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 void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp, 640 struct rdma_route *route) 641 { 642 struct rdma_dev_addr *dev_addr; 643 644 dev_addr = &route->addr.dev_addr; 645 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid); 646 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid); 647 } 648 649 static ssize_t ucma_query_route(struct ucma_file *file, 650 const char __user *inbuf, 651 int in_len, int out_len) 652 { 653 struct rdma_ucm_query_route cmd; 654 struct rdma_ucm_query_route_resp resp; 655 struct ucma_context *ctx; 656 struct sockaddr *addr; 657 int ret = 0; 658 659 if (out_len < sizeof(resp)) 660 return -ENOSPC; 661 662 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 663 return -EFAULT; 664 665 ctx = ucma_get_ctx(file, cmd.id); 666 if (IS_ERR(ctx)) 667 return PTR_ERR(ctx); 668 669 memset(&resp, 0, sizeof resp); 670 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr; 671 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ? 672 sizeof(struct sockaddr_in) : 673 sizeof(struct sockaddr_in6)); 674 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr; 675 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ? 676 sizeof(struct sockaddr_in) : 677 sizeof(struct sockaddr_in6)); 678 if (!ctx->cm_id->device) 679 goto out; 680 681 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid; 682 resp.port_num = ctx->cm_id->port_num; 683 switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) { 684 case RDMA_TRANSPORT_IB: 685 switch (rdma_port_get_link_layer(ctx->cm_id->device, 686 ctx->cm_id->port_num)) { 687 case IB_LINK_LAYER_INFINIBAND: 688 ucma_copy_ib_route(&resp, &ctx->cm_id->route); 689 break; 690 case IB_LINK_LAYER_ETHERNET: 691 ucma_copy_iboe_route(&resp, &ctx->cm_id->route); 692 break; 693 default: 694 break; 695 } 696 break; 697 case RDMA_TRANSPORT_IWARP: 698 ucma_copy_iw_route(&resp, &ctx->cm_id->route); 699 break; 700 default: 701 break; 702 } 703 704 out: 705 if (copy_to_user((void __user *)(unsigned long)cmd.response, 706 &resp, sizeof(resp))) 707 ret = -EFAULT; 708 709 ucma_put_ctx(ctx); 710 return ret; 711 } 712 713 static void ucma_copy_conn_param(struct rdma_conn_param *dst, 714 struct rdma_ucm_conn_param *src) 715 { 716 dst->private_data = src->private_data; 717 dst->private_data_len = src->private_data_len; 718 dst->responder_resources =src->responder_resources; 719 dst->initiator_depth = src->initiator_depth; 720 dst->flow_control = src->flow_control; 721 dst->retry_count = src->retry_count; 722 dst->rnr_retry_count = src->rnr_retry_count; 723 dst->srq = src->srq; 724 dst->qp_num = src->qp_num; 725 } 726 727 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf, 728 int in_len, int out_len) 729 { 730 struct rdma_ucm_connect cmd; 731 struct rdma_conn_param conn_param; 732 struct ucma_context *ctx; 733 int ret; 734 735 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 736 return -EFAULT; 737 738 if (!cmd.conn_param.valid) 739 return -EINVAL; 740 741 ctx = ucma_get_ctx(file, cmd.id); 742 if (IS_ERR(ctx)) 743 return PTR_ERR(ctx); 744 745 ucma_copy_conn_param(&conn_param, &cmd.conn_param); 746 ret = rdma_connect(ctx->cm_id, &conn_param); 747 ucma_put_ctx(ctx); 748 return ret; 749 } 750 751 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf, 752 int in_len, int out_len) 753 { 754 struct rdma_ucm_listen cmd; 755 struct ucma_context *ctx; 756 int ret; 757 758 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 759 return -EFAULT; 760 761 ctx = ucma_get_ctx(file, cmd.id); 762 if (IS_ERR(ctx)) 763 return PTR_ERR(ctx); 764 765 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ? 766 cmd.backlog : max_backlog; 767 ret = rdma_listen(ctx->cm_id, ctx->backlog); 768 ucma_put_ctx(ctx); 769 return ret; 770 } 771 772 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf, 773 int in_len, int out_len) 774 { 775 struct rdma_ucm_accept cmd; 776 struct rdma_conn_param conn_param; 777 struct ucma_context *ctx; 778 int ret; 779 780 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 781 return -EFAULT; 782 783 ctx = ucma_get_ctx(file, cmd.id); 784 if (IS_ERR(ctx)) 785 return PTR_ERR(ctx); 786 787 if (cmd.conn_param.valid) { 788 ctx->uid = cmd.uid; 789 ucma_copy_conn_param(&conn_param, &cmd.conn_param); 790 ret = rdma_accept(ctx->cm_id, &conn_param); 791 } else 792 ret = rdma_accept(ctx->cm_id, NULL); 793 794 ucma_put_ctx(ctx); 795 return ret; 796 } 797 798 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf, 799 int in_len, int out_len) 800 { 801 struct rdma_ucm_reject cmd; 802 struct ucma_context *ctx; 803 int ret; 804 805 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 806 return -EFAULT; 807 808 ctx = ucma_get_ctx(file, cmd.id); 809 if (IS_ERR(ctx)) 810 return PTR_ERR(ctx); 811 812 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len); 813 ucma_put_ctx(ctx); 814 return ret; 815 } 816 817 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf, 818 int in_len, int out_len) 819 { 820 struct rdma_ucm_disconnect cmd; 821 struct ucma_context *ctx; 822 int ret; 823 824 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 825 return -EFAULT; 826 827 ctx = ucma_get_ctx(file, cmd.id); 828 if (IS_ERR(ctx)) 829 return PTR_ERR(ctx); 830 831 ret = rdma_disconnect(ctx->cm_id); 832 ucma_put_ctx(ctx); 833 return ret; 834 } 835 836 static ssize_t ucma_init_qp_attr(struct ucma_file *file, 837 const char __user *inbuf, 838 int in_len, int out_len) 839 { 840 struct rdma_ucm_init_qp_attr cmd; 841 struct ib_uverbs_qp_attr resp; 842 struct ucma_context *ctx; 843 struct ib_qp_attr qp_attr; 844 int ret; 845 846 if (out_len < sizeof(resp)) 847 return -ENOSPC; 848 849 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 850 return -EFAULT; 851 852 ctx = ucma_get_ctx(file, cmd.id); 853 if (IS_ERR(ctx)) 854 return PTR_ERR(ctx); 855 856 resp.qp_attr_mask = 0; 857 memset(&qp_attr, 0, sizeof qp_attr); 858 qp_attr.qp_state = cmd.qp_state; 859 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask); 860 if (ret) 861 goto out; 862 863 ib_copy_qp_attr_to_user(&resp, &qp_attr); 864 if (copy_to_user((void __user *)(unsigned long)cmd.response, 865 &resp, sizeof(resp))) 866 ret = -EFAULT; 867 868 out: 869 ucma_put_ctx(ctx); 870 return ret; 871 } 872 873 static int ucma_set_option_id(struct ucma_context *ctx, int optname, 874 void *optval, size_t optlen) 875 { 876 int ret = 0; 877 878 switch (optname) { 879 case RDMA_OPTION_ID_TOS: 880 if (optlen != sizeof(u8)) { 881 ret = -EINVAL; 882 break; 883 } 884 rdma_set_service_type(ctx->cm_id, *((u8 *) optval)); 885 break; 886 case RDMA_OPTION_ID_REUSEADDR: 887 if (optlen != sizeof(int)) { 888 ret = -EINVAL; 889 break; 890 } 891 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0); 892 break; 893 default: 894 ret = -ENOSYS; 895 } 896 897 return ret; 898 } 899 900 static int ucma_set_ib_path(struct ucma_context *ctx, 901 struct ib_path_rec_data *path_data, size_t optlen) 902 { 903 struct ib_sa_path_rec sa_path; 904 struct rdma_cm_event event; 905 int ret; 906 907 if (optlen % sizeof(*path_data)) 908 return -EINVAL; 909 910 for (; optlen; optlen -= sizeof(*path_data), path_data++) { 911 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY | 912 IB_PATH_BIDIRECTIONAL)) 913 break; 914 } 915 916 if (!optlen) 917 return -EINVAL; 918 919 ib_sa_unpack_path(path_data->path_rec, &sa_path); 920 ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1); 921 if (ret) 922 return ret; 923 924 memset(&event, 0, sizeof event); 925 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 926 return ucma_event_handler(ctx->cm_id, &event); 927 } 928 929 static int ucma_set_option_ib(struct ucma_context *ctx, int optname, 930 void *optval, size_t optlen) 931 { 932 int ret; 933 934 switch (optname) { 935 case RDMA_OPTION_IB_PATH: 936 ret = ucma_set_ib_path(ctx, optval, optlen); 937 break; 938 default: 939 ret = -ENOSYS; 940 } 941 942 return ret; 943 } 944 945 static int ucma_set_option_level(struct ucma_context *ctx, int level, 946 int optname, void *optval, size_t optlen) 947 { 948 int ret; 949 950 switch (level) { 951 case RDMA_OPTION_ID: 952 ret = ucma_set_option_id(ctx, optname, optval, optlen); 953 break; 954 case RDMA_OPTION_IB: 955 ret = ucma_set_option_ib(ctx, optname, optval, optlen); 956 break; 957 default: 958 ret = -ENOSYS; 959 } 960 961 return ret; 962 } 963 964 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf, 965 int in_len, int out_len) 966 { 967 struct rdma_ucm_set_option cmd; 968 struct ucma_context *ctx; 969 void *optval; 970 int ret; 971 972 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 973 return -EFAULT; 974 975 ctx = ucma_get_ctx(file, cmd.id); 976 if (IS_ERR(ctx)) 977 return PTR_ERR(ctx); 978 979 optval = kmalloc(cmd.optlen, GFP_KERNEL); 980 if (!optval) { 981 ret = -ENOMEM; 982 goto out1; 983 } 984 985 if (copy_from_user(optval, (void __user *) (unsigned long) cmd.optval, 986 cmd.optlen)) { 987 ret = -EFAULT; 988 goto out2; 989 } 990 991 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval, 992 cmd.optlen); 993 out2: 994 kfree(optval); 995 out1: 996 ucma_put_ctx(ctx); 997 return ret; 998 } 999 1000 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf, 1001 int in_len, int out_len) 1002 { 1003 struct rdma_ucm_notify cmd; 1004 struct ucma_context *ctx; 1005 int ret; 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 ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event); 1015 ucma_put_ctx(ctx); 1016 return ret; 1017 } 1018 1019 static ssize_t ucma_join_multicast(struct ucma_file *file, 1020 const char __user *inbuf, 1021 int in_len, int out_len) 1022 { 1023 struct rdma_ucm_join_mcast cmd; 1024 struct rdma_ucm_create_id_resp resp; 1025 struct ucma_context *ctx; 1026 struct ucma_multicast *mc; 1027 int ret; 1028 1029 if (out_len < sizeof(resp)) 1030 return -ENOSPC; 1031 1032 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1033 return -EFAULT; 1034 1035 ctx = ucma_get_ctx(file, cmd.id); 1036 if (IS_ERR(ctx)) 1037 return PTR_ERR(ctx); 1038 1039 mutex_lock(&file->mut); 1040 mc = ucma_alloc_multicast(ctx); 1041 if (!mc) { 1042 ret = -ENOMEM; 1043 goto err1; 1044 } 1045 1046 mc->uid = cmd.uid; 1047 memcpy(&mc->addr, &cmd.addr, sizeof cmd.addr); 1048 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc); 1049 if (ret) 1050 goto err2; 1051 1052 resp.id = mc->id; 1053 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1054 &resp, sizeof(resp))) { 1055 ret = -EFAULT; 1056 goto err3; 1057 } 1058 1059 mutex_unlock(&file->mut); 1060 ucma_put_ctx(ctx); 1061 return 0; 1062 1063 err3: 1064 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr); 1065 ucma_cleanup_mc_events(mc); 1066 err2: 1067 mutex_lock(&mut); 1068 idr_remove(&multicast_idr, mc->id); 1069 mutex_unlock(&mut); 1070 list_del(&mc->list); 1071 kfree(mc); 1072 err1: 1073 mutex_unlock(&file->mut); 1074 ucma_put_ctx(ctx); 1075 return ret; 1076 } 1077 1078 static ssize_t ucma_leave_multicast(struct ucma_file *file, 1079 const char __user *inbuf, 1080 int in_len, int out_len) 1081 { 1082 struct rdma_ucm_destroy_id cmd; 1083 struct rdma_ucm_destroy_id_resp resp; 1084 struct ucma_multicast *mc; 1085 int ret = 0; 1086 1087 if (out_len < sizeof(resp)) 1088 return -ENOSPC; 1089 1090 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1091 return -EFAULT; 1092 1093 mutex_lock(&mut); 1094 mc = idr_find(&multicast_idr, cmd.id); 1095 if (!mc) 1096 mc = ERR_PTR(-ENOENT); 1097 else if (mc->ctx->file != file) 1098 mc = ERR_PTR(-EINVAL); 1099 else { 1100 idr_remove(&multicast_idr, mc->id); 1101 atomic_inc(&mc->ctx->ref); 1102 } 1103 mutex_unlock(&mut); 1104 1105 if (IS_ERR(mc)) { 1106 ret = PTR_ERR(mc); 1107 goto out; 1108 } 1109 1110 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr); 1111 mutex_lock(&mc->ctx->file->mut); 1112 ucma_cleanup_mc_events(mc); 1113 list_del(&mc->list); 1114 mutex_unlock(&mc->ctx->file->mut); 1115 1116 ucma_put_ctx(mc->ctx); 1117 resp.events_reported = mc->events_reported; 1118 kfree(mc); 1119 1120 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1121 &resp, sizeof(resp))) 1122 ret = -EFAULT; 1123 out: 1124 return ret; 1125 } 1126 1127 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2) 1128 { 1129 /* Acquire mutex's based on pointer comparison to prevent deadlock. */ 1130 if (file1 < file2) { 1131 mutex_lock(&file1->mut); 1132 mutex_lock(&file2->mut); 1133 } else { 1134 mutex_lock(&file2->mut); 1135 mutex_lock(&file1->mut); 1136 } 1137 } 1138 1139 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2) 1140 { 1141 if (file1 < file2) { 1142 mutex_unlock(&file2->mut); 1143 mutex_unlock(&file1->mut); 1144 } else { 1145 mutex_unlock(&file1->mut); 1146 mutex_unlock(&file2->mut); 1147 } 1148 } 1149 1150 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file) 1151 { 1152 struct ucma_event *uevent, *tmp; 1153 1154 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) 1155 if (uevent->ctx == ctx) 1156 list_move_tail(&uevent->list, &file->event_list); 1157 } 1158 1159 static ssize_t ucma_migrate_id(struct ucma_file *new_file, 1160 const char __user *inbuf, 1161 int in_len, int out_len) 1162 { 1163 struct rdma_ucm_migrate_id cmd; 1164 struct rdma_ucm_migrate_resp resp; 1165 struct ucma_context *ctx; 1166 struct file *filp; 1167 struct ucma_file *cur_file; 1168 int ret = 0; 1169 1170 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1171 return -EFAULT; 1172 1173 /* Get current fd to protect against it being closed */ 1174 filp = fget(cmd.fd); 1175 if (!filp) 1176 return -ENOENT; 1177 1178 /* Validate current fd and prevent destruction of id. */ 1179 ctx = ucma_get_ctx(filp->private_data, cmd.id); 1180 if (IS_ERR(ctx)) { 1181 ret = PTR_ERR(ctx); 1182 goto file_put; 1183 } 1184 1185 cur_file = ctx->file; 1186 if (cur_file == new_file) { 1187 resp.events_reported = ctx->events_reported; 1188 goto response; 1189 } 1190 1191 /* 1192 * Migrate events between fd's, maintaining order, and avoiding new 1193 * events being added before existing events. 1194 */ 1195 ucma_lock_files(cur_file, new_file); 1196 mutex_lock(&mut); 1197 1198 list_move_tail(&ctx->list, &new_file->ctx_list); 1199 ucma_move_events(ctx, new_file); 1200 ctx->file = new_file; 1201 resp.events_reported = ctx->events_reported; 1202 1203 mutex_unlock(&mut); 1204 ucma_unlock_files(cur_file, new_file); 1205 1206 response: 1207 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1208 &resp, sizeof(resp))) 1209 ret = -EFAULT; 1210 1211 ucma_put_ctx(ctx); 1212 file_put: 1213 fput(filp); 1214 return ret; 1215 } 1216 1217 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file, 1218 const char __user *inbuf, 1219 int in_len, int out_len) = { 1220 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id, 1221 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id, 1222 [RDMA_USER_CM_CMD_BIND_ADDR] = ucma_bind_addr, 1223 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr, 1224 [RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route, 1225 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route, 1226 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect, 1227 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen, 1228 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept, 1229 [RDMA_USER_CM_CMD_REJECT] = ucma_reject, 1230 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect, 1231 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr, 1232 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event, 1233 [RDMA_USER_CM_CMD_GET_OPTION] = NULL, 1234 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option, 1235 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify, 1236 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast, 1237 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast, 1238 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id 1239 }; 1240 1241 static ssize_t ucma_write(struct file *filp, const char __user *buf, 1242 size_t len, loff_t *pos) 1243 { 1244 struct ucma_file *file = filp->private_data; 1245 struct rdma_ucm_cmd_hdr hdr; 1246 ssize_t ret; 1247 1248 if (len < sizeof(hdr)) 1249 return -EINVAL; 1250 1251 if (copy_from_user(&hdr, buf, sizeof(hdr))) 1252 return -EFAULT; 1253 1254 if (hdr.cmd < 0 || hdr.cmd >= ARRAY_SIZE(ucma_cmd_table)) 1255 return -EINVAL; 1256 1257 if (hdr.in + sizeof(hdr) > len) 1258 return -EINVAL; 1259 1260 if (!ucma_cmd_table[hdr.cmd]) 1261 return -ENOSYS; 1262 1263 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out); 1264 if (!ret) 1265 ret = len; 1266 1267 return ret; 1268 } 1269 1270 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait) 1271 { 1272 struct ucma_file *file = filp->private_data; 1273 unsigned int mask = 0; 1274 1275 poll_wait(filp, &file->poll_wait, wait); 1276 1277 if (!list_empty(&file->event_list)) 1278 mask = POLLIN | POLLRDNORM; 1279 1280 return mask; 1281 } 1282 1283 /* 1284 * ucma_open() does not need the BKL: 1285 * 1286 * - no global state is referred to; 1287 * - there is no ioctl method to race against; 1288 * - no further module initialization is required for open to work 1289 * after the device is registered. 1290 */ 1291 static int ucma_open(struct inode *inode, struct file *filp) 1292 { 1293 struct ucma_file *file; 1294 1295 file = kmalloc(sizeof *file, GFP_KERNEL); 1296 if (!file) 1297 return -ENOMEM; 1298 1299 INIT_LIST_HEAD(&file->event_list); 1300 INIT_LIST_HEAD(&file->ctx_list); 1301 init_waitqueue_head(&file->poll_wait); 1302 mutex_init(&file->mut); 1303 1304 filp->private_data = file; 1305 file->filp = filp; 1306 1307 return nonseekable_open(inode, filp); 1308 } 1309 1310 static int ucma_close(struct inode *inode, struct file *filp) 1311 { 1312 struct ucma_file *file = filp->private_data; 1313 struct ucma_context *ctx, *tmp; 1314 1315 mutex_lock(&file->mut); 1316 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) { 1317 mutex_unlock(&file->mut); 1318 1319 mutex_lock(&mut); 1320 idr_remove(&ctx_idr, ctx->id); 1321 mutex_unlock(&mut); 1322 1323 ucma_free_ctx(ctx); 1324 mutex_lock(&file->mut); 1325 } 1326 mutex_unlock(&file->mut); 1327 kfree(file); 1328 return 0; 1329 } 1330 1331 static const struct file_operations ucma_fops = { 1332 .owner = THIS_MODULE, 1333 .open = ucma_open, 1334 .release = ucma_close, 1335 .write = ucma_write, 1336 .poll = ucma_poll, 1337 .llseek = no_llseek, 1338 }; 1339 1340 static struct miscdevice ucma_misc = { 1341 .minor = MISC_DYNAMIC_MINOR, 1342 .name = "rdma_cm", 1343 .fops = &ucma_fops, 1344 }; 1345 1346 static ssize_t show_abi_version(struct device *dev, 1347 struct device_attribute *attr, 1348 char *buf) 1349 { 1350 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION); 1351 } 1352 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL); 1353 1354 static int __init ucma_init(void) 1355 { 1356 int ret; 1357 1358 ret = misc_register(&ucma_misc); 1359 if (ret) 1360 return ret; 1361 1362 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version); 1363 if (ret) { 1364 printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n"); 1365 goto err1; 1366 } 1367 1368 ucma_ctl_table_hdr = register_sysctl_paths(ucma_ctl_path, ucma_ctl_table); 1369 if (!ucma_ctl_table_hdr) { 1370 printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n"); 1371 ret = -ENOMEM; 1372 goto err2; 1373 } 1374 return 0; 1375 err2: 1376 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version); 1377 err1: 1378 misc_deregister(&ucma_misc); 1379 return ret; 1380 } 1381 1382 static void __exit ucma_cleanup(void) 1383 { 1384 unregister_sysctl_table(ucma_ctl_table_hdr); 1385 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version); 1386 misc_deregister(&ucma_misc); 1387 idr_destroy(&ctx_idr); 1388 } 1389 1390 module_init(ucma_init); 1391 module_exit(ucma_cleanup); 1392