1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0 3 * 4 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved. 5 * 6 * This software is available to you under a choice of one of two 7 * licenses. You may choose to be licensed under the terms of the GNU 8 * General Public License (GPL) Version 2, available from the file 9 * COPYING in the main directory of this source tree, or the 10 * OpenIB.org BSD license below: 11 * 12 * Redistribution and use in source and binary forms, with or 13 * without modification, are permitted provided that the following 14 * conditions are met: 15 * 16 * - Redistributions of source code must retain the above 17 * copyright notice, this list of conditions and the following 18 * disclaimer. 19 * 20 * - Redistributions in binary form must reproduce the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer in the documentation and/or other materials 23 * provided with the distribution. 24 * 25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 32 * SOFTWARE. 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include <linux/completion.h> 39 #include <linux/file.h> 40 #include <linux/mutex.h> 41 #include <linux/poll.h> 42 #include <linux/sched.h> 43 #include <linux/idr.h> 44 #include <linux/in.h> 45 #include <linux/in6.h> 46 #include <linux/miscdevice.h> 47 #include <linux/slab.h> 48 #include <linux/module.h> 49 50 #include <sys/filio.h> 51 52 #include <rdma/rdma_user_cm.h> 53 #include <rdma/ib_marshall.h> 54 #include <rdma/rdma_cm.h> 55 #include <rdma/rdma_cm_ib.h> 56 #include <rdma/ib_addr.h> 57 #include <rdma/ib.h> 58 59 MODULE_AUTHOR("Sean Hefty"); 60 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access"); 61 MODULE_LICENSE("Dual BSD/GPL"); 62 63 static unsigned int max_backlog = 1024; 64 65 struct ucma_file { 66 struct mutex mut; 67 struct file *filp; 68 struct list_head ctx_list; 69 struct list_head event_list; 70 wait_queue_head_t poll_wait; 71 struct workqueue_struct *close_wq; 72 }; 73 74 struct ucma_context { 75 int id; 76 struct completion comp; 77 atomic_t ref; 78 int events_reported; 79 int backlog; 80 81 struct ucma_file *file; 82 struct rdma_cm_id *cm_id; 83 u64 uid; 84 85 struct list_head list; 86 struct list_head mc_list; 87 /* mark that device is in process of destroying the internal HW 88 * resources, protected by the global mut 89 */ 90 int closing; 91 /* sync between removal event and id destroy, protected by file mut */ 92 int destroying; 93 struct work_struct close_work; 94 }; 95 96 struct ucma_multicast { 97 struct ucma_context *ctx; 98 int id; 99 int events_reported; 100 101 u64 uid; 102 u8 join_state; 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 struct work_struct close_work; 114 }; 115 116 static DEFINE_MUTEX(mut); 117 static DEFINE_IDR(ctx_idr); 118 static DEFINE_IDR(multicast_idr); 119 120 static inline struct ucma_context *_ucma_find_context(int id, 121 struct ucma_file *file) 122 { 123 struct ucma_context *ctx; 124 125 ctx = idr_find(&ctx_idr, id); 126 if (!ctx) 127 ctx = ERR_PTR(-ENOENT); 128 else if (ctx->file != file || !ctx->cm_id) 129 ctx = ERR_PTR(-EINVAL); 130 return ctx; 131 } 132 133 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id) 134 { 135 struct ucma_context *ctx; 136 137 mutex_lock(&mut); 138 ctx = _ucma_find_context(id, file); 139 if (!IS_ERR(ctx)) { 140 if (ctx->closing) 141 ctx = ERR_PTR(-EIO); 142 else 143 atomic_inc(&ctx->ref); 144 } 145 mutex_unlock(&mut); 146 return ctx; 147 } 148 149 static void ucma_put_ctx(struct ucma_context *ctx) 150 { 151 if (atomic_dec_and_test(&ctx->ref)) 152 complete(&ctx->comp); 153 } 154 155 /* 156 * Same as ucm_get_ctx but requires that ->cm_id->device is valid, eg that the 157 * CM_ID is bound. 158 */ 159 static struct ucma_context *ucma_get_ctx_dev(struct ucma_file *file, int id) 160 { 161 struct ucma_context *ctx = ucma_get_ctx(file, id); 162 163 if (IS_ERR(ctx)) 164 return ctx; 165 if (!ctx->cm_id->device) { 166 ucma_put_ctx(ctx); 167 return ERR_PTR(-EINVAL); 168 } 169 return ctx; 170 } 171 172 static void ucma_close_event_id(struct work_struct *work) 173 { 174 struct ucma_event *uevent_close = container_of(work, struct ucma_event, close_work); 175 176 rdma_destroy_id(uevent_close->cm_id); 177 kfree(uevent_close); 178 } 179 180 static void ucma_close_id(struct work_struct *work) 181 { 182 struct ucma_context *ctx = container_of(work, struct ucma_context, close_work); 183 184 /* once all inflight tasks are finished, we close all underlying 185 * resources. The context is still alive till its explicit destryoing 186 * by its creator. 187 */ 188 ucma_put_ctx(ctx); 189 wait_for_completion(&ctx->comp); 190 /* No new events will be generated after destroying the id. */ 191 rdma_destroy_id(ctx->cm_id); 192 } 193 194 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file) 195 { 196 struct ucma_context *ctx; 197 198 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 199 if (!ctx) 200 return NULL; 201 202 INIT_WORK(&ctx->close_work, ucma_close_id); 203 atomic_set(&ctx->ref, 1); 204 init_completion(&ctx->comp); 205 INIT_LIST_HEAD(&ctx->mc_list); 206 ctx->file = file; 207 208 mutex_lock(&mut); 209 ctx->id = idr_alloc(&ctx_idr, ctx, 0, 0, GFP_KERNEL); 210 mutex_unlock(&mut); 211 if (ctx->id < 0) 212 goto error; 213 214 list_add_tail(&ctx->list, &file->ctx_list); 215 return ctx; 216 217 error: 218 kfree(ctx); 219 return NULL; 220 } 221 222 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx) 223 { 224 struct ucma_multicast *mc; 225 226 mc = kzalloc(sizeof(*mc), GFP_KERNEL); 227 if (!mc) 228 return NULL; 229 230 mutex_lock(&mut); 231 mc->id = idr_alloc(&multicast_idr, mc, 0, 0, GFP_KERNEL); 232 mutex_unlock(&mut); 233 if (mc->id < 0) 234 goto error; 235 236 mc->ctx = ctx; 237 list_add_tail(&mc->list, &ctx->mc_list); 238 return mc; 239 240 error: 241 kfree(mc); 242 return NULL; 243 } 244 245 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst, 246 struct rdma_conn_param *src) 247 { 248 if (src->private_data_len) 249 memcpy(dst->private_data, src->private_data, 250 src->private_data_len); 251 dst->private_data_len = src->private_data_len; 252 dst->responder_resources =src->responder_resources; 253 dst->initiator_depth = src->initiator_depth; 254 dst->flow_control = src->flow_control; 255 dst->retry_count = src->retry_count; 256 dst->rnr_retry_count = src->rnr_retry_count; 257 dst->srq = src->srq; 258 dst->qp_num = src->qp_num; 259 } 260 261 static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst, 262 struct rdma_ud_param *src) 263 { 264 if (src->private_data_len) 265 memcpy(dst->private_data, src->private_data, 266 src->private_data_len); 267 dst->private_data_len = src->private_data_len; 268 ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr); 269 dst->qp_num = src->qp_num; 270 dst->qkey = src->qkey; 271 } 272 273 static void ucma_set_event_context(struct ucma_context *ctx, 274 struct rdma_cm_event *event, 275 struct ucma_event *uevent) 276 { 277 uevent->ctx = ctx; 278 switch (event->event) { 279 case RDMA_CM_EVENT_MULTICAST_JOIN: 280 case RDMA_CM_EVENT_MULTICAST_ERROR: 281 uevent->mc = __DECONST(struct ucma_multicast *, 282 event->param.ud.private_data); 283 uevent->resp.uid = uevent->mc->uid; 284 uevent->resp.id = uevent->mc->id; 285 break; 286 default: 287 uevent->resp.uid = ctx->uid; 288 uevent->resp.id = ctx->id; 289 break; 290 } 291 } 292 293 /* Called with file->mut locked for the relevant context. */ 294 static void ucma_removal_event_handler(struct rdma_cm_id *cm_id) 295 { 296 struct ucma_context *ctx = cm_id->context; 297 struct ucma_event *con_req_eve; 298 int event_found = 0; 299 300 if (ctx->destroying) 301 return; 302 303 /* only if context is pointing to cm_id that it owns it and can be 304 * queued to be closed, otherwise that cm_id is an inflight one that 305 * is part of that context event list pending to be detached and 306 * reattached to its new context as part of ucma_get_event, 307 * handled separately below. 308 */ 309 if (ctx->cm_id == cm_id) { 310 mutex_lock(&mut); 311 ctx->closing = 1; 312 mutex_unlock(&mut); 313 queue_work(ctx->file->close_wq, &ctx->close_work); 314 return; 315 } 316 317 list_for_each_entry(con_req_eve, &ctx->file->event_list, list) { 318 if (con_req_eve->cm_id == cm_id && 319 con_req_eve->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) { 320 list_del(&con_req_eve->list); 321 INIT_WORK(&con_req_eve->close_work, ucma_close_event_id); 322 queue_work(ctx->file->close_wq, &con_req_eve->close_work); 323 event_found = 1; 324 break; 325 } 326 } 327 if (!event_found) 328 pr_err("ucma_removal_event_handler: warning: connect request event wasn't found\n"); 329 } 330 331 static int ucma_event_handler(struct rdma_cm_id *cm_id, 332 struct rdma_cm_event *event) 333 { 334 struct ucma_event *uevent; 335 struct ucma_context *ctx = cm_id->context; 336 int ret = 0; 337 338 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL); 339 if (!uevent) 340 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST; 341 342 mutex_lock(&ctx->file->mut); 343 uevent->cm_id = cm_id; 344 ucma_set_event_context(ctx, event, uevent); 345 uevent->resp.event = event->event; 346 uevent->resp.status = event->status; 347 if (cm_id->qp_type == IB_QPT_UD) 348 ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud); 349 else 350 ucma_copy_conn_event(&uevent->resp.param.conn, 351 &event->param.conn); 352 353 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) { 354 if (!ctx->backlog) { 355 ret = -ENOMEM; 356 kfree(uevent); 357 goto out; 358 } 359 ctx->backlog--; 360 } else if (!ctx->uid || ctx->cm_id != cm_id) { 361 /* 362 * We ignore events for new connections until userspace has set 363 * their context. This can only happen if an error occurs on a 364 * new connection before the user accepts it. This is okay, 365 * since the accept will just fail later. However, we do need 366 * to release the underlying HW resources in case of a device 367 * removal event. 368 */ 369 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL) 370 ucma_removal_event_handler(cm_id); 371 372 kfree(uevent); 373 goto out; 374 } 375 376 list_add_tail(&uevent->list, &ctx->file->event_list); 377 wake_up_interruptible(&ctx->file->poll_wait); 378 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL) 379 ucma_removal_event_handler(cm_id); 380 out: 381 mutex_unlock(&ctx->file->mut); 382 return ret; 383 } 384 385 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf, 386 int in_len, int out_len) 387 { 388 struct ucma_context *ctx; 389 struct rdma_ucm_get_event cmd; 390 struct ucma_event *uevent; 391 int ret = 0; 392 393 if (out_len < sizeof uevent->resp) 394 return -ENOSPC; 395 396 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 397 return -EFAULT; 398 399 mutex_lock(&file->mut); 400 while (list_empty(&file->event_list)) { 401 mutex_unlock(&file->mut); 402 403 if (file->filp->f_flags & O_NONBLOCK) 404 return -EAGAIN; 405 406 if (wait_event_interruptible(file->poll_wait, 407 !list_empty(&file->event_list))) 408 return -ERESTARTSYS; 409 410 mutex_lock(&file->mut); 411 } 412 413 uevent = list_entry(file->event_list.next, struct ucma_event, list); 414 415 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) { 416 ctx = ucma_alloc_ctx(file); 417 if (!ctx) { 418 ret = -ENOMEM; 419 goto done; 420 } 421 uevent->ctx->backlog++; 422 ctx->cm_id = uevent->cm_id; 423 ctx->cm_id->context = ctx; 424 uevent->resp.id = ctx->id; 425 } 426 427 if (copy_to_user((void __user *)(unsigned long)cmd.response, 428 &uevent->resp, sizeof uevent->resp)) { 429 ret = -EFAULT; 430 goto done; 431 } 432 433 list_del(&uevent->list); 434 uevent->ctx->events_reported++; 435 if (uevent->mc) 436 uevent->mc->events_reported++; 437 kfree(uevent); 438 done: 439 mutex_unlock(&file->mut); 440 return ret; 441 } 442 443 static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type) 444 { 445 switch (cmd->ps) { 446 case RDMA_PS_TCP: 447 *qp_type = IB_QPT_RC; 448 return 0; 449 case RDMA_PS_UDP: 450 case RDMA_PS_IPOIB: 451 *qp_type = IB_QPT_UD; 452 return 0; 453 case RDMA_PS_IB: 454 *qp_type = cmd->qp_type; 455 return 0; 456 default: 457 return -EINVAL; 458 } 459 } 460 461 static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf, 462 int in_len, int out_len) 463 { 464 struct rdma_ucm_create_id cmd; 465 struct rdma_ucm_create_id_resp resp; 466 struct ucma_context *ctx; 467 struct rdma_cm_id *cm_id; 468 enum ib_qp_type qp_type; 469 int ret; 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 ret = ucma_get_qp_type(&cmd, &qp_type); 478 if (ret) 479 return ret; 480 481 mutex_lock(&file->mut); 482 ctx = ucma_alloc_ctx(file); 483 mutex_unlock(&file->mut); 484 if (!ctx) 485 return -ENOMEM; 486 487 ctx->uid = cmd.uid; 488 cm_id = rdma_create_id(TD_TO_VNET(curthread), 489 ucma_event_handler, ctx, cmd.ps, qp_type); 490 if (IS_ERR(cm_id)) { 491 ret = PTR_ERR(cm_id); 492 goto err1; 493 } 494 495 resp.id = ctx->id; 496 if (copy_to_user((void __user *)(unsigned long)cmd.response, 497 &resp, sizeof(resp))) { 498 ret = -EFAULT; 499 goto err2; 500 } 501 502 ctx->cm_id = cm_id; 503 return 0; 504 505 err2: 506 rdma_destroy_id(cm_id); 507 err1: 508 mutex_lock(&mut); 509 idr_remove(&ctx_idr, ctx->id); 510 mutex_unlock(&mut); 511 mutex_lock(&file->mut); 512 list_del(&ctx->list); 513 mutex_unlock(&file->mut); 514 kfree(ctx); 515 return ret; 516 } 517 518 static void ucma_cleanup_multicast(struct ucma_context *ctx) 519 { 520 struct ucma_multicast *mc, *tmp; 521 522 mutex_lock(&mut); 523 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) { 524 list_del(&mc->list); 525 idr_remove(&multicast_idr, mc->id); 526 kfree(mc); 527 } 528 mutex_unlock(&mut); 529 } 530 531 static void ucma_cleanup_mc_events(struct ucma_multicast *mc) 532 { 533 struct ucma_event *uevent, *tmp; 534 535 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) { 536 if (uevent->mc != mc) 537 continue; 538 539 list_del(&uevent->list); 540 kfree(uevent); 541 } 542 } 543 544 /* 545 * ucma_free_ctx is called after the underlying rdma CM-ID is destroyed. At 546 * this point, no new events will be reported from the hardware. However, we 547 * still need to cleanup the UCMA context for this ID. Specifically, there 548 * might be events that have not yet been consumed by the user space software. 549 * These might include pending connect requests which we have not completed 550 * processing. We cannot call rdma_destroy_id while holding the lock of the 551 * context (file->mut), as it might cause a deadlock. We therefore extract all 552 * relevant events from the context pending events list while holding the 553 * mutex. After that we release them as needed. 554 */ 555 static int ucma_free_ctx(struct ucma_context *ctx) 556 { 557 int events_reported; 558 struct ucma_event *uevent, *tmp; 559 LIST_HEAD(list); 560 561 562 ucma_cleanup_multicast(ctx); 563 564 /* Cleanup events not yet reported to the user. */ 565 mutex_lock(&ctx->file->mut); 566 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) { 567 if (uevent->ctx == ctx) 568 list_move_tail(&uevent->list, &list); 569 } 570 list_del(&ctx->list); 571 mutex_unlock(&ctx->file->mut); 572 573 list_for_each_entry_safe(uevent, tmp, &list, list) { 574 list_del(&uevent->list); 575 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) 576 rdma_destroy_id(uevent->cm_id); 577 kfree(uevent); 578 } 579 580 events_reported = ctx->events_reported; 581 kfree(ctx); 582 return events_reported; 583 } 584 585 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf, 586 int in_len, int out_len) 587 { 588 struct rdma_ucm_destroy_id cmd; 589 struct rdma_ucm_destroy_id_resp resp; 590 struct ucma_context *ctx; 591 int ret = 0; 592 593 if (out_len < sizeof(resp)) 594 return -ENOSPC; 595 596 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 597 return -EFAULT; 598 599 mutex_lock(&mut); 600 ctx = _ucma_find_context(cmd.id, file); 601 if (!IS_ERR(ctx)) 602 idr_remove(&ctx_idr, ctx->id); 603 mutex_unlock(&mut); 604 605 if (IS_ERR(ctx)) 606 return PTR_ERR(ctx); 607 608 mutex_lock(&ctx->file->mut); 609 ctx->destroying = 1; 610 mutex_unlock(&ctx->file->mut); 611 612 flush_workqueue(ctx->file->close_wq); 613 /* At this point it's guaranteed that there is no inflight 614 * closing task */ 615 mutex_lock(&mut); 616 if (!ctx->closing) { 617 mutex_unlock(&mut); 618 ucma_put_ctx(ctx); 619 wait_for_completion(&ctx->comp); 620 rdma_destroy_id(ctx->cm_id); 621 } else { 622 mutex_unlock(&mut); 623 } 624 625 resp.events_reported = ucma_free_ctx(ctx); 626 if (copy_to_user((void __user *)(unsigned long)cmd.response, 627 &resp, sizeof(resp))) 628 ret = -EFAULT; 629 630 return ret; 631 } 632 633 static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf, 634 int in_len, int out_len) 635 { 636 struct rdma_ucm_bind_ip cmd; 637 struct ucma_context *ctx; 638 int ret; 639 640 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 641 return -EFAULT; 642 643 if (!rdma_addr_size_in6(&cmd.addr)) 644 return -EINVAL; 645 646 ctx = ucma_get_ctx(file, cmd.id); 647 if (IS_ERR(ctx)) 648 return PTR_ERR(ctx); 649 650 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr); 651 ucma_put_ctx(ctx); 652 return ret; 653 } 654 655 static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf, 656 int in_len, int out_len) 657 { 658 struct rdma_ucm_bind cmd; 659 struct ucma_context *ctx; 660 int ret; 661 662 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 663 return -EFAULT; 664 665 if (cmd.reserved || !cmd.addr_size || 666 cmd.addr_size != rdma_addr_size_kss(&cmd.addr)) 667 return -EINVAL; 668 669 ctx = ucma_get_ctx(file, cmd.id); 670 if (IS_ERR(ctx)) 671 return PTR_ERR(ctx); 672 673 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr); 674 ucma_put_ctx(ctx); 675 return ret; 676 } 677 678 static ssize_t ucma_resolve_ip(struct ucma_file *file, 679 const char __user *inbuf, 680 int in_len, int out_len) 681 { 682 struct rdma_ucm_resolve_ip cmd; 683 struct ucma_context *ctx; 684 int ret; 685 686 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 687 return -EFAULT; 688 689 if ((cmd.src_addr.sin6_family && !rdma_addr_size_in6(&cmd.src_addr)) || 690 !rdma_addr_size_in6(&cmd.dst_addr)) 691 return -EINVAL; 692 693 ctx = ucma_get_ctx(file, cmd.id); 694 if (IS_ERR(ctx)) 695 return PTR_ERR(ctx); 696 697 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr, 698 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms); 699 ucma_put_ctx(ctx); 700 return ret; 701 } 702 703 static ssize_t ucma_resolve_addr(struct ucma_file *file, 704 const char __user *inbuf, 705 int in_len, int out_len) 706 { 707 struct rdma_ucm_resolve_addr cmd; 708 struct ucma_context *ctx; 709 int ret; 710 711 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 712 return -EFAULT; 713 714 if (cmd.reserved || 715 (cmd.src_size && (cmd.src_size != rdma_addr_size_kss(&cmd.src_addr))) || 716 !cmd.dst_size || (cmd.dst_size != rdma_addr_size_kss(&cmd.dst_addr))) 717 return -EINVAL; 718 719 ctx = ucma_get_ctx(file, cmd.id); 720 if (IS_ERR(ctx)) 721 return PTR_ERR(ctx); 722 723 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr, 724 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms); 725 ucma_put_ctx(ctx); 726 return ret; 727 } 728 729 static ssize_t ucma_resolve_route(struct ucma_file *file, 730 const char __user *inbuf, 731 int in_len, int out_len) 732 { 733 struct rdma_ucm_resolve_route cmd; 734 struct ucma_context *ctx; 735 int ret; 736 737 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 738 return -EFAULT; 739 740 ctx = ucma_get_ctx_dev(file, cmd.id); 741 if (IS_ERR(ctx)) 742 return PTR_ERR(ctx); 743 744 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms); 745 ucma_put_ctx(ctx); 746 return ret; 747 } 748 749 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp, 750 struct rdma_route *route) 751 { 752 struct rdma_dev_addr *dev_addr; 753 754 resp->num_paths = route->num_paths; 755 switch (route->num_paths) { 756 case 0: 757 dev_addr = &route->addr.dev_addr; 758 rdma_addr_get_dgid(dev_addr, 759 (union ib_gid *) &resp->ib_route[0].dgid); 760 rdma_addr_get_sgid(dev_addr, 761 (union ib_gid *) &resp->ib_route[0].sgid); 762 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 763 break; 764 case 2: 765 ib_copy_path_rec_to_user(&resp->ib_route[1], 766 &route->path_rec[1]); 767 /* fall through */ 768 case 1: 769 ib_copy_path_rec_to_user(&resp->ib_route[0], 770 &route->path_rec[0]); 771 break; 772 default: 773 break; 774 } 775 } 776 777 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp, 778 struct rdma_route *route) 779 { 780 781 resp->num_paths = route->num_paths; 782 switch (route->num_paths) { 783 case 0: 784 rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr, 785 (union ib_gid *)&resp->ib_route[0].dgid); 786 rdma_ip2gid((struct sockaddr *)&route->addr.src_addr, 787 (union ib_gid *)&resp->ib_route[0].sgid); 788 resp->ib_route[0].pkey = cpu_to_be16(0xffff); 789 break; 790 case 2: 791 ib_copy_path_rec_to_user(&resp->ib_route[1], 792 &route->path_rec[1]); 793 /* fall through */ 794 case 1: 795 ib_copy_path_rec_to_user(&resp->ib_route[0], 796 &route->path_rec[0]); 797 break; 798 default: 799 break; 800 } 801 } 802 803 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp, 804 struct rdma_route *route) 805 { 806 struct rdma_dev_addr *dev_addr; 807 808 dev_addr = &route->addr.dev_addr; 809 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid); 810 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid); 811 } 812 813 static ssize_t ucma_query_route(struct ucma_file *file, 814 const char __user *inbuf, 815 int in_len, int out_len) 816 { 817 struct rdma_ucm_query cmd; 818 struct rdma_ucm_query_route_resp resp; 819 struct ucma_context *ctx; 820 struct sockaddr *addr; 821 int ret = 0; 822 823 if (out_len < sizeof(resp)) 824 return -ENOSPC; 825 826 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 827 return -EFAULT; 828 829 ctx = ucma_get_ctx(file, cmd.id); 830 if (IS_ERR(ctx)) 831 return PTR_ERR(ctx); 832 833 memset(&resp, 0, sizeof resp); 834 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr; 835 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ? 836 sizeof(struct sockaddr_in) : 837 sizeof(struct sockaddr_in6)); 838 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr; 839 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ? 840 sizeof(struct sockaddr_in) : 841 sizeof(struct sockaddr_in6)); 842 if (!ctx->cm_id->device) 843 goto out; 844 845 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid; 846 resp.port_num = ctx->cm_id->port_num; 847 848 if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num)) 849 ucma_copy_ib_route(&resp, &ctx->cm_id->route); 850 else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num)) 851 ucma_copy_iboe_route(&resp, &ctx->cm_id->route); 852 else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num)) 853 ucma_copy_iw_route(&resp, &ctx->cm_id->route); 854 855 out: 856 if (copy_to_user((void __user *)(unsigned long)cmd.response, 857 &resp, sizeof(resp))) 858 ret = -EFAULT; 859 860 ucma_put_ctx(ctx); 861 return ret; 862 } 863 864 static void ucma_query_device_addr(struct rdma_cm_id *cm_id, 865 struct rdma_ucm_query_addr_resp *resp) 866 { 867 if (!cm_id->device) 868 return; 869 870 resp->node_guid = (__force __u64) cm_id->device->node_guid; 871 resp->port_num = cm_id->port_num; 872 resp->pkey = (__force __u16) cpu_to_be16( 873 ib_addr_get_pkey(&cm_id->route.addr.dev_addr)); 874 } 875 876 static ssize_t ucma_query_addr(struct ucma_context *ctx, 877 void __user *response, int out_len) 878 { 879 struct rdma_ucm_query_addr_resp resp; 880 struct sockaddr *addr; 881 int ret = 0; 882 883 if (out_len < sizeof(resp)) 884 return -ENOSPC; 885 886 memset(&resp, 0, sizeof resp); 887 888 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr; 889 resp.src_size = rdma_addr_size(addr); 890 memcpy(&resp.src_addr, addr, resp.src_size); 891 892 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr; 893 resp.dst_size = rdma_addr_size(addr); 894 memcpy(&resp.dst_addr, addr, resp.dst_size); 895 896 ucma_query_device_addr(ctx->cm_id, &resp); 897 898 if (copy_to_user(response, &resp, sizeof(resp))) 899 ret = -EFAULT; 900 901 return ret; 902 } 903 904 static ssize_t ucma_query_path(struct ucma_context *ctx, 905 void __user *response, int out_len) 906 { 907 struct rdma_ucm_query_path_resp *resp; 908 int i, ret = 0; 909 910 if (out_len < sizeof(*resp)) 911 return -ENOSPC; 912 913 resp = kzalloc(out_len, GFP_KERNEL); 914 if (!resp) 915 return -ENOMEM; 916 917 resp->num_paths = ctx->cm_id->route.num_paths; 918 for (i = 0, out_len -= sizeof(*resp); 919 i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data); 920 i++, out_len -= sizeof(struct ib_path_rec_data)) { 921 922 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY | 923 IB_PATH_BIDIRECTIONAL; 924 ib_sa_pack_path(&ctx->cm_id->route.path_rec[i], 925 &resp->path_data[i].path_rec); 926 } 927 928 if (copy_to_user(response, resp, 929 sizeof(*resp) + (i * sizeof(struct ib_path_rec_data)))) 930 ret = -EFAULT; 931 932 kfree(resp); 933 return ret; 934 } 935 936 static ssize_t ucma_query_gid(struct ucma_context *ctx, 937 void __user *response, int out_len) 938 { 939 struct rdma_ucm_query_addr_resp resp; 940 struct sockaddr_ib *addr; 941 int ret = 0; 942 943 if (out_len < sizeof(resp)) 944 return -ENOSPC; 945 946 memset(&resp, 0, sizeof resp); 947 948 ucma_query_device_addr(ctx->cm_id, &resp); 949 950 addr = (struct sockaddr_ib *) &resp.src_addr; 951 resp.src_size = sizeof(*addr); 952 if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) { 953 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size); 954 } else { 955 addr->sib_family = AF_IB; 956 addr->sib_pkey = (__force __be16) resp.pkey; 957 rdma_addr_get_sgid(&ctx->cm_id->route.addr.dev_addr, 958 (union ib_gid *) &addr->sib_addr); 959 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *) 960 &ctx->cm_id->route.addr.src_addr); 961 } 962 963 addr = (struct sockaddr_ib *) &resp.dst_addr; 964 resp.dst_size = sizeof(*addr); 965 if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) { 966 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size); 967 } else { 968 addr->sib_family = AF_IB; 969 addr->sib_pkey = (__force __be16) resp.pkey; 970 rdma_addr_get_dgid(&ctx->cm_id->route.addr.dev_addr, 971 (union ib_gid *) &addr->sib_addr); 972 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *) 973 &ctx->cm_id->route.addr.dst_addr); 974 } 975 976 if (copy_to_user(response, &resp, sizeof(resp))) 977 ret = -EFAULT; 978 979 return ret; 980 } 981 982 static ssize_t ucma_query(struct ucma_file *file, 983 const char __user *inbuf, 984 int in_len, int out_len) 985 { 986 struct rdma_ucm_query cmd; 987 struct ucma_context *ctx; 988 void __user *response; 989 int ret; 990 991 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 992 return -EFAULT; 993 994 response = (void __user *)(unsigned long) cmd.response; 995 ctx = ucma_get_ctx(file, cmd.id); 996 if (IS_ERR(ctx)) 997 return PTR_ERR(ctx); 998 999 switch (cmd.option) { 1000 case RDMA_USER_CM_QUERY_ADDR: 1001 ret = ucma_query_addr(ctx, response, out_len); 1002 break; 1003 case RDMA_USER_CM_QUERY_PATH: 1004 ret = ucma_query_path(ctx, response, out_len); 1005 break; 1006 case RDMA_USER_CM_QUERY_GID: 1007 ret = ucma_query_gid(ctx, response, out_len); 1008 break; 1009 default: 1010 ret = -ENOSYS; 1011 break; 1012 } 1013 1014 ucma_put_ctx(ctx); 1015 return ret; 1016 } 1017 1018 static void ucma_copy_conn_param(struct rdma_cm_id *id, 1019 struct rdma_conn_param *dst, 1020 struct rdma_ucm_conn_param *src) 1021 { 1022 dst->private_data = src->private_data; 1023 dst->private_data_len = src->private_data_len; 1024 dst->responder_resources =src->responder_resources; 1025 dst->initiator_depth = src->initiator_depth; 1026 dst->flow_control = src->flow_control; 1027 dst->retry_count = src->retry_count; 1028 dst->rnr_retry_count = src->rnr_retry_count; 1029 dst->srq = src->srq; 1030 dst->qp_num = src->qp_num; 1031 dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0; 1032 } 1033 1034 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf, 1035 int in_len, int out_len) 1036 { 1037 struct rdma_ucm_connect cmd; 1038 struct rdma_conn_param conn_param; 1039 struct ucma_context *ctx; 1040 int ret; 1041 1042 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1043 return -EFAULT; 1044 1045 if (!cmd.conn_param.valid) 1046 return -EINVAL; 1047 1048 ctx = ucma_get_ctx_dev(file, cmd.id); 1049 if (IS_ERR(ctx)) 1050 return PTR_ERR(ctx); 1051 1052 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param); 1053 ret = rdma_connect(ctx->cm_id, &conn_param); 1054 ucma_put_ctx(ctx); 1055 return ret; 1056 } 1057 1058 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf, 1059 int in_len, int out_len) 1060 { 1061 struct rdma_ucm_listen cmd; 1062 struct ucma_context *ctx; 1063 int ret; 1064 1065 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1066 return -EFAULT; 1067 1068 ctx = ucma_get_ctx(file, cmd.id); 1069 if (IS_ERR(ctx)) 1070 return PTR_ERR(ctx); 1071 1072 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ? 1073 cmd.backlog : max_backlog; 1074 ret = rdma_listen(ctx->cm_id, ctx->backlog); 1075 ucma_put_ctx(ctx); 1076 return ret; 1077 } 1078 1079 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf, 1080 int in_len, int out_len) 1081 { 1082 struct rdma_ucm_accept cmd; 1083 struct rdma_conn_param conn_param; 1084 struct ucma_context *ctx; 1085 int ret; 1086 1087 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1088 return -EFAULT; 1089 1090 ctx = ucma_get_ctx_dev(file, cmd.id); 1091 if (IS_ERR(ctx)) 1092 return PTR_ERR(ctx); 1093 1094 if (cmd.conn_param.valid) { 1095 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param); 1096 mutex_lock(&file->mut); 1097 ret = rdma_accept(ctx->cm_id, &conn_param); 1098 if (!ret) 1099 ctx->uid = cmd.uid; 1100 mutex_unlock(&file->mut); 1101 } else 1102 ret = rdma_accept(ctx->cm_id, NULL); 1103 1104 ucma_put_ctx(ctx); 1105 return ret; 1106 } 1107 1108 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf, 1109 int in_len, int out_len) 1110 { 1111 struct rdma_ucm_reject cmd; 1112 struct ucma_context *ctx; 1113 int ret; 1114 1115 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1116 return -EFAULT; 1117 1118 ctx = ucma_get_ctx_dev(file, cmd.id); 1119 if (IS_ERR(ctx)) 1120 return PTR_ERR(ctx); 1121 1122 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len); 1123 ucma_put_ctx(ctx); 1124 return ret; 1125 } 1126 1127 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf, 1128 int in_len, int out_len) 1129 { 1130 struct rdma_ucm_disconnect cmd; 1131 struct ucma_context *ctx; 1132 int ret; 1133 1134 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1135 return -EFAULT; 1136 1137 ctx = ucma_get_ctx_dev(file, cmd.id); 1138 if (IS_ERR(ctx)) 1139 return PTR_ERR(ctx); 1140 1141 ret = rdma_disconnect(ctx->cm_id); 1142 ucma_put_ctx(ctx); 1143 return ret; 1144 } 1145 1146 static ssize_t ucma_init_qp_attr(struct ucma_file *file, 1147 const char __user *inbuf, 1148 int in_len, int out_len) 1149 { 1150 struct rdma_ucm_init_qp_attr cmd; 1151 struct ib_uverbs_qp_attr resp; 1152 struct ucma_context *ctx; 1153 struct ib_qp_attr qp_attr; 1154 int ret; 1155 1156 if (out_len < sizeof(resp)) 1157 return -ENOSPC; 1158 1159 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1160 return -EFAULT; 1161 1162 ctx = ucma_get_ctx_dev(file, cmd.id); 1163 if (IS_ERR(ctx)) 1164 return PTR_ERR(ctx); 1165 1166 resp.qp_attr_mask = 0; 1167 memset(&qp_attr, 0, sizeof qp_attr); 1168 qp_attr.qp_state = cmd.qp_state; 1169 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask); 1170 if (ret) 1171 goto out; 1172 1173 ib_copy_qp_attr_to_user(&resp, &qp_attr); 1174 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1175 &resp, sizeof(resp))) 1176 ret = -EFAULT; 1177 1178 out: 1179 ucma_put_ctx(ctx); 1180 return ret; 1181 } 1182 1183 static int ucma_set_option_id(struct ucma_context *ctx, int optname, 1184 void *optval, size_t optlen) 1185 { 1186 int ret = 0; 1187 1188 switch (optname) { 1189 case RDMA_OPTION_ID_TOS: 1190 if (optlen != sizeof(u8)) { 1191 ret = -EINVAL; 1192 break; 1193 } 1194 rdma_set_service_type(ctx->cm_id, *((u8 *) optval)); 1195 break; 1196 case RDMA_OPTION_ID_REUSEADDR: 1197 if (optlen != sizeof(int)) { 1198 ret = -EINVAL; 1199 break; 1200 } 1201 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0); 1202 break; 1203 case RDMA_OPTION_ID_AFONLY: 1204 if (optlen != sizeof(int)) { 1205 ret = -EINVAL; 1206 break; 1207 } 1208 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0); 1209 break; 1210 case RDMA_OPTION_ID_ACK_TIMEOUT: 1211 if (optlen != sizeof(u8)) { 1212 ret = -EINVAL; 1213 break; 1214 } 1215 ret = rdma_set_ack_timeout(ctx->cm_id, *((u8 *)optval)); 1216 break; 1217 default: 1218 ret = -ENOSYS; 1219 } 1220 1221 return ret; 1222 } 1223 1224 static int ucma_set_ib_path(struct ucma_context *ctx, 1225 struct ib_path_rec_data *path_data, size_t optlen) 1226 { 1227 struct ib_sa_path_rec sa_path; 1228 struct rdma_cm_event event; 1229 int ret; 1230 1231 if (optlen % sizeof(*path_data)) 1232 return -EINVAL; 1233 1234 for (; optlen; optlen -= sizeof(*path_data), path_data++) { 1235 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY | 1236 IB_PATH_BIDIRECTIONAL)) 1237 break; 1238 } 1239 1240 if (!optlen) 1241 return -EINVAL; 1242 1243 memset(&sa_path, 0, sizeof(sa_path)); 1244 1245 ib_sa_unpack_path(path_data->path_rec, &sa_path); 1246 ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1); 1247 if (ret) 1248 return ret; 1249 1250 memset(&event, 0, sizeof event); 1251 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 1252 return ucma_event_handler(ctx->cm_id, &event); 1253 } 1254 1255 static int ucma_set_option_ib(struct ucma_context *ctx, int optname, 1256 void *optval, size_t optlen) 1257 { 1258 int ret; 1259 1260 switch (optname) { 1261 case RDMA_OPTION_IB_PATH: 1262 ret = ucma_set_ib_path(ctx, optval, optlen); 1263 break; 1264 default: 1265 ret = -ENOSYS; 1266 } 1267 1268 return ret; 1269 } 1270 1271 static int ucma_set_option_level(struct ucma_context *ctx, int level, 1272 int optname, void *optval, size_t optlen) 1273 { 1274 int ret; 1275 1276 switch (level) { 1277 case RDMA_OPTION_ID: 1278 ret = ucma_set_option_id(ctx, optname, optval, optlen); 1279 break; 1280 case RDMA_OPTION_IB: 1281 ret = ucma_set_option_ib(ctx, optname, optval, optlen); 1282 break; 1283 default: 1284 ret = -ENOSYS; 1285 } 1286 1287 return ret; 1288 } 1289 1290 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf, 1291 int in_len, int out_len) 1292 { 1293 struct rdma_ucm_set_option cmd; 1294 struct ucma_context *ctx; 1295 void *optval; 1296 int ret; 1297 1298 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1299 return -EFAULT; 1300 1301 ctx = ucma_get_ctx(file, cmd.id); 1302 if (IS_ERR(ctx)) 1303 return PTR_ERR(ctx); 1304 1305 optval = memdup_user((void __user *) (unsigned long) cmd.optval, 1306 cmd.optlen); 1307 if (IS_ERR(optval)) { 1308 ret = PTR_ERR(optval); 1309 goto out; 1310 } 1311 1312 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval, 1313 cmd.optlen); 1314 kfree(optval); 1315 1316 out: 1317 ucma_put_ctx(ctx); 1318 return ret; 1319 } 1320 1321 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf, 1322 int in_len, int out_len) 1323 { 1324 struct rdma_ucm_notify cmd; 1325 struct ucma_context *ctx; 1326 int ret; 1327 1328 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1329 return -EFAULT; 1330 1331 ctx = ucma_get_ctx(file, cmd.id); 1332 if (IS_ERR(ctx)) 1333 return PTR_ERR(ctx); 1334 1335 ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event); 1336 ucma_put_ctx(ctx); 1337 return ret; 1338 } 1339 1340 static ssize_t ucma_process_join(struct ucma_file *file, 1341 struct rdma_ucm_join_mcast *cmd, int out_len) 1342 { 1343 struct rdma_ucm_create_id_resp resp; 1344 struct ucma_context *ctx; 1345 struct ucma_multicast *mc; 1346 struct sockaddr *addr; 1347 int ret; 1348 u8 join_state; 1349 1350 if (out_len < sizeof(resp)) 1351 return -ENOSPC; 1352 1353 addr = (struct sockaddr *) &cmd->addr; 1354 if (!cmd->addr_size || (cmd->addr_size != rdma_addr_size(addr))) 1355 return -EINVAL; 1356 1357 if (cmd->join_flags == RDMA_MC_JOIN_FLAG_FULLMEMBER) 1358 join_state = BIT(FULLMEMBER_JOIN); 1359 else if (cmd->join_flags == RDMA_MC_JOIN_FLAG_SENDONLY_FULLMEMBER) 1360 join_state = BIT(SENDONLY_FULLMEMBER_JOIN); 1361 else 1362 return -EINVAL; 1363 1364 ctx = ucma_get_ctx_dev(file, cmd->id); 1365 if (IS_ERR(ctx)) 1366 return PTR_ERR(ctx); 1367 1368 mutex_lock(&file->mut); 1369 mc = ucma_alloc_multicast(ctx); 1370 if (!mc) { 1371 ret = -ENOMEM; 1372 goto err1; 1373 } 1374 mc->join_state = join_state; 1375 mc->uid = cmd->uid; 1376 memcpy(&mc->addr, addr, cmd->addr_size); 1377 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *)&mc->addr, 1378 join_state, mc); 1379 if (ret) 1380 goto err2; 1381 1382 resp.id = mc->id; 1383 if (copy_to_user((void __user *)(unsigned long) cmd->response, 1384 &resp, sizeof(resp))) { 1385 ret = -EFAULT; 1386 goto err3; 1387 } 1388 1389 mutex_unlock(&file->mut); 1390 ucma_put_ctx(ctx); 1391 return 0; 1392 1393 err3: 1394 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr); 1395 ucma_cleanup_mc_events(mc); 1396 err2: 1397 mutex_lock(&mut); 1398 idr_remove(&multicast_idr, mc->id); 1399 mutex_unlock(&mut); 1400 list_del(&mc->list); 1401 kfree(mc); 1402 err1: 1403 mutex_unlock(&file->mut); 1404 ucma_put_ctx(ctx); 1405 return ret; 1406 } 1407 1408 static ssize_t ucma_join_ip_multicast(struct ucma_file *file, 1409 const char __user *inbuf, 1410 int in_len, int out_len) 1411 { 1412 struct rdma_ucm_join_ip_mcast cmd; 1413 struct rdma_ucm_join_mcast join_cmd; 1414 1415 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1416 return -EFAULT; 1417 1418 join_cmd.response = cmd.response; 1419 join_cmd.uid = cmd.uid; 1420 join_cmd.id = cmd.id; 1421 join_cmd.addr_size = rdma_addr_size_in6(&cmd.addr); 1422 join_cmd.join_flags = RDMA_MC_JOIN_FLAG_FULLMEMBER; 1423 memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size); 1424 1425 return ucma_process_join(file, &join_cmd, out_len); 1426 } 1427 1428 static ssize_t ucma_join_multicast(struct ucma_file *file, 1429 const char __user *inbuf, 1430 int in_len, int out_len) 1431 { 1432 struct rdma_ucm_join_mcast cmd; 1433 1434 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1435 return -EFAULT; 1436 1437 if (!rdma_addr_size_kss(&cmd.addr)) 1438 return -EINVAL; 1439 1440 return ucma_process_join(file, &cmd, out_len); 1441 } 1442 1443 static ssize_t ucma_leave_multicast(struct ucma_file *file, 1444 const char __user *inbuf, 1445 int in_len, int out_len) 1446 { 1447 struct rdma_ucm_destroy_id cmd; 1448 struct rdma_ucm_destroy_id_resp resp; 1449 struct ucma_multicast *mc; 1450 int ret = 0; 1451 1452 if (out_len < sizeof(resp)) 1453 return -ENOSPC; 1454 1455 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1456 return -EFAULT; 1457 1458 mutex_lock(&mut); 1459 mc = idr_find(&multicast_idr, cmd.id); 1460 if (!mc) 1461 mc = ERR_PTR(-ENOENT); 1462 else if (mc->ctx->file != file) 1463 mc = ERR_PTR(-EINVAL); 1464 else if (!atomic_inc_not_zero(&mc->ctx->ref)) 1465 mc = ERR_PTR(-ENXIO); 1466 else 1467 idr_remove(&multicast_idr, mc->id); 1468 mutex_unlock(&mut); 1469 1470 if (IS_ERR(mc)) { 1471 ret = PTR_ERR(mc); 1472 goto out; 1473 } 1474 1475 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr); 1476 mutex_lock(&mc->ctx->file->mut); 1477 ucma_cleanup_mc_events(mc); 1478 list_del(&mc->list); 1479 mutex_unlock(&mc->ctx->file->mut); 1480 1481 ucma_put_ctx(mc->ctx); 1482 resp.events_reported = mc->events_reported; 1483 kfree(mc); 1484 1485 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1486 &resp, sizeof(resp))) 1487 ret = -EFAULT; 1488 out: 1489 return ret; 1490 } 1491 1492 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2) 1493 { 1494 /* Acquire mutex's based on pointer comparison to prevent deadlock. */ 1495 if (file1 < file2) { 1496 mutex_lock(&file1->mut); 1497 mutex_lock_nested(&file2->mut, SINGLE_DEPTH_NESTING); 1498 } else { 1499 mutex_lock(&file2->mut); 1500 mutex_lock_nested(&file1->mut, SINGLE_DEPTH_NESTING); 1501 } 1502 } 1503 1504 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2) 1505 { 1506 if (file1 < file2) { 1507 mutex_unlock(&file2->mut); 1508 mutex_unlock(&file1->mut); 1509 } else { 1510 mutex_unlock(&file1->mut); 1511 mutex_unlock(&file2->mut); 1512 } 1513 } 1514 1515 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file) 1516 { 1517 struct ucma_event *uevent, *tmp; 1518 1519 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) 1520 if (uevent->ctx == ctx) 1521 list_move_tail(&uevent->list, &file->event_list); 1522 } 1523 1524 static ssize_t ucma_migrate_id(struct ucma_file *new_file, 1525 const char __user *inbuf, 1526 int in_len, int out_len) 1527 { 1528 struct rdma_ucm_migrate_id cmd; 1529 struct rdma_ucm_migrate_resp resp; 1530 struct ucma_context *ctx; 1531 struct fd f; 1532 struct ucma_file *cur_file; 1533 int ret = 0; 1534 1535 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1536 return -EFAULT; 1537 1538 /* Get current fd to protect against it being closed */ 1539 f = fdget(cmd.fd); 1540 if (!f.file) 1541 return -ENOENT; 1542 1543 /* Validate current fd and prevent destruction of id. */ 1544 ctx = ucma_get_ctx(f.file->private_data, cmd.id); 1545 if (IS_ERR(ctx)) { 1546 ret = PTR_ERR(ctx); 1547 goto file_put; 1548 } 1549 1550 cur_file = ctx->file; 1551 if (cur_file == new_file) { 1552 resp.events_reported = ctx->events_reported; 1553 goto response; 1554 } 1555 1556 /* 1557 * Migrate events between fd's, maintaining order, and avoiding new 1558 * events being added before existing events. 1559 */ 1560 ucma_lock_files(cur_file, new_file); 1561 mutex_lock(&mut); 1562 1563 list_move_tail(&ctx->list, &new_file->ctx_list); 1564 ucma_move_events(ctx, new_file); 1565 ctx->file = new_file; 1566 resp.events_reported = ctx->events_reported; 1567 1568 mutex_unlock(&mut); 1569 ucma_unlock_files(cur_file, new_file); 1570 1571 response: 1572 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1573 &resp, sizeof(resp))) 1574 ret = -EFAULT; 1575 1576 ucma_put_ctx(ctx); 1577 file_put: 1578 fdput(f); 1579 return ret; 1580 } 1581 1582 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file, 1583 const char __user *inbuf, 1584 int in_len, int out_len) = { 1585 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id, 1586 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id, 1587 [RDMA_USER_CM_CMD_BIND_IP] = ucma_bind_ip, 1588 [RDMA_USER_CM_CMD_RESOLVE_IP] = ucma_resolve_ip, 1589 [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route, 1590 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route, 1591 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect, 1592 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen, 1593 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept, 1594 [RDMA_USER_CM_CMD_REJECT] = ucma_reject, 1595 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect, 1596 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr, 1597 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event, 1598 [RDMA_USER_CM_CMD_GET_OPTION] = NULL, 1599 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option, 1600 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify, 1601 [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast, 1602 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast, 1603 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id, 1604 [RDMA_USER_CM_CMD_QUERY] = ucma_query, 1605 [RDMA_USER_CM_CMD_BIND] = ucma_bind, 1606 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr, 1607 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast 1608 }; 1609 1610 static ssize_t ucma_write(struct file *filp, const char __user *buf, 1611 size_t len, loff_t *pos) 1612 { 1613 struct ucma_file *file = filp->private_data; 1614 struct rdma_ucm_cmd_hdr hdr; 1615 ssize_t ret; 1616 1617 if (WARN_ON_ONCE(!ib_safe_file_access(filp))) 1618 return -EACCES; 1619 1620 if (len < sizeof(hdr)) 1621 return -EINVAL; 1622 1623 if (copy_from_user(&hdr, buf, sizeof(hdr))) 1624 return -EFAULT; 1625 1626 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table)) 1627 return -EINVAL; 1628 1629 if (hdr.in + sizeof(hdr) > len) 1630 return -EINVAL; 1631 1632 if (!ucma_cmd_table[hdr.cmd]) 1633 return -ENOSYS; 1634 1635 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out); 1636 if (!ret) 1637 ret = len; 1638 1639 return ret; 1640 } 1641 1642 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait) 1643 { 1644 struct ucma_file *file = filp->private_data; 1645 unsigned int mask = 0; 1646 1647 poll_wait(filp, &file->poll_wait, wait); 1648 1649 if (!list_empty(&file->event_list)) 1650 mask = POLLIN | POLLRDNORM; 1651 1652 return mask; 1653 } 1654 1655 /* 1656 * ucma_open() does not need the BKL: 1657 * 1658 * - no global state is referred to; 1659 * - there is no ioctl method to race against; 1660 * - no further module initialization is required for open to work 1661 * after the device is registered. 1662 */ 1663 static int ucma_open(struct inode *inode, struct file *filp) 1664 { 1665 struct ucma_file *file; 1666 1667 file = kmalloc(sizeof *file, GFP_KERNEL); 1668 if (!file) 1669 return -ENOMEM; 1670 1671 file->close_wq = alloc_ordered_workqueue("ucma_close_id", 1672 WQ_MEM_RECLAIM); 1673 if (!file->close_wq) { 1674 kfree(file); 1675 return -ENOMEM; 1676 } 1677 1678 INIT_LIST_HEAD(&file->event_list); 1679 INIT_LIST_HEAD(&file->ctx_list); 1680 init_waitqueue_head(&file->poll_wait); 1681 mutex_init(&file->mut); 1682 1683 filp->private_data = file; 1684 file->filp = filp; 1685 1686 return nonseekable_open(inode, filp); 1687 } 1688 1689 static int ucma_close(struct inode *inode, struct file *filp) 1690 { 1691 struct ucma_file *file = filp->private_data; 1692 struct ucma_context *ctx, *tmp; 1693 1694 mutex_lock(&file->mut); 1695 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) { 1696 ctx->destroying = 1; 1697 mutex_unlock(&file->mut); 1698 1699 mutex_lock(&mut); 1700 idr_remove(&ctx_idr, ctx->id); 1701 mutex_unlock(&mut); 1702 1703 flush_workqueue(file->close_wq); 1704 /* At that step once ctx was marked as destroying and workqueue 1705 * was flushed we are safe from any inflights handlers that 1706 * might put other closing task. 1707 */ 1708 mutex_lock(&mut); 1709 if (!ctx->closing) { 1710 mutex_unlock(&mut); 1711 ucma_put_ctx(ctx); 1712 wait_for_completion(&ctx->comp); 1713 /* rdma_destroy_id ensures that no event handlers are 1714 * inflight for that id before releasing it. 1715 */ 1716 rdma_destroy_id(ctx->cm_id); 1717 } else { 1718 mutex_unlock(&mut); 1719 } 1720 1721 ucma_free_ctx(ctx); 1722 mutex_lock(&file->mut); 1723 } 1724 mutex_unlock(&file->mut); 1725 destroy_workqueue(file->close_wq); 1726 kfree(file); 1727 return 0; 1728 } 1729 1730 static long 1731 ucma_ioctl(struct file *filp, unsigned int cmd, unsigned long arg) 1732 { 1733 1734 switch (cmd) { 1735 case FIONBIO: 1736 case FIOASYNC: 1737 return (0); 1738 default: 1739 return (-ENOTTY); 1740 } 1741 } 1742 1743 static const struct file_operations ucma_fops = { 1744 .owner = THIS_MODULE, 1745 .open = ucma_open, 1746 .release = ucma_close, 1747 .write = ucma_write, 1748 .unlocked_ioctl = ucma_ioctl, 1749 .poll = ucma_poll, 1750 .llseek = no_llseek, 1751 }; 1752 1753 static struct miscdevice ucma_misc = { 1754 .minor = MISC_DYNAMIC_MINOR, 1755 .name = "rdma_cm", 1756 .nodename = "infiniband/rdma_cm", 1757 .mode = 0666, 1758 .fops = &ucma_fops, 1759 }; 1760 1761 static ssize_t show_abi_version(struct device *dev, 1762 struct device_attribute *attr, 1763 char *buf) 1764 { 1765 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION); 1766 } 1767 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL); 1768 1769 static int __init ucma_init(void) 1770 { 1771 int ret; 1772 1773 ret = misc_register(&ucma_misc); 1774 if (ret) 1775 return ret; 1776 1777 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version); 1778 if (ret) { 1779 pr_err("rdma_ucm: couldn't create abi_version attr\n"); 1780 goto err1; 1781 } 1782 1783 return 0; 1784 err1: 1785 misc_deregister(&ucma_misc); 1786 return ret; 1787 } 1788 1789 static void __exit ucma_cleanup(void) 1790 { 1791 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version); 1792 misc_deregister(&ucma_misc); 1793 idr_destroy(&ctx_idr); 1794 idr_destroy(&multicast_idr); 1795 } 1796 1797 module_init_order(ucma_init, SI_ORDER_FIFTH); 1798 module_exit_order(ucma_cleanup, SI_ORDER_FIFTH); 1799