1/* 2 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 */ 32 33#include <linux/completion.h> 34#include <linux/file.h> 35#include <linux/mutex.h> 36#include <linux/poll.h> 37#include <linux/idr.h> 38#include <linux/in.h> 39#include <linux/in6.h> 40#include <linux/miscdevice.h> 41 42#include <rdma/rdma_user_cm.h> 43#include <rdma/ib_marshall.h> 44#include <rdma/rdma_cm.h> 45#include <rdma/rdma_cm_ib.h> 46 47MODULE_AUTHOR("Sean Hefty"); 48MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access"); 49MODULE_LICENSE("Dual BSD/GPL"); 50 51enum { 52 UCMA_MAX_BACKLOG = 1024 53}; 54 55struct ucma_file { 56 struct mutex mut; 57 struct file *filp; 58 struct list_head ctx_list; 59 struct list_head event_list; 60 wait_queue_head_t poll_wait; 61}; 62 63struct ucma_context { 64 int id; 65 struct completion comp; 66 atomic_t ref; 67 int events_reported; 68 int backlog; 69 70 struct ucma_file *file; 71 struct rdma_cm_id *cm_id; 72 u64 uid; 73 74 struct list_head list; 75 struct list_head mc_list; 76}; 77 78struct ucma_multicast { 79 struct ucma_context *ctx; 80 int id; 81 int events_reported; 82 83 u64 uid; 84 struct list_head list; 85 struct sockaddr_storage addr; 86}; 87 88struct ucma_event { 89 struct ucma_context *ctx; 90 struct ucma_multicast *mc; 91 struct list_head list; 92 struct rdma_cm_id *cm_id; 93 struct rdma_ucm_event_resp resp; 94}; 95 96static DEFINE_MUTEX(mut); 97static DEFINE_IDR(ctx_idr); 98static DEFINE_IDR(multicast_idr); 99 100static inline struct ucma_context *_ucma_find_context(int id, 101 struct ucma_file *file) 102{ 103 struct ucma_context *ctx; 104 105 ctx = idr_find(&ctx_idr, id); 106 if (!ctx) 107 ctx = ERR_PTR(-ENOENT); 108 else if (ctx->file != file) 109 ctx = ERR_PTR(-EINVAL); 110 return ctx; 111} 112 113static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id) 114{ 115 struct ucma_context *ctx; 116 117 mutex_lock(&mut); 118 ctx = _ucma_find_context(id, file); 119 if (!IS_ERR(ctx)) 120 atomic_inc(&ctx->ref); 121 mutex_unlock(&mut); 122 return ctx; 123} 124 125static void ucma_put_ctx(struct ucma_context *ctx) 126{ 127 if (atomic_dec_and_test(&ctx->ref)) 128 complete(&ctx->comp); 129} 130 131static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file) 132{ 133 struct ucma_context *ctx; 134 int ret; 135 136 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 137 if (!ctx) 138 return NULL; 139 140 atomic_set(&ctx->ref, 1); 141 init_completion(&ctx->comp); 142 INIT_LIST_HEAD(&ctx->mc_list); 143 ctx->file = file; 144 145 do { 146 ret = idr_pre_get(&ctx_idr, GFP_KERNEL); 147 if (!ret) 148 goto error; 149 150 mutex_lock(&mut); 151 ret = idr_get_new(&ctx_idr, ctx, &ctx->id); 152 mutex_unlock(&mut); 153 } while (ret == -EAGAIN); 154 155 if (ret) 156 goto error; 157 158 list_add_tail(&ctx->list, &file->ctx_list); 159 return ctx; 160 161error: 162 kfree(ctx); 163 return NULL; 164} 165 166static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx) 167{ 168 struct ucma_multicast *mc; 169 int ret; 170 171 mc = kzalloc(sizeof(*mc), GFP_KERNEL); 172 if (!mc) 173 return NULL; 174 175 do { 176 ret = idr_pre_get(&multicast_idr, GFP_KERNEL); 177 if (!ret) 178 goto error; 179 180 mutex_lock(&mut); 181 ret = idr_get_new(&multicast_idr, mc, &mc->id); 182 mutex_unlock(&mut); 183 } while (ret == -EAGAIN); 184 185 if (ret) 186 goto error; 187 188 mc->ctx = ctx; 189 list_add_tail(&mc->list, &ctx->mc_list); 190 return mc; 191 192error: 193 kfree(mc); 194 return NULL; 195} 196 197static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst, 198 struct rdma_conn_param *src) 199{ 200 if (src->private_data_len) 201 memcpy(dst->private_data, src->private_data, 202 src->private_data_len); 203 dst->private_data_len = src->private_data_len; 204 dst->responder_resources =src->responder_resources; 205 dst->initiator_depth = src->initiator_depth; 206 dst->flow_control = src->flow_control; 207 dst->retry_count = src->retry_count; 208 dst->rnr_retry_count = src->rnr_retry_count; 209 dst->srq = src->srq; 210 dst->qp_num = src->qp_num; 211} 212 213static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst, 214 struct rdma_ud_param *src) 215{ 216 if (src->private_data_len) 217 memcpy(dst->private_data, src->private_data, 218 src->private_data_len); 219 dst->private_data_len = src->private_data_len; 220 ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr); 221 dst->qp_num = src->qp_num; 222 dst->qkey = src->qkey; 223} 224 225static void ucma_set_event_context(struct ucma_context *ctx, 226 struct rdma_cm_event *event, 227 struct ucma_event *uevent) 228{ 229 uevent->ctx = ctx; 230 switch (event->event) { 231 case RDMA_CM_EVENT_MULTICAST_JOIN: 232 case RDMA_CM_EVENT_MULTICAST_ERROR: 233 uevent->mc = (struct ucma_multicast *) 234 event->param.ud.private_data; 235 uevent->resp.uid = uevent->mc->uid; 236 uevent->resp.id = uevent->mc->id; 237 break; 238 default: 239 uevent->resp.uid = ctx->uid; 240 uevent->resp.id = ctx->id; 241 break; 242 } 243} 244 245static int ucma_event_handler(struct rdma_cm_id *cm_id, 246 struct rdma_cm_event *event) 247{ 248 struct ucma_event *uevent; 249 struct ucma_context *ctx = cm_id->context; 250 int ret = 0; 251 252 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL); 253 if (!uevent) 254 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST; 255 256 uevent->cm_id = cm_id; 257 ucma_set_event_context(ctx, event, uevent); 258 uevent->resp.event = event->event; 259 uevent->resp.status = event->status; 260 if (cm_id->ps == RDMA_PS_UDP || cm_id->ps == RDMA_PS_IPOIB) 261 ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud); 262 else 263 ucma_copy_conn_event(&uevent->resp.param.conn, 264 &event->param.conn); 265 266 mutex_lock(&ctx->file->mut); 267 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) { 268 if (!ctx->backlog) { 269 ret = -ENOMEM; 270 kfree(uevent); 271 goto out; 272 } 273 ctx->backlog--; 274 } else if (!ctx->uid) { 275 /* 276 * We ignore events for new connections until userspace has set 277 * their context. This can only happen if an error occurs on a 278 * new connection before the user accepts it. This is okay, 279 * since the accept will just fail later. 280 */ 281 kfree(uevent); 282 goto out; 283 } 284 285 list_add_tail(&uevent->list, &ctx->file->event_list); 286 wake_up_interruptible(&ctx->file->poll_wait); 287 if (ctx->file->filp) 288 selwakeup(&ctx->file->filp->f_selinfo); 289out: 290 mutex_unlock(&ctx->file->mut); 291 return ret; 292} 293 294static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf, 295 int in_len, int out_len) 296{ 297 struct ucma_context *ctx; 298 struct rdma_ucm_get_event cmd; 299 struct ucma_event *uevent; 300 int ret = 0; 301 DEFINE_WAIT(wait); 302 303 if (out_len < sizeof uevent->resp) 304 return -ENOSPC; 305 306 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 307 return -EFAULT; 308 309 mutex_lock(&file->mut); 310 while (list_empty(&file->event_list)) { 311 mutex_unlock(&file->mut); 312 313 if (file->filp->f_flags & O_NONBLOCK) 314 return -EAGAIN; 315 316 if (wait_event_interruptible(file->poll_wait, 317 !list_empty(&file->event_list))) 318 return -ERESTARTSYS; 319 320 mutex_lock(&file->mut); 321 } 322 323 uevent = list_entry(file->event_list.next, struct ucma_event, list); 324 325 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) { 326 ctx = ucma_alloc_ctx(file); 327 if (!ctx) { 328 ret = -ENOMEM; 329 goto done; 330 } 331 uevent->ctx->backlog++; 332 ctx->cm_id = uevent->cm_id; 333 ctx->cm_id->context = ctx; 334 uevent->resp.id = ctx->id; 335 } 336 337 if (copy_to_user((void __user *)(unsigned long)cmd.response, 338 &uevent->resp, sizeof uevent->resp)) { 339 ret = -EFAULT; 340 goto done; 341 } 342 343 list_del(&uevent->list); 344 uevent->ctx->events_reported++; 345 if (uevent->mc) 346 uevent->mc->events_reported++; 347 kfree(uevent); 348done: 349 mutex_unlock(&file->mut); 350 return ret; 351} 352 353static ssize_t ucma_create_id(struct ucma_file *file, 354 const char __user *inbuf, 355 int in_len, int out_len) 356{ 357 struct rdma_ucm_create_id cmd; 358 struct rdma_ucm_create_id_resp resp; 359 struct ucma_context *ctx; 360 int ret; 361 362 if (out_len < sizeof(resp)) 363 return -ENOSPC; 364 365 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 366 return -EFAULT; 367 368 mutex_lock(&file->mut); 369 ctx = ucma_alloc_ctx(file); 370 mutex_unlock(&file->mut); 371 if (!ctx) 372 return -ENOMEM; 373 374 ctx->uid = cmd.uid; 375 ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps); 376 if (IS_ERR(ctx->cm_id)) { 377 ret = PTR_ERR(ctx->cm_id); 378 goto err1; 379 } 380 381 resp.id = ctx->id; 382 if (copy_to_user((void __user *)(unsigned long)cmd.response, 383 &resp, sizeof(resp))) { 384 ret = -EFAULT; 385 goto err2; 386 } 387 return 0; 388 389err2: 390 rdma_destroy_id(ctx->cm_id); 391err1: 392 mutex_lock(&mut); 393 idr_remove(&ctx_idr, ctx->id); 394 mutex_unlock(&mut); 395 kfree(ctx); 396 return ret; 397} 398 399static void ucma_cleanup_multicast(struct ucma_context *ctx) 400{ 401 struct ucma_multicast *mc, *tmp; 402 403 mutex_lock(&mut); 404 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) { 405 list_del(&mc->list); 406 idr_remove(&multicast_idr, mc->id); 407 kfree(mc); 408 } 409 mutex_unlock(&mut); 410} 411 412static void ucma_cleanup_events(struct ucma_context *ctx) 413{ 414 struct ucma_event *uevent, *tmp; 415 416 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) { 417 if (uevent->ctx != ctx) 418 continue; 419 420 list_del(&uevent->list); 421 422 /* clear incoming connections. */ 423 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) 424 rdma_destroy_id(uevent->cm_id); 425 426 kfree(uevent); 427 } 428} 429 430static void ucma_cleanup_mc_events(struct ucma_multicast *mc) 431{ 432 struct ucma_event *uevent, *tmp; 433 434 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) { 435 if (uevent->mc != mc) 436 continue; 437 438 list_del(&uevent->list); 439 kfree(uevent); 440 } 441} 442 443static int ucma_free_ctx(struct ucma_context *ctx) 444{ 445 int events_reported; 446 447 /* No new events will be generated after destroying the id. */ 448 rdma_destroy_id(ctx->cm_id); 449 450 ucma_cleanup_multicast(ctx); 451 452 /* Cleanup events not yet reported to the user. */ 453 mutex_lock(&ctx->file->mut); 454 ucma_cleanup_events(ctx); 455 list_del(&ctx->list); 456 mutex_unlock(&ctx->file->mut); 457 458 events_reported = ctx->events_reported; 459 kfree(ctx); 460 return events_reported; 461} 462 463static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf, 464 int in_len, int out_len) 465{ 466 struct rdma_ucm_destroy_id cmd; 467 struct rdma_ucm_destroy_id_resp resp; 468 struct ucma_context *ctx; 469 int ret = 0; 470 471 if (out_len < sizeof(resp)) 472 return -ENOSPC; 473 474 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 475 return -EFAULT; 476 477 mutex_lock(&mut); 478 ctx = _ucma_find_context(cmd.id, file); 479 if (!IS_ERR(ctx)) 480 idr_remove(&ctx_idr, ctx->id); 481 mutex_unlock(&mut); 482 483 if (IS_ERR(ctx)) 484 return PTR_ERR(ctx); 485 486 ucma_put_ctx(ctx); 487 wait_for_completion(&ctx->comp); 488 resp.events_reported = ucma_free_ctx(ctx); 489 490 if (copy_to_user((void __user *)(unsigned long)cmd.response, 491 &resp, sizeof(resp))) 492 ret = -EFAULT; 493 494 return ret; 495} 496 497static ssize_t ucma_bind_addr(struct ucma_file *file, const char __user *inbuf, 498 int in_len, int out_len) 499{ 500 struct rdma_ucm_bind_addr cmd; 501 struct ucma_context *ctx; 502 int ret; 503 504 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 505 return -EFAULT; 506 507 ctx = ucma_get_ctx(file, cmd.id); 508 if (IS_ERR(ctx)) 509 return PTR_ERR(ctx); 510 511 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr); 512 ucma_put_ctx(ctx); 513 return ret; 514} 515 516static ssize_t ucma_resolve_addr(struct ucma_file *file, 517 const char __user *inbuf, 518 int in_len, int out_len) 519{ 520 struct rdma_ucm_resolve_addr cmd; 521 struct ucma_context *ctx; 522 int ret; 523 524 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 525 return -EFAULT; 526 527 ctx = ucma_get_ctx(file, cmd.id); 528 if (IS_ERR(ctx)) 529 return PTR_ERR(ctx); 530 531 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr, 532 (struct sockaddr *) &cmd.dst_addr, 533 cmd.timeout_ms); 534 ucma_put_ctx(ctx); 535 return ret; 536} 537 538static ssize_t ucma_resolve_route(struct ucma_file *file, 539 const char __user *inbuf, 540 int in_len, int out_len) 541{ 542 struct rdma_ucm_resolve_route cmd; 543 struct ucma_context *ctx; 544 int ret; 545 546 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 547 return -EFAULT; 548 549 ctx = ucma_get_ctx(file, cmd.id); 550 if (IS_ERR(ctx)) 551 return PTR_ERR(ctx); 552 553 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms); 554 ucma_put_ctx(ctx); 555 return ret; 556} 557 558static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp, 559 struct rdma_route *route) 560{ 561 struct rdma_dev_addr *dev_addr; 562 563 resp->num_paths = route->num_paths; 564 switch (route->num_paths) { 565 case 0: 566 dev_addr = &route->addr.dev_addr; 567 rdma_addr_get_dgid(dev_addr, 568 (union ib_gid *) &resp->ib_route[0].dgid); 569 rdma_addr_get_sgid(dev_addr, 570 (union ib_gid *) &resp->ib_route[0].sgid); 571 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 572 break; 573 case 2: 574 ib_copy_path_rec_to_user(&resp->ib_route[1], 575 &route->path_rec[1]); 576 /* fall through */ 577 case 1: 578 ib_copy_path_rec_to_user(&resp->ib_route[0], 579 &route->path_rec[0]); 580 break; 581 default: 582 break; 583 } 584} 585 586static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp, 587 struct rdma_route *route) 588{ 589 struct rdma_dev_addr *dev_addr; 590 struct net_device *dev; 591 u16 vid = 0; 592 593 resp->num_paths = route->num_paths; 594 switch (route->num_paths) { 595 case 0: 596 dev_addr = &route->addr.dev_addr; 597 dev = dev_get_by_index(&init_net, dev_addr->bound_dev_if); 598 if (dev) { 599 vid = rdma_vlan_dev_vlan_id(dev); 600 dev_put(dev); 601 } 602 603 iboe_mac_vlan_to_ll((union ib_gid *) &resp->ib_route[0].dgid, 604 dev_addr->dst_dev_addr, vid); 605 iboe_addr_get_sgid(dev_addr, 606 (union ib_gid *) &resp->ib_route[0].sgid); 607 resp->ib_route[0].pkey = cpu_to_be16(0xffff); 608 break; 609 case 2: 610 ib_copy_path_rec_to_user(&resp->ib_route[1], 611 &route->path_rec[1]); 612 /* fall through */ 613 case 1: 614 ib_copy_path_rec_to_user(&resp->ib_route[0], 615 &route->path_rec[0]); 616 break; 617 default: 618 break; 619 } 620} 621 622static ssize_t ucma_query_route(struct ucma_file *file, 623 const char __user *inbuf, 624 int in_len, int out_len) 625{ 626 struct rdma_ucm_query_route cmd; 627 struct rdma_ucm_query_route_resp resp; 628 struct ucma_context *ctx; 629 struct sockaddr *addr; 630 int ret = 0; 631 632 if (out_len < sizeof(resp)) 633 return -ENOSPC; 634 635 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 636 return -EFAULT; 637 638 ctx = ucma_get_ctx(file, cmd.id); 639 if (IS_ERR(ctx)) 640 return PTR_ERR(ctx); 641 642 memset(&resp, 0, sizeof resp); 643 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr; 644 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ? 645 sizeof(struct sockaddr_in) : 646 sizeof(struct sockaddr_in6)); 647 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr; 648 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ? 649 sizeof(struct sockaddr_in) : 650 sizeof(struct sockaddr_in6)); 651 if (!ctx->cm_id->device) 652 goto out; 653 654 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid; 655 resp.port_num = ctx->cm_id->port_num; 656 if (rdma_node_get_transport(ctx->cm_id->device->node_type) == RDMA_TRANSPORT_IB) { 657 switch (rdma_port_get_link_layer(ctx->cm_id->device, ctx->cm_id->port_num)) { 658 case IB_LINK_LAYER_INFINIBAND: 659 ucma_copy_ib_route(&resp, &ctx->cm_id->route); 660 break; 661 case IB_LINK_LAYER_ETHERNET: 662 ucma_copy_iboe_route(&resp, &ctx->cm_id->route); 663 break; 664 default: 665 break; 666 } 667 } 668 669out: 670 if (copy_to_user((void __user *)(unsigned long)cmd.response, 671 &resp, sizeof(resp))) 672 ret = -EFAULT; 673 674 ucma_put_ctx(ctx); 675 return ret; 676} 677 678static void ucma_copy_conn_param(struct rdma_conn_param *dst, 679 struct rdma_ucm_conn_param *src) 680{ 681 dst->private_data = src->private_data; 682 dst->private_data_len = src->private_data_len; 683 dst->responder_resources =src->responder_resources; 684 dst->initiator_depth = src->initiator_depth; 685 dst->flow_control = src->flow_control; 686 dst->retry_count = src->retry_count; 687 dst->rnr_retry_count = src->rnr_retry_count; 688 dst->srq = src->srq; 689 dst->qp_num = src->qp_num; 690} 691 692static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf, 693 int in_len, int out_len) 694{ 695 struct rdma_ucm_connect cmd; 696 struct rdma_conn_param conn_param; 697 struct ucma_context *ctx; 698 int ret; 699 700 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 701 return -EFAULT; 702 703 if (!cmd.conn_param.valid) 704 return -EINVAL; 705 706 ctx = ucma_get_ctx(file, cmd.id); 707 if (IS_ERR(ctx)) 708 return PTR_ERR(ctx); 709 710 ucma_copy_conn_param(&conn_param, &cmd.conn_param); 711 ret = rdma_connect(ctx->cm_id, &conn_param); 712 ucma_put_ctx(ctx); 713 return ret; 714} 715 716static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf, 717 int in_len, int out_len) 718{ 719 struct rdma_ucm_listen cmd; 720 struct ucma_context *ctx; 721 int ret; 722 723 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 724 return -EFAULT; 725 726 ctx = ucma_get_ctx(file, cmd.id); 727 if (IS_ERR(ctx)) 728 return PTR_ERR(ctx); 729 730 ctx->backlog = cmd.backlog > 0 && cmd.backlog < UCMA_MAX_BACKLOG ? 731 cmd.backlog : UCMA_MAX_BACKLOG; 732 ret = rdma_listen(ctx->cm_id, ctx->backlog); 733 ucma_put_ctx(ctx); 734 return ret; 735} 736 737static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf, 738 int in_len, int out_len) 739{ 740 struct rdma_ucm_accept cmd; 741 struct rdma_conn_param conn_param; 742 struct ucma_context *ctx; 743 int ret; 744 745 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 746 return -EFAULT; 747 748 ctx = ucma_get_ctx(file, cmd.id); 749 if (IS_ERR(ctx)) 750 return PTR_ERR(ctx); 751 752 if (cmd.conn_param.valid) { 753 ctx->uid = cmd.uid; 754 ucma_copy_conn_param(&conn_param, &cmd.conn_param); 755 ret = rdma_accept(ctx->cm_id, &conn_param); 756 } else 757 ret = rdma_accept(ctx->cm_id, NULL); 758 759 ucma_put_ctx(ctx); 760 return ret; 761} 762 763static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf, 764 int in_len, int out_len) 765{ 766 struct rdma_ucm_reject cmd; 767 struct ucma_context *ctx; 768 int ret; 769 770 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 771 return -EFAULT; 772 773 ctx = ucma_get_ctx(file, cmd.id); 774 if (IS_ERR(ctx)) 775 return PTR_ERR(ctx); 776 777 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len); 778 ucma_put_ctx(ctx); 779 return ret; 780} 781 782static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf, 783 int in_len, int out_len) 784{ 785 struct rdma_ucm_disconnect cmd; 786 struct ucma_context *ctx; 787 int ret; 788 789 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 790 return -EFAULT; 791 792 ctx = ucma_get_ctx(file, cmd.id); 793 if (IS_ERR(ctx)) 794 return PTR_ERR(ctx); 795 796 ret = rdma_disconnect(ctx->cm_id); 797 ucma_put_ctx(ctx); 798 return ret; 799} 800 801static ssize_t ucma_init_qp_attr(struct ucma_file *file, 802 const char __user *inbuf, 803 int in_len, int out_len) 804{ 805 struct rdma_ucm_init_qp_attr cmd; 806 struct ib_uverbs_qp_attr resp; 807 struct ucma_context *ctx; 808 struct ib_qp_attr qp_attr; 809 int ret; 810 811 if (out_len < sizeof(resp)) 812 return -ENOSPC; 813 814 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 815 return -EFAULT; 816 817 ctx = ucma_get_ctx(file, cmd.id); 818 if (IS_ERR(ctx)) 819 return PTR_ERR(ctx); 820 821 resp.qp_attr_mask = 0; 822 memset(&qp_attr, 0, sizeof qp_attr); 823 qp_attr.qp_state = cmd.qp_state; 824 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask); 825 if (ret) 826 goto out; 827 828 ib_copy_qp_attr_to_user(&resp, &qp_attr); 829 if (copy_to_user((void __user *)(unsigned long)cmd.response, 830 &resp, sizeof(resp))) 831 ret = -EFAULT; 832 833out: 834 ucma_put_ctx(ctx); 835 return ret; 836} 837 838static int ucma_set_option_id(struct ucma_context *ctx, int optname, 839 void *optval, size_t optlen) 840{ 841 int ret = 0; 842 843 switch (optname) { 844 case RDMA_OPTION_ID_TOS: 845 if (optlen != sizeof(u8)) { 846 ret = -EINVAL; 847 break; 848 } 849 rdma_set_service_type(ctx->cm_id, *((u8 *) optval)); 850 break; 851 default: 852 ret = -ENOSYS; 853 } 854 855 return ret; 856} 857 858static int ucma_set_ib_path(struct ucma_context *ctx, 859 struct ib_path_rec_data *path_data, size_t optlen) 860{ 861 struct ib_sa_path_rec sa_path; 862 struct rdma_cm_event event; 863 int ret; 864 865 if (optlen % sizeof(*path_data)) 866 return -EINVAL; 867 868 for (; optlen; optlen -= sizeof(*path_data), path_data++) { 869 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY | 870 IB_PATH_BIDIRECTIONAL)) 871 break; 872 } 873 874 if (!optlen) 875 return -EINVAL; 876 877 ib_sa_unpack_path(path_data->path_rec, &sa_path); 878 ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1); 879 if (ret) 880 return ret; 881 882 memset(&event, 0, sizeof event); 883 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 884 return ucma_event_handler(ctx->cm_id, &event); 885} 886 887static int ucma_set_option_ib(struct ucma_context *ctx, int optname, 888 void *optval, size_t optlen) 889{ 890 int ret; 891 892 switch (optname) { 893 case RDMA_OPTION_IB_PATH: 894 ret = ucma_set_ib_path(ctx, optval, optlen); 895 break; 896 default: 897 ret = -ENOSYS; 898 } 899 900 return ret; 901} 902 903static int ucma_set_option_level(struct ucma_context *ctx, int level, 904 int optname, void *optval, size_t optlen) 905{ 906 int ret; 907 908 switch (level) { 909 case RDMA_OPTION_ID: 910 ret = ucma_set_option_id(ctx, optname, optval, optlen); 911 break; 912 case RDMA_OPTION_IB: 913 ret = ucma_set_option_ib(ctx, optname, optval, optlen); 914 break; 915 default: 916 ret = -ENOSYS; 917 } 918 919 return ret; 920} 921 922static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf, 923 int in_len, int out_len) 924{ 925 struct rdma_ucm_set_option cmd; 926 struct ucma_context *ctx; 927 void *optval; 928 int ret; 929 930 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 931 return -EFAULT; 932 933 ctx = ucma_get_ctx(file, cmd.id); 934 if (IS_ERR(ctx)) 935 return PTR_ERR(ctx); 936 937 optval = kmalloc(cmd.optlen, GFP_KERNEL); 938 if (!optval) { 939 ret = -ENOMEM; 940 goto out1; 941 } 942 943 if (copy_from_user(optval, (void __user *) (unsigned long) cmd.optval, 944 cmd.optlen)) { 945 ret = -EFAULT; 946 goto out2; 947 } 948 949 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval, 950 cmd.optlen); 951out2: 952 kfree(optval); 953out1: 954 ucma_put_ctx(ctx); 955 return ret; 956} 957 958static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf, 959 int in_len, int out_len) 960{ 961 struct rdma_ucm_notify cmd; 962 struct ucma_context *ctx; 963 int ret; 964 965 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 966 return -EFAULT; 967 968 ctx = ucma_get_ctx(file, cmd.id); 969 if (IS_ERR(ctx)) 970 return PTR_ERR(ctx); 971 972 ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event); 973 ucma_put_ctx(ctx); 974 return ret; 975} 976 977static ssize_t ucma_join_multicast(struct ucma_file *file, 978 const char __user *inbuf, 979 int in_len, int out_len) 980{ 981 struct rdma_ucm_join_mcast cmd; 982 struct rdma_ucm_create_id_resp resp; 983 struct ucma_context *ctx; 984 struct ucma_multicast *mc; 985 int ret; 986 987 if (out_len < sizeof(resp)) 988 return -ENOSPC; 989 990 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 991 return -EFAULT; 992 993 ctx = ucma_get_ctx(file, cmd.id); 994 if (IS_ERR(ctx)) 995 return PTR_ERR(ctx); 996 997 mutex_lock(&file->mut); 998 mc = ucma_alloc_multicast(ctx); 999 if (!mc) { 1000 ret = -ENOMEM; 1001 goto err1; 1002 } 1003 1004 mc->uid = cmd.uid; 1005 memcpy(&mc->addr, &cmd.addr, sizeof cmd.addr); 1006 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc); 1007 if (ret) 1008 goto err2; 1009 1010 resp.id = mc->id; 1011 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1012 &resp, sizeof(resp))) { 1013 ret = -EFAULT; 1014 goto err3; 1015 } 1016 1017 mutex_unlock(&file->mut); 1018 ucma_put_ctx(ctx); 1019 return 0; 1020 1021err3: 1022 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr); 1023 ucma_cleanup_mc_events(mc); 1024err2: 1025 mutex_lock(&mut); 1026 idr_remove(&multicast_idr, mc->id); 1027 mutex_unlock(&mut); 1028 list_del(&mc->list); 1029 kfree(mc); 1030err1: 1031 mutex_unlock(&file->mut); 1032 ucma_put_ctx(ctx); 1033 return ret; 1034} 1035 1036static ssize_t ucma_leave_multicast(struct ucma_file *file, 1037 const char __user *inbuf, 1038 int in_len, int out_len) 1039{ 1040 struct rdma_ucm_destroy_id cmd; 1041 struct rdma_ucm_destroy_id_resp resp; 1042 struct ucma_multicast *mc; 1043 int ret = 0; 1044 1045 if (out_len < sizeof(resp)) 1046 return -ENOSPC; 1047 1048 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1049 return -EFAULT; 1050 1051 mutex_lock(&mut); 1052 mc = idr_find(&multicast_idr, cmd.id); 1053 if (!mc) 1054 mc = ERR_PTR(-ENOENT); 1055 else if (mc->ctx->file != file) 1056 mc = ERR_PTR(-EINVAL); 1057 else { 1058 idr_remove(&multicast_idr, mc->id); 1059 atomic_inc(&mc->ctx->ref); 1060 } 1061 mutex_unlock(&mut); 1062 1063 if (IS_ERR(mc)) { 1064 ret = PTR_ERR(mc); 1065 goto out; 1066 } 1067 1068 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr); 1069 mutex_lock(&mc->ctx->file->mut); 1070 ucma_cleanup_mc_events(mc); 1071 list_del(&mc->list); 1072 mutex_unlock(&mc->ctx->file->mut); 1073 1074 ucma_put_ctx(mc->ctx); 1075 resp.events_reported = mc->events_reported; 1076 kfree(mc); 1077 1078 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1079 &resp, sizeof(resp))) 1080 ret = -EFAULT; 1081out: 1082 return ret; 1083} 1084 1085static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2) 1086{ 1087 /* Acquire mutex's based on pointer comparison to prevent deadlock. */ 1088 if (file1 < file2) { 1089 mutex_lock(&file1->mut); 1090 mutex_lock(&file2->mut); 1091 } else { 1092 mutex_lock(&file2->mut); 1093 mutex_lock(&file1->mut); 1094 } 1095} 1096 1097static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2) 1098{ 1099 if (file1 < file2) { 1100 mutex_unlock(&file2->mut); 1101 mutex_unlock(&file1->mut); 1102 } else { 1103 mutex_unlock(&file1->mut); 1104 mutex_unlock(&file2->mut); 1105 } 1106} 1107 1108static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file) 1109{ 1110 struct ucma_event *uevent, *tmp; 1111 1112 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) 1113 if (uevent->ctx == ctx) 1114 list_move_tail(&uevent->list, &file->event_list); 1115} 1116 1117static ssize_t ucma_migrate_id(struct ucma_file *new_file, 1118 const char __user *inbuf, 1119 int in_len, int out_len) 1120{ 1121 struct rdma_ucm_migrate_id cmd; 1122 struct rdma_ucm_migrate_resp resp; 1123 struct ucma_context *ctx; 1124 struct file *filp; 1125 struct ucma_file *cur_file; 1126 int ret = 0; 1127 1128 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1129 return -EFAULT; 1130 1131 /* Get current fd to protect against it being closed */ 1132 filp = fget(cmd.fd); 1133 if (!filp) 1134 return -ENOENT; 1135 1136 /* Validate current fd and prevent destruction of id. */ 1137 ctx = ucma_get_ctx(filp->private_data, cmd.id); 1138 if (IS_ERR(ctx)) { 1139 ret = PTR_ERR(ctx); 1140 goto file_put; 1141 } 1142 1143 cur_file = ctx->file; 1144 if (cur_file == new_file) { 1145 resp.events_reported = ctx->events_reported; 1146 goto response; 1147 } 1148 1149 /* 1150 * Migrate events between fd's, maintaining order, and avoiding new 1151 * events being added before existing events. 1152 */ 1153 ucma_lock_files(cur_file, new_file); 1154 mutex_lock(&mut); 1155 1156 list_move_tail(&ctx->list, &new_file->ctx_list); 1157 ucma_move_events(ctx, new_file); 1158 ctx->file = new_file; 1159 resp.events_reported = ctx->events_reported; 1160 1161 mutex_unlock(&mut); 1162 ucma_unlock_files(cur_file, new_file); 1163 1164response: 1165 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1166 &resp, sizeof(resp))) 1167 ret = -EFAULT; 1168 1169 ucma_put_ctx(ctx); 1170file_put: 1171 fput(filp); 1172 return ret; 1173} 1174 1175static ssize_t (*ucma_cmd_table[])(struct ucma_file *file, 1176 const char __user *inbuf, 1177 int in_len, int out_len) = { 1178 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id, 1179 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id, 1180 [RDMA_USER_CM_CMD_BIND_ADDR] = ucma_bind_addr, 1181 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr, 1182 [RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route, 1183 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route, 1184 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect, 1185 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen, 1186 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept, 1187 [RDMA_USER_CM_CMD_REJECT] = ucma_reject, 1188 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect, 1189 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr, 1190 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event, 1191 [RDMA_USER_CM_CMD_GET_OPTION] = NULL, 1192 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option, 1193 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify, 1194 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast, 1195 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast, 1196 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id 1197}; 1198 1199static ssize_t ucma_write(struct file *filp, const char __user *buf, 1200 size_t len, loff_t *pos) 1201{ 1202 struct ucma_file *file = filp->private_data; 1203 struct rdma_ucm_cmd_hdr hdr; 1204 ssize_t ret; 1205 1206 if (len < sizeof(hdr)) 1207 return -EINVAL; 1208 1209 if (copy_from_user(&hdr, buf, sizeof(hdr))) 1210 return -EFAULT; 1211 1212 if (hdr.cmd < 0 || hdr.cmd >= ARRAY_SIZE(ucma_cmd_table)) 1213 return -EINVAL; 1214 1215 if (hdr.in + sizeof(hdr) > len) 1216 return -EINVAL; 1217 1218 if (!ucma_cmd_table[hdr.cmd]) 1219 return -ENOSYS; 1220 1221 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out); 1222 if (!ret) 1223 ret = len; 1224 1225 return ret; 1226} 1227 1228static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait) 1229{ 1230 struct ucma_file *file = filp->private_data; 1231 unsigned int mask = 0; 1232 1233 poll_wait(filp, &file->poll_wait, wait); 1234 1235 if (!list_empty(&file->event_list)) 1236 mask = POLLIN | POLLRDNORM; 1237 1238 return mask; 1239} 1240 1241/* 1242 * ucma_open() does not need the BKL: 1243 * 1244 * - no global state is referred to; 1245 * - there is no ioctl method to race against; 1246 * - no further module initialization is required for open to work 1247 * after the device is registered. 1248 */ 1249static int ucma_open(struct inode *inode, struct file *filp) 1250{ 1251 struct ucma_file *file; 1252 1253 file = kmalloc(sizeof *file, GFP_KERNEL); 1254 if (!file) 1255 return -ENOMEM; 1256 1257 INIT_LIST_HEAD(&file->event_list); 1258 INIT_LIST_HEAD(&file->ctx_list); 1259 init_waitqueue_head(&file->poll_wait); 1260 mutex_init(&file->mut); 1261 1262 filp->private_data = file; 1263 file->filp = filp; 1264 return 0; 1265} 1266 1267static int ucma_close(struct inode *inode, struct file *filp) 1268{ 1269 struct ucma_file *file = filp->private_data; 1270 struct ucma_context *ctx, *tmp; 1271 1272 mutex_lock(&file->mut); 1273 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) { 1274 mutex_unlock(&file->mut); 1275 1276 mutex_lock(&mut); 1277 idr_remove(&ctx_idr, ctx->id); 1278 mutex_unlock(&mut); 1279 1280 ucma_free_ctx(ctx); 1281 mutex_lock(&file->mut); 1282 } 1283 mutex_unlock(&file->mut); 1284 kfree(file); 1285 return 0; 1286} 1287 1288static const struct file_operations ucma_fops = { 1289 .owner = THIS_MODULE, 1290 .open = ucma_open, 1291 .release = ucma_close, 1292 .write = ucma_write, 1293 .poll = ucma_poll, 1294}; 1295 1296static struct miscdevice ucma_misc = { 1297 .minor = MISC_DYNAMIC_MINOR, 1298 .name = "rdma_cm", 1299 .fops = &ucma_fops, 1300}; 1301 1302static ssize_t show_abi_version(struct device *dev, 1303 struct device_attribute *attr, 1304 char *buf) 1305{ 1306 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION); 1307} 1308static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL); 1309 1310static int __init ucma_init(void) 1311{ 1312 int ret; 1313 1314 ret = misc_register(&ucma_misc); 1315 if (ret) 1316 return ret; 1317 1318 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version); 1319 if (ret) { 1320 printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n"); 1321 goto err; 1322 } 1323 return 0; 1324err: 1325 misc_deregister(&ucma_misc); 1326 return ret; 1327} 1328 1329static void __exit ucma_cleanup(void) 1330{ 1331 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version); 1332 misc_deregister(&ucma_misc); 1333 idr_destroy(&ctx_idr); 1334} 1335 1336module_init(ucma_init); 1337module_exit(ucma_cleanup); 1338