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#include <rdma/rdma_netlink.h>
  55#include <rdma/ib_ucaps.h>
  56
  57#include "uverbs.h"
  58#include "core_priv.h"
  59#include "rdma_core.h"
  60
  61MODULE_AUTHOR("Roland Dreier");
  62MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  63MODULE_LICENSE("Dual BSD/GPL");
  64
  65enum {
  66	IB_UVERBS_MAJOR       = 231,
  67	IB_UVERBS_BASE_MINOR  = 192,
  68	IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS,
  69	IB_UVERBS_NUM_FIXED_MINOR = 32,
  70	IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR,
  71};
  72
  73#define IB_UVERBS_BASE_DEV	MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  74
  75static dev_t dynamic_uverbs_dev;
  76
  77static DEFINE_IDA(uverbs_ida);
  78static int ib_uverbs_add_one(struct ib_device *device);
  79static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
  80static struct ib_client uverbs_client;
  81
  82static char *uverbs_devnode(const struct device *dev, umode_t *mode)
  83{
  84	if (mode)
  85		*mode = 0666;
  86	return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
  87}
  88
  89static const struct class uverbs_class = {
  90	.name = "infiniband_verbs",
  91	.devnode = uverbs_devnode,
  92};
  93
  94/*
  95 * Must be called with the ufile->device->disassociate_srcu held, and the lock
  96 * must be held until use of the ucontext is finished.
  97 */
  98struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile)
  99{
 100	/*
 101	 * We do not hold the hw_destroy_rwsem lock for this flow, instead
 102	 * srcu is used. It does not matter if someone races this with
 103	 * get_context, we get NULL or valid ucontext.
 104	 */
 105	struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext);
 106
 107	if (!srcu_dereference(ufile->device->ib_dev,
 108			      &ufile->device->disassociate_srcu))
 109		return ERR_PTR(-EIO);
 110
 111	if (!ucontext)
 112		return ERR_PTR(-EINVAL);
 113
 114	return ucontext;
 115}
 116EXPORT_SYMBOL(ib_uverbs_get_ucontext_file);
 117
 118int uverbs_dealloc_mw(struct ib_mw *mw)
 119{
 120	struct ib_pd *pd = mw->pd;
 121	int ret;
 122
 123	ret = mw->device->ops.dealloc_mw(mw);
 124	if (ret)
 125		return ret;
 126
 127	atomic_dec(&pd->usecnt);
 128	kfree(mw);
 129	return ret;
 130}
 131
 132static void ib_uverbs_release_dev(struct device *device)
 133{
 134	struct ib_uverbs_device *dev =
 135			container_of(device, struct ib_uverbs_device, dev);
 136
 137	uverbs_destroy_api(dev->uapi);
 138	cleanup_srcu_struct(&dev->disassociate_srcu);
 139	mutex_destroy(&dev->lists_mutex);
 140	mutex_destroy(&dev->xrcd_tree_mutex);
 141	kfree(dev);
 142}
 143
 144void ib_uverbs_release_ucq(struct ib_uverbs_completion_event_file *ev_file,
 145			   struct ib_ucq_object *uobj)
 146{
 147	struct ib_uverbs_event *evt, *tmp;
 148
 149	if (ev_file) {
 150		spin_lock_irq(&ev_file->ev_queue.lock);
 151		list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 152			list_del(&evt->list);
 153			kfree(evt);
 154		}
 155		spin_unlock_irq(&ev_file->ev_queue.lock);
 156
 157		uverbs_uobject_put(&ev_file->uobj);
 158	}
 159
 160	ib_uverbs_release_uevent(&uobj->uevent);
 161}
 162
 163void ib_uverbs_release_uevent(struct ib_uevent_object *uobj)
 164{
 165	struct ib_uverbs_async_event_file *async_file = uobj->event_file;
 166	struct ib_uverbs_event *evt, *tmp;
 167
 168	if (!async_file)
 169		return;
 170
 171	spin_lock_irq(&async_file->ev_queue.lock);
 172	list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 173		list_del(&evt->list);
 174		kfree(evt);
 175	}
 176	spin_unlock_irq(&async_file->ev_queue.lock);
 177	uverbs_uobject_put(&async_file->uobj);
 178}
 179
 180void ib_uverbs_detach_umcast(struct ib_qp *qp,
 181			     struct ib_uqp_object *uobj)
 182{
 183	struct ib_uverbs_mcast_entry *mcast, *tmp;
 184
 185	list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 186		ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 187		list_del(&mcast->list);
 188		kfree(mcast);
 189	}
 190}
 191
 192static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
 193{
 194	complete(&dev->comp);
 195}
 196
 197void ib_uverbs_release_file(struct kref *ref)
 198{
 199	struct ib_uverbs_file *file =
 200		container_of(ref, struct ib_uverbs_file, ref);
 201	struct ib_device *ib_dev;
 202	int srcu_key;
 203
 204	release_ufile_idr_uobject(file);
 205
 206	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 207	ib_dev = srcu_dereference(file->device->ib_dev,
 208				  &file->device->disassociate_srcu);
 209	if (ib_dev && !ib_dev->ops.disassociate_ucontext)
 210		module_put(ib_dev->ops.owner);
 211	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 212
 213	if (refcount_dec_and_test(&file->device->refcount))
 214		ib_uverbs_comp_dev(file->device);
 215
 216	if (file->default_async_file)
 217		uverbs_uobject_put(&file->default_async_file->uobj);
 218	put_device(&file->device->dev);
 219
 220	if (file->disassociate_page)
 221		__free_pages(file->disassociate_page, 0);
 222	mutex_destroy(&file->disassociation_lock);
 223	mutex_destroy(&file->umap_lock);
 224	mutex_destroy(&file->ucontext_lock);
 225	kfree(file);
 226}
 227
 228static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
 229				    struct file *filp, char __user *buf,
 230				    size_t count, loff_t *pos,
 231				    size_t eventsz)
 232{
 233	struct ib_uverbs_event *event;
 234	int ret = 0;
 235
 236	spin_lock_irq(&ev_queue->lock);
 237
 238	while (list_empty(&ev_queue->event_list)) {
 239		if (ev_queue->is_closed) {
 240			spin_unlock_irq(&ev_queue->lock);
 241			return -EIO;
 242		}
 243
 244		spin_unlock_irq(&ev_queue->lock);
 245		if (filp->f_flags & O_NONBLOCK)
 246			return -EAGAIN;
 247
 248		if (wait_event_interruptible(ev_queue->poll_wait,
 249					     (!list_empty(&ev_queue->event_list) ||
 250					      ev_queue->is_closed)))
 251			return -ERESTARTSYS;
 252
 253		spin_lock_irq(&ev_queue->lock);
 254	}
 255
 256	event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
 257
 258	if (eventsz > count) {
 259		ret   = -EINVAL;
 260		event = NULL;
 261	} else {
 262		list_del(ev_queue->event_list.next);
 263		if (event->counter) {
 264			++(*event->counter);
 265			list_del(&event->obj_list);
 266		}
 267	}
 268
 269	spin_unlock_irq(&ev_queue->lock);
 270
 271	if (event) {
 272		if (copy_to_user(buf, event, eventsz))
 273			ret = -EFAULT;
 274		else
 275			ret = eventsz;
 276	}
 277
 278	kfree(event);
 279
 280	return ret;
 281}
 282
 283static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
 284					  size_t count, loff_t *pos)
 285{
 286	struct ib_uverbs_async_event_file *file = filp->private_data;
 287
 288	return ib_uverbs_event_read(&file->ev_queue, filp, buf, count, pos,
 289				    sizeof(struct ib_uverbs_async_event_desc));
 290}
 291
 292static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
 293					 size_t count, loff_t *pos)
 294{
 295	struct ib_uverbs_completion_event_file *comp_ev_file =
 296		filp->private_data;
 297
 298	return ib_uverbs_event_read(&comp_ev_file->ev_queue, filp, buf, count,
 299				    pos,
 300				    sizeof(struct ib_uverbs_comp_event_desc));
 301}
 302
 303static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
 304					 struct file *filp,
 305					 struct poll_table_struct *wait)
 306{
 307	__poll_t pollflags = 0;
 308
 309	poll_wait(filp, &ev_queue->poll_wait, wait);
 310
 311	spin_lock_irq(&ev_queue->lock);
 312	if (!list_empty(&ev_queue->event_list))
 313		pollflags = EPOLLIN | EPOLLRDNORM;
 314	else if (ev_queue->is_closed)
 315		pollflags = EPOLLERR;
 316	spin_unlock_irq(&ev_queue->lock);
 317
 318	return pollflags;
 319}
 320
 321static __poll_t ib_uverbs_async_event_poll(struct file *filp,
 322					       struct poll_table_struct *wait)
 323{
 324	struct ib_uverbs_async_event_file *file = filp->private_data;
 325
 326	return ib_uverbs_event_poll(&file->ev_queue, filp, wait);
 327}
 328
 329static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
 330					      struct poll_table_struct *wait)
 331{
 332	struct ib_uverbs_completion_event_file *comp_ev_file =
 333		filp->private_data;
 334
 335	return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
 336}
 337
 338static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
 339{
 340	struct ib_uverbs_async_event_file *file = filp->private_data;
 341
 342	return fasync_helper(fd, filp, on, &file->ev_queue.async_queue);
 343}
 344
 345static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
 346{
 347	struct ib_uverbs_completion_event_file *comp_ev_file =
 348		filp->private_data;
 349
 350	return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
 351}
 352
 353const struct file_operations uverbs_event_fops = {
 354	.owner	 = THIS_MODULE,
 355	.read	 = ib_uverbs_comp_event_read,
 356	.poll    = ib_uverbs_comp_event_poll,
 357	.release = uverbs_uobject_fd_release,
 358	.fasync  = ib_uverbs_comp_event_fasync,
 359};
 360
 361const struct file_operations uverbs_async_event_fops = {
 362	.owner	 = THIS_MODULE,
 363	.read	 = ib_uverbs_async_event_read,
 364	.poll    = ib_uverbs_async_event_poll,
 365	.release = uverbs_async_event_release,
 366	.fasync  = ib_uverbs_async_event_fasync,
 367};
 368
 369void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 370{
 371	struct ib_uverbs_event_queue   *ev_queue = cq_context;
 372	struct ib_ucq_object	       *uobj;
 373	struct ib_uverbs_event	       *entry;
 374	unsigned long			flags;
 375
 376	if (!ev_queue)
 377		return;
 378
 379	spin_lock_irqsave(&ev_queue->lock, flags);
 380	if (ev_queue->is_closed) {
 381		spin_unlock_irqrestore(&ev_queue->lock, flags);
 382		return;
 383	}
 384
 385	entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 386	if (!entry) {
 387		spin_unlock_irqrestore(&ev_queue->lock, flags);
 388		return;
 389	}
 390
 391	uobj = cq->uobject;
 392
 393	entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle;
 394	entry->counter		   = &uobj->comp_events_reported;
 395
 396	list_add_tail(&entry->list, &ev_queue->event_list);
 397	list_add_tail(&entry->obj_list, &uobj->comp_list);
 398	spin_unlock_irqrestore(&ev_queue->lock, flags);
 399
 400	wake_up_interruptible(&ev_queue->poll_wait);
 401	kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
 402}
 403
 404void ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file,
 405			     __u64 element, __u64 event,
 406			     struct list_head *obj_list, u32 *counter)
 407{
 408	struct ib_uverbs_event *entry;
 409	unsigned long flags;
 410
 411	if (!async_file)
 412		return;
 413
 414	spin_lock_irqsave(&async_file->ev_queue.lock, flags);
 415	if (async_file->ev_queue.is_closed) {
 416		spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
 417		return;
 418	}
 419
 420	entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 421	if (!entry) {
 422		spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
 423		return;
 424	}
 425
 426	entry->desc.async.element = element;
 427	entry->desc.async.event_type = event;
 428	entry->desc.async.reserved = 0;
 429	entry->counter = counter;
 430
 431	list_add_tail(&entry->list, &async_file->ev_queue.event_list);
 432	if (obj_list)
 433		list_add_tail(&entry->obj_list, obj_list);
 434	spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
 435
 436	wake_up_interruptible(&async_file->ev_queue.poll_wait);
 437	kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN);
 438}
 439
 440static void uverbs_uobj_event(struct ib_uevent_object *eobj,
 441			      struct ib_event *event)
 442{
 443	ib_uverbs_async_handler(eobj->event_file,
 444				eobj->uobject.user_handle, event->event,
 445				&eobj->event_list, &eobj->events_reported);
 446}
 447
 448void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 449{
 450	uverbs_uobj_event(&event->element.cq->uobject->uevent, event);
 451}
 452
 453void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 454{
 455	/* for XRC target qp's, check that qp is live */
 456	if (!event->element.qp->uobject)
 457		return;
 458
 459	uverbs_uobj_event(&event->element.qp->uobject->uevent, event);
 460}
 461
 462void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
 463{
 464	uverbs_uobj_event(&event->element.wq->uobject->uevent, event);
 465}
 466
 467void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 468{
 469	uverbs_uobj_event(&event->element.srq->uobject->uevent, event);
 470}
 471
 472static void ib_uverbs_event_handler(struct ib_event_handler *handler,
 473				    struct ib_event *event)
 474{
 475	ib_uverbs_async_handler(
 476		container_of(handler, struct ib_uverbs_async_event_file,
 477			     event_handler),
 478		event->element.port_num, event->event, NULL, NULL);
 479}
 480
 481void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
 482{
 483	spin_lock_init(&ev_queue->lock);
 484	INIT_LIST_HEAD(&ev_queue->event_list);
 485	init_waitqueue_head(&ev_queue->poll_wait);
 486	ev_queue->is_closed   = 0;
 487	ev_queue->async_queue = NULL;
 488}
 489
 490void ib_uverbs_init_async_event_file(
 491	struct ib_uverbs_async_event_file *async_file)
 492{
 493	struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile;
 494	struct ib_device *ib_dev = async_file->uobj.context->device;
 495
 496	ib_uverbs_init_event_queue(&async_file->ev_queue);
 497
 498	/* The first async_event_file becomes the default one for the file. */
 499	mutex_lock(&uverbs_file->ucontext_lock);
 500	if (!uverbs_file->default_async_file) {
 501		/* Pairs with the put in ib_uverbs_release_file */
 502		uverbs_uobject_get(&async_file->uobj);
 503		smp_store_release(&uverbs_file->default_async_file, async_file);
 504	}
 505	mutex_unlock(&uverbs_file->ucontext_lock);
 506
 507	INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev,
 508			      ib_uverbs_event_handler);
 509	ib_register_event_handler(&async_file->event_handler);
 510}
 511
 512static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
 513			  struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count,
 514			  const struct uverbs_api_write_method *method_elm)
 515{
 516	if (method_elm->is_ex) {
 517		count -= sizeof(*hdr) + sizeof(*ex_hdr);
 518
 519		if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
 520			return -EINVAL;
 521
 522		if (hdr->in_words * 8 < method_elm->req_size)
 523			return -ENOSPC;
 524
 525		if (ex_hdr->cmd_hdr_reserved)
 526			return -EINVAL;
 527
 528		if (ex_hdr->response) {
 529			if (!hdr->out_words && !ex_hdr->provider_out_words)
 530				return -EINVAL;
 531
 532			if (hdr->out_words * 8 < method_elm->resp_size)
 533				return -ENOSPC;
 534
 535			if (!access_ok(u64_to_user_ptr(ex_hdr->response),
 536				       (hdr->out_words + ex_hdr->provider_out_words) * 8))
 537				return -EFAULT;
 538		} else {
 539			if (hdr->out_words || ex_hdr->provider_out_words)
 540				return -EINVAL;
 541		}
 542
 543		return 0;
 544	}
 545
 546	/* not extended command */
 547	if (hdr->in_words * 4 != count)
 548		return -EINVAL;
 549
 550	if (count < method_elm->req_size + sizeof(*hdr)) {
 551		/*
 552		 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
 553		 * with a 16 byte write instead of 24. Old kernels didn't
 554		 * check the size so they allowed this. Now that the size is
 555		 * checked provide a compatibility work around to not break
 556		 * those userspaces.
 557		 */
 558		if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
 559		    count == 16) {
 560			hdr->in_words = 6;
 561			return 0;
 562		}
 563		return -ENOSPC;
 564	}
 565	if (hdr->out_words * 4 < method_elm->resp_size)
 566		return -ENOSPC;
 567
 568	return 0;
 569}
 570
 571static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 572			     size_t count, loff_t *pos)
 573{
 574	struct ib_uverbs_file *file = filp->private_data;
 575	const struct uverbs_api_write_method *method_elm;
 576	struct uverbs_api *uapi = file->device->uapi;
 577	struct ib_uverbs_ex_cmd_hdr ex_hdr;
 578	struct ib_uverbs_cmd_hdr hdr;
 579	struct uverbs_attr_bundle bundle;
 580	int srcu_key;
 581	ssize_t ret;
 582
 583	if (!ib_safe_file_access(filp)) {
 584		pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
 585			    task_tgid_vnr(current), current->comm);
 586		return -EACCES;
 587	}
 588
 589	if (count < sizeof(hdr))
 590		return -EINVAL;
 591
 592	if (copy_from_user(&hdr, buf, sizeof(hdr)))
 593		return -EFAULT;
 594
 595	method_elm = uapi_get_method(uapi, hdr.command);
 596	if (IS_ERR(method_elm))
 597		return PTR_ERR(method_elm);
 598
 599	if (method_elm->is_ex) {
 600		if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 601			return -EINVAL;
 602		if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
 603			return -EFAULT;
 604	}
 605
 606	ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
 607	if (ret)
 608		return ret;
 609
 610	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 611
 612	buf += sizeof(hdr);
 613
 614	memset(bundle.attr_present, 0, sizeof(bundle.attr_present));
 615	bundle.ufile = file;
 616	bundle.context = NULL; /* only valid if bundle has uobject */
 617	bundle.uobject = NULL;
 618	if (!method_elm->is_ex) {
 619		size_t in_len = hdr.in_words * 4 - sizeof(hdr);
 620		size_t out_len = hdr.out_words * 4;
 621		u64 response = 0;
 622
 623		if (method_elm->has_udata) {
 624			bundle.driver_udata.inlen =
 625				in_len - method_elm->req_size;
 626			in_len = method_elm->req_size;
 627			if (bundle.driver_udata.inlen)
 628				bundle.driver_udata.inbuf = buf + in_len;
 629			else
 630				bundle.driver_udata.inbuf = NULL;
 631		} else {
 632			memset(&bundle.driver_udata, 0,
 633			       sizeof(bundle.driver_udata));
 634		}
 635
 636		if (method_elm->has_resp) {
 637			/*
 638			 * The macros check that if has_resp is set
 639			 * then the command request structure starts
 640			 * with a '__aligned u64 response' member.
 641			 */
 642			ret = get_user(response, (const u64 __user *)buf);
 643			if (ret)
 644				goto out_unlock;
 645
 646			if (method_elm->has_udata) {
 647				bundle.driver_udata.outlen =
 648					out_len - method_elm->resp_size;
 649				out_len = method_elm->resp_size;
 650				if (bundle.driver_udata.outlen)
 651					bundle.driver_udata.outbuf =
 652						u64_to_user_ptr(response +
 653								out_len);
 654				else
 655					bundle.driver_udata.outbuf = NULL;
 656			}
 657		} else {
 658			bundle.driver_udata.outlen = 0;
 659			bundle.driver_udata.outbuf = NULL;
 660		}
 661
 662		ib_uverbs_init_udata_buf_or_null(
 663			&bundle.ucore, buf, u64_to_user_ptr(response),
 664			in_len, out_len);
 665	} else {
 666		buf += sizeof(ex_hdr);
 667
 668		ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
 669					u64_to_user_ptr(ex_hdr.response),
 670					hdr.in_words * 8, hdr.out_words * 8);
 671
 672		ib_uverbs_init_udata_buf_or_null(
 673			&bundle.driver_udata, buf + bundle.ucore.inlen,
 674			u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
 675			ex_hdr.provider_in_words * 8,
 676			ex_hdr.provider_out_words * 8);
 677
 678	}
 679
 680	ret = method_elm->handler(&bundle);
 681	if (bundle.uobject)
 682		uverbs_finalize_object(bundle.uobject, UVERBS_ACCESS_NEW, true,
 683				       !ret, &bundle);
 684out_unlock:
 685	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 686	return (ret) ? : count;
 687}
 688
 689static const struct vm_operations_struct rdma_umap_ops;
 690
 691static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 692{
 693	struct ib_uverbs_file *file = filp->private_data;
 694	struct ib_ucontext *ucontext;
 695	int ret = 0;
 696	int srcu_key;
 697
 698	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 699	ucontext = ib_uverbs_get_ucontext_file(file);
 700	if (IS_ERR(ucontext)) {
 701		ret = PTR_ERR(ucontext);
 702		goto out;
 703	}
 704
 705	mutex_lock(&file->disassociation_lock);
 706
 707	vma->vm_ops = &rdma_umap_ops;
 708	ret = ucontext->device->ops.mmap(ucontext, vma);
 709
 710	mutex_unlock(&file->disassociation_lock);
 711out:
 712	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 713	return ret;
 714}
 715
 716/*
 717 * The VMA has been dup'd, initialize the vm_private_data with a new tracking
 718 * struct
 719 */
 720static void rdma_umap_open(struct vm_area_struct *vma)
 721{
 722	struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 723	struct rdma_umap_priv *opriv = vma->vm_private_data;
 724	struct rdma_umap_priv *priv;
 725
 726	if (!opriv)
 727		return;
 728
 729	/* We are racing with disassociation */
 730	if (!down_read_trylock(&ufile->hw_destroy_rwsem))
 731		goto out_zap;
 732	mutex_lock(&ufile->disassociation_lock);
 733
 734	/*
 735	 * Disassociation already completed, the VMA should already be zapped.
 736	 */
 737	if (!ufile->ucontext)
 738		goto out_unlock;
 739
 740	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 741	if (!priv)
 742		goto out_unlock;
 743	rdma_umap_priv_init(priv, vma, opriv->entry);
 744
 745	mutex_unlock(&ufile->disassociation_lock);
 746	up_read(&ufile->hw_destroy_rwsem);
 747	return;
 748
 749out_unlock:
 750	mutex_unlock(&ufile->disassociation_lock);
 751	up_read(&ufile->hw_destroy_rwsem);
 752out_zap:
 753	/*
 754	 * We can't allow the VMA to be created with the actual IO pages, that
 755	 * would break our API contract, and it can't be stopped at this
 756	 * point, so zap it.
 757	 */
 758	vma->vm_private_data = NULL;
 759	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
 760}
 761
 762static void rdma_umap_close(struct vm_area_struct *vma)
 763{
 764	struct ib_uverbs_file *ufile = vma->vm_file->private_data;
 765	struct rdma_umap_priv *priv = vma->vm_private_data;
 766
 767	if (!priv)
 768		return;
 769
 770	/*
 771	 * The vma holds a reference on the struct file that created it, which
 772	 * in turn means that the ib_uverbs_file is guaranteed to exist at
 773	 * this point.
 774	 */
 775	mutex_lock(&ufile->umap_lock);
 776	if (priv->entry)
 777		rdma_user_mmap_entry_put(priv->entry);
 778
 779	list_del(&priv->list);
 780	mutex_unlock(&ufile->umap_lock);
 781	kfree(priv);
 782}
 783
 784/*
 785 * Once the zap_vma_ptes has been called touches to the VMA will come here and
 786 * we return a dummy writable zero page for all the pfns.
 787 */
 788static vm_fault_t rdma_umap_fault(struct vm_fault *vmf)
 789{
 790	struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data;
 791	struct rdma_umap_priv *priv = vmf->vma->vm_private_data;
 792	vm_fault_t ret = 0;
 793
 794	if (!priv)
 795		return VM_FAULT_SIGBUS;
 796
 797	/* Read only pages can just use the system zero page. */
 798	if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) {
 799		vmf->page = ZERO_PAGE(vmf->address);
 800		get_page(vmf->page);
 801		return 0;
 802	}
 803
 804	mutex_lock(&ufile->umap_lock);
 805	if (!ufile->disassociate_page)
 806		ufile->disassociate_page =
 807			alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0);
 808
 809	if (ufile->disassociate_page) {
 810		/*
 811		 * This VMA is forced to always be shared so this doesn't have
 812		 * to worry about COW.
 813		 */
 814		vmf->page = ufile->disassociate_page;
 815		get_page(vmf->page);
 816	} else {
 817		ret = VM_FAULT_SIGBUS;
 818	}
 819	mutex_unlock(&ufile->umap_lock);
 820
 821	return ret;
 822}
 823
 824static const struct vm_operations_struct rdma_umap_ops = {
 825	.open = rdma_umap_open,
 826	.close = rdma_umap_close,
 827	.fault = rdma_umap_fault,
 828};
 829
 830void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
 831{
 832	struct rdma_umap_priv *priv, *next_priv;
 833
 834	mutex_lock(&ufile->disassociation_lock);
 835
 836	while (1) {
 837		struct mm_struct *mm = NULL;
 838
 839		/* Get an arbitrary mm pointer that hasn't been cleaned yet */
 840		mutex_lock(&ufile->umap_lock);
 841		while (!list_empty(&ufile->umaps)) {
 842			int ret;
 843
 844			priv = list_first_entry(&ufile->umaps,
 845						struct rdma_umap_priv, list);
 846			mm = priv->vma->vm_mm;
 847			ret = mmget_not_zero(mm);
 848			if (!ret) {
 849				list_del_init(&priv->list);
 850				if (priv->entry) {
 851					rdma_user_mmap_entry_put(priv->entry);
 852					priv->entry = NULL;
 853				}
 854				mm = NULL;
 855				continue;
 856			}
 857			break;
 858		}
 859		mutex_unlock(&ufile->umap_lock);
 860		if (!mm) {
 861			mutex_unlock(&ufile->disassociation_lock);
 862			return;
 863		}
 864
 865		/*
 866		 * The umap_lock is nested under mmap_lock since it used within
 867		 * the vma_ops callbacks, so we have to clean the list one mm
 868		 * at a time to get the lock ordering right. Typically there
 869		 * will only be one mm, so no big deal.
 870		 */
 871		mmap_read_lock(mm);
 872		mutex_lock(&ufile->umap_lock);
 873		list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
 874					  list) {
 875			struct vm_area_struct *vma = priv->vma;
 876
 877			if (vma->vm_mm != mm)
 878				continue;
 879			list_del_init(&priv->list);
 880
 881			zap_vma_ptes(vma, vma->vm_start,
 882				     vma->vm_end - vma->vm_start);
 883
 884			if (priv->entry) {
 885				rdma_user_mmap_entry_put(priv->entry);
 886				priv->entry = NULL;
 887			}
 888		}
 889		mutex_unlock(&ufile->umap_lock);
 890		mmap_read_unlock(mm);
 891		mmput(mm);
 892	}
 893
 894	mutex_unlock(&ufile->disassociation_lock);
 895}
 896
 897/**
 898 * rdma_user_mmap_disassociate() - Revoke mmaps for a device
 899 * @device: device to revoke
 900 *
 901 * This function should be called by drivers that need to disable mmaps for the
 902 * device, for instance because it is going to be reset.
 903 */
 904void rdma_user_mmap_disassociate(struct ib_device *device)
 905{
 906	struct ib_uverbs_device *uverbs_dev =
 907		ib_get_client_data(device, &uverbs_client);
 908	struct ib_uverbs_file *ufile;
 909
 910	mutex_lock(&uverbs_dev->lists_mutex);
 911	list_for_each_entry(ufile, &uverbs_dev->uverbs_file_list, list) {
 912		if (ufile->ucontext)
 913			uverbs_user_mmap_disassociate(ufile);
 914	}
 915	mutex_unlock(&uverbs_dev->lists_mutex);
 916}
 917EXPORT_SYMBOL(rdma_user_mmap_disassociate);
 918
 919/*
 920 * ib_uverbs_open() does not need the BKL:
 921 *
 922 *  - the ib_uverbs_device structures are properly reference counted and
 923 *    everything else is purely local to the file being created, so
 924 *    races against other open calls are not a problem;
 925 *  - there is no ioctl method to race against;
 926 *  - the open method will either immediately run -ENXIO, or all
 927 *    required initialization will be done.
 928 */
 929static int ib_uverbs_open(struct inode *inode, struct file *filp)
 930{
 931	struct ib_uverbs_device *dev;
 932	struct ib_uverbs_file *file;
 933	struct ib_device *ib_dev;
 934	int ret;
 935	int module_dependent;
 936	int srcu_key;
 937
 938	dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 939	if (!refcount_inc_not_zero(&dev->refcount))
 940		return -ENXIO;
 941
 942	get_device(&dev->dev);
 943	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
 944	mutex_lock(&dev->lists_mutex);
 945	ib_dev = srcu_dereference(dev->ib_dev,
 946				  &dev->disassociate_srcu);
 947	if (!ib_dev) {
 948		ret = -EIO;
 949		goto err;
 950	}
 951
 952	if (!rdma_dev_access_netns(ib_dev, current->nsproxy->net_ns)) {
 953		ret = -EPERM;
 954		goto err;
 955	}
 956
 957	/* In case IB device supports disassociate ucontext, there is no hard
 958	 * dependency between uverbs device and its low level device.
 959	 */
 960	module_dependent = !(ib_dev->ops.disassociate_ucontext);
 961
 962	if (module_dependent) {
 963		if (!try_module_get(ib_dev->ops.owner)) {
 964			ret = -ENODEV;
 965			goto err;
 966		}
 967	}
 968
 969	file = kzalloc(sizeof(*file), GFP_KERNEL);
 970	if (!file) {
 971		ret = -ENOMEM;
 972		if (module_dependent)
 973			goto err_module;
 974
 975		goto err;
 976	}
 977
 978	file->device	 = dev;
 979	kref_init(&file->ref);
 980	mutex_init(&file->ucontext_lock);
 981
 982	spin_lock_init(&file->uobjects_lock);
 983	INIT_LIST_HEAD(&file->uobjects);
 984	init_rwsem(&file->hw_destroy_rwsem);
 985	mutex_init(&file->umap_lock);
 986	INIT_LIST_HEAD(&file->umaps);
 987
 988	mutex_init(&file->disassociation_lock);
 989
 990	filp->private_data = file;
 991	list_add_tail(&file->list, &dev->uverbs_file_list);
 992	mutex_unlock(&dev->lists_mutex);
 993	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 994
 995	setup_ufile_idr_uobject(file);
 996
 997	return stream_open(inode, filp);
 998
 999err_module:
1000	module_put(ib_dev->ops.owner);
1001
1002err:
1003	mutex_unlock(&dev->lists_mutex);
1004	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1005	if (refcount_dec_and_test(&dev->refcount))
1006		ib_uverbs_comp_dev(dev);
1007
1008	put_device(&dev->dev);
1009	return ret;
1010}
1011
1012static int ib_uverbs_close(struct inode *inode, struct file *filp)
1013{
1014	struct ib_uverbs_file *file = filp->private_data;
1015
1016	uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
1017
1018	mutex_lock(&file->device->lists_mutex);
1019	list_del_init(&file->list);
1020	mutex_unlock(&file->device->lists_mutex);
1021
1022	kref_put(&file->ref, ib_uverbs_release_file);
1023
1024	return 0;
1025}
1026
1027static const struct file_operations uverbs_fops = {
1028	.owner	 = THIS_MODULE,
1029	.write	 = ib_uverbs_write,
1030	.open	 = ib_uverbs_open,
1031	.release = ib_uverbs_close,
1032	.unlocked_ioctl = ib_uverbs_ioctl,
1033	.compat_ioctl = compat_ptr_ioctl,
1034};
1035
1036static const struct file_operations uverbs_mmap_fops = {
1037	.owner	 = THIS_MODULE,
1038	.write	 = ib_uverbs_write,
1039	.mmap    = ib_uverbs_mmap,
1040	.open	 = ib_uverbs_open,
1041	.release = ib_uverbs_close,
1042	.unlocked_ioctl = ib_uverbs_ioctl,
1043	.compat_ioctl = compat_ptr_ioctl,
1044};
1045
1046static int ib_uverbs_get_nl_info(struct ib_device *ibdev, void *client_data,
1047				 struct ib_client_nl_info *res)
1048{
1049	struct ib_uverbs_device *uverbs_dev = client_data;
1050	int ret;
1051
1052	if (res->port != -1)
1053		return -EINVAL;
1054
1055	res->abi = ibdev->ops.uverbs_abi_ver;
1056	res->cdev = &uverbs_dev->dev;
1057
1058	/*
1059	 * To support DRIVER_ID binding in userspace some of the driver need
1060	 * upgrading to expose their PCI dependent revision information
1061	 * through get_context instead of relying on modalias matching. When
1062	 * the drivers are fixed they can drop this flag.
1063	 */
1064	if (!ibdev->ops.uverbs_no_driver_id_binding) {
1065		ret = nla_put_u32(res->nl_msg, RDMA_NLDEV_ATTR_UVERBS_DRIVER_ID,
1066				  ibdev->ops.driver_id);
1067		if (ret)
1068			return ret;
1069	}
1070	return 0;
1071}
1072
1073static struct ib_client uverbs_client = {
1074	.name   = "uverbs",
1075	.no_kverbs_req = true,
1076	.add    = ib_uverbs_add_one,
1077	.remove = ib_uverbs_remove_one,
1078	.get_nl_info = ib_uverbs_get_nl_info,
1079};
1080MODULE_ALIAS_RDMA_CLIENT("uverbs");
1081
1082static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1083			  char *buf)
1084{
1085	struct ib_uverbs_device *dev =
1086			container_of(device, struct ib_uverbs_device, dev);
1087	int ret = -ENODEV;
1088	int srcu_key;
1089	struct ib_device *ib_dev;
1090
1091	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1092	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1093	if (ib_dev)
1094		ret = sysfs_emit(buf, "%s\n", dev_name(&ib_dev->dev));
1095	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1096
1097	return ret;
1098}
1099static DEVICE_ATTR_RO(ibdev);
1100
1101static ssize_t abi_version_show(struct device *device,
1102				struct device_attribute *attr, char *buf)
1103{
1104	struct ib_uverbs_device *dev =
1105			container_of(device, struct ib_uverbs_device, dev);
1106	int ret = -ENODEV;
1107	int srcu_key;
1108	struct ib_device *ib_dev;
1109
1110	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1111	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1112	if (ib_dev)
1113		ret = sysfs_emit(buf, "%u\n", ib_dev->ops.uverbs_abi_ver);
1114	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1115
1116	return ret;
1117}
1118static DEVICE_ATTR_RO(abi_version);
1119
1120static struct attribute *ib_dev_attrs[] = {
1121	&dev_attr_abi_version.attr,
1122	&dev_attr_ibdev.attr,
1123	NULL,
1124};
1125
1126static const struct attribute_group dev_attr_group = {
1127	.attrs = ib_dev_attrs,
1128};
1129
1130static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1131			 __stringify(IB_USER_VERBS_ABI_VERSION));
1132
1133static int ib_uverbs_create_uapi(struct ib_device *device,
1134				 struct ib_uverbs_device *uverbs_dev)
1135{
1136	struct uverbs_api *uapi;
1137
1138	uapi = uverbs_alloc_api(device);
1139	if (IS_ERR(uapi))
1140		return PTR_ERR(uapi);
1141
1142	uverbs_dev->uapi = uapi;
1143	return 0;
1144}
1145
1146static int ib_uverbs_add_one(struct ib_device *device)
1147{
1148	int devnum;
1149	dev_t base;
1150	struct ib_uverbs_device *uverbs_dev;
1151	int ret;
1152
1153	if (!device->ops.alloc_ucontext ||
1154	    device->type == RDMA_DEVICE_TYPE_SMI)
1155		return -EOPNOTSUPP;
1156
1157	uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1158	if (!uverbs_dev)
1159		return -ENOMEM;
1160
1161	ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1162	if (ret) {
1163		kfree(uverbs_dev);
1164		return -ENOMEM;
1165	}
1166
1167	device_initialize(&uverbs_dev->dev);
1168	uverbs_dev->dev.class = &uverbs_class;
1169	uverbs_dev->dev.parent = device->dev.parent;
1170	uverbs_dev->dev.release = ib_uverbs_release_dev;
1171	uverbs_dev->groups[0] = &dev_attr_group;
1172	uverbs_dev->dev.groups = uverbs_dev->groups;
1173	refcount_set(&uverbs_dev->refcount, 1);
1174	init_completion(&uverbs_dev->comp);
1175	uverbs_dev->xrcd_tree = RB_ROOT;
1176	mutex_init(&uverbs_dev->xrcd_tree_mutex);
1177	mutex_init(&uverbs_dev->lists_mutex);
1178	INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1179	rcu_assign_pointer(uverbs_dev->ib_dev, device);
1180	uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1181
1182	devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1183			       GFP_KERNEL);
1184	if (devnum < 0) {
1185		ret = -ENOMEM;
1186		goto err;
1187	}
1188	uverbs_dev->devnum = devnum;
1189	if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1190		base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1191	else
1192		base = IB_UVERBS_BASE_DEV + devnum;
1193
1194	ret = ib_uverbs_create_uapi(device, uverbs_dev);
1195	if (ret)
1196		goto err_uapi;
1197
1198	uverbs_dev->dev.devt = base;
1199	dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1200
1201	cdev_init(&uverbs_dev->cdev,
1202		  device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1203	uverbs_dev->cdev.owner = THIS_MODULE;
1204
1205	ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1206	if (ret)
1207		goto err_uapi;
1208
1209	ib_set_client_data(device, &uverbs_client, uverbs_dev);
1210	return 0;
1211
1212err_uapi:
1213	ida_free(&uverbs_ida, devnum);
1214err:
1215	if (refcount_dec_and_test(&uverbs_dev->refcount))
1216		ib_uverbs_comp_dev(uverbs_dev);
1217	wait_for_completion(&uverbs_dev->comp);
1218	put_device(&uverbs_dev->dev);
1219	return ret;
1220}
1221
1222static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1223					struct ib_device *ib_dev)
1224{
1225	struct ib_uverbs_file *file;
1226
1227	/* Pending running commands to terminate */
1228	uverbs_disassociate_api_pre(uverbs_dev);
1229
1230	mutex_lock(&uverbs_dev->lists_mutex);
1231	while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1232		file = list_first_entry(&uverbs_dev->uverbs_file_list,
1233					struct ib_uverbs_file, list);
1234		list_del_init(&file->list);
1235		kref_get(&file->ref);
1236
1237		/* We must release the mutex before going ahead and calling
1238		 * uverbs_cleanup_ufile, as it might end up indirectly calling
1239		 * uverbs_close, for example due to freeing the resources (e.g
1240		 * mmput).
1241		 */
1242		mutex_unlock(&uverbs_dev->lists_mutex);
1243
1244		uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1245		kref_put(&file->ref, ib_uverbs_release_file);
1246
1247		mutex_lock(&uverbs_dev->lists_mutex);
1248	}
1249	mutex_unlock(&uverbs_dev->lists_mutex);
1250
1251	uverbs_disassociate_api(uverbs_dev->uapi);
1252}
1253
1254static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1255{
1256	struct ib_uverbs_device *uverbs_dev = client_data;
1257	int wait_clients = 1;
1258
1259	cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1260	ida_free(&uverbs_ida, uverbs_dev->devnum);
1261
1262	if (device->ops.disassociate_ucontext) {
1263		/* We disassociate HW resources and immediately return.
1264		 * Userspace will see a EIO errno for all future access.
1265		 * Upon returning, ib_device may be freed internally and is not
1266		 * valid any more.
1267		 * uverbs_device is still available until all clients close
1268		 * their files, then the uverbs device ref count will be zero
1269		 * and its resources will be freed.
1270		 * Note: At this point no more files can be opened since the
1271		 * cdev was deleted, however active clients can still issue
1272		 * commands and close their open files.
1273		 */
1274		ib_uverbs_free_hw_resources(uverbs_dev, device);
1275		wait_clients = 0;
1276	}
1277
1278	if (refcount_dec_and_test(&uverbs_dev->refcount))
1279		ib_uverbs_comp_dev(uverbs_dev);
1280	if (wait_clients)
1281		wait_for_completion(&uverbs_dev->comp);
1282
1283	put_device(&uverbs_dev->dev);
1284}
1285
1286static int __init ib_uverbs_init(void)
1287{
1288	int ret;
1289
1290	ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1291				     IB_UVERBS_NUM_FIXED_MINOR,
1292				     "infiniband_verbs");
1293	if (ret) {
1294		pr_err("user_verbs: couldn't register device number\n");
1295		goto out;
1296	}
1297
1298	ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1299				  IB_UVERBS_NUM_DYNAMIC_MINOR,
1300				  "infiniband_verbs");
1301	if (ret) {
1302		pr_err("couldn't register dynamic device number\n");
1303		goto out_alloc;
1304	}
1305
1306	ret = class_register(&uverbs_class);
1307	if (ret) {
1308		pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1309		goto out_chrdev;
1310	}
1311
1312	ret = class_create_file(&uverbs_class, &class_attr_abi_version.attr);
1313	if (ret) {
1314		pr_err("user_verbs: couldn't create abi_version attribute\n");
1315		goto out_class;
1316	}
1317
1318	ret = ib_register_client(&uverbs_client);
1319	if (ret) {
1320		pr_err("user_verbs: couldn't register client\n");
1321		goto out_class;
1322	}
1323
1324	return 0;
1325
1326out_class:
1327	class_unregister(&uverbs_class);
1328
1329out_chrdev:
1330	unregister_chrdev_region(dynamic_uverbs_dev,
1331				 IB_UVERBS_NUM_DYNAMIC_MINOR);
1332
1333out_alloc:
1334	unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1335				 IB_UVERBS_NUM_FIXED_MINOR);
1336
1337out:
1338	return ret;
1339}
1340
1341static void __exit ib_uverbs_cleanup(void)
1342{
1343	ib_unregister_client(&uverbs_client);
1344	class_unregister(&uverbs_class);
1345	unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1346				 IB_UVERBS_NUM_FIXED_MINOR);
1347	unregister_chrdev_region(dynamic_uverbs_dev,
1348				 IB_UVERBS_NUM_DYNAMIC_MINOR);
1349	ib_cleanup_ucaps();
1350	mmu_notifier_synchronize();
1351}
1352
1353module_init(ib_uverbs_init);
1354module_exit(ib_uverbs_cleanup);