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