drivers/infiniband/core/uverbs_main.c

   1 /*
   2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
   6  * Copyright (c) 2005 PathScale, Inc. All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  */
  36
  37 #include <linux/module.h>
  38 #include <linux/init.h>
  39 #include <linux/device.h>
  40 #include <linux/err.h>
  41 #include <linux/fs.h>
  42 #include <linux/poll.h>
  43 #include <linux/sched.h>
  44 #include <linux/file.h>
  45 #include <linux/cdev.h>
  46 #include <linux/anon_inodes.h>
  47 #include <linux/slab.h>
  48
  49 #include <asm/uaccess.h>
  50
  51 #include <rdma/ib.h>
  52
  53 #include "uverbs.h"
  54
  55 MODULE_AUTHOR("Roland Dreier");
  56 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  57 MODULE_LICENSE("Dual BSD/GPL");
  58
  59 enum {
  60         IB_UVERBS_MAJOR       = 231,
  61         IB_UVERBS_BASE_MINOR  = 192,
  62         IB_UVERBS_MAX_DEVICES = 32
  63 };
  64
  65 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  66
  67 static struct class *uverbs_class;
  68
  69 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  70 DEFINE_IDR(ib_uverbs_pd_idr);
  71 DEFINE_IDR(ib_uverbs_mr_idr);
  72 DEFINE_IDR(ib_uverbs_mw_idr);
  73 DEFINE_IDR(ib_uverbs_ah_idr);
  74 DEFINE_IDR(ib_uverbs_cq_idr);
  75 DEFINE_IDR(ib_uverbs_qp_idr);
  76 DEFINE_IDR(ib_uverbs_srq_idr);
  77 DEFINE_IDR(ib_uverbs_xrcd_idr);
  78 DEFINE_IDR(ib_uverbs_rule_idr);
  79
  80 static DEFINE_SPINLOCK(map_lock);
  81 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  82
  83 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  84                                      struct ib_device *ib_dev,
  85                                      const char __user *buf, int in_len,
  86                                      int out_len) = {
  87         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  88         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  89         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  90         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  91         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  92         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  93         [IB_USER_VERBS_CMD_REREG_MR]            = ib_uverbs_rereg_mr,
  94         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  95         [IB_USER_VERBS_CMD_ALLOC_MW]            = ib_uverbs_alloc_mw,
  96         [IB_USER_VERBS_CMD_DEALLOC_MW]          = ib_uverbs_dealloc_mw,
  97         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  98         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  99         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
 100         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
 101         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
 102         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
 103         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
 104         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 105         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 106         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 107         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 108         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 109         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 110         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 111         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 112         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 113         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 114         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 115         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 116         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 117         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 118         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 119         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 120         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 121         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp,
 122 };
 123
 124 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
 125                                     struct ib_device *ib_dev,
 126                                     struct ib_udata *ucore,
 127                                     struct ib_udata *uhw) = {
 128         [IB_USER_VERBS_EX_CMD_CREATE_FLOW]      = ib_uverbs_ex_create_flow,
 129         [IB_USER_VERBS_EX_CMD_DESTROY_FLOW]     = ib_uverbs_ex_destroy_flow,
 130         [IB_USER_VERBS_EX_CMD_QUERY_DEVICE]     = ib_uverbs_ex_query_device,
 131         [IB_USER_VERBS_EX_CMD_CREATE_CQ]        = ib_uverbs_ex_create_cq,
 132         [IB_USER_VERBS_EX_CMD_CREATE_QP]        = ib_uverbs_ex_create_qp,
 133 };
 134
 135 static void ib_uverbs_add_one(struct ib_device *device);
 136 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
 137
 138 int uverbs_dealloc_mw(struct ib_mw *mw)
 139 {
 140         struct ib_pd *pd = mw->pd;
 141         int ret;
 142
 143         ret = mw->device->dealloc_mw(mw);
 144         if (!ret)
 145                 atomic_dec(&pd->usecnt);
 146         return ret;
 147 }
 148
 149 static void ib_uverbs_release_dev(struct kobject *kobj)
 150 {
 151         struct ib_uverbs_device *dev =
 152                 container_of(kobj, struct ib_uverbs_device, kobj);
 153
 154         cleanup_srcu_struct(&dev->disassociate_srcu);
 155         kfree(dev);
 156 }
 157
 158 static struct kobj_type ib_uverbs_dev_ktype = {
 159         .release = ib_uverbs_release_dev,
 160 };
 161
 162 static void ib_uverbs_release_event_file(struct kref *ref)
 163 {
 164         struct ib_uverbs_event_file *file =
 165                 container_of(ref, struct ib_uverbs_event_file, ref);
 166
 167         kfree(file);
 168 }
 169
 170 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 171                           struct ib_uverbs_event_file *ev_file,
 172                           struct ib_ucq_object *uobj)
 173 {
 174         struct ib_uverbs_event *evt, *tmp;
 175
 176         if (ev_file) {
 177                 spin_lock_irq(&ev_file->lock);
 178                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 179                         list_del(&evt->list);
 180                         kfree(evt);
 181                 }
 182                 spin_unlock_irq(&ev_file->lock);
 183
 184                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 185         }
 186
 187         spin_lock_irq(&file->async_file->lock);
 188         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 189                 list_del(&evt->list);
 190                 kfree(evt);
 191         }
 192         spin_unlock_irq(&file->async_file->lock);
 193 }
 194
 195 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 196                               struct ib_uevent_object *uobj)
 197 {
 198         struct ib_uverbs_event *evt, *tmp;
 199
 200         spin_lock_irq(&file->async_file->lock);
 201         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 202                 list_del(&evt->list);
 203                 kfree(evt);
 204         }
 205         spin_unlock_irq(&file->async_file->lock);
 206 }
 207
 208 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 209                                     struct ib_uqp_object *uobj)
 210 {
 211         struct ib_uverbs_mcast_entry *mcast, *tmp;
 212
 213         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 214                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 215                 list_del(&mcast->list);
 216                 kfree(mcast);
 217         }
 218 }
 219
 220 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 221                                       struct ib_ucontext *context)
 222 {
 223         struct ib_uobject *uobj, *tmp;
 224
 225         context->closing = 1;
 226
 227         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 228                 struct ib_ah *ah = uobj->object;
 229
 230                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 231                 ib_destroy_ah(ah);
 232                 kfree(uobj);
 233         }
 234
 235         /* Remove MWs before QPs, in order to support type 2A MWs. */
 236         list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
 237                 struct ib_mw *mw = uobj->object;
 238
 239                 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
 240                 uverbs_dealloc_mw(mw);
 241                 kfree(uobj);
 242         }
 243
 244         list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
 245                 struct ib_flow *flow_id = uobj->object;
 246
 247                 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
 248                 ib_destroy_flow(flow_id);
 249                 kfree(uobj);
 250         }
 251
 252         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 253                 struct ib_qp *qp = uobj->object;
 254                 struct ib_uqp_object *uqp =
 255                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 256
 257                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 258                 if (qp != qp->real_qp) {
 259                         ib_close_qp(qp);
 260                 } else {
 261                         ib_uverbs_detach_umcast(qp, uqp);
 262                         ib_destroy_qp(qp);
 263                 }
 264                 ib_uverbs_release_uevent(file, &uqp->uevent);
 265                 kfree(uqp);
 266         }
 267
 268         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 269                 struct ib_srq *srq = uobj->object;
 270                 struct ib_uevent_object *uevent =
 271                         container_of(uobj, struct ib_uevent_object, uobject);
 272
 273                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 274                 ib_destroy_srq(srq);
 275                 ib_uverbs_release_uevent(file, uevent);
 276                 kfree(uevent);
 277         }
 278
 279         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 280                 struct ib_cq *cq = uobj->object;
 281                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 282                 struct ib_ucq_object *ucq =
 283                         container_of(uobj, struct ib_ucq_object, uobject);
 284
 285                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 286                 ib_destroy_cq(cq);
 287                 ib_uverbs_release_ucq(file, ev_file, ucq);
 288                 kfree(ucq);
 289         }
 290
 291         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 292                 struct ib_mr *mr = uobj->object;
 293
 294                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 295                 ib_dereg_mr(mr);
 296                 kfree(uobj);
 297         }
 298
 299         mutex_lock(&file->device->xrcd_tree_mutex);
 300         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 301                 struct ib_xrcd *xrcd = uobj->object;
 302                 struct ib_uxrcd_object *uxrcd =
 303                         container_of(uobj, struct ib_uxrcd_object, uobject);
 304
 305                 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 306                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 307                 kfree(uxrcd);
 308         }
 309         mutex_unlock(&file->device->xrcd_tree_mutex);
 310
 311         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 312                 struct ib_pd *pd = uobj->object;
 313
 314                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 315                 ib_dealloc_pd(pd);
 316                 kfree(uobj);
 317         }
 318
 319         put_pid(context->tgid);
 320
 321         return context->device->dealloc_ucontext(context);
 322 }
 323
 324 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
 325 {
 326         complete(&dev->comp);
 327 }
 328
 329 static void ib_uverbs_release_file(struct kref *ref)
 330 {
 331         struct ib_uverbs_file *file =
 332                 container_of(ref, struct ib_uverbs_file, ref);
 333         struct ib_device *ib_dev;
 334         int srcu_key;
 335
 336         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 337         ib_dev = srcu_dereference(file->device->ib_dev,
 338                                   &file->device->disassociate_srcu);
 339         if (ib_dev && !ib_dev->disassociate_ucontext)
 340                 module_put(ib_dev->owner);
 341         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 342
 343         if (atomic_dec_and_test(&file->device->refcount))
 344                 ib_uverbs_comp_dev(file->device);
 345
 346         kfree(file);
 347 }
 348
 349 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 350                                     size_t count, loff_t *pos)
 351 {
 352         struct ib_uverbs_event_file *file = filp->private_data;
 353         struct ib_uverbs_event *event;
 354         int eventsz;
 355         int ret = 0;
 356
 357         spin_lock_irq(&file->lock);
 358
 359         while (list_empty(&file->event_list)) {
 360                 spin_unlock_irq(&file->lock);
 361
 362                 if (filp->f_flags & O_NONBLOCK)
 363                         return -EAGAIN;
 364
 365                 if (wait_event_interruptible(file->poll_wait,
 366                                              (!list_empty(&file->event_list) ||
 367                         /* The barriers built into wait_event_interruptible()
 368                          * and wake_up() guarentee this will see the null set
 369                          * without using RCU
 370                          */
 371                                              !file->uverbs_file->device->ib_dev)))
 372                         return -ERESTARTSYS;
 373
 374                 /* If device was disassociated and no event exists set an error */
 375                 if (list_empty(&file->event_list) &&
 376                     !file->uverbs_file->device->ib_dev)
 377                         return -EIO;
 378
 379                 spin_lock_irq(&file->lock);
 380         }
 381
 382         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 383
 384         if (file->is_async)
 385                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 386         else
 387                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 388
 389         if (eventsz > count) {
 390                 ret   = -EINVAL;
 391                 event = NULL;
 392         } else {
 393                 list_del(file->event_list.next);
 394                 if (event->counter) {
 395                         ++(*event->counter);
 396                         list_del(&event->obj_list);
 397                 }
 398         }
 399
 400         spin_unlock_irq(&file->lock);
 401
 402         if (event) {
 403                 if (copy_to_user(buf, event, eventsz))
 404                         ret = -EFAULT;
 405                 else
 406                         ret = eventsz;
 407         }
 408
 409         kfree(event);
 410
 411         return ret;
 412 }
 413
 414 static unsigned int ib_uverbs_event_poll(struct file *filp,
 415                                          struct poll_table_struct *wait)
 416 {
 417         unsigned int pollflags = 0;
 418         struct ib_uverbs_event_file *file = filp->private_data;
 419
 420         poll_wait(filp, &file->poll_wait, wait);
 421
 422         spin_lock_irq(&file->lock);
 423         if (!list_empty(&file->event_list))
 424                 pollflags = POLLIN | POLLRDNORM;
 425         spin_unlock_irq(&file->lock);
 426
 427         return pollflags;
 428 }
 429
 430 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 431 {
 432         struct ib_uverbs_event_file *file = filp->private_data;
 433
 434         return fasync_helper(fd, filp, on, &file->async_queue);
 435 }
 436
 437 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 438 {
 439         struct ib_uverbs_event_file *file = filp->private_data;
 440         struct ib_uverbs_event *entry, *tmp;
 441         int closed_already = 0;
 442
 443         mutex_lock(&file->uverbs_file->device->lists_mutex);
 444         spin_lock_irq(&file->lock);
 445         closed_already = file->is_closed;
 446         file->is_closed = 1;
 447         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 448                 if (entry->counter)
 449                         list_del(&entry->obj_list);
 450                 kfree(entry);
 451         }
 452         spin_unlock_irq(&file->lock);
 453         if (!closed_already) {
 454                 list_del(&file->list);
 455                 if (file->is_async)
 456                         ib_unregister_event_handler(&file->uverbs_file->
 457                                 event_handler);
 458         }
 459         mutex_unlock(&file->uverbs_file->device->lists_mutex);
 460
 461         kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 462         kref_put(&file->ref, ib_uverbs_release_event_file);
 463
 464         return 0;
 465 }
 466
 467 static const struct file_operations uverbs_event_fops = {
 468         .owner   = THIS_MODULE,
 469         .read    = ib_uverbs_event_read,
 470         .poll    = ib_uverbs_event_poll,
 471         .release = ib_uverbs_event_close,
 472         .fasync  = ib_uverbs_event_fasync,
 473         .llseek  = no_llseek,
 474 };
 475
 476 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 477 {
 478         struct ib_uverbs_event_file    *file = cq_context;
 479         struct ib_ucq_object           *uobj;
 480         struct ib_uverbs_event         *entry;
 481         unsigned long                   flags;
 482
 483         if (!file)
 484                 return;
 485
 486         spin_lock_irqsave(&file->lock, flags);
 487         if (file->is_closed) {
 488                 spin_unlock_irqrestore(&file->lock, flags);
 489                 return;
 490         }
 491
 492         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 493         if (!entry) {
 494                 spin_unlock_irqrestore(&file->lock, flags);
 495                 return;
 496         }
 497
 498         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 499
 500         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 501         entry->counter             = &uobj->comp_events_reported;
 502
 503         list_add_tail(&entry->list, &file->event_list);
 504         list_add_tail(&entry->obj_list, &uobj->comp_list);
 505         spin_unlock_irqrestore(&file->lock, flags);
 506
 507         wake_up_interruptible(&file->poll_wait);
 508         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 509 }
 510
 511 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 512                                     __u64 element, __u64 event,
 513                                     struct list_head *obj_list,
 514                                     u32 *counter)
 515 {
 516         struct ib_uverbs_event *entry;
 517         unsigned long flags;
 518
 519         spin_lock_irqsave(&file->async_file->lock, flags);
 520         if (file->async_file->is_closed) {
 521                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 522                 return;
 523         }
 524
 525         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 526         if (!entry) {
 527                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 528                 return;
 529         }
 530
 531         entry->desc.async.element    = element;
 532         entry->desc.async.event_type = event;
 533         entry->desc.async.reserved   = 0;
 534         entry->counter               = counter;
 535
 536         list_add_tail(&entry->list, &file->async_file->event_list);
 537         if (obj_list)
 538                 list_add_tail(&entry->obj_list, obj_list);
 539         spin_unlock_irqrestore(&file->async_file->lock, flags);
 540
 541         wake_up_interruptible(&file->async_file->poll_wait);
 542         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 543 }
 544
 545 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 546 {
 547         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 548                                                   struct ib_ucq_object, uobject);
 549
 550         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 551                                 event->event, &uobj->async_list,
 552                                 &uobj->async_events_reported);
 553 }
 554
 555 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 556 {
 557         struct ib_uevent_object *uobj;
 558
 559         /* for XRC target qp's, check that qp is live */
 560         if (!event->element.qp->uobject || !event->element.qp->uobject->live)
 561                 return;
 562
 563         uobj = container_of(event->element.qp->uobject,
 564                             struct ib_uevent_object, uobject);
 565
 566         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 567                                 event->event, &uobj->event_list,
 568                                 &uobj->events_reported);
 569 }
 570
 571 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 572 {
 573         struct ib_uevent_object *uobj;
 574
 575         uobj = container_of(event->element.srq->uobject,
 576                             struct ib_uevent_object, uobject);
 577
 578         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 579                                 event->event, &uobj->event_list,
 580                                 &uobj->events_reported);
 581 }
 582
 583 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 584                              struct ib_event *event)
 585 {
 586         struct ib_uverbs_file *file =
 587                 container_of(handler, struct ib_uverbs_file, event_handler);
 588
 589         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 590                                 NULL, NULL);
 591 }
 592
 593 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
 594 {
 595         kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 596         file->async_file = NULL;
 597 }
 598
 599 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 600                                         struct ib_device        *ib_dev,
 601                                         int is_async)
 602 {
 603         struct ib_uverbs_event_file *ev_file;
 604         struct file *filp;
 605         int ret;
 606
 607         ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
 608         if (!ev_file)
 609                 return ERR_PTR(-ENOMEM);
 610
 611         kref_init(&ev_file->ref);
 612         spin_lock_init(&ev_file->lock);
 613         INIT_LIST_HEAD(&ev_file->event_list);
 614         init_waitqueue_head(&ev_file->poll_wait);
 615         ev_file->uverbs_file = uverbs_file;
 616         kref_get(&ev_file->uverbs_file->ref);
 617         ev_file->async_queue = NULL;
 618         ev_file->is_closed   = 0;
 619
 620         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 621                                   ev_file, O_RDONLY);
 622         if (IS_ERR(filp))
 623                 goto err_put_refs;
 624
 625         mutex_lock(&uverbs_file->device->lists_mutex);
 626         list_add_tail(&ev_file->list,
 627                       &uverbs_file->device->uverbs_events_file_list);
 628         mutex_unlock(&uverbs_file->device->lists_mutex);
 629
 630         if (is_async) {
 631                 WARN_ON(uverbs_file->async_file);
 632                 uverbs_file->async_file = ev_file;
 633                 kref_get(&uverbs_file->async_file->ref);
 634                 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
 635                                       ib_dev,
 636                                       ib_uverbs_event_handler);
 637                 ret = ib_register_event_handler(&uverbs_file->event_handler);
 638                 if (ret)
 639                         goto err_put_file;
 640
 641                 /* At that point async file stuff was fully set */
 642                 ev_file->is_async = 1;
 643         }
 644
 645         return filp;
 646
 647 err_put_file:
 648         fput(filp);
 649         kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
 650         uverbs_file->async_file = NULL;
 651         return ERR_PTR(ret);
 652
 653 err_put_refs:
 654         kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
 655         kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 656         return filp;
 657 }
 658
 659 /*
 660  * Look up a completion event file by FD.  If lookup is successful,
 661  * takes a ref to the event file struct that it returns; if
 662  * unsuccessful, returns NULL.
 663  */
 664 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 665 {
 666         struct ib_uverbs_event_file *ev_file = NULL;
 667         struct fd f = fdget(fd);
 668
 669         if (!f.file)
 670                 return NULL;
 671
 672         if (f.file->f_op != &uverbs_event_fops)
 673                 goto out;
 674
 675         ev_file = f.file->private_data;
 676         if (ev_file->is_async) {
 677                 ev_file = NULL;
 678                 goto out;
 679         }
 680
 681         kref_get(&ev_file->ref);
 682
 683 out:
 684         fdput(f);
 685         return ev_file;
 686 }
 687
 688 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
 689 {
 690         u64 mask;
 691
 692         if (command <= IB_USER_VERBS_CMD_OPEN_QP)
 693                 mask = ib_dev->uverbs_cmd_mask;
 694         else
 695                 mask = ib_dev->uverbs_ex_cmd_mask;
 696
 697         if (mask & ((u64)1 << command))
 698                 return 0;
 699
 700         return -1;
 701 }
 702
 703 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 704                              size_t count, loff_t *pos)
 705 {
 706         struct ib_uverbs_file *file = filp->private_data;
 707         struct ib_device *ib_dev;
 708         struct ib_uverbs_cmd_hdr hdr;
 709         __u32 command;
 710         __u32 flags;
 711         int srcu_key;
 712         ssize_t ret;
 713
 714         if (WARN_ON_ONCE(!ib_safe_file_access(filp)))
 715                 return -EACCES;
 716
 717         if (count < sizeof hdr)
 718                 return -EINVAL;
 719
 720         if (copy_from_user(&hdr, buf, sizeof hdr))
 721                 return -EFAULT;
 722
 723         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 724         ib_dev = srcu_dereference(file->device->ib_dev,
 725                                   &file->device->disassociate_srcu);
 726         if (!ib_dev) {
 727                 ret = -EIO;
 728                 goto out;
 729         }
 730
 731         if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 732                                    IB_USER_VERBS_CMD_COMMAND_MASK)) {
 733                 ret = -EINVAL;
 734                 goto out;
 735         }
 736
 737         command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 738         if (verify_command_mask(ib_dev, command)) {
 739                 ret = -EOPNOTSUPP;
 740                 goto out;
 741         }
 742
 743         if (!file->ucontext &&
 744             command != IB_USER_VERBS_CMD_GET_CONTEXT) {
 745                 ret = -EINVAL;
 746                 goto out;
 747         }
 748
 749         flags = (hdr.command &
 750                  IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
 751
 752         if (!flags) {
 753                 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
 754                     !uverbs_cmd_table[command]) {
 755                         ret = -EINVAL;
 756                         goto out;
 757                 }
 758
 759                 if (hdr.in_words * 4 != count) {
 760                         ret = -EINVAL;
 761                         goto out;
 762                 }
 763
 764                 ret = uverbs_cmd_table[command](file, ib_dev,
 765                                                  buf + sizeof(hdr),
 766                                                  hdr.in_words * 4,
 767                                                  hdr.out_words * 4);
 768
 769         } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
 770                 struct ib_uverbs_ex_cmd_hdr ex_hdr;
 771                 struct ib_udata ucore;
 772                 struct ib_udata uhw;
 773                 size_t written_count = count;
 774
 775                 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
 776                     !uverbs_ex_cmd_table[command]) {
 777                         ret = -ENOSYS;
 778                         goto out;
 779                 }
 780
 781                 if (!file->ucontext) {
 782                         ret = -EINVAL;
 783                         goto out;
 784                 }
 785
 786                 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
 787                         ret = -EINVAL;
 788                         goto out;
 789                 }
 790
 791                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
 792                         ret = -EFAULT;
 793                         goto out;
 794                 }
 795
 796                 count -= sizeof(hdr) + sizeof(ex_hdr);
 797                 buf += sizeof(hdr) + sizeof(ex_hdr);
 798
 799                 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
 800                         ret = -EINVAL;
 801                         goto out;
 802                 }
 803
 804                 if (ex_hdr.cmd_hdr_reserved) {
 805                         ret = -EINVAL;
 806                         goto out;
 807                 }
 808
 809                 if (ex_hdr.response) {
 810                         if (!hdr.out_words && !ex_hdr.provider_out_words) {
 811                                 ret = -EINVAL;
 812                                 goto out;
 813                         }
 814
 815                         if (!access_ok(VERIFY_WRITE,
 816                                        (void __user *) (unsigned long) ex_hdr.response,
 817                                        (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
 818                                 ret = -EFAULT;
 819                                 goto out;
 820                         }
 821                 } else {
 822                         if (hdr.out_words || ex_hdr.provider_out_words) {
 823                                 ret = -EINVAL;
 824                                 goto out;
 825                         }
 826                 }
 827
 828                 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
 829                                        hdr.in_words * 8, hdr.out_words * 8);
 830
 831                 INIT_UDATA_BUF_OR_NULL(&uhw,
 832                                        buf + ucore.inlen,
 833                                        (unsigned long) ex_hdr.response + ucore.outlen,
 834                                        ex_hdr.provider_in_words * 8,
 835                                        ex_hdr.provider_out_words * 8);
 836
 837                 ret = uverbs_ex_cmd_table[command](file,
 838                                                    ib_dev,
 839                                                    &ucore,
 840                                                    &uhw);
 841                 if (!ret)
 842                         ret = written_count;
 843         } else {
 844                 ret = -ENOSYS;
 845         }
 846
 847 out:
 848         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 849         return ret;
 850 }
 851
 852 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 853 {
 854         struct ib_uverbs_file *file = filp->private_data;
 855         struct ib_device *ib_dev;
 856         int ret = 0;
 857         int srcu_key;
 858
 859         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 860         ib_dev = srcu_dereference(file->device->ib_dev,
 861                                   &file->device->disassociate_srcu);
 862         if (!ib_dev) {
 863                 ret = -EIO;
 864                 goto out;
 865         }
 866
 867         if (!file->ucontext)
 868                 ret = -ENODEV;
 869         else
 870                 ret = ib_dev->mmap(file->ucontext, vma);
 871 out:
 872         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 873         return ret;
 874 }
 875
 876 /*
 877  * ib_uverbs_open() does not need the BKL:
 878  *
 879  *  - the ib_uverbs_device structures are properly reference counted and
 880  *    everything else is purely local to the file being created, so
 881  *    races against other open calls are not a problem;
 882  *  - there is no ioctl method to race against;
 883  *  - the open method will either immediately run -ENXIO, or all
 884  *    required initialization will be done.
 885  */
 886 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 887 {
 888         struct ib_uverbs_device *dev;
 889         struct ib_uverbs_file *file;
 890         struct ib_device *ib_dev;
 891         int ret;
 892         int module_dependent;
 893         int srcu_key;
 894
 895         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 896         if (!atomic_inc_not_zero(&dev->refcount))
 897                 return -ENXIO;
 898
 899         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
 900         mutex_lock(&dev->lists_mutex);
 901         ib_dev = srcu_dereference(dev->ib_dev,
 902                                   &dev->disassociate_srcu);
 903         if (!ib_dev) {
 904                 ret = -EIO;
 905                 goto err;
 906         }
 907
 908         /* In case IB device supports disassociate ucontext, there is no hard
 909          * dependency between uverbs device and its low level device.
 910          */
 911         module_dependent = !(ib_dev->disassociate_ucontext);
 912
 913         if (module_dependent) {
 914                 if (!try_module_get(ib_dev->owner)) {
 915                         ret = -ENODEV;
 916                         goto err;
 917                 }
 918         }
 919
 920         file = kzalloc(sizeof(*file), GFP_KERNEL);
 921         if (!file) {
 922                 ret = -ENOMEM;
 923                 if (module_dependent)
 924                         goto err_module;
 925
 926                 goto err;
 927         }
 928
 929         file->device     = dev;
 930         file->ucontext   = NULL;
 931         file->async_file = NULL;
 932         kref_init(&file->ref);
 933         mutex_init(&file->mutex);
 934         mutex_init(&file->cleanup_mutex);
 935
 936         filp->private_data = file;
 937         kobject_get(&dev->kobj);
 938         list_add_tail(&file->list, &dev->uverbs_file_list);
 939         mutex_unlock(&dev->lists_mutex);
 940         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 941
 942         return nonseekable_open(inode, filp);
 943
 944 err_module:
 945         module_put(ib_dev->owner);
 946
 947 err:
 948         mutex_unlock(&dev->lists_mutex);
 949         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 950         if (atomic_dec_and_test(&dev->refcount))
 951                 ib_uverbs_comp_dev(dev);
 952
 953         return ret;
 954 }
 955
 956 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 957 {
 958         struct ib_uverbs_file *file = filp->private_data;
 959         struct ib_uverbs_device *dev = file->device;
 960
 961         mutex_lock(&file->cleanup_mutex);
 962         if (file->ucontext) {
 963                 ib_uverbs_cleanup_ucontext(file, file->ucontext);
 964                 file->ucontext = NULL;
 965         }
 966         mutex_unlock(&file->cleanup_mutex);
 967
 968         mutex_lock(&file->device->lists_mutex);
 969         if (!file->is_closed) {
 970                 list_del(&file->list);
 971                 file->is_closed = 1;
 972         }
 973         mutex_unlock(&file->device->lists_mutex);
 974
 975         if (file->async_file)
 976                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 977
 978         kref_put(&file->ref, ib_uverbs_release_file);
 979         kobject_put(&dev->kobj);
 980
 981         return 0;
 982 }
 983
 984 static const struct file_operations uverbs_fops = {
 985         .owner   = THIS_MODULE,
 986         .write   = ib_uverbs_write,
 987         .open    = ib_uverbs_open,
 988         .release = ib_uverbs_close,
 989         .llseek  = no_llseek,
 990 };
 991
 992 static const struct file_operations uverbs_mmap_fops = {
 993         .owner   = THIS_MODULE,
 994         .write   = ib_uverbs_write,
 995         .mmap    = ib_uverbs_mmap,
 996         .open    = ib_uverbs_open,
 997         .release = ib_uverbs_close,
 998         .llseek  = no_llseek,
 999 };
1000
1001 static struct ib_client uverbs_client = {
1002         .name   = "uverbs",
1003         .add    = ib_uverbs_add_one,
1004         .remove = ib_uverbs_remove_one
1005 };
1006
1007 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
1008                           char *buf)
1009 {
1010         int ret = -ENODEV;
1011         int srcu_key;
1012         struct ib_uverbs_device *dev = dev_get_drvdata(device);
1013         struct ib_device *ib_dev;
1014
1015         if (!dev)
1016                 return -ENODEV;
1017
1018         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1019         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1020         if (ib_dev)
1021                 ret = sprintf(buf, "%s\n", ib_dev->name);
1022         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1023
1024         return ret;
1025 }
1026 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1027
1028 static ssize_t show_dev_abi_version(struct device *device,
1029                                     struct device_attribute *attr, char *buf)
1030 {
1031         struct ib_uverbs_device *dev = dev_get_drvdata(device);
1032         int ret = -ENODEV;
1033         int srcu_key;
1034         struct ib_device *ib_dev;
1035
1036         if (!dev)
1037                 return -ENODEV;
1038         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1039         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1040         if (ib_dev)
1041                 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1042         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1043
1044         return ret;
1045 }
1046 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1047
1048 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1049                          __stringify(IB_USER_VERBS_ABI_VERSION));
1050
1051 static dev_t overflow_maj;
1052 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1053
1054 /*
1055  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1056  * requesting a new major number and doubling the number of max devices we
1057  * support. It's stupid, but simple.
1058  */
1059 static int find_overflow_devnum(void)
1060 {
1061         int ret;
1062
1063         if (!overflow_maj) {
1064                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1065                                           "infiniband_verbs");
1066                 if (ret) {
1067                         pr_err("user_verbs: couldn't register dynamic device number\n");
1068                         return ret;
1069                 }
1070         }
1071
1072         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1073         if (ret >= IB_UVERBS_MAX_DEVICES)
1074                 return -1;
1075
1076         return ret;
1077 }
1078
1079 static void ib_uverbs_add_one(struct ib_device *device)
1080 {
1081         int devnum;
1082         dev_t base;
1083         struct ib_uverbs_device *uverbs_dev;
1084         int ret;
1085
1086         if (!device->alloc_ucontext)
1087                 return;
1088
1089         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1090         if (!uverbs_dev)
1091                 return;
1092
1093         ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1094         if (ret) {
1095                 kfree(uverbs_dev);
1096                 return;
1097         }
1098
1099         atomic_set(&uverbs_dev->refcount, 1);
1100         init_completion(&uverbs_dev->comp);
1101         uverbs_dev->xrcd_tree = RB_ROOT;
1102         mutex_init(&uverbs_dev->xrcd_tree_mutex);
1103         kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1104         mutex_init(&uverbs_dev->lists_mutex);
1105         INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1106         INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1107
1108         spin_lock(&map_lock);
1109         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1110         if (devnum >= IB_UVERBS_MAX_DEVICES) {
1111                 spin_unlock(&map_lock);
1112                 devnum = find_overflow_devnum();
1113                 if (devnum < 0)
1114                         goto err;
1115
1116                 spin_lock(&map_lock);
1117                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1118                 base = devnum + overflow_maj;
1119                 set_bit(devnum, overflow_map);
1120         } else {
1121                 uverbs_dev->devnum = devnum;
1122                 base = devnum + IB_UVERBS_BASE_DEV;
1123                 set_bit(devnum, dev_map);
1124         }
1125         spin_unlock(&map_lock);
1126
1127         rcu_assign_pointer(uverbs_dev->ib_dev, device);
1128         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1129
1130         cdev_init(&uverbs_dev->cdev, NULL);
1131         uverbs_dev->cdev.owner = THIS_MODULE;
1132         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1133         uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1134         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1135         if (cdev_add(&uverbs_dev->cdev, base, 1))
1136                 goto err_cdev;
1137
1138         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
1139                                         uverbs_dev->cdev.dev, uverbs_dev,
1140                                         "uverbs%d", uverbs_dev->devnum);
1141         if (IS_ERR(uverbs_dev->dev))
1142                 goto err_cdev;
1143
1144         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1145                 goto err_class;
1146         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1147                 goto err_class;
1148
1149         ib_set_client_data(device, &uverbs_client, uverbs_dev);
1150
1151         return;
1152
1153 err_class:
1154         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1155
1156 err_cdev:
1157         cdev_del(&uverbs_dev->cdev);
1158         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1159                 clear_bit(devnum, dev_map);
1160         else
1161                 clear_bit(devnum, overflow_map);
1162
1163 err:
1164         if (atomic_dec_and_test(&uverbs_dev->refcount))
1165                 ib_uverbs_comp_dev(uverbs_dev);
1166         wait_for_completion(&uverbs_dev->comp);
1167         kobject_put(&uverbs_dev->kobj);
1168         return;
1169 }
1170
1171 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1172                                         struct ib_device *ib_dev)
1173 {
1174         struct ib_uverbs_file *file;
1175         struct ib_uverbs_event_file *event_file;
1176         struct ib_event event;
1177
1178         /* Pending running commands to terminate */
1179         synchronize_srcu(&uverbs_dev->disassociate_srcu);
1180         event.event = IB_EVENT_DEVICE_FATAL;
1181         event.element.port_num = 0;
1182         event.device = ib_dev;
1183
1184         mutex_lock(&uverbs_dev->lists_mutex);
1185         while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1186                 struct ib_ucontext *ucontext;
1187                 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1188                                         struct ib_uverbs_file, list);
1189                 file->is_closed = 1;
1190                 list_del(&file->list);
1191                 kref_get(&file->ref);
1192                 mutex_unlock(&uverbs_dev->lists_mutex);
1193
1194                 ib_uverbs_event_handler(&file->event_handler, &event);
1195
1196                 mutex_lock(&file->cleanup_mutex);
1197                 ucontext = file->ucontext;
1198                 file->ucontext = NULL;
1199                 mutex_unlock(&file->cleanup_mutex);
1200
1201                 /* At this point ib_uverbs_close cannot be running
1202                  * ib_uverbs_cleanup_ucontext
1203                  */
1204                 if (ucontext) {
1205                         /* We must release the mutex before going ahead and
1206                          * calling disassociate_ucontext. disassociate_ucontext
1207                          * might end up indirectly calling uverbs_close,
1208                          * for example due to freeing the resources
1209                          * (e.g mmput).
1210                          */
1211                         ib_dev->disassociate_ucontext(ucontext);
1212                         ib_uverbs_cleanup_ucontext(file, ucontext);
1213                 }
1214
1215                 mutex_lock(&uverbs_dev->lists_mutex);
1216                 kref_put(&file->ref, ib_uverbs_release_file);
1217         }
1218
1219         while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1220                 event_file = list_first_entry(&uverbs_dev->
1221                                               uverbs_events_file_list,
1222                                               struct ib_uverbs_event_file,
1223                                               list);
1224                 spin_lock_irq(&event_file->lock);
1225                 event_file->is_closed = 1;
1226                 spin_unlock_irq(&event_file->lock);
1227
1228                 list_del(&event_file->list);
1229                 if (event_file->is_async) {
1230                         ib_unregister_event_handler(&event_file->uverbs_file->
1231                                                     event_handler);
1232                         event_file->uverbs_file->event_handler.device = NULL;
1233                 }
1234
1235                 wake_up_interruptible(&event_file->poll_wait);
1236                 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1237         }
1238         mutex_unlock(&uverbs_dev->lists_mutex);
1239 }
1240
1241 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1242 {
1243         struct ib_uverbs_device *uverbs_dev = client_data;
1244         int wait_clients = 1;
1245
1246         if (!uverbs_dev)
1247                 return;
1248
1249         dev_set_drvdata(uverbs_dev->dev, NULL);
1250         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1251         cdev_del(&uverbs_dev->cdev);
1252
1253         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1254                 clear_bit(uverbs_dev->devnum, dev_map);
1255         else
1256                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1257
1258         if (device->disassociate_ucontext) {
1259                 /* We disassociate HW resources and immediately return.
1260                  * Userspace will see a EIO errno for all future access.
1261                  * Upon returning, ib_device may be freed internally and is not
1262                  * valid any more.
1263                  * uverbs_device is still available until all clients close
1264                  * their files, then the uverbs device ref count will be zero
1265                  * and its resources will be freed.
1266                  * Note: At this point no more files can be opened since the
1267                  * cdev was deleted, however active clients can still issue
1268                  * commands and close their open files.
1269                  */
1270                 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1271                 ib_uverbs_free_hw_resources(uverbs_dev, device);
1272                 wait_clients = 0;
1273         }
1274
1275         if (atomic_dec_and_test(&uverbs_dev->refcount))
1276                 ib_uverbs_comp_dev(uverbs_dev);
1277         if (wait_clients)
1278                 wait_for_completion(&uverbs_dev->comp);
1279         kobject_put(&uverbs_dev->kobj);
1280 }
1281
1282 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1283 {
1284         if (mode)
1285                 *mode = 0666;
1286         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1287 }
1288
1289 static int __init ib_uverbs_init(void)
1290 {
1291         int ret;
1292
1293         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1294                                      "infiniband_verbs");
1295         if (ret) {
1296                 pr_err("user_verbs: couldn't register device number\n");
1297                 goto out;
1298         }
1299
1300         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1301         if (IS_ERR(uverbs_class)) {
1302                 ret = PTR_ERR(uverbs_class);
1303                 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1304                 goto out_chrdev;
1305         }
1306
1307         uverbs_class->devnode = uverbs_devnode;
1308
1309         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1310         if (ret) {
1311                 pr_err("user_verbs: couldn't create abi_version attribute\n");
1312                 goto out_class;
1313         }
1314
1315         ret = ib_register_client(&uverbs_client);
1316         if (ret) {
1317                 pr_err("user_verbs: couldn't register client\n");
1318                 goto out_class;
1319         }
1320
1321         return 0;
1322
1323 out_class:
1324         class_destroy(uverbs_class);
1325
1326 out_chrdev:
1327         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1328
1329 out:
1330         return ret;
1331 }
1332
1333 static void __exit ib_uverbs_cleanup(void)
1334 {
1335         ib_unregister_client(&uverbs_client);
1336         class_destroy(uverbs_class);
1337         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1338         if (overflow_maj)
1339                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1340         idr_destroy(&ib_uverbs_pd_idr);
1341         idr_destroy(&ib_uverbs_mr_idr);
1342         idr_destroy(&ib_uverbs_mw_idr);
1343         idr_destroy(&ib_uverbs_ah_idr);
1344         idr_destroy(&ib_uverbs_cq_idr);
1345         idr_destroy(&ib_uverbs_qp_idr);
1346         idr_destroy(&ib_uverbs_srq_idr);
1347 }
1348
1349 module_init(ib_uverbs_init);
1350 module_exit(ib_uverbs_cleanup);