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 #include <linux/sched/mm.h>
  49
  50 #include <linux/uaccess.h>
  51
  52 #include <rdma/ib.h>
  53 #include <rdma/uverbs_std_types.h>
  54
  55 #include "uverbs.h"
  56 #include "core_priv.h"
  57 #include "rdma_core.h"
  58
  59 MODULE_AUTHOR("Roland Dreier");
  60 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  61 MODULE_LICENSE("Dual BSD/GPL");
  62
  63 enum {
  64         IB_UVERBS_MAJOR       = 231,
  65         IB_UVERBS_BASE_MINOR  = 192,
  66         IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS,
  67         IB_UVERBS_NUM_FIXED_MINOR = 32,
  68         IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR,
  69 };
  70
  71 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  72
  73 static dev_t dynamic_uverbs_dev;
  74 static struct class *uverbs_class;
  75
  76 static DEFINE_IDA(uverbs_ida);
  77 static void ib_uverbs_add_one(struct ib_device *device);
  78 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
  79
  80 /*
  81  * Must be called with the ufile->device->disassociate_srcu held, and the lock
  82  * must be held until use of the ucontext is finished.
  83  */
  84 struct ib_ucontext *ib_uverbs_get_ucontext(struct ib_uverbs_file *ufile)
  85 {
  86         /*
  87          * We do not hold the hw_destroy_rwsem lock for this flow, instead
  88          * srcu is used. It does not matter if someone races this with
  89          * get_context, we get NULL or valid ucontext.
  90          */
  91         struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext);
  92
  93         if (!srcu_dereference(ufile->device->ib_dev,
  94                               &ufile->device->disassociate_srcu))
  95                 return ERR_PTR(-EIO);
  96
  97         if (!ucontext)
  98                 return ERR_PTR(-EINVAL);
  99
 100         return ucontext;
 101 }
 102 EXPORT_SYMBOL(ib_uverbs_get_ucontext);
 103
 104 int uverbs_dealloc_mw(struct ib_mw *mw)
 105 {
 106         struct ib_pd *pd = mw->pd;
 107         int ret;
 108
 109         ret = mw->device->dealloc_mw(mw);
 110         if (!ret)
 111                 atomic_dec(&pd->usecnt);
 112         return ret;
 113 }
 114
 115 static void ib_uverbs_release_dev(struct device *device)
 116 {
 117         struct ib_uverbs_device *dev =
 118                         container_of(device, struct ib_uverbs_device, dev);
 119
 120         uverbs_destroy_api(dev->uapi);
 121         cleanup_srcu_struct(&dev->disassociate_srcu);
 122         kfree(dev);
 123 }
 124
 125 static void ib_uverbs_release_async_event_file(struct kref *ref)
 126 {
 127         struct ib_uverbs_async_event_file *file =
 128                 container_of(ref, struct ib_uverbs_async_event_file, ref);
 129
 130         kfree(file);
 131 }
 132
 133 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 134                           struct ib_uverbs_completion_event_file *ev_file,
 135                           struct ib_ucq_object *uobj)
 136 {
 137         struct ib_uverbs_event *evt, *tmp;
 138
 139         if (ev_file) {
 140                 spin_lock_irq(&ev_file->ev_queue.lock);
 141                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 142                         list_del(&evt->list);
 143                         kfree(evt);
 144                 }
 145                 spin_unlock_irq(&ev_file->ev_queue.lock);
 146
 147                 uverbs_uobject_put(&ev_file->uobj);
 148         }
 149
 150         spin_lock_irq(&file->async_file->ev_queue.lock);
 151         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 152                 list_del(&evt->list);
 153                 kfree(evt);
 154         }
 155         spin_unlock_irq(&file->async_file->ev_queue.lock);
 156 }
 157
 158 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 159                               struct ib_uevent_object *uobj)
 160 {
 161         struct ib_uverbs_event *evt, *tmp;
 162
 163         spin_lock_irq(&file->async_file->ev_queue.lock);
 164         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 165                 list_del(&evt->list);
 166                 kfree(evt);
 167         }
 168         spin_unlock_irq(&file->async_file->ev_queue.lock);
 169 }
 170
 171 void ib_uverbs_detach_umcast(struct ib_qp *qp,
 172                              struct ib_uqp_object *uobj)
 173 {
 174         struct ib_uverbs_mcast_entry *mcast, *tmp;
 175
 176         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 177                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 178                 list_del(&mcast->list);
 179                 kfree(mcast);
 180         }
 181 }
 182
 183 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
 184 {
 185         complete(&dev->comp);
 186 }
 187
 188 void ib_uverbs_release_file(struct kref *ref)
 189 {
 190         struct ib_uverbs_file *file =
 191                 container_of(ref, struct ib_uverbs_file, ref);
 192         struct ib_device *ib_dev;
 193         int srcu_key;
 194
 195         release_ufile_idr_uobject(file);
 196
 197         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 198         ib_dev = srcu_dereference(file->device->ib_dev,
 199                                   &file->device->disassociate_srcu);
 200         if (ib_dev && !ib_dev->disassociate_ucontext)
 201                 module_put(ib_dev->owner);
 202         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 203
 204         if (atomic_dec_and_test(&file->device->refcount))
 205                 ib_uverbs_comp_dev(file->device);
 206
 207         put_device(&file->device->dev);
 208         kfree(file);
 209 }
 210
 211 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
 212                                     struct ib_uverbs_file *uverbs_file,
 213                                     struct file *filp, char __user *buf,
 214                                     size_t count, loff_t *pos,
 215                                     size_t eventsz)
 216 {
 217         struct ib_uverbs_event *event;
 218         int ret = 0;
 219
 220         spin_lock_irq(&ev_queue->lock);
 221
 222         while (list_empty(&ev_queue->event_list)) {
 223                 spin_unlock_irq(&ev_queue->lock);
 224
 225                 if (filp->f_flags & O_NONBLOCK)
 226                         return -EAGAIN;
 227
 228                 if (wait_event_interruptible(ev_queue->poll_wait,
 229                                              (!list_empty(&ev_queue->event_list) ||
 230                         /* The barriers built into wait_event_interruptible()
 231                          * and wake_up() guarentee this will see the null set
 232                          * without using RCU
 233                          */
 234                                              !uverbs_file->device->ib_dev)))
 235                         return -ERESTARTSYS;
 236
 237                 /* If device was disassociated and no event exists set an error */
 238                 if (list_empty(&ev_queue->event_list) &&
 239                     !uverbs_file->device->ib_dev)
 240                         return -EIO;
 241
 242                 spin_lock_irq(&ev_queue->lock);
 243         }
 244
 245         event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
 246
 247         if (eventsz > count) {
 248                 ret   = -EINVAL;
 249                 event = NULL;
 250         } else {
 251                 list_del(ev_queue->event_list.next);
 252                 if (event->counter) {
 253                         ++(*event->counter);
 254                         list_del(&event->obj_list);
 255                 }
 256         }
 257
 258         spin_unlock_irq(&ev_queue->lock);
 259
 260         if (event) {
 261                 if (copy_to_user(buf, event, eventsz))
 262                         ret = -EFAULT;
 263                 else
 264                         ret = eventsz;
 265         }
 266
 267         kfree(event);
 268
 269         return ret;
 270 }
 271
 272 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
 273                                           size_t count, loff_t *pos)
 274 {
 275         struct ib_uverbs_async_event_file *file = filp->private_data;
 276
 277         return ib_uverbs_event_read(&file->ev_queue, file->uverbs_file, filp,
 278                                     buf, count, pos,
 279                                     sizeof(struct ib_uverbs_async_event_desc));
 280 }
 281
 282 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
 283                                          size_t count, loff_t *pos)
 284 {
 285         struct ib_uverbs_completion_event_file *comp_ev_file =
 286                 filp->private_data;
 287
 288         return ib_uverbs_event_read(&comp_ev_file->ev_queue,
 289                                     comp_ev_file->uobj.ufile, filp,
 290                                     buf, count, pos,
 291                                     sizeof(struct ib_uverbs_comp_event_desc));
 292 }
 293
 294 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
 295                                          struct file *filp,
 296                                          struct poll_table_struct *wait)
 297 {
 298         __poll_t pollflags = 0;
 299
 300         poll_wait(filp, &ev_queue->poll_wait, wait);
 301
 302         spin_lock_irq(&ev_queue->lock);
 303         if (!list_empty(&ev_queue->event_list))
 304                 pollflags = EPOLLIN | EPOLLRDNORM;
 305         spin_unlock_irq(&ev_queue->lock);
 306
 307         return pollflags;
 308 }
 309
 310 static __poll_t ib_uverbs_async_event_poll(struct file *filp,
 311                                                struct poll_table_struct *wait)
 312 {
 313         return ib_uverbs_event_poll(filp->private_data, filp, wait);
 314 }
 315
 316 static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
 317                                               struct poll_table_struct *wait)
 318 {
 319         struct ib_uverbs_completion_event_file *comp_ev_file =
 320                 filp->private_data;
 321
 322         return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
 323 }
 324
 325 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
 326 {
 327         struct ib_uverbs_event_queue *ev_queue = filp->private_data;
 328
 329         return fasync_helper(fd, filp, on, &ev_queue->async_queue);
 330 }
 331
 332 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
 333 {
 334         struct ib_uverbs_completion_event_file *comp_ev_file =
 335                 filp->private_data;
 336
 337         return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
 338 }
 339
 340 static int ib_uverbs_async_event_close(struct inode *inode, struct file *filp)
 341 {
 342         struct ib_uverbs_async_event_file *file = filp->private_data;
 343         struct ib_uverbs_file *uverbs_file = file->uverbs_file;
 344         struct ib_uverbs_event *entry, *tmp;
 345         int closed_already = 0;
 346
 347         mutex_lock(&uverbs_file->device->lists_mutex);
 348         spin_lock_irq(&file->ev_queue.lock);
 349         closed_already = file->ev_queue.is_closed;
 350         file->ev_queue.is_closed = 1;
 351         list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
 352                 if (entry->counter)
 353                         list_del(&entry->obj_list);
 354                 kfree(entry);
 355         }
 356         spin_unlock_irq(&file->ev_queue.lock);
 357         if (!closed_already) {
 358                 list_del(&file->list);
 359                 ib_unregister_event_handler(&uverbs_file->event_handler);
 360         }
 361         mutex_unlock(&uverbs_file->device->lists_mutex);
 362
 363         kref_put(&uverbs_file->ref, ib_uverbs_release_file);
 364         kref_put(&file->ref, ib_uverbs_release_async_event_file);
 365
 366         return 0;
 367 }
 368
 369 static int ib_uverbs_comp_event_close(struct inode *inode, struct file *filp)
 370 {
 371         struct ib_uobject *uobj = filp->private_data;
 372         struct ib_uverbs_completion_event_file *file = container_of(
 373                 uobj, struct ib_uverbs_completion_event_file, uobj);
 374         struct ib_uverbs_event *entry, *tmp;
 375
 376         spin_lock_irq(&file->ev_queue.lock);
 377         list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
 378                 if (entry->counter)
 379                         list_del(&entry->obj_list);
 380                 kfree(entry);
 381         }
 382         file->ev_queue.is_closed = 1;
 383         spin_unlock_irq(&file->ev_queue.lock);
 384
 385         uverbs_close_fd(filp);
 386
 387         return 0;
 388 }
 389
 390 const struct file_operations uverbs_event_fops = {
 391         .owner   = THIS_MODULE,
 392         .read    = ib_uverbs_comp_event_read,
 393         .poll    = ib_uverbs_comp_event_poll,
 394         .release = ib_uverbs_comp_event_close,
 395         .fasync  = ib_uverbs_comp_event_fasync,
 396         .llseek  = no_llseek,
 397 };
 398
 399 static const struct file_operations uverbs_async_event_fops = {
 400         .owner   = THIS_MODULE,
 401         .read    = ib_uverbs_async_event_read,
 402         .poll    = ib_uverbs_async_event_poll,
 403         .release = ib_uverbs_async_event_close,
 404         .fasync  = ib_uverbs_async_event_fasync,
 405         .llseek  = no_llseek,
 406 };
 407
 408 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 409 {
 410         struct ib_uverbs_event_queue   *ev_queue = cq_context;
 411         struct ib_ucq_object           *uobj;
 412         struct ib_uverbs_event         *entry;
 413         unsigned long                   flags;
 414
 415         if (!ev_queue)
 416                 return;
 417
 418         spin_lock_irqsave(&ev_queue->lock, flags);
 419         if (ev_queue->is_closed) {
 420                 spin_unlock_irqrestore(&ev_queue->lock, flags);
 421                 return;
 422         }
 423
 424         entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 425         if (!entry) {
 426                 spin_unlock_irqrestore(&ev_queue->lock, flags);
 427                 return;
 428         }
 429
 430         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 431
 432         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 433         entry->counter             = &uobj->comp_events_reported;
 434
 435         list_add_tail(&entry->list, &ev_queue->event_list);
 436         list_add_tail(&entry->obj_list, &uobj->comp_list);
 437         spin_unlock_irqrestore(&ev_queue->lock, flags);
 438
 439         wake_up_interruptible(&ev_queue->poll_wait);
 440         kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
 441 }
 442
 443 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 444                                     __u64 element, __u64 event,
 445                                     struct list_head *obj_list,
 446                                     u32 *counter)
 447 {
 448         struct ib_uverbs_event *entry;
 449         unsigned long flags;
 450
 451         spin_lock_irqsave(&file->async_file->ev_queue.lock, flags);
 452         if (file->async_file->ev_queue.is_closed) {
 453                 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 454                 return;
 455         }
 456
 457         entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 458         if (!entry) {
 459                 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 460                 return;
 461         }
 462
 463         entry->desc.async.element    = element;
 464         entry->desc.async.event_type = event;
 465         entry->desc.async.reserved   = 0;
 466         entry->counter               = counter;
 467
 468         list_add_tail(&entry->list, &file->async_file->ev_queue.event_list);
 469         if (obj_list)
 470                 list_add_tail(&entry->obj_list, obj_list);
 471         spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 472
 473         wake_up_interruptible(&file->async_file->ev_queue.poll_wait);
 474         kill_fasync(&file->async_file->ev_queue.async_queue, SIGIO, POLL_IN);
 475 }
 476
 477 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 478 {
 479         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 480                                                   struct ib_ucq_object, uobject);
 481
 482         ib_uverbs_async_handler(uobj->uobject.ufile, uobj->uobject.user_handle,
 483                                 event->event, &uobj->async_list,
 484                                 &uobj->async_events_reported);
 485 }
 486
 487 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 488 {
 489         struct ib_uevent_object *uobj;
 490
 491         /* for XRC target qp's, check that qp is live */
 492         if (!event->element.qp->uobject)
 493                 return;
 494
 495         uobj = container_of(event->element.qp->uobject,
 496                             struct ib_uevent_object, uobject);
 497
 498         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 499                                 event->event, &uobj->event_list,
 500                                 &uobj->events_reported);
 501 }
 502
 503 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
 504 {
 505         struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
 506                                                   struct ib_uevent_object, uobject);
 507
 508         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 509                                 event->event, &uobj->event_list,
 510                                 &uobj->events_reported);
 511 }
 512
 513 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 514 {
 515         struct ib_uevent_object *uobj;
 516
 517         uobj = container_of(event->element.srq->uobject,
 518                             struct ib_uevent_object, uobject);
 519
 520         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 521                                 event->event, &uobj->event_list,
 522                                 &uobj->events_reported);
 523 }
 524
 525 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 526                              struct ib_event *event)
 527 {
 528         struct ib_uverbs_file *file =
 529                 container_of(handler, struct ib_uverbs_file, event_handler);
 530
 531         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 532                                 NULL, NULL);
 533 }
 534
 535 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
 536 {
 537         kref_put(&file->async_file->ref, ib_uverbs_release_async_event_file);
 538         file->async_file = NULL;
 539 }
 540
 541 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
 542 {
 543         spin_lock_init(&ev_queue->lock);
 544         INIT_LIST_HEAD(&ev_queue->event_list);
 545         init_waitqueue_head(&ev_queue->poll_wait);
 546         ev_queue->is_closed   = 0;
 547         ev_queue->async_queue = NULL;
 548 }
 549
 550 struct file *ib_uverbs_alloc_async_event_file(struct ib_uverbs_file *uverbs_file,
 551                                               struct ib_device  *ib_dev)
 552 {
 553         struct ib_uverbs_async_event_file *ev_file;
 554         struct file *filp;
 555
 556         ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
 557         if (!ev_file)
 558                 return ERR_PTR(-ENOMEM);
 559
 560         ib_uverbs_init_event_queue(&ev_file->ev_queue);
 561         ev_file->uverbs_file = uverbs_file;
 562         kref_get(&ev_file->uverbs_file->ref);
 563         kref_init(&ev_file->ref);
 564         filp = anon_inode_getfile("[infinibandevent]", &uverbs_async_event_fops,
 565                                   ev_file, O_RDONLY);
 566         if (IS_ERR(filp))
 567                 goto err_put_refs;
 568
 569         mutex_lock(&uverbs_file->device->lists_mutex);
 570         list_add_tail(&ev_file->list,
 571                       &uverbs_file->device->uverbs_events_file_list);
 572         mutex_unlock(&uverbs_file->device->lists_mutex);
 573
 574         WARN_ON(uverbs_file->async_file);
 575         uverbs_file->async_file = ev_file;
 576         kref_get(&uverbs_file->async_file->ref);
 577         INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
 578                               ib_dev,
 579                               ib_uverbs_event_handler);
 580         ib_register_event_handler(&uverbs_file->event_handler);
 581         /* At that point async file stuff was fully set */
 582
 583         return filp;
 584
 585 err_put_refs:
 586         kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
 587         kref_put(&ev_file->ref, ib_uverbs_release_async_event_file);
 588         return filp;
 589 }
 590
 591 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
 592                           struct ib_uverbs_ex_cmd_hdr *ex_hdr,
 593                           size_t count, bool extended)
 594 {
 595         if (extended) {
 596                 count -= sizeof(*hdr) + sizeof(*ex_hdr);
 597
 598                 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
 599                         return -EINVAL;
 600
 601                 if (ex_hdr->cmd_hdr_reserved)
 602                         return -EINVAL;
 603
 604                 if (ex_hdr->response) {
 605                         if (!hdr->out_words && !ex_hdr->provider_out_words)
 606                                 return -EINVAL;
 607
 608                         if (!access_ok(VERIFY_WRITE,
 609                                        u64_to_user_ptr(ex_hdr->response),
 610                                        (hdr->out_words + ex_hdr->provider_out_words) * 8))
 611                                 return -EFAULT;
 612                 } else {
 613                         if (hdr->out_words || ex_hdr->provider_out_words)
 614                                 return -EINVAL;
 615                 }
 616
 617                 return 0;
 618         }
 619
 620         /* not extended command */
 621         if (hdr->in_words * 4 != count)
 622                 return -EINVAL;
 623
 624         return 0;
 625 }
 626
 627 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 628                              size_t count, loff_t *pos)
 629 {
 630         struct ib_uverbs_file *file = filp->private_data;
 631         const struct uverbs_api_write_method *method_elm;
 632         struct uverbs_api *uapi = file->device->uapi;
 633         struct ib_uverbs_ex_cmd_hdr ex_hdr;
 634         struct ib_uverbs_cmd_hdr hdr;
 635         int srcu_key;
 636         ssize_t ret;
 637
 638         if (!ib_safe_file_access(filp)) {
 639                 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
 640                             task_tgid_vnr(current), current->comm);
 641                 return -EACCES;
 642         }
 643
 644         if (count < sizeof(hdr))
 645                 return -EINVAL;
 646
 647         if (copy_from_user(&hdr, buf, sizeof(hdr)))
 648                 return -EFAULT;
 649
 650         method_elm = uapi_get_method(uapi, hdr.command);
 651         if (IS_ERR(method_elm))
 652                 return PTR_ERR(method_elm);
 653
 654         if (method_elm->is_ex) {
 655                 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 656                         return -EINVAL;
 657                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
 658                         return -EFAULT;
 659         }
 660
 661         ret = verify_hdr(&hdr, &ex_hdr, count, method_elm->is_ex);
 662         if (ret)
 663                 return ret;
 664
 665         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 666
 667         buf += sizeof(hdr);
 668
 669         if (!method_elm->is_ex) {
 670                 ret = method_elm->handler(file, buf, hdr.in_words * 4,
 671                                           hdr.out_words * 4);
 672         } else {
 673                 struct ib_udata ucore;
 674                 struct ib_udata uhw;
 675
 676                 buf += sizeof(ex_hdr);
 677
 678                 ib_uverbs_init_udata_buf_or_null(&ucore, buf,
 679                                         u64_to_user_ptr(ex_hdr.response),
 680                                         hdr.in_words * 8, hdr.out_words * 8);
 681
 682                 ib_uverbs_init_udata_buf_or_null(&uhw,
 683                                         buf + ucore.inlen,
 684                                         u64_to_user_ptr(ex_hdr.response) + ucore.outlen,
 685                                         ex_hdr.provider_in_words * 8,
 686                                         ex_hdr.provider_out_words * 8);
 687
 688                 ret = method_elm->handler_ex(file, &ucore, &uhw);
 689                 ret = (ret) ? : count;
 690         }
 691
 692         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 693         return ret;
 694 }
 695
 696 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 697 {
 698         struct ib_uverbs_file *file = filp->private_data;
 699         struct ib_ucontext *ucontext;
 700         int ret = 0;
 701         int srcu_key;
 702
 703         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 704         ucontext = ib_uverbs_get_ucontext(file);
 705         if (IS_ERR(ucontext)) {
 706                 ret = PTR_ERR(ucontext);
 707                 goto out;
 708         }
 709
 710         ret = ucontext->device->mmap(ucontext, vma);
 711 out:
 712         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 713         return ret;
 714 }
 715
 716 /*
 717  * Each time we map IO memory into user space this keeps track of the mapping.
 718  * When the device is hot-unplugged we 'zap' the mmaps in user space to point
 719  * to the zero page and allow the hot unplug to proceed.
 720  *
 721  * This is necessary for cases like PCI physical hot unplug as the actual BAR
 722  * memory may vanish after this and access to it from userspace could MCE.
 723  *
 724  * RDMA drivers supporting disassociation must have their user space designed
 725  * to cope in some way with their IO pages going to the zero page.
 726  */
 727 struct rdma_umap_priv {
 728         struct vm_area_struct *vma;
 729         struct list_head list;
 730 };
 731
 732 static const struct vm_operations_struct rdma_umap_ops;
 733
 734 static void rdma_umap_priv_init(struct rdma_umap_priv *priv,
 735                                 struct vm_area_struct *vma)
 736 {
 737         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 738
 739         priv->vma = vma;
 740         vma->vm_private_data = priv;
 741         vma->vm_ops = &rdma_umap_ops;
 742
 743         mutex_lock(&ufile->umap_lock);
 744         list_add(&priv->list, &ufile->umaps);
 745         mutex_unlock(&ufile->umap_lock);
 746 }
 747
 748 /*
 749  * The VMA has been dup'd, initialize the vm_private_data with a new tracking
 750  * struct
 751  */
 752 static void rdma_umap_open(struct vm_area_struct *vma)
 753 {
 754         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 755         struct rdma_umap_priv *opriv = vma->vm_private_data;
 756         struct rdma_umap_priv *priv;
 757
 758         if (!opriv)
 759                 return;
 760
 761         /* We are racing with disassociation */
 762         if (!down_read_trylock(&ufile->hw_destroy_rwsem))
 763                 goto out_zap;
 764         /*
 765          * Disassociation already completed, the VMA should already be zapped.
 766          */
 767         if (!ufile->ucontext)
 768                 goto out_unlock;
 769
 770         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 771         if (!priv)
 772                 goto out_unlock;
 773         rdma_umap_priv_init(priv, vma);
 774
 775         up_read(&ufile->hw_destroy_rwsem);
 776         return;
 777
 778 out_unlock:
 779         up_read(&ufile->hw_destroy_rwsem);
 780 out_zap:
 781         /*
 782          * We can't allow the VMA to be created with the actual IO pages, that
 783          * would break our API contract, and it can't be stopped at this
 784          * point, so zap it.
 785          */
 786         vma->vm_private_data = NULL;
 787         zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
 788 }
 789
 790 static void rdma_umap_close(struct vm_area_struct *vma)
 791 {
 792         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 793         struct rdma_umap_priv *priv = vma->vm_private_data;
 794
 795         if (!priv)
 796                 return;
 797
 798         /*
 799          * The vma holds a reference on the struct file that created it, which
 800          * in turn means that the ib_uverbs_file is guaranteed to exist at
 801          * this point.
 802          */
 803         mutex_lock(&ufile->umap_lock);
 804         list_del(&priv->list);
 805         mutex_unlock(&ufile->umap_lock);
 806         kfree(priv);
 807 }
 808
 809 static const struct vm_operations_struct rdma_umap_ops = {
 810         .open = rdma_umap_open,
 811         .close = rdma_umap_close,
 812 };
 813
 814 static struct rdma_umap_priv *rdma_user_mmap_pre(struct ib_ucontext *ucontext,
 815                                                  struct vm_area_struct *vma,
 816                                                  unsigned long size)
 817 {
 818         struct ib_uverbs_file *ufile = ucontext->ufile;
 819         struct rdma_umap_priv *priv;
 820
 821         if (vma->vm_end - vma->vm_start != size)
 822                 return ERR_PTR(-EINVAL);
 823
 824         /* Driver is using this wrong, must be called by ib_uverbs_mmap */
 825         if (WARN_ON(!vma->vm_file ||
 826                     vma->vm_file->private_data != ufile))
 827                 return ERR_PTR(-EINVAL);
 828         lockdep_assert_held(&ufile->device->disassociate_srcu);
 829
 830         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 831         if (!priv)
 832                 return ERR_PTR(-ENOMEM);
 833         return priv;
 834 }
 835
 836 /*
 837  * Map IO memory into a process. This is to be called by drivers as part of
 838  * their mmap() functions if they wish to send something like PCI-E BAR memory
 839  * to userspace.
 840  */
 841 int rdma_user_mmap_io(struct ib_ucontext *ucontext, struct vm_area_struct *vma,
 842                       unsigned long pfn, unsigned long size, pgprot_t prot)
 843 {
 844         struct rdma_umap_priv *priv = rdma_user_mmap_pre(ucontext, vma, size);
 845
 846         if (IS_ERR(priv))
 847                 return PTR_ERR(priv);
 848
 849         vma->vm_page_prot = prot;
 850         if (io_remap_pfn_range(vma, vma->vm_start, pfn, size, prot)) {
 851                 kfree(priv);
 852                 return -EAGAIN;
 853         }
 854
 855         rdma_umap_priv_init(priv, vma);
 856         return 0;
 857 }
 858 EXPORT_SYMBOL(rdma_user_mmap_io);
 859
 860 /*
 861  * The page case is here for a slightly different reason, the driver expects
 862  * to be able to free the page it is sharing to user space when it destroys
 863  * its ucontext, which means we need to zap the user space references.
 864  *
 865  * We could handle this differently by providing an API to allocate a shared
 866  * page and then only freeing the shared page when the last ufile is
 867  * destroyed.
 868  */
 869 int rdma_user_mmap_page(struct ib_ucontext *ucontext,
 870                         struct vm_area_struct *vma, struct page *page,
 871                         unsigned long size)
 872 {
 873         struct rdma_umap_priv *priv = rdma_user_mmap_pre(ucontext, vma, size);
 874
 875         if (IS_ERR(priv))
 876                 return PTR_ERR(priv);
 877
 878         if (remap_pfn_range(vma, vma->vm_start, page_to_pfn(page), size,
 879                             vma->vm_page_prot)) {
 880                 kfree(priv);
 881                 return -EAGAIN;
 882         }
 883
 884         rdma_umap_priv_init(priv, vma);
 885         return 0;
 886 }
 887 EXPORT_SYMBOL(rdma_user_mmap_page);
 888
 889 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
 890 {
 891         struct rdma_umap_priv *priv, *next_priv;
 892
 893         lockdep_assert_held(&ufile->hw_destroy_rwsem);
 894
 895         while (1) {
 896                 struct mm_struct *mm = NULL;
 897
 898                 /* Get an arbitrary mm pointer that hasn't been cleaned yet */
 899                 mutex_lock(&ufile->umap_lock);
 900                 if (!list_empty(&ufile->umaps)) {
 901                         mm = list_first_entry(&ufile->umaps,
 902                                               struct rdma_umap_priv, list)
 903                                      ->vma->vm_mm;
 904                         mmget(mm);
 905                 }
 906                 mutex_unlock(&ufile->umap_lock);
 907                 if (!mm)
 908                         return;
 909
 910                 /*
 911                  * The umap_lock is nested under mmap_sem since it used within
 912                  * the vma_ops callbacks, so we have to clean the list one mm
 913                  * at a time to get the lock ordering right. Typically there
 914                  * will only be one mm, so no big deal.
 915                  */
 916                 down_write(&mm->mmap_sem);
 917                 mutex_lock(&ufile->umap_lock);
 918                 list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
 919                                           list) {
 920                         struct vm_area_struct *vma = priv->vma;
 921
 922                         if (vma->vm_mm != mm)
 923                                 continue;
 924                         list_del_init(&priv->list);
 925
 926                         zap_vma_ptes(vma, vma->vm_start,
 927                                      vma->vm_end - vma->vm_start);
 928                         vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
 929                 }
 930                 mutex_unlock(&ufile->umap_lock);
 931                 up_write(&mm->mmap_sem);
 932                 mmput(mm);
 933         }
 934 }
 935
 936 /*
 937  * ib_uverbs_open() does not need the BKL:
 938  *
 939  *  - the ib_uverbs_device structures are properly reference counted and
 940  *    everything else is purely local to the file being created, so
 941  *    races against other open calls are not a problem;
 942  *  - there is no ioctl method to race against;
 943  *  - the open method will either immediately run -ENXIO, or all
 944  *    required initialization will be done.
 945  */
 946 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 947 {
 948         struct ib_uverbs_device *dev;
 949         struct ib_uverbs_file *file;
 950         struct ib_device *ib_dev;
 951         int ret;
 952         int module_dependent;
 953         int srcu_key;
 954
 955         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 956         if (!atomic_inc_not_zero(&dev->refcount))
 957                 return -ENXIO;
 958
 959         get_device(&dev->dev);
 960         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
 961         mutex_lock(&dev->lists_mutex);
 962         ib_dev = srcu_dereference(dev->ib_dev,
 963                                   &dev->disassociate_srcu);
 964         if (!ib_dev) {
 965                 ret = -EIO;
 966                 goto err;
 967         }
 968
 969         /* In case IB device supports disassociate ucontext, there is no hard
 970          * dependency between uverbs device and its low level device.
 971          */
 972         module_dependent = !(ib_dev->disassociate_ucontext);
 973
 974         if (module_dependent) {
 975                 if (!try_module_get(ib_dev->owner)) {
 976                         ret = -ENODEV;
 977                         goto err;
 978                 }
 979         }
 980
 981         file = kzalloc(sizeof(*file), GFP_KERNEL);
 982         if (!file) {
 983                 ret = -ENOMEM;
 984                 if (module_dependent)
 985                         goto err_module;
 986
 987                 goto err;
 988         }
 989
 990         file->device     = dev;
 991         kref_init(&file->ref);
 992         mutex_init(&file->ucontext_lock);
 993
 994         spin_lock_init(&file->uobjects_lock);
 995         INIT_LIST_HEAD(&file->uobjects);
 996         init_rwsem(&file->hw_destroy_rwsem);
 997         mutex_init(&file->umap_lock);
 998         INIT_LIST_HEAD(&file->umaps);
 999
1000         filp->private_data = file;
1001         list_add_tail(&file->list, &dev->uverbs_file_list);
1002         mutex_unlock(&dev->lists_mutex);
1003         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1004
1005         setup_ufile_idr_uobject(file);
1006
1007         return nonseekable_open(inode, filp);
1008
1009 err_module:
1010         module_put(ib_dev->owner);
1011
1012 err:
1013         mutex_unlock(&dev->lists_mutex);
1014         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1015         if (atomic_dec_and_test(&dev->refcount))
1016                 ib_uverbs_comp_dev(dev);
1017
1018         put_device(&dev->dev);
1019         return ret;
1020 }
1021
1022 static int ib_uverbs_close(struct inode *inode, struct file *filp)
1023 {
1024         struct ib_uverbs_file *file = filp->private_data;
1025
1026         uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
1027
1028         mutex_lock(&file->device->lists_mutex);
1029         list_del_init(&file->list);
1030         mutex_unlock(&file->device->lists_mutex);
1031
1032         if (file->async_file)
1033                 kref_put(&file->async_file->ref,
1034                          ib_uverbs_release_async_event_file);
1035
1036         kref_put(&file->ref, ib_uverbs_release_file);
1037
1038         return 0;
1039 }
1040
1041 static const struct file_operations uverbs_fops = {
1042         .owner   = THIS_MODULE,
1043         .write   = ib_uverbs_write,
1044         .open    = ib_uverbs_open,
1045         .release = ib_uverbs_close,
1046         .llseek  = no_llseek,
1047         .unlocked_ioctl = ib_uverbs_ioctl,
1048         .compat_ioctl = ib_uverbs_ioctl,
1049 };
1050
1051 static const struct file_operations uverbs_mmap_fops = {
1052         .owner   = THIS_MODULE,
1053         .write   = ib_uverbs_write,
1054         .mmap    = ib_uverbs_mmap,
1055         .open    = ib_uverbs_open,
1056         .release = ib_uverbs_close,
1057         .llseek  = no_llseek,
1058         .unlocked_ioctl = ib_uverbs_ioctl,
1059         .compat_ioctl = ib_uverbs_ioctl,
1060 };
1061
1062 static struct ib_client uverbs_client = {
1063         .name   = "uverbs",
1064         .add    = ib_uverbs_add_one,
1065         .remove = ib_uverbs_remove_one
1066 };
1067
1068 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1069                           char *buf)
1070 {
1071         struct ib_uverbs_device *dev =
1072                         container_of(device, struct ib_uverbs_device, dev);
1073         int ret = -ENODEV;
1074         int srcu_key;
1075         struct ib_device *ib_dev;
1076
1077         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1078         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1079         if (ib_dev)
1080                 ret = sprintf(buf, "%s\n", dev_name(&ib_dev->dev));
1081         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1082
1083         return ret;
1084 }
1085 static DEVICE_ATTR_RO(ibdev);
1086
1087 static ssize_t abi_version_show(struct device *device,
1088                                 struct device_attribute *attr, char *buf)
1089 {
1090         struct ib_uverbs_device *dev =
1091                         container_of(device, struct ib_uverbs_device, dev);
1092         int ret = -ENODEV;
1093         int srcu_key;
1094         struct ib_device *ib_dev;
1095
1096         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1097         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1098         if (ib_dev)
1099                 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1100         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1101
1102         return ret;
1103 }
1104 static DEVICE_ATTR_RO(abi_version);
1105
1106 static struct attribute *ib_dev_attrs[] = {
1107         &dev_attr_abi_version.attr,
1108         &dev_attr_ibdev.attr,
1109         NULL,
1110 };
1111
1112 static const struct attribute_group dev_attr_group = {
1113         .attrs = ib_dev_attrs,
1114 };
1115
1116 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1117                          __stringify(IB_USER_VERBS_ABI_VERSION));
1118
1119 static int ib_uverbs_create_uapi(struct ib_device *device,
1120                                  struct ib_uverbs_device *uverbs_dev)
1121 {
1122         struct uverbs_api *uapi;
1123
1124         uapi = uverbs_alloc_api(device);
1125         if (IS_ERR(uapi))
1126                 return PTR_ERR(uapi);
1127
1128         uverbs_dev->uapi = uapi;
1129         return 0;
1130 }
1131
1132 static void ib_uverbs_add_one(struct ib_device *device)
1133 {
1134         int devnum;
1135         dev_t base;
1136         struct ib_uverbs_device *uverbs_dev;
1137         int ret;
1138
1139         if (!device->alloc_ucontext)
1140                 return;
1141
1142         uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1143         if (!uverbs_dev)
1144                 return;
1145
1146         ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1147         if (ret) {
1148                 kfree(uverbs_dev);
1149                 return;
1150         }
1151
1152         device_initialize(&uverbs_dev->dev);
1153         uverbs_dev->dev.class = uverbs_class;
1154         uverbs_dev->dev.parent = device->dev.parent;
1155         uverbs_dev->dev.release = ib_uverbs_release_dev;
1156         uverbs_dev->groups[0] = &dev_attr_group;
1157         uverbs_dev->dev.groups = uverbs_dev->groups;
1158         atomic_set(&uverbs_dev->refcount, 1);
1159         init_completion(&uverbs_dev->comp);
1160         uverbs_dev->xrcd_tree = RB_ROOT;
1161         mutex_init(&uverbs_dev->xrcd_tree_mutex);
1162         mutex_init(&uverbs_dev->lists_mutex);
1163         INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1164         INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1165         rcu_assign_pointer(uverbs_dev->ib_dev, device);
1166         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1167
1168         devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1169                                GFP_KERNEL);
1170         if (devnum < 0)
1171                 goto err;
1172         uverbs_dev->devnum = devnum;
1173         if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1174                 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1175         else
1176                 base = IB_UVERBS_BASE_DEV + devnum;
1177
1178         if (ib_uverbs_create_uapi(device, uverbs_dev))
1179                 goto err_uapi;
1180
1181         uverbs_dev->dev.devt = base;
1182         dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1183
1184         cdev_init(&uverbs_dev->cdev,
1185                   device->mmap ? &uverbs_mmap_fops : &uverbs_fops);
1186         uverbs_dev->cdev.owner = THIS_MODULE;
1187
1188         ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1189         if (ret)
1190                 goto err_uapi;
1191
1192         ib_set_client_data(device, &uverbs_client, uverbs_dev);
1193         return;
1194
1195 err_uapi:
1196         ida_free(&uverbs_ida, devnum);
1197 err:
1198         if (atomic_dec_and_test(&uverbs_dev->refcount))
1199                 ib_uverbs_comp_dev(uverbs_dev);
1200         wait_for_completion(&uverbs_dev->comp);
1201         put_device(&uverbs_dev->dev);
1202         return;
1203 }
1204
1205 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1206                                         struct ib_device *ib_dev)
1207 {
1208         struct ib_uverbs_file *file;
1209         struct ib_uverbs_async_event_file *event_file;
1210         struct ib_event event;
1211
1212         /* Pending running commands to terminate */
1213         uverbs_disassociate_api_pre(uverbs_dev);
1214         event.event = IB_EVENT_DEVICE_FATAL;
1215         event.element.port_num = 0;
1216         event.device = ib_dev;
1217
1218         mutex_lock(&uverbs_dev->lists_mutex);
1219         while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1220                 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1221                                         struct ib_uverbs_file, list);
1222                 list_del_init(&file->list);
1223                 kref_get(&file->ref);
1224
1225                 /* We must release the mutex before going ahead and calling
1226                  * uverbs_cleanup_ufile, as it might end up indirectly calling
1227                  * uverbs_close, for example due to freeing the resources (e.g
1228                  * mmput).
1229                  */
1230                 mutex_unlock(&uverbs_dev->lists_mutex);
1231
1232                 ib_uverbs_event_handler(&file->event_handler, &event);
1233                 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1234                 kref_put(&file->ref, ib_uverbs_release_file);
1235
1236                 mutex_lock(&uverbs_dev->lists_mutex);
1237         }
1238
1239         while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1240                 event_file = list_first_entry(&uverbs_dev->
1241                                               uverbs_events_file_list,
1242                                               struct ib_uverbs_async_event_file,
1243                                               list);
1244                 spin_lock_irq(&event_file->ev_queue.lock);
1245                 event_file->ev_queue.is_closed = 1;
1246                 spin_unlock_irq(&event_file->ev_queue.lock);
1247
1248                 list_del(&event_file->list);
1249                 ib_unregister_event_handler(
1250                         &event_file->uverbs_file->event_handler);
1251                 event_file->uverbs_file->event_handler.device =
1252                         NULL;
1253
1254                 wake_up_interruptible(&event_file->ev_queue.poll_wait);
1255                 kill_fasync(&event_file->ev_queue.async_queue, SIGIO, POLL_IN);
1256         }
1257         mutex_unlock(&uverbs_dev->lists_mutex);
1258
1259         uverbs_disassociate_api(uverbs_dev->uapi);
1260 }
1261
1262 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1263 {
1264         struct ib_uverbs_device *uverbs_dev = client_data;
1265         int wait_clients = 1;
1266
1267         if (!uverbs_dev)
1268                 return;
1269
1270         cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1271         ida_free(&uverbs_ida, uverbs_dev->devnum);
1272
1273         if (device->disassociate_ucontext) {
1274                 /* We disassociate HW resources and immediately return.
1275                  * Userspace will see a EIO errno for all future access.
1276                  * Upon returning, ib_device may be freed internally and is not
1277                  * valid any more.
1278                  * uverbs_device is still available until all clients close
1279                  * their files, then the uverbs device ref count will be zero
1280                  * and its resources will be freed.
1281                  * Note: At this point no more files can be opened since the
1282                  * cdev was deleted, however active clients can still issue
1283                  * commands and close their open files.
1284                  */
1285                 ib_uverbs_free_hw_resources(uverbs_dev, device);
1286                 wait_clients = 0;
1287         }
1288
1289         if (atomic_dec_and_test(&uverbs_dev->refcount))
1290                 ib_uverbs_comp_dev(uverbs_dev);
1291         if (wait_clients)
1292                 wait_for_completion(&uverbs_dev->comp);
1293
1294         put_device(&uverbs_dev->dev);
1295 }
1296
1297 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1298 {
1299         if (mode)
1300                 *mode = 0666;
1301         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1302 }
1303
1304 static int __init ib_uverbs_init(void)
1305 {
1306         int ret;
1307
1308         ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1309                                      IB_UVERBS_NUM_FIXED_MINOR,
1310                                      "infiniband_verbs");
1311         if (ret) {
1312                 pr_err("user_verbs: couldn't register device number\n");
1313                 goto out;
1314         }
1315
1316         ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1317                                   IB_UVERBS_NUM_DYNAMIC_MINOR,
1318                                   "infiniband_verbs");
1319         if (ret) {
1320                 pr_err("couldn't register dynamic device number\n");
1321                 goto out_alloc;
1322         }
1323
1324         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1325         if (IS_ERR(uverbs_class)) {
1326                 ret = PTR_ERR(uverbs_class);
1327                 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1328                 goto out_chrdev;
1329         }
1330
1331         uverbs_class->devnode = uverbs_devnode;
1332
1333         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1334         if (ret) {
1335                 pr_err("user_verbs: couldn't create abi_version attribute\n");
1336                 goto out_class;
1337         }
1338
1339         ret = ib_register_client(&uverbs_client);
1340         if (ret) {
1341                 pr_err("user_verbs: couldn't register client\n");
1342                 goto out_class;
1343         }
1344
1345         return 0;
1346
1347 out_class:
1348         class_destroy(uverbs_class);
1349
1350 out_chrdev:
1351         unregister_chrdev_region(dynamic_uverbs_dev,
1352                                  IB_UVERBS_NUM_DYNAMIC_MINOR);
1353
1354 out_alloc:
1355         unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1356                                  IB_UVERBS_NUM_FIXED_MINOR);
1357
1358 out:
1359         return ret;
1360 }
1361
1362 static void __exit ib_uverbs_cleanup(void)
1363 {
1364         ib_unregister_client(&uverbs_client);
1365         class_destroy(uverbs_class);
1366         unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1367                                  IB_UVERBS_NUM_FIXED_MINOR);
1368         unregister_chrdev_region(dynamic_uverbs_dev,
1369                                  IB_UVERBS_NUM_DYNAMIC_MINOR);
1370 }
1371
1372 module_init(ib_uverbs_init);
1373 module_exit(ib_uverbs_cleanup);