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
 212         if (file->disassociate_page)
 213                 __free_pages(file->disassociate_page, 0);
 214         kfree(file);
 215 }
 216
 217 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
 218                                     struct ib_uverbs_file *uverbs_file,
 219                                     struct file *filp, char __user *buf,
 220                                     size_t count, loff_t *pos,
 221                                     size_t eventsz)
 222 {
 223         struct ib_uverbs_event *event;
 224         int ret = 0;
 225
 226         spin_lock_irq(&ev_queue->lock);
 227
 228         while (list_empty(&ev_queue->event_list)) {
 229                 spin_unlock_irq(&ev_queue->lock);
 230
 231                 if (filp->f_flags & O_NONBLOCK)
 232                         return -EAGAIN;
 233
 234                 if (wait_event_interruptible(ev_queue->poll_wait,
 235                                              (!list_empty(&ev_queue->event_list) ||
 236                         /* The barriers built into wait_event_interruptible()
 237                          * and wake_up() guarentee this will see the null set
 238                          * without using RCU
 239                          */
 240                                              !uverbs_file->device->ib_dev)))
 241                         return -ERESTARTSYS;
 242
 243                 /* If device was disassociated and no event exists set an error */
 244                 if (list_empty(&ev_queue->event_list) &&
 245                     !uverbs_file->device->ib_dev)
 246                         return -EIO;
 247
 248                 spin_lock_irq(&ev_queue->lock);
 249         }
 250
 251         event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
 252
 253         if (eventsz > count) {
 254                 ret   = -EINVAL;
 255                 event = NULL;
 256         } else {
 257                 list_del(ev_queue->event_list.next);
 258                 if (event->counter) {
 259                         ++(*event->counter);
 260                         list_del(&event->obj_list);
 261                 }
 262         }
 263
 264         spin_unlock_irq(&ev_queue->lock);
 265
 266         if (event) {
 267                 if (copy_to_user(buf, event, eventsz))
 268                         ret = -EFAULT;
 269                 else
 270                         ret = eventsz;
 271         }
 272
 273         kfree(event);
 274
 275         return ret;
 276 }
 277
 278 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
 279                                           size_t count, loff_t *pos)
 280 {
 281         struct ib_uverbs_async_event_file *file = filp->private_data;
 282
 283         return ib_uverbs_event_read(&file->ev_queue, file->uverbs_file, filp,
 284                                     buf, count, pos,
 285                                     sizeof(struct ib_uverbs_async_event_desc));
 286 }
 287
 288 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
 289                                          size_t count, loff_t *pos)
 290 {
 291         struct ib_uverbs_completion_event_file *comp_ev_file =
 292                 filp->private_data;
 293
 294         return ib_uverbs_event_read(&comp_ev_file->ev_queue,
 295                                     comp_ev_file->uobj.ufile, filp,
 296                                     buf, count, pos,
 297                                     sizeof(struct ib_uverbs_comp_event_desc));
 298 }
 299
 300 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
 301                                          struct file *filp,
 302                                          struct poll_table_struct *wait)
 303 {
 304         __poll_t pollflags = 0;
 305
 306         poll_wait(filp, &ev_queue->poll_wait, wait);
 307
 308         spin_lock_irq(&ev_queue->lock);
 309         if (!list_empty(&ev_queue->event_list))
 310                 pollflags = EPOLLIN | EPOLLRDNORM;
 311         spin_unlock_irq(&ev_queue->lock);
 312
 313         return pollflags;
 314 }
 315
 316 static __poll_t ib_uverbs_async_event_poll(struct file *filp,
 317                                                struct poll_table_struct *wait)
 318 {
 319         return ib_uverbs_event_poll(filp->private_data, filp, wait);
 320 }
 321
 322 static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
 323                                               struct poll_table_struct *wait)
 324 {
 325         struct ib_uverbs_completion_event_file *comp_ev_file =
 326                 filp->private_data;
 327
 328         return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
 329 }
 330
 331 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
 332 {
 333         struct ib_uverbs_event_queue *ev_queue = filp->private_data;
 334
 335         return fasync_helper(fd, filp, on, &ev_queue->async_queue);
 336 }
 337
 338 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
 339 {
 340         struct ib_uverbs_completion_event_file *comp_ev_file =
 341                 filp->private_data;
 342
 343         return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
 344 }
 345
 346 static int ib_uverbs_async_event_close(struct inode *inode, struct file *filp)
 347 {
 348         struct ib_uverbs_async_event_file *file = filp->private_data;
 349         struct ib_uverbs_file *uverbs_file = file->uverbs_file;
 350         struct ib_uverbs_event *entry, *tmp;
 351         int closed_already = 0;
 352
 353         mutex_lock(&uverbs_file->device->lists_mutex);
 354         spin_lock_irq(&file->ev_queue.lock);
 355         closed_already = file->ev_queue.is_closed;
 356         file->ev_queue.is_closed = 1;
 357         list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
 358                 if (entry->counter)
 359                         list_del(&entry->obj_list);
 360                 kfree(entry);
 361         }
 362         spin_unlock_irq(&file->ev_queue.lock);
 363         if (!closed_already) {
 364                 list_del(&file->list);
 365                 ib_unregister_event_handler(&uverbs_file->event_handler);
 366         }
 367         mutex_unlock(&uverbs_file->device->lists_mutex);
 368
 369         kref_put(&uverbs_file->ref, ib_uverbs_release_file);
 370         kref_put(&file->ref, ib_uverbs_release_async_event_file);
 371
 372         return 0;
 373 }
 374
 375 static int ib_uverbs_comp_event_close(struct inode *inode, struct file *filp)
 376 {
 377         struct ib_uobject *uobj = filp->private_data;
 378         struct ib_uverbs_completion_event_file *file = container_of(
 379                 uobj, struct ib_uverbs_completion_event_file, uobj);
 380         struct ib_uverbs_event *entry, *tmp;
 381
 382         spin_lock_irq(&file->ev_queue.lock);
 383         list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
 384                 if (entry->counter)
 385                         list_del(&entry->obj_list);
 386                 kfree(entry);
 387         }
 388         file->ev_queue.is_closed = 1;
 389         spin_unlock_irq(&file->ev_queue.lock);
 390
 391         uverbs_close_fd(filp);
 392
 393         return 0;
 394 }
 395
 396 const struct file_operations uverbs_event_fops = {
 397         .owner   = THIS_MODULE,
 398         .read    = ib_uverbs_comp_event_read,
 399         .poll    = ib_uverbs_comp_event_poll,
 400         .release = ib_uverbs_comp_event_close,
 401         .fasync  = ib_uverbs_comp_event_fasync,
 402         .llseek  = no_llseek,
 403 };
 404
 405 static const struct file_operations uverbs_async_event_fops = {
 406         .owner   = THIS_MODULE,
 407         .read    = ib_uverbs_async_event_read,
 408         .poll    = ib_uverbs_async_event_poll,
 409         .release = ib_uverbs_async_event_close,
 410         .fasync  = ib_uverbs_async_event_fasync,
 411         .llseek  = no_llseek,
 412 };
 413
 414 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 415 {
 416         struct ib_uverbs_event_queue   *ev_queue = cq_context;
 417         struct ib_ucq_object           *uobj;
 418         struct ib_uverbs_event         *entry;
 419         unsigned long                   flags;
 420
 421         if (!ev_queue)
 422                 return;
 423
 424         spin_lock_irqsave(&ev_queue->lock, flags);
 425         if (ev_queue->is_closed) {
 426                 spin_unlock_irqrestore(&ev_queue->lock, flags);
 427                 return;
 428         }
 429
 430         entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 431         if (!entry) {
 432                 spin_unlock_irqrestore(&ev_queue->lock, flags);
 433                 return;
 434         }
 435
 436         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 437
 438         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 439         entry->counter             = &uobj->comp_events_reported;
 440
 441         list_add_tail(&entry->list, &ev_queue->event_list);
 442         list_add_tail(&entry->obj_list, &uobj->comp_list);
 443         spin_unlock_irqrestore(&ev_queue->lock, flags);
 444
 445         wake_up_interruptible(&ev_queue->poll_wait);
 446         kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
 447 }
 448
 449 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 450                                     __u64 element, __u64 event,
 451                                     struct list_head *obj_list,
 452                                     u32 *counter)
 453 {
 454         struct ib_uverbs_event *entry;
 455         unsigned long flags;
 456
 457         spin_lock_irqsave(&file->async_file->ev_queue.lock, flags);
 458         if (file->async_file->ev_queue.is_closed) {
 459                 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 460                 return;
 461         }
 462
 463         entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 464         if (!entry) {
 465                 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 466                 return;
 467         }
 468
 469         entry->desc.async.element    = element;
 470         entry->desc.async.event_type = event;
 471         entry->desc.async.reserved   = 0;
 472         entry->counter               = counter;
 473
 474         list_add_tail(&entry->list, &file->async_file->ev_queue.event_list);
 475         if (obj_list)
 476                 list_add_tail(&entry->obj_list, obj_list);
 477         spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 478
 479         wake_up_interruptible(&file->async_file->ev_queue.poll_wait);
 480         kill_fasync(&file->async_file->ev_queue.async_queue, SIGIO, POLL_IN);
 481 }
 482
 483 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 484 {
 485         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 486                                                   struct ib_ucq_object, uobject);
 487
 488         ib_uverbs_async_handler(uobj->uobject.ufile, uobj->uobject.user_handle,
 489                                 event->event, &uobj->async_list,
 490                                 &uobj->async_events_reported);
 491 }
 492
 493 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 494 {
 495         struct ib_uevent_object *uobj;
 496
 497         /* for XRC target qp's, check that qp is live */
 498         if (!event->element.qp->uobject)
 499                 return;
 500
 501         uobj = container_of(event->element.qp->uobject,
 502                             struct ib_uevent_object, uobject);
 503
 504         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 505                                 event->event, &uobj->event_list,
 506                                 &uobj->events_reported);
 507 }
 508
 509 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
 510 {
 511         struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
 512                                                   struct ib_uevent_object, uobject);
 513
 514         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 515                                 event->event, &uobj->event_list,
 516                                 &uobj->events_reported);
 517 }
 518
 519 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 520 {
 521         struct ib_uevent_object *uobj;
 522
 523         uobj = container_of(event->element.srq->uobject,
 524                             struct ib_uevent_object, uobject);
 525
 526         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 527                                 event->event, &uobj->event_list,
 528                                 &uobj->events_reported);
 529 }
 530
 531 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 532                              struct ib_event *event)
 533 {
 534         struct ib_uverbs_file *file =
 535                 container_of(handler, struct ib_uverbs_file, event_handler);
 536
 537         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 538                                 NULL, NULL);
 539 }
 540
 541 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
 542 {
 543         kref_put(&file->async_file->ref, ib_uverbs_release_async_event_file);
 544         file->async_file = NULL;
 545 }
 546
 547 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
 548 {
 549         spin_lock_init(&ev_queue->lock);
 550         INIT_LIST_HEAD(&ev_queue->event_list);
 551         init_waitqueue_head(&ev_queue->poll_wait);
 552         ev_queue->is_closed   = 0;
 553         ev_queue->async_queue = NULL;
 554 }
 555
 556 struct file *ib_uverbs_alloc_async_event_file(struct ib_uverbs_file *uverbs_file,
 557                                               struct ib_device  *ib_dev)
 558 {
 559         struct ib_uverbs_async_event_file *ev_file;
 560         struct file *filp;
 561
 562         ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
 563         if (!ev_file)
 564                 return ERR_PTR(-ENOMEM);
 565
 566         ib_uverbs_init_event_queue(&ev_file->ev_queue);
 567         ev_file->uverbs_file = uverbs_file;
 568         kref_get(&ev_file->uverbs_file->ref);
 569         kref_init(&ev_file->ref);
 570         filp = anon_inode_getfile("[infinibandevent]", &uverbs_async_event_fops,
 571                                   ev_file, O_RDONLY);
 572         if (IS_ERR(filp))
 573                 goto err_put_refs;
 574
 575         mutex_lock(&uverbs_file->device->lists_mutex);
 576         list_add_tail(&ev_file->list,
 577                       &uverbs_file->device->uverbs_events_file_list);
 578         mutex_unlock(&uverbs_file->device->lists_mutex);
 579
 580         WARN_ON(uverbs_file->async_file);
 581         uverbs_file->async_file = ev_file;
 582         kref_get(&uverbs_file->async_file->ref);
 583         INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
 584                               ib_dev,
 585                               ib_uverbs_event_handler);
 586         ib_register_event_handler(&uverbs_file->event_handler);
 587         /* At that point async file stuff was fully set */
 588
 589         return filp;
 590
 591 err_put_refs:
 592         kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
 593         kref_put(&ev_file->ref, ib_uverbs_release_async_event_file);
 594         return filp;
 595 }
 596
 597 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
 598                           struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count,
 599                           const struct uverbs_api_write_method *method_elm)
 600 {
 601         if (method_elm->is_ex) {
 602                 count -= sizeof(*hdr) + sizeof(*ex_hdr);
 603
 604                 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
 605                         return -EINVAL;
 606
 607                 if (hdr->in_words * 8 < method_elm->req_size)
 608                         return -ENOSPC;
 609
 610                 if (ex_hdr->cmd_hdr_reserved)
 611                         return -EINVAL;
 612
 613                 if (ex_hdr->response) {
 614                         if (!hdr->out_words && !ex_hdr->provider_out_words)
 615                                 return -EINVAL;
 616
 617                         if (hdr->out_words * 8 < method_elm->resp_size)
 618                                 return -ENOSPC;
 619
 620                         if (!access_ok(u64_to_user_ptr(ex_hdr->response),
 621                                        (hdr->out_words + ex_hdr->provider_out_words) * 8))
 622                                 return -EFAULT;
 623                 } else {
 624                         if (hdr->out_words || ex_hdr->provider_out_words)
 625                                 return -EINVAL;
 626                 }
 627
 628                 return 0;
 629         }
 630
 631         /* not extended command */
 632         if (hdr->in_words * 4 != count)
 633                 return -EINVAL;
 634
 635         if (count < method_elm->req_size + sizeof(hdr)) {
 636                 /*
 637                  * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
 638                  * with a 16 byte write instead of 24. Old kernels didn't
 639                  * check the size so they allowed this. Now that the size is
 640                  * checked provide a compatibility work around to not break
 641                  * those userspaces.
 642                  */
 643                 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
 644                     count == 16) {
 645                         hdr->in_words = 6;
 646                         return 0;
 647                 }
 648                 return -ENOSPC;
 649         }
 650         if (hdr->out_words * 4 < method_elm->resp_size)
 651                 return -ENOSPC;
 652
 653         return 0;
 654 }
 655
 656 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 657                              size_t count, loff_t *pos)
 658 {
 659         struct ib_uverbs_file *file = filp->private_data;
 660         const struct uverbs_api_write_method *method_elm;
 661         struct uverbs_api *uapi = file->device->uapi;
 662         struct ib_uverbs_ex_cmd_hdr ex_hdr;
 663         struct ib_uverbs_cmd_hdr hdr;
 664         struct uverbs_attr_bundle bundle;
 665         int srcu_key;
 666         ssize_t ret;
 667
 668         if (!ib_safe_file_access(filp)) {
 669                 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
 670                             task_tgid_vnr(current), current->comm);
 671                 return -EACCES;
 672         }
 673
 674         if (count < sizeof(hdr))
 675                 return -EINVAL;
 676
 677         if (copy_from_user(&hdr, buf, sizeof(hdr)))
 678                 return -EFAULT;
 679
 680         method_elm = uapi_get_method(uapi, hdr.command);
 681         if (IS_ERR(method_elm))
 682                 return PTR_ERR(method_elm);
 683
 684         if (method_elm->is_ex) {
 685                 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 686                         return -EINVAL;
 687                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
 688                         return -EFAULT;
 689         }
 690
 691         ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
 692         if (ret)
 693                 return ret;
 694
 695         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 696
 697         buf += sizeof(hdr);
 698
 699         memset(bundle.attr_present, 0, sizeof(bundle.attr_present));
 700         bundle.ufile = file;
 701         bundle.context = NULL; /* only valid if bundle has uobject */
 702         if (!method_elm->is_ex) {
 703                 size_t in_len = hdr.in_words * 4 - sizeof(hdr);
 704                 size_t out_len = hdr.out_words * 4;
 705                 u64 response = 0;
 706
 707                 if (method_elm->has_udata) {
 708                         bundle.driver_udata.inlen =
 709                                 in_len - method_elm->req_size;
 710                         in_len = method_elm->req_size;
 711                         if (bundle.driver_udata.inlen)
 712                                 bundle.driver_udata.inbuf = buf + in_len;
 713                         else
 714                                 bundle.driver_udata.inbuf = NULL;
 715                 } else {
 716                         memset(&bundle.driver_udata, 0,
 717                                sizeof(bundle.driver_udata));
 718                 }
 719
 720                 if (method_elm->has_resp) {
 721                         /*
 722                          * The macros check that if has_resp is set
 723                          * then the command request structure starts
 724                          * with a '__aligned u64 response' member.
 725                          */
 726                         ret = get_user(response, (const u64 *)buf);
 727                         if (ret)
 728                                 goto out_unlock;
 729
 730                         if (method_elm->has_udata) {
 731                                 bundle.driver_udata.outlen =
 732                                         out_len - method_elm->resp_size;
 733                                 out_len = method_elm->resp_size;
 734                                 if (bundle.driver_udata.outlen)
 735                                         bundle.driver_udata.outbuf =
 736                                                 u64_to_user_ptr(response +
 737                                                                 out_len);
 738                                 else
 739                                         bundle.driver_udata.outbuf = NULL;
 740                         }
 741                 } else {
 742                         bundle.driver_udata.outlen = 0;
 743                         bundle.driver_udata.outbuf = NULL;
 744                 }
 745
 746                 ib_uverbs_init_udata_buf_or_null(
 747                         &bundle.ucore, buf, u64_to_user_ptr(response),
 748                         in_len, out_len);
 749         } else {
 750                 buf += sizeof(ex_hdr);
 751
 752                 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
 753                                         u64_to_user_ptr(ex_hdr.response),
 754                                         hdr.in_words * 8, hdr.out_words * 8);
 755
 756                 ib_uverbs_init_udata_buf_or_null(
 757                         &bundle.driver_udata, buf + bundle.ucore.inlen,
 758                         u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
 759                         ex_hdr.provider_in_words * 8,
 760                         ex_hdr.provider_out_words * 8);
 761
 762         }
 763
 764         ret = method_elm->handler(&bundle);
 765 out_unlock:
 766         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 767         return (ret) ? : count;
 768 }
 769
 770 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 771 {
 772         struct ib_uverbs_file *file = filp->private_data;
 773         struct ib_ucontext *ucontext;
 774         int ret = 0;
 775         int srcu_key;
 776
 777         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 778         ucontext = ib_uverbs_get_ucontext_file(file);
 779         if (IS_ERR(ucontext)) {
 780                 ret = PTR_ERR(ucontext);
 781                 goto out;
 782         }
 783
 784         ret = ucontext->device->ops.mmap(ucontext, vma);
 785 out:
 786         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 787         return ret;
 788 }
 789
 790 /*
 791  * Each time we map IO memory into user space this keeps track of the mapping.
 792  * When the device is hot-unplugged we 'zap' the mmaps in user space to point
 793  * to the zero page and allow the hot unplug to proceed.
 794  *
 795  * This is necessary for cases like PCI physical hot unplug as the actual BAR
 796  * memory may vanish after this and access to it from userspace could MCE.
 797  *
 798  * RDMA drivers supporting disassociation must have their user space designed
 799  * to cope in some way with their IO pages going to the zero page.
 800  */
 801 struct rdma_umap_priv {
 802         struct vm_area_struct *vma;
 803         struct list_head list;
 804 };
 805
 806 static const struct vm_operations_struct rdma_umap_ops;
 807
 808 static void rdma_umap_priv_init(struct rdma_umap_priv *priv,
 809                                 struct vm_area_struct *vma)
 810 {
 811         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 812
 813         priv->vma = vma;
 814         vma->vm_private_data = priv;
 815         vma->vm_ops = &rdma_umap_ops;
 816
 817         mutex_lock(&ufile->umap_lock);
 818         list_add(&priv->list, &ufile->umaps);
 819         mutex_unlock(&ufile->umap_lock);
 820 }
 821
 822 /*
 823  * The VMA has been dup'd, initialize the vm_private_data with a new tracking
 824  * struct
 825  */
 826 static void rdma_umap_open(struct vm_area_struct *vma)
 827 {
 828         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 829         struct rdma_umap_priv *opriv = vma->vm_private_data;
 830         struct rdma_umap_priv *priv;
 831
 832         if (!opriv)
 833                 return;
 834
 835         /* We are racing with disassociation */
 836         if (!down_read_trylock(&ufile->hw_destroy_rwsem))
 837                 goto out_zap;
 838         /*
 839          * Disassociation already completed, the VMA should already be zapped.
 840          */
 841         if (!ufile->ucontext)
 842                 goto out_unlock;
 843
 844         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 845         if (!priv)
 846                 goto out_unlock;
 847         rdma_umap_priv_init(priv, vma);
 848
 849         up_read(&ufile->hw_destroy_rwsem);
 850         return;
 851
 852 out_unlock:
 853         up_read(&ufile->hw_destroy_rwsem);
 854 out_zap:
 855         /*
 856          * We can't allow the VMA to be created with the actual IO pages, that
 857          * would break our API contract, and it can't be stopped at this
 858          * point, so zap it.
 859          */
 860         vma->vm_private_data = NULL;
 861         zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
 862 }
 863
 864 static void rdma_umap_close(struct vm_area_struct *vma)
 865 {
 866         struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 867         struct rdma_umap_priv *priv = vma->vm_private_data;
 868
 869         if (!priv)
 870                 return;
 871
 872         /*
 873          * The vma holds a reference on the struct file that created it, which
 874          * in turn means that the ib_uverbs_file is guaranteed to exist at
 875          * this point.
 876          */
 877         mutex_lock(&ufile->umap_lock);
 878         list_del(&priv->list);
 879         mutex_unlock(&ufile->umap_lock);
 880         kfree(priv);
 881 }
 882
 883 /*
 884  * Once the zap_vma_ptes has been called touches to the VMA will come here and
 885  * we return a dummy writable zero page for all the pfns.
 886  */
 887 static vm_fault_t rdma_umap_fault(struct vm_fault *vmf)
 888 {
 889         struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data;
 890         struct rdma_umap_priv *priv = vmf->vma->vm_private_data;
 891         vm_fault_t ret = 0;
 892
 893         if (!priv)
 894                 return VM_FAULT_SIGBUS;
 895
 896         /* Read only pages can just use the system zero page. */
 897         if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) {
 898                 vmf->page = ZERO_PAGE(vmf->address);
 899                 get_page(vmf->page);
 900                 return 0;
 901         }
 902
 903         mutex_lock(&ufile->umap_lock);
 904         if (!ufile->disassociate_page)
 905                 ufile->disassociate_page =
 906                         alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0);
 907
 908         if (ufile->disassociate_page) {
 909                 /*
 910                  * This VMA is forced to always be shared so this doesn't have
 911                  * to worry about COW.
 912                  */
 913                 vmf->page = ufile->disassociate_page;
 914                 get_page(vmf->page);
 915         } else {
 916                 ret = VM_FAULT_SIGBUS;
 917         }
 918         mutex_unlock(&ufile->umap_lock);
 919
 920         return ret;
 921 }
 922
 923 static const struct vm_operations_struct rdma_umap_ops = {
 924         .open = rdma_umap_open,
 925         .close = rdma_umap_close,
 926         .fault = rdma_umap_fault,
 927 };
 928
 929 static struct rdma_umap_priv *rdma_user_mmap_pre(struct ib_ucontext *ucontext,
 930                                                  struct vm_area_struct *vma,
 931                                                  unsigned long size)
 932 {
 933         struct ib_uverbs_file *ufile = ucontext->ufile;
 934         struct rdma_umap_priv *priv;
 935
 936         if (!(vma->vm_flags & VM_SHARED))
 937                 return ERR_PTR(-EINVAL);
 938
 939         if (vma->vm_end - vma->vm_start != size)
 940                 return ERR_PTR(-EINVAL);
 941
 942         /* Driver is using this wrong, must be called by ib_uverbs_mmap */
 943         if (WARN_ON(!vma->vm_file ||
 944                     vma->vm_file->private_data != ufile))
 945                 return ERR_PTR(-EINVAL);
 946         lockdep_assert_held(&ufile->device->disassociate_srcu);
 947
 948         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 949         if (!priv)
 950                 return ERR_PTR(-ENOMEM);
 951         return priv;
 952 }
 953
 954 /*
 955  * Map IO memory into a process. This is to be called by drivers as part of
 956  * their mmap() functions if they wish to send something like PCI-E BAR memory
 957  * to userspace.
 958  */
 959 int rdma_user_mmap_io(struct ib_ucontext *ucontext, struct vm_area_struct *vma,
 960                       unsigned long pfn, unsigned long size, pgprot_t prot)
 961 {
 962         struct rdma_umap_priv *priv = rdma_user_mmap_pre(ucontext, vma, size);
 963
 964         if (IS_ERR(priv))
 965                 return PTR_ERR(priv);
 966
 967         vma->vm_page_prot = prot;
 968         if (io_remap_pfn_range(vma, vma->vm_start, pfn, size, prot)) {
 969                 kfree(priv);
 970                 return -EAGAIN;
 971         }
 972
 973         rdma_umap_priv_init(priv, vma);
 974         return 0;
 975 }
 976 EXPORT_SYMBOL(rdma_user_mmap_io);
 977
 978 /*
 979  * The page case is here for a slightly different reason, the driver expects
 980  * to be able to free the page it is sharing to user space when it destroys
 981  * its ucontext, which means we need to zap the user space references.
 982  *
 983  * We could handle this differently by providing an API to allocate a shared
 984  * page and then only freeing the shared page when the last ufile is
 985  * destroyed.
 986  */
 987 int rdma_user_mmap_page(struct ib_ucontext *ucontext,
 988                         struct vm_area_struct *vma, struct page *page,
 989                         unsigned long size)
 990 {
 991         struct rdma_umap_priv *priv = rdma_user_mmap_pre(ucontext, vma, size);
 992
 993         if (IS_ERR(priv))
 994                 return PTR_ERR(priv);
 995
 996         if (remap_pfn_range(vma, vma->vm_start, page_to_pfn(page), size,
 997                             vma->vm_page_prot)) {
 998                 kfree(priv);
 999                 return -EAGAIN;
1000         }
1001
1002         rdma_umap_priv_init(priv, vma);
1003         return 0;
1004 }
1005 EXPORT_SYMBOL(rdma_user_mmap_page);
1006
1007 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
1008 {
1009         struct rdma_umap_priv *priv, *next_priv;
1010
1011         lockdep_assert_held(&ufile->hw_destroy_rwsem);
1012
1013         while (1) {
1014                 struct mm_struct *mm = NULL;
1015
1016                 /* Get an arbitrary mm pointer that hasn't been cleaned yet */
1017                 mutex_lock(&ufile->umap_lock);
1018                 while (!list_empty(&ufile->umaps)) {
1019                         int ret;
1020
1021                         priv = list_first_entry(&ufile->umaps,
1022                                                 struct rdma_umap_priv, list);
1023                         mm = priv->vma->vm_mm;
1024                         ret = mmget_not_zero(mm);
1025                         if (!ret) {
1026                                 list_del_init(&priv->list);
1027                                 mm = NULL;
1028                                 continue;
1029                         }
1030                         break;
1031                 }
1032                 mutex_unlock(&ufile->umap_lock);
1033                 if (!mm)
1034                         return;
1035
1036                 /*
1037                  * The umap_lock is nested under mmap_sem since it used within
1038                  * the vma_ops callbacks, so we have to clean the list one mm
1039                  * at a time to get the lock ordering right. Typically there
1040                  * will only be one mm, so no big deal.
1041                  */
1042                 down_read(&mm->mmap_sem);
1043                 if (!mmget_still_valid(mm))
1044                         goto skip_mm;
1045                 mutex_lock(&ufile->umap_lock);
1046                 list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
1047                                           list) {
1048                         struct vm_area_struct *vma = priv->vma;
1049
1050                         if (vma->vm_mm != mm)
1051                                 continue;
1052                         list_del_init(&priv->list);
1053
1054                         zap_vma_ptes(vma, vma->vm_start,
1055                                      vma->vm_end - vma->vm_start);
1056                 }
1057                 mutex_unlock(&ufile->umap_lock);
1058         skip_mm:
1059                 up_read(&mm->mmap_sem);
1060                 mmput(mm);
1061         }
1062 }
1063
1064 /*
1065  * ib_uverbs_open() does not need the BKL:
1066  *
1067  *  - the ib_uverbs_device structures are properly reference counted and
1068  *    everything else is purely local to the file being created, so
1069  *    races against other open calls are not a problem;
1070  *  - there is no ioctl method to race against;
1071  *  - the open method will either immediately run -ENXIO, or all
1072  *    required initialization will be done.
1073  */
1074 static int ib_uverbs_open(struct inode *inode, struct file *filp)
1075 {
1076         struct ib_uverbs_device *dev;
1077         struct ib_uverbs_file *file;
1078         struct ib_device *ib_dev;
1079         int ret;
1080         int module_dependent;
1081         int srcu_key;
1082
1083         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
1084         if (!atomic_inc_not_zero(&dev->refcount))
1085                 return -ENXIO;
1086
1087         get_device(&dev->dev);
1088         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1089         mutex_lock(&dev->lists_mutex);
1090         ib_dev = srcu_dereference(dev->ib_dev,
1091                                   &dev->disassociate_srcu);
1092         if (!ib_dev) {
1093                 ret = -EIO;
1094                 goto err;
1095         }
1096
1097         /* In case IB device supports disassociate ucontext, there is no hard
1098          * dependency between uverbs device and its low level device.
1099          */
1100         module_dependent = !(ib_dev->ops.disassociate_ucontext);
1101
1102         if (module_dependent) {
1103                 if (!try_module_get(ib_dev->owner)) {
1104                         ret = -ENODEV;
1105                         goto err;
1106                 }
1107         }
1108
1109         file = kzalloc(sizeof(*file), GFP_KERNEL);
1110         if (!file) {
1111                 ret = -ENOMEM;
1112                 if (module_dependent)
1113                         goto err_module;
1114
1115                 goto err;
1116         }
1117
1118         file->device     = dev;
1119         kref_init(&file->ref);
1120         mutex_init(&file->ucontext_lock);
1121
1122         spin_lock_init(&file->uobjects_lock);
1123         INIT_LIST_HEAD(&file->uobjects);
1124         init_rwsem(&file->hw_destroy_rwsem);
1125         mutex_init(&file->umap_lock);
1126         INIT_LIST_HEAD(&file->umaps);
1127
1128         filp->private_data = file;
1129         list_add_tail(&file->list, &dev->uverbs_file_list);
1130         mutex_unlock(&dev->lists_mutex);
1131         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1132
1133         setup_ufile_idr_uobject(file);
1134
1135         return nonseekable_open(inode, filp);
1136
1137 err_module:
1138         module_put(ib_dev->owner);
1139
1140 err:
1141         mutex_unlock(&dev->lists_mutex);
1142         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1143         if (atomic_dec_and_test(&dev->refcount))
1144                 ib_uverbs_comp_dev(dev);
1145
1146         put_device(&dev->dev);
1147         return ret;
1148 }
1149
1150 static int ib_uverbs_close(struct inode *inode, struct file *filp)
1151 {
1152         struct ib_uverbs_file *file = filp->private_data;
1153
1154         uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
1155
1156         mutex_lock(&file->device->lists_mutex);
1157         list_del_init(&file->list);
1158         mutex_unlock(&file->device->lists_mutex);
1159
1160         kref_put(&file->ref, ib_uverbs_release_file);
1161
1162         return 0;
1163 }
1164
1165 static const struct file_operations uverbs_fops = {
1166         .owner   = THIS_MODULE,
1167         .write   = ib_uverbs_write,
1168         .open    = ib_uverbs_open,
1169         .release = ib_uverbs_close,
1170         .llseek  = no_llseek,
1171         .unlocked_ioctl = ib_uverbs_ioctl,
1172         .compat_ioctl = ib_uverbs_ioctl,
1173 };
1174
1175 static const struct file_operations uverbs_mmap_fops = {
1176         .owner   = THIS_MODULE,
1177         .write   = ib_uverbs_write,
1178         .mmap    = ib_uverbs_mmap,
1179         .open    = ib_uverbs_open,
1180         .release = ib_uverbs_close,
1181         .llseek  = no_llseek,
1182         .unlocked_ioctl = ib_uverbs_ioctl,
1183         .compat_ioctl = ib_uverbs_ioctl,
1184 };
1185
1186 static struct ib_client uverbs_client = {
1187         .name   = "uverbs",
1188         .no_kverbs_req = true,
1189         .add    = ib_uverbs_add_one,
1190         .remove = ib_uverbs_remove_one
1191 };
1192
1193 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1194                           char *buf)
1195 {
1196         struct ib_uverbs_device *dev =
1197                         container_of(device, struct ib_uverbs_device, dev);
1198         int ret = -ENODEV;
1199         int srcu_key;
1200         struct ib_device *ib_dev;
1201
1202         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1203         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1204         if (ib_dev)
1205                 ret = sprintf(buf, "%s\n", dev_name(&ib_dev->dev));
1206         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1207
1208         return ret;
1209 }
1210 static DEVICE_ATTR_RO(ibdev);
1211
1212 static ssize_t abi_version_show(struct device *device,
1213                                 struct device_attribute *attr, char *buf)
1214 {
1215         struct ib_uverbs_device *dev =
1216                         container_of(device, struct ib_uverbs_device, dev);
1217         int ret = -ENODEV;
1218         int srcu_key;
1219         struct ib_device *ib_dev;
1220
1221         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1222         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1223         if (ib_dev)
1224                 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1225         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1226
1227         return ret;
1228 }
1229 static DEVICE_ATTR_RO(abi_version);
1230
1231 static struct attribute *ib_dev_attrs[] = {
1232         &dev_attr_abi_version.attr,
1233         &dev_attr_ibdev.attr,
1234         NULL,
1235 };
1236
1237 static const struct attribute_group dev_attr_group = {
1238         .attrs = ib_dev_attrs,
1239 };
1240
1241 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1242                          __stringify(IB_USER_VERBS_ABI_VERSION));
1243
1244 static int ib_uverbs_create_uapi(struct ib_device *device,
1245                                  struct ib_uverbs_device *uverbs_dev)
1246 {
1247         struct uverbs_api *uapi;
1248
1249         uapi = uverbs_alloc_api(device);
1250         if (IS_ERR(uapi))
1251                 return PTR_ERR(uapi);
1252
1253         uverbs_dev->uapi = uapi;
1254         return 0;
1255 }
1256
1257 static void ib_uverbs_add_one(struct ib_device *device)
1258 {
1259         int devnum;
1260         dev_t base;
1261         struct ib_uverbs_device *uverbs_dev;
1262         int ret;
1263
1264         if (!device->ops.alloc_ucontext)
1265                 return;
1266
1267         uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1268         if (!uverbs_dev)
1269                 return;
1270
1271         ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1272         if (ret) {
1273                 kfree(uverbs_dev);
1274                 return;
1275         }
1276
1277         device_initialize(&uverbs_dev->dev);
1278         uverbs_dev->dev.class = uverbs_class;
1279         uverbs_dev->dev.parent = device->dev.parent;
1280         uverbs_dev->dev.release = ib_uverbs_release_dev;
1281         uverbs_dev->groups[0] = &dev_attr_group;
1282         uverbs_dev->dev.groups = uverbs_dev->groups;
1283         atomic_set(&uverbs_dev->refcount, 1);
1284         init_completion(&uverbs_dev->comp);
1285         uverbs_dev->xrcd_tree = RB_ROOT;
1286         mutex_init(&uverbs_dev->xrcd_tree_mutex);
1287         mutex_init(&uverbs_dev->lists_mutex);
1288         INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1289         INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1290         rcu_assign_pointer(uverbs_dev->ib_dev, device);
1291         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1292
1293         devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1294                                GFP_KERNEL);
1295         if (devnum < 0)
1296                 goto err;
1297         uverbs_dev->devnum = devnum;
1298         if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1299                 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1300         else
1301                 base = IB_UVERBS_BASE_DEV + devnum;
1302
1303         if (ib_uverbs_create_uapi(device, uverbs_dev))
1304                 goto err_uapi;
1305
1306         uverbs_dev->dev.devt = base;
1307         dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1308
1309         cdev_init(&uverbs_dev->cdev,
1310                   device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1311         uverbs_dev->cdev.owner = THIS_MODULE;
1312
1313         ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1314         if (ret)
1315                 goto err_uapi;
1316
1317         ib_set_client_data(device, &uverbs_client, uverbs_dev);
1318         return;
1319
1320 err_uapi:
1321         ida_free(&uverbs_ida, devnum);
1322 err:
1323         if (atomic_dec_and_test(&uverbs_dev->refcount))
1324                 ib_uverbs_comp_dev(uverbs_dev);
1325         wait_for_completion(&uverbs_dev->comp);
1326         put_device(&uverbs_dev->dev);
1327         return;
1328 }
1329
1330 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1331                                         struct ib_device *ib_dev)
1332 {
1333         struct ib_uverbs_file *file;
1334         struct ib_uverbs_async_event_file *event_file;
1335         struct ib_event event;
1336
1337         /* Pending running commands to terminate */
1338         uverbs_disassociate_api_pre(uverbs_dev);
1339         event.event = IB_EVENT_DEVICE_FATAL;
1340         event.element.port_num = 0;
1341         event.device = ib_dev;
1342
1343         mutex_lock(&uverbs_dev->lists_mutex);
1344         while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1345                 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1346                                         struct ib_uverbs_file, list);
1347                 list_del_init(&file->list);
1348                 kref_get(&file->ref);
1349
1350                 /* We must release the mutex before going ahead and calling
1351                  * uverbs_cleanup_ufile, as it might end up indirectly calling
1352                  * uverbs_close, for example due to freeing the resources (e.g
1353                  * mmput).
1354                  */
1355                 mutex_unlock(&uverbs_dev->lists_mutex);
1356
1357                 ib_uverbs_event_handler(&file->event_handler, &event);
1358                 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1359                 kref_put(&file->ref, ib_uverbs_release_file);
1360
1361                 mutex_lock(&uverbs_dev->lists_mutex);
1362         }
1363
1364         while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1365                 event_file = list_first_entry(&uverbs_dev->
1366                                               uverbs_events_file_list,
1367                                               struct ib_uverbs_async_event_file,
1368                                               list);
1369                 spin_lock_irq(&event_file->ev_queue.lock);
1370                 event_file->ev_queue.is_closed = 1;
1371                 spin_unlock_irq(&event_file->ev_queue.lock);
1372
1373                 list_del(&event_file->list);
1374                 ib_unregister_event_handler(
1375                         &event_file->uverbs_file->event_handler);
1376                 event_file->uverbs_file->event_handler.device =
1377                         NULL;
1378
1379                 wake_up_interruptible(&event_file->ev_queue.poll_wait);
1380                 kill_fasync(&event_file->ev_queue.async_queue, SIGIO, POLL_IN);
1381         }
1382         mutex_unlock(&uverbs_dev->lists_mutex);
1383
1384         uverbs_disassociate_api(uverbs_dev->uapi);
1385 }
1386
1387 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1388 {
1389         struct ib_uverbs_device *uverbs_dev = client_data;
1390         int wait_clients = 1;
1391
1392         if (!uverbs_dev)
1393                 return;
1394
1395         cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1396         ida_free(&uverbs_ida, uverbs_dev->devnum);
1397
1398         if (device->ops.disassociate_ucontext) {
1399                 /* We disassociate HW resources and immediately return.
1400                  * Userspace will see a EIO errno for all future access.
1401                  * Upon returning, ib_device may be freed internally and is not
1402                  * valid any more.
1403                  * uverbs_device is still available until all clients close
1404                  * their files, then the uverbs device ref count will be zero
1405                  * and its resources will be freed.
1406                  * Note: At this point no more files can be opened since the
1407                  * cdev was deleted, however active clients can still issue
1408                  * commands and close their open files.
1409                  */
1410                 ib_uverbs_free_hw_resources(uverbs_dev, device);
1411                 wait_clients = 0;
1412         }
1413
1414         if (atomic_dec_and_test(&uverbs_dev->refcount))
1415                 ib_uverbs_comp_dev(uverbs_dev);
1416         if (wait_clients)
1417                 wait_for_completion(&uverbs_dev->comp);
1418
1419         put_device(&uverbs_dev->dev);
1420 }
1421
1422 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1423 {
1424         if (mode)
1425                 *mode = 0666;
1426         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1427 }
1428
1429 static int __init ib_uverbs_init(void)
1430 {
1431         int ret;
1432
1433         ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1434                                      IB_UVERBS_NUM_FIXED_MINOR,
1435                                      "infiniband_verbs");
1436         if (ret) {
1437                 pr_err("user_verbs: couldn't register device number\n");
1438                 goto out;
1439         }
1440
1441         ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1442                                   IB_UVERBS_NUM_DYNAMIC_MINOR,
1443                                   "infiniband_verbs");
1444         if (ret) {
1445                 pr_err("couldn't register dynamic device number\n");
1446                 goto out_alloc;
1447         }
1448
1449         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1450         if (IS_ERR(uverbs_class)) {
1451                 ret = PTR_ERR(uverbs_class);
1452                 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1453                 goto out_chrdev;
1454         }
1455
1456         uverbs_class->devnode = uverbs_devnode;
1457
1458         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1459         if (ret) {
1460                 pr_err("user_verbs: couldn't create abi_version attribute\n");
1461                 goto out_class;
1462         }
1463
1464         ret = ib_register_client(&uverbs_client);
1465         if (ret) {
1466                 pr_err("user_verbs: couldn't register client\n");
1467                 goto out_class;
1468         }
1469
1470         return 0;
1471
1472 out_class:
1473         class_destroy(uverbs_class);
1474
1475 out_chrdev:
1476         unregister_chrdev_region(dynamic_uverbs_dev,
1477                                  IB_UVERBS_NUM_DYNAMIC_MINOR);
1478
1479 out_alloc:
1480         unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1481                                  IB_UVERBS_NUM_FIXED_MINOR);
1482
1483 out:
1484         return ret;
1485 }
1486
1487 static void __exit ib_uverbs_cleanup(void)
1488 {
1489         ib_unregister_client(&uverbs_client);
1490         class_destroy(uverbs_class);
1491         unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1492                                  IB_UVERBS_NUM_FIXED_MINOR);
1493         unregister_chrdev_region(dynamic_uverbs_dev,
1494                                  IB_UVERBS_NUM_DYNAMIC_MINOR);
1495 }
1496
1497 module_init(ib_uverbs_init);
1498 module_exit(ib_uverbs_cleanup);