 330 } 331 EXPORT_SYMBOL(ib_create_ah); 332 333 static int ib_get_header_version(const union rdma_network_hdr *hdr) 334 { 335 const struct ip *ip4h = (const struct ip *)&hdr->roce4grh; 336 struct ip ip4h_checked; 337 const struct ip6_hdr *ip6h = (const struct ip6_hdr *)&hdr->ibgrh; 338 339 /* If it's IPv6, the version must be 6, otherwise, the first 340 * 20 bytes (before the IPv4 header) are garbled. 341 */ 342 if ((ip6h->ip6_vfc & IPV6_VERSION_MASK) != IPV6_VERSION) 343 return (ip4h->ip_v == 4) ? 4 : 0; 344 /* version may be 6 or 4 because the first 20 bytes could be garbled */ 345 346 /* RoCE v2 requires no options, thus header length 347 * must be 5 words 348 */ 349 if (ip4h->ip_hl != 5) 350 return 6; 351 352 /* Verify checksum. 353 * We can't write on scattered buffers so we need to copy to 354 * temp buffer. 355 */ 356 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked)); 357 ip4h_checked.ip_sum = 0; 358 ip4h_checked.ip_sum = in_cksum_hdr(&ip4h_checked); 359 /* if IPv4 header checksum is OK, believe it */ 360 if (ip4h->ip_sum == ip4h_checked.ip_sum) 361 return 4; 362 return 6; 363 } 364 365 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device, 366 u8 port_num, 367 const struct ib_grh *grh) 368 { 369 int grh_version; 370 371 if (rdma_protocol_ib(device, port_num)) 372 return RDMA_NETWORK_IB; 373 374 grh_version = ib_get_header_version((const union rdma_network_hdr *)grh); 375 376 if (grh_version == 4) 377 return RDMA_NETWORK_IPV4; 378 379 if (grh->next_hdr == IPPROTO_UDP) 380 return RDMA_NETWORK_IPV6; 381 382 return RDMA_NETWORK_ROCE_V1; 383 } 384 385 struct find_gid_index_context { 386 u16 vlan_id; 387 enum ib_gid_type gid_type; 388 }; 389 390 static bool find_gid_index(const union ib_gid *gid, 391 const struct ib_gid_attr *gid_attr, 392 void *context) 393 { 394 struct find_gid_index_context *ctx = 395 (struct find_gid_index_context *)context; 396 397 if (ctx->gid_type != gid_attr->gid_type) 398 return false; 399 400 if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) || 401 (is_vlan_dev(gid_attr->ndev) && 402 vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id)) 403 return false; 404 405 return true; 406 } 407 408 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num, 409 u16 vlan_id, const union ib_gid *sgid, 410 enum ib_gid_type gid_type, 411 u16 *gid_index) 412 { 413 struct find_gid_index_context context = {.vlan_id = vlan_id, 414 .gid_type = gid_type}; 415 416 return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index, 417 &context, gid_index); 418 } 419 420 static int get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr, 421 enum rdma_network_type net_type, 422 union ib_gid *sgid, union ib_gid *dgid) 423 { 424 struct sockaddr_in src_in; 425 struct sockaddr_in dst_in; 426 __be32 src_saddr, dst_saddr; 427 428 if (!sgid || !dgid) 429 return -EINVAL; 430 431 if (net_type == RDMA_NETWORK_IPV4) { 432 memcpy(&src_in.sin_addr.s_addr, 433 &hdr->roce4grh.ip_src, 4); 434 memcpy(&dst_in.sin_addr.s_addr, 435 &hdr->roce4grh.ip_dst, 4); 436 src_saddr = src_in.sin_addr.s_addr; 437 dst_saddr = dst_in.sin_addr.s_addr; 438 ipv6_addr_set_v4mapped(src_saddr, 439 (struct in6_addr *)sgid); 440 ipv6_addr_set_v4mapped(dst_saddr, 441 (struct in6_addr *)dgid); 442 return 0; 443 } else if (net_type == RDMA_NETWORK_IPV6 || 444 net_type == RDMA_NETWORK_IB) { 445 *dgid = hdr->ibgrh.dgid; 446 *sgid = hdr->ibgrh.sgid; 447 return 0; 448 } else { 449 return -EINVAL; 450 } 451 } 452 453 int ib_init_ah_from_wc(struct ib_device *device, u8 port_num, 454 const struct ib_wc *wc, const struct ib_grh *grh, 455 struct ib_ah_attr *ah_attr) 456 { 457 u32 flow_class; 458 u16 gid_index; 459 int ret; 460 enum rdma_network_type net_type = RDMA_NETWORK_IB; 461 enum ib_gid_type gid_type = IB_GID_TYPE_IB; 462 int hoplimit = 0xff; 463 union ib_gid dgid; 464 union ib_gid sgid; 465 466 memset(ah_attr, 0, sizeof *ah_attr); 467 if (rdma_cap_eth_ah(device, port_num)) { 468 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE) 469 net_type = wc->network_hdr_type; 470 else 471 net_type = ib_get_net_type_by_grh(device, port_num, grh); 472 gid_type = ib_network_to_gid_type(net_type); 473 } 474 ret = get_gids_from_rdma_hdr((const union rdma_network_hdr *)grh, net_type, 475 &sgid, &dgid); 476 if (ret) 477 return ret; 478 479 if (rdma_protocol_roce(device, port_num)) { 480 int if_index = 0; 481 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ? 482 wc->vlan_id : 0xffff; 483 struct net_device *idev; 484 struct net_device *resolved_dev; 485 486 if (!(wc->wc_flags & IB_WC_GRH)) 487 return -EPROTOTYPE; 488 489 if (!device->get_netdev) 490 return -EOPNOTSUPP; 491 492 idev = device->get_netdev(device, port_num); 493 if (!idev) 494 return -ENODEV; 495 496 ret = rdma_addr_find_l2_eth_by_grh(&dgid, &sgid, 497 ah_attr->dmac, 498 wc->wc_flags & IB_WC_WITH_VLAN ? 499 NULL : &vlan_id, 500 &if_index, &hoplimit); 501 if (ret) { 502 dev_put(idev); 503 return ret; 504 } 505 506 resolved_dev = dev_get_by_index(&init_net, if_index); 507 if (resolved_dev->if_flags & IFF_LOOPBACK) { 508 dev_put(resolved_dev); 509 resolved_dev = idev; 510 dev_hold(resolved_dev); 511 } 512 rcu_read_lock(); 513 if (resolved_dev != idev && !rdma_is_upper_dev_rcu(idev, 514 resolved_dev)) 515 ret = -EHOSTUNREACH; 516 rcu_read_unlock(); 517 dev_put(idev); 518 dev_put(resolved_dev); 519 if (ret) 520 return ret; 521 522 ret = get_sgid_index_from_eth(device, port_num, vlan_id, 523 &dgid, gid_type, &gid_index); 524 if (ret) 525 return ret; 526 } 527 528 ah_attr->dlid = wc->slid; 529 ah_attr->sl = wc->sl; 530 ah_attr->src_path_bits = wc->dlid_path_bits; 531 ah_attr->port_num = port_num; 532 533 if (wc->wc_flags & IB_WC_GRH) { 534 ah_attr->ah_flags = IB_AH_GRH; 535 ah_attr->grh.dgid = sgid; 536 537 if (!rdma_cap_eth_ah(device, port_num)) { 538 if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) { 539 ret = ib_find_cached_gid_by_port(device, &dgid, 540 IB_GID_TYPE_IB, 541 port_num, NULL, 542 &gid_index); 543 if (ret) 544 return ret; 545 } else { 546 gid_index = 0; 547 } 548 } 549 550 ah_attr->grh.sgid_index = (u8) gid_index; 551 flow_class = be32_to_cpu(grh->version_tclass_flow); 552 ah_attr->grh.flow_label = flow_class & 0xFFFFF; 553 ah_attr->grh.hop_limit = hoplimit; 554 ah_attr->grh.traffic_class = (flow_class >> 20) & 0xFF; 555 } 556 return 0; 557 } 558 EXPORT_SYMBOL(ib_init_ah_from_wc); 559 560 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc, 561 const struct ib_grh *grh, u8 port_num) 562 { 563 struct ib_ah_attr ah_attr; 564 int ret; 565 566 ret = ib_init_ah_from_wc(pd->device, port_num, wc, grh, &ah_attr); 567 if (ret) 568 return ERR_PTR(ret); 569 570 return ib_create_ah(pd, &ah_attr); 571 } 572 EXPORT_SYMBOL(ib_create_ah_from_wc); 573 574 int ib_modify_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr) 575 { 576 return ah->device->modify_ah ? 577 ah->device->modify_ah(ah, ah_attr) : 578 -ENOSYS; 579 } 580 EXPORT_SYMBOL(ib_modify_ah); 581 582 int ib_query_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr) 583 { 584 return ah->device->query_ah ? 585 ah->device->query_ah(ah, ah_attr) : 586 -ENOSYS; 587 } 588 EXPORT_SYMBOL(ib_query_ah); 589 590 int ib_destroy_ah(struct ib_ah *ah) 591 { 592 struct ib_pd *pd; 593 int ret; 594 595 pd = ah->pd; 596 ret = ah->device->destroy_ah(ah); 597 if (!ret) 598 atomic_dec(&pd->usecnt); 599 600 return ret; 601 } 602 EXPORT_SYMBOL(ib_destroy_ah); 603 604 /* Shared receive queues */ 605 606 struct ib_srq *ib_create_srq(struct ib_pd *pd, 607 struct ib_srq_init_attr *srq_init_attr) 608 { 609 struct ib_srq *srq; 610 611 if (!pd->device->create_srq) 612 return ERR_PTR(-ENOSYS); 613 614 srq = pd->device->create_srq(pd, srq_init_attr, NULL); 615 616 if (!IS_ERR(srq)) { 617 srq->device = pd->device; 618 srq->pd = pd; 619 srq->uobject = NULL; 620 srq->event_handler = srq_init_attr->event_handler; 621 srq->srq_context = srq_init_attr->srq_context; 622 srq->srq_type = srq_init_attr->srq_type; 623 if (srq->srq_type == IB_SRQT_XRC) { 624 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd; 625 srq->ext.xrc.cq = srq_init_attr->ext.xrc.cq; 626 atomic_inc(&srq->ext.xrc.xrcd->usecnt); 627 atomic_inc(&srq->ext.xrc.cq->usecnt); 628 } 629 atomic_inc(&pd->usecnt); 630 atomic_set(&srq->usecnt, 0); 631 } 632 633 return srq; 634 } 635 EXPORT_SYMBOL(ib_create_srq); 636 637 int ib_modify_srq(struct ib_srq *srq, 638 struct ib_srq_attr *srq_attr, 639 enum ib_srq_attr_mask srq_attr_mask) 640 { 641 return srq->device->modify_srq ? 642 srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) : 643 -ENOSYS; 644 } 645 EXPORT_SYMBOL(ib_modify_srq); 646 647 int ib_query_srq(struct ib_srq *srq, 648 struct ib_srq_attr *srq_attr) 649 { 650 return srq->device->query_srq ? 651 srq->device->query_srq(srq, srq_attr) : -ENOSYS; 652 } 653 EXPORT_SYMBOL(ib_query_srq); 654 655 int ib_destroy_srq(struct ib_srq *srq) 656 { 657 struct ib_pd *pd; 658 enum ib_srq_type srq_type; 659 struct ib_xrcd *uninitialized_var(xrcd); 660 struct ib_cq *uninitialized_var(cq); 661 int ret; 662 663 if (atomic_read(&srq->usecnt)) 664 return -EBUSY; 665 666 pd = srq->pd; 667 srq_type = srq->srq_type; 668 if (srq_type == IB_SRQT_XRC) { 669 xrcd = srq->ext.xrc.xrcd; 670 cq = srq->ext.xrc.cq; 671 } 672 673 ret = srq->device->destroy_srq(srq); 674 if (!ret) { 675 atomic_dec(&pd->usecnt); 676 if (srq_type == IB_SRQT_XRC) { 677 atomic_dec(&xrcd->usecnt); 678 atomic_dec(&cq->usecnt); 679 } 680 } 681 682 return ret; 683 } 684 EXPORT_SYMBOL(ib_destroy_srq); 685 686 /* Queue pairs */ 687 688 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context) 689 { 690 struct ib_qp *qp = context; 691 unsigned long flags; 692 693 spin_lock_irqsave(&qp->device->event_handler_lock, flags); 694 list_for_each_entry(event->element.qp, &qp->open_list, open_list) 695 if (event->element.qp->event_handler) 696 event->element.qp->event_handler(event, event->element.qp->qp_context); 697 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags); 698 } 699 700 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp) 701 { 702 mutex_lock(&xrcd->tgt_qp_mutex); 703 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list); 704 mutex_unlock(&xrcd->tgt_qp_mutex); 705 } 706 707 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp, 708 void (*event_handler)(struct ib_event *, void *), 709 void *qp_context) 710 { 711 struct ib_qp *qp; 712 unsigned long flags; 713 714 qp = kzalloc(sizeof *qp, GFP_KERNEL); 715 if (!qp) 716 return ERR_PTR(-ENOMEM); 717 718 qp->real_qp = real_qp; 719 atomic_inc(&real_qp->usecnt); 720 qp->device = real_qp->device; 721 qp->event_handler = event_handler; 722 qp->qp_context = qp_context; 723 qp->qp_num = real_qp->qp_num; 724 qp->qp_type = real_qp->qp_type; 725 726 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags); 727 list_add(&qp->open_list, &real_qp->open_list); 728 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags); 729 730 return qp; 731 } 732 733 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd, 734 struct ib_qp_open_attr *qp_open_attr) 735 { 736 struct ib_qp *qp, *real_qp; 737 738 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT) 739 return ERR_PTR(-EINVAL); 740 741 qp = ERR_PTR(-EINVAL); 742 mutex_lock(&xrcd->tgt_qp_mutex); 743 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) { 744 if (real_qp->qp_num == qp_open_attr->qp_num) { 745 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler, 746 qp_open_attr->qp_context); 747 break; 748 } 749 } 750 mutex_unlock(&xrcd->tgt_qp_mutex); 751 return qp; 752 } 753 EXPORT_SYMBOL(ib_open_qp); 754 755 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp, 756 struct ib_qp_init_attr *qp_init_attr) 757 { 758 struct ib_qp *real_qp = qp; 759 760 qp->event_handler = __ib_shared_qp_event_handler; 761 qp->qp_context = qp; 762 qp->pd = NULL; 763 qp->send_cq = qp->recv_cq = NULL; 764 qp->srq = NULL; 765 qp->xrcd = qp_init_attr->xrcd; 766 atomic_inc(&qp_init_attr->xrcd->usecnt); 767 INIT_LIST_HEAD(&qp->open_list); 768 769 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler, 770 qp_init_attr->qp_context); 771 if (!IS_ERR(qp)) 772 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp); 773 else 774 real_qp->device->destroy_qp(real_qp); 775 return qp; 776 } 777 778 struct ib_qp *ib_create_qp(struct ib_pd *pd, 779 struct ib_qp_init_attr *qp_init_attr) 780 { 781 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device; 782 struct ib_qp *qp; 783 784 if (qp_init_attr->rwq_ind_tbl && 785 (qp_init_attr->recv_cq || 786 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr || 787 qp_init_attr->cap.max_recv_sge)) 788 return ERR_PTR(-EINVAL); 789 790 qp = device->create_qp(pd, qp_init_attr, NULL); 791 if (IS_ERR(qp)) 792 return qp; 793 794 qp->device = device; 795 qp->real_qp = qp; 796 qp->uobject = NULL; 797 qp->qp_type = qp_init_attr->qp_type; 798 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl; 799 800 atomic_set(&qp->usecnt, 0); 801 spin_lock_init(&qp->mr_lock); 802 803 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) 804 return ib_create_xrc_qp(qp, qp_init_attr); 805 806 qp->event_handler = qp_init_attr->event_handler; 807 qp->qp_context = qp_init_attr->qp_context; 808 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) { 809 qp->recv_cq = NULL; 810 qp->srq = NULL; 811 } else { 812 qp->recv_cq = qp_init_attr->recv_cq; 813 if (qp_init_attr->recv_cq) 814 atomic_inc(&qp_init_attr->recv_cq->usecnt); 815 qp->srq = qp_init_attr->srq; 816 if (qp->srq) 817 atomic_inc(&qp_init_attr->srq->usecnt); 818 } 819 820 qp->pd = pd; 821 qp->send_cq = qp_init_attr->send_cq; 822 qp->xrcd = NULL; 823 824 atomic_inc(&pd->usecnt); 825 if (qp_init_attr->send_cq) 826 atomic_inc(&qp_init_attr->send_cq->usecnt); 827 if (qp_init_attr->rwq_ind_tbl) 828 atomic_inc(&qp->rwq_ind_tbl->usecnt); 829 830 /* 831 * Note: all hw drivers guarantee that max_send_sge is lower than 832 * the device RDMA WRITE SGE limit but not all hw drivers ensure that 833 * max_send_sge <= max_sge_rd. 834 */ 835 qp->max_write_sge = qp_init_attr->cap.max_send_sge; 836 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge, 837 device->attrs.max_sge_rd); 838 839 return qp; 840 } 841 EXPORT_SYMBOL(ib_create_qp); 842 843 static const struct { 844 int valid; 845 enum ib_qp_attr_mask req_param[IB_QPT_MAX]; 846 enum ib_qp_attr_mask opt_param[IB_QPT_MAX]; 847 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = { 848 [IB_QPS_RESET] = { 849 [IB_QPS_RESET] = { .valid = 1 }, 850 [IB_QPS_INIT] = { 851 .valid = 1, 852 .req_param = { 853 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 854 IB_QP_PORT | 855 IB_QP_QKEY), 856 [IB_QPT_RAW_PACKET] = IB_QP_PORT, 857 [IB_QPT_UC] = (IB_QP_PKEY_INDEX | 858 IB_QP_PORT | 859 IB_QP_ACCESS_FLAGS), 860 [IB_QPT_RC] = (IB_QP_PKEY_INDEX | 861 IB_QP_PORT | 862 IB_QP_ACCESS_FLAGS), 863 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX | 864 IB_QP_PORT | 865 IB_QP_ACCESS_FLAGS), 866 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX | 867 IB_QP_PORT | 868 IB_QP_ACCESS_FLAGS), 869 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 870 IB_QP_QKEY), 871 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 872 IB_QP_QKEY), 873 } 874 }, 875 }, 876 [IB_QPS_INIT] = { 877 [IB_QPS_RESET] = { .valid = 1 }, 878 [IB_QPS_ERR] = { .valid = 1 }, 879 [IB_QPS_INIT] = { 880 .valid = 1, 881 .opt_param = { 882 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 883 IB_QP_PORT | 884 IB_QP_QKEY), 885 [IB_QPT_UC] = (IB_QP_PKEY_INDEX | 886 IB_QP_PORT | 887 IB_QP_ACCESS_FLAGS), 888 [IB_QPT_RC] = (IB_QP_PKEY_INDEX | 889 IB_QP_PORT | 890 IB_QP_ACCESS_FLAGS), 891 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX | 892 IB_QP_PORT | 893 IB_QP_ACCESS_FLAGS), 894 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX | 895 IB_QP_PORT | 896 IB_QP_ACCESS_FLAGS), 897 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 898 IB_QP_QKEY), 899 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 900 IB_QP_QKEY), 901 } 902 }, 903 [IB_QPS_RTR] = { 904 .valid = 1, 905 .req_param = { 906 [IB_QPT_UC] = (IB_QP_AV | 907 IB_QP_PATH_MTU | 908 IB_QP_DEST_QPN | 909 IB_QP_RQ_PSN), 910 [IB_QPT_RC] = (IB_QP_AV | 911 IB_QP_PATH_MTU | 912 IB_QP_DEST_QPN | 913 IB_QP_RQ_PSN | 914 IB_QP_MAX_DEST_RD_ATOMIC | 915 IB_QP_MIN_RNR_TIMER), 916 [IB_QPT_XRC_INI] = (IB_QP_AV | 917 IB_QP_PATH_MTU | 918 IB_QP_DEST_QPN | 919 IB_QP_RQ_PSN), 920 [IB_QPT_XRC_TGT] = (IB_QP_AV | 921 IB_QP_PATH_MTU | 922 IB_QP_DEST_QPN | 923 IB_QP_RQ_PSN | 924 IB_QP_MAX_DEST_RD_ATOMIC | 925 IB_QP_MIN_RNR_TIMER), 926 }, 927 .opt_param = { 928 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 929 IB_QP_QKEY), 930 [IB_QPT_UC] = (IB_QP_ALT_PATH | 931 IB_QP_ACCESS_FLAGS | 932 IB_QP_PKEY_INDEX), 933 [IB_QPT_RC] = (IB_QP_ALT_PATH | 934 IB_QP_ACCESS_FLAGS | 935 IB_QP_PKEY_INDEX), 936 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH | 937 IB_QP_ACCESS_FLAGS | 938 IB_QP_PKEY_INDEX), 939 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH | 940 IB_QP_ACCESS_FLAGS | 941 IB_QP_PKEY_INDEX), 942 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 943 IB_QP_QKEY), 944 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 945 IB_QP_QKEY), 946 }, 947 }, 948 }, 949 [IB_QPS_RTR] = { 950 [IB_QPS_RESET] = { .valid = 1 }, 951 [IB_QPS_ERR] = { .valid = 1 }, 952 [IB_QPS_RTS] = { 953 .valid = 1, 954 .req_param = { 955 [IB_QPT_UD] = IB_QP_SQ_PSN, 956 [IB_QPT_UC] = IB_QP_SQ_PSN, 957 [IB_QPT_RC] = (IB_QP_TIMEOUT | 958 IB_QP_RETRY_CNT | 959 IB_QP_RNR_RETRY | 960 IB_QP_SQ_PSN | 961 IB_QP_MAX_QP_RD_ATOMIC), 962 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT | 963 IB_QP_RETRY_CNT | 964 IB_QP_RNR_RETRY | 965 IB_QP_SQ_PSN | 966 IB_QP_MAX_QP_RD_ATOMIC), 967 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT | 968 IB_QP_SQ_PSN), 969 [IB_QPT_SMI] = IB_QP_SQ_PSN, 970 [IB_QPT_GSI] = IB_QP_SQ_PSN, 971 }, 972 .opt_param = { 973 [IB_QPT_UD] = (IB_QP_CUR_STATE | 974 IB_QP_QKEY), 975 [IB_QPT_UC] = (IB_QP_CUR_STATE | 976 IB_QP_ALT_PATH | 977 IB_QP_ACCESS_FLAGS | 978 IB_QP_PATH_MIG_STATE), 979 [IB_QPT_RC] = (IB_QP_CUR_STATE | 980 IB_QP_ALT_PATH | 981 IB_QP_ACCESS_FLAGS | 982 IB_QP_MIN_RNR_TIMER | 983 IB_QP_PATH_MIG_STATE), 984 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 985 IB_QP_ALT_PATH | 986 IB_QP_ACCESS_FLAGS | 987 IB_QP_PATH_MIG_STATE), 988 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 989 IB_QP_ALT_PATH | 990 IB_QP_ACCESS_FLAGS | 991 IB_QP_MIN_RNR_TIMER | 992 IB_QP_PATH_MIG_STATE), 993 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 994 IB_QP_QKEY), 995 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 996 IB_QP_QKEY), 997 } 998 } 999 }, 1000 [IB_QPS_RTS] = { 1001 [IB_QPS_RESET] = { .valid = 1 }, 1002 [IB_QPS_ERR] = { .valid = 1 }, 1003 [IB_QPS_RTS] = { 1004 .valid = 1, 1005 .opt_param = { 1006 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1007 IB_QP_QKEY), 1008 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1009 IB_QP_ACCESS_FLAGS | 1010 IB_QP_ALT_PATH | 1011 IB_QP_PATH_MIG_STATE), 1012 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1013 IB_QP_ACCESS_FLAGS | 1014 IB_QP_ALT_PATH | 1015 IB_QP_PATH_MIG_STATE | 1016 IB_QP_MIN_RNR_TIMER), 1017 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1018 IB_QP_ACCESS_FLAGS | 1019 IB_QP_ALT_PATH | 1020 IB_QP_PATH_MIG_STATE), 1021 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1022 IB_QP_ACCESS_FLAGS | 1023 IB_QP_ALT_PATH | 1024 IB_QP_PATH_MIG_STATE | 1025 IB_QP_MIN_RNR_TIMER), 1026 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1027 IB_QP_QKEY), 1028 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1029 IB_QP_QKEY), 1030 } 1031 }, 1032 [IB_QPS_SQD] = { 1033 .valid = 1, 1034 .opt_param = { 1035 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1036 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1037 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1038 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1039 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */ 1040 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY, 1041 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY 1042 } 1043 }, 1044 }, 1045 [IB_QPS_SQD] = { 1046 [IB_QPS_RESET] = { .valid = 1 }, 1047 [IB_QPS_ERR] = { .valid = 1 }, 1048 [IB_QPS_RTS] = { 1049 .valid = 1, 1050 .opt_param = { 1051 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1052 IB_QP_QKEY), 1053 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1054 IB_QP_ALT_PATH | 1055 IB_QP_ACCESS_FLAGS | 1056 IB_QP_PATH_MIG_STATE), 1057 [IB_QPT_RC] = (IB_QP_CUR_STATE | 1058 IB_QP_ALT_PATH | 1059 IB_QP_ACCESS_FLAGS | 1060 IB_QP_MIN_RNR_TIMER | 1061 IB_QP_PATH_MIG_STATE), 1062 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE | 1063 IB_QP_ALT_PATH | 1064 IB_QP_ACCESS_FLAGS | 1065 IB_QP_PATH_MIG_STATE), 1066 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE | 1067 IB_QP_ALT_PATH | 1068 IB_QP_ACCESS_FLAGS | 1069 IB_QP_MIN_RNR_TIMER | 1070 IB_QP_PATH_MIG_STATE), 1071 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1072 IB_QP_QKEY), 1073 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1074 IB_QP_QKEY), 1075 } 1076 }, 1077 [IB_QPS_SQD] = { 1078 .valid = 1, 1079 .opt_param = { 1080 [IB_QPT_UD] = (IB_QP_PKEY_INDEX | 1081 IB_QP_QKEY), 1082 [IB_QPT_UC] = (IB_QP_AV | 1083 IB_QP_ALT_PATH | 1084 IB_QP_ACCESS_FLAGS | 1085 IB_QP_PKEY_INDEX | 1086 IB_QP_PATH_MIG_STATE), 1087 [IB_QPT_RC] = (IB_QP_PORT | 1088 IB_QP_AV | 1089 IB_QP_TIMEOUT | 1090 IB_QP_RETRY_CNT | 1091 IB_QP_RNR_RETRY | 1092 IB_QP_MAX_QP_RD_ATOMIC | 1093 IB_QP_MAX_DEST_RD_ATOMIC | 1094 IB_QP_ALT_PATH | 1095 IB_QP_ACCESS_FLAGS | 1096 IB_QP_PKEY_INDEX | 1097 IB_QP_MIN_RNR_TIMER | 1098 IB_QP_PATH_MIG_STATE), 1099 [IB_QPT_XRC_INI] = (IB_QP_PORT | 1100 IB_QP_AV | 1101 IB_QP_TIMEOUT | 1102 IB_QP_RETRY_CNT | 1103 IB_QP_RNR_RETRY | 1104 IB_QP_MAX_QP_RD_ATOMIC | 1105 IB_QP_ALT_PATH | 1106 IB_QP_ACCESS_FLAGS | 1107 IB_QP_PKEY_INDEX | 1108 IB_QP_PATH_MIG_STATE), 1109 [IB_QPT_XRC_TGT] = (IB_QP_PORT | 1110 IB_QP_AV | 1111 IB_QP_TIMEOUT | 1112 IB_QP_MAX_DEST_RD_ATOMIC | 1113 IB_QP_ALT_PATH | 1114 IB_QP_ACCESS_FLAGS | 1115 IB_QP_PKEY_INDEX | 1116 IB_QP_MIN_RNR_TIMER | 1117 IB_QP_PATH_MIG_STATE), 1118 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX | 1119 IB_QP_QKEY), 1120 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX | 1121 IB_QP_QKEY), 1122 } 1123 } 1124 }, 1125 [IB_QPS_SQE] = { 1126 [IB_QPS_RESET] = { .valid = 1 }, 1127 [IB_QPS_ERR] = { .valid = 1 }, 1128 [IB_QPS_RTS] = { 1129 .valid = 1, 1130 .opt_param = { 1131 [IB_QPT_UD] = (IB_QP_CUR_STATE | 1132 IB_QP_QKEY), 1133 [IB_QPT_UC] = (IB_QP_CUR_STATE | 1134 IB_QP_ACCESS_FLAGS), 1135 [IB_QPT_SMI] = (IB_QP_CUR_STATE | 1136 IB_QP_QKEY), 1137 [IB_QPT_GSI] = (IB_QP_CUR_STATE | 1138 IB_QP_QKEY), 1139 } 1140 } 1141 }, 1142 [IB_QPS_ERR] = { 1143 [IB_QPS_RESET] = { .valid = 1 }, 1144 [IB_QPS_ERR] = { .valid = 1 } 1145 } 1146 }; 1147 1148 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state, 1149 enum ib_qp_type type, enum ib_qp_attr_mask mask, 1150 enum rdma_link_layer ll) 1151 { 1152 enum ib_qp_attr_mask req_param, opt_param; 1153 1154 if (cur_state < 0 || cur_state > IB_QPS_ERR || 1155 next_state < 0 || next_state > IB_QPS_ERR) 1156 return 0; 1157 1158 if (mask & IB_QP_CUR_STATE && 1159 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS && 1160 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE) 1161 return 0; 1162 1163 if (!qp_state_table[cur_state][next_state].valid) 1164 return 0; 1165 1166 req_param = qp_state_table[cur_state][next_state].req_param[type]; 1167 opt_param = qp_state_table[cur_state][next_state].opt_param[type]; 1168 1169 if ((mask & req_param) != req_param) 1170 return 0; 1171 1172 if (mask & ~(req_param | opt_param | IB_QP_STATE)) 1173 return 0; 1174 1175 return 1; 1176 } 1177 EXPORT_SYMBOL(ib_modify_qp_is_ok); 1178 1179 int ib_resolve_eth_dmac(struct ib_qp *qp, 1180 struct ib_qp_attr *qp_attr, int *qp_attr_mask) 1181 { 1182 int ret = 0; 1183 1184 if (*qp_attr_mask & IB_QP_AV) { 1185 if (qp_attr->ah_attr.port_num < rdma_start_port(qp->device) || 1186 qp_attr->ah_attr.port_num > rdma_end_port(qp->device)) 1187 return -EINVAL; 1188 1189 if (!rdma_cap_eth_ah(qp->device, qp_attr->ah_attr.port_num)) 1190 return 0; 1191 1192 if (rdma_link_local_addr((struct in6_addr *)qp_attr->ah_attr.grh.dgid.raw)) { 1193 rdma_get_ll_mac((struct in6_addr *)qp_attr->ah_attr.grh.dgid.raw, 1194 qp_attr->ah_attr.dmac); 1195 } else { 1196 union ib_gid sgid; 1197 struct ib_gid_attr sgid_attr; 1198 int ifindex; 1199 int hop_limit; 1200 1201 ret = ib_query_gid(qp->device, 1202 qp_attr->ah_attr.port_num, 1203 qp_attr->ah_attr.grh.sgid_index, 1204 &sgid, &sgid_attr); 1205 1206 if (ret || !sgid_attr.ndev) { 1207 if (!ret) 1208 ret = -ENXIO; 1209 goto out; 1210 } 1211 1212 ifindex = sgid_attr.ndev->if_index; 1213 1214 ret = rdma_addr_find_l2_eth_by_grh(&sgid, 1215 &qp_attr->ah_attr.grh.dgid, 1216 qp_attr->ah_attr.dmac, 1217 NULL, &ifindex, &hop_limit); 1218 1219 dev_put(sgid_attr.ndev); 1220 1221 qp_attr->ah_attr.grh.hop_limit = hop_limit; 1222 } 1223 } 1224 out: 1225 return ret; 1226 } 1227 EXPORT_SYMBOL(ib_resolve_eth_dmac); 1228 1229 1230 int ib_modify_qp(struct ib_qp *qp, 1231 struct ib_qp_attr *qp_attr, 1232 int qp_attr_mask) 1233 { 1234 int ret; 1235 1236 ret = ib_resolve_eth_dmac(qp, qp_attr, &qp_attr_mask); 1237 if (ret) 1238 return ret; 1239 1240 return qp->device->modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL); 1241 } 1242 EXPORT_SYMBOL(ib_modify_qp); 1243 1244 int ib_query_qp(struct ib_qp *qp, 1245 struct ib_qp_attr *qp_attr, 1246 int qp_attr_mask, 1247 struct ib_qp_init_attr *qp_init_attr) 1248 { 1249 return qp->device->query_qp ? 1250 qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) : 1251 -ENOSYS; 1252 } 1253 EXPORT_SYMBOL(ib_query_qp); 1254 1255 int ib_close_qp(struct ib_qp *qp) 1256 { 1257 struct ib_qp *real_qp; 1258 unsigned long flags; 1259 1260 real_qp = qp->real_qp; 1261 if (real_qp == qp) 1262 return -EINVAL; 1263 1264 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags); 1265 list_del(&qp->open_list); 1266 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags); 1267 1268 atomic_dec(&real_qp->usecnt); 1269 kfree(qp); 1270 1271 return 0; 1272 } 1273 EXPORT_SYMBOL(ib_close_qp); 1274 1275 static int __ib_destroy_shared_qp(struct ib_qp *qp) 1276 { 1277 struct ib_xrcd *xrcd; 1278 struct ib_qp *real_qp; 1279 int ret; 1280 1281 real_qp = qp->real_qp; 1282 xrcd = real_qp->xrcd; 1283 1284 mutex_lock(&xrcd->tgt_qp_mutex); 1285 ib_close_qp(qp); 1286 if (atomic_read(&real_qp->usecnt) == 0) 1287 list_del(&real_qp->xrcd_list); 1288 else 1289 real_qp = NULL; 1290 mutex_unlock(&xrcd->tgt_qp_mutex); 1291 1292 if (real_qp) { 1293 ret = ib_destroy_qp(real_qp); 1294 if (!ret) 1295 atomic_dec(&xrcd->usecnt); 1296 else 1297 __ib_insert_xrcd_qp(xrcd, real_qp); 1298 } 1299 1300 return 0; 1301 } 1302 1303 int ib_destroy_qp(struct ib_qp *qp) 1304 { 1305 struct ib_pd *pd; 1306 struct ib_cq *scq, *rcq; 1307 struct ib_srq *srq; 1308 struct ib_rwq_ind_table *ind_tbl; 1309 int ret; 1310 1311 if (atomic_read(&qp->usecnt)) 1312 return -EBUSY; 1313 1314 if (qp->real_qp != qp) 1315 return __ib_destroy_shared_qp(qp); 1316 1317 pd = qp->pd; 1318 scq = qp->send_cq; 1319 rcq = qp->recv_cq; 1320 srq = qp->srq; 1321 ind_tbl = qp->rwq_ind_tbl; 1322 1323 ret = qp->device->destroy_qp(qp); 1324 if (!ret) { 1325 if (pd) 1326 atomic_dec(&pd->usecnt); 1327 if (scq) 1328 atomic_dec(&scq->usecnt); 1329 if (rcq) 1330 atomic_dec(&rcq->usecnt); 1331 if (srq) 1332 atomic_dec(&srq->usecnt); 1333 if (ind_tbl) 1334 atomic_dec(&ind_tbl->usecnt); 1335 } 1336 1337 return ret; 1338 } 1339 EXPORT_SYMBOL(ib_destroy_qp); 1340 1341 /* Completion queues */ 1342 1343 struct ib_cq *ib_create_cq(struct ib_device *device, 1344 ib_comp_handler comp_handler, 1345 void (*event_handler)(struct ib_event *, void *), 1346 void *cq_context, 1347 const struct ib_cq_init_attr *cq_attr) 1348 { 1349 struct ib_cq *cq; 1350 1351 cq = device->create_cq(device, cq_attr, NULL, NULL); 1352 1353 if (!IS_ERR(cq)) { 1354 cq->device = device; 1355 cq->uobject = NULL; 1356 cq->comp_handler = comp_handler; 1357 cq->event_handler = event_handler; 1358 cq->cq_context = cq_context; 1359 atomic_set(&cq->usecnt, 0); 1360 } 1361 1362 return cq; 1363 } 1364 EXPORT_SYMBOL(ib_create_cq); 1365 1366 int ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period) 1367 { 1368 return cq->device->modify_cq ? 1369 cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS; 1370 } 1371 EXPORT_SYMBOL(ib_modify_cq); 1372 1373 int ib_destroy_cq(struct ib_cq *cq) 1374 { 1375 if (atomic_read(&cq->usecnt)) 1376 return -EBUSY; 1377 1378 return cq->device->destroy_cq(cq); 1379 } 1380 EXPORT_SYMBOL(ib_destroy_cq); 1381 1382 int ib_resize_cq(struct ib_cq *cq, int cqe) 1383 { 1384 return cq->device->resize_cq ? 1385 cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS; 1386 } 1387 EXPORT_SYMBOL(ib_resize_cq); 1388 1389 /* Memory regions */ 1390 1391 int ib_dereg_mr(struct ib_mr *mr) 1392 { 1393 struct ib_pd *pd = mr->pd; 1394 int ret; 1395 1396 ret = mr->device->dereg_mr(mr); 1397 if (!ret) 1398 atomic_dec(&pd->usecnt); 1399 1400 return ret; 1401 } 1402 EXPORT_SYMBOL(ib_dereg_mr); 1403 1404 /** 1405 * ib_alloc_mr() - Allocates a memory region 1406 * @pd: protection domain associated with the region 1407 * @mr_type: memory region type 1408 * @max_num_sg: maximum sg entries available for registration. 1409 * 1410 * Notes: 1411 * Memory registeration page/sg lists must not exceed max_num_sg. 1412 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed 1413 * max_num_sg * used_page_size. 1414 * 1415 */ 1416 struct ib_mr *ib_alloc_mr(struct ib_pd *pd, 1417 enum ib_mr_type mr_type, 1418 u32 max_num_sg) 1419 { 1420 struct ib_mr *mr; 1421 1422 if (!pd->device->alloc_mr) 1423 return ERR_PTR(-ENOSYS); 1424 1425 mr = pd->device->alloc_mr(pd, mr_type, max_num_sg); 1426 if (!IS_ERR(mr)) { 1427 mr->device = pd->device; 1428 mr->pd = pd; 1429 mr->uobject = NULL; 1430 atomic_inc(&pd->usecnt); 1431 mr->need_inval = false; 1432 } 1433 1434 return mr; 1435 } 1436 EXPORT_SYMBOL(ib_alloc_mr); 1437 1438 /* "Fast" memory regions */ 1439 1440 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd, 1441 int mr_access_flags, 1442 struct ib_fmr_attr *fmr_attr) 1443 { 1444 struct ib_fmr *fmr; 1445 1446 if (!pd->device->alloc_fmr) 1447 return ERR_PTR(-ENOSYS); 1448 1449 fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr); 1450 if (!IS_ERR(fmr)) { 1451 fmr->device = pd->device; 1452 fmr->pd = pd; 1453 atomic_inc(&pd->usecnt); 1454 } 1455 1456 return fmr; 1457 } 1458 EXPORT_SYMBOL(ib_alloc_fmr); 1459 1460 int ib_unmap_fmr(struct list_head *fmr_list) 1461 { 1462 struct ib_fmr *fmr; 1463 1464 if (list_empty(fmr_list)) 1465 return 0; 1466 1467 fmr = list_entry(fmr_list->next, struct ib_fmr, list); 1468 return fmr->device->unmap_fmr(fmr_list); 1469 } 1470 EXPORT_SYMBOL(ib_unmap_fmr); 1471 1472 int ib_dealloc_fmr(struct ib_fmr *fmr) 1473 { 1474 struct ib_pd *pd; 1475 int ret; 1476 1477 pd = fmr->pd; 1478 ret = fmr->device->dealloc_fmr(fmr); 1479 if (!ret) 1480 atomic_dec(&pd->usecnt); 1481 1482 return ret; 1483 } 1484 EXPORT_SYMBOL(ib_dealloc_fmr); 1485 1486 /* Multicast groups */ 1487 1488 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid) 1489 { 1490 int ret; 1491 1492 if (!qp->device->attach_mcast) 1493 return -ENOSYS; 1494 if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD) 1495 return -EINVAL; 1496 1497 ret = qp->device->attach_mcast(qp, gid, lid); 1498 if (!ret) 1499 atomic_inc(&qp->usecnt); 1500 return ret; 1501 } 1502 EXPORT_SYMBOL(ib_attach_mcast); 1503 1504 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid) 1505 { 1506 int ret; 1507 1508 if (!qp->device->detach_mcast) 1509 return -ENOSYS; 1510 if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD) 1511 return -EINVAL; 1512 1513 ret = qp->device->detach_mcast(qp, gid, lid); 1514 if (!ret) 1515 atomic_dec(&qp->usecnt); 1516 return ret; 1517 } 1518 EXPORT_SYMBOL(ib_detach_mcast); 1519 1520 struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device) 1521 { 1522 struct ib_xrcd *xrcd; 1523 1524 if (!device->alloc_xrcd) 1525 return ERR_PTR(-ENOSYS); 1526 1527 xrcd = device->alloc_xrcd(device, NULL, NULL); 1528 if (!IS_ERR(xrcd)) { 1529 xrcd->device = device; 1530 xrcd->inode = NULL; 1531 atomic_set(&xrcd->usecnt, 0); 1532 mutex_init(&xrcd->tgt_qp_mutex); 1533 INIT_LIST_HEAD(&xrcd->tgt_qp_list); 1534 } 1535 1536 return xrcd; 1537 } 1538 EXPORT_SYMBOL(ib_alloc_xrcd); 1539 1540 int ib_dealloc_xrcd(struct ib_xrcd *xrcd) 1541 { 1542 struct ib_qp *qp; 1543 int ret; 1544 1545 if (atomic_read(&xrcd->usecnt)) 1546 return -EBUSY; 1547 1548 while (!list_empty(&xrcd->tgt_qp_list)) { 1549 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list); 1550 ret = ib_destroy_qp(qp); 1551 if (ret) 1552 return ret; 1553 } 1554 1555 return xrcd->device->dealloc_xrcd(xrcd); 1556 } 1557 EXPORT_SYMBOL(ib_dealloc_xrcd); 1558 1559 /** 1560 * ib_create_wq - Creates a WQ associated with the specified protection 1561 * domain. 1562 * @pd: The protection domain associated with the WQ. 1563 * @wq_init_attr: A list of initial attributes required to create the 1564 * WQ. If WQ creation succeeds, then the attributes are updated to 1565 * the actual capabilities of the created WQ. 1566 * 1567 * wq_init_attr->max_wr and wq_init_attr->max_sge determine 1568 * the requested size of the WQ, and set to the actual values allocated 1569 * on return. 1570 * If ib_create_wq() succeeds, then max_wr and max_sge will always be 1571 * at least as large as the requested values. 1572 */ 1573 struct ib_wq *ib_create_wq(struct ib_pd *pd, 1574 struct ib_wq_init_attr *wq_attr) 1575 { 1576 struct ib_wq *wq; 1577 1578 if (!pd->device->create_wq) 1579 return ERR_PTR(-ENOSYS); 1580 1581 wq = pd->device->create_wq(pd, wq_attr, NULL); 1582 if (!IS_ERR(wq)) { 1583 wq->event_handler = wq_attr->event_handler; 1584 wq->wq_context = wq_attr->wq_context; 1585 wq->wq_type = wq_attr->wq_type; 1586 wq->cq = wq_attr->cq; 1587 wq->device = pd->device; 1588 wq->pd = pd; 1589 wq->uobject = NULL; 1590 atomic_inc(&pd->usecnt); 1591 atomic_inc(&wq_attr->cq->usecnt); 1592 atomic_set(&wq->usecnt, 0); 1593 } 1594 return wq; 1595 } 1596 EXPORT_SYMBOL(ib_create_wq); 1597 1598 /** 1599 * ib_destroy_wq - Destroys the specified WQ. 1600 * @wq: The WQ to destroy. 1601 */ 1602 int ib_destroy_wq(struct ib_wq *wq) 1603 { 1604 int err; 1605 struct ib_cq *cq = wq->cq; 1606 struct ib_pd *pd = wq->pd; 1607 1608 if (atomic_read(&wq->usecnt)) 1609 return -EBUSY; 1610 1611 err = wq->device->destroy_wq(wq); 1612 if (!err) { 1613 atomic_dec(&pd->usecnt); 1614 atomic_dec(&cq->usecnt); 1615 } 1616 return err; 1617 } 1618 EXPORT_SYMBOL(ib_destroy_wq); 1619 1620 /** 1621 * ib_modify_wq - Modifies the specified WQ. 1622 * @wq: The WQ to modify. 1623 * @wq_attr: On input, specifies the WQ attributes to modify. 1624 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ 1625 * are being modified. 1626 * On output, the current values of selected WQ attributes are returned. 1627 */ 1628 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr, 1629 u32 wq_attr_mask) 1630 { 1631 int err; 1632 1633 if (!wq->device->modify_wq) 1634 return -ENOSYS; 1635 1636 err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL); 1637 return err; 1638 } 1639 EXPORT_SYMBOL(ib_modify_wq); 1640 1641 /* 1642 * ib_create_rwq_ind_table - Creates a RQ Indirection Table. 1643 * @device: The device on which to create the rwq indirection table. 1644 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to 1645 * create the Indirection Table. 1646 * 1647 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less 1648 * than the created ib_rwq_ind_table object and the caller is responsible 1649 * for its memory allocation/free. 1650 */ 1651 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device, 1652 struct ib_rwq_ind_table_init_attr *init_attr) 1653 { 1654 struct ib_rwq_ind_table *rwq_ind_table; 1655 int i; 1656 u32 table_size; 1657 1658 if (!device->create_rwq_ind_table) 1659 return ERR_PTR(-ENOSYS); 1660 1661 table_size = (1 << init_attr->log_ind_tbl_size); 1662 rwq_ind_table = device->create_rwq_ind_table(device, 1663 init_attr, NULL); 1664 if (IS_ERR(rwq_ind_table)) 1665 return rwq_ind_table; 1666 1667 rwq_ind_table->ind_tbl = init_attr->ind_tbl; 1668 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size; 1669 rwq_ind_table->device = device; 1670 rwq_ind_table->uobject = NULL; 1671 atomic_set(&rwq_ind_table->usecnt, 0); 1672 1673 for (i = 0; i < table_size; i++) 1674 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt); 1675 1676 return rwq_ind_table; 1677 } 1678 EXPORT_SYMBOL(ib_create_rwq_ind_table); 1679 1680 /* 1681 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table. 1682 * @wq_ind_table: The Indirection Table to destroy. 1683 */ 1684 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table) 1685 { 1686 int err, i; 1687 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size); 1688 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl; 1689 1690 if (atomic_read(&rwq_ind_table->usecnt)) 1691 return -EBUSY; 1692 1693 err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table); 1694 if (!err) { 1695 for (i = 0; i < table_size; i++) 1696 atomic_dec(&ind_tbl[i]->usecnt); 1697 } 1698 1699 return err; 1700 } 1701 EXPORT_SYMBOL(ib_destroy_rwq_ind_table); 1702 1703 struct ib_flow *ib_create_flow(struct ib_qp *qp, 1704 struct ib_flow_attr *flow_attr, 1705 int domain) 1706 { 1707 struct ib_flow *flow_id; 1708 if (!qp->device->create_flow) 1709 return ERR_PTR(-ENOSYS); 1710 1711 flow_id = qp->device->create_flow(qp, flow_attr, domain); 1712 if (!IS_ERR(flow_id)) 1713 atomic_inc(&qp->usecnt); 1714 return flow_id; 1715 } 1716 EXPORT_SYMBOL(ib_create_flow); 1717 1718 int ib_destroy_flow(struct ib_flow *flow_id) 1719 { 1720 int err; 1721 struct ib_qp *qp = flow_id->qp; 1722 1723 err = qp->device->destroy_flow(flow_id); 1724 if (!err) 1725 atomic_dec(&qp->usecnt); 1726 return err; 1727 } 1728 EXPORT_SYMBOL(ib_destroy_flow); 1729 1730 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask, 1731 struct ib_mr_status *mr_status) 1732 { 1733 return mr->device->check_mr_status ? 1734 mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS; 1735 } 1736 EXPORT_SYMBOL(ib_check_mr_status); 1737 1738 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port, 1739 int state) 1740 { 1741 if (!device->set_vf_link_state) 1742 return -ENOSYS; 1743 1744 return device->set_vf_link_state(device, vf, port, state); 1745 } 1746 EXPORT_SYMBOL(ib_set_vf_link_state); 1747 1748 int ib_get_vf_config(struct ib_device *device, int vf, u8 port, 1749 struct ifla_vf_info *info) 1750 { 1751 if (!device->get_vf_config) 1752 return -ENOSYS; 1753 1754 return device->get_vf_config(device, vf, port, info); 1755 } 1756 EXPORT_SYMBOL(ib_get_vf_config); 1757 1758 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port, 1759 struct ifla_vf_stats *stats) 1760 { 1761 if (!device->get_vf_stats) 1762 return -ENOSYS; 1763 1764 return device->get_vf_stats(device, vf, port, stats); 1765 } 1766 EXPORT_SYMBOL(ib_get_vf_stats); 1767 1768 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid, 1769 int type) 1770 { 1771 if (!device->set_vf_guid) 1772 return -ENOSYS; 1773 1774 return device->set_vf_guid(device, vf, port, guid, type); 1775 } 1776 EXPORT_SYMBOL(ib_set_vf_guid); 1777 1778 /** 1779 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list 1780 * and set it the memory region. 1781 * @mr: memory region 1782 * @sg: dma mapped scatterlist 1783 * @sg_nents: number of entries in sg 1784 * @sg_offset: offset in bytes into sg 1785 * @page_size: page vector desired page size 1786 * 1787 * Constraints: 1788 * - The first sg element is allowed to have an offset. 1789 * - Each sg element must either be aligned to page_size or virtually 1790 * contiguous to the previous element. In case an sg element has a 1791 * non-contiguous offset, the mapping prefix will not include it. 1792 * - The last sg element is allowed to have length less than page_size. 1793 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size 1794 * then only max_num_sg entries will be mapped. 1795 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these 1796 * constraints holds and the page_size argument is ignored. 1797 * 1798 * Returns the number of sg elements that were mapped to the memory region. 1799 * 1800 * After this completes successfully, the memory region 1801 * is ready for registration. 1802 */ 1803 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents, 1804 unsigned int *sg_offset, unsigned int page_size) 1805 { 1806 if (unlikely(!mr->device->map_mr_sg)) 1807 return -ENOSYS; 1808 1809 mr->page_size = page_size; 1810 1811 return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset); 1812 } 1813 EXPORT_SYMBOL(ib_map_mr_sg); 1814 1815 /** 1816 * ib_sg_to_pages() - Convert the largest prefix of a sg list 1817 * to a page vector 1818 * @mr: memory region 1819 * @sgl: dma mapped scatterlist 1820 * @sg_nents: number of entries in sg 1821 * @sg_offset_p: IN: start offset in bytes into sg 1822 * OUT: offset in bytes for element n of the sg of the first 1823 * byte that has not been processed where n is the return 1824 * value of this function. 1825 * @set_page: driver page assignment function pointer 1826 * 1827 * Core service helper for drivers to convert the largest 1828 * prefix of given sg list to a page vector. The sg list 1829 * prefix converted is the prefix that meet the requirements 1830 * of ib_map_mr_sg. 1831 * 1832 * Returns the number of sg elements that were assigned to 1833 * a page vector. 1834 */ 1835 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents, 1836 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64)) 1837 { 1838 struct scatterlist *sg; 1839 u64 last_end_dma_addr = 0; 1840 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0; 1841 unsigned int last_page_off = 0; 1842 u64 page_mask = ~((u64)mr->page_size - 1); 1843 int i, ret; 1844 1845 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0]))) 1846 return -EINVAL; 1847 1848 mr->iova = sg_dma_address(&sgl[0]) + sg_offset; 1849 mr->length = 0; 1850 1851 for_each_sg(sgl, sg, sg_nents, i) { 1852 u64 dma_addr = sg_dma_address(sg) + sg_offset; 1853 u64 prev_addr = dma_addr; 1854 unsigned int dma_len = sg_dma_len(sg) - sg_offset; 1855 u64 end_dma_addr = dma_addr + dma_len; 1856 u64 page_addr = dma_addr & page_mask; 1857 1858 /* 1859 * For the second and later elements, check whether either the 1860 * end of element i-1 or the start of element i is not aligned 1861 * on a page boundary. 1862 */ 1863 if (i && (last_page_off != 0 || page_addr != dma_addr)) { 1864 /* Stop mapping if there is a gap. */ 1865 if (last_end_dma_addr != dma_addr) 1866 break; 1867 1868 /* 1869 * Coalesce this element with the last. If it is small 1870 * enough just update mr->length. Otherwise start 1871 * mapping from the next page. 1872 */ 1873 goto next_page; 1874 } 1875 1876 do { 1877 ret = set_page(mr, page_addr); 1878 if (unlikely(ret < 0)) { 1879 sg_offset = prev_addr - sg_dma_address(sg); 1880 mr->length += prev_addr - dma_addr; 1881 if (sg_offset_p) 1882 *sg_offset_p = sg_offset; 1883 return i || sg_offset ? i : ret; 1884 } 1885 prev_addr = page_addr; 1886 next_page: 1887 page_addr += mr->page_size; 1888 } while (page_addr < end_dma_addr); 1889 1890 mr->length += dma_len; 1891 last_end_dma_addr = end_dma_addr; 1892 last_page_off = end_dma_addr & ~page_mask; 1893 1894 sg_offset = 0; 1895 } 1896 1897 if (sg_offset_p) 1898 *sg_offset_p = 0; 1899 return i; 1900 } 1901 EXPORT_SYMBOL(ib_sg_to_pages); 1902 1903 struct ib_drain_cqe { 1904 struct ib_cqe cqe; 1905 struct completion done; 1906 }; 1907 1908 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc) 1909 { 1910 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe, 1911 cqe); 1912 1913 complete(&cqe->done); 1914 } 1915 1916 /* 1917 * Post a WR and block until its completion is reaped for the SQ. 1918 */ 1919 static void __ib_drain_sq(struct ib_qp *qp) 1920 { 1921 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 1922 struct ib_drain_cqe sdrain; 1923 struct ib_send_wr swr = {}, *bad_swr; 1924 int ret; 1925 1926 if (qp->send_cq->poll_ctx == IB_POLL_DIRECT) { 1927 WARN_ONCE(qp->send_cq->poll_ctx == IB_POLL_DIRECT, 1928 "IB_POLL_DIRECT poll_ctx not supported for drain\n"); 1929 return; 1930 } 1931 1932 swr.wr_cqe = &sdrain.cqe; 1933 sdrain.cqe.done = ib_drain_qp_done; 1934 init_completion(&sdrain.done); 1935 1936 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 1937 if (ret) { 1938 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 1939 return; 1940 } 1941 1942 ret = ib_post_send(qp, &swr, &bad_swr); 1943 if (ret) { 1944 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 1945 return; 1946 } 1947 1948 wait_for_completion(&sdrain.done); 1949 } 1950 1951 /* 1952 * Post a WR and block until its completion is reaped for the RQ. 1953 */ 1954 static void __ib_drain_rq(struct ib_qp *qp) 1955 { 1956 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 1957 struct ib_drain_cqe rdrain; 1958 struct ib_recv_wr rwr = {}, *bad_rwr; 1959 int ret; 1960 1961 if (qp->recv_cq->poll_ctx == IB_POLL_DIRECT) { 1962 WARN_ONCE(qp->recv_cq->poll_ctx == IB_POLL_DIRECT, 1963 "IB_POLL_DIRECT poll_ctx not supported for drain\n"); 1964 return; 1965 } 1966 1967 rwr.wr_cqe = &rdrain.cqe; 1968 rdrain.cqe.done = ib_drain_qp_done; 1969 init_completion(&rdrain.done); 1970 1971 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 1972 if (ret) { 1973 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 1974 return; 1975 } 1976 1977 ret = ib_post_recv(qp, &rwr, &bad_rwr); 1978 if (ret) { 1979 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 1980 return; 1981 } 1982 1983 wait_for_completion(&rdrain.done); 1984 } 1985 1986 /** 1987 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the 1988 * application. 1989 * @qp: queue pair to drain 1990 * 1991 * If the device has a provider-specific drain function, then 1992 * call that. Otherwise call the generic drain function 1993 * __ib_drain_sq(). 1994 * 1995 * The caller must: 1996 * 1997 * ensure there is room in the CQ and SQ for the drain work request and 1998 * completion. 1999 * 2000 * allocate the CQ using ib_alloc_cq() and the CQ poll context cannot be 2001 * IB_POLL_DIRECT. 2002 * 2003 * ensure that there are no other contexts that are posting WRs concurrently. 2004 * Otherwise the drain is not guaranteed. 2005 */ 2006 void ib_drain_sq(struct ib_qp *qp) 2007 { 2008 if (qp->device->drain_sq) 2009 qp->device->drain_sq(qp); 2010 else 2011 __ib_drain_sq(qp); 2012 } 2013 EXPORT_SYMBOL(ib_drain_sq); 2014 2015 /** 2016 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the 2017 * application. 2018 * @qp: queue pair to drain 2019 * 2020 * If the device has a provider-specific drain function, then 2021 * call that. Otherwise call the generic drain function 2022 * __ib_drain_rq(). 2023 * 2024 * The caller must: 2025 * 2026 * ensure there is room in the CQ and RQ for the drain work request and 2027 * completion. 2028 * 2029 * allocate the CQ using ib_alloc_cq() and the CQ poll context cannot be 2030 * IB_POLL_DIRECT. 2031 * 2032 * ensure that there are no other contexts that are posting WRs concurrently. 2033 * Otherwise the drain is not guaranteed. 2034 */ 2035 void ib_drain_rq(struct ib_qp *qp) 2036 { 2037 if (qp->device->drain_rq) 2038 qp->device->drain_rq(qp); 2039 else 2040 __ib_drain_rq(qp); 2041 } 2042 EXPORT_SYMBOL(ib_drain_rq); 2043 2044 /** 2045 * ib_drain_qp() - Block until all CQEs have been consumed by the 2046 * application on both the RQ and SQ. 2047 * @qp: queue pair to drain 2048 * 2049 * The caller must: 2050 * 2051 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests 2052 * and completions. 2053 * 2054 * allocate the CQs using ib_alloc_cq() and the CQ poll context cannot be 2055 * IB_POLL_DIRECT. 2056 * 2057 * ensure that there are no other contexts that are posting WRs concurrently. 2058 * Otherwise the drain is not guaranteed. 2059 */ 2060 void ib_drain_qp(struct ib_qp *qp) 2061 { 2062 ib_drain_sq(qp); 2063 if (!qp->srq) 2064 ib_drain_rq(qp); 2065 } 2066 EXPORT_SYMBOL(ib_drain_qp); 2067