drivers/infiniband/core/ucma.c

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