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/sched.h> 38 #include <linux/idr.h> 39 #include <linux/in.h> 40 #include <linux/in6.h> 41 #include <linux/miscdevice.h> 42 #include <linux/slab.h> 43 #include <linux/sysctl.h> 44 45 #include <rdma/rdma_user_cm.h> 46 #include <rdma/ib_marshall.h> 47 #include <rdma/rdma_cm.h> 48 #include <rdma/rdma_cm_ib.h> 49 50 MODULE_AUTHOR("Sean Hefty"); 51 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access"); 52 MODULE_LICENSE("Dual BSD/GPL"); 53 54 static unsigned int max_backlog = 1024; 55 56 static struct ctl_table_header *ucma_ctl_table_hdr; 57 static ctl_table ucma_ctl_table[] = { 58 { 59 .procname = "max_backlog", 60 .data = &max_backlog, 61 .maxlen = sizeof max_backlog, 62 .mode = 0644, 63 .proc_handler = proc_dointvec, 64 }, 65 { } 66 }; 67 68 static struct ctl_path ucma_ctl_path[] = { 69 { .procname = "net" }, 70 { .procname = "rdma_ucm" }, 71 { } 72 }; 73 74 struct ucma_file { 75 struct mutex mut; 76 struct file *filp; 77 struct list_head ctx_list; 78 struct list_head event_list; 79 wait_queue_head_t poll_wait; 80 }; 81 82 struct ucma_context { 83 int id; 84 struct completion comp; 85 atomic_t ref; 86 int events_reported; 87 int backlog; 88 89 struct ucma_file *file; 90 struct rdma_cm_id *cm_id; 91 u64 uid; 92 93 struct list_head list; 94 struct list_head mc_list; 95 }; 96 97 struct ucma_multicast { 98 struct ucma_context *ctx; 99 int id; 100 int events_reported; 101 102 u64 uid; 103 struct list_head list; 104 struct sockaddr_storage addr; 105 }; 106 107 struct ucma_event { 108 struct ucma_context *ctx; 109 struct ucma_multicast *mc; 110 struct list_head list; 111 struct rdma_cm_id *cm_id; 112 struct rdma_ucm_event_resp resp; 113 }; 114 115 static DEFINE_MUTEX(mut); 116 static DEFINE_IDR(ctx_idr); 117 static DEFINE_IDR(multicast_idr); 118 119 static inline struct ucma_context *_ucma_find_context(int id, 120 struct ucma_file *file) 121 { 122 struct ucma_context *ctx; 123 124 ctx = idr_find(&ctx_idr, id); 125 if (!ctx) 126 ctx = ERR_PTR(-ENOENT); 127 else if (ctx->file != file) 128 ctx = ERR_PTR(-EINVAL); 129 return ctx; 130 } 131 132 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id) 133 { 134 struct ucma_context *ctx; 135 136 mutex_lock(&mut); 137 ctx = _ucma_find_context(id, file); 138 if (!IS_ERR(ctx)) 139 atomic_inc(&ctx->ref); 140 mutex_unlock(&mut); 141 return ctx; 142 } 143 144 static void ucma_put_ctx(struct ucma_context *ctx) 145 { 146 if (atomic_dec_and_test(&ctx->ref)) 147 complete(&ctx->comp); 148 } 149 150 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file) 151 { 152 struct ucma_context *ctx; 153 int ret; 154 155 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 156 if (!ctx) 157 return NULL; 158 159 atomic_set(&ctx->ref, 1); 160 init_completion(&ctx->comp); 161 INIT_LIST_HEAD(&ctx->mc_list); 162 ctx->file = file; 163 164 do { 165 ret = idr_pre_get(&ctx_idr, GFP_KERNEL); 166 if (!ret) 167 goto error; 168 169 mutex_lock(&mut); 170 ret = idr_get_new(&ctx_idr, ctx, &ctx->id); 171 mutex_unlock(&mut); 172 } while (ret == -EAGAIN); 173 174 if (ret) 175 goto error; 176 177 list_add_tail(&ctx->list, &file->ctx_list); 178 return ctx; 179 180 error: 181 kfree(ctx); 182 return NULL; 183 } 184 185 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx) 186 { 187 struct ucma_multicast *mc; 188 int ret; 189 190 mc = kzalloc(sizeof(*mc), GFP_KERNEL); 191 if (!mc) 192 return NULL; 193 194 do { 195 ret = idr_pre_get(&multicast_idr, GFP_KERNEL); 196 if (!ret) 197 goto error; 198 199 mutex_lock(&mut); 200 ret = idr_get_new(&multicast_idr, mc, &mc->id); 201 mutex_unlock(&mut); 202 } while (ret == -EAGAIN); 203 204 if (ret) 205 goto error; 206 207 mc->ctx = ctx; 208 list_add_tail(&mc->list, &ctx->mc_list); 209 return mc; 210 211 error: 212 kfree(mc); 213 return NULL; 214 } 215 216 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst, 217 struct rdma_conn_param *src) 218 { 219 if (src->private_data_len) 220 memcpy(dst->private_data, src->private_data, 221 src->private_data_len); 222 dst->private_data_len = src->private_data_len; 223 dst->responder_resources =src->responder_resources; 224 dst->initiator_depth = src->initiator_depth; 225 dst->flow_control = src->flow_control; 226 dst->retry_count = src->retry_count; 227 dst->rnr_retry_count = src->rnr_retry_count; 228 dst->srq = src->srq; 229 dst->qp_num = src->qp_num; 230 } 231 232 static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst, 233 struct rdma_ud_param *src) 234 { 235 if (src->private_data_len) 236 memcpy(dst->private_data, src->private_data, 237 src->private_data_len); 238 dst->private_data_len = src->private_data_len; 239 ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr); 240 dst->qp_num = src->qp_num; 241 dst->qkey = src->qkey; 242 } 243 244 static void ucma_set_event_context(struct ucma_context *ctx, 245 struct rdma_cm_event *event, 246 struct ucma_event *uevent) 247 { 248 uevent->ctx = ctx; 249 switch (event->event) { 250 case RDMA_CM_EVENT_MULTICAST_JOIN: 251 case RDMA_CM_EVENT_MULTICAST_ERROR: 252 uevent->mc = (struct ucma_multicast *) 253 event->param.ud.private_data; 254 uevent->resp.uid = uevent->mc->uid; 255 uevent->resp.id = uevent->mc->id; 256 break; 257 default: 258 uevent->resp.uid = ctx->uid; 259 uevent->resp.id = ctx->id; 260 break; 261 } 262 } 263 264 static int ucma_event_handler(struct rdma_cm_id *cm_id, 265 struct rdma_cm_event *event) 266 { 267 struct ucma_event *uevent; 268 struct ucma_context *ctx = cm_id->context; 269 int ret = 0; 270 271 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL); 272 if (!uevent) 273 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST; 274 275 uevent->cm_id = cm_id; 276 ucma_set_event_context(ctx, event, uevent); 277 uevent->resp.event = event->event; 278 uevent->resp.status = event->status; 279 if (cm_id->ps == RDMA_PS_UDP || cm_id->ps == RDMA_PS_IPOIB) 280 ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud); 281 else 282 ucma_copy_conn_event(&uevent->resp.param.conn, 283 &event->param.conn); 284 285 mutex_lock(&ctx->file->mut); 286 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) { 287 if (!ctx->backlog) { 288 ret = -ENOMEM; 289 kfree(uevent); 290 goto out; 291 } 292 ctx->backlog--; 293 } else if (!ctx->uid) { 294 /* 295 * We ignore events for new connections until userspace has set 296 * their context. This can only happen if an error occurs on a 297 * new connection before the user accepts it. This is okay, 298 * since the accept will just fail later. 299 */ 300 kfree(uevent); 301 goto out; 302 } 303 304 list_add_tail(&uevent->list, &ctx->file->event_list); 305 wake_up_interruptible(&ctx->file->poll_wait); 306 out: 307 mutex_unlock(&ctx->file->mut); 308 return ret; 309 } 310 311 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf, 312 int in_len, int out_len) 313 { 314 struct ucma_context *ctx; 315 struct rdma_ucm_get_event cmd; 316 struct ucma_event *uevent; 317 int ret = 0; 318 DEFINE_WAIT(wait); 319 320 if (out_len < sizeof uevent->resp) 321 return -ENOSPC; 322 323 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 324 return -EFAULT; 325 326 mutex_lock(&file->mut); 327 while (list_empty(&file->event_list)) { 328 mutex_unlock(&file->mut); 329 330 if (file->filp->f_flags & O_NONBLOCK) 331 return -EAGAIN; 332 333 if (wait_event_interruptible(file->poll_wait, 334 !list_empty(&file->event_list))) 335 return -ERESTARTSYS; 336 337 mutex_lock(&file->mut); 338 } 339 340 uevent = list_entry(file->event_list.next, struct ucma_event, list); 341 342 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) { 343 ctx = ucma_alloc_ctx(file); 344 if (!ctx) { 345 ret = -ENOMEM; 346 goto done; 347 } 348 uevent->ctx->backlog++; 349 ctx->cm_id = uevent->cm_id; 350 ctx->cm_id->context = ctx; 351 uevent->resp.id = ctx->id; 352 } 353 354 if (copy_to_user((void __user *)(unsigned long)cmd.response, 355 &uevent->resp, sizeof uevent->resp)) { 356 ret = -EFAULT; 357 goto done; 358 } 359 360 list_del(&uevent->list); 361 uevent->ctx->events_reported++; 362 if (uevent->mc) 363 uevent->mc->events_reported++; 364 kfree(uevent); 365 done: 366 mutex_unlock(&file->mut); 367 return ret; 368 } 369 370 static ssize_t ucma_create_id(struct ucma_file *file, 371 const char __user *inbuf, 372 int in_len, int out_len) 373 { 374 struct rdma_ucm_create_id cmd; 375 struct rdma_ucm_create_id_resp resp; 376 struct ucma_context *ctx; 377 int ret; 378 379 if (out_len < sizeof(resp)) 380 return -ENOSPC; 381 382 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 383 return -EFAULT; 384 385 mutex_lock(&file->mut); 386 ctx = ucma_alloc_ctx(file); 387 mutex_unlock(&file->mut); 388 if (!ctx) 389 return -ENOMEM; 390 391 ctx->uid = cmd.uid; 392 ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps); 393 if (IS_ERR(ctx->cm_id)) { 394 ret = PTR_ERR(ctx->cm_id); 395 goto err1; 396 } 397 398 resp.id = ctx->id; 399 if (copy_to_user((void __user *)(unsigned long)cmd.response, 400 &resp, sizeof(resp))) { 401 ret = -EFAULT; 402 goto err2; 403 } 404 return 0; 405 406 err2: 407 rdma_destroy_id(ctx->cm_id); 408 err1: 409 mutex_lock(&mut); 410 idr_remove(&ctx_idr, ctx->id); 411 mutex_unlock(&mut); 412 kfree(ctx); 413 return ret; 414 } 415 416 static void ucma_cleanup_multicast(struct ucma_context *ctx) 417 { 418 struct ucma_multicast *mc, *tmp; 419 420 mutex_lock(&mut); 421 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) { 422 list_del(&mc->list); 423 idr_remove(&multicast_idr, mc->id); 424 kfree(mc); 425 } 426 mutex_unlock(&mut); 427 } 428 429 static void ucma_cleanup_events(struct ucma_context *ctx) 430 { 431 struct ucma_event *uevent, *tmp; 432 433 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) { 434 if (uevent->ctx != ctx) 435 continue; 436 437 list_del(&uevent->list); 438 439 /* clear incoming connections. */ 440 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) 441 rdma_destroy_id(uevent->cm_id); 442 443 kfree(uevent); 444 } 445 } 446 447 static void ucma_cleanup_mc_events(struct ucma_multicast *mc) 448 { 449 struct ucma_event *uevent, *tmp; 450 451 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) { 452 if (uevent->mc != mc) 453 continue; 454 455 list_del(&uevent->list); 456 kfree(uevent); 457 } 458 } 459 460 static int ucma_free_ctx(struct ucma_context *ctx) 461 { 462 int events_reported; 463 464 /* No new events will be generated after destroying the id. */ 465 rdma_destroy_id(ctx->cm_id); 466 467 ucma_cleanup_multicast(ctx); 468 469 /* Cleanup events not yet reported to the user. */ 470 mutex_lock(&ctx->file->mut); 471 ucma_cleanup_events(ctx); 472 list_del(&ctx->list); 473 mutex_unlock(&ctx->file->mut); 474 475 events_reported = ctx->events_reported; 476 kfree(ctx); 477 return events_reported; 478 } 479 480 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf, 481 int in_len, int out_len) 482 { 483 struct rdma_ucm_destroy_id cmd; 484 struct rdma_ucm_destroy_id_resp resp; 485 struct ucma_context *ctx; 486 int ret = 0; 487 488 if (out_len < sizeof(resp)) 489 return -ENOSPC; 490 491 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 492 return -EFAULT; 493 494 mutex_lock(&mut); 495 ctx = _ucma_find_context(cmd.id, file); 496 if (!IS_ERR(ctx)) 497 idr_remove(&ctx_idr, ctx->id); 498 mutex_unlock(&mut); 499 500 if (IS_ERR(ctx)) 501 return PTR_ERR(ctx); 502 503 ucma_put_ctx(ctx); 504 wait_for_completion(&ctx->comp); 505 resp.events_reported = ucma_free_ctx(ctx); 506 507 if (copy_to_user((void __user *)(unsigned long)cmd.response, 508 &resp, sizeof(resp))) 509 ret = -EFAULT; 510 511 return ret; 512 } 513 514 static ssize_t ucma_bind_addr(struct ucma_file *file, const char __user *inbuf, 515 int in_len, int out_len) 516 { 517 struct rdma_ucm_bind_addr cmd; 518 struct ucma_context *ctx; 519 int ret; 520 521 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 522 return -EFAULT; 523 524 ctx = ucma_get_ctx(file, cmd.id); 525 if (IS_ERR(ctx)) 526 return PTR_ERR(ctx); 527 528 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr); 529 ucma_put_ctx(ctx); 530 return ret; 531 } 532 533 static ssize_t ucma_resolve_addr(struct ucma_file *file, 534 const char __user *inbuf, 535 int in_len, int out_len) 536 { 537 struct rdma_ucm_resolve_addr cmd; 538 struct ucma_context *ctx; 539 int ret; 540 541 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 542 return -EFAULT; 543 544 ctx = ucma_get_ctx(file, cmd.id); 545 if (IS_ERR(ctx)) 546 return PTR_ERR(ctx); 547 548 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr, 549 (struct sockaddr *) &cmd.dst_addr, 550 cmd.timeout_ms); 551 ucma_put_ctx(ctx); 552 return ret; 553 } 554 555 static ssize_t ucma_resolve_route(struct ucma_file *file, 556 const char __user *inbuf, 557 int in_len, int out_len) 558 { 559 struct rdma_ucm_resolve_route cmd; 560 struct ucma_context *ctx; 561 int ret; 562 563 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 564 return -EFAULT; 565 566 ctx = ucma_get_ctx(file, cmd.id); 567 if (IS_ERR(ctx)) 568 return PTR_ERR(ctx); 569 570 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms); 571 ucma_put_ctx(ctx); 572 return ret; 573 } 574 575 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp, 576 struct rdma_route *route) 577 { 578 struct rdma_dev_addr *dev_addr; 579 580 resp->num_paths = route->num_paths; 581 switch (route->num_paths) { 582 case 0: 583 dev_addr = &route->addr.dev_addr; 584 rdma_addr_get_dgid(dev_addr, 585 (union ib_gid *) &resp->ib_route[0].dgid); 586 rdma_addr_get_sgid(dev_addr, 587 (union ib_gid *) &resp->ib_route[0].sgid); 588 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 589 break; 590 case 2: 591 ib_copy_path_rec_to_user(&resp->ib_route[1], 592 &route->path_rec[1]); 593 /* fall through */ 594 case 1: 595 ib_copy_path_rec_to_user(&resp->ib_route[0], 596 &route->path_rec[0]); 597 break; 598 default: 599 break; 600 } 601 } 602 603 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp, 604 struct rdma_route *route) 605 { 606 struct rdma_dev_addr *dev_addr; 607 struct net_device *dev; 608 u16 vid = 0; 609 610 resp->num_paths = route->num_paths; 611 switch (route->num_paths) { 612 case 0: 613 dev_addr = &route->addr.dev_addr; 614 dev = dev_get_by_index(&init_net, dev_addr->bound_dev_if); 615 if (dev) { 616 vid = rdma_vlan_dev_vlan_id(dev); 617 dev_put(dev); 618 } 619 620 iboe_mac_vlan_to_ll((union ib_gid *) &resp->ib_route[0].dgid, 621 dev_addr->dst_dev_addr, vid); 622 iboe_addr_get_sgid(dev_addr, 623 (union ib_gid *) &resp->ib_route[0].sgid); 624 resp->ib_route[0].pkey = cpu_to_be16(0xffff); 625 break; 626 case 2: 627 ib_copy_path_rec_to_user(&resp->ib_route[1], 628 &route->path_rec[1]); 629 /* fall through */ 630 case 1: 631 ib_copy_path_rec_to_user(&resp->ib_route[0], 632 &route->path_rec[0]); 633 break; 634 default: 635 break; 636 } 637 } 638 639 static ssize_t ucma_query_route(struct ucma_file *file, 640 const char __user *inbuf, 641 int in_len, int out_len) 642 { 643 struct rdma_ucm_query_route cmd; 644 struct rdma_ucm_query_route_resp resp; 645 struct ucma_context *ctx; 646 struct sockaddr *addr; 647 int ret = 0; 648 649 if (out_len < sizeof(resp)) 650 return -ENOSPC; 651 652 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 653 return -EFAULT; 654 655 ctx = ucma_get_ctx(file, cmd.id); 656 if (IS_ERR(ctx)) 657 return PTR_ERR(ctx); 658 659 memset(&resp, 0, sizeof resp); 660 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr; 661 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ? 662 sizeof(struct sockaddr_in) : 663 sizeof(struct sockaddr_in6)); 664 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr; 665 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ? 666 sizeof(struct sockaddr_in) : 667 sizeof(struct sockaddr_in6)); 668 if (!ctx->cm_id->device) 669 goto out; 670 671 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid; 672 resp.port_num = ctx->cm_id->port_num; 673 if (rdma_node_get_transport(ctx->cm_id->device->node_type) == RDMA_TRANSPORT_IB) { 674 switch (rdma_port_get_link_layer(ctx->cm_id->device, ctx->cm_id->port_num)) { 675 case IB_LINK_LAYER_INFINIBAND: 676 ucma_copy_ib_route(&resp, &ctx->cm_id->route); 677 break; 678 case IB_LINK_LAYER_ETHERNET: 679 ucma_copy_iboe_route(&resp, &ctx->cm_id->route); 680 break; 681 default: 682 break; 683 } 684 } 685 686 out: 687 if (copy_to_user((void __user *)(unsigned long)cmd.response, 688 &resp, sizeof(resp))) 689 ret = -EFAULT; 690 691 ucma_put_ctx(ctx); 692 return ret; 693 } 694 695 static void ucma_copy_conn_param(struct rdma_conn_param *dst, 696 struct rdma_ucm_conn_param *src) 697 { 698 dst->private_data = src->private_data; 699 dst->private_data_len = src->private_data_len; 700 dst->responder_resources =src->responder_resources; 701 dst->initiator_depth = src->initiator_depth; 702 dst->flow_control = src->flow_control; 703 dst->retry_count = src->retry_count; 704 dst->rnr_retry_count = src->rnr_retry_count; 705 dst->srq = src->srq; 706 dst->qp_num = src->qp_num; 707 } 708 709 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf, 710 int in_len, int out_len) 711 { 712 struct rdma_ucm_connect cmd; 713 struct rdma_conn_param conn_param; 714 struct ucma_context *ctx; 715 int ret; 716 717 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 718 return -EFAULT; 719 720 if (!cmd.conn_param.valid) 721 return -EINVAL; 722 723 ctx = ucma_get_ctx(file, cmd.id); 724 if (IS_ERR(ctx)) 725 return PTR_ERR(ctx); 726 727 ucma_copy_conn_param(&conn_param, &cmd.conn_param); 728 ret = rdma_connect(ctx->cm_id, &conn_param); 729 ucma_put_ctx(ctx); 730 return ret; 731 } 732 733 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf, 734 int in_len, int out_len) 735 { 736 struct rdma_ucm_listen cmd; 737 struct ucma_context *ctx; 738 int ret; 739 740 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 741 return -EFAULT; 742 743 ctx = ucma_get_ctx(file, cmd.id); 744 if (IS_ERR(ctx)) 745 return PTR_ERR(ctx); 746 747 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ? 748 cmd.backlog : max_backlog; 749 ret = rdma_listen(ctx->cm_id, ctx->backlog); 750 ucma_put_ctx(ctx); 751 return ret; 752 } 753 754 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf, 755 int in_len, int out_len) 756 { 757 struct rdma_ucm_accept cmd; 758 struct rdma_conn_param conn_param; 759 struct ucma_context *ctx; 760 int ret; 761 762 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 763 return -EFAULT; 764 765 ctx = ucma_get_ctx(file, cmd.id); 766 if (IS_ERR(ctx)) 767 return PTR_ERR(ctx); 768 769 if (cmd.conn_param.valid) { 770 ctx->uid = cmd.uid; 771 ucma_copy_conn_param(&conn_param, &cmd.conn_param); 772 ret = rdma_accept(ctx->cm_id, &conn_param); 773 } else 774 ret = rdma_accept(ctx->cm_id, NULL); 775 776 ucma_put_ctx(ctx); 777 return ret; 778 } 779 780 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf, 781 int in_len, int out_len) 782 { 783 struct rdma_ucm_reject cmd; 784 struct ucma_context *ctx; 785 int ret; 786 787 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 788 return -EFAULT; 789 790 ctx = ucma_get_ctx(file, cmd.id); 791 if (IS_ERR(ctx)) 792 return PTR_ERR(ctx); 793 794 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len); 795 ucma_put_ctx(ctx); 796 return ret; 797 } 798 799 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf, 800 int in_len, int out_len) 801 { 802 struct rdma_ucm_disconnect cmd; 803 struct ucma_context *ctx; 804 int ret; 805 806 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 807 return -EFAULT; 808 809 ctx = ucma_get_ctx(file, cmd.id); 810 if (IS_ERR(ctx)) 811 return PTR_ERR(ctx); 812 813 ret = rdma_disconnect(ctx->cm_id); 814 ucma_put_ctx(ctx); 815 return ret; 816 } 817 818 static ssize_t ucma_init_qp_attr(struct ucma_file *file, 819 const char __user *inbuf, 820 int in_len, int out_len) 821 { 822 struct rdma_ucm_init_qp_attr cmd; 823 struct ib_uverbs_qp_attr resp; 824 struct ucma_context *ctx; 825 struct ib_qp_attr qp_attr; 826 int ret; 827 828 if (out_len < sizeof(resp)) 829 return -ENOSPC; 830 831 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 832 return -EFAULT; 833 834 ctx = ucma_get_ctx(file, cmd.id); 835 if (IS_ERR(ctx)) 836 return PTR_ERR(ctx); 837 838 resp.qp_attr_mask = 0; 839 memset(&qp_attr, 0, sizeof qp_attr); 840 qp_attr.qp_state = cmd.qp_state; 841 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask); 842 if (ret) 843 goto out; 844 845 ib_copy_qp_attr_to_user(&resp, &qp_attr); 846 if (copy_to_user((void __user *)(unsigned long)cmd.response, 847 &resp, sizeof(resp))) 848 ret = -EFAULT; 849 850 out: 851 ucma_put_ctx(ctx); 852 return ret; 853 } 854 855 static int ucma_set_option_id(struct ucma_context *ctx, int optname, 856 void *optval, size_t optlen) 857 { 858 int ret = 0; 859 860 switch (optname) { 861 case RDMA_OPTION_ID_TOS: 862 if (optlen != sizeof(u8)) { 863 ret = -EINVAL; 864 break; 865 } 866 rdma_set_service_type(ctx->cm_id, *((u8 *) optval)); 867 break; 868 default: 869 ret = -ENOSYS; 870 } 871 872 return ret; 873 } 874 875 static int ucma_set_ib_path(struct ucma_context *ctx, 876 struct ib_path_rec_data *path_data, size_t optlen) 877 { 878 struct ib_sa_path_rec sa_path; 879 struct rdma_cm_event event; 880 int ret; 881 882 if (optlen % sizeof(*path_data)) 883 return -EINVAL; 884 885 for (; optlen; optlen -= sizeof(*path_data), path_data++) { 886 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY | 887 IB_PATH_BIDIRECTIONAL)) 888 break; 889 } 890 891 if (!optlen) 892 return -EINVAL; 893 894 ib_sa_unpack_path(path_data->path_rec, &sa_path); 895 ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1); 896 if (ret) 897 return ret; 898 899 memset(&event, 0, sizeof event); 900 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 901 return ucma_event_handler(ctx->cm_id, &event); 902 } 903 904 static int ucma_set_option_ib(struct ucma_context *ctx, int optname, 905 void *optval, size_t optlen) 906 { 907 int ret; 908 909 switch (optname) { 910 case RDMA_OPTION_IB_PATH: 911 ret = ucma_set_ib_path(ctx, optval, optlen); 912 break; 913 default: 914 ret = -ENOSYS; 915 } 916 917 return ret; 918 } 919 920 static int ucma_set_option_level(struct ucma_context *ctx, int level, 921 int optname, void *optval, size_t optlen) 922 { 923 int ret; 924 925 switch (level) { 926 case RDMA_OPTION_ID: 927 ret = ucma_set_option_id(ctx, optname, optval, optlen); 928 break; 929 case RDMA_OPTION_IB: 930 ret = ucma_set_option_ib(ctx, optname, optval, optlen); 931 break; 932 default: 933 ret = -ENOSYS; 934 } 935 936 return ret; 937 } 938 939 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf, 940 int in_len, int out_len) 941 { 942 struct rdma_ucm_set_option cmd; 943 struct ucma_context *ctx; 944 void *optval; 945 int ret; 946 947 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 948 return -EFAULT; 949 950 ctx = ucma_get_ctx(file, cmd.id); 951 if (IS_ERR(ctx)) 952 return PTR_ERR(ctx); 953 954 optval = kmalloc(cmd.optlen, GFP_KERNEL); 955 if (!optval) { 956 ret = -ENOMEM; 957 goto out1; 958 } 959 960 if (copy_from_user(optval, (void __user *) (unsigned long) cmd.optval, 961 cmd.optlen)) { 962 ret = -EFAULT; 963 goto out2; 964 } 965 966 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval, 967 cmd.optlen); 968 out2: 969 kfree(optval); 970 out1: 971 ucma_put_ctx(ctx); 972 return ret; 973 } 974 975 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf, 976 int in_len, int out_len) 977 { 978 struct rdma_ucm_notify cmd; 979 struct ucma_context *ctx; 980 int ret; 981 982 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 983 return -EFAULT; 984 985 ctx = ucma_get_ctx(file, cmd.id); 986 if (IS_ERR(ctx)) 987 return PTR_ERR(ctx); 988 989 ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event); 990 ucma_put_ctx(ctx); 991 return ret; 992 } 993 994 static ssize_t ucma_join_multicast(struct ucma_file *file, 995 const char __user *inbuf, 996 int in_len, int out_len) 997 { 998 struct rdma_ucm_join_mcast cmd; 999 struct rdma_ucm_create_id_resp resp; 1000 struct ucma_context *ctx; 1001 struct ucma_multicast *mc; 1002 int ret; 1003 1004 if (out_len < sizeof(resp)) 1005 return -ENOSPC; 1006 1007 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1008 return -EFAULT; 1009 1010 ctx = ucma_get_ctx(file, cmd.id); 1011 if (IS_ERR(ctx)) 1012 return PTR_ERR(ctx); 1013 1014 mutex_lock(&file->mut); 1015 mc = ucma_alloc_multicast(ctx); 1016 if (!mc) { 1017 ret = -ENOMEM; 1018 goto err1; 1019 } 1020 1021 mc->uid = cmd.uid; 1022 memcpy(&mc->addr, &cmd.addr, sizeof cmd.addr); 1023 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc); 1024 if (ret) 1025 goto err2; 1026 1027 resp.id = mc->id; 1028 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1029 &resp, sizeof(resp))) { 1030 ret = -EFAULT; 1031 goto err3; 1032 } 1033 1034 mutex_unlock(&file->mut); 1035 ucma_put_ctx(ctx); 1036 return 0; 1037 1038 err3: 1039 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr); 1040 ucma_cleanup_mc_events(mc); 1041 err2: 1042 mutex_lock(&mut); 1043 idr_remove(&multicast_idr, mc->id); 1044 mutex_unlock(&mut); 1045 list_del(&mc->list); 1046 kfree(mc); 1047 err1: 1048 mutex_unlock(&file->mut); 1049 ucma_put_ctx(ctx); 1050 return ret; 1051 } 1052 1053 static ssize_t ucma_leave_multicast(struct ucma_file *file, 1054 const char __user *inbuf, 1055 int in_len, int out_len) 1056 { 1057 struct rdma_ucm_destroy_id cmd; 1058 struct rdma_ucm_destroy_id_resp resp; 1059 struct ucma_multicast *mc; 1060 int ret = 0; 1061 1062 if (out_len < sizeof(resp)) 1063 return -ENOSPC; 1064 1065 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1066 return -EFAULT; 1067 1068 mutex_lock(&mut); 1069 mc = idr_find(&multicast_idr, cmd.id); 1070 if (!mc) 1071 mc = ERR_PTR(-ENOENT); 1072 else if (mc->ctx->file != file) 1073 mc = ERR_PTR(-EINVAL); 1074 else { 1075 idr_remove(&multicast_idr, mc->id); 1076 atomic_inc(&mc->ctx->ref); 1077 } 1078 mutex_unlock(&mut); 1079 1080 if (IS_ERR(mc)) { 1081 ret = PTR_ERR(mc); 1082 goto out; 1083 } 1084 1085 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr); 1086 mutex_lock(&mc->ctx->file->mut); 1087 ucma_cleanup_mc_events(mc); 1088 list_del(&mc->list); 1089 mutex_unlock(&mc->ctx->file->mut); 1090 1091 ucma_put_ctx(mc->ctx); 1092 resp.events_reported = mc->events_reported; 1093 kfree(mc); 1094 1095 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1096 &resp, sizeof(resp))) 1097 ret = -EFAULT; 1098 out: 1099 return ret; 1100 } 1101 1102 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2) 1103 { 1104 /* Acquire mutex's based on pointer comparison to prevent deadlock. */ 1105 if (file1 < file2) { 1106 mutex_lock(&file1->mut); 1107 mutex_lock(&file2->mut); 1108 } else { 1109 mutex_lock(&file2->mut); 1110 mutex_lock(&file1->mut); 1111 } 1112 } 1113 1114 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2) 1115 { 1116 if (file1 < file2) { 1117 mutex_unlock(&file2->mut); 1118 mutex_unlock(&file1->mut); 1119 } else { 1120 mutex_unlock(&file1->mut); 1121 mutex_unlock(&file2->mut); 1122 } 1123 } 1124 1125 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file) 1126 { 1127 struct ucma_event *uevent, *tmp; 1128 1129 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) 1130 if (uevent->ctx == ctx) 1131 list_move_tail(&uevent->list, &file->event_list); 1132 } 1133 1134 static ssize_t ucma_migrate_id(struct ucma_file *new_file, 1135 const char __user *inbuf, 1136 int in_len, int out_len) 1137 { 1138 struct rdma_ucm_migrate_id cmd; 1139 struct rdma_ucm_migrate_resp resp; 1140 struct ucma_context *ctx; 1141 struct file *filp; 1142 struct ucma_file *cur_file; 1143 int ret = 0; 1144 1145 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1146 return -EFAULT; 1147 1148 /* Get current fd to protect against it being closed */ 1149 filp = fget(cmd.fd); 1150 if (!filp) 1151 return -ENOENT; 1152 1153 /* Validate current fd and prevent destruction of id. */ 1154 ctx = ucma_get_ctx(filp->private_data, cmd.id); 1155 if (IS_ERR(ctx)) { 1156 ret = PTR_ERR(ctx); 1157 goto file_put; 1158 } 1159 1160 cur_file = ctx->file; 1161 if (cur_file == new_file) { 1162 resp.events_reported = ctx->events_reported; 1163 goto response; 1164 } 1165 1166 /* 1167 * Migrate events between fd's, maintaining order, and avoiding new 1168 * events being added before existing events. 1169 */ 1170 ucma_lock_files(cur_file, new_file); 1171 mutex_lock(&mut); 1172 1173 list_move_tail(&ctx->list, &new_file->ctx_list); 1174 ucma_move_events(ctx, new_file); 1175 ctx->file = new_file; 1176 resp.events_reported = ctx->events_reported; 1177 1178 mutex_unlock(&mut); 1179 ucma_unlock_files(cur_file, new_file); 1180 1181 response: 1182 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1183 &resp, sizeof(resp))) 1184 ret = -EFAULT; 1185 1186 ucma_put_ctx(ctx); 1187 file_put: 1188 fput(filp); 1189 return ret; 1190 } 1191 1192 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file, 1193 const char __user *inbuf, 1194 int in_len, int out_len) = { 1195 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id, 1196 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id, 1197 [RDMA_USER_CM_CMD_BIND_ADDR] = ucma_bind_addr, 1198 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr, 1199 [RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route, 1200 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route, 1201 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect, 1202 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen, 1203 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept, 1204 [RDMA_USER_CM_CMD_REJECT] = ucma_reject, 1205 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect, 1206 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr, 1207 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event, 1208 [RDMA_USER_CM_CMD_GET_OPTION] = NULL, 1209 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option, 1210 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify, 1211 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast, 1212 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast, 1213 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id 1214 }; 1215 1216 static ssize_t ucma_write(struct file *filp, const char __user *buf, 1217 size_t len, loff_t *pos) 1218 { 1219 struct ucma_file *file = filp->private_data; 1220 struct rdma_ucm_cmd_hdr hdr; 1221 ssize_t ret; 1222 1223 if (len < sizeof(hdr)) 1224 return -EINVAL; 1225 1226 if (copy_from_user(&hdr, buf, sizeof(hdr))) 1227 return -EFAULT; 1228 1229 if (hdr.cmd < 0 || hdr.cmd >= ARRAY_SIZE(ucma_cmd_table)) 1230 return -EINVAL; 1231 1232 if (hdr.in + sizeof(hdr) > len) 1233 return -EINVAL; 1234 1235 if (!ucma_cmd_table[hdr.cmd]) 1236 return -ENOSYS; 1237 1238 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out); 1239 if (!ret) 1240 ret = len; 1241 1242 return ret; 1243 } 1244 1245 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait) 1246 { 1247 struct ucma_file *file = filp->private_data; 1248 unsigned int mask = 0; 1249 1250 poll_wait(filp, &file->poll_wait, wait); 1251 1252 if (!list_empty(&file->event_list)) 1253 mask = POLLIN | POLLRDNORM; 1254 1255 return mask; 1256 } 1257 1258 /* 1259 * ucma_open() does not need the BKL: 1260 * 1261 * - no global state is referred to; 1262 * - there is no ioctl method to race against; 1263 * - no further module initialization is required for open to work 1264 * after the device is registered. 1265 */ 1266 static int ucma_open(struct inode *inode, struct file *filp) 1267 { 1268 struct ucma_file *file; 1269 1270 file = kmalloc(sizeof *file, GFP_KERNEL); 1271 if (!file) 1272 return -ENOMEM; 1273 1274 INIT_LIST_HEAD(&file->event_list); 1275 INIT_LIST_HEAD(&file->ctx_list); 1276 init_waitqueue_head(&file->poll_wait); 1277 mutex_init(&file->mut); 1278 1279 filp->private_data = file; 1280 file->filp = filp; 1281 1282 return nonseekable_open(inode, filp); 1283 } 1284 1285 static int ucma_close(struct inode *inode, struct file *filp) 1286 { 1287 struct ucma_file *file = filp->private_data; 1288 struct ucma_context *ctx, *tmp; 1289 1290 mutex_lock(&file->mut); 1291 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) { 1292 mutex_unlock(&file->mut); 1293 1294 mutex_lock(&mut); 1295 idr_remove(&ctx_idr, ctx->id); 1296 mutex_unlock(&mut); 1297 1298 ucma_free_ctx(ctx); 1299 mutex_lock(&file->mut); 1300 } 1301 mutex_unlock(&file->mut); 1302 kfree(file); 1303 return 0; 1304 } 1305 1306 static const struct file_operations ucma_fops = { 1307 .owner = THIS_MODULE, 1308 .open = ucma_open, 1309 .release = ucma_close, 1310 .write = ucma_write, 1311 .poll = ucma_poll, 1312 .llseek = no_llseek, 1313 }; 1314 1315 static struct miscdevice ucma_misc = { 1316 .minor = MISC_DYNAMIC_MINOR, 1317 .name = "rdma_cm", 1318 .fops = &ucma_fops, 1319 }; 1320 1321 static ssize_t show_abi_version(struct device *dev, 1322 struct device_attribute *attr, 1323 char *buf) 1324 { 1325 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION); 1326 } 1327 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL); 1328 1329 static int __init ucma_init(void) 1330 { 1331 int ret; 1332 1333 ret = misc_register(&ucma_misc); 1334 if (ret) 1335 return ret; 1336 1337 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version); 1338 if (ret) { 1339 printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n"); 1340 goto err1; 1341 } 1342 1343 ucma_ctl_table_hdr = register_sysctl_paths(ucma_ctl_path, ucma_ctl_table); 1344 if (!ucma_ctl_table_hdr) { 1345 printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n"); 1346 ret = -ENOMEM; 1347 goto err2; 1348 } 1349 return 0; 1350 err2: 1351 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version); 1352 err1: 1353 misc_deregister(&ucma_misc); 1354 return ret; 1355 } 1356 1357 static void __exit ucma_cleanup(void) 1358 { 1359 unregister_sysctl_table(ucma_ctl_table_hdr); 1360 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version); 1361 misc_deregister(&ucma_misc); 1362 idr_destroy(&ctx_idr); 1363 } 1364 1365 module_init(ucma_init); 1366 module_exit(ucma_cleanup); 1367