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_file(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_file);
 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->ops.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->ops.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, size_t count,
 593                           const struct uverbs_api_write_method *method_elm)
 594 {
 595         if (method_elm->is_ex) {
 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 (hdr->in_words * 8 < method_elm->req_size)
 602                         return -ENOSPC;
 603
 604                 if (ex_hdr->cmd_hdr_reserved)
 605                         return -EINVAL;
 606
 607                 if (ex_hdr->response) {
 608                         if (!hdr->out_words && !ex_hdr->provider_out_words)
 609                                 return -EINVAL;
 610
 611                         if (hdr->out_words * 8 < method_elm->resp_size)
 612                                 return -ENOSPC;
 613
 614                         if (!access_ok(u64_to_user_ptr(ex_hdr->response),
 615                                        (hdr->out_words + ex_hdr->provider_out_words) * 8))
 616                                 return -EFAULT;
 617                 } else {
 618                         if (hdr->out_words || ex_hdr->provider_out_words)
 619                                 return -EINVAL;
 620                 }
 621
 622                 return 0;
 623         }
 624
 625         /* not extended command */
 626         if (hdr->in_words * 4 != count)
 627                 return -EINVAL;
 628
 629         if (count < method_elm->req_size + sizeof(hdr)) {
 630                 /*
 631                  * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
 632                  * with a 16 byte write instead of 24. Old kernels didn't
 633                  * check the size so they allowed this. Now that the size is
 634                  * checked provide a compatibility work around to not break
 635                  * those userspaces.
 636                  */
 637                 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
 638                     count == 16) {
 639                         hdr->in_words = 6;
 640                         return 0;
 641                 }
 642                 return -ENOSPC;
 643         }
 644         if (hdr->out_words * 4 < method_elm->resp_size)
 645                 return -ENOSPC;
 646
 647         return 0;
 648 }
 649
 650 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 651                              size_t count, loff_t *pos)
 652 {
 653         struct ib_uverbs_file *file = filp->private_data;
 654         const struct uverbs_api_write_method *method_elm;
 655         struct uverbs_api *uapi = file->device->uapi;
 656         struct ib_uverbs_ex_cmd_hdr ex_hdr;
 657         struct ib_uverbs_cmd_hdr hdr;
 658         struct uverbs_attr_bundle bundle;
 659         int srcu_key;
 660         ssize_t ret;
 661
 662         if (!ib_safe_file_access(filp)) {
 663                 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
 664                             task_tgid_vnr(current), current->comm);
 665                 return -EACCES;
 666         }
 667
 668         if (count < sizeof(hdr))
 669                 return -EINVAL;
 670
 671         if (copy_from_user(&hdr, buf, sizeof(hdr)))
 672                 return -EFAULT;
 673
 674         method_elm = uapi_get_method(uapi, hdr.command);
 675         if (IS_ERR(method_elm))
 676                 return PTR_ERR(method_elm);
 677
 678         if (method_elm->is_ex) {
 679                 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 680                         return -EINVAL;
 681                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
 682                         return -EFAULT;
 683         }
 684
 685         ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
 686         if (ret)
 687                 return ret;
 688
 689         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 690
 691         buf += sizeof(hdr);
 692
 693         memset(bundle.attr_present, 0, sizeof(bundle.attr_present));
 694         bundle.ufile = file;
 695         if (!method_elm->is_ex) {
 696                 size_t in_len = hdr.in_words * 4 - sizeof(hdr);
 697                 size_t out_len = hdr.out_words * 4;
 698                 u64 response = 0;
 699
 700                 if (method_elm->has_udata) {
 701                         bundle.driver_udata.inlen =
 702                                 in_len - method_elm->req_size;
 703                         in_len = method_elm->req_size;
 704                         if (bundle.driver_udata.inlen)
 705                                 bundle.driver_udata.inbuf = buf + in_len;
 706                         else
 707                                 bundle.driver_udata.inbuf = NULL;
 708                 } else {
 709                         memset(&bundle.driver_udata, 0,
 710                                sizeof(bundle.driver_udata));
 711                 }
 712
 713                 if (method_elm->has_resp) {
 714                         /*
 715                          * The macros check that if has_resp is set
 716                          * then the command request structure starts
 717                          * with a '__aligned u64 response' member.
 718                          */
 719                         ret = get_user(response, (const u64 *)buf);
 720                         if (ret)
 721                                 goto out_unlock;
 722
 723                         if (method_elm->has_udata) {
 724                                 bundle.driver_udata.outlen =
 725                                         out_len - method_elm->resp_size;
 726                                 out_len = method_elm->resp_size;
 727                                 if (bundle.driver_udata.outlen)
 728                                         bundle.driver_udata.outbuf =
 729                                                 u64_to_user_ptr(response +
 730                                                                 out_len);
 731                                 else
 732                                         bundle.driver_udata.outbuf = NULL;
 733                         }
 734                 } else {
 735                         bundle.driver_udata.outlen = 0;
 736                         bundle.driver_udata.outbuf = NULL;
 737                 }
 738
 739                 ib_uverbs_init_udata_buf_or_null(
 740                         &bundle.ucore, buf, u64_to_user_ptr(response),
 741                         in_len, out_len);
 742         } else {
 743                 buf += sizeof(ex_hdr);
 744
 745                 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
 746                                         u64_to_user_ptr(ex_hdr.response),
 747                                         hdr.in_words * 8, hdr.out_words * 8);
 748
 749                 ib_uverbs_init_udata_buf_or_null(
 750                         &bundle.driver_udata, buf + bundle.ucore.inlen,
 751                         u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
 752                         ex_hdr.provider_in_words * 8,
 753                         ex_hdr.provider_out_words * 8);
 754
 755         }
 756
 757         ret = method_elm->handler(&bundle);
 758 out_unlock:
 759         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 760         return (ret) ? : count;
 761 }
 762
 763 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 764 {
 765         struct ib_uverbs_file *file = filp->private_data;
 766         struct ib_ucontext *ucontext;
 767         int ret = 0;
 768         int srcu_key;
 769
 770         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 771         ucontext = ib_uverbs_get_ucontext_file(file);
 772         if (IS_ERR(ucontext)) {
 773                 ret = PTR_ERR(ucontext);
 774                 goto out;
 775         }
 776
 777         ret = ucontext->device->ops.mmap(ucontext, vma);
 778 out:
 779         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 780         return ret;
 781 }
 782
 783 /*
 784  * Each time we map IO memory into user space this keeps track of the mapping.
 785  * When the device is hot-unplugged we 'zap' the mmaps in user space to point
 786  * to the zero page and allow the hot unplug to proceed.
 787  *
 788  * This is necessary for cases like PCI physical hot unplug as the actual BAR
 789  * memory may vanish after this and access to it from userspace could MCE.
 790  *
 791  * RDMA drivers supporting disassociation must have their user space designed
 792  * to cope in some way with their IO pages going to the zero page.
 793  */
 794 struct rdma_umap_priv {
 795         struct vm_area_struct *vma;
 796         struct list_head list;
 797 };
 798
 799 static const struct vm_operations_struct rdma_umap_ops;
 800
 801 static void rdma_umap_priv_init(struct rdma_umap_priv *priv,
 802                                 struct vm_area_struct *vma)
 803 {
 804         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 805
 806         priv->vma = vma;
 807         vma->vm_private_data = priv;
 808         vma->vm_ops = &rdma_umap_ops;
 809
 810         mutex_lock(&ufile->umap_lock);
 811         list_add(&priv->list, &ufile->umaps);
 812         mutex_unlock(&ufile->umap_lock);
 813 }
 814
 815 /*
 816  * The VMA has been dup'd, initialize the vm_private_data with a new tracking
 817  * struct
 818  */
 819 static void rdma_umap_open(struct vm_area_struct *vma)
 820 {
 821         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 822         struct rdma_umap_priv *opriv = vma->vm_private_data;
 823         struct rdma_umap_priv *priv;
 824
 825         if (!opriv)
 826                 return;
 827
 828         /* We are racing with disassociation */
 829         if (!down_read_trylock(&ufile->hw_destroy_rwsem))
 830                 goto out_zap;
 831         /*
 832          * Disassociation already completed, the VMA should already be zapped.
 833          */
 834         if (!ufile->ucontext)
 835                 goto out_unlock;
 836
 837         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 838         if (!priv)
 839                 goto out_unlock;
 840         rdma_umap_priv_init(priv, vma);
 841
 842         up_read(&ufile->hw_destroy_rwsem);
 843         return;
 844
 845 out_unlock:
 846         up_read(&ufile->hw_destroy_rwsem);
 847 out_zap:
 848         /*
 849          * We can't allow the VMA to be created with the actual IO pages, that
 850          * would break our API contract, and it can't be stopped at this
 851          * point, so zap it.
 852          */
 853         vma->vm_private_data = NULL;
 854         zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
 855 }
 856
 857 static void rdma_umap_close(struct vm_area_struct *vma)
 858 {
 859         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 860         struct rdma_umap_priv *priv = vma->vm_private_data;
 861
 862         if (!priv)
 863                 return;
 864
 865         /*
 866          * The vma holds a reference on the struct file that created it, which
 867          * in turn means that the ib_uverbs_file is guaranteed to exist at
 868          * this point.
 869          */
 870         mutex_lock(&ufile->umap_lock);
 871         list_del(&priv->list);
 872         mutex_unlock(&ufile->umap_lock);
 873         kfree(priv);
 874 }
 875
 876 static const struct vm_operations_struct rdma_umap_ops = {
 877         .open = rdma_umap_open,
 878         .close = rdma_umap_close,
 879 };
 880
 881 static struct rdma_umap_priv *rdma_user_mmap_pre(struct ib_ucontext *ucontext,
 882                                                  struct vm_area_struct *vma,
 883                                                  unsigned long size)
 884 {
 885         struct ib_uverbs_file *ufile = ucontext->ufile;
 886         struct rdma_umap_priv *priv;
 887
 888         if (vma->vm_end - vma->vm_start != size)
 889                 return ERR_PTR(-EINVAL);
 890
 891         /* Driver is using this wrong, must be called by ib_uverbs_mmap */
 892         if (WARN_ON(!vma->vm_file ||
 893                     vma->vm_file->private_data != ufile))
 894                 return ERR_PTR(-EINVAL);
 895         lockdep_assert_held(&ufile->device->disassociate_srcu);
 896
 897         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 898         if (!priv)
 899                 return ERR_PTR(-ENOMEM);
 900         return priv;
 901 }
 902
 903 /*
 904  * Map IO memory into a process. This is to be called by drivers as part of
 905  * their mmap() functions if they wish to send something like PCI-E BAR memory
 906  * to userspace.
 907  */
 908 int rdma_user_mmap_io(struct ib_ucontext *ucontext, struct vm_area_struct *vma,
 909                       unsigned long pfn, unsigned long size, pgprot_t prot)
 910 {
 911         struct rdma_umap_priv *priv = rdma_user_mmap_pre(ucontext, vma, size);
 912
 913         if (IS_ERR(priv))
 914                 return PTR_ERR(priv);
 915
 916         vma->vm_page_prot = prot;
 917         if (io_remap_pfn_range(vma, vma->vm_start, pfn, size, prot)) {
 918                 kfree(priv);
 919                 return -EAGAIN;
 920         }
 921
 922         rdma_umap_priv_init(priv, vma);
 923         return 0;
 924 }
 925 EXPORT_SYMBOL(rdma_user_mmap_io);
 926
 927 /*
 928  * The page case is here for a slightly different reason, the driver expects
 929  * to be able to free the page it is sharing to user space when it destroys
 930  * its ucontext, which means we need to zap the user space references.
 931  *
 932  * We could handle this differently by providing an API to allocate a shared
 933  * page and then only freeing the shared page when the last ufile is
 934  * destroyed.
 935  */
 936 int rdma_user_mmap_page(struct ib_ucontext *ucontext,
 937                         struct vm_area_struct *vma, struct page *page,
 938                         unsigned long size)
 939 {
 940         struct rdma_umap_priv *priv = rdma_user_mmap_pre(ucontext, vma, size);
 941
 942         if (IS_ERR(priv))
 943                 return PTR_ERR(priv);
 944
 945         if (remap_pfn_range(vma, vma->vm_start, page_to_pfn(page), size,
 946                             vma->vm_page_prot)) {
 947                 kfree(priv);
 948                 return -EAGAIN;
 949         }
 950
 951         rdma_umap_priv_init(priv, vma);
 952         return 0;
 953 }
 954 EXPORT_SYMBOL(rdma_user_mmap_page);
 955
 956 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
 957 {
 958         struct rdma_umap_priv *priv, *next_priv;
 959
 960         lockdep_assert_held(&ufile->hw_destroy_rwsem);
 961
 962         while (1) {
 963                 struct mm_struct *mm = NULL;
 964
 965                 /* Get an arbitrary mm pointer that hasn't been cleaned yet */
 966                 mutex_lock(&ufile->umap_lock);
 967                 if (!list_empty(&ufile->umaps)) {
 968                         mm = list_first_entry(&ufile->umaps,
 969                                               struct rdma_umap_priv, list)
 970                                      ->vma->vm_mm;
 971                         mmget(mm);
 972                 }
 973                 mutex_unlock(&ufile->umap_lock);
 974                 if (!mm)
 975                         return;
 976
 977                 /*
 978                  * The umap_lock is nested under mmap_sem since it used within
 979                  * the vma_ops callbacks, so we have to clean the list one mm
 980                  * at a time to get the lock ordering right. Typically there
 981                  * will only be one mm, so no big deal.
 982                  */
 983                 down_write(&mm->mmap_sem);
 984                 mutex_lock(&ufile->umap_lock);
 985                 list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
 986                                           list) {
 987                         struct vm_area_struct *vma = priv->vma;
 988
 989                         if (vma->vm_mm != mm)
 990                                 continue;
 991                         list_del_init(&priv->list);
 992
 993                         zap_vma_ptes(vma, vma->vm_start,
 994                                      vma->vm_end - vma->vm_start);
 995                         vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
 996                 }
 997                 mutex_unlock(&ufile->umap_lock);
 998                 up_write(&mm->mmap_sem);
 999                 mmput(mm);
1000         }
1001 }
1002
1003 /*
1004  * ib_uverbs_open() does not need the BKL:
1005  *
1006  *  - the ib_uverbs_device structures are properly reference counted and
1007  *    everything else is purely local to the file being created, so
1008  *    races against other open calls are not a problem;
1009  *  - there is no ioctl method to race against;
1010  *  - the open method will either immediately run -ENXIO, or all
1011  *    required initialization will be done.
1012  */
1013 static int ib_uverbs_open(struct inode *inode, struct file *filp)
1014 {
1015         struct ib_uverbs_device *dev;
1016         struct ib_uverbs_file *file;
1017         struct ib_device *ib_dev;
1018         int ret;
1019         int module_dependent;
1020         int srcu_key;
1021
1022         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
1023         if (!atomic_inc_not_zero(&dev->refcount))
1024                 return -ENXIO;
1025
1026         get_device(&dev->dev);
1027         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1028         mutex_lock(&dev->lists_mutex);
1029         ib_dev = srcu_dereference(dev->ib_dev,
1030                                   &dev->disassociate_srcu);
1031         if (!ib_dev) {
1032                 ret = -EIO;
1033                 goto err;
1034         }
1035
1036         /* In case IB device supports disassociate ucontext, there is no hard
1037          * dependency between uverbs device and its low level device.
1038          */
1039         module_dependent = !(ib_dev->ops.disassociate_ucontext);
1040
1041         if (module_dependent) {
1042                 if (!try_module_get(ib_dev->owner)) {
1043                         ret = -ENODEV;
1044                         goto err;
1045                 }
1046         }
1047
1048         file = kzalloc(sizeof(*file), GFP_KERNEL);
1049         if (!file) {
1050                 ret = -ENOMEM;
1051                 if (module_dependent)
1052                         goto err_module;
1053
1054                 goto err;
1055         }
1056
1057         file->device     = dev;
1058         kref_init(&file->ref);
1059         mutex_init(&file->ucontext_lock);
1060
1061         spin_lock_init(&file->uobjects_lock);
1062         INIT_LIST_HEAD(&file->uobjects);
1063         init_rwsem(&file->hw_destroy_rwsem);
1064         mutex_init(&file->umap_lock);
1065         INIT_LIST_HEAD(&file->umaps);
1066
1067         filp->private_data = file;
1068         list_add_tail(&file->list, &dev->uverbs_file_list);
1069         mutex_unlock(&dev->lists_mutex);
1070         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1071
1072         setup_ufile_idr_uobject(file);
1073
1074         return nonseekable_open(inode, filp);
1075
1076 err_module:
1077         module_put(ib_dev->owner);
1078
1079 err:
1080         mutex_unlock(&dev->lists_mutex);
1081         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1082         if (atomic_dec_and_test(&dev->refcount))
1083                 ib_uverbs_comp_dev(dev);
1084
1085         put_device(&dev->dev);
1086         return ret;
1087 }
1088
1089 static int ib_uverbs_close(struct inode *inode, struct file *filp)
1090 {
1091         struct ib_uverbs_file *file = filp->private_data;
1092
1093         uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
1094
1095         mutex_lock(&file->device->lists_mutex);
1096         list_del_init(&file->list);
1097         mutex_unlock(&file->device->lists_mutex);
1098
1099         if (file->async_file)
1100                 kref_put(&file->async_file->ref,
1101                          ib_uverbs_release_async_event_file);
1102
1103         kref_put(&file->ref, ib_uverbs_release_file);
1104
1105         return 0;
1106 }
1107
1108 static const struct file_operations uverbs_fops = {
1109         .owner   = THIS_MODULE,
1110         .write   = ib_uverbs_write,
1111         .open    = ib_uverbs_open,
1112         .release = ib_uverbs_close,
1113         .llseek  = no_llseek,
1114         .unlocked_ioctl = ib_uverbs_ioctl,
1115         .compat_ioctl = ib_uverbs_ioctl,
1116 };
1117
1118 static const struct file_operations uverbs_mmap_fops = {
1119         .owner   = THIS_MODULE,
1120         .write   = ib_uverbs_write,
1121         .mmap    = ib_uverbs_mmap,
1122         .open    = ib_uverbs_open,
1123         .release = ib_uverbs_close,
1124         .llseek  = no_llseek,
1125         .unlocked_ioctl = ib_uverbs_ioctl,
1126         .compat_ioctl = ib_uverbs_ioctl,
1127 };
1128
1129 static struct ib_client uverbs_client = {
1130         .name   = "uverbs",
1131         .add    = ib_uverbs_add_one,
1132         .remove = ib_uverbs_remove_one
1133 };
1134
1135 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1136                           char *buf)
1137 {
1138         struct ib_uverbs_device *dev =
1139                         container_of(device, struct ib_uverbs_device, dev);
1140         int ret = -ENODEV;
1141         int srcu_key;
1142         struct ib_device *ib_dev;
1143
1144         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1145         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1146         if (ib_dev)
1147                 ret = sprintf(buf, "%s\n", dev_name(&ib_dev->dev));
1148         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1149
1150         return ret;
1151 }
1152 static DEVICE_ATTR_RO(ibdev);
1153
1154 static ssize_t abi_version_show(struct device *device,
1155                                 struct device_attribute *attr, char *buf)
1156 {
1157         struct ib_uverbs_device *dev =
1158                         container_of(device, struct ib_uverbs_device, dev);
1159         int ret = -ENODEV;
1160         int srcu_key;
1161         struct ib_device *ib_dev;
1162
1163         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1164         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1165         if (ib_dev)
1166                 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1167         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1168
1169         return ret;
1170 }
1171 static DEVICE_ATTR_RO(abi_version);
1172
1173 static struct attribute *ib_dev_attrs[] = {
1174         &dev_attr_abi_version.attr,
1175         &dev_attr_ibdev.attr,
1176         NULL,
1177 };
1178
1179 static const struct attribute_group dev_attr_group = {
1180         .attrs = ib_dev_attrs,
1181 };
1182
1183 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1184                          __stringify(IB_USER_VERBS_ABI_VERSION));
1185
1186 static int ib_uverbs_create_uapi(struct ib_device *device,
1187                                  struct ib_uverbs_device *uverbs_dev)
1188 {
1189         struct uverbs_api *uapi;
1190
1191         uapi = uverbs_alloc_api(device);
1192         if (IS_ERR(uapi))
1193                 return PTR_ERR(uapi);
1194
1195         uverbs_dev->uapi = uapi;
1196         return 0;
1197 }
1198
1199 static void ib_uverbs_add_one(struct ib_device *device)
1200 {
1201         int devnum;
1202         dev_t base;
1203         struct ib_uverbs_device *uverbs_dev;
1204         int ret;
1205
1206         if (!device->ops.alloc_ucontext)
1207                 return;
1208
1209         uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1210         if (!uverbs_dev)
1211                 return;
1212
1213         ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1214         if (ret) {
1215                 kfree(uverbs_dev);
1216                 return;
1217         }
1218
1219         device_initialize(&uverbs_dev->dev);
1220         uverbs_dev->dev.class = uverbs_class;
1221         uverbs_dev->dev.parent = device->dev.parent;
1222         uverbs_dev->dev.release = ib_uverbs_release_dev;
1223         uverbs_dev->groups[0] = &dev_attr_group;
1224         uverbs_dev->dev.groups = uverbs_dev->groups;
1225         atomic_set(&uverbs_dev->refcount, 1);
1226         init_completion(&uverbs_dev->comp);
1227         uverbs_dev->xrcd_tree = RB_ROOT;
1228         mutex_init(&uverbs_dev->xrcd_tree_mutex);
1229         mutex_init(&uverbs_dev->lists_mutex);
1230         INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1231         INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1232         rcu_assign_pointer(uverbs_dev->ib_dev, device);
1233         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1234
1235         devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1236                                GFP_KERNEL);
1237         if (devnum < 0)
1238                 goto err;
1239         uverbs_dev->devnum = devnum;
1240         if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1241                 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1242         else
1243                 base = IB_UVERBS_BASE_DEV + devnum;
1244
1245         if (ib_uverbs_create_uapi(device, uverbs_dev))
1246                 goto err_uapi;
1247
1248         uverbs_dev->dev.devt = base;
1249         dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1250
1251         cdev_init(&uverbs_dev->cdev,
1252                   device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1253         uverbs_dev->cdev.owner = THIS_MODULE;
1254
1255         ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1256         if (ret)
1257                 goto err_uapi;
1258
1259         ib_set_client_data(device, &uverbs_client, uverbs_dev);
1260         return;
1261
1262 err_uapi:
1263         ida_free(&uverbs_ida, devnum);
1264 err:
1265         if (atomic_dec_and_test(&uverbs_dev->refcount))
1266                 ib_uverbs_comp_dev(uverbs_dev);
1267         wait_for_completion(&uverbs_dev->comp);
1268         put_device(&uverbs_dev->dev);
1269         return;
1270 }
1271
1272 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1273                                         struct ib_device *ib_dev)
1274 {
1275         struct ib_uverbs_file *file;
1276         struct ib_uverbs_async_event_file *event_file;
1277         struct ib_event event;
1278
1279         /* Pending running commands to terminate */
1280         uverbs_disassociate_api_pre(uverbs_dev);
1281         event.event = IB_EVENT_DEVICE_FATAL;
1282         event.element.port_num = 0;
1283         event.device = ib_dev;
1284
1285         mutex_lock(&uverbs_dev->lists_mutex);
1286         while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1287                 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1288                                         struct ib_uverbs_file, list);
1289                 list_del_init(&file->list);
1290                 kref_get(&file->ref);
1291
1292                 /* We must release the mutex before going ahead and calling
1293                  * uverbs_cleanup_ufile, as it might end up indirectly calling
1294                  * uverbs_close, for example due to freeing the resources (e.g
1295                  * mmput).
1296                  */
1297                 mutex_unlock(&uverbs_dev->lists_mutex);
1298
1299                 ib_uverbs_event_handler(&file->event_handler, &event);
1300                 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1301                 kref_put(&file->ref, ib_uverbs_release_file);
1302
1303                 mutex_lock(&uverbs_dev->lists_mutex);
1304         }
1305
1306         while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1307                 event_file = list_first_entry(&uverbs_dev->
1308                                               uverbs_events_file_list,
1309                                               struct ib_uverbs_async_event_file,
1310                                               list);
1311                 spin_lock_irq(&event_file->ev_queue.lock);
1312                 event_file->ev_queue.is_closed = 1;
1313                 spin_unlock_irq(&event_file->ev_queue.lock);
1314
1315                 list_del(&event_file->list);
1316                 ib_unregister_event_handler(
1317                         &event_file->uverbs_file->event_handler);
1318                 event_file->uverbs_file->event_handler.device =
1319                         NULL;
1320
1321                 wake_up_interruptible(&event_file->ev_queue.poll_wait);
1322                 kill_fasync(&event_file->ev_queue.async_queue, SIGIO, POLL_IN);
1323         }
1324         mutex_unlock(&uverbs_dev->lists_mutex);
1325
1326         uverbs_disassociate_api(uverbs_dev->uapi);
1327 }
1328
1329 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1330 {
1331         struct ib_uverbs_device *uverbs_dev = client_data;
1332         int wait_clients = 1;
1333
1334         if (!uverbs_dev)
1335                 return;
1336
1337         cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1338         ida_free(&uverbs_ida, uverbs_dev->devnum);
1339
1340         if (device->ops.disassociate_ucontext) {
1341                 /* We disassociate HW resources and immediately return.
1342                  * Userspace will see a EIO errno for all future access.
1343                  * Upon returning, ib_device may be freed internally and is not
1344                  * valid any more.
1345                  * uverbs_device is still available until all clients close
1346                  * their files, then the uverbs device ref count will be zero
1347                  * and its resources will be freed.
1348                  * Note: At this point no more files can be opened since the
1349                  * cdev was deleted, however active clients can still issue
1350                  * commands and close their open files.
1351                  */
1352                 ib_uverbs_free_hw_resources(uverbs_dev, device);
1353                 wait_clients = 0;
1354         }
1355
1356         if (atomic_dec_and_test(&uverbs_dev->refcount))
1357                 ib_uverbs_comp_dev(uverbs_dev);
1358         if (wait_clients)
1359                 wait_for_completion(&uverbs_dev->comp);
1360
1361         put_device(&uverbs_dev->dev);
1362 }
1363
1364 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1365 {
1366         if (mode)
1367                 *mode = 0666;
1368         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1369 }
1370
1371 static int __init ib_uverbs_init(void)
1372 {
1373         int ret;
1374
1375         ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1376                                      IB_UVERBS_NUM_FIXED_MINOR,
1377                                      "infiniband_verbs");
1378         if (ret) {
1379                 pr_err("user_verbs: couldn't register device number\n");
1380                 goto out;
1381         }
1382
1383         ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1384                                   IB_UVERBS_NUM_DYNAMIC_MINOR,
1385                                   "infiniband_verbs");
1386         if (ret) {
1387                 pr_err("couldn't register dynamic device number\n");
1388                 goto out_alloc;
1389         }
1390
1391         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1392         if (IS_ERR(uverbs_class)) {
1393                 ret = PTR_ERR(uverbs_class);
1394                 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1395                 goto out_chrdev;
1396         }
1397
1398         uverbs_class->devnode = uverbs_devnode;
1399
1400         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1401         if (ret) {
1402                 pr_err("user_verbs: couldn't create abi_version attribute\n");
1403                 goto out_class;
1404         }
1405
1406         ret = ib_register_client(&uverbs_client);
1407         if (ret) {
1408                 pr_err("user_verbs: couldn't register client\n");
1409                 goto out_class;
1410         }
1411
1412         return 0;
1413
1414 out_class:
1415         class_destroy(uverbs_class);
1416
1417 out_chrdev:
1418         unregister_chrdev_region(dynamic_uverbs_dev,
1419                                  IB_UVERBS_NUM_DYNAMIC_MINOR);
1420
1421 out_alloc:
1422         unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1423                                  IB_UVERBS_NUM_FIXED_MINOR);
1424
1425 out:
1426         return ret;
1427 }
1428
1429 static void __exit ib_uverbs_cleanup(void)
1430 {
1431         ib_unregister_client(&uverbs_client);
1432         class_destroy(uverbs_class);
1433         unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1434                                  IB_UVERBS_NUM_FIXED_MINOR);
1435         unregister_chrdev_region(dynamic_uverbs_dev,
1436                                  IB_UVERBS_NUM_DYNAMIC_MINOR);
1437 }
1438
1439 module_init(ib_uverbs_init);
1440 module_exit(ib_uverbs_cleanup);