linux/drivers/infiniband/core/ucma.c
<<
>>
Prefs
   1/*
   2 * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   3 *
   4 * This software is available to you under a choice of one of two
   5 * licenses.  You may choose to be licensed under the terms of the GNU
   6 * General Public License (GPL) Version 2, available from the file
   7 * COPYING in the main directory of this source tree, or the
   8 * OpenIB.org BSD license below:
   9 *
  10 *     Redistribution and use in source and binary forms, with or
  11 *     without modification, are permitted provided that the following
  12 *     conditions are met:
  13 *
  14 *      - Redistributions of source code must retain the above
  15 *      copyright notice, this list of conditions and the following
  16 *      disclaimer.
  17 *
  18 *      - Redistributions in binary form must reproduce the above
  19 *      copyright notice, this list of conditions and the following
  20 *      disclaimer in the documentation and/or other materials
  21 *      provided with the distribution.
  22 *
  23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30 * SOFTWARE.
  31 */
  32
  33#include <linux/completion.h>
  34#include <linux/file.h>
  35#include <linux/mutex.h>
  36#include <linux/poll.h>
  37#include <linux/sched.h>
  38#include <linux/idr.h>
  39#include <linux/in.h>
  40#include <linux/in6.h>
  41#include <linux/miscdevice.h>
  42#include <linux/slab.h>
  43#include <linux/sysctl.h>
  44#include <linux/module.h>
  45
  46#include <rdma/rdma_user_cm.h>
  47#include <rdma/ib_marshall.h>
  48#include <rdma/rdma_cm.h>
  49#include <rdma/rdma_cm_ib.h>
  50#include <rdma/ib_addr.h>
  51#include <rdma/ib.h>
  52
  53MODULE_AUTHOR("Sean Hefty");
  54MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
  55MODULE_LICENSE("Dual BSD/GPL");
  56
  57static unsigned int max_backlog = 1024;
  58
  59static struct ctl_table_header *ucma_ctl_table_hdr;
  60static struct ctl_table ucma_ctl_table[] = {
  61        {
  62                .procname       = "max_backlog",
  63                .data           = &max_backlog,
  64                .maxlen         = sizeof max_backlog,
  65                .mode           = 0644,
  66                .proc_handler   = proc_dointvec,
  67        },
  68        { }
  69};
  70
  71struct ucma_file {
  72        struct mutex            mut;
  73        struct file             *filp;
  74        struct list_head        ctx_list;
  75        struct list_head        event_list;
  76        wait_queue_head_t       poll_wait;
  77};
  78
  79struct ucma_context {
  80        int                     id;
  81        struct completion       comp;
  82        atomic_t                ref;
  83        int                     events_reported;
  84        int                     backlog;
  85
  86        struct ucma_file        *file;
  87        struct rdma_cm_id       *cm_id;
  88        u64                     uid;
  89
  90        struct list_head        list;
  91        struct list_head        mc_list;
  92};
  93
  94struct ucma_multicast {
  95        struct ucma_context     *ctx;
  96        int                     id;
  97        int                     events_reported;
  98
  99        u64                     uid;
 100        struct list_head        list;
 101        struct sockaddr_storage addr;
 102};
 103
 104struct ucma_event {
 105        struct ucma_context     *ctx;
 106        struct ucma_multicast   *mc;
 107        struct list_head        list;
 108        struct rdma_cm_id       *cm_id;
 109        struct rdma_ucm_event_resp resp;
 110};
 111
 112static DEFINE_MUTEX(mut);
 113static DEFINE_IDR(ctx_idr);
 114static DEFINE_IDR(multicast_idr);
 115
 116static inline struct ucma_context *_ucma_find_context(int id,
 117                                                      struct ucma_file *file)
 118{
 119        struct ucma_context *ctx;
 120
 121        ctx = idr_find(&ctx_idr, id);
 122        if (!ctx)
 123                ctx = ERR_PTR(-ENOENT);
 124        else if (ctx->file != file)
 125                ctx = ERR_PTR(-EINVAL);
 126        return ctx;
 127}
 128
 129static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
 130{
 131        struct ucma_context *ctx;
 132
 133        mutex_lock(&mut);
 134        ctx = _ucma_find_context(id, file);
 135        if (!IS_ERR(ctx))
 136                atomic_inc(&ctx->ref);
 137        mutex_unlock(&mut);
 138        return ctx;
 139}
 140
 141static void ucma_put_ctx(struct ucma_context *ctx)
 142{
 143        if (atomic_dec_and_test(&ctx->ref))
 144                complete(&ctx->comp);
 145}
 146
 147static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
 148{
 149        struct ucma_context *ctx;
 150
 151        ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 152        if (!ctx)
 153                return NULL;
 154
 155        atomic_set(&ctx->ref, 1);
 156        init_completion(&ctx->comp);
 157        INIT_LIST_HEAD(&ctx->mc_list);
 158        ctx->file = file;
 159
 160        mutex_lock(&mut);
 161        ctx->id = idr_alloc(&ctx_idr, ctx, 0, 0, GFP_KERNEL);
 162        mutex_unlock(&mut);
 163        if (ctx->id < 0)
 164                goto error;
 165
 166        list_add_tail(&ctx->list, &file->ctx_list);
 167        return ctx;
 168
 169error:
 170        kfree(ctx);
 171        return NULL;
 172}
 173
 174static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
 175{
 176        struct ucma_multicast *mc;
 177
 178        mc = kzalloc(sizeof(*mc), GFP_KERNEL);
 179        if (!mc)
 180                return NULL;
 181
 182        mutex_lock(&mut);
 183        mc->id = idr_alloc(&multicast_idr, mc, 0, 0, GFP_KERNEL);
 184        mutex_unlock(&mut);
 185        if (mc->id < 0)
 186                goto error;
 187
 188        mc->ctx = ctx;
 189        list_add_tail(&mc->list, &ctx->mc_list);
 190        return mc;
 191
 192error:
 193        kfree(mc);
 194        return NULL;
 195}
 196
 197static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
 198                                 struct rdma_conn_param *src)
 199{
 200        if (src->private_data_len)
 201                memcpy(dst->private_data, src->private_data,
 202                       src->private_data_len);
 203        dst->private_data_len = src->private_data_len;
 204        dst->responder_resources =src->responder_resources;
 205        dst->initiator_depth = src->initiator_depth;
 206        dst->flow_control = src->flow_control;
 207        dst->retry_count = src->retry_count;
 208        dst->rnr_retry_count = src->rnr_retry_count;
 209        dst->srq = src->srq;
 210        dst->qp_num = src->qp_num;
 211}
 212
 213static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst,
 214                               struct rdma_ud_param *src)
 215{
 216        if (src->private_data_len)
 217                memcpy(dst->private_data, src->private_data,
 218                       src->private_data_len);
 219        dst->private_data_len = src->private_data_len;
 220        ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr);
 221        dst->qp_num = src->qp_num;
 222        dst->qkey = src->qkey;
 223}
 224
 225static void ucma_set_event_context(struct ucma_context *ctx,
 226                                   struct rdma_cm_event *event,
 227                                   struct ucma_event *uevent)
 228{
 229        uevent->ctx = ctx;
 230        switch (event->event) {
 231        case RDMA_CM_EVENT_MULTICAST_JOIN:
 232        case RDMA_CM_EVENT_MULTICAST_ERROR:
 233                uevent->mc = (struct ucma_multicast *)
 234                             event->param.ud.private_data;
 235                uevent->resp.uid = uevent->mc->uid;
 236                uevent->resp.id = uevent->mc->id;
 237                break;
 238        default:
 239                uevent->resp.uid = ctx->uid;
 240                uevent->resp.id = ctx->id;
 241                break;
 242        }
 243}
 244
 245static int ucma_event_handler(struct rdma_cm_id *cm_id,
 246                              struct rdma_cm_event *event)
 247{
 248        struct ucma_event *uevent;
 249        struct ucma_context *ctx = cm_id->context;
 250        int ret = 0;
 251
 252        uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
 253        if (!uevent)
 254                return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
 255
 256        mutex_lock(&ctx->file->mut);
 257        uevent->cm_id = cm_id;
 258        ucma_set_event_context(ctx, event, uevent);
 259        uevent->resp.event = event->event;
 260        uevent->resp.status = event->status;
 261        if (cm_id->qp_type == IB_QPT_UD)
 262                ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud);
 263        else
 264                ucma_copy_conn_event(&uevent->resp.param.conn,
 265                                     &event->param.conn);
 266
 267        if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
 268                if (!ctx->backlog) {
 269                        ret = -ENOMEM;
 270                        kfree(uevent);
 271                        goto out;
 272                }
 273                ctx->backlog--;
 274        } else if (!ctx->uid || ctx->cm_id != cm_id) {
 275                /*
 276                 * We ignore events for new connections until userspace has set
 277                 * their context.  This can only happen if an error occurs on a
 278                 * new connection before the user accepts it.  This is okay,
 279                 * since the accept will just fail later.
 280                 */
 281                kfree(uevent);
 282                goto out;
 283        }
 284
 285        list_add_tail(&uevent->list, &ctx->file->event_list);
 286        wake_up_interruptible(&ctx->file->poll_wait);
 287out:
 288        mutex_unlock(&ctx->file->mut);
 289        return ret;
 290}
 291
 292static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
 293                              int in_len, int out_len)
 294{
 295        struct ucma_context *ctx;
 296        struct rdma_ucm_get_event cmd;
 297        struct ucma_event *uevent;
 298        int ret = 0;
 299
 300        if (out_len < sizeof uevent->resp)
 301                return -ENOSPC;
 302
 303        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 304                return -EFAULT;
 305
 306        mutex_lock(&file->mut);
 307        while (list_empty(&file->event_list)) {
 308                mutex_unlock(&file->mut);
 309
 310                if (file->filp->f_flags & O_NONBLOCK)
 311                        return -EAGAIN;
 312
 313                if (wait_event_interruptible(file->poll_wait,
 314                                             !list_empty(&file->event_list)))
 315                        return -ERESTARTSYS;
 316
 317                mutex_lock(&file->mut);
 318        }
 319
 320        uevent = list_entry(file->event_list.next, struct ucma_event, list);
 321
 322        if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
 323                ctx = ucma_alloc_ctx(file);
 324                if (!ctx) {
 325                        ret = -ENOMEM;
 326                        goto done;
 327                }
 328                uevent->ctx->backlog++;
 329                ctx->cm_id = uevent->cm_id;
 330                ctx->cm_id->context = ctx;
 331                uevent->resp.id = ctx->id;
 332        }
 333
 334        if (copy_to_user((void __user *)(unsigned long)cmd.response,
 335                         &uevent->resp, sizeof uevent->resp)) {
 336                ret = -EFAULT;
 337                goto done;
 338        }
 339
 340        list_del(&uevent->list);
 341        uevent->ctx->events_reported++;
 342        if (uevent->mc)
 343                uevent->mc->events_reported++;
 344        kfree(uevent);
 345done:
 346        mutex_unlock(&file->mut);
 347        return ret;
 348}
 349
 350static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
 351{
 352        switch (cmd->ps) {
 353        case RDMA_PS_TCP:
 354                *qp_type = IB_QPT_RC;
 355                return 0;
 356        case RDMA_PS_UDP:
 357        case RDMA_PS_IPOIB:
 358                *qp_type = IB_QPT_UD;
 359                return 0;
 360        case RDMA_PS_IB:
 361                *qp_type = cmd->qp_type;
 362                return 0;
 363        default:
 364                return -EINVAL;
 365        }
 366}
 367
 368static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
 369                              int in_len, int out_len)
 370{
 371        struct rdma_ucm_create_id cmd;
 372        struct rdma_ucm_create_id_resp resp;
 373        struct ucma_context *ctx;
 374        enum ib_qp_type qp_type;
 375        int ret;
 376
 377        if (out_len < sizeof(resp))
 378                return -ENOSPC;
 379
 380        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 381                return -EFAULT;
 382
 383        ret = ucma_get_qp_type(&cmd, &qp_type);
 384        if (ret)
 385                return ret;
 386
 387        mutex_lock(&file->mut);
 388        ctx = ucma_alloc_ctx(file);
 389        mutex_unlock(&file->mut);
 390        if (!ctx)
 391                return -ENOMEM;
 392
 393        ctx->uid = cmd.uid;
 394        ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps, qp_type);
 395        if (IS_ERR(ctx->cm_id)) {
 396                ret = PTR_ERR(ctx->cm_id);
 397                goto err1;
 398        }
 399
 400        resp.id = ctx->id;
 401        if (copy_to_user((void __user *)(unsigned long)cmd.response,
 402                         &resp, sizeof(resp))) {
 403                ret = -EFAULT;
 404                goto err2;
 405        }
 406        return 0;
 407
 408err2:
 409        rdma_destroy_id(ctx->cm_id);
 410err1:
 411        mutex_lock(&mut);
 412        idr_remove(&ctx_idr, ctx->id);
 413        mutex_unlock(&mut);
 414        kfree(ctx);
 415        return ret;
 416}
 417
 418static void ucma_cleanup_multicast(struct ucma_context *ctx)
 419{
 420        struct ucma_multicast *mc, *tmp;
 421
 422        mutex_lock(&mut);
 423        list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
 424                list_del(&mc->list);
 425                idr_remove(&multicast_idr, mc->id);
 426                kfree(mc);
 427        }
 428        mutex_unlock(&mut);
 429}
 430
 431static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
 432{
 433        struct ucma_event *uevent, *tmp;
 434
 435        list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
 436                if (uevent->mc != mc)
 437                        continue;
 438
 439                list_del(&uevent->list);
 440                kfree(uevent);
 441        }
 442}
 443
 444/*
 445 * We cannot hold file->mut when calling rdma_destroy_id() or we can
 446 * deadlock.  We also acquire file->mut in ucma_event_handler(), and
 447 * rdma_destroy_id() will wait until all callbacks have completed.
 448 */
 449static int ucma_free_ctx(struct ucma_context *ctx)
 450{
 451        int events_reported;
 452        struct ucma_event *uevent, *tmp;
 453        LIST_HEAD(list);
 454
 455        /* No new events will be generated after destroying the id. */
 456        rdma_destroy_id(ctx->cm_id);
 457
 458        ucma_cleanup_multicast(ctx);
 459
 460        /* Cleanup events not yet reported to the user. */
 461        mutex_lock(&ctx->file->mut);
 462        list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
 463                if (uevent->ctx == ctx)
 464                        list_move_tail(&uevent->list, &list);
 465        }
 466        list_del(&ctx->list);
 467        mutex_unlock(&ctx->file->mut);
 468
 469        list_for_each_entry_safe(uevent, tmp, &list, list) {
 470                list_del(&uevent->list);
 471                if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
 472                        rdma_destroy_id(uevent->cm_id);
 473                kfree(uevent);
 474        }
 475
 476        events_reported = ctx->events_reported;
 477        kfree(ctx);
 478        return events_reported;
 479}
 480
 481static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
 482                               int in_len, int out_len)
 483{
 484        struct rdma_ucm_destroy_id cmd;
 485        struct rdma_ucm_destroy_id_resp resp;
 486        struct ucma_context *ctx;
 487        int ret = 0;
 488
 489        if (out_len < sizeof(resp))
 490                return -ENOSPC;
 491
 492        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 493                return -EFAULT;
 494
 495        mutex_lock(&mut);
 496        ctx = _ucma_find_context(cmd.id, file);
 497        if (!IS_ERR(ctx))
 498                idr_remove(&ctx_idr, ctx->id);
 499        mutex_unlock(&mut);
 500
 501        if (IS_ERR(ctx))
 502                return PTR_ERR(ctx);
 503
 504        ucma_put_ctx(ctx);
 505        wait_for_completion(&ctx->comp);
 506        resp.events_reported = ucma_free_ctx(ctx);
 507
 508        if (copy_to_user((void __user *)(unsigned long)cmd.response,
 509                         &resp, sizeof(resp)))
 510                ret = -EFAULT;
 511
 512        return ret;
 513}
 514
 515static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
 516                              int in_len, int out_len)
 517{
 518        struct rdma_ucm_bind_ip cmd;
 519        struct ucma_context *ctx;
 520        int ret;
 521
 522        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 523                return -EFAULT;
 524
 525        ctx = ucma_get_ctx(file, cmd.id);
 526        if (IS_ERR(ctx))
 527                return PTR_ERR(ctx);
 528
 529        ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
 530        ucma_put_ctx(ctx);
 531        return ret;
 532}
 533
 534static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
 535                         int in_len, int out_len)
 536{
 537        struct rdma_ucm_bind cmd;
 538        struct sockaddr *addr;
 539        struct ucma_context *ctx;
 540        int ret;
 541
 542        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 543                return -EFAULT;
 544
 545        addr = (struct sockaddr *) &cmd.addr;
 546        if (cmd.reserved || !cmd.addr_size || (cmd.addr_size != rdma_addr_size(addr)))
 547                return -EINVAL;
 548
 549        ctx = ucma_get_ctx(file, cmd.id);
 550        if (IS_ERR(ctx))
 551                return PTR_ERR(ctx);
 552
 553        ret = rdma_bind_addr(ctx->cm_id, addr);
 554        ucma_put_ctx(ctx);
 555        return ret;
 556}
 557
 558static ssize_t ucma_resolve_ip(struct ucma_file *file,
 559                               const char __user *inbuf,
 560                               int in_len, int out_len)
 561{
 562        struct rdma_ucm_resolve_ip cmd;
 563        struct ucma_context *ctx;
 564        int ret;
 565
 566        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 567                return -EFAULT;
 568
 569        ctx = ucma_get_ctx(file, cmd.id);
 570        if (IS_ERR(ctx))
 571                return PTR_ERR(ctx);
 572
 573        ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
 574                                (struct sockaddr *) &cmd.dst_addr,
 575                                cmd.timeout_ms);
 576        ucma_put_ctx(ctx);
 577        return ret;
 578}
 579
 580static ssize_t ucma_resolve_addr(struct ucma_file *file,
 581                                 const char __user *inbuf,
 582                                 int in_len, int out_len)
 583{
 584        struct rdma_ucm_resolve_addr cmd;
 585        struct sockaddr *src, *dst;
 586        struct ucma_context *ctx;
 587        int ret;
 588
 589        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 590                return -EFAULT;
 591
 592        src = (struct sockaddr *) &cmd.src_addr;
 593        dst = (struct sockaddr *) &cmd.dst_addr;
 594        if (cmd.reserved || (cmd.src_size && (cmd.src_size != rdma_addr_size(src))) ||
 595            !cmd.dst_size || (cmd.dst_size != rdma_addr_size(dst)))
 596                return -EINVAL;
 597
 598        ctx = ucma_get_ctx(file, cmd.id);
 599        if (IS_ERR(ctx))
 600                return PTR_ERR(ctx);
 601
 602        ret = rdma_resolve_addr(ctx->cm_id, src, dst, cmd.timeout_ms);
 603        ucma_put_ctx(ctx);
 604        return ret;
 605}
 606
 607static ssize_t ucma_resolve_route(struct ucma_file *file,
 608                                  const char __user *inbuf,
 609                                  int in_len, int out_len)
 610{
 611        struct rdma_ucm_resolve_route cmd;
 612        struct ucma_context *ctx;
 613        int ret;
 614
 615        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 616                return -EFAULT;
 617
 618        ctx = ucma_get_ctx(file, cmd.id);
 619        if (IS_ERR(ctx))
 620                return PTR_ERR(ctx);
 621
 622        ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
 623        ucma_put_ctx(ctx);
 624        return ret;
 625}
 626
 627static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
 628                               struct rdma_route *route)
 629{
 630        struct rdma_dev_addr *dev_addr;
 631
 632        resp->num_paths = route->num_paths;
 633        switch (route->num_paths) {
 634        case 0:
 635                dev_addr = &route->addr.dev_addr;
 636                rdma_addr_get_dgid(dev_addr,
 637                                   (union ib_gid *) &resp->ib_route[0].dgid);
 638                rdma_addr_get_sgid(dev_addr,
 639                                   (union ib_gid *) &resp->ib_route[0].sgid);
 640                resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
 641                break;
 642        case 2:
 643                ib_copy_path_rec_to_user(&resp->ib_route[1],
 644                                         &route->path_rec[1]);
 645                /* fall through */
 646        case 1:
 647                ib_copy_path_rec_to_user(&resp->ib_route[0],
 648                                         &route->path_rec[0]);
 649                break;
 650        default:
 651                break;
 652        }
 653}
 654
 655static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
 656                                 struct rdma_route *route)
 657{
 658
 659        resp->num_paths = route->num_paths;
 660        switch (route->num_paths) {
 661        case 0:
 662                rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
 663                            (union ib_gid *)&resp->ib_route[0].dgid);
 664                rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
 665                            (union ib_gid *)&resp->ib_route[0].sgid);
 666                resp->ib_route[0].pkey = cpu_to_be16(0xffff);
 667                break;
 668        case 2:
 669                ib_copy_path_rec_to_user(&resp->ib_route[1],
 670                                         &route->path_rec[1]);
 671                /* fall through */
 672        case 1:
 673                ib_copy_path_rec_to_user(&resp->ib_route[0],
 674                                         &route->path_rec[0]);
 675                break;
 676        default:
 677                break;
 678        }
 679}
 680
 681static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
 682                               struct rdma_route *route)
 683{
 684        struct rdma_dev_addr *dev_addr;
 685
 686        dev_addr = &route->addr.dev_addr;
 687        rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
 688        rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
 689}
 690
 691static ssize_t ucma_query_route(struct ucma_file *file,
 692                                const char __user *inbuf,
 693                                int in_len, int out_len)
 694{
 695        struct rdma_ucm_query cmd;
 696        struct rdma_ucm_query_route_resp resp;
 697        struct ucma_context *ctx;
 698        struct sockaddr *addr;
 699        int ret = 0;
 700
 701        if (out_len < sizeof(resp))
 702                return -ENOSPC;
 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        memset(&resp, 0, sizeof resp);
 712        addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
 713        memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
 714                                     sizeof(struct sockaddr_in) :
 715                                     sizeof(struct sockaddr_in6));
 716        addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
 717        memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
 718                                     sizeof(struct sockaddr_in) :
 719                                     sizeof(struct sockaddr_in6));
 720        if (!ctx->cm_id->device)
 721                goto out;
 722
 723        resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
 724        resp.port_num = ctx->cm_id->port_num;
 725        switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) {
 726        case RDMA_TRANSPORT_IB:
 727                switch (rdma_port_get_link_layer(ctx->cm_id->device,
 728                        ctx->cm_id->port_num)) {
 729                case IB_LINK_LAYER_INFINIBAND:
 730                        ucma_copy_ib_route(&resp, &ctx->cm_id->route);
 731                        break;
 732                case IB_LINK_LAYER_ETHERNET:
 733                        ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
 734                        break;
 735                default:
 736                        break;
 737                }
 738                break;
 739        case RDMA_TRANSPORT_IWARP:
 740                ucma_copy_iw_route(&resp, &ctx->cm_id->route);
 741                break;
 742        default:
 743                break;
 744        }
 745
 746out:
 747        if (copy_to_user((void __user *)(unsigned long)cmd.response,
 748                         &resp, sizeof(resp)))
 749                ret = -EFAULT;
 750
 751        ucma_put_ctx(ctx);
 752        return ret;
 753}
 754
 755static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
 756                                   struct rdma_ucm_query_addr_resp *resp)
 757{
 758        if (!cm_id->device)
 759                return;
 760
 761        resp->node_guid = (__force __u64) cm_id->device->node_guid;
 762        resp->port_num = cm_id->port_num;
 763        resp->pkey = (__force __u16) cpu_to_be16(
 764                     ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
 765}
 766
 767static ssize_t ucma_query_addr(struct ucma_context *ctx,
 768                               void __user *response, int out_len)
 769{
 770        struct rdma_ucm_query_addr_resp resp;
 771        struct sockaddr *addr;
 772        int ret = 0;
 773
 774        if (out_len < sizeof(resp))
 775                return -ENOSPC;
 776
 777        memset(&resp, 0, sizeof resp);
 778
 779        addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
 780        resp.src_size = rdma_addr_size(addr);
 781        memcpy(&resp.src_addr, addr, resp.src_size);
 782
 783        addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
 784        resp.dst_size = rdma_addr_size(addr);
 785        memcpy(&resp.dst_addr, addr, resp.dst_size);
 786
 787        ucma_query_device_addr(ctx->cm_id, &resp);
 788
 789        if (copy_to_user(response, &resp, sizeof(resp)))
 790                ret = -EFAULT;
 791
 792        return ret;
 793}
 794
 795static ssize_t ucma_query_path(struct ucma_context *ctx,
 796                               void __user *response, int out_len)
 797{
 798        struct rdma_ucm_query_path_resp *resp;
 799        int i, ret = 0;
 800
 801        if (out_len < sizeof(*resp))
 802                return -ENOSPC;
 803
 804        resp = kzalloc(out_len, GFP_KERNEL);
 805        if (!resp)
 806                return -ENOMEM;
 807
 808        resp->num_paths = ctx->cm_id->route.num_paths;
 809        for (i = 0, out_len -= sizeof(*resp);
 810             i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
 811             i++, out_len -= sizeof(struct ib_path_rec_data)) {
 812
 813                resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
 814                                           IB_PATH_BIDIRECTIONAL;
 815                ib_sa_pack_path(&ctx->cm_id->route.path_rec[i],
 816                                &resp->path_data[i].path_rec);
 817        }
 818
 819        if (copy_to_user(response, resp,
 820                         sizeof(*resp) + (i * sizeof(struct ib_path_rec_data))))
 821                ret = -EFAULT;
 822
 823        kfree(resp);
 824        return ret;
 825}
 826
 827static ssize_t ucma_query_gid(struct ucma_context *ctx,
 828                              void __user *response, int out_len)
 829{
 830        struct rdma_ucm_query_addr_resp resp;
 831        struct sockaddr_ib *addr;
 832        int ret = 0;
 833
 834        if (out_len < sizeof(resp))
 835                return -ENOSPC;
 836
 837        memset(&resp, 0, sizeof resp);
 838
 839        ucma_query_device_addr(ctx->cm_id, &resp);
 840
 841        addr = (struct sockaddr_ib *) &resp.src_addr;
 842        resp.src_size = sizeof(*addr);
 843        if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
 844                memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
 845        } else {
 846                addr->sib_family = AF_IB;
 847                addr->sib_pkey = (__force __be16) resp.pkey;
 848                rdma_addr_get_sgid(&ctx->cm_id->route.addr.dev_addr,
 849                                   (union ib_gid *) &addr->sib_addr);
 850                addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
 851                                                    &ctx->cm_id->route.addr.src_addr);
 852        }
 853
 854        addr = (struct sockaddr_ib *) &resp.dst_addr;
 855        resp.dst_size = sizeof(*addr);
 856        if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
 857                memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
 858        } else {
 859                addr->sib_family = AF_IB;
 860                addr->sib_pkey = (__force __be16) resp.pkey;
 861                rdma_addr_get_dgid(&ctx->cm_id->route.addr.dev_addr,
 862                                   (union ib_gid *) &addr->sib_addr);
 863                addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
 864                                                    &ctx->cm_id->route.addr.dst_addr);
 865        }
 866
 867        if (copy_to_user(response, &resp, sizeof(resp)))
 868                ret = -EFAULT;
 869
 870        return ret;
 871}
 872
 873static ssize_t ucma_query(struct ucma_file *file,
 874                          const char __user *inbuf,
 875                          int in_len, int out_len)
 876{
 877        struct rdma_ucm_query cmd;
 878        struct ucma_context *ctx;
 879        void __user *response;
 880        int ret;
 881
 882        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 883                return -EFAULT;
 884
 885        response = (void __user *)(unsigned long) cmd.response;
 886        ctx = ucma_get_ctx(file, cmd.id);
 887        if (IS_ERR(ctx))
 888                return PTR_ERR(ctx);
 889
 890        switch (cmd.option) {
 891        case RDMA_USER_CM_QUERY_ADDR:
 892                ret = ucma_query_addr(ctx, response, out_len);
 893                break;
 894        case RDMA_USER_CM_QUERY_PATH:
 895                ret = ucma_query_path(ctx, response, out_len);
 896                break;
 897        case RDMA_USER_CM_QUERY_GID:
 898                ret = ucma_query_gid(ctx, response, out_len);
 899                break;
 900        default:
 901                ret = -ENOSYS;
 902                break;
 903        }
 904
 905        ucma_put_ctx(ctx);
 906        return ret;
 907}
 908
 909static void ucma_copy_conn_param(struct rdma_cm_id *id,
 910                                 struct rdma_conn_param *dst,
 911                                 struct rdma_ucm_conn_param *src)
 912{
 913        dst->private_data = src->private_data;
 914        dst->private_data_len = src->private_data_len;
 915        dst->responder_resources =src->responder_resources;
 916        dst->initiator_depth = src->initiator_depth;
 917        dst->flow_control = src->flow_control;
 918        dst->retry_count = src->retry_count;
 919        dst->rnr_retry_count = src->rnr_retry_count;
 920        dst->srq = src->srq;
 921        dst->qp_num = src->qp_num;
 922        dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
 923}
 924
 925static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
 926                            int in_len, int out_len)
 927{
 928        struct rdma_ucm_connect cmd;
 929        struct rdma_conn_param conn_param;
 930        struct ucma_context *ctx;
 931        int ret;
 932
 933        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 934                return -EFAULT;
 935
 936        if (!cmd.conn_param.valid)
 937                return -EINVAL;
 938
 939        ctx = ucma_get_ctx(file, cmd.id);
 940        if (IS_ERR(ctx))
 941                return PTR_ERR(ctx);
 942
 943        ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
 944        ret = rdma_connect(ctx->cm_id, &conn_param);
 945        ucma_put_ctx(ctx);
 946        return ret;
 947}
 948
 949static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
 950                           int in_len, int out_len)
 951{
 952        struct rdma_ucm_listen cmd;
 953        struct ucma_context *ctx;
 954        int ret;
 955
 956        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 957                return -EFAULT;
 958
 959        ctx = ucma_get_ctx(file, cmd.id);
 960        if (IS_ERR(ctx))
 961                return PTR_ERR(ctx);
 962
 963        ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
 964                       cmd.backlog : max_backlog;
 965        ret = rdma_listen(ctx->cm_id, ctx->backlog);
 966        ucma_put_ctx(ctx);
 967        return ret;
 968}
 969
 970static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
 971                           int in_len, int out_len)
 972{
 973        struct rdma_ucm_accept cmd;
 974        struct rdma_conn_param conn_param;
 975        struct ucma_context *ctx;
 976        int ret;
 977
 978        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 979                return -EFAULT;
 980
 981        ctx = ucma_get_ctx(file, cmd.id);
 982        if (IS_ERR(ctx))
 983                return PTR_ERR(ctx);
 984
 985        if (cmd.conn_param.valid) {
 986                ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
 987                mutex_lock(&file->mut);
 988                ret = rdma_accept(ctx->cm_id, &conn_param);
 989                if (!ret)
 990                        ctx->uid = cmd.uid;
 991                mutex_unlock(&file->mut);
 992        } else
 993                ret = rdma_accept(ctx->cm_id, NULL);
 994
 995        ucma_put_ctx(ctx);
 996        return ret;
 997}
 998
 999static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1000                           int in_len, int out_len)
1001{
1002        struct rdma_ucm_reject cmd;
1003        struct ucma_context *ctx;
1004        int ret;
1005
1006        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1007                return -EFAULT;
1008
1009        ctx = ucma_get_ctx(file, cmd.id);
1010        if (IS_ERR(ctx))
1011                return PTR_ERR(ctx);
1012
1013        ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
1014        ucma_put_ctx(ctx);
1015        return ret;
1016}
1017
1018static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1019                               int in_len, int out_len)
1020{
1021        struct rdma_ucm_disconnect cmd;
1022        struct ucma_context *ctx;
1023        int ret;
1024
1025        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1026                return -EFAULT;
1027
1028        ctx = ucma_get_ctx(file, cmd.id);
1029        if (IS_ERR(ctx))
1030                return PTR_ERR(ctx);
1031
1032        ret = rdma_disconnect(ctx->cm_id);
1033        ucma_put_ctx(ctx);
1034        return ret;
1035}
1036
1037static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1038                                 const char __user *inbuf,
1039                                 int in_len, int out_len)
1040{
1041        struct rdma_ucm_init_qp_attr cmd;
1042        struct ib_uverbs_qp_attr resp;
1043        struct ucma_context *ctx;
1044        struct ib_qp_attr qp_attr;
1045        int ret;
1046
1047        if (out_len < sizeof(resp))
1048                return -ENOSPC;
1049
1050        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1051                return -EFAULT;
1052
1053        ctx = ucma_get_ctx(file, cmd.id);
1054        if (IS_ERR(ctx))
1055                return PTR_ERR(ctx);
1056
1057        resp.qp_attr_mask = 0;
1058        memset(&qp_attr, 0, sizeof qp_attr);
1059        qp_attr.qp_state = cmd.qp_state;
1060        ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1061        if (ret)
1062                goto out;
1063
1064        ib_copy_qp_attr_to_user(&resp, &qp_attr);
1065        if (copy_to_user((void __user *)(unsigned long)cmd.response,
1066                         &resp, sizeof(resp)))
1067                ret = -EFAULT;
1068
1069out:
1070        ucma_put_ctx(ctx);
1071        return ret;
1072}
1073
1074static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1075                              void *optval, size_t optlen)
1076{
1077        int ret = 0;
1078
1079        switch (optname) {
1080        case RDMA_OPTION_ID_TOS:
1081                if (optlen != sizeof(u8)) {
1082                        ret = -EINVAL;
1083                        break;
1084                }
1085                rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1086                break;
1087        case RDMA_OPTION_ID_REUSEADDR:
1088                if (optlen != sizeof(int)) {
1089                        ret = -EINVAL;
1090                        break;
1091                }
1092                ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1093                break;
1094        case RDMA_OPTION_ID_AFONLY:
1095                if (optlen != sizeof(int)) {
1096                        ret = -EINVAL;
1097                        break;
1098                }
1099                ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1100                break;
1101        default:
1102                ret = -ENOSYS;
1103        }
1104
1105        return ret;
1106}
1107
1108static int ucma_set_ib_path(struct ucma_context *ctx,
1109                            struct ib_path_rec_data *path_data, size_t optlen)
1110{
1111        struct ib_sa_path_rec sa_path;
1112        struct rdma_cm_event event;
1113        int ret;
1114
1115        if (optlen % sizeof(*path_data))
1116                return -EINVAL;
1117
1118        for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1119                if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1120                                         IB_PATH_BIDIRECTIONAL))
1121                        break;
1122        }
1123
1124        if (!optlen)
1125                return -EINVAL;
1126
1127        ib_sa_unpack_path(path_data->path_rec, &sa_path);
1128        ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1);
1129        if (ret)
1130                return ret;
1131
1132        memset(&event, 0, sizeof event);
1133        event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1134        return ucma_event_handler(ctx->cm_id, &event);
1135}
1136
1137static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1138                              void *optval, size_t optlen)
1139{
1140        int ret;
1141
1142        switch (optname) {
1143        case RDMA_OPTION_IB_PATH:
1144                ret = ucma_set_ib_path(ctx, optval, optlen);
1145                break;
1146        default:
1147                ret = -ENOSYS;
1148        }
1149
1150        return ret;
1151}
1152
1153static int ucma_set_option_level(struct ucma_context *ctx, int level,
1154                                 int optname, void *optval, size_t optlen)
1155{
1156        int ret;
1157
1158        switch (level) {
1159        case RDMA_OPTION_ID:
1160                ret = ucma_set_option_id(ctx, optname, optval, optlen);
1161                break;
1162        case RDMA_OPTION_IB:
1163                ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1164                break;
1165        default:
1166                ret = -ENOSYS;
1167        }
1168
1169        return ret;
1170}
1171
1172static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1173                               int in_len, int out_len)
1174{
1175        struct rdma_ucm_set_option cmd;
1176        struct ucma_context *ctx;
1177        void *optval;
1178        int ret;
1179
1180        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1181                return -EFAULT;
1182
1183        ctx = ucma_get_ctx(file, cmd.id);
1184        if (IS_ERR(ctx))
1185                return PTR_ERR(ctx);
1186
1187        optval = memdup_user((void __user *) (unsigned long) cmd.optval,
1188                             cmd.optlen);
1189        if (IS_ERR(optval)) {
1190                ret = PTR_ERR(optval);
1191                goto out;
1192        }
1193
1194        ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1195                                    cmd.optlen);
1196        kfree(optval);
1197
1198out:
1199        ucma_put_ctx(ctx);
1200        return ret;
1201}
1202
1203static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1204                           int in_len, int out_len)
1205{
1206        struct rdma_ucm_notify cmd;
1207        struct ucma_context *ctx;
1208        int ret;
1209
1210        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1211                return -EFAULT;
1212
1213        ctx = ucma_get_ctx(file, cmd.id);
1214        if (IS_ERR(ctx))
1215                return PTR_ERR(ctx);
1216
1217        ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
1218        ucma_put_ctx(ctx);
1219        return ret;
1220}
1221
1222static ssize_t ucma_process_join(struct ucma_file *file,
1223                                 struct rdma_ucm_join_mcast *cmd,  int out_len)
1224{
1225        struct rdma_ucm_create_id_resp resp;
1226        struct ucma_context *ctx;
1227        struct ucma_multicast *mc;
1228        struct sockaddr *addr;
1229        int ret;
1230
1231        if (out_len < sizeof(resp))
1232                return -ENOSPC;
1233
1234        addr = (struct sockaddr *) &cmd->addr;
1235        if (cmd->reserved || !cmd->addr_size || (cmd->addr_size != rdma_addr_size(addr)))
1236                return -EINVAL;
1237
1238        ctx = ucma_get_ctx(file, cmd->id);
1239        if (IS_ERR(ctx))
1240                return PTR_ERR(ctx);
1241
1242        mutex_lock(&file->mut);
1243        mc = ucma_alloc_multicast(ctx);
1244        if (!mc) {
1245                ret = -ENOMEM;
1246                goto err1;
1247        }
1248
1249        mc->uid = cmd->uid;
1250        memcpy(&mc->addr, addr, cmd->addr_size);
1251        ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc);
1252        if (ret)
1253                goto err2;
1254
1255        resp.id = mc->id;
1256        if (copy_to_user((void __user *)(unsigned long) cmd->response,
1257                         &resp, sizeof(resp))) {
1258                ret = -EFAULT;
1259                goto err3;
1260        }
1261
1262        mutex_unlock(&file->mut);
1263        ucma_put_ctx(ctx);
1264        return 0;
1265
1266err3:
1267        rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1268        ucma_cleanup_mc_events(mc);
1269err2:
1270        mutex_lock(&mut);
1271        idr_remove(&multicast_idr, mc->id);
1272        mutex_unlock(&mut);
1273        list_del(&mc->list);
1274        kfree(mc);
1275err1:
1276        mutex_unlock(&file->mut);
1277        ucma_put_ctx(ctx);
1278        return ret;
1279}
1280
1281static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1282                                      const char __user *inbuf,
1283                                      int in_len, int out_len)
1284{
1285        struct rdma_ucm_join_ip_mcast cmd;
1286        struct rdma_ucm_join_mcast join_cmd;
1287
1288        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1289                return -EFAULT;
1290
1291        join_cmd.response = cmd.response;
1292        join_cmd.uid = cmd.uid;
1293        join_cmd.id = cmd.id;
1294        join_cmd.addr_size = rdma_addr_size((struct sockaddr *) &cmd.addr);
1295        join_cmd.reserved = 0;
1296        memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1297
1298        return ucma_process_join(file, &join_cmd, out_len);
1299}
1300
1301static ssize_t ucma_join_multicast(struct ucma_file *file,
1302                                   const char __user *inbuf,
1303                                   int in_len, int out_len)
1304{
1305        struct rdma_ucm_join_mcast cmd;
1306
1307        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1308                return -EFAULT;
1309
1310        return ucma_process_join(file, &cmd, out_len);
1311}
1312
1313static ssize_t ucma_leave_multicast(struct ucma_file *file,
1314                                    const char __user *inbuf,
1315                                    int in_len, int out_len)
1316{
1317        struct rdma_ucm_destroy_id cmd;
1318        struct rdma_ucm_destroy_id_resp resp;
1319        struct ucma_multicast *mc;
1320        int ret = 0;
1321
1322        if (out_len < sizeof(resp))
1323                return -ENOSPC;
1324
1325        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1326                return -EFAULT;
1327
1328        mutex_lock(&mut);
1329        mc = idr_find(&multicast_idr, cmd.id);
1330        if (!mc)
1331                mc = ERR_PTR(-ENOENT);
1332        else if (mc->ctx->file != file)
1333                mc = ERR_PTR(-EINVAL);
1334        else {
1335                idr_remove(&multicast_idr, mc->id);
1336                atomic_inc(&mc->ctx->ref);
1337        }
1338        mutex_unlock(&mut);
1339
1340        if (IS_ERR(mc)) {
1341                ret = PTR_ERR(mc);
1342                goto out;
1343        }
1344
1345        rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1346        mutex_lock(&mc->ctx->file->mut);
1347        ucma_cleanup_mc_events(mc);
1348        list_del(&mc->list);
1349        mutex_unlock(&mc->ctx->file->mut);
1350
1351        ucma_put_ctx(mc->ctx);
1352        resp.events_reported = mc->events_reported;
1353        kfree(mc);
1354
1355        if (copy_to_user((void __user *)(unsigned long)cmd.response,
1356                         &resp, sizeof(resp)))
1357                ret = -EFAULT;
1358out:
1359        return ret;
1360}
1361
1362static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1363{
1364        /* Acquire mutex's based on pointer comparison to prevent deadlock. */
1365        if (file1 < file2) {
1366                mutex_lock(&file1->mut);
1367                mutex_lock(&file2->mut);
1368        } else {
1369                mutex_lock(&file2->mut);
1370                mutex_lock(&file1->mut);
1371        }
1372}
1373
1374static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1375{
1376        if (file1 < file2) {
1377                mutex_unlock(&file2->mut);
1378                mutex_unlock(&file1->mut);
1379        } else {
1380                mutex_unlock(&file1->mut);
1381                mutex_unlock(&file2->mut);
1382        }
1383}
1384
1385static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1386{
1387        struct ucma_event *uevent, *tmp;
1388
1389        list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1390                if (uevent->ctx == ctx)
1391                        list_move_tail(&uevent->list, &file->event_list);
1392}
1393
1394static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1395                               const char __user *inbuf,
1396                               int in_len, int out_len)
1397{
1398        struct rdma_ucm_migrate_id cmd;
1399        struct rdma_ucm_migrate_resp resp;
1400        struct ucma_context *ctx;
1401        struct fd f;
1402        struct ucma_file *cur_file;
1403        int ret = 0;
1404
1405        if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1406                return -EFAULT;
1407
1408        /* Get current fd to protect against it being closed */
1409        f = fdget(cmd.fd);
1410        if (!f.file)
1411                return -ENOENT;
1412
1413        /* Validate current fd and prevent destruction of id. */
1414        ctx = ucma_get_ctx(f.file->private_data, cmd.id);
1415        if (IS_ERR(ctx)) {
1416                ret = PTR_ERR(ctx);
1417                goto file_put;
1418        }
1419
1420        cur_file = ctx->file;
1421        if (cur_file == new_file) {
1422                resp.events_reported = ctx->events_reported;
1423                goto response;
1424        }
1425
1426        /*
1427         * Migrate events between fd's, maintaining order, and avoiding new
1428         * events being added before existing events.
1429         */
1430        ucma_lock_files(cur_file, new_file);
1431        mutex_lock(&mut);
1432
1433        list_move_tail(&ctx->list, &new_file->ctx_list);
1434        ucma_move_events(ctx, new_file);
1435        ctx->file = new_file;
1436        resp.events_reported = ctx->events_reported;
1437
1438        mutex_unlock(&mut);
1439        ucma_unlock_files(cur_file, new_file);
1440
1441response:
1442        if (copy_to_user((void __user *)(unsigned long)cmd.response,
1443                         &resp, sizeof(resp)))
1444                ret = -EFAULT;
1445
1446        ucma_put_ctx(ctx);
1447file_put:
1448        fdput(f);
1449        return ret;
1450}
1451
1452static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1453                                   const char __user *inbuf,
1454                                   int in_len, int out_len) = {
1455        [RDMA_USER_CM_CMD_CREATE_ID]     = ucma_create_id,
1456        [RDMA_USER_CM_CMD_DESTROY_ID]    = ucma_destroy_id,
1457        [RDMA_USER_CM_CMD_BIND_IP]       = ucma_bind_ip,
1458        [RDMA_USER_CM_CMD_RESOLVE_IP]    = ucma_resolve_ip,
1459        [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1460        [RDMA_USER_CM_CMD_QUERY_ROUTE]   = ucma_query_route,
1461        [RDMA_USER_CM_CMD_CONNECT]       = ucma_connect,
1462        [RDMA_USER_CM_CMD_LISTEN]        = ucma_listen,
1463        [RDMA_USER_CM_CMD_ACCEPT]        = ucma_accept,
1464        [RDMA_USER_CM_CMD_REJECT]        = ucma_reject,
1465        [RDMA_USER_CM_CMD_DISCONNECT]    = ucma_disconnect,
1466        [RDMA_USER_CM_CMD_INIT_QP_ATTR]  = ucma_init_qp_attr,
1467        [RDMA_USER_CM_CMD_GET_EVENT]     = ucma_get_event,
1468        [RDMA_USER_CM_CMD_GET_OPTION]    = NULL,
1469        [RDMA_USER_CM_CMD_SET_OPTION]    = ucma_set_option,
1470        [RDMA_USER_CM_CMD_NOTIFY]        = ucma_notify,
1471        [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1472        [RDMA_USER_CM_CMD_LEAVE_MCAST]   = ucma_leave_multicast,
1473        [RDMA_USER_CM_CMD_MIGRATE_ID]    = ucma_migrate_id,
1474        [RDMA_USER_CM_CMD_QUERY]         = ucma_query,
1475        [RDMA_USER_CM_CMD_BIND]          = ucma_bind,
1476        [RDMA_USER_CM_CMD_RESOLVE_ADDR]  = ucma_resolve_addr,
1477        [RDMA_USER_CM_CMD_JOIN_MCAST]    = ucma_join_multicast
1478};
1479
1480static ssize_t ucma_write(struct file *filp, const char __user *buf,
1481                          size_t len, loff_t *pos)
1482{
1483        struct ucma_file *file = filp->private_data;
1484        struct rdma_ucm_cmd_hdr hdr;
1485        ssize_t ret;
1486
1487        if (len < sizeof(hdr))
1488                return -EINVAL;
1489
1490        if (copy_from_user(&hdr, buf, sizeof(hdr)))
1491                return -EFAULT;
1492
1493        if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1494                return -EINVAL;
1495
1496        if (hdr.in + sizeof(hdr) > len)
1497                return -EINVAL;
1498
1499        if (!ucma_cmd_table[hdr.cmd])
1500                return -ENOSYS;
1501
1502        ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1503        if (!ret)
1504                ret = len;
1505
1506        return ret;
1507}
1508
1509static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
1510{
1511        struct ucma_file *file = filp->private_data;
1512        unsigned int mask = 0;
1513
1514        poll_wait(filp, &file->poll_wait, wait);
1515
1516        if (!list_empty(&file->event_list))
1517                mask = POLLIN | POLLRDNORM;
1518
1519        return mask;
1520}
1521
1522/*
1523 * ucma_open() does not need the BKL:
1524 *
1525 *  - no global state is referred to;
1526 *  - there is no ioctl method to race against;
1527 *  - no further module initialization is required for open to work
1528 *    after the device is registered.
1529 */
1530static int ucma_open(struct inode *inode, struct file *filp)
1531{
1532        struct ucma_file *file;
1533
1534        file = kmalloc(sizeof *file, GFP_KERNEL);
1535        if (!file)
1536                return -ENOMEM;
1537
1538        INIT_LIST_HEAD(&file->event_list);
1539        INIT_LIST_HEAD(&file->ctx_list);
1540        init_waitqueue_head(&file->poll_wait);
1541        mutex_init(&file->mut);
1542
1543        filp->private_data = file;
1544        file->filp = filp;
1545
1546        return nonseekable_open(inode, filp);
1547}
1548
1549static int ucma_close(struct inode *inode, struct file *filp)
1550{
1551        struct ucma_file *file = filp->private_data;
1552        struct ucma_context *ctx, *tmp;
1553
1554        mutex_lock(&file->mut);
1555        list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1556                mutex_unlock(&file->mut);
1557
1558                mutex_lock(&mut);
1559                idr_remove(&ctx_idr, ctx->id);
1560                mutex_unlock(&mut);
1561
1562                ucma_free_ctx(ctx);
1563                mutex_lock(&file->mut);
1564        }
1565        mutex_unlock(&file->mut);
1566        kfree(file);
1567        return 0;
1568}
1569
1570static const struct file_operations ucma_fops = {
1571        .owner   = THIS_MODULE,
1572        .open    = ucma_open,
1573        .release = ucma_close,
1574        .write   = ucma_write,
1575        .poll    = ucma_poll,
1576        .llseek  = no_llseek,
1577};
1578
1579static struct miscdevice ucma_misc = {
1580        .minor          = MISC_DYNAMIC_MINOR,
1581        .name           = "rdma_cm",
1582        .nodename       = "infiniband/rdma_cm",
1583        .mode           = 0666,
1584        .fops           = &ucma_fops,
1585};
1586
1587static ssize_t show_abi_version(struct device *dev,
1588                                struct device_attribute *attr,
1589                                char *buf)
1590{
1591        return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1592}
1593static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1594
1595static int __init ucma_init(void)
1596{
1597        int ret;
1598
1599        ret = misc_register(&ucma_misc);
1600        if (ret)
1601                return ret;
1602
1603        ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1604        if (ret) {
1605                printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n");
1606                goto err1;
1607        }
1608
1609        ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1610        if (!ucma_ctl_table_hdr) {
1611                printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n");
1612                ret = -ENOMEM;
1613                goto err2;
1614        }
1615        return 0;
1616err2:
1617        device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1618err1:
1619        misc_deregister(&ucma_misc);
1620        return ret;
1621}
1622
1623static void __exit ucma_cleanup(void)
1624{
1625        unregister_net_sysctl_table(ucma_ctl_table_hdr);
1626        device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1627        misc_deregister(&ucma_misc);
1628        idr_destroy(&ctx_idr);
1629}
1630
1631module_init(ucma_init);
1632module_exit(ucma_cleanup);
1633