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         if (file->async_file)
 208                 kref_put(&file->async_file->ref,
 209                          ib_uverbs_release_async_event_file);
 210         put_device(&file->device->dev);
 211         kfree(file);
 212 }
 213
 214 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
 215                                     struct ib_uverbs_file *uverbs_file,
 216                                     struct file *filp, char __user *buf,
 217                                     size_t count, loff_t *pos,
 218                                     size_t eventsz)
 219 {
 220         struct ib_uverbs_event *event;
 221         int ret = 0;
 222
 223         spin_lock_irq(&ev_queue->lock);
 224
 225         while (list_empty(&ev_queue->event_list)) {
 226                 spin_unlock_irq(&ev_queue->lock);
 227
 228                 if (filp->f_flags & O_NONBLOCK)
 229                         return -EAGAIN;
 230
 231                 if (wait_event_interruptible(ev_queue->poll_wait,
 232                                              (!list_empty(&ev_queue->event_list) ||
 233                         /* The barriers built into wait_event_interruptible()
 234                          * and wake_up() guarentee this will see the null set
 235                          * without using RCU
 236                          */
 237                                              !uverbs_file->device->ib_dev)))
 238                         return -ERESTARTSYS;
 239
 240                 /* If device was disassociated and no event exists set an error */
 241                 if (list_empty(&ev_queue->event_list) &&
 242                     !uverbs_file->device->ib_dev)
 243                         return -EIO;
 244
 245                 spin_lock_irq(&ev_queue->lock);
 246         }
 247
 248         event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
 249
 250         if (eventsz > count) {
 251                 ret   = -EINVAL;
 252                 event = NULL;
 253         } else {
 254                 list_del(ev_queue->event_list.next);
 255                 if (event->counter) {
 256                         ++(*event->counter);
 257                         list_del(&event->obj_list);
 258                 }
 259         }
 260
 261         spin_unlock_irq(&ev_queue->lock);
 262
 263         if (event) {
 264                 if (copy_to_user(buf, event, eventsz))
 265                         ret = -EFAULT;
 266                 else
 267                         ret = eventsz;
 268         }
 269
 270         kfree(event);
 271
 272         return ret;
 273 }
 274
 275 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
 276                                           size_t count, loff_t *pos)
 277 {
 278         struct ib_uverbs_async_event_file *file = filp->private_data;
 279
 280         return ib_uverbs_event_read(&file->ev_queue, file->uverbs_file, filp,
 281                                     buf, count, pos,
 282                                     sizeof(struct ib_uverbs_async_event_desc));
 283 }
 284
 285 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
 286                                          size_t count, loff_t *pos)
 287 {
 288         struct ib_uverbs_completion_event_file *comp_ev_file =
 289                 filp->private_data;
 290
 291         return ib_uverbs_event_read(&comp_ev_file->ev_queue,
 292                                     comp_ev_file->uobj.ufile, filp,
 293                                     buf, count, pos,
 294                                     sizeof(struct ib_uverbs_comp_event_desc));
 295 }
 296
 297 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
 298                                          struct file *filp,
 299                                          struct poll_table_struct *wait)
 300 {
 301         __poll_t pollflags = 0;
 302
 303         poll_wait(filp, &ev_queue->poll_wait, wait);
 304
 305         spin_lock_irq(&ev_queue->lock);
 306         if (!list_empty(&ev_queue->event_list))
 307                 pollflags = EPOLLIN | EPOLLRDNORM;
 308         spin_unlock_irq(&ev_queue->lock);
 309
 310         return pollflags;
 311 }
 312
 313 static __poll_t ib_uverbs_async_event_poll(struct file *filp,
 314                                                struct poll_table_struct *wait)
 315 {
 316         return ib_uverbs_event_poll(filp->private_data, filp, wait);
 317 }
 318
 319 static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
 320                                               struct poll_table_struct *wait)
 321 {
 322         struct ib_uverbs_completion_event_file *comp_ev_file =
 323                 filp->private_data;
 324
 325         return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
 326 }
 327
 328 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
 329 {
 330         struct ib_uverbs_event_queue *ev_queue = filp->private_data;
 331
 332         return fasync_helper(fd, filp, on, &ev_queue->async_queue);
 333 }
 334
 335 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
 336 {
 337         struct ib_uverbs_completion_event_file *comp_ev_file =
 338                 filp->private_data;
 339
 340         return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
 341 }
 342
 343 static int ib_uverbs_async_event_close(struct inode *inode, struct file *filp)
 344 {
 345         struct ib_uverbs_async_event_file *file = filp->private_data;
 346         struct ib_uverbs_file *uverbs_file = file->uverbs_file;
 347         struct ib_uverbs_event *entry, *tmp;
 348         int closed_already = 0;
 349
 350         mutex_lock(&uverbs_file->device->lists_mutex);
 351         spin_lock_irq(&file->ev_queue.lock);
 352         closed_already = file->ev_queue.is_closed;
 353         file->ev_queue.is_closed = 1;
 354         list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
 355                 if (entry->counter)
 356                         list_del(&entry->obj_list);
 357                 kfree(entry);
 358         }
 359         spin_unlock_irq(&file->ev_queue.lock);
 360         if (!closed_already) {
 361                 list_del(&file->list);
 362                 ib_unregister_event_handler(&uverbs_file->event_handler);
 363         }
 364         mutex_unlock(&uverbs_file->device->lists_mutex);
 365
 366         kref_put(&uverbs_file->ref, ib_uverbs_release_file);
 367         kref_put(&file->ref, ib_uverbs_release_async_event_file);
 368
 369         return 0;
 370 }
 371
 372 static int ib_uverbs_comp_event_close(struct inode *inode, struct file *filp)
 373 {
 374         struct ib_uobject *uobj = filp->private_data;
 375         struct ib_uverbs_completion_event_file *file = container_of(
 376                 uobj, struct ib_uverbs_completion_event_file, uobj);
 377         struct ib_uverbs_event *entry, *tmp;
 378
 379         spin_lock_irq(&file->ev_queue.lock);
 380         list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
 381                 if (entry->counter)
 382                         list_del(&entry->obj_list);
 383                 kfree(entry);
 384         }
 385         file->ev_queue.is_closed = 1;
 386         spin_unlock_irq(&file->ev_queue.lock);
 387
 388         uverbs_close_fd(filp);
 389
 390         return 0;
 391 }
 392
 393 const struct file_operations uverbs_event_fops = {
 394         .owner   = THIS_MODULE,
 395         .read    = ib_uverbs_comp_event_read,
 396         .poll    = ib_uverbs_comp_event_poll,
 397         .release = ib_uverbs_comp_event_close,
 398         .fasync  = ib_uverbs_comp_event_fasync,
 399         .llseek  = no_llseek,
 400 };
 401
 402 static const struct file_operations uverbs_async_event_fops = {
 403         .owner   = THIS_MODULE,
 404         .read    = ib_uverbs_async_event_read,
 405         .poll    = ib_uverbs_async_event_poll,
 406         .release = ib_uverbs_async_event_close,
 407         .fasync  = ib_uverbs_async_event_fasync,
 408         .llseek  = no_llseek,
 409 };
 410
 411 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 412 {
 413         struct ib_uverbs_event_queue   *ev_queue = cq_context;
 414         struct ib_ucq_object           *uobj;
 415         struct ib_uverbs_event         *entry;
 416         unsigned long                   flags;
 417
 418         if (!ev_queue)
 419                 return;
 420
 421         spin_lock_irqsave(&ev_queue->lock, flags);
 422         if (ev_queue->is_closed) {
 423                 spin_unlock_irqrestore(&ev_queue->lock, flags);
 424                 return;
 425         }
 426
 427         entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 428         if (!entry) {
 429                 spin_unlock_irqrestore(&ev_queue->lock, flags);
 430                 return;
 431         }
 432
 433         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 434
 435         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 436         entry->counter             = &uobj->comp_events_reported;
 437
 438         list_add_tail(&entry->list, &ev_queue->event_list);
 439         list_add_tail(&entry->obj_list, &uobj->comp_list);
 440         spin_unlock_irqrestore(&ev_queue->lock, flags);
 441
 442         wake_up_interruptible(&ev_queue->poll_wait);
 443         kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
 444 }
 445
 446 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 447                                     __u64 element, __u64 event,
 448                                     struct list_head *obj_list,
 449                                     u32 *counter)
 450 {
 451         struct ib_uverbs_event *entry;
 452         unsigned long flags;
 453
 454         spin_lock_irqsave(&file->async_file->ev_queue.lock, flags);
 455         if (file->async_file->ev_queue.is_closed) {
 456                 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 457                 return;
 458         }
 459
 460         entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 461         if (!entry) {
 462                 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 463                 return;
 464         }
 465
 466         entry->desc.async.element    = element;
 467         entry->desc.async.event_type = event;
 468         entry->desc.async.reserved   = 0;
 469         entry->counter               = counter;
 470
 471         list_add_tail(&entry->list, &file->async_file->ev_queue.event_list);
 472         if (obj_list)
 473                 list_add_tail(&entry->obj_list, obj_list);
 474         spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 475
 476         wake_up_interruptible(&file->async_file->ev_queue.poll_wait);
 477         kill_fasync(&file->async_file->ev_queue.async_queue, SIGIO, POLL_IN);
 478 }
 479
 480 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 481 {
 482         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 483                                                   struct ib_ucq_object, uobject);
 484
 485         ib_uverbs_async_handler(uobj->uobject.ufile, uobj->uobject.user_handle,
 486                                 event->event, &uobj->async_list,
 487                                 &uobj->async_events_reported);
 488 }
 489
 490 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 491 {
 492         struct ib_uevent_object *uobj;
 493
 494         /* for XRC target qp's, check that qp is live */
 495         if (!event->element.qp->uobject)
 496                 return;
 497
 498         uobj = container_of(event->element.qp->uobject,
 499                             struct ib_uevent_object, uobject);
 500
 501         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 502                                 event->event, &uobj->event_list,
 503                                 &uobj->events_reported);
 504 }
 505
 506 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
 507 {
 508         struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
 509                                                   struct ib_uevent_object, uobject);
 510
 511         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 512                                 event->event, &uobj->event_list,
 513                                 &uobj->events_reported);
 514 }
 515
 516 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 517 {
 518         struct ib_uevent_object *uobj;
 519
 520         uobj = container_of(event->element.srq->uobject,
 521                             struct ib_uevent_object, uobject);
 522
 523         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 524                                 event->event, &uobj->event_list,
 525                                 &uobj->events_reported);
 526 }
 527
 528 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 529                              struct ib_event *event)
 530 {
 531         struct ib_uverbs_file *file =
 532                 container_of(handler, struct ib_uverbs_file, event_handler);
 533
 534         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 535                                 NULL, NULL);
 536 }
 537
 538 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
 539 {
 540         kref_put(&file->async_file->ref, ib_uverbs_release_async_event_file);
 541         file->async_file = NULL;
 542 }
 543
 544 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
 545 {
 546         spin_lock_init(&ev_queue->lock);
 547         INIT_LIST_HEAD(&ev_queue->event_list);
 548         init_waitqueue_head(&ev_queue->poll_wait);
 549         ev_queue->is_closed   = 0;
 550         ev_queue->async_queue = NULL;
 551 }
 552
 553 struct file *ib_uverbs_alloc_async_event_file(struct ib_uverbs_file *uverbs_file,
 554                                               struct ib_device  *ib_dev)
 555 {
 556         struct ib_uverbs_async_event_file *ev_file;
 557         struct file *filp;
 558
 559         ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
 560         if (!ev_file)
 561                 return ERR_PTR(-ENOMEM);
 562
 563         ib_uverbs_init_event_queue(&ev_file->ev_queue);
 564         ev_file->uverbs_file = uverbs_file;
 565         kref_get(&ev_file->uverbs_file->ref);
 566         kref_init(&ev_file->ref);
 567         filp = anon_inode_getfile("[infinibandevent]", &uverbs_async_event_fops,
 568                                   ev_file, O_RDONLY);
 569         if (IS_ERR(filp))
 570                 goto err_put_refs;
 571
 572         mutex_lock(&uverbs_file->device->lists_mutex);
 573         list_add_tail(&ev_file->list,
 574                       &uverbs_file->device->uverbs_events_file_list);
 575         mutex_unlock(&uverbs_file->device->lists_mutex);
 576
 577         WARN_ON(uverbs_file->async_file);
 578         uverbs_file->async_file = ev_file;
 579         kref_get(&uverbs_file->async_file->ref);
 580         INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
 581                               ib_dev,
 582                               ib_uverbs_event_handler);
 583         ib_register_event_handler(&uverbs_file->event_handler);
 584         /* At that point async file stuff was fully set */
 585
 586         return filp;
 587
 588 err_put_refs:
 589         kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
 590         kref_put(&ev_file->ref, ib_uverbs_release_async_event_file);
 591         return filp;
 592 }
 593
 594 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
 595                           struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count,
 596                           const struct uverbs_api_write_method *method_elm)
 597 {
 598         if (method_elm->is_ex) {
 599                 count -= sizeof(*hdr) + sizeof(*ex_hdr);
 600
 601                 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
 602                         return -EINVAL;
 603
 604                 if (hdr->in_words * 8 < method_elm->req_size)
 605                         return -ENOSPC;
 606
 607                 if (ex_hdr->cmd_hdr_reserved)
 608                         return -EINVAL;
 609
 610                 if (ex_hdr->response) {
 611                         if (!hdr->out_words && !ex_hdr->provider_out_words)
 612                                 return -EINVAL;
 613
 614                         if (hdr->out_words * 8 < method_elm->resp_size)
 615                                 return -ENOSPC;
 616
 617                         if (!access_ok(u64_to_user_ptr(ex_hdr->response),
 618                                        (hdr->out_words + ex_hdr->provider_out_words) * 8))
 619                                 return -EFAULT;
 620                 } else {
 621                         if (hdr->out_words || ex_hdr->provider_out_words)
 622                                 return -EINVAL;
 623                 }
 624
 625                 return 0;
 626         }
 627
 628         /* not extended command */
 629         if (hdr->in_words * 4 != count)
 630                 return -EINVAL;
 631
 632         if (count < method_elm->req_size + sizeof(hdr)) {
 633                 /*
 634                  * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
 635                  * with a 16 byte write instead of 24. Old kernels didn't
 636                  * check the size so they allowed this. Now that the size is
 637                  * checked provide a compatibility work around to not break
 638                  * those userspaces.
 639                  */
 640                 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
 641                     count == 16) {
 642                         hdr->in_words = 6;
 643                         return 0;
 644                 }
 645                 return -ENOSPC;
 646         }
 647         if (hdr->out_words * 4 < method_elm->resp_size)
 648                 return -ENOSPC;
 649
 650         return 0;
 651 }
 652
 653 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 654                              size_t count, loff_t *pos)
 655 {
 656         struct ib_uverbs_file *file = filp->private_data;
 657         const struct uverbs_api_write_method *method_elm;
 658         struct uverbs_api *uapi = file->device->uapi;
 659         struct ib_uverbs_ex_cmd_hdr ex_hdr;
 660         struct ib_uverbs_cmd_hdr hdr;
 661         struct uverbs_attr_bundle bundle;
 662         int srcu_key;
 663         ssize_t ret;
 664
 665         if (!ib_safe_file_access(filp)) {
 666                 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
 667                             task_tgid_vnr(current), current->comm);
 668                 return -EACCES;
 669         }
 670
 671         if (count < sizeof(hdr))
 672                 return -EINVAL;
 673
 674         if (copy_from_user(&hdr, buf, sizeof(hdr)))
 675                 return -EFAULT;
 676
 677         method_elm = uapi_get_method(uapi, hdr.command);
 678         if (IS_ERR(method_elm))
 679                 return PTR_ERR(method_elm);
 680
 681         if (method_elm->is_ex) {
 682                 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 683                         return -EINVAL;
 684                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
 685                         return -EFAULT;
 686         }
 687
 688         ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
 689         if (ret)
 690                 return ret;
 691
 692         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 693
 694         buf += sizeof(hdr);
 695
 696         memset(bundle.attr_present, 0, sizeof(bundle.attr_present));
 697         bundle.ufile = file;
 698         if (!method_elm->is_ex) {
 699                 size_t in_len = hdr.in_words * 4 - sizeof(hdr);
 700                 size_t out_len = hdr.out_words * 4;
 701                 u64 response = 0;
 702
 703                 if (method_elm->has_udata) {
 704                         bundle.driver_udata.inlen =
 705                                 in_len - method_elm->req_size;
 706                         in_len = method_elm->req_size;
 707                         if (bundle.driver_udata.inlen)
 708                                 bundle.driver_udata.inbuf = buf + in_len;
 709                         else
 710                                 bundle.driver_udata.inbuf = NULL;
 711                 } else {
 712                         memset(&bundle.driver_udata, 0,
 713                                sizeof(bundle.driver_udata));
 714                 }
 715
 716                 if (method_elm->has_resp) {
 717                         /*
 718                          * The macros check that if has_resp is set
 719                          * then the command request structure starts
 720                          * with a '__aligned u64 response' member.
 721                          */
 722                         ret = get_user(response, (const u64 *)buf);
 723                         if (ret)
 724                                 goto out_unlock;
 725
 726                         if (method_elm->has_udata) {
 727                                 bundle.driver_udata.outlen =
 728                                         out_len - method_elm->resp_size;
 729                                 out_len = method_elm->resp_size;
 730                                 if (bundle.driver_udata.outlen)
 731                                         bundle.driver_udata.outbuf =
 732                                                 u64_to_user_ptr(response +
 733                                                                 out_len);
 734                                 else
 735                                         bundle.driver_udata.outbuf = NULL;
 736                         }
 737                 } else {
 738                         bundle.driver_udata.outlen = 0;
 739                         bundle.driver_udata.outbuf = NULL;
 740                 }
 741
 742                 ib_uverbs_init_udata_buf_or_null(
 743                         &bundle.ucore, buf, u64_to_user_ptr(response),
 744                         in_len, out_len);
 745         } else {
 746                 buf += sizeof(ex_hdr);
 747
 748                 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
 749                                         u64_to_user_ptr(ex_hdr.response),
 750                                         hdr.in_words * 8, hdr.out_words * 8);
 751
 752                 ib_uverbs_init_udata_buf_or_null(
 753                         &bundle.driver_udata, buf + bundle.ucore.inlen,
 754                         u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
 755                         ex_hdr.provider_in_words * 8,
 756                         ex_hdr.provider_out_words * 8);
 757
 758         }
 759
 760         ret = method_elm->handler(&bundle);
 761 out_unlock:
 762         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 763         return (ret) ? : count;
 764 }
 765
 766 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 767 {
 768         struct ib_uverbs_file *file = filp->private_data;
 769         struct ib_ucontext *ucontext;
 770         int ret = 0;
 771         int srcu_key;
 772
 773         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 774         ucontext = ib_uverbs_get_ucontext_file(file);
 775         if (IS_ERR(ucontext)) {
 776                 ret = PTR_ERR(ucontext);
 777                 goto out;
 778         }
 779
 780         ret = ucontext->device->ops.mmap(ucontext, vma);
 781 out:
 782         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 783         return ret;
 784 }
 785
 786 /*
 787  * Each time we map IO memory into user space this keeps track of the mapping.
 788  * When the device is hot-unplugged we 'zap' the mmaps in user space to point
 789  * to the zero page and allow the hot unplug to proceed.
 790  *
 791  * This is necessary for cases like PCI physical hot unplug as the actual BAR
 792  * memory may vanish after this and access to it from userspace could MCE.
 793  *
 794  * RDMA drivers supporting disassociation must have their user space designed
 795  * to cope in some way with their IO pages going to the zero page.
 796  */
 797 struct rdma_umap_priv {
 798         struct vm_area_struct *vma;
 799         struct list_head list;
 800 };
 801
 802 static const struct vm_operations_struct rdma_umap_ops;
 803
 804 static void rdma_umap_priv_init(struct rdma_umap_priv *priv,
 805                                 struct vm_area_struct *vma)
 806 {
 807         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 808
 809         priv->vma = vma;
 810         vma->vm_private_data = priv;
 811         vma->vm_ops = &rdma_umap_ops;
 812
 813         mutex_lock(&ufile->umap_lock);
 814         list_add(&priv->list, &ufile->umaps);
 815         mutex_unlock(&ufile->umap_lock);
 816 }
 817
 818 /*
 819  * The VMA has been dup'd, initialize the vm_private_data with a new tracking
 820  * struct
 821  */
 822 static void rdma_umap_open(struct vm_area_struct *vma)
 823 {
 824         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 825         struct rdma_umap_priv *opriv = vma->vm_private_data;
 826         struct rdma_umap_priv *priv;
 827
 828         if (!opriv)
 829                 return;
 830
 831         /* We are racing with disassociation */
 832         if (!down_read_trylock(&ufile->hw_destroy_rwsem))
 833                 goto out_zap;
 834         /*
 835          * Disassociation already completed, the VMA should already be zapped.
 836          */
 837         if (!ufile->ucontext)
 838                 goto out_unlock;
 839
 840         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 841         if (!priv)
 842                 goto out_unlock;
 843         rdma_umap_priv_init(priv, vma);
 844
 845         up_read(&ufile->hw_destroy_rwsem);
 846         return;
 847
 848 out_unlock:
 849         up_read(&ufile->hw_destroy_rwsem);
 850 out_zap:
 851         /*
 852          * We can't allow the VMA to be created with the actual IO pages, that
 853          * would break our API contract, and it can't be stopped at this
 854          * point, so zap it.
 855          */
 856         vma->vm_private_data = NULL;
 857         zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
 858 }
 859
 860 static void rdma_umap_close(struct vm_area_struct *vma)
 861 {
 862         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 863         struct rdma_umap_priv *priv = vma->vm_private_data;
 864
 865         if (!priv)
 866                 return;
 867
 868         /*
 869          * The vma holds a reference on the struct file that created it, which
 870          * in turn means that the ib_uverbs_file is guaranteed to exist at
 871          * this point.
 872          */
 873         mutex_lock(&ufile->umap_lock);
 874         list_del(&priv->list);
 875         mutex_unlock(&ufile->umap_lock);
 876         kfree(priv);
 877 }
 878
 879 static const struct vm_operations_struct rdma_umap_ops = {
 880         .open = rdma_umap_open,
 881         .close = rdma_umap_close,
 882 };
 883
 884 static struct rdma_umap_priv *rdma_user_mmap_pre(struct ib_ucontext *ucontext,
 885                                                  struct vm_area_struct *vma,
 886                                                  unsigned long size)
 887 {
 888         struct ib_uverbs_file *ufile = ucontext->ufile;
 889         struct rdma_umap_priv *priv;
 890
 891         if (vma->vm_end - vma->vm_start != size)
 892                 return ERR_PTR(-EINVAL);
 893
 894         /* Driver is using this wrong, must be called by ib_uverbs_mmap */
 895         if (WARN_ON(!vma->vm_file ||
 896                     vma->vm_file->private_data != ufile))
 897                 return ERR_PTR(-EINVAL);
 898         lockdep_assert_held(&ufile->device->disassociate_srcu);
 899
 900         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 901         if (!priv)
 902                 return ERR_PTR(-ENOMEM);
 903         return priv;
 904 }
 905
 906 /*
 907  * Map IO memory into a process. This is to be called by drivers as part of
 908  * their mmap() functions if they wish to send something like PCI-E BAR memory
 909  * to userspace.
 910  */
 911 int rdma_user_mmap_io(struct ib_ucontext *ucontext, struct vm_area_struct *vma,
 912                       unsigned long pfn, unsigned long size, pgprot_t prot)
 913 {
 914         struct rdma_umap_priv *priv = rdma_user_mmap_pre(ucontext, vma, size);
 915
 916         if (IS_ERR(priv))
 917                 return PTR_ERR(priv);
 918
 919         vma->vm_page_prot = prot;
 920         if (io_remap_pfn_range(vma, vma->vm_start, pfn, size, prot)) {
 921                 kfree(priv);
 922                 return -EAGAIN;
 923         }
 924
 925         rdma_umap_priv_init(priv, vma);
 926         return 0;
 927 }
 928 EXPORT_SYMBOL(rdma_user_mmap_io);
 929
 930 /*
 931  * The page case is here for a slightly different reason, the driver expects
 932  * to be able to free the page it is sharing to user space when it destroys
 933  * its ucontext, which means we need to zap the user space references.
 934  *
 935  * We could handle this differently by providing an API to allocate a shared
 936  * page and then only freeing the shared page when the last ufile is
 937  * destroyed.
 938  */
 939 int rdma_user_mmap_page(struct ib_ucontext *ucontext,
 940                         struct vm_area_struct *vma, struct page *page,
 941                         unsigned long size)
 942 {
 943         struct rdma_umap_priv *priv = rdma_user_mmap_pre(ucontext, vma, size);
 944
 945         if (IS_ERR(priv))
 946                 return PTR_ERR(priv);
 947
 948         if (remap_pfn_range(vma, vma->vm_start, page_to_pfn(page), size,
 949                             vma->vm_page_prot)) {
 950                 kfree(priv);
 951                 return -EAGAIN;
 952         }
 953
 954         rdma_umap_priv_init(priv, vma);
 955         return 0;
 956 }
 957 EXPORT_SYMBOL(rdma_user_mmap_page);
 958
 959 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
 960 {
 961         struct rdma_umap_priv *priv, *next_priv;
 962
 963         lockdep_assert_held(&ufile->hw_destroy_rwsem);
 964
 965         while (1) {
 966                 struct mm_struct *mm = NULL;
 967
 968                 /* Get an arbitrary mm pointer that hasn't been cleaned yet */
 969                 mutex_lock(&ufile->umap_lock);
 970                 while (!list_empty(&ufile->umaps)) {
 971                         int ret;
 972
 973                         priv = list_first_entry(&ufile->umaps,
 974                                                 struct rdma_umap_priv, list);
 975                         mm = priv->vma->vm_mm;
 976                         ret = mmget_not_zero(mm);
 977                         if (!ret) {
 978                                 list_del_init(&priv->list);
 979                                 mm = NULL;
 980                                 continue;
 981                         }
 982                         break;
 983                 }
 984                 mutex_unlock(&ufile->umap_lock);
 985                 if (!mm)
 986                         return;
 987
 988                 /*
 989                  * The umap_lock is nested under mmap_sem since it used within
 990                  * the vma_ops callbacks, so we have to clean the list one mm
 991                  * at a time to get the lock ordering right. Typically there
 992                  * will only be one mm, so no big deal.
 993                  */
 994                 down_write(&mm->mmap_sem);
 995                 mutex_lock(&ufile->umap_lock);
 996                 list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
 997                                           list) {
 998                         struct vm_area_struct *vma = priv->vma;
 999
1000                         if (vma->vm_mm != mm)
1001                                 continue;
1002                         list_del_init(&priv->list);
1003
1004                         zap_vma_ptes(vma, vma->vm_start,
1005                                      vma->vm_end - vma->vm_start);
1006                         vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
1007                 }
1008                 mutex_unlock(&ufile->umap_lock);
1009                 up_write(&mm->mmap_sem);
1010                 mmput(mm);
1011         }
1012 }
1013
1014 /*
1015  * ib_uverbs_open() does not need the BKL:
1016  *
1017  *  - the ib_uverbs_device structures are properly reference counted and
1018  *    everything else is purely local to the file being created, so
1019  *    races against other open calls are not a problem;
1020  *  - there is no ioctl method to race against;
1021  *  - the open method will either immediately run -ENXIO, or all
1022  *    required initialization will be done.
1023  */
1024 static int ib_uverbs_open(struct inode *inode, struct file *filp)
1025 {
1026         struct ib_uverbs_device *dev;
1027         struct ib_uverbs_file *file;
1028         struct ib_device *ib_dev;
1029         int ret;
1030         int module_dependent;
1031         int srcu_key;
1032
1033         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
1034         if (!atomic_inc_not_zero(&dev->refcount))
1035                 return -ENXIO;
1036
1037         get_device(&dev->dev);
1038         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1039         mutex_lock(&dev->lists_mutex);
1040         ib_dev = srcu_dereference(dev->ib_dev,
1041                                   &dev->disassociate_srcu);
1042         if (!ib_dev) {
1043                 ret = -EIO;
1044                 goto err;
1045         }
1046
1047         /* In case IB device supports disassociate ucontext, there is no hard
1048          * dependency between uverbs device and its low level device.
1049          */
1050         module_dependent = !(ib_dev->ops.disassociate_ucontext);
1051
1052         if (module_dependent) {
1053                 if (!try_module_get(ib_dev->owner)) {
1054                         ret = -ENODEV;
1055                         goto err;
1056                 }
1057         }
1058
1059         file = kzalloc(sizeof(*file), GFP_KERNEL);
1060         if (!file) {
1061                 ret = -ENOMEM;
1062                 if (module_dependent)
1063                         goto err_module;
1064
1065                 goto err;
1066         }
1067
1068         file->device     = dev;
1069         kref_init(&file->ref);
1070         mutex_init(&file->ucontext_lock);
1071
1072         spin_lock_init(&file->uobjects_lock);
1073         INIT_LIST_HEAD(&file->uobjects);
1074         init_rwsem(&file->hw_destroy_rwsem);
1075         mutex_init(&file->umap_lock);
1076         INIT_LIST_HEAD(&file->umaps);
1077
1078         filp->private_data = file;
1079         list_add_tail(&file->list, &dev->uverbs_file_list);
1080         mutex_unlock(&dev->lists_mutex);
1081         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1082
1083         setup_ufile_idr_uobject(file);
1084
1085         return nonseekable_open(inode, filp);
1086
1087 err_module:
1088         module_put(ib_dev->owner);
1089
1090 err:
1091         mutex_unlock(&dev->lists_mutex);
1092         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1093         if (atomic_dec_and_test(&dev->refcount))
1094                 ib_uverbs_comp_dev(dev);
1095
1096         put_device(&dev->dev);
1097         return ret;
1098 }
1099
1100 static int ib_uverbs_close(struct inode *inode, struct file *filp)
1101 {
1102         struct ib_uverbs_file *file = filp->private_data;
1103
1104         uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
1105
1106         mutex_lock(&file->device->lists_mutex);
1107         list_del_init(&file->list);
1108         mutex_unlock(&file->device->lists_mutex);
1109
1110         kref_put(&file->ref, ib_uverbs_release_file);
1111
1112         return 0;
1113 }
1114
1115 static const struct file_operations uverbs_fops = {
1116         .owner   = THIS_MODULE,
1117         .write   = ib_uverbs_write,
1118         .open    = ib_uverbs_open,
1119         .release = ib_uverbs_close,
1120         .llseek  = no_llseek,
1121         .unlocked_ioctl = ib_uverbs_ioctl,
1122         .compat_ioctl = ib_uverbs_ioctl,
1123 };
1124
1125 static const struct file_operations uverbs_mmap_fops = {
1126         .owner   = THIS_MODULE,
1127         .write   = ib_uverbs_write,
1128         .mmap    = ib_uverbs_mmap,
1129         .open    = ib_uverbs_open,
1130         .release = ib_uverbs_close,
1131         .llseek  = no_llseek,
1132         .unlocked_ioctl = ib_uverbs_ioctl,
1133         .compat_ioctl = ib_uverbs_ioctl,
1134 };
1135
1136 static struct ib_client uverbs_client = {
1137         .name   = "uverbs",
1138         .add    = ib_uverbs_add_one,
1139         .remove = ib_uverbs_remove_one
1140 };
1141
1142 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1143                           char *buf)
1144 {
1145         struct ib_uverbs_device *dev =
1146                         container_of(device, struct ib_uverbs_device, dev);
1147         int ret = -ENODEV;
1148         int srcu_key;
1149         struct ib_device *ib_dev;
1150
1151         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1152         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1153         if (ib_dev)
1154                 ret = sprintf(buf, "%s\n", dev_name(&ib_dev->dev));
1155         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1156
1157         return ret;
1158 }
1159 static DEVICE_ATTR_RO(ibdev);
1160
1161 static ssize_t abi_version_show(struct device *device,
1162                                 struct device_attribute *attr, char *buf)
1163 {
1164         struct ib_uverbs_device *dev =
1165                         container_of(device, struct ib_uverbs_device, dev);
1166         int ret = -ENODEV;
1167         int srcu_key;
1168         struct ib_device *ib_dev;
1169
1170         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1171         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1172         if (ib_dev)
1173                 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1174         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1175
1176         return ret;
1177 }
1178 static DEVICE_ATTR_RO(abi_version);
1179
1180 static struct attribute *ib_dev_attrs[] = {
1181         &dev_attr_abi_version.attr,
1182         &dev_attr_ibdev.attr,
1183         NULL,
1184 };
1185
1186 static const struct attribute_group dev_attr_group = {
1187         .attrs = ib_dev_attrs,
1188 };
1189
1190 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1191                          __stringify(IB_USER_VERBS_ABI_VERSION));
1192
1193 static int ib_uverbs_create_uapi(struct ib_device *device,
1194                                  struct ib_uverbs_device *uverbs_dev)
1195 {
1196         struct uverbs_api *uapi;
1197
1198         uapi = uverbs_alloc_api(device);
1199         if (IS_ERR(uapi))
1200                 return PTR_ERR(uapi);
1201
1202         uverbs_dev->uapi = uapi;
1203         return 0;
1204 }
1205
1206 static void ib_uverbs_add_one(struct ib_device *device)
1207 {
1208         int devnum;
1209         dev_t base;
1210         struct ib_uverbs_device *uverbs_dev;
1211         int ret;
1212
1213         if (!device->ops.alloc_ucontext)
1214                 return;
1215
1216         uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1217         if (!uverbs_dev)
1218                 return;
1219
1220         ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1221         if (ret) {
1222                 kfree(uverbs_dev);
1223                 return;
1224         }
1225
1226         device_initialize(&uverbs_dev->dev);
1227         uverbs_dev->dev.class = uverbs_class;
1228         uverbs_dev->dev.parent = device->dev.parent;
1229         uverbs_dev->dev.release = ib_uverbs_release_dev;
1230         uverbs_dev->groups[0] = &dev_attr_group;
1231         uverbs_dev->dev.groups = uverbs_dev->groups;
1232         atomic_set(&uverbs_dev->refcount, 1);
1233         init_completion(&uverbs_dev->comp);
1234         uverbs_dev->xrcd_tree = RB_ROOT;
1235         mutex_init(&uverbs_dev->xrcd_tree_mutex);
1236         mutex_init(&uverbs_dev->lists_mutex);
1237         INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1238         INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1239         rcu_assign_pointer(uverbs_dev->ib_dev, device);
1240         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1241
1242         devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1243                                GFP_KERNEL);
1244         if (devnum < 0)
1245                 goto err;
1246         uverbs_dev->devnum = devnum;
1247         if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1248                 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1249         else
1250                 base = IB_UVERBS_BASE_DEV + devnum;
1251
1252         if (ib_uverbs_create_uapi(device, uverbs_dev))
1253                 goto err_uapi;
1254
1255         uverbs_dev->dev.devt = base;
1256         dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1257
1258         cdev_init(&uverbs_dev->cdev,
1259                   device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1260         uverbs_dev->cdev.owner = THIS_MODULE;
1261
1262         ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1263         if (ret)
1264                 goto err_uapi;
1265
1266         ib_set_client_data(device, &uverbs_client, uverbs_dev);
1267         return;
1268
1269 err_uapi:
1270         ida_free(&uverbs_ida, devnum);
1271 err:
1272         if (atomic_dec_and_test(&uverbs_dev->refcount))
1273                 ib_uverbs_comp_dev(uverbs_dev);
1274         wait_for_completion(&uverbs_dev->comp);
1275         put_device(&uverbs_dev->dev);
1276         return;
1277 }
1278
1279 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1280                                         struct ib_device *ib_dev)
1281 {
1282         struct ib_uverbs_file *file;
1283         struct ib_uverbs_async_event_file *event_file;
1284         struct ib_event event;
1285
1286         /* Pending running commands to terminate */
1287         uverbs_disassociate_api_pre(uverbs_dev);
1288         event.event = IB_EVENT_DEVICE_FATAL;
1289         event.element.port_num = 0;
1290         event.device = ib_dev;
1291
1292         mutex_lock(&uverbs_dev->lists_mutex);
1293         while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1294                 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1295                                         struct ib_uverbs_file, list);
1296                 list_del_init(&file->list);
1297                 kref_get(&file->ref);
1298
1299                 /* We must release the mutex before going ahead and calling
1300                  * uverbs_cleanup_ufile, as it might end up indirectly calling
1301                  * uverbs_close, for example due to freeing the resources (e.g
1302                  * mmput).
1303                  */
1304                 mutex_unlock(&uverbs_dev->lists_mutex);
1305
1306                 ib_uverbs_event_handler(&file->event_handler, &event);
1307                 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1308                 kref_put(&file->ref, ib_uverbs_release_file);
1309
1310                 mutex_lock(&uverbs_dev->lists_mutex);
1311         }
1312
1313         while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1314                 event_file = list_first_entry(&uverbs_dev->
1315                                               uverbs_events_file_list,
1316                                               struct ib_uverbs_async_event_file,
1317                                               list);
1318                 spin_lock_irq(&event_file->ev_queue.lock);
1319                 event_file->ev_queue.is_closed = 1;
1320                 spin_unlock_irq(&event_file->ev_queue.lock);
1321
1322                 list_del(&event_file->list);
1323                 ib_unregister_event_handler(
1324                         &event_file->uverbs_file->event_handler);
1325                 event_file->uverbs_file->event_handler.device =
1326                         NULL;
1327
1328                 wake_up_interruptible(&event_file->ev_queue.poll_wait);
1329                 kill_fasync(&event_file->ev_queue.async_queue, SIGIO, POLL_IN);
1330         }
1331         mutex_unlock(&uverbs_dev->lists_mutex);
1332
1333         uverbs_disassociate_api(uverbs_dev->uapi);
1334 }
1335
1336 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1337 {
1338         struct ib_uverbs_device *uverbs_dev = client_data;
1339         int wait_clients = 1;
1340
1341         if (!uverbs_dev)
1342                 return;
1343
1344         cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1345         ida_free(&uverbs_ida, uverbs_dev->devnum);
1346
1347         if (device->ops.disassociate_ucontext) {
1348                 /* We disassociate HW resources and immediately return.
1349                  * Userspace will see a EIO errno for all future access.
1350                  * Upon returning, ib_device may be freed internally and is not
1351                  * valid any more.
1352                  * uverbs_device is still available until all clients close
1353                  * their files, then the uverbs device ref count will be zero
1354                  * and its resources will be freed.
1355                  * Note: At this point no more files can be opened since the
1356                  * cdev was deleted, however active clients can still issue
1357                  * commands and close their open files.
1358                  */
1359                 ib_uverbs_free_hw_resources(uverbs_dev, device);
1360                 wait_clients = 0;
1361         }
1362
1363         if (atomic_dec_and_test(&uverbs_dev->refcount))
1364                 ib_uverbs_comp_dev(uverbs_dev);
1365         if (wait_clients)
1366                 wait_for_completion(&uverbs_dev->comp);
1367
1368         put_device(&uverbs_dev->dev);
1369 }
1370
1371 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1372 {
1373         if (mode)
1374                 *mode = 0666;
1375         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1376 }
1377
1378 static int __init ib_uverbs_init(void)
1379 {
1380         int ret;
1381
1382         ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1383                                      IB_UVERBS_NUM_FIXED_MINOR,
1384                                      "infiniband_verbs");
1385         if (ret) {
1386                 pr_err("user_verbs: couldn't register device number\n");
1387                 goto out;
1388         }
1389
1390         ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1391                                   IB_UVERBS_NUM_DYNAMIC_MINOR,
1392                                   "infiniband_verbs");
1393         if (ret) {
1394                 pr_err("couldn't register dynamic device number\n");
1395                 goto out_alloc;
1396         }
1397
1398         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1399         if (IS_ERR(uverbs_class)) {
1400                 ret = PTR_ERR(uverbs_class);
1401                 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1402                 goto out_chrdev;
1403         }
1404
1405         uverbs_class->devnode = uverbs_devnode;
1406
1407         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1408         if (ret) {
1409                 pr_err("user_verbs: couldn't create abi_version attribute\n");
1410                 goto out_class;
1411         }
1412
1413         ret = ib_register_client(&uverbs_client);
1414         if (ret) {
1415                 pr_err("user_verbs: couldn't register client\n");
1416                 goto out_class;
1417         }
1418
1419         return 0;
1420
1421 out_class:
1422         class_destroy(uverbs_class);
1423
1424 out_chrdev:
1425         unregister_chrdev_region(dynamic_uverbs_dev,
1426                                  IB_UVERBS_NUM_DYNAMIC_MINOR);
1427
1428 out_alloc:
1429         unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1430                                  IB_UVERBS_NUM_FIXED_MINOR);
1431
1432 out:
1433         return ret;
1434 }
1435
1436 static void __exit ib_uverbs_cleanup(void)
1437 {
1438         ib_unregister_client(&uverbs_client);
1439         class_destroy(uverbs_class);
1440         unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1441                                  IB_UVERBS_NUM_FIXED_MINOR);
1442         unregister_chrdev_region(dynamic_uverbs_dev,
1443                                  IB_UVERBS_NUM_DYNAMIC_MINOR);
1444 }
1445
1446 module_init(ib_uverbs_init);
1447 module_exit(ib_uverbs_cleanup);