drivers/infiniband/core/ucma.c

   1 /*
   2  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *      copyright notice, this list of conditions and the following
  16  *      disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *      copyright notice, this list of conditions and the following
  20  *      disclaimer in the documentation and/or other materials
  21  *      provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/completion.h>
  34 #include <linux/file.h>
  35 #include <linux/mutex.h>
  36 #include <linux/poll.h>
  37 #include <linux/sched.h>
  38 #include <linux/idr.h>
  39 #include <linux/in.h>
  40 #include <linux/in6.h>
  41 #include <linux/miscdevice.h>
  42 #include <linux/slab.h>
  43 #include <linux/sysctl.h>
  44 #include <linux/module.h>
  45 #include <linux/nsproxy.h>
  46
  47 #include <rdma/rdma_user_cm.h>
  48 #include <rdma/ib_marshall.h>
  49 #include <rdma/rdma_cm.h>
  50 #include <rdma/rdma_cm_ib.h>
  51 #include <rdma/ib_addr.h>
  52 #include <rdma/ib.h>
  53
  54 MODULE_AUTHOR("Sean Hefty");
  55 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
  56 MODULE_LICENSE("Dual BSD/GPL");
  57
  58 static unsigned int max_backlog = 1024;
  59
  60 static struct ctl_table_header *ucma_ctl_table_hdr;
  61 static struct ctl_table ucma_ctl_table[] = {
  62         {
  63                 .procname       = "max_backlog",
  64                 .data           = &max_backlog,
  65                 .maxlen         = sizeof max_backlog,
  66                 .mode           = 0644,
  67                 .proc_handler   = proc_dointvec,
  68         },
  69         { }
  70 };
  71
  72 struct ucma_file {
  73         struct mutex            mut;
  74         struct file             *filp;
  75         struct list_head        ctx_list;
  76         struct list_head        event_list;
  77         wait_queue_head_t       poll_wait;
  78         struct workqueue_struct *close_wq;
  79 };
  80
  81 struct ucma_context {
  82         int                     id;
  83         struct completion       comp;
  84         atomic_t                ref;
  85         int                     events_reported;
  86         int                     backlog;
  87
  88         struct ucma_file        *file;
  89         struct rdma_cm_id       *cm_id;
  90         u64                     uid;
  91
  92         struct list_head        list;
  93         struct list_head        mc_list;
  94         /* mark that device is in process of destroying the internal HW
  95          * resources, protected by the global mut
  96          */
  97         int                     closing;
  98         /* sync between removal event and id destroy, protected by file mut */
  99         int                     destroying;
 100         struct work_struct      close_work;
 101 };
 102
 103 struct ucma_multicast {
 104         struct ucma_context     *ctx;
 105         int                     id;
 106         int                     events_reported;
 107
 108         u64                     uid;
 109         struct list_head        list;
 110         struct sockaddr_storage addr;
 111 };
 112
 113 struct ucma_event {
 114         struct ucma_context     *ctx;
 115         struct ucma_multicast   *mc;
 116         struct list_head        list;
 117         struct rdma_cm_id       *cm_id;
 118         struct rdma_ucm_event_resp resp;
 119         struct work_struct      close_work;
 120 };
 121
 122 static DEFINE_MUTEX(mut);
 123 static DEFINE_IDR(ctx_idr);
 124 static DEFINE_IDR(multicast_idr);
 125
 126 static inline struct ucma_context *_ucma_find_context(int id,
 127                                                       struct ucma_file *file)
 128 {
 129         struct ucma_context *ctx;
 130
 131         ctx = idr_find(&ctx_idr, id);
 132         if (!ctx)
 133                 ctx = ERR_PTR(-ENOENT);
 134         else if (ctx->file != file)
 135                 ctx = ERR_PTR(-EINVAL);
 136         return ctx;
 137 }
 138
 139 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
 140 {
 141         struct ucma_context *ctx;
 142
 143         mutex_lock(&mut);
 144         ctx = _ucma_find_context(id, file);
 145         if (!IS_ERR(ctx)) {
 146                 if (ctx->closing)
 147                         ctx = ERR_PTR(-EIO);
 148                 else
 149                         atomic_inc(&ctx->ref);
 150         }
 151         mutex_unlock(&mut);
 152         return ctx;
 153 }
 154
 155 static void ucma_put_ctx(struct ucma_context *ctx)
 156 {
 157         if (atomic_dec_and_test(&ctx->ref))
 158                 complete(&ctx->comp);
 159 }
 160
 161 static void ucma_close_event_id(struct work_struct *work)
 162 {
 163         struct ucma_event *uevent_close =  container_of(work, struct ucma_event, close_work);
 164
 165         rdma_destroy_id(uevent_close->cm_id);
 166         kfree(uevent_close);
 167 }
 168
 169 static void ucma_close_id(struct work_struct *work)
 170 {
 171         struct ucma_context *ctx =  container_of(work, struct ucma_context, close_work);
 172
 173         /* once all inflight tasks are finished, we close all underlying
 174          * resources. The context is still alive till its explicit destryoing
 175          * by its creator.
 176          */
 177         ucma_put_ctx(ctx);
 178         wait_for_completion(&ctx->comp);
 179         /* No new events will be generated after destroying the id. */
 180         rdma_destroy_id(ctx->cm_id);
 181 }
 182
 183 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
 184 {
 185         struct ucma_context *ctx;
 186
 187         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 188         if (!ctx)
 189                 return NULL;
 190
 191         INIT_WORK(&ctx->close_work, ucma_close_id);
 192         atomic_set(&ctx->ref, 1);
 193         init_completion(&ctx->comp);
 194         INIT_LIST_HEAD(&ctx->mc_list);
 195         ctx->file = file;
 196
 197         mutex_lock(&mut);
 198         ctx->id = idr_alloc(&ctx_idr, ctx, 0, 0, GFP_KERNEL);
 199         mutex_unlock(&mut);
 200         if (ctx->id < 0)
 201                 goto error;
 202
 203         list_add_tail(&ctx->list, &file->ctx_list);
 204         return ctx;
 205
 206 error:
 207         kfree(ctx);
 208         return NULL;
 209 }
 210
 211 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
 212 {
 213         struct ucma_multicast *mc;
 214
 215         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
 216         if (!mc)
 217                 return NULL;
 218
 219         mutex_lock(&mut);
 220         mc->id = idr_alloc(&multicast_idr, mc, 0, 0, GFP_KERNEL);
 221         mutex_unlock(&mut);
 222         if (mc->id < 0)
 223                 goto error;
 224
 225         mc->ctx = ctx;
 226         list_add_tail(&mc->list, &ctx->mc_list);
 227         return mc;
 228
 229 error:
 230         kfree(mc);
 231         return NULL;
 232 }
 233
 234 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
 235                                  struct rdma_conn_param *src)
 236 {
 237         if (src->private_data_len)
 238                 memcpy(dst->private_data, src->private_data,
 239                        src->private_data_len);
 240         dst->private_data_len = src->private_data_len;
 241         dst->responder_resources =src->responder_resources;
 242         dst->initiator_depth = src->initiator_depth;
 243         dst->flow_control = src->flow_control;
 244         dst->retry_count = src->retry_count;
 245         dst->rnr_retry_count = src->rnr_retry_count;
 246         dst->srq = src->srq;
 247         dst->qp_num = src->qp_num;
 248 }
 249
 250 static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst,
 251                                struct rdma_ud_param *src)
 252 {
 253         if (src->private_data_len)
 254                 memcpy(dst->private_data, src->private_data,
 255                        src->private_data_len);
 256         dst->private_data_len = src->private_data_len;
 257         ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr);
 258         dst->qp_num = src->qp_num;
 259         dst->qkey = src->qkey;
 260 }
 261
 262 static void ucma_set_event_context(struct ucma_context *ctx,
 263                                    struct rdma_cm_event *event,
 264                                    struct ucma_event *uevent)
 265 {
 266         uevent->ctx = ctx;
 267         switch (event->event) {
 268         case RDMA_CM_EVENT_MULTICAST_JOIN:
 269         case RDMA_CM_EVENT_MULTICAST_ERROR:
 270                 uevent->mc = (struct ucma_multicast *)
 271                              event->param.ud.private_data;
 272                 uevent->resp.uid = uevent->mc->uid;
 273                 uevent->resp.id = uevent->mc->id;
 274                 break;
 275         default:
 276                 uevent->resp.uid = ctx->uid;
 277                 uevent->resp.id = ctx->id;
 278                 break;
 279         }
 280 }
 281
 282 /* Called with file->mut locked for the relevant context. */
 283 static void ucma_removal_event_handler(struct rdma_cm_id *cm_id)
 284 {
 285         struct ucma_context *ctx = cm_id->context;
 286         struct ucma_event *con_req_eve;
 287         int event_found = 0;
 288
 289         if (ctx->destroying)
 290                 return;
 291
 292         /* only if context is pointing to cm_id that it owns it and can be
 293          * queued to be closed, otherwise that cm_id is an inflight one that
 294          * is part of that context event list pending to be detached and
 295          * reattached to its new context as part of ucma_get_event,
 296          * handled separately below.
 297          */
 298         if (ctx->cm_id == cm_id) {
 299                 mutex_lock(&mut);
 300                 ctx->closing = 1;
 301                 mutex_unlock(&mut);
 302                 queue_work(ctx->file->close_wq, &ctx->close_work);
 303                 return;
 304         }
 305
 306         list_for_each_entry(con_req_eve, &ctx->file->event_list, list) {
 307                 if (con_req_eve->cm_id == cm_id &&
 308                     con_req_eve->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
 309                         list_del(&con_req_eve->list);
 310                         INIT_WORK(&con_req_eve->close_work, ucma_close_event_id);
 311                         queue_work(ctx->file->close_wq, &con_req_eve->close_work);
 312                         event_found = 1;
 313                         break;
 314                 }
 315         }
 316         if (!event_found)
 317                 pr_err("ucma_removal_event_handler: warning: connect request event wasn't found\n");
 318 }
 319
 320 static int ucma_event_handler(struct rdma_cm_id *cm_id,
 321                               struct rdma_cm_event *event)
 322 {
 323         struct ucma_event *uevent;
 324         struct ucma_context *ctx = cm_id->context;
 325         int ret = 0;
 326
 327         uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
 328         if (!uevent)
 329                 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
 330
 331         mutex_lock(&ctx->file->mut);
 332         uevent->cm_id = cm_id;
 333         ucma_set_event_context(ctx, event, uevent);
 334         uevent->resp.event = event->event;
 335         uevent->resp.status = event->status;
 336         if (cm_id->qp_type == IB_QPT_UD)
 337                 ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud);
 338         else
 339                 ucma_copy_conn_event(&uevent->resp.param.conn,
 340                                      &event->param.conn);
 341
 342         if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
 343                 if (!ctx->backlog) {
 344                         ret = -ENOMEM;
 345                         kfree(uevent);
 346                         goto out;
 347                 }
 348                 ctx->backlog--;
 349         } else if (!ctx->uid || ctx->cm_id != cm_id) {
 350                 /*
 351                  * We ignore events for new connections until userspace has set
 352                  * their context.  This can only happen if an error occurs on a
 353                  * new connection before the user accepts it.  This is okay,
 354                  * since the accept will just fail later. However, we do need
 355                  * to release the underlying HW resources in case of a device
 356                  * removal event.
 357                  */
 358                 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
 359                         ucma_removal_event_handler(cm_id);
 360
 361                 kfree(uevent);
 362                 goto out;
 363         }
 364
 365         list_add_tail(&uevent->list, &ctx->file->event_list);
 366         wake_up_interruptible(&ctx->file->poll_wait);
 367         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
 368                 ucma_removal_event_handler(cm_id);
 369 out:
 370         mutex_unlock(&ctx->file->mut);
 371         return ret;
 372 }
 373
 374 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
 375                               int in_len, int out_len)
 376 {
 377         struct ucma_context *ctx;
 378         struct rdma_ucm_get_event cmd;
 379         struct ucma_event *uevent;
 380         int ret = 0;
 381
 382         if (out_len < sizeof uevent->resp)
 383                 return -ENOSPC;
 384
 385         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 386                 return -EFAULT;
 387
 388         mutex_lock(&file->mut);
 389         while (list_empty(&file->event_list)) {
 390                 mutex_unlock(&file->mut);
 391
 392                 if (file->filp->f_flags & O_NONBLOCK)
 393                         return -EAGAIN;
 394
 395                 if (wait_event_interruptible(file->poll_wait,
 396                                              !list_empty(&file->event_list)))
 397                         return -ERESTARTSYS;
 398
 399                 mutex_lock(&file->mut);
 400         }
 401
 402         uevent = list_entry(file->event_list.next, struct ucma_event, list);
 403
 404         if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
 405                 ctx = ucma_alloc_ctx(file);
 406                 if (!ctx) {
 407                         ret = -ENOMEM;
 408                         goto done;
 409                 }
 410                 uevent->ctx->backlog++;
 411                 ctx->cm_id = uevent->cm_id;
 412                 ctx->cm_id->context = ctx;
 413                 uevent->resp.id = ctx->id;
 414         }
 415
 416         if (copy_to_user((void __user *)(unsigned long)cmd.response,
 417                          &uevent->resp, sizeof uevent->resp)) {
 418                 ret = -EFAULT;
 419                 goto done;
 420         }
 421
 422         list_del(&uevent->list);
 423         uevent->ctx->events_reported++;
 424         if (uevent->mc)
 425                 uevent->mc->events_reported++;
 426         kfree(uevent);
 427 done:
 428         mutex_unlock(&file->mut);
 429         return ret;
 430 }
 431
 432 static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
 433 {
 434         switch (cmd->ps) {
 435         case RDMA_PS_TCP:
 436                 *qp_type = IB_QPT_RC;
 437                 return 0;
 438         case RDMA_PS_UDP:
 439         case RDMA_PS_IPOIB:
 440                 *qp_type = IB_QPT_UD;
 441                 return 0;
 442         case RDMA_PS_IB:
 443                 *qp_type = cmd->qp_type;
 444                 return 0;
 445         default:
 446                 return -EINVAL;
 447         }
 448 }
 449
 450 static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
 451                               int in_len, int out_len)
 452 {
 453         struct rdma_ucm_create_id cmd;
 454         struct rdma_ucm_create_id_resp resp;
 455         struct ucma_context *ctx;
 456         enum ib_qp_type qp_type;
 457         int ret;
 458
 459         if (out_len < sizeof(resp))
 460                 return -ENOSPC;
 461
 462         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 463                 return -EFAULT;
 464
 465         ret = ucma_get_qp_type(&cmd, &qp_type);
 466         if (ret)
 467                 return ret;
 468
 469         mutex_lock(&file->mut);
 470         ctx = ucma_alloc_ctx(file);
 471         mutex_unlock(&file->mut);
 472         if (!ctx)
 473                 return -ENOMEM;
 474
 475         ctx->uid = cmd.uid;
 476         ctx->cm_id = rdma_create_id(current->nsproxy->net_ns,
 477                                     ucma_event_handler, ctx, cmd.ps, qp_type);
 478         if (IS_ERR(ctx->cm_id)) {
 479                 ret = PTR_ERR(ctx->cm_id);
 480                 goto err1;
 481         }
 482
 483         resp.id = ctx->id;
 484         if (copy_to_user((void __user *)(unsigned long)cmd.response,
 485                          &resp, sizeof(resp))) {
 486                 ret = -EFAULT;
 487                 goto err2;
 488         }
 489         return 0;
 490
 491 err2:
 492         rdma_destroy_id(ctx->cm_id);
 493 err1:
 494         mutex_lock(&mut);
 495         idr_remove(&ctx_idr, ctx->id);
 496         mutex_unlock(&mut);
 497         kfree(ctx);
 498         return ret;
 499 }
 500
 501 static void ucma_cleanup_multicast(struct ucma_context *ctx)
 502 {
 503         struct ucma_multicast *mc, *tmp;
 504
 505         mutex_lock(&mut);
 506         list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
 507                 list_del(&mc->list);
 508                 idr_remove(&multicast_idr, mc->id);
 509                 kfree(mc);
 510         }
 511         mutex_unlock(&mut);
 512 }
 513
 514 static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
 515 {
 516         struct ucma_event *uevent, *tmp;
 517
 518         list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
 519                 if (uevent->mc != mc)
 520                         continue;
 521
 522                 list_del(&uevent->list);
 523                 kfree(uevent);
 524         }
 525 }
 526
 527 /*
 528  * ucma_free_ctx is called after the underlying rdma CM-ID is destroyed. At
 529  * this point, no new events will be reported from the hardware. However, we
 530  * still need to cleanup the UCMA context for this ID. Specifically, there
 531  * might be events that have not yet been consumed by the user space software.
 532  * These might include pending connect requests which we have not completed
 533  * processing.  We cannot call rdma_destroy_id while holding the lock of the
 534  * context (file->mut), as it might cause a deadlock. We therefore extract all
 535  * relevant events from the context pending events list while holding the
 536  * mutex. After that we release them as needed.
 537  */
 538 static int ucma_free_ctx(struct ucma_context *ctx)
 539 {
 540         int events_reported;
 541         struct ucma_event *uevent, *tmp;
 542         LIST_HEAD(list);
 543
 544
 545         ucma_cleanup_multicast(ctx);
 546
 547         /* Cleanup events not yet reported to the user. */
 548         mutex_lock(&ctx->file->mut);
 549         list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
 550                 if (uevent->ctx == ctx)
 551                         list_move_tail(&uevent->list, &list);
 552         }
 553         list_del(&ctx->list);
 554         mutex_unlock(&ctx->file->mut);
 555
 556         list_for_each_entry_safe(uevent, tmp, &list, list) {
 557                 list_del(&uevent->list);
 558                 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
 559                         rdma_destroy_id(uevent->cm_id);
 560                 kfree(uevent);
 561         }
 562
 563         events_reported = ctx->events_reported;
 564         kfree(ctx);
 565         return events_reported;
 566 }
 567
 568 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
 569                                int in_len, int out_len)
 570 {
 571         struct rdma_ucm_destroy_id cmd;
 572         struct rdma_ucm_destroy_id_resp resp;
 573         struct ucma_context *ctx;
 574         int ret = 0;
 575
 576         if (out_len < sizeof(resp))
 577                 return -ENOSPC;
 578
 579         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 580                 return -EFAULT;
 581
 582         mutex_lock(&mut);
 583         ctx = _ucma_find_context(cmd.id, file);
 584         if (!IS_ERR(ctx))
 585                 idr_remove(&ctx_idr, ctx->id);
 586         mutex_unlock(&mut);
 587
 588         if (IS_ERR(ctx))
 589                 return PTR_ERR(ctx);
 590
 591         mutex_lock(&ctx->file->mut);
 592         ctx->destroying = 1;
 593         mutex_unlock(&ctx->file->mut);
 594
 595         flush_workqueue(ctx->file->close_wq);
 596         /* At this point it's guaranteed that there is no inflight
 597          * closing task */
 598         mutex_lock(&mut);
 599         if (!ctx->closing) {
 600                 mutex_unlock(&mut);
 601                 ucma_put_ctx(ctx);
 602                 wait_for_completion(&ctx->comp);
 603                 rdma_destroy_id(ctx->cm_id);
 604         } else {
 605                 mutex_unlock(&mut);
 606         }
 607
 608         resp.events_reported = ucma_free_ctx(ctx);
 609         if (copy_to_user((void __user *)(unsigned long)cmd.response,
 610                          &resp, sizeof(resp)))
 611                 ret = -EFAULT;
 612
 613         return ret;
 614 }
 615
 616 static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
 617                               int in_len, int out_len)
 618 {
 619         struct rdma_ucm_bind_ip cmd;
 620         struct ucma_context *ctx;
 621         int ret;
 622
 623         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 624                 return -EFAULT;
 625
 626         ctx = ucma_get_ctx(file, cmd.id);
 627         if (IS_ERR(ctx))
 628                 return PTR_ERR(ctx);
 629
 630         ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
 631         ucma_put_ctx(ctx);
 632         return ret;
 633 }
 634
 635 static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
 636                          int in_len, int out_len)
 637 {
 638         struct rdma_ucm_bind cmd;
 639         struct sockaddr *addr;
 640         struct ucma_context *ctx;
 641         int ret;
 642
 643         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 644                 return -EFAULT;
 645
 646         addr = (struct sockaddr *) &cmd.addr;
 647         if (cmd.reserved || !cmd.addr_size || (cmd.addr_size != rdma_addr_size(addr)))
 648                 return -EINVAL;
 649
 650         ctx = ucma_get_ctx(file, cmd.id);
 651         if (IS_ERR(ctx))
 652                 return PTR_ERR(ctx);
 653
 654         ret = rdma_bind_addr(ctx->cm_id, addr);
 655         ucma_put_ctx(ctx);
 656         return ret;
 657 }
 658
 659 static ssize_t ucma_resolve_ip(struct ucma_file *file,
 660                                const char __user *inbuf,
 661                                int in_len, int out_len)
 662 {
 663         struct rdma_ucm_resolve_ip cmd;
 664         struct ucma_context *ctx;
 665         int ret;
 666
 667         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 668                 return -EFAULT;
 669
 670         ctx = ucma_get_ctx(file, cmd.id);
 671         if (IS_ERR(ctx))
 672                 return PTR_ERR(ctx);
 673
 674         ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
 675                                 (struct sockaddr *) &cmd.dst_addr,
 676                                 cmd.timeout_ms);
 677         ucma_put_ctx(ctx);
 678         return ret;
 679 }
 680
 681 static ssize_t ucma_resolve_addr(struct ucma_file *file,
 682                                  const char __user *inbuf,
 683                                  int in_len, int out_len)
 684 {
 685         struct rdma_ucm_resolve_addr cmd;
 686         struct sockaddr *src, *dst;
 687         struct ucma_context *ctx;
 688         int ret;
 689
 690         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 691                 return -EFAULT;
 692
 693         src = (struct sockaddr *) &cmd.src_addr;
 694         dst = (struct sockaddr *) &cmd.dst_addr;
 695         if (cmd.reserved || (cmd.src_size && (cmd.src_size != rdma_addr_size(src))) ||
 696             !cmd.dst_size || (cmd.dst_size != rdma_addr_size(dst)))
 697                 return -EINVAL;
 698
 699         ctx = ucma_get_ctx(file, cmd.id);
 700         if (IS_ERR(ctx))
 701                 return PTR_ERR(ctx);
 702
 703         ret = rdma_resolve_addr(ctx->cm_id, src, dst, cmd.timeout_ms);
 704         ucma_put_ctx(ctx);
 705         return ret;
 706 }
 707
 708 static ssize_t ucma_resolve_route(struct ucma_file *file,
 709                                   const char __user *inbuf,
 710                                   int in_len, int out_len)
 711 {
 712         struct rdma_ucm_resolve_route cmd;
 713         struct ucma_context *ctx;
 714         int ret;
 715
 716         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 717                 return -EFAULT;
 718
 719         ctx = ucma_get_ctx(file, cmd.id);
 720         if (IS_ERR(ctx))
 721                 return PTR_ERR(ctx);
 722
 723         ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
 724         ucma_put_ctx(ctx);
 725         return ret;
 726 }
 727
 728 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
 729                                struct rdma_route *route)
 730 {
 731         struct rdma_dev_addr *dev_addr;
 732
 733         resp->num_paths = route->num_paths;
 734         switch (route->num_paths) {
 735         case 0:
 736                 dev_addr = &route->addr.dev_addr;
 737                 rdma_addr_get_dgid(dev_addr,
 738                                    (union ib_gid *) &resp->ib_route[0].dgid);
 739                 rdma_addr_get_sgid(dev_addr,
 740                                    (union ib_gid *) &resp->ib_route[0].sgid);
 741                 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
 742                 break;
 743         case 2:
 744                 ib_copy_path_rec_to_user(&resp->ib_route[1],
 745                                          &route->path_rec[1]);
 746                 /* fall through */
 747         case 1:
 748                 ib_copy_path_rec_to_user(&resp->ib_route[0],
 749                                          &route->path_rec[0]);
 750                 break;
 751         default:
 752                 break;
 753         }
 754 }
 755
 756 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
 757                                  struct rdma_route *route)
 758 {
 759
 760         resp->num_paths = route->num_paths;
 761         switch (route->num_paths) {
 762         case 0:
 763                 rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
 764                             (union ib_gid *)&resp->ib_route[0].dgid);
 765                 rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
 766                             (union ib_gid *)&resp->ib_route[0].sgid);
 767                 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
 768                 break;
 769         case 2:
 770                 ib_copy_path_rec_to_user(&resp->ib_route[1],
 771                                          &route->path_rec[1]);
 772                 /* fall through */
 773         case 1:
 774                 ib_copy_path_rec_to_user(&resp->ib_route[0],
 775                                          &route->path_rec[0]);
 776                 break;
 777         default:
 778                 break;
 779         }
 780 }
 781
 782 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
 783                                struct rdma_route *route)
 784 {
 785         struct rdma_dev_addr *dev_addr;
 786
 787         dev_addr = &route->addr.dev_addr;
 788         rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
 789         rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
 790 }
 791
 792 static ssize_t ucma_query_route(struct ucma_file *file,
 793                                 const char __user *inbuf,
 794                                 int in_len, int out_len)
 795 {
 796         struct rdma_ucm_query cmd;
 797         struct rdma_ucm_query_route_resp resp;
 798         struct ucma_context *ctx;
 799         struct sockaddr *addr;
 800         int ret = 0;
 801
 802         if (out_len < sizeof(resp))
 803                 return -ENOSPC;
 804
 805         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 806                 return -EFAULT;
 807
 808         ctx = ucma_get_ctx(file, cmd.id);
 809         if (IS_ERR(ctx))
 810                 return PTR_ERR(ctx);
 811
 812         memset(&resp, 0, sizeof resp);
 813         addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
 814         memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
 815                                      sizeof(struct sockaddr_in) :
 816                                      sizeof(struct sockaddr_in6));
 817         addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
 818         memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
 819                                      sizeof(struct sockaddr_in) :
 820                                      sizeof(struct sockaddr_in6));
 821         if (!ctx->cm_id->device)
 822                 goto out;
 823
 824         resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
 825         resp.port_num = ctx->cm_id->port_num;
 826
 827         if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num))
 828                 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
 829         else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num))
 830                 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
 831         else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num))
 832                 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
 833
 834 out:
 835         if (copy_to_user((void __user *)(unsigned long)cmd.response,
 836                          &resp, sizeof(resp)))
 837                 ret = -EFAULT;
 838
 839         ucma_put_ctx(ctx);
 840         return ret;
 841 }
 842
 843 static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
 844                                    struct rdma_ucm_query_addr_resp *resp)
 845 {
 846         if (!cm_id->device)
 847                 return;
 848
 849         resp->node_guid = (__force __u64) cm_id->device->node_guid;
 850         resp->port_num = cm_id->port_num;
 851         resp->pkey = (__force __u16) cpu_to_be16(
 852                      ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
 853 }
 854
 855 static ssize_t ucma_query_addr(struct ucma_context *ctx,
 856                                void __user *response, int out_len)
 857 {
 858         struct rdma_ucm_query_addr_resp resp;
 859         struct sockaddr *addr;
 860         int ret = 0;
 861
 862         if (out_len < sizeof(resp))
 863                 return -ENOSPC;
 864
 865         memset(&resp, 0, sizeof resp);
 866
 867         addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
 868         resp.src_size = rdma_addr_size(addr);
 869         memcpy(&resp.src_addr, addr, resp.src_size);
 870
 871         addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
 872         resp.dst_size = rdma_addr_size(addr);
 873         memcpy(&resp.dst_addr, addr, resp.dst_size);
 874
 875         ucma_query_device_addr(ctx->cm_id, &resp);
 876
 877         if (copy_to_user(response, &resp, sizeof(resp)))
 878                 ret = -EFAULT;
 879
 880         return ret;
 881 }
 882
 883 static ssize_t ucma_query_path(struct ucma_context *ctx,
 884                                void __user *response, int out_len)
 885 {
 886         struct rdma_ucm_query_path_resp *resp;
 887         int i, ret = 0;
 888
 889         if (out_len < sizeof(*resp))
 890                 return -ENOSPC;
 891
 892         resp = kzalloc(out_len, GFP_KERNEL);
 893         if (!resp)
 894                 return -ENOMEM;
 895
 896         resp->num_paths = ctx->cm_id->route.num_paths;
 897         for (i = 0, out_len -= sizeof(*resp);
 898              i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
 899              i++, out_len -= sizeof(struct ib_path_rec_data)) {
 900
 901                 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
 902                                            IB_PATH_BIDIRECTIONAL;
 903                 ib_sa_pack_path(&ctx->cm_id->route.path_rec[i],
 904                                 &resp->path_data[i].path_rec);
 905         }
 906
 907         if (copy_to_user(response, resp,
 908                          sizeof(*resp) + (i * sizeof(struct ib_path_rec_data))))
 909                 ret = -EFAULT;
 910
 911         kfree(resp);
 912         return ret;
 913 }
 914
 915 static ssize_t ucma_query_gid(struct ucma_context *ctx,
 916                               void __user *response, int out_len)
 917 {
 918         struct rdma_ucm_query_addr_resp resp;
 919         struct sockaddr_ib *addr;
 920         int ret = 0;
 921
 922         if (out_len < sizeof(resp))
 923                 return -ENOSPC;
 924
 925         memset(&resp, 0, sizeof resp);
 926
 927         ucma_query_device_addr(ctx->cm_id, &resp);
 928
 929         addr = (struct sockaddr_ib *) &resp.src_addr;
 930         resp.src_size = sizeof(*addr);
 931         if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
 932                 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
 933         } else {
 934                 addr->sib_family = AF_IB;
 935                 addr->sib_pkey = (__force __be16) resp.pkey;
 936                 rdma_addr_get_sgid(&ctx->cm_id->route.addr.dev_addr,
 937                                    (union ib_gid *) &addr->sib_addr);
 938                 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
 939                                                     &ctx->cm_id->route.addr.src_addr);
 940         }
 941
 942         addr = (struct sockaddr_ib *) &resp.dst_addr;
 943         resp.dst_size = sizeof(*addr);
 944         if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
 945                 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
 946         } else {
 947                 addr->sib_family = AF_IB;
 948                 addr->sib_pkey = (__force __be16) resp.pkey;
 949                 rdma_addr_get_dgid(&ctx->cm_id->route.addr.dev_addr,
 950                                    (union ib_gid *) &addr->sib_addr);
 951                 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
 952                                                     &ctx->cm_id->route.addr.dst_addr);
 953         }
 954
 955         if (copy_to_user(response, &resp, sizeof(resp)))
 956                 ret = -EFAULT;
 957
 958         return ret;
 959 }
 960
 961 static ssize_t ucma_query(struct ucma_file *file,
 962                           const char __user *inbuf,
 963                           int in_len, int out_len)
 964 {
 965         struct rdma_ucm_query cmd;
 966         struct ucma_context *ctx;
 967         void __user *response;
 968         int ret;
 969
 970         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
 971                 return -EFAULT;
 972
 973         response = (void __user *)(unsigned long) cmd.response;
 974         ctx = ucma_get_ctx(file, cmd.id);
 975         if (IS_ERR(ctx))
 976                 return PTR_ERR(ctx);
 977
 978         switch (cmd.option) {
 979         case RDMA_USER_CM_QUERY_ADDR:
 980                 ret = ucma_query_addr(ctx, response, out_len);
 981                 break;
 982         case RDMA_USER_CM_QUERY_PATH:
 983                 ret = ucma_query_path(ctx, response, out_len);
 984                 break;
 985         case RDMA_USER_CM_QUERY_GID:
 986                 ret = ucma_query_gid(ctx, response, out_len);
 987                 break;
 988         default:
 989                 ret = -ENOSYS;
 990                 break;
 991         }
 992
 993         ucma_put_ctx(ctx);
 994         return ret;
 995 }
 996
 997 static void ucma_copy_conn_param(struct rdma_cm_id *id,
 998                                  struct rdma_conn_param *dst,
 999                                  struct rdma_ucm_conn_param *src)
1000 {
1001         dst->private_data = src->private_data;
1002         dst->private_data_len = src->private_data_len;
1003         dst->responder_resources =src->responder_resources;
1004         dst->initiator_depth = src->initiator_depth;
1005         dst->flow_control = src->flow_control;
1006         dst->retry_count = src->retry_count;
1007         dst->rnr_retry_count = src->rnr_retry_count;
1008         dst->srq = src->srq;
1009         dst->qp_num = src->qp_num;
1010         dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
1011 }
1012
1013 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
1014                             int in_len, int out_len)
1015 {
1016         struct rdma_ucm_connect cmd;
1017         struct rdma_conn_param conn_param;
1018         struct ucma_context *ctx;
1019         int ret;
1020
1021         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1022                 return -EFAULT;
1023
1024         if (!cmd.conn_param.valid)
1025                 return -EINVAL;
1026
1027         ctx = ucma_get_ctx(file, cmd.id);
1028         if (IS_ERR(ctx))
1029                 return PTR_ERR(ctx);
1030
1031         ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1032         ret = rdma_connect(ctx->cm_id, &conn_param);
1033         ucma_put_ctx(ctx);
1034         return ret;
1035 }
1036
1037 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
1038                            int in_len, int out_len)
1039 {
1040         struct rdma_ucm_listen cmd;
1041         struct ucma_context *ctx;
1042         int ret;
1043
1044         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1045                 return -EFAULT;
1046
1047         ctx = ucma_get_ctx(file, cmd.id);
1048         if (IS_ERR(ctx))
1049                 return PTR_ERR(ctx);
1050
1051         ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
1052                        cmd.backlog : max_backlog;
1053         ret = rdma_listen(ctx->cm_id, ctx->backlog);
1054         ucma_put_ctx(ctx);
1055         return ret;
1056 }
1057
1058 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
1059                            int in_len, int out_len)
1060 {
1061         struct rdma_ucm_accept cmd;
1062         struct rdma_conn_param conn_param;
1063         struct ucma_context *ctx;
1064         int ret;
1065
1066         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1067                 return -EFAULT;
1068
1069         ctx = ucma_get_ctx(file, cmd.id);
1070         if (IS_ERR(ctx))
1071                 return PTR_ERR(ctx);
1072
1073         if (cmd.conn_param.valid) {
1074                 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1075                 mutex_lock(&file->mut);
1076                 ret = rdma_accept(ctx->cm_id, &conn_param);
1077                 if (!ret)
1078                         ctx->uid = cmd.uid;
1079                 mutex_unlock(&file->mut);
1080         } else
1081                 ret = rdma_accept(ctx->cm_id, NULL);
1082
1083         ucma_put_ctx(ctx);
1084         return ret;
1085 }
1086
1087 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1088                            int in_len, int out_len)
1089 {
1090         struct rdma_ucm_reject cmd;
1091         struct ucma_context *ctx;
1092         int ret;
1093
1094         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1095                 return -EFAULT;
1096
1097         ctx = ucma_get_ctx(file, cmd.id);
1098         if (IS_ERR(ctx))
1099                 return PTR_ERR(ctx);
1100
1101         ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
1102         ucma_put_ctx(ctx);
1103         return ret;
1104 }
1105
1106 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1107                                int in_len, int out_len)
1108 {
1109         struct rdma_ucm_disconnect cmd;
1110         struct ucma_context *ctx;
1111         int ret;
1112
1113         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1114                 return -EFAULT;
1115
1116         ctx = ucma_get_ctx(file, cmd.id);
1117         if (IS_ERR(ctx))
1118                 return PTR_ERR(ctx);
1119
1120         ret = rdma_disconnect(ctx->cm_id);
1121         ucma_put_ctx(ctx);
1122         return ret;
1123 }
1124
1125 static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1126                                  const char __user *inbuf,
1127                                  int in_len, int out_len)
1128 {
1129         struct rdma_ucm_init_qp_attr cmd;
1130         struct ib_uverbs_qp_attr resp;
1131         struct ucma_context *ctx;
1132         struct ib_qp_attr qp_attr;
1133         int ret;
1134
1135         if (out_len < sizeof(resp))
1136                 return -ENOSPC;
1137
1138         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1139                 return -EFAULT;
1140
1141         ctx = ucma_get_ctx(file, cmd.id);
1142         if (IS_ERR(ctx))
1143                 return PTR_ERR(ctx);
1144
1145         resp.qp_attr_mask = 0;
1146         memset(&qp_attr, 0, sizeof qp_attr);
1147         qp_attr.qp_state = cmd.qp_state;
1148         ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1149         if (ret)
1150                 goto out;
1151
1152         ib_copy_qp_attr_to_user(&resp, &qp_attr);
1153         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1154                          &resp, sizeof(resp)))
1155                 ret = -EFAULT;
1156
1157 out:
1158         ucma_put_ctx(ctx);
1159         return ret;
1160 }
1161
1162 static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1163                               void *optval, size_t optlen)
1164 {
1165         int ret = 0;
1166
1167         switch (optname) {
1168         case RDMA_OPTION_ID_TOS:
1169                 if (optlen != sizeof(u8)) {
1170                         ret = -EINVAL;
1171                         break;
1172                 }
1173                 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1174                 break;
1175         case RDMA_OPTION_ID_REUSEADDR:
1176                 if (optlen != sizeof(int)) {
1177                         ret = -EINVAL;
1178                         break;
1179                 }
1180                 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1181                 break;
1182         case RDMA_OPTION_ID_AFONLY:
1183                 if (optlen != sizeof(int)) {
1184                         ret = -EINVAL;
1185                         break;
1186                 }
1187                 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1188                 break;
1189         default:
1190                 ret = -ENOSYS;
1191         }
1192
1193         return ret;
1194 }
1195
1196 static int ucma_set_ib_path(struct ucma_context *ctx,
1197                             struct ib_path_rec_data *path_data, size_t optlen)
1198 {
1199         struct ib_sa_path_rec sa_path;
1200         struct rdma_cm_event event;
1201         int ret;
1202
1203         if (optlen % sizeof(*path_data))
1204                 return -EINVAL;
1205
1206         for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1207                 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1208                                          IB_PATH_BIDIRECTIONAL))
1209                         break;
1210         }
1211
1212         if (!optlen)
1213                 return -EINVAL;
1214
1215         memset(&sa_path, 0, sizeof(sa_path));
1216
1217         ib_sa_unpack_path(path_data->path_rec, &sa_path);
1218         ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1);
1219         if (ret)
1220                 return ret;
1221
1222         memset(&event, 0, sizeof event);
1223         event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1224         return ucma_event_handler(ctx->cm_id, &event);
1225 }
1226
1227 static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1228                               void *optval, size_t optlen)
1229 {
1230         int ret;
1231
1232         switch (optname) {
1233         case RDMA_OPTION_IB_PATH:
1234                 ret = ucma_set_ib_path(ctx, optval, optlen);
1235                 break;
1236         default:
1237                 ret = -ENOSYS;
1238         }
1239
1240         return ret;
1241 }
1242
1243 static int ucma_set_option_level(struct ucma_context *ctx, int level,
1244                                  int optname, void *optval, size_t optlen)
1245 {
1246         int ret;
1247
1248         switch (level) {
1249         case RDMA_OPTION_ID:
1250                 ret = ucma_set_option_id(ctx, optname, optval, optlen);
1251                 break;
1252         case RDMA_OPTION_IB:
1253                 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1254                 break;
1255         default:
1256                 ret = -ENOSYS;
1257         }
1258
1259         return ret;
1260 }
1261
1262 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1263                                int in_len, int out_len)
1264 {
1265         struct rdma_ucm_set_option cmd;
1266         struct ucma_context *ctx;
1267         void *optval;
1268         int ret;
1269
1270         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1271                 return -EFAULT;
1272
1273         ctx = ucma_get_ctx(file, cmd.id);
1274         if (IS_ERR(ctx))
1275                 return PTR_ERR(ctx);
1276
1277         optval = memdup_user((void __user *) (unsigned long) cmd.optval,
1278                              cmd.optlen);
1279         if (IS_ERR(optval)) {
1280                 ret = PTR_ERR(optval);
1281                 goto out;
1282         }
1283
1284         ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1285                                     cmd.optlen);
1286         kfree(optval);
1287
1288 out:
1289         ucma_put_ctx(ctx);
1290         return ret;
1291 }
1292
1293 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1294                            int in_len, int out_len)
1295 {
1296         struct rdma_ucm_notify cmd;
1297         struct ucma_context *ctx;
1298         int ret;
1299
1300         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1301                 return -EFAULT;
1302
1303         ctx = ucma_get_ctx(file, cmd.id);
1304         if (IS_ERR(ctx))
1305                 return PTR_ERR(ctx);
1306
1307         ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
1308         ucma_put_ctx(ctx);
1309         return ret;
1310 }
1311
1312 static ssize_t ucma_process_join(struct ucma_file *file,
1313                                  struct rdma_ucm_join_mcast *cmd,  int out_len)
1314 {
1315         struct rdma_ucm_create_id_resp resp;
1316         struct ucma_context *ctx;
1317         struct ucma_multicast *mc;
1318         struct sockaddr *addr;
1319         int ret;
1320
1321         if (out_len < sizeof(resp))
1322                 return -ENOSPC;
1323
1324         addr = (struct sockaddr *) &cmd->addr;
1325         if (cmd->reserved || !cmd->addr_size || (cmd->addr_size != rdma_addr_size(addr)))
1326                 return -EINVAL;
1327
1328         ctx = ucma_get_ctx(file, cmd->id);
1329         if (IS_ERR(ctx))
1330                 return PTR_ERR(ctx);
1331
1332         mutex_lock(&file->mut);
1333         mc = ucma_alloc_multicast(ctx);
1334         if (!mc) {
1335                 ret = -ENOMEM;
1336                 goto err1;
1337         }
1338
1339         mc->uid = cmd->uid;
1340         memcpy(&mc->addr, addr, cmd->addr_size);
1341         ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc);
1342         if (ret)
1343                 goto err2;
1344
1345         resp.id = mc->id;
1346         if (copy_to_user((void __user *)(unsigned long) cmd->response,
1347                          &resp, sizeof(resp))) {
1348                 ret = -EFAULT;
1349                 goto err3;
1350         }
1351
1352         mutex_unlock(&file->mut);
1353         ucma_put_ctx(ctx);
1354         return 0;
1355
1356 err3:
1357         rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1358         ucma_cleanup_mc_events(mc);
1359 err2:
1360         mutex_lock(&mut);
1361         idr_remove(&multicast_idr, mc->id);
1362         mutex_unlock(&mut);
1363         list_del(&mc->list);
1364         kfree(mc);
1365 err1:
1366         mutex_unlock(&file->mut);
1367         ucma_put_ctx(ctx);
1368         return ret;
1369 }
1370
1371 static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1372                                       const char __user *inbuf,
1373                                       int in_len, int out_len)
1374 {
1375         struct rdma_ucm_join_ip_mcast cmd;
1376         struct rdma_ucm_join_mcast join_cmd;
1377
1378         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1379                 return -EFAULT;
1380
1381         join_cmd.response = cmd.response;
1382         join_cmd.uid = cmd.uid;
1383         join_cmd.id = cmd.id;
1384         join_cmd.addr_size = rdma_addr_size((struct sockaddr *) &cmd.addr);
1385         join_cmd.reserved = 0;
1386         memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1387
1388         return ucma_process_join(file, &join_cmd, out_len);
1389 }
1390
1391 static ssize_t ucma_join_multicast(struct ucma_file *file,
1392                                    const char __user *inbuf,
1393                                    int in_len, int out_len)
1394 {
1395         struct rdma_ucm_join_mcast cmd;
1396
1397         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1398                 return -EFAULT;
1399
1400         return ucma_process_join(file, &cmd, out_len);
1401 }
1402
1403 static ssize_t ucma_leave_multicast(struct ucma_file *file,
1404                                     const char __user *inbuf,
1405                                     int in_len, int out_len)
1406 {
1407         struct rdma_ucm_destroy_id cmd;
1408         struct rdma_ucm_destroy_id_resp resp;
1409         struct ucma_multicast *mc;
1410         int ret = 0;
1411
1412         if (out_len < sizeof(resp))
1413                 return -ENOSPC;
1414
1415         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1416                 return -EFAULT;
1417
1418         mutex_lock(&mut);
1419         mc = idr_find(&multicast_idr, cmd.id);
1420         if (!mc)
1421                 mc = ERR_PTR(-ENOENT);
1422         else if (mc->ctx->file != file)
1423                 mc = ERR_PTR(-EINVAL);
1424         else if (!atomic_inc_not_zero(&mc->ctx->ref))
1425                 mc = ERR_PTR(-ENXIO);
1426         else
1427                 idr_remove(&multicast_idr, mc->id);
1428         mutex_unlock(&mut);
1429
1430         if (IS_ERR(mc)) {
1431                 ret = PTR_ERR(mc);
1432                 goto out;
1433         }
1434
1435         rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1436         mutex_lock(&mc->ctx->file->mut);
1437         ucma_cleanup_mc_events(mc);
1438         list_del(&mc->list);
1439         mutex_unlock(&mc->ctx->file->mut);
1440
1441         ucma_put_ctx(mc->ctx);
1442         resp.events_reported = mc->events_reported;
1443         kfree(mc);
1444
1445         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1446                          &resp, sizeof(resp)))
1447                 ret = -EFAULT;
1448 out:
1449         return ret;
1450 }
1451
1452 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1453 {
1454         /* Acquire mutex's based on pointer comparison to prevent deadlock. */
1455         if (file1 < file2) {
1456                 mutex_lock(&file1->mut);
1457                 mutex_lock_nested(&file2->mut, SINGLE_DEPTH_NESTING);
1458         } else {
1459                 mutex_lock(&file2->mut);
1460                 mutex_lock_nested(&file1->mut, SINGLE_DEPTH_NESTING);
1461         }
1462 }
1463
1464 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1465 {
1466         if (file1 < file2) {
1467                 mutex_unlock(&file2->mut);
1468                 mutex_unlock(&file1->mut);
1469         } else {
1470                 mutex_unlock(&file1->mut);
1471                 mutex_unlock(&file2->mut);
1472         }
1473 }
1474
1475 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1476 {
1477         struct ucma_event *uevent, *tmp;
1478
1479         list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1480                 if (uevent->ctx == ctx)
1481                         list_move_tail(&uevent->list, &file->event_list);
1482 }
1483
1484 static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1485                                const char __user *inbuf,
1486                                int in_len, int out_len)
1487 {
1488         struct rdma_ucm_migrate_id cmd;
1489         struct rdma_ucm_migrate_resp resp;
1490         struct ucma_context *ctx;
1491         struct fd f;
1492         struct ucma_file *cur_file;
1493         int ret = 0;
1494
1495         if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1496                 return -EFAULT;
1497
1498         /* Get current fd to protect against it being closed */
1499         f = fdget(cmd.fd);
1500         if (!f.file)
1501                 return -ENOENT;
1502
1503         /* Validate current fd and prevent destruction of id. */
1504         ctx = ucma_get_ctx(f.file->private_data, cmd.id);
1505         if (IS_ERR(ctx)) {
1506                 ret = PTR_ERR(ctx);
1507                 goto file_put;
1508         }
1509
1510         cur_file = ctx->file;
1511         if (cur_file == new_file) {
1512                 resp.events_reported = ctx->events_reported;
1513                 goto response;
1514         }
1515
1516         /*
1517          * Migrate events between fd's, maintaining order, and avoiding new
1518          * events being added before existing events.
1519          */
1520         ucma_lock_files(cur_file, new_file);
1521         mutex_lock(&mut);
1522
1523         list_move_tail(&ctx->list, &new_file->ctx_list);
1524         ucma_move_events(ctx, new_file);
1525         ctx->file = new_file;
1526         resp.events_reported = ctx->events_reported;
1527
1528         mutex_unlock(&mut);
1529         ucma_unlock_files(cur_file, new_file);
1530
1531 response:
1532         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1533                          &resp, sizeof(resp)))
1534                 ret = -EFAULT;
1535
1536         ucma_put_ctx(ctx);
1537 file_put:
1538         fdput(f);
1539         return ret;
1540 }
1541
1542 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1543                                    const char __user *inbuf,
1544                                    int in_len, int out_len) = {
1545         [RDMA_USER_CM_CMD_CREATE_ID]     = ucma_create_id,
1546         [RDMA_USER_CM_CMD_DESTROY_ID]    = ucma_destroy_id,
1547         [RDMA_USER_CM_CMD_BIND_IP]       = ucma_bind_ip,
1548         [RDMA_USER_CM_CMD_RESOLVE_IP]    = ucma_resolve_ip,
1549         [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1550         [RDMA_USER_CM_CMD_QUERY_ROUTE]   = ucma_query_route,
1551         [RDMA_USER_CM_CMD_CONNECT]       = ucma_connect,
1552         [RDMA_USER_CM_CMD_LISTEN]        = ucma_listen,
1553         [RDMA_USER_CM_CMD_ACCEPT]        = ucma_accept,
1554         [RDMA_USER_CM_CMD_REJECT]        = ucma_reject,
1555         [RDMA_USER_CM_CMD_DISCONNECT]    = ucma_disconnect,
1556         [RDMA_USER_CM_CMD_INIT_QP_ATTR]  = ucma_init_qp_attr,
1557         [RDMA_USER_CM_CMD_GET_EVENT]     = ucma_get_event,
1558         [RDMA_USER_CM_CMD_GET_OPTION]    = NULL,
1559         [RDMA_USER_CM_CMD_SET_OPTION]    = ucma_set_option,
1560         [RDMA_USER_CM_CMD_NOTIFY]        = ucma_notify,
1561         [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1562         [RDMA_USER_CM_CMD_LEAVE_MCAST]   = ucma_leave_multicast,
1563         [RDMA_USER_CM_CMD_MIGRATE_ID]    = ucma_migrate_id,
1564         [RDMA_USER_CM_CMD_QUERY]         = ucma_query,
1565         [RDMA_USER_CM_CMD_BIND]          = ucma_bind,
1566         [RDMA_USER_CM_CMD_RESOLVE_ADDR]  = ucma_resolve_addr,
1567         [RDMA_USER_CM_CMD_JOIN_MCAST]    = ucma_join_multicast
1568 };
1569
1570 static ssize_t ucma_write(struct file *filp, const char __user *buf,
1571                           size_t len, loff_t *pos)
1572 {
1573         struct ucma_file *file = filp->private_data;
1574         struct rdma_ucm_cmd_hdr hdr;
1575         ssize_t ret;
1576
1577         if (len < sizeof(hdr))
1578                 return -EINVAL;
1579
1580         if (copy_from_user(&hdr, buf, sizeof(hdr)))
1581                 return -EFAULT;
1582
1583         if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1584                 return -EINVAL;
1585
1586         if (hdr.in + sizeof(hdr) > len)
1587                 return -EINVAL;
1588
1589         if (!ucma_cmd_table[hdr.cmd])
1590                 return -ENOSYS;
1591
1592         ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1593         if (!ret)
1594                 ret = len;
1595
1596         return ret;
1597 }
1598
1599 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
1600 {
1601         struct ucma_file *file = filp->private_data;
1602         unsigned int mask = 0;
1603
1604         poll_wait(filp, &file->poll_wait, wait);
1605
1606         if (!list_empty(&file->event_list))
1607                 mask = POLLIN | POLLRDNORM;
1608
1609         return mask;
1610 }
1611
1612 /*
1613  * ucma_open() does not need the BKL:
1614  *
1615  *  - no global state is referred to;
1616  *  - there is no ioctl method to race against;
1617  *  - no further module initialization is required for open to work
1618  *    after the device is registered.
1619  */
1620 static int ucma_open(struct inode *inode, struct file *filp)
1621 {
1622         struct ucma_file *file;
1623
1624         file = kmalloc(sizeof *file, GFP_KERNEL);
1625         if (!file)
1626                 return -ENOMEM;
1627
1628         file->close_wq = create_singlethread_workqueue("ucma_close_id");
1629         if (!file->close_wq) {
1630                 kfree(file);
1631                 return -ENOMEM;
1632         }
1633
1634         INIT_LIST_HEAD(&file->event_list);
1635         INIT_LIST_HEAD(&file->ctx_list);
1636         init_waitqueue_head(&file->poll_wait);
1637         mutex_init(&file->mut);
1638
1639         filp->private_data = file;
1640         file->filp = filp;
1641
1642         return nonseekable_open(inode, filp);
1643 }
1644
1645 static int ucma_close(struct inode *inode, struct file *filp)
1646 {
1647         struct ucma_file *file = filp->private_data;
1648         struct ucma_context *ctx, *tmp;
1649
1650         mutex_lock(&file->mut);
1651         list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1652                 ctx->destroying = 1;
1653                 mutex_unlock(&file->mut);
1654
1655                 mutex_lock(&mut);
1656                 idr_remove(&ctx_idr, ctx->id);
1657                 mutex_unlock(&mut);
1658
1659                 flush_workqueue(file->close_wq);
1660                 /* At that step once ctx was marked as destroying and workqueue
1661                  * was flushed we are safe from any inflights handlers that
1662                  * might put other closing task.
1663                  */
1664                 mutex_lock(&mut);
1665                 if (!ctx->closing) {
1666                         mutex_unlock(&mut);
1667                         /* rdma_destroy_id ensures that no event handlers are
1668                          * inflight for that id before releasing it.
1669                          */
1670                         rdma_destroy_id(ctx->cm_id);
1671                 } else {
1672                         mutex_unlock(&mut);
1673                 }
1674
1675                 ucma_free_ctx(ctx);
1676                 mutex_lock(&file->mut);
1677         }
1678         mutex_unlock(&file->mut);
1679         destroy_workqueue(file->close_wq);
1680         kfree(file);
1681         return 0;
1682 }
1683
1684 static const struct file_operations ucma_fops = {
1685         .owner   = THIS_MODULE,
1686         .open    = ucma_open,
1687         .release = ucma_close,
1688         .write   = ucma_write,
1689         .poll    = ucma_poll,
1690         .llseek  = no_llseek,
1691 };
1692
1693 static struct miscdevice ucma_misc = {
1694         .minor          = MISC_DYNAMIC_MINOR,
1695         .name           = "rdma_cm",
1696         .nodename       = "infiniband/rdma_cm",
1697         .mode           = 0666,
1698         .fops           = &ucma_fops,
1699 };
1700
1701 static ssize_t show_abi_version(struct device *dev,
1702                                 struct device_attribute *attr,
1703                                 char *buf)
1704 {
1705         return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1706 }
1707 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1708
1709 static int __init ucma_init(void)
1710 {
1711         int ret;
1712
1713         ret = misc_register(&ucma_misc);
1714         if (ret)
1715                 return ret;
1716
1717         ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1718         if (ret) {
1719                 pr_err("rdma_ucm: couldn't create abi_version attr\n");
1720                 goto err1;
1721         }
1722
1723         ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1724         if (!ucma_ctl_table_hdr) {
1725                 pr_err("rdma_ucm: couldn't register sysctl paths\n");
1726                 ret = -ENOMEM;
1727                 goto err2;
1728         }
1729         return 0;
1730 err2:
1731         device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1732 err1:
1733         misc_deregister(&ucma_misc);
1734         return ret;
1735 }
1736
1737 static void __exit ucma_cleanup(void)
1738 {
1739         unregister_net_sysctl_table(ucma_ctl_table_hdr);
1740         device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1741         misc_deregister(&ucma_misc);
1742         idr_destroy(&ctx_idr);
1743         idr_destroy(&multicast_idr);
1744 }
1745
1746 module_init(ucma_init);
1747 module_exit(ucma_cleanup);