Linux Audio

Check our new training course

Loading...
Note: File does not exist in v6.8.
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * VFIO core
   4 *
   5 * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
   6 *     Author: Alex Williamson <alex.williamson@redhat.com>
   7 *
   8 * Derived from original vfio:
   9 * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
  10 * Author: Tom Lyon, pugs@cisco.com
  11 */
  12
  13#include <linux/cdev.h>
  14#include <linux/compat.h>
  15#include <linux/device.h>
  16#include <linux/file.h>
  17#include <linux/anon_inodes.h>
  18#include <linux/fs.h>
  19#include <linux/idr.h>
  20#include <linux/iommu.h>
  21#include <linux/list.h>
  22#include <linux/miscdevice.h>
  23#include <linux/module.h>
  24#include <linux/mutex.h>
  25#include <linux/pci.h>
  26#include <linux/rwsem.h>
  27#include <linux/sched.h>
  28#include <linux/slab.h>
  29#include <linux/stat.h>
  30#include <linux/string.h>
  31#include <linux/uaccess.h>
  32#include <linux/vfio.h>
  33#include <linux/wait.h>
  34#include <linux/sched/signal.h>
  35
  36#define DRIVER_VERSION	"0.3"
  37#define DRIVER_AUTHOR	"Alex Williamson <alex.williamson@redhat.com>"
  38#define DRIVER_DESC	"VFIO - User Level meta-driver"
  39
  40static struct vfio {
  41	struct class			*class;
  42	struct list_head		iommu_drivers_list;
  43	struct mutex			iommu_drivers_lock;
  44	struct list_head		group_list;
  45	struct idr			group_idr;
  46	struct mutex			group_lock;
  47	struct cdev			group_cdev;
  48	dev_t				group_devt;
  49	wait_queue_head_t		release_q;
  50} vfio;
  51
  52struct vfio_iommu_driver {
  53	const struct vfio_iommu_driver_ops	*ops;
  54	struct list_head			vfio_next;
  55};
  56
  57struct vfio_container {
  58	struct kref			kref;
  59	struct list_head		group_list;
  60	struct rw_semaphore		group_lock;
  61	struct vfio_iommu_driver	*iommu_driver;
  62	void				*iommu_data;
  63	bool				noiommu;
  64};
  65
  66struct vfio_unbound_dev {
  67	struct device			*dev;
  68	struct list_head		unbound_next;
  69};
  70
  71struct vfio_group {
  72	struct kref			kref;
  73	int				minor;
  74	atomic_t			container_users;
  75	struct iommu_group		*iommu_group;
  76	struct vfio_container		*container;
  77	struct list_head		device_list;
  78	struct mutex			device_lock;
  79	struct device			*dev;
  80	struct notifier_block		nb;
  81	struct list_head		vfio_next;
  82	struct list_head		container_next;
  83	struct list_head		unbound_list;
  84	struct mutex			unbound_lock;
  85	atomic_t			opened;
  86	wait_queue_head_t		container_q;
  87	bool				noiommu;
  88	unsigned int			dev_counter;
  89	struct kvm			*kvm;
  90	struct blocking_notifier_head	notifier;
  91};
  92
  93struct vfio_device {
  94	struct kref			kref;
  95	struct device			*dev;
  96	const struct vfio_device_ops	*ops;
  97	struct vfio_group		*group;
  98	struct list_head		group_next;
  99	void				*device_data;
 100};
 101
 102#ifdef CONFIG_VFIO_NOIOMMU
 103static bool noiommu __read_mostly;
 104module_param_named(enable_unsafe_noiommu_mode,
 105		   noiommu, bool, S_IRUGO | S_IWUSR);
 106MODULE_PARM_DESC(enable_unsafe_noiommu_mode, "Enable UNSAFE, no-IOMMU mode.  This mode provides no device isolation, no DMA translation, no host kernel protection, cannot be used for device assignment to virtual machines, requires RAWIO permissions, and will taint the kernel.  If you do not know what this is for, step away. (default: false)");
 107#endif
 108
 109/*
 110 * vfio_iommu_group_{get,put} are only intended for VFIO bus driver probe
 111 * and remove functions, any use cases other than acquiring the first
 112 * reference for the purpose of calling vfio_add_group_dev() or removing
 113 * that symmetric reference after vfio_del_group_dev() should use the raw
 114 * iommu_group_{get,put} functions.  In particular, vfio_iommu_group_put()
 115 * removes the device from the dummy group and cannot be nested.
 116 */
 117struct iommu_group *vfio_iommu_group_get(struct device *dev)
 118{
 119	struct iommu_group *group;
 120	int __maybe_unused ret;
 121
 122	group = iommu_group_get(dev);
 123
 124#ifdef CONFIG_VFIO_NOIOMMU
 125	/*
 126	 * With noiommu enabled, an IOMMU group will be created for a device
 127	 * that doesn't already have one and doesn't have an iommu_ops on their
 128	 * bus.  We set iommudata simply to be able to identify these groups
 129	 * as special use and for reclamation later.
 130	 */
 131	if (group || !noiommu || iommu_present(dev->bus))
 132		return group;
 133
 134	group = iommu_group_alloc();
 135	if (IS_ERR(group))
 136		return NULL;
 137
 138	iommu_group_set_name(group, "vfio-noiommu");
 139	iommu_group_set_iommudata(group, &noiommu, NULL);
 140	ret = iommu_group_add_device(group, dev);
 141	if (ret) {
 142		iommu_group_put(group);
 143		return NULL;
 144	}
 145
 146	/*
 147	 * Where to taint?  At this point we've added an IOMMU group for a
 148	 * device that is not backed by iommu_ops, therefore any iommu_
 149	 * callback using iommu_ops can legitimately Oops.  So, while we may
 150	 * be about to give a DMA capable device to a user without IOMMU
 151	 * protection, which is clearly taint-worthy, let's go ahead and do
 152	 * it here.
 153	 */
 154	add_taint(TAINT_USER, LOCKDEP_STILL_OK);
 155	dev_warn(dev, "Adding kernel taint for vfio-noiommu group on device\n");
 156#endif
 157
 158	return group;
 159}
 160EXPORT_SYMBOL_GPL(vfio_iommu_group_get);
 161
 162void vfio_iommu_group_put(struct iommu_group *group, struct device *dev)
 163{
 164#ifdef CONFIG_VFIO_NOIOMMU
 165	if (iommu_group_get_iommudata(group) == &noiommu)
 166		iommu_group_remove_device(dev);
 167#endif
 168
 169	iommu_group_put(group);
 170}
 171EXPORT_SYMBOL_GPL(vfio_iommu_group_put);
 172
 173#ifdef CONFIG_VFIO_NOIOMMU
 174static void *vfio_noiommu_open(unsigned long arg)
 175{
 176	if (arg != VFIO_NOIOMMU_IOMMU)
 177		return ERR_PTR(-EINVAL);
 178	if (!capable(CAP_SYS_RAWIO))
 179		return ERR_PTR(-EPERM);
 180
 181	return NULL;
 182}
 183
 184static void vfio_noiommu_release(void *iommu_data)
 185{
 186}
 187
 188static long vfio_noiommu_ioctl(void *iommu_data,
 189			       unsigned int cmd, unsigned long arg)
 190{
 191	if (cmd == VFIO_CHECK_EXTENSION)
 192		return noiommu && (arg == VFIO_NOIOMMU_IOMMU) ? 1 : 0;
 193
 194	return -ENOTTY;
 195}
 196
 197static int vfio_noiommu_attach_group(void *iommu_data,
 198				     struct iommu_group *iommu_group)
 199{
 200	return iommu_group_get_iommudata(iommu_group) == &noiommu ? 0 : -EINVAL;
 201}
 202
 203static void vfio_noiommu_detach_group(void *iommu_data,
 204				      struct iommu_group *iommu_group)
 205{
 206}
 207
 208static const struct vfio_iommu_driver_ops vfio_noiommu_ops = {
 209	.name = "vfio-noiommu",
 210	.owner = THIS_MODULE,
 211	.open = vfio_noiommu_open,
 212	.release = vfio_noiommu_release,
 213	.ioctl = vfio_noiommu_ioctl,
 214	.attach_group = vfio_noiommu_attach_group,
 215	.detach_group = vfio_noiommu_detach_group,
 216};
 217#endif
 218
 219
 220/**
 221 * IOMMU driver registration
 222 */
 223int vfio_register_iommu_driver(const struct vfio_iommu_driver_ops *ops)
 224{
 225	struct vfio_iommu_driver *driver, *tmp;
 226
 227	driver = kzalloc(sizeof(*driver), GFP_KERNEL);
 228	if (!driver)
 229		return -ENOMEM;
 230
 231	driver->ops = ops;
 232
 233	mutex_lock(&vfio.iommu_drivers_lock);
 234
 235	/* Check for duplicates */
 236	list_for_each_entry(tmp, &vfio.iommu_drivers_list, vfio_next) {
 237		if (tmp->ops == ops) {
 238			mutex_unlock(&vfio.iommu_drivers_lock);
 239			kfree(driver);
 240			return -EINVAL;
 241		}
 242	}
 243
 244	list_add(&driver->vfio_next, &vfio.iommu_drivers_list);
 245
 246	mutex_unlock(&vfio.iommu_drivers_lock);
 247
 248	return 0;
 249}
 250EXPORT_SYMBOL_GPL(vfio_register_iommu_driver);
 251
 252void vfio_unregister_iommu_driver(const struct vfio_iommu_driver_ops *ops)
 253{
 254	struct vfio_iommu_driver *driver;
 255
 256	mutex_lock(&vfio.iommu_drivers_lock);
 257	list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
 258		if (driver->ops == ops) {
 259			list_del(&driver->vfio_next);
 260			mutex_unlock(&vfio.iommu_drivers_lock);
 261			kfree(driver);
 262			return;
 263		}
 264	}
 265	mutex_unlock(&vfio.iommu_drivers_lock);
 266}
 267EXPORT_SYMBOL_GPL(vfio_unregister_iommu_driver);
 268
 269/**
 270 * Group minor allocation/free - both called with vfio.group_lock held
 271 */
 272static int vfio_alloc_group_minor(struct vfio_group *group)
 273{
 274	return idr_alloc(&vfio.group_idr, group, 0, MINORMASK + 1, GFP_KERNEL);
 275}
 276
 277static void vfio_free_group_minor(int minor)
 278{
 279	idr_remove(&vfio.group_idr, minor);
 280}
 281
 282static int vfio_iommu_group_notifier(struct notifier_block *nb,
 283				     unsigned long action, void *data);
 284static void vfio_group_get(struct vfio_group *group);
 285
 286/**
 287 * Container objects - containers are created when /dev/vfio/vfio is
 288 * opened, but their lifecycle extends until the last user is done, so
 289 * it's freed via kref.  Must support container/group/device being
 290 * closed in any order.
 291 */
 292static void vfio_container_get(struct vfio_container *container)
 293{
 294	kref_get(&container->kref);
 295}
 296
 297static void vfio_container_release(struct kref *kref)
 298{
 299	struct vfio_container *container;
 300	container = container_of(kref, struct vfio_container, kref);
 301
 302	kfree(container);
 303}
 304
 305static void vfio_container_put(struct vfio_container *container)
 306{
 307	kref_put(&container->kref, vfio_container_release);
 308}
 309
 310static void vfio_group_unlock_and_free(struct vfio_group *group)
 311{
 312	mutex_unlock(&vfio.group_lock);
 313	/*
 314	 * Unregister outside of lock.  A spurious callback is harmless now
 315	 * that the group is no longer in vfio.group_list.
 316	 */
 317	iommu_group_unregister_notifier(group->iommu_group, &group->nb);
 318	kfree(group);
 319}
 320
 321/**
 322 * Group objects - create, release, get, put, search
 323 */
 324static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group)
 325{
 326	struct vfio_group *group, *tmp;
 327	struct device *dev;
 328	int ret, minor;
 329
 330	group = kzalloc(sizeof(*group), GFP_KERNEL);
 331	if (!group)
 332		return ERR_PTR(-ENOMEM);
 333
 334	kref_init(&group->kref);
 335	INIT_LIST_HEAD(&group->device_list);
 336	mutex_init(&group->device_lock);
 337	INIT_LIST_HEAD(&group->unbound_list);
 338	mutex_init(&group->unbound_lock);
 339	atomic_set(&group->container_users, 0);
 340	atomic_set(&group->opened, 0);
 341	init_waitqueue_head(&group->container_q);
 342	group->iommu_group = iommu_group;
 343#ifdef CONFIG_VFIO_NOIOMMU
 344	group->noiommu = (iommu_group_get_iommudata(iommu_group) == &noiommu);
 345#endif
 346	BLOCKING_INIT_NOTIFIER_HEAD(&group->notifier);
 347
 348	group->nb.notifier_call = vfio_iommu_group_notifier;
 349
 350	/*
 351	 * blocking notifiers acquire a rwsem around registering and hold
 352	 * it around callback.  Therefore, need to register outside of
 353	 * vfio.group_lock to avoid A-B/B-A contention.  Our callback won't
 354	 * do anything unless it can find the group in vfio.group_list, so
 355	 * no harm in registering early.
 356	 */
 357	ret = iommu_group_register_notifier(iommu_group, &group->nb);
 358	if (ret) {
 359		kfree(group);
 360		return ERR_PTR(ret);
 361	}
 362
 363	mutex_lock(&vfio.group_lock);
 364
 365	/* Did we race creating this group? */
 366	list_for_each_entry(tmp, &vfio.group_list, vfio_next) {
 367		if (tmp->iommu_group == iommu_group) {
 368			vfio_group_get(tmp);
 369			vfio_group_unlock_and_free(group);
 370			return tmp;
 371		}
 372	}
 373
 374	minor = vfio_alloc_group_minor(group);
 375	if (minor < 0) {
 376		vfio_group_unlock_and_free(group);
 377		return ERR_PTR(minor);
 378	}
 379
 380	dev = device_create(vfio.class, NULL,
 381			    MKDEV(MAJOR(vfio.group_devt), minor),
 382			    group, "%s%d", group->noiommu ? "noiommu-" : "",
 383			    iommu_group_id(iommu_group));
 384	if (IS_ERR(dev)) {
 385		vfio_free_group_minor(minor);
 386		vfio_group_unlock_and_free(group);
 387		return ERR_CAST(dev);
 388	}
 389
 390	group->minor = minor;
 391	group->dev = dev;
 392
 393	list_add(&group->vfio_next, &vfio.group_list);
 394
 395	mutex_unlock(&vfio.group_lock);
 396
 397	return group;
 398}
 399
 400/* called with vfio.group_lock held */
 401static void vfio_group_release(struct kref *kref)
 402{
 403	struct vfio_group *group = container_of(kref, struct vfio_group, kref);
 404	struct vfio_unbound_dev *unbound, *tmp;
 405	struct iommu_group *iommu_group = group->iommu_group;
 406
 407	WARN_ON(!list_empty(&group->device_list));
 408	WARN_ON(group->notifier.head);
 409
 410	list_for_each_entry_safe(unbound, tmp,
 411				 &group->unbound_list, unbound_next) {
 412		list_del(&unbound->unbound_next);
 413		kfree(unbound);
 414	}
 415
 416	device_destroy(vfio.class, MKDEV(MAJOR(vfio.group_devt), group->minor));
 417	list_del(&group->vfio_next);
 418	vfio_free_group_minor(group->minor);
 419	vfio_group_unlock_and_free(group);
 420	iommu_group_put(iommu_group);
 421}
 422
 423static void vfio_group_put(struct vfio_group *group)
 424{
 425	kref_put_mutex(&group->kref, vfio_group_release, &vfio.group_lock);
 426}
 427
 428struct vfio_group_put_work {
 429	struct work_struct work;
 430	struct vfio_group *group;
 431};
 432
 433static void vfio_group_put_bg(struct work_struct *work)
 434{
 435	struct vfio_group_put_work *do_work;
 436
 437	do_work = container_of(work, struct vfio_group_put_work, work);
 438
 439	vfio_group_put(do_work->group);
 440	kfree(do_work);
 441}
 442
 443static void vfio_group_schedule_put(struct vfio_group *group)
 444{
 445	struct vfio_group_put_work *do_work;
 446
 447	do_work = kmalloc(sizeof(*do_work), GFP_KERNEL);
 448	if (WARN_ON(!do_work))
 449		return;
 450
 451	INIT_WORK(&do_work->work, vfio_group_put_bg);
 452	do_work->group = group;
 453	schedule_work(&do_work->work);
 454}
 455
 456/* Assume group_lock or group reference is held */
 457static void vfio_group_get(struct vfio_group *group)
 458{
 459	kref_get(&group->kref);
 460}
 461
 462/*
 463 * Not really a try as we will sleep for mutex, but we need to make
 464 * sure the group pointer is valid under lock and get a reference.
 465 */
 466static struct vfio_group *vfio_group_try_get(struct vfio_group *group)
 467{
 468	struct vfio_group *target = group;
 469
 470	mutex_lock(&vfio.group_lock);
 471	list_for_each_entry(group, &vfio.group_list, vfio_next) {
 472		if (group == target) {
 473			vfio_group_get(group);
 474			mutex_unlock(&vfio.group_lock);
 475			return group;
 476		}
 477	}
 478	mutex_unlock(&vfio.group_lock);
 479
 480	return NULL;
 481}
 482
 483static
 484struct vfio_group *vfio_group_get_from_iommu(struct iommu_group *iommu_group)
 485{
 486	struct vfio_group *group;
 487
 488	mutex_lock(&vfio.group_lock);
 489	list_for_each_entry(group, &vfio.group_list, vfio_next) {
 490		if (group->iommu_group == iommu_group) {
 491			vfio_group_get(group);
 492			mutex_unlock(&vfio.group_lock);
 493			return group;
 494		}
 495	}
 496	mutex_unlock(&vfio.group_lock);
 497
 498	return NULL;
 499}
 500
 501static struct vfio_group *vfio_group_get_from_minor(int minor)
 502{
 503	struct vfio_group *group;
 504
 505	mutex_lock(&vfio.group_lock);
 506	group = idr_find(&vfio.group_idr, minor);
 507	if (!group) {
 508		mutex_unlock(&vfio.group_lock);
 509		return NULL;
 510	}
 511	vfio_group_get(group);
 512	mutex_unlock(&vfio.group_lock);
 513
 514	return group;
 515}
 516
 517static struct vfio_group *vfio_group_get_from_dev(struct device *dev)
 518{
 519	struct iommu_group *iommu_group;
 520	struct vfio_group *group;
 521
 522	iommu_group = iommu_group_get(dev);
 523	if (!iommu_group)
 524		return NULL;
 525
 526	group = vfio_group_get_from_iommu(iommu_group);
 527	iommu_group_put(iommu_group);
 528
 529	return group;
 530}
 531
 532/**
 533 * Device objects - create, release, get, put, search
 534 */
 535static
 536struct vfio_device *vfio_group_create_device(struct vfio_group *group,
 537					     struct device *dev,
 538					     const struct vfio_device_ops *ops,
 539					     void *device_data)
 540{
 541	struct vfio_device *device;
 542
 543	device = kzalloc(sizeof(*device), GFP_KERNEL);
 544	if (!device)
 545		return ERR_PTR(-ENOMEM);
 546
 547	kref_init(&device->kref);
 548	device->dev = dev;
 549	device->group = group;
 550	device->ops = ops;
 551	device->device_data = device_data;
 552	dev_set_drvdata(dev, device);
 553
 554	/* No need to get group_lock, caller has group reference */
 555	vfio_group_get(group);
 556
 557	mutex_lock(&group->device_lock);
 558	list_add(&device->group_next, &group->device_list);
 559	group->dev_counter++;
 560	mutex_unlock(&group->device_lock);
 561
 562	return device;
 563}
 564
 565static void vfio_device_release(struct kref *kref)
 566{
 567	struct vfio_device *device = container_of(kref,
 568						  struct vfio_device, kref);
 569	struct vfio_group *group = device->group;
 570
 571	list_del(&device->group_next);
 572	group->dev_counter--;
 573	mutex_unlock(&group->device_lock);
 574
 575	dev_set_drvdata(device->dev, NULL);
 576
 577	kfree(device);
 578
 579	/* vfio_del_group_dev may be waiting for this device */
 580	wake_up(&vfio.release_q);
 581}
 582
 583/* Device reference always implies a group reference */
 584void vfio_device_put(struct vfio_device *device)
 585{
 586	struct vfio_group *group = device->group;
 587	kref_put_mutex(&device->kref, vfio_device_release, &group->device_lock);
 588	vfio_group_put(group);
 589}
 590EXPORT_SYMBOL_GPL(vfio_device_put);
 591
 592static void vfio_device_get(struct vfio_device *device)
 593{
 594	vfio_group_get(device->group);
 595	kref_get(&device->kref);
 596}
 597
 598static struct vfio_device *vfio_group_get_device(struct vfio_group *group,
 599						 struct device *dev)
 600{
 601	struct vfio_device *device;
 602
 603	mutex_lock(&group->device_lock);
 604	list_for_each_entry(device, &group->device_list, group_next) {
 605		if (device->dev == dev) {
 606			vfio_device_get(device);
 607			mutex_unlock(&group->device_lock);
 608			return device;
 609		}
 610	}
 611	mutex_unlock(&group->device_lock);
 612	return NULL;
 613}
 614
 615/*
 616 * Some drivers, like pci-stub, are only used to prevent other drivers from
 617 * claiming a device and are therefore perfectly legitimate for a user owned
 618 * group.  The pci-stub driver has no dependencies on DMA or the IOVA mapping
 619 * of the device, but it does prevent the user from having direct access to
 620 * the device, which is useful in some circumstances.
 621 *
 622 * We also assume that we can include PCI interconnect devices, ie. bridges.
 623 * IOMMU grouping on PCI necessitates that if we lack isolation on a bridge
 624 * then all of the downstream devices will be part of the same IOMMU group as
 625 * the bridge.  Thus, if placing the bridge into the user owned IOVA space
 626 * breaks anything, it only does so for user owned devices downstream.  Note
 627 * that error notification via MSI can be affected for platforms that handle
 628 * MSI within the same IOVA space as DMA.
 629 */
 630static const char * const vfio_driver_allowed[] = { "pci-stub" };
 631
 632static bool vfio_dev_driver_allowed(struct device *dev,
 633				    struct device_driver *drv)
 634{
 635	if (dev_is_pci(dev)) {
 636		struct pci_dev *pdev = to_pci_dev(dev);
 637
 638		if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
 639			return true;
 640	}
 641
 642	return match_string(vfio_driver_allowed,
 643			    ARRAY_SIZE(vfio_driver_allowed),
 644			    drv->name) >= 0;
 645}
 646
 647/*
 648 * A vfio group is viable for use by userspace if all devices are in
 649 * one of the following states:
 650 *  - driver-less
 651 *  - bound to a vfio driver
 652 *  - bound to an otherwise allowed driver
 653 *  - a PCI interconnect device
 654 *
 655 * We use two methods to determine whether a device is bound to a vfio
 656 * driver.  The first is to test whether the device exists in the vfio
 657 * group.  The second is to test if the device exists on the group
 658 * unbound_list, indicating it's in the middle of transitioning from
 659 * a vfio driver to driver-less.
 660 */
 661static int vfio_dev_viable(struct device *dev, void *data)
 662{
 663	struct vfio_group *group = data;
 664	struct vfio_device *device;
 665	struct device_driver *drv = READ_ONCE(dev->driver);
 666	struct vfio_unbound_dev *unbound;
 667	int ret = -EINVAL;
 668
 669	mutex_lock(&group->unbound_lock);
 670	list_for_each_entry(unbound, &group->unbound_list, unbound_next) {
 671		if (dev == unbound->dev) {
 672			ret = 0;
 673			break;
 674		}
 675	}
 676	mutex_unlock(&group->unbound_lock);
 677
 678	if (!ret || !drv || vfio_dev_driver_allowed(dev, drv))
 679		return 0;
 680
 681	device = vfio_group_get_device(group, dev);
 682	if (device) {
 683		vfio_device_put(device);
 684		return 0;
 685	}
 686
 687	return ret;
 688}
 689
 690/**
 691 * Async device support
 692 */
 693static int vfio_group_nb_add_dev(struct vfio_group *group, struct device *dev)
 694{
 695	struct vfio_device *device;
 696
 697	/* Do we already know about it?  We shouldn't */
 698	device = vfio_group_get_device(group, dev);
 699	if (WARN_ON_ONCE(device)) {
 700		vfio_device_put(device);
 701		return 0;
 702	}
 703
 704	/* Nothing to do for idle groups */
 705	if (!atomic_read(&group->container_users))
 706		return 0;
 707
 708	/* TODO Prevent device auto probing */
 709	dev_WARN(dev, "Device added to live group %d!\n",
 710		 iommu_group_id(group->iommu_group));
 711
 712	return 0;
 713}
 714
 715static int vfio_group_nb_verify(struct vfio_group *group, struct device *dev)
 716{
 717	/* We don't care what happens when the group isn't in use */
 718	if (!atomic_read(&group->container_users))
 719		return 0;
 720
 721	return vfio_dev_viable(dev, group);
 722}
 723
 724static int vfio_iommu_group_notifier(struct notifier_block *nb,
 725				     unsigned long action, void *data)
 726{
 727	struct vfio_group *group = container_of(nb, struct vfio_group, nb);
 728	struct device *dev = data;
 729	struct vfio_unbound_dev *unbound;
 730
 731	/*
 732	 * Need to go through a group_lock lookup to get a reference or we
 733	 * risk racing a group being removed.  Ignore spurious notifies.
 734	 */
 735	group = vfio_group_try_get(group);
 736	if (!group)
 737		return NOTIFY_OK;
 738
 739	switch (action) {
 740	case IOMMU_GROUP_NOTIFY_ADD_DEVICE:
 741		vfio_group_nb_add_dev(group, dev);
 742		break;
 743	case IOMMU_GROUP_NOTIFY_DEL_DEVICE:
 744		/*
 745		 * Nothing to do here.  If the device is in use, then the
 746		 * vfio sub-driver should block the remove callback until
 747		 * it is unused.  If the device is unused or attached to a
 748		 * stub driver, then it should be released and we don't
 749		 * care that it will be going away.
 750		 */
 751		break;
 752	case IOMMU_GROUP_NOTIFY_BIND_DRIVER:
 753		dev_dbg(dev, "%s: group %d binding to driver\n", __func__,
 754			iommu_group_id(group->iommu_group));
 755		break;
 756	case IOMMU_GROUP_NOTIFY_BOUND_DRIVER:
 757		dev_dbg(dev, "%s: group %d bound to driver %s\n", __func__,
 758			iommu_group_id(group->iommu_group), dev->driver->name);
 759		BUG_ON(vfio_group_nb_verify(group, dev));
 760		break;
 761	case IOMMU_GROUP_NOTIFY_UNBIND_DRIVER:
 762		dev_dbg(dev, "%s: group %d unbinding from driver %s\n",
 763			__func__, iommu_group_id(group->iommu_group),
 764			dev->driver->name);
 765		break;
 766	case IOMMU_GROUP_NOTIFY_UNBOUND_DRIVER:
 767		dev_dbg(dev, "%s: group %d unbound from driver\n", __func__,
 768			iommu_group_id(group->iommu_group));
 769		/*
 770		 * XXX An unbound device in a live group is ok, but we'd
 771		 * really like to avoid the above BUG_ON by preventing other
 772		 * drivers from binding to it.  Once that occurs, we have to
 773		 * stop the system to maintain isolation.  At a minimum, we'd
 774		 * want a toggle to disable driver auto probe for this device.
 775		 */
 776
 777		mutex_lock(&group->unbound_lock);
 778		list_for_each_entry(unbound,
 779				    &group->unbound_list, unbound_next) {
 780			if (dev == unbound->dev) {
 781				list_del(&unbound->unbound_next);
 782				kfree(unbound);
 783				break;
 784			}
 785		}
 786		mutex_unlock(&group->unbound_lock);
 787		break;
 788	}
 789
 790	/*
 791	 * If we're the last reference to the group, the group will be
 792	 * released, which includes unregistering the iommu group notifier.
 793	 * We hold a read-lock on that notifier list, unregistering needs
 794	 * a write-lock... deadlock.  Release our reference asynchronously
 795	 * to avoid that situation.
 796	 */
 797	vfio_group_schedule_put(group);
 798	return NOTIFY_OK;
 799}
 800
 801/**
 802 * VFIO driver API
 803 */
 804int vfio_add_group_dev(struct device *dev,
 805		       const struct vfio_device_ops *ops, void *device_data)
 806{
 807	struct iommu_group *iommu_group;
 808	struct vfio_group *group;
 809	struct vfio_device *device;
 810
 811	iommu_group = iommu_group_get(dev);
 812	if (!iommu_group)
 813		return -EINVAL;
 814
 815	group = vfio_group_get_from_iommu(iommu_group);
 816	if (!group) {
 817		group = vfio_create_group(iommu_group);
 818		if (IS_ERR(group)) {
 819			iommu_group_put(iommu_group);
 820			return PTR_ERR(group);
 821		}
 822	} else {
 823		/*
 824		 * A found vfio_group already holds a reference to the
 825		 * iommu_group.  A created vfio_group keeps the reference.
 826		 */
 827		iommu_group_put(iommu_group);
 828	}
 829
 830	device = vfio_group_get_device(group, dev);
 831	if (device) {
 832		dev_WARN(dev, "Device already exists on group %d\n",
 833			 iommu_group_id(iommu_group));
 834		vfio_device_put(device);
 835		vfio_group_put(group);
 836		return -EBUSY;
 837	}
 838
 839	device = vfio_group_create_device(group, dev, ops, device_data);
 840	if (IS_ERR(device)) {
 841		vfio_group_put(group);
 842		return PTR_ERR(device);
 843	}
 844
 845	/*
 846	 * Drop all but the vfio_device reference.  The vfio_device holds
 847	 * a reference to the vfio_group, which holds a reference to the
 848	 * iommu_group.
 849	 */
 850	vfio_group_put(group);
 851
 852	return 0;
 853}
 854EXPORT_SYMBOL_GPL(vfio_add_group_dev);
 855
 856/**
 857 * Get a reference to the vfio_device for a device.  Even if the
 858 * caller thinks they own the device, they could be racing with a
 859 * release call path, so we can't trust drvdata for the shortcut.
 860 * Go the long way around, from the iommu_group to the vfio_group
 861 * to the vfio_device.
 862 */
 863struct vfio_device *vfio_device_get_from_dev(struct device *dev)
 864{
 865	struct vfio_group *group;
 866	struct vfio_device *device;
 867
 868	group = vfio_group_get_from_dev(dev);
 869	if (!group)
 870		return NULL;
 871
 872	device = vfio_group_get_device(group, dev);
 873	vfio_group_put(group);
 874
 875	return device;
 876}
 877EXPORT_SYMBOL_GPL(vfio_device_get_from_dev);
 878
 879static struct vfio_device *vfio_device_get_from_name(struct vfio_group *group,
 880						     char *buf)
 881{
 882	struct vfio_device *it, *device = ERR_PTR(-ENODEV);
 883
 884	mutex_lock(&group->device_lock);
 885	list_for_each_entry(it, &group->device_list, group_next) {
 886		int ret;
 887
 888		if (it->ops->match) {
 889			ret = it->ops->match(it->device_data, buf);
 890			if (ret < 0) {
 891				device = ERR_PTR(ret);
 892				break;
 893			}
 894		} else {
 895			ret = !strcmp(dev_name(it->dev), buf);
 896		}
 897
 898		if (ret) {
 899			device = it;
 900			vfio_device_get(device);
 901			break;
 902		}
 903	}
 904	mutex_unlock(&group->device_lock);
 905
 906	return device;
 907}
 908
 909/*
 910 * Caller must hold a reference to the vfio_device
 911 */
 912void *vfio_device_data(struct vfio_device *device)
 913{
 914	return device->device_data;
 915}
 916EXPORT_SYMBOL_GPL(vfio_device_data);
 917
 918/*
 919 * Decrement the device reference count and wait for the device to be
 920 * removed.  Open file descriptors for the device... */
 921void *vfio_del_group_dev(struct device *dev)
 922{
 923	DEFINE_WAIT_FUNC(wait, woken_wake_function);
 924	struct vfio_device *device = dev_get_drvdata(dev);
 925	struct vfio_group *group = device->group;
 926	void *device_data = device->device_data;
 927	struct vfio_unbound_dev *unbound;
 928	unsigned int i = 0;
 929	bool interrupted = false;
 930
 931	/*
 932	 * The group exists so long as we have a device reference.  Get
 933	 * a group reference and use it to scan for the device going away.
 934	 */
 935	vfio_group_get(group);
 936
 937	/*
 938	 * When the device is removed from the group, the group suddenly
 939	 * becomes non-viable; the device has a driver (until the unbind
 940	 * completes), but it's not present in the group.  This is bad news
 941	 * for any external users that need to re-acquire a group reference
 942	 * in order to match and release their existing reference.  To
 943	 * solve this, we track such devices on the unbound_list to bridge
 944	 * the gap until they're fully unbound.
 945	 */
 946	unbound = kzalloc(sizeof(*unbound), GFP_KERNEL);
 947	if (unbound) {
 948		unbound->dev = dev;
 949		mutex_lock(&group->unbound_lock);
 950		list_add(&unbound->unbound_next, &group->unbound_list);
 951		mutex_unlock(&group->unbound_lock);
 952	}
 953	WARN_ON(!unbound);
 954
 955	vfio_device_put(device);
 956
 957	/*
 958	 * If the device is still present in the group after the above
 959	 * 'put', then it is in use and we need to request it from the
 960	 * bus driver.  The driver may in turn need to request the
 961	 * device from the user.  We send the request on an arbitrary
 962	 * interval with counter to allow the driver to take escalating
 963	 * measures to release the device if it has the ability to do so.
 964	 */
 965	add_wait_queue(&vfio.release_q, &wait);
 966
 967	do {
 968		device = vfio_group_get_device(group, dev);
 969		if (!device)
 970			break;
 971
 972		if (device->ops->request)
 973			device->ops->request(device_data, i++);
 974
 975		vfio_device_put(device);
 976
 977		if (interrupted) {
 978			wait_woken(&wait, TASK_UNINTERRUPTIBLE, HZ * 10);
 979		} else {
 980			wait_woken(&wait, TASK_INTERRUPTIBLE, HZ * 10);
 981			if (signal_pending(current)) {
 982				interrupted = true;
 983				dev_warn(dev,
 984					 "Device is currently in use, task"
 985					 " \"%s\" (%d) "
 986					 "blocked until device is released",
 987					 current->comm, task_pid_nr(current));
 988			}
 989		}
 990
 991	} while (1);
 992
 993	remove_wait_queue(&vfio.release_q, &wait);
 994	/*
 995	 * In order to support multiple devices per group, devices can be
 996	 * plucked from the group while other devices in the group are still
 997	 * in use.  The container persists with this group and those remaining
 998	 * devices still attached.  If the user creates an isolation violation
 999	 * by binding this device to another driver while the group is still in
1000	 * use, that's their fault.  However, in the case of removing the last,
1001	 * or potentially the only, device in the group there can be no other
1002	 * in-use devices in the group.  The user has done their due diligence
1003	 * and we should lay no claims to those devices.  In order to do that,
1004	 * we need to make sure the group is detached from the container.
1005	 * Without this stall, we're potentially racing with a user process
1006	 * that may attempt to immediately bind this device to another driver.
1007	 */
1008	if (list_empty(&group->device_list))
1009		wait_event(group->container_q, !group->container);
1010
1011	vfio_group_put(group);
1012
1013	return device_data;
1014}
1015EXPORT_SYMBOL_GPL(vfio_del_group_dev);
1016
1017/**
1018 * VFIO base fd, /dev/vfio/vfio
1019 */
1020static long vfio_ioctl_check_extension(struct vfio_container *container,
1021				       unsigned long arg)
1022{
1023	struct vfio_iommu_driver *driver;
1024	long ret = 0;
1025
1026	down_read(&container->group_lock);
1027
1028	driver = container->iommu_driver;
1029
1030	switch (arg) {
1031		/* No base extensions yet */
1032	default:
1033		/*
1034		 * If no driver is set, poll all registered drivers for
1035		 * extensions and return the first positive result.  If
1036		 * a driver is already set, further queries will be passed
1037		 * only to that driver.
1038		 */
1039		if (!driver) {
1040			mutex_lock(&vfio.iommu_drivers_lock);
1041			list_for_each_entry(driver, &vfio.iommu_drivers_list,
1042					    vfio_next) {
1043
1044#ifdef CONFIG_VFIO_NOIOMMU
1045				if (!list_empty(&container->group_list) &&
1046				    (container->noiommu !=
1047				     (driver->ops == &vfio_noiommu_ops)))
1048					continue;
1049#endif
1050
1051				if (!try_module_get(driver->ops->owner))
1052					continue;
1053
1054				ret = driver->ops->ioctl(NULL,
1055							 VFIO_CHECK_EXTENSION,
1056							 arg);
1057				module_put(driver->ops->owner);
1058				if (ret > 0)
1059					break;
1060			}
1061			mutex_unlock(&vfio.iommu_drivers_lock);
1062		} else
1063			ret = driver->ops->ioctl(container->iommu_data,
1064						 VFIO_CHECK_EXTENSION, arg);
1065	}
1066
1067	up_read(&container->group_lock);
1068
1069	return ret;
1070}
1071
1072/* hold write lock on container->group_lock */
1073static int __vfio_container_attach_groups(struct vfio_container *container,
1074					  struct vfio_iommu_driver *driver,
1075					  void *data)
1076{
1077	struct vfio_group *group;
1078	int ret = -ENODEV;
1079
1080	list_for_each_entry(group, &container->group_list, container_next) {
1081		ret = driver->ops->attach_group(data, group->iommu_group);
1082		if (ret)
1083			goto unwind;
1084	}
1085
1086	return ret;
1087
1088unwind:
1089	list_for_each_entry_continue_reverse(group, &container->group_list,
1090					     container_next) {
1091		driver->ops->detach_group(data, group->iommu_group);
1092	}
1093
1094	return ret;
1095}
1096
1097static long vfio_ioctl_set_iommu(struct vfio_container *container,
1098				 unsigned long arg)
1099{
1100	struct vfio_iommu_driver *driver;
1101	long ret = -ENODEV;
1102
1103	down_write(&container->group_lock);
1104
1105	/*
1106	 * The container is designed to be an unprivileged interface while
1107	 * the group can be assigned to specific users.  Therefore, only by
1108	 * adding a group to a container does the user get the privilege of
1109	 * enabling the iommu, which may allocate finite resources.  There
1110	 * is no unset_iommu, but by removing all the groups from a container,
1111	 * the container is deprivileged and returns to an unset state.
1112	 */
1113	if (list_empty(&container->group_list) || container->iommu_driver) {
1114		up_write(&container->group_lock);
1115		return -EINVAL;
1116	}
1117
1118	mutex_lock(&vfio.iommu_drivers_lock);
1119	list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
1120		void *data;
1121
1122#ifdef CONFIG_VFIO_NOIOMMU
1123		/*
1124		 * Only noiommu containers can use vfio-noiommu and noiommu
1125		 * containers can only use vfio-noiommu.
1126		 */
1127		if (container->noiommu != (driver->ops == &vfio_noiommu_ops))
1128			continue;
1129#endif
1130
1131		if (!try_module_get(driver->ops->owner))
1132			continue;
1133
1134		/*
1135		 * The arg magic for SET_IOMMU is the same as CHECK_EXTENSION,
1136		 * so test which iommu driver reported support for this
1137		 * extension and call open on them.  We also pass them the
1138		 * magic, allowing a single driver to support multiple
1139		 * interfaces if they'd like.
1140		 */
1141		if (driver->ops->ioctl(NULL, VFIO_CHECK_EXTENSION, arg) <= 0) {
1142			module_put(driver->ops->owner);
1143			continue;
1144		}
1145
1146		data = driver->ops->open(arg);
1147		if (IS_ERR(data)) {
1148			ret = PTR_ERR(data);
1149			module_put(driver->ops->owner);
1150			continue;
1151		}
1152
1153		ret = __vfio_container_attach_groups(container, driver, data);
1154		if (ret) {
1155			driver->ops->release(data);
1156			module_put(driver->ops->owner);
1157			continue;
1158		}
1159
1160		container->iommu_driver = driver;
1161		container->iommu_data = data;
1162		break;
1163	}
1164
1165	mutex_unlock(&vfio.iommu_drivers_lock);
1166	up_write(&container->group_lock);
1167
1168	return ret;
1169}
1170
1171static long vfio_fops_unl_ioctl(struct file *filep,
1172				unsigned int cmd, unsigned long arg)
1173{
1174	struct vfio_container *container = filep->private_data;
1175	struct vfio_iommu_driver *driver;
1176	void *data;
1177	long ret = -EINVAL;
1178
1179	if (!container)
1180		return ret;
1181
1182	switch (cmd) {
1183	case VFIO_GET_API_VERSION:
1184		ret = VFIO_API_VERSION;
1185		break;
1186	case VFIO_CHECK_EXTENSION:
1187		ret = vfio_ioctl_check_extension(container, arg);
1188		break;
1189	case VFIO_SET_IOMMU:
1190		ret = vfio_ioctl_set_iommu(container, arg);
1191		break;
1192	default:
1193		driver = container->iommu_driver;
1194		data = container->iommu_data;
1195
1196		if (driver) /* passthrough all unrecognized ioctls */
1197			ret = driver->ops->ioctl(data, cmd, arg);
1198	}
1199
1200	return ret;
1201}
1202
1203static int vfio_fops_open(struct inode *inode, struct file *filep)
1204{
1205	struct vfio_container *container;
1206
1207	container = kzalloc(sizeof(*container), GFP_KERNEL);
1208	if (!container)
1209		return -ENOMEM;
1210
1211	INIT_LIST_HEAD(&container->group_list);
1212	init_rwsem(&container->group_lock);
1213	kref_init(&container->kref);
1214
1215	filep->private_data = container;
1216
1217	return 0;
1218}
1219
1220static int vfio_fops_release(struct inode *inode, struct file *filep)
1221{
1222	struct vfio_container *container = filep->private_data;
1223
1224	filep->private_data = NULL;
1225
1226	vfio_container_put(container);
1227
1228	return 0;
1229}
1230
1231/*
1232 * Once an iommu driver is set, we optionally pass read/write/mmap
1233 * on to the driver, allowing management interfaces beyond ioctl.
1234 */
1235static ssize_t vfio_fops_read(struct file *filep, char __user *buf,
1236			      size_t count, loff_t *ppos)
1237{
1238	struct vfio_container *container = filep->private_data;
1239	struct vfio_iommu_driver *driver;
1240	ssize_t ret = -EINVAL;
1241
1242	driver = container->iommu_driver;
1243	if (likely(driver && driver->ops->read))
1244		ret = driver->ops->read(container->iommu_data,
1245					buf, count, ppos);
1246
1247	return ret;
1248}
1249
1250static ssize_t vfio_fops_write(struct file *filep, const char __user *buf,
1251			       size_t count, loff_t *ppos)
1252{
1253	struct vfio_container *container = filep->private_data;
1254	struct vfio_iommu_driver *driver;
1255	ssize_t ret = -EINVAL;
1256
1257	driver = container->iommu_driver;
1258	if (likely(driver && driver->ops->write))
1259		ret = driver->ops->write(container->iommu_data,
1260					 buf, count, ppos);
1261
1262	return ret;
1263}
1264
1265static int vfio_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1266{
1267	struct vfio_container *container = filep->private_data;
1268	struct vfio_iommu_driver *driver;
1269	int ret = -EINVAL;
1270
1271	driver = container->iommu_driver;
1272	if (likely(driver && driver->ops->mmap))
1273		ret = driver->ops->mmap(container->iommu_data, vma);
1274
1275	return ret;
1276}
1277
1278static const struct file_operations vfio_fops = {
1279	.owner		= THIS_MODULE,
1280	.open		= vfio_fops_open,
1281	.release	= vfio_fops_release,
1282	.read		= vfio_fops_read,
1283	.write		= vfio_fops_write,
1284	.unlocked_ioctl	= vfio_fops_unl_ioctl,
1285	.compat_ioctl	= compat_ptr_ioctl,
1286	.mmap		= vfio_fops_mmap,
1287};
1288
1289/**
1290 * VFIO Group fd, /dev/vfio/$GROUP
1291 */
1292static void __vfio_group_unset_container(struct vfio_group *group)
1293{
1294	struct vfio_container *container = group->container;
1295	struct vfio_iommu_driver *driver;
1296
1297	down_write(&container->group_lock);
1298
1299	driver = container->iommu_driver;
1300	if (driver)
1301		driver->ops->detach_group(container->iommu_data,
1302					  group->iommu_group);
1303
1304	group->container = NULL;
1305	wake_up(&group->container_q);
1306	list_del(&group->container_next);
1307
1308	/* Detaching the last group deprivileges a container, remove iommu */
1309	if (driver && list_empty(&container->group_list)) {
1310		driver->ops->release(container->iommu_data);
1311		module_put(driver->ops->owner);
1312		container->iommu_driver = NULL;
1313		container->iommu_data = NULL;
1314	}
1315
1316	up_write(&container->group_lock);
1317
1318	vfio_container_put(container);
1319}
1320
1321/*
1322 * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or
1323 * if there was no container to unset.  Since the ioctl is called on
1324 * the group, we know that still exists, therefore the only valid
1325 * transition here is 1->0.
1326 */
1327static int vfio_group_unset_container(struct vfio_group *group)
1328{
1329	int users = atomic_cmpxchg(&group->container_users, 1, 0);
1330
1331	if (!users)
1332		return -EINVAL;
1333	if (users != 1)
1334		return -EBUSY;
1335
1336	__vfio_group_unset_container(group);
1337
1338	return 0;
1339}
1340
1341/*
1342 * When removing container users, anything that removes the last user
1343 * implicitly removes the group from the container.  That is, if the
1344 * group file descriptor is closed, as well as any device file descriptors,
1345 * the group is free.
1346 */
1347static void vfio_group_try_dissolve_container(struct vfio_group *group)
1348{
1349	if (0 == atomic_dec_if_positive(&group->container_users))
1350		__vfio_group_unset_container(group);
1351}
1352
1353static int vfio_group_set_container(struct vfio_group *group, int container_fd)
1354{
1355	struct fd f;
1356	struct vfio_container *container;
1357	struct vfio_iommu_driver *driver;
1358	int ret = 0;
1359
1360	if (atomic_read(&group->container_users))
1361		return -EINVAL;
1362
1363	if (group->noiommu && !capable(CAP_SYS_RAWIO))
1364		return -EPERM;
1365
1366	f = fdget(container_fd);
1367	if (!f.file)
1368		return -EBADF;
1369
1370	/* Sanity check, is this really our fd? */
1371	if (f.file->f_op != &vfio_fops) {
1372		fdput(f);
1373		return -EINVAL;
1374	}
1375
1376	container = f.file->private_data;
1377	WARN_ON(!container); /* fget ensures we don't race vfio_release */
1378
1379	down_write(&container->group_lock);
1380
1381	/* Real groups and fake groups cannot mix */
1382	if (!list_empty(&container->group_list) &&
1383	    container->noiommu != group->noiommu) {
1384		ret = -EPERM;
1385		goto unlock_out;
1386	}
1387
1388	driver = container->iommu_driver;
1389	if (driver) {
1390		ret = driver->ops->attach_group(container->iommu_data,
1391						group->iommu_group);
1392		if (ret)
1393			goto unlock_out;
1394	}
1395
1396	group->container = container;
1397	container->noiommu = group->noiommu;
1398	list_add(&group->container_next, &container->group_list);
1399
1400	/* Get a reference on the container and mark a user within the group */
1401	vfio_container_get(container);
1402	atomic_inc(&group->container_users);
1403
1404unlock_out:
1405	up_write(&container->group_lock);
1406	fdput(f);
1407	return ret;
1408}
1409
1410static bool vfio_group_viable(struct vfio_group *group)
1411{
1412	return (iommu_group_for_each_dev(group->iommu_group,
1413					 group, vfio_dev_viable) == 0);
1414}
1415
1416static int vfio_group_add_container_user(struct vfio_group *group)
1417{
1418	if (!atomic_inc_not_zero(&group->container_users))
1419		return -EINVAL;
1420
1421	if (group->noiommu) {
1422		atomic_dec(&group->container_users);
1423		return -EPERM;
1424	}
1425	if (!group->container->iommu_driver || !vfio_group_viable(group)) {
1426		atomic_dec(&group->container_users);
1427		return -EINVAL;
1428	}
1429
1430	return 0;
1431}
1432
1433static const struct file_operations vfio_device_fops;
1434
1435static int vfio_group_get_device_fd(struct vfio_group *group, char *buf)
1436{
1437	struct vfio_device *device;
1438	struct file *filep;
1439	int ret;
1440
1441	if (0 == atomic_read(&group->container_users) ||
1442	    !group->container->iommu_driver || !vfio_group_viable(group))
1443		return -EINVAL;
1444
1445	if (group->noiommu && !capable(CAP_SYS_RAWIO))
1446		return -EPERM;
1447
1448	device = vfio_device_get_from_name(group, buf);
1449	if (IS_ERR(device))
1450		return PTR_ERR(device);
1451
1452	ret = device->ops->open(device->device_data);
1453	if (ret) {
1454		vfio_device_put(device);
1455		return ret;
1456	}
1457
1458	/*
1459	 * We can't use anon_inode_getfd() because we need to modify
1460	 * the f_mode flags directly to allow more than just ioctls
1461	 */
1462	ret = get_unused_fd_flags(O_CLOEXEC);
1463	if (ret < 0) {
1464		device->ops->release(device->device_data);
1465		vfio_device_put(device);
1466		return ret;
1467	}
1468
1469	filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
1470				   device, O_RDWR);
1471	if (IS_ERR(filep)) {
1472		put_unused_fd(ret);
1473		ret = PTR_ERR(filep);
1474		device->ops->release(device->device_data);
1475		vfio_device_put(device);
1476		return ret;
1477	}
1478
1479	/*
1480	 * TODO: add an anon_inode interface to do this.
1481	 * Appears to be missing by lack of need rather than
1482	 * explicitly prevented.  Now there's need.
1483	 */
1484	filep->f_mode |= (FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE);
1485
1486	atomic_inc(&group->container_users);
1487
1488	fd_install(ret, filep);
1489
1490	if (group->noiommu)
1491		dev_warn(device->dev, "vfio-noiommu device opened by user "
1492			 "(%s:%d)\n", current->comm, task_pid_nr(current));
1493
1494	return ret;
1495}
1496
1497static long vfio_group_fops_unl_ioctl(struct file *filep,
1498				      unsigned int cmd, unsigned long arg)
1499{
1500	struct vfio_group *group = filep->private_data;
1501	long ret = -ENOTTY;
1502
1503	switch (cmd) {
1504	case VFIO_GROUP_GET_STATUS:
1505	{
1506		struct vfio_group_status status;
1507		unsigned long minsz;
1508
1509		minsz = offsetofend(struct vfio_group_status, flags);
1510
1511		if (copy_from_user(&status, (void __user *)arg, minsz))
1512			return -EFAULT;
1513
1514		if (status.argsz < minsz)
1515			return -EINVAL;
1516
1517		status.flags = 0;
1518
1519		if (vfio_group_viable(group))
1520			status.flags |= VFIO_GROUP_FLAGS_VIABLE;
1521
1522		if (group->container)
1523			status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET;
1524
1525		if (copy_to_user((void __user *)arg, &status, minsz))
1526			return -EFAULT;
1527
1528		ret = 0;
1529		break;
1530	}
1531	case VFIO_GROUP_SET_CONTAINER:
1532	{
1533		int fd;
1534
1535		if (get_user(fd, (int __user *)arg))
1536			return -EFAULT;
1537
1538		if (fd < 0)
1539			return -EINVAL;
1540
1541		ret = vfio_group_set_container(group, fd);
1542		break;
1543	}
1544	case VFIO_GROUP_UNSET_CONTAINER:
1545		ret = vfio_group_unset_container(group);
1546		break;
1547	case VFIO_GROUP_GET_DEVICE_FD:
1548	{
1549		char *buf;
1550
1551		buf = strndup_user((const char __user *)arg, PAGE_SIZE);
1552		if (IS_ERR(buf))
1553			return PTR_ERR(buf);
1554
1555		ret = vfio_group_get_device_fd(group, buf);
1556		kfree(buf);
1557		break;
1558	}
1559	}
1560
1561	return ret;
1562}
1563
1564static int vfio_group_fops_open(struct inode *inode, struct file *filep)
1565{
1566	struct vfio_group *group;
1567	int opened;
1568
1569	group = vfio_group_get_from_minor(iminor(inode));
1570	if (!group)
1571		return -ENODEV;
1572
1573	if (group->noiommu && !capable(CAP_SYS_RAWIO)) {
1574		vfio_group_put(group);
1575		return -EPERM;
1576	}
1577
1578	/* Do we need multiple instances of the group open?  Seems not. */
1579	opened = atomic_cmpxchg(&group->opened, 0, 1);
1580	if (opened) {
1581		vfio_group_put(group);
1582		return -EBUSY;
1583	}
1584
1585	/* Is something still in use from a previous open? */
1586	if (group->container) {
1587		atomic_dec(&group->opened);
1588		vfio_group_put(group);
1589		return -EBUSY;
1590	}
1591
1592	/* Warn if previous user didn't cleanup and re-init to drop them */
1593	if (WARN_ON(group->notifier.head))
1594		BLOCKING_INIT_NOTIFIER_HEAD(&group->notifier);
1595
1596	filep->private_data = group;
1597
1598	return 0;
1599}
1600
1601static int vfio_group_fops_release(struct inode *inode, struct file *filep)
1602{
1603	struct vfio_group *group = filep->private_data;
1604
1605	filep->private_data = NULL;
1606
1607	vfio_group_try_dissolve_container(group);
1608
1609	atomic_dec(&group->opened);
1610
1611	vfio_group_put(group);
1612
1613	return 0;
1614}
1615
1616static const struct file_operations vfio_group_fops = {
1617	.owner		= THIS_MODULE,
1618	.unlocked_ioctl	= vfio_group_fops_unl_ioctl,
1619	.compat_ioctl	= compat_ptr_ioctl,
1620	.open		= vfio_group_fops_open,
1621	.release	= vfio_group_fops_release,
1622};
1623
1624/**
1625 * VFIO Device fd
1626 */
1627static int vfio_device_fops_release(struct inode *inode, struct file *filep)
1628{
1629	struct vfio_device *device = filep->private_data;
1630
1631	device->ops->release(device->device_data);
1632
1633	vfio_group_try_dissolve_container(device->group);
1634
1635	vfio_device_put(device);
1636
1637	return 0;
1638}
1639
1640static long vfio_device_fops_unl_ioctl(struct file *filep,
1641				       unsigned int cmd, unsigned long arg)
1642{
1643	struct vfio_device *device = filep->private_data;
1644
1645	if (unlikely(!device->ops->ioctl))
1646		return -EINVAL;
1647
1648	return device->ops->ioctl(device->device_data, cmd, arg);
1649}
1650
1651static ssize_t vfio_device_fops_read(struct file *filep, char __user *buf,
1652				     size_t count, loff_t *ppos)
1653{
1654	struct vfio_device *device = filep->private_data;
1655
1656	if (unlikely(!device->ops->read))
1657		return -EINVAL;
1658
1659	return device->ops->read(device->device_data, buf, count, ppos);
1660}
1661
1662static ssize_t vfio_device_fops_write(struct file *filep,
1663				      const char __user *buf,
1664				      size_t count, loff_t *ppos)
1665{
1666	struct vfio_device *device = filep->private_data;
1667
1668	if (unlikely(!device->ops->write))
1669		return -EINVAL;
1670
1671	return device->ops->write(device->device_data, buf, count, ppos);
1672}
1673
1674static int vfio_device_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1675{
1676	struct vfio_device *device = filep->private_data;
1677
1678	if (unlikely(!device->ops->mmap))
1679		return -EINVAL;
1680
1681	return device->ops->mmap(device->device_data, vma);
1682}
1683
1684static const struct file_operations vfio_device_fops = {
1685	.owner		= THIS_MODULE,
1686	.release	= vfio_device_fops_release,
1687	.read		= vfio_device_fops_read,
1688	.write		= vfio_device_fops_write,
1689	.unlocked_ioctl	= vfio_device_fops_unl_ioctl,
1690	.compat_ioctl	= compat_ptr_ioctl,
1691	.mmap		= vfio_device_fops_mmap,
1692};
1693
1694/**
1695 * External user API, exported by symbols to be linked dynamically.
1696 *
1697 * The protocol includes:
1698 *  1. do normal VFIO init operation:
1699 *	- opening a new container;
1700 *	- attaching group(s) to it;
1701 *	- setting an IOMMU driver for a container.
1702 * When IOMMU is set for a container, all groups in it are
1703 * considered ready to use by an external user.
1704 *
1705 * 2. User space passes a group fd to an external user.
1706 * The external user calls vfio_group_get_external_user()
1707 * to verify that:
1708 *	- the group is initialized;
1709 *	- IOMMU is set for it.
1710 * If both checks passed, vfio_group_get_external_user()
1711 * increments the container user counter to prevent
1712 * the VFIO group from disposal before KVM exits.
1713 *
1714 * 3. The external user calls vfio_external_user_iommu_id()
1715 * to know an IOMMU ID.
1716 *
1717 * 4. When the external KVM finishes, it calls
1718 * vfio_group_put_external_user() to release the VFIO group.
1719 * This call decrements the container user counter.
1720 */
1721struct vfio_group *vfio_group_get_external_user(struct file *filep)
1722{
1723	struct vfio_group *group = filep->private_data;
1724	int ret;
1725
1726	if (filep->f_op != &vfio_group_fops)
1727		return ERR_PTR(-EINVAL);
1728
1729	ret = vfio_group_add_container_user(group);
1730	if (ret)
1731		return ERR_PTR(ret);
1732
1733	vfio_group_get(group);
1734
1735	return group;
1736}
1737EXPORT_SYMBOL_GPL(vfio_group_get_external_user);
1738
1739/**
1740 * External user API, exported by symbols to be linked dynamically.
1741 * The external user passes in a device pointer
1742 * to verify that:
1743 *	- A VFIO group is assiciated with the device;
1744 *	- IOMMU is set for the group.
1745 * If both checks passed, vfio_group_get_external_user_from_dev()
1746 * increments the container user counter to prevent the VFIO group
1747 * from disposal before external user exits and returns the pointer
1748 * to the VFIO group.
1749 *
1750 * When the external user finishes using the VFIO group, it calls
1751 * vfio_group_put_external_user() to release the VFIO group and
1752 * decrement the container user counter.
1753 *
1754 * @dev [in]	: device
1755 * Return error PTR or pointer to VFIO group.
1756 */
1757
1758struct vfio_group *vfio_group_get_external_user_from_dev(struct device *dev)
1759{
1760	struct vfio_group *group;
1761	int ret;
1762
1763	group = vfio_group_get_from_dev(dev);
1764	if (!group)
1765		return ERR_PTR(-ENODEV);
1766
1767	ret = vfio_group_add_container_user(group);
1768	if (ret) {
1769		vfio_group_put(group);
1770		return ERR_PTR(ret);
1771	}
1772
1773	return group;
1774}
1775EXPORT_SYMBOL_GPL(vfio_group_get_external_user_from_dev);
1776
1777void vfio_group_put_external_user(struct vfio_group *group)
1778{
1779	vfio_group_try_dissolve_container(group);
1780	vfio_group_put(group);
1781}
1782EXPORT_SYMBOL_GPL(vfio_group_put_external_user);
1783
1784bool vfio_external_group_match_file(struct vfio_group *test_group,
1785				    struct file *filep)
1786{
1787	struct vfio_group *group = filep->private_data;
1788
1789	return (filep->f_op == &vfio_group_fops) && (group == test_group);
1790}
1791EXPORT_SYMBOL_GPL(vfio_external_group_match_file);
1792
1793int vfio_external_user_iommu_id(struct vfio_group *group)
1794{
1795	return iommu_group_id(group->iommu_group);
1796}
1797EXPORT_SYMBOL_GPL(vfio_external_user_iommu_id);
1798
1799long vfio_external_check_extension(struct vfio_group *group, unsigned long arg)
1800{
1801	return vfio_ioctl_check_extension(group->container, arg);
1802}
1803EXPORT_SYMBOL_GPL(vfio_external_check_extension);
1804
1805/**
1806 * Sub-module support
1807 */
1808/*
1809 * Helper for managing a buffer of info chain capabilities, allocate or
1810 * reallocate a buffer with additional @size, filling in @id and @version
1811 * of the capability.  A pointer to the new capability is returned.
1812 *
1813 * NB. The chain is based at the head of the buffer, so new entries are
1814 * added to the tail, vfio_info_cap_shift() should be called to fixup the
1815 * next offsets prior to copying to the user buffer.
1816 */
1817struct vfio_info_cap_header *vfio_info_cap_add(struct vfio_info_cap *caps,
1818					       size_t size, u16 id, u16 version)
1819{
1820	void *buf;
1821	struct vfio_info_cap_header *header, *tmp;
1822
1823	buf = krealloc(caps->buf, caps->size + size, GFP_KERNEL);
1824	if (!buf) {
1825		kfree(caps->buf);
1826		caps->size = 0;
1827		return ERR_PTR(-ENOMEM);
1828	}
1829
1830	caps->buf = buf;
1831	header = buf + caps->size;
1832
1833	/* Eventually copied to user buffer, zero */
1834	memset(header, 0, size);
1835
1836	header->id = id;
1837	header->version = version;
1838
1839	/* Add to the end of the capability chain */
1840	for (tmp = buf; tmp->next; tmp = buf + tmp->next)
1841		; /* nothing */
1842
1843	tmp->next = caps->size;
1844	caps->size += size;
1845
1846	return header;
1847}
1848EXPORT_SYMBOL_GPL(vfio_info_cap_add);
1849
1850void vfio_info_cap_shift(struct vfio_info_cap *caps, size_t offset)
1851{
1852	struct vfio_info_cap_header *tmp;
1853	void *buf = (void *)caps->buf;
1854
1855	for (tmp = buf; tmp->next; tmp = buf + tmp->next - offset)
1856		tmp->next += offset;
1857}
1858EXPORT_SYMBOL(vfio_info_cap_shift);
1859
1860int vfio_info_add_capability(struct vfio_info_cap *caps,
1861			     struct vfio_info_cap_header *cap, size_t size)
1862{
1863	struct vfio_info_cap_header *header;
1864
1865	header = vfio_info_cap_add(caps, size, cap->id, cap->version);
1866	if (IS_ERR(header))
1867		return PTR_ERR(header);
1868
1869	memcpy(header + 1, cap + 1, size - sizeof(*header));
1870
1871	return 0;
1872}
1873EXPORT_SYMBOL(vfio_info_add_capability);
1874
1875int vfio_set_irqs_validate_and_prepare(struct vfio_irq_set *hdr, int num_irqs,
1876				       int max_irq_type, size_t *data_size)
1877{
1878	unsigned long minsz;
1879	size_t size;
1880
1881	minsz = offsetofend(struct vfio_irq_set, count);
1882
1883	if ((hdr->argsz < minsz) || (hdr->index >= max_irq_type) ||
1884	    (hdr->count >= (U32_MAX - hdr->start)) ||
1885	    (hdr->flags & ~(VFIO_IRQ_SET_DATA_TYPE_MASK |
1886				VFIO_IRQ_SET_ACTION_TYPE_MASK)))
1887		return -EINVAL;
1888
1889	if (data_size)
1890		*data_size = 0;
1891
1892	if (hdr->start >= num_irqs || hdr->start + hdr->count > num_irqs)
1893		return -EINVAL;
1894
1895	switch (hdr->flags & VFIO_IRQ_SET_DATA_TYPE_MASK) {
1896	case VFIO_IRQ_SET_DATA_NONE:
1897		size = 0;
1898		break;
1899	case VFIO_IRQ_SET_DATA_BOOL:
1900		size = sizeof(uint8_t);
1901		break;
1902	case VFIO_IRQ_SET_DATA_EVENTFD:
1903		size = sizeof(int32_t);
1904		break;
1905	default:
1906		return -EINVAL;
1907	}
1908
1909	if (size) {
1910		if (hdr->argsz - minsz < hdr->count * size)
1911			return -EINVAL;
1912
1913		if (!data_size)
1914			return -EINVAL;
1915
1916		*data_size = hdr->count * size;
1917	}
1918
1919	return 0;
1920}
1921EXPORT_SYMBOL(vfio_set_irqs_validate_and_prepare);
1922
1923/*
1924 * Pin a set of guest PFNs and return their associated host PFNs for local
1925 * domain only.
1926 * @dev [in]     : device
1927 * @user_pfn [in]: array of user/guest PFNs to be pinned.
1928 * @npage [in]   : count of elements in user_pfn array.  This count should not
1929 *		   be greater VFIO_PIN_PAGES_MAX_ENTRIES.
1930 * @prot [in]    : protection flags
1931 * @phys_pfn[out]: array of host PFNs
1932 * Return error or number of pages pinned.
1933 */
1934int vfio_pin_pages(struct device *dev, unsigned long *user_pfn, int npage,
1935		   int prot, unsigned long *phys_pfn)
1936{
1937	struct vfio_container *container;
1938	struct vfio_group *group;
1939	struct vfio_iommu_driver *driver;
1940	int ret;
1941
1942	if (!dev || !user_pfn || !phys_pfn || !npage)
1943		return -EINVAL;
1944
1945	if (npage > VFIO_PIN_PAGES_MAX_ENTRIES)
1946		return -E2BIG;
1947
1948	group = vfio_group_get_from_dev(dev);
1949	if (!group)
1950		return -ENODEV;
1951
1952	if (group->dev_counter > 1)
1953		return -EINVAL;
1954
1955	ret = vfio_group_add_container_user(group);
1956	if (ret)
1957		goto err_pin_pages;
1958
1959	container = group->container;
1960	driver = container->iommu_driver;
1961	if (likely(driver && driver->ops->pin_pages))
1962		ret = driver->ops->pin_pages(container->iommu_data,
1963					     group->iommu_group, user_pfn,
1964					     npage, prot, phys_pfn);
1965	else
1966		ret = -ENOTTY;
1967
1968	vfio_group_try_dissolve_container(group);
1969
1970err_pin_pages:
1971	vfio_group_put(group);
1972	return ret;
1973}
1974EXPORT_SYMBOL(vfio_pin_pages);
1975
1976/*
1977 * Unpin set of host PFNs for local domain only.
1978 * @dev [in]     : device
1979 * @user_pfn [in]: array of user/guest PFNs to be unpinned. Number of user/guest
1980 *		   PFNs should not be greater than VFIO_PIN_PAGES_MAX_ENTRIES.
1981 * @npage [in]   : count of elements in user_pfn array.  This count should not
1982 *                 be greater than VFIO_PIN_PAGES_MAX_ENTRIES.
1983 * Return error or number of pages unpinned.
1984 */
1985int vfio_unpin_pages(struct device *dev, unsigned long *user_pfn, int npage)
1986{
1987	struct vfio_container *container;
1988	struct vfio_group *group;
1989	struct vfio_iommu_driver *driver;
1990	int ret;
1991
1992	if (!dev || !user_pfn || !npage)
1993		return -EINVAL;
1994
1995	if (npage > VFIO_PIN_PAGES_MAX_ENTRIES)
1996		return -E2BIG;
1997
1998	group = vfio_group_get_from_dev(dev);
1999	if (!group)
2000		return -ENODEV;
2001
2002	ret = vfio_group_add_container_user(group);
2003	if (ret)
2004		goto err_unpin_pages;
2005
2006	container = group->container;
2007	driver = container->iommu_driver;
2008	if (likely(driver && driver->ops->unpin_pages))
2009		ret = driver->ops->unpin_pages(container->iommu_data, user_pfn,
2010					       npage);
2011	else
2012		ret = -ENOTTY;
2013
2014	vfio_group_try_dissolve_container(group);
2015
2016err_unpin_pages:
2017	vfio_group_put(group);
2018	return ret;
2019}
2020EXPORT_SYMBOL(vfio_unpin_pages);
2021
2022/*
2023 * Pin a set of guest IOVA PFNs and return their associated host PFNs for a
2024 * VFIO group.
2025 *
2026 * The caller needs to call vfio_group_get_external_user() or
2027 * vfio_group_get_external_user_from_dev() prior to calling this interface,
2028 * so as to prevent the VFIO group from disposal in the middle of the call.
2029 * But it can keep the reference to the VFIO group for several calls into
2030 * this interface.
2031 * After finishing using of the VFIO group, the caller needs to release the
2032 * VFIO group by calling vfio_group_put_external_user().
2033 *
2034 * @group [in]		: VFIO group
2035 * @user_iova_pfn [in]	: array of user/guest IOVA PFNs to be pinned.
2036 * @npage [in]		: count of elements in user_iova_pfn array.
2037 *			  This count should not be greater
2038 *			  VFIO_PIN_PAGES_MAX_ENTRIES.
2039 * @prot [in]		: protection flags
2040 * @phys_pfn [out]	: array of host PFNs
2041 * Return error or number of pages pinned.
2042 */
2043int vfio_group_pin_pages(struct vfio_group *group,
2044			 unsigned long *user_iova_pfn, int npage,
2045			 int prot, unsigned long *phys_pfn)
2046{
2047	struct vfio_container *container;
2048	struct vfio_iommu_driver *driver;
2049	int ret;
2050
2051	if (!group || !user_iova_pfn || !phys_pfn || !npage)
2052		return -EINVAL;
2053
2054	if (npage > VFIO_PIN_PAGES_MAX_ENTRIES)
2055		return -E2BIG;
2056
2057	container = group->container;
2058	driver = container->iommu_driver;
2059	if (likely(driver && driver->ops->pin_pages))
2060		ret = driver->ops->pin_pages(container->iommu_data,
2061					     group->iommu_group, user_iova_pfn,
2062					     npage, prot, phys_pfn);
2063	else
2064		ret = -ENOTTY;
2065
2066	return ret;
2067}
2068EXPORT_SYMBOL(vfio_group_pin_pages);
2069
2070/*
2071 * Unpin a set of guest IOVA PFNs for a VFIO group.
2072 *
2073 * The caller needs to call vfio_group_get_external_user() or
2074 * vfio_group_get_external_user_from_dev() prior to calling this interface,
2075 * so as to prevent the VFIO group from disposal in the middle of the call.
2076 * But it can keep the reference to the VFIO group for several calls into
2077 * this interface.
2078 * After finishing using of the VFIO group, the caller needs to release the
2079 * VFIO group by calling vfio_group_put_external_user().
2080 *
2081 * @group [in]		: vfio group
2082 * @user_iova_pfn [in]	: array of user/guest IOVA PFNs to be unpinned.
2083 * @npage [in]		: count of elements in user_iova_pfn array.
2084 *			  This count should not be greater than
2085 *			  VFIO_PIN_PAGES_MAX_ENTRIES.
2086 * Return error or number of pages unpinned.
2087 */
2088int vfio_group_unpin_pages(struct vfio_group *group,
2089			   unsigned long *user_iova_pfn, int npage)
2090{
2091	struct vfio_container *container;
2092	struct vfio_iommu_driver *driver;
2093	int ret;
2094
2095	if (!group || !user_iova_pfn || !npage)
2096		return -EINVAL;
2097
2098	if (npage > VFIO_PIN_PAGES_MAX_ENTRIES)
2099		return -E2BIG;
2100
2101	container = group->container;
2102	driver = container->iommu_driver;
2103	if (likely(driver && driver->ops->unpin_pages))
2104		ret = driver->ops->unpin_pages(container->iommu_data,
2105					       user_iova_pfn, npage);
2106	else
2107		ret = -ENOTTY;
2108
2109	return ret;
2110}
2111EXPORT_SYMBOL(vfio_group_unpin_pages);
2112
2113
2114/*
2115 * This interface allows the CPUs to perform some sort of virtual DMA on
2116 * behalf of the device.
2117 *
2118 * CPUs read/write from/into a range of IOVAs pointing to user space memory
2119 * into/from a kernel buffer.
2120 *
2121 * As the read/write of user space memory is conducted via the CPUs and is
2122 * not a real device DMA, it is not necessary to pin the user space memory.
2123 *
2124 * The caller needs to call vfio_group_get_external_user() or
2125 * vfio_group_get_external_user_from_dev() prior to calling this interface,
2126 * so as to prevent the VFIO group from disposal in the middle of the call.
2127 * But it can keep the reference to the VFIO group for several calls into
2128 * this interface.
2129 * After finishing using of the VFIO group, the caller needs to release the
2130 * VFIO group by calling vfio_group_put_external_user().
2131 *
2132 * @group [in]		: VFIO group
2133 * @user_iova [in]	: base IOVA of a user space buffer
2134 * @data [in]		: pointer to kernel buffer
2135 * @len [in]		: kernel buffer length
2136 * @write		: indicate read or write
2137 * Return error code on failure or 0 on success.
2138 */
2139int vfio_dma_rw(struct vfio_group *group, dma_addr_t user_iova,
2140		void *data, size_t len, bool write)
2141{
2142	struct vfio_container *container;
2143	struct vfio_iommu_driver *driver;
2144	int ret = 0;
2145
2146	if (!group || !data || len <= 0)
2147		return -EINVAL;
2148
2149	container = group->container;
2150	driver = container->iommu_driver;
2151
2152	if (likely(driver && driver->ops->dma_rw))
2153		ret = driver->ops->dma_rw(container->iommu_data,
2154					  user_iova, data, len, write);
2155	else
2156		ret = -ENOTTY;
2157
2158	return ret;
2159}
2160EXPORT_SYMBOL(vfio_dma_rw);
2161
2162static int vfio_register_iommu_notifier(struct vfio_group *group,
2163					unsigned long *events,
2164					struct notifier_block *nb)
2165{
2166	struct vfio_container *container;
2167	struct vfio_iommu_driver *driver;
2168	int ret;
2169
2170	ret = vfio_group_add_container_user(group);
2171	if (ret)
2172		return -EINVAL;
2173
2174	container = group->container;
2175	driver = container->iommu_driver;
2176	if (likely(driver && driver->ops->register_notifier))
2177		ret = driver->ops->register_notifier(container->iommu_data,
2178						     events, nb);
2179	else
2180		ret = -ENOTTY;
2181
2182	vfio_group_try_dissolve_container(group);
2183
2184	return ret;
2185}
2186
2187static int vfio_unregister_iommu_notifier(struct vfio_group *group,
2188					  struct notifier_block *nb)
2189{
2190	struct vfio_container *container;
2191	struct vfio_iommu_driver *driver;
2192	int ret;
2193
2194	ret = vfio_group_add_container_user(group);
2195	if (ret)
2196		return -EINVAL;
2197
2198	container = group->container;
2199	driver = container->iommu_driver;
2200	if (likely(driver && driver->ops->unregister_notifier))
2201		ret = driver->ops->unregister_notifier(container->iommu_data,
2202						       nb);
2203	else
2204		ret = -ENOTTY;
2205
2206	vfio_group_try_dissolve_container(group);
2207
2208	return ret;
2209}
2210
2211void vfio_group_set_kvm(struct vfio_group *group, struct kvm *kvm)
2212{
2213	group->kvm = kvm;
2214	blocking_notifier_call_chain(&group->notifier,
2215				VFIO_GROUP_NOTIFY_SET_KVM, kvm);
2216}
2217EXPORT_SYMBOL_GPL(vfio_group_set_kvm);
2218
2219static int vfio_register_group_notifier(struct vfio_group *group,
2220					unsigned long *events,
2221					struct notifier_block *nb)
2222{
2223	int ret;
2224	bool set_kvm = false;
2225
2226	if (*events & VFIO_GROUP_NOTIFY_SET_KVM)
2227		set_kvm = true;
2228
2229	/* clear known events */
2230	*events &= ~VFIO_GROUP_NOTIFY_SET_KVM;
2231
2232	/* refuse to continue if still events remaining */
2233	if (*events)
2234		return -EINVAL;
2235
2236	ret = vfio_group_add_container_user(group);
2237	if (ret)
2238		return -EINVAL;
2239
2240	ret = blocking_notifier_chain_register(&group->notifier, nb);
2241
2242	/*
2243	 * The attaching of kvm and vfio_group might already happen, so
2244	 * here we replay once upon registration.
2245	 */
2246	if (!ret && set_kvm && group->kvm)
2247		blocking_notifier_call_chain(&group->notifier,
2248					VFIO_GROUP_NOTIFY_SET_KVM, group->kvm);
2249
2250	vfio_group_try_dissolve_container(group);
2251
2252	return ret;
2253}
2254
2255static int vfio_unregister_group_notifier(struct vfio_group *group,
2256					 struct notifier_block *nb)
2257{
2258	int ret;
2259
2260	ret = vfio_group_add_container_user(group);
2261	if (ret)
2262		return -EINVAL;
2263
2264	ret = blocking_notifier_chain_unregister(&group->notifier, nb);
2265
2266	vfio_group_try_dissolve_container(group);
2267
2268	return ret;
2269}
2270
2271int vfio_register_notifier(struct device *dev, enum vfio_notify_type type,
2272			   unsigned long *events, struct notifier_block *nb)
2273{
2274	struct vfio_group *group;
2275	int ret;
2276
2277	if (!dev || !nb || !events || (*events == 0))
2278		return -EINVAL;
2279
2280	group = vfio_group_get_from_dev(dev);
2281	if (!group)
2282		return -ENODEV;
2283
2284	switch (type) {
2285	case VFIO_IOMMU_NOTIFY:
2286		ret = vfio_register_iommu_notifier(group, events, nb);
2287		break;
2288	case VFIO_GROUP_NOTIFY:
2289		ret = vfio_register_group_notifier(group, events, nb);
2290		break;
2291	default:
2292		ret = -EINVAL;
2293	}
2294
2295	vfio_group_put(group);
2296	return ret;
2297}
2298EXPORT_SYMBOL(vfio_register_notifier);
2299
2300int vfio_unregister_notifier(struct device *dev, enum vfio_notify_type type,
2301			     struct notifier_block *nb)
2302{
2303	struct vfio_group *group;
2304	int ret;
2305
2306	if (!dev || !nb)
2307		return -EINVAL;
2308
2309	group = vfio_group_get_from_dev(dev);
2310	if (!group)
2311		return -ENODEV;
2312
2313	switch (type) {
2314	case VFIO_IOMMU_NOTIFY:
2315		ret = vfio_unregister_iommu_notifier(group, nb);
2316		break;
2317	case VFIO_GROUP_NOTIFY:
2318		ret = vfio_unregister_group_notifier(group, nb);
2319		break;
2320	default:
2321		ret = -EINVAL;
2322	}
2323
2324	vfio_group_put(group);
2325	return ret;
2326}
2327EXPORT_SYMBOL(vfio_unregister_notifier);
2328
2329/**
2330 * Module/class support
2331 */
2332static char *vfio_devnode(struct device *dev, umode_t *mode)
2333{
2334	return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
2335}
2336
2337static struct miscdevice vfio_dev = {
2338	.minor = VFIO_MINOR,
2339	.name = "vfio",
2340	.fops = &vfio_fops,
2341	.nodename = "vfio/vfio",
2342	.mode = S_IRUGO | S_IWUGO,
2343};
2344
2345static int __init vfio_init(void)
2346{
2347	int ret;
2348
2349	idr_init(&vfio.group_idr);
2350	mutex_init(&vfio.group_lock);
2351	mutex_init(&vfio.iommu_drivers_lock);
2352	INIT_LIST_HEAD(&vfio.group_list);
2353	INIT_LIST_HEAD(&vfio.iommu_drivers_list);
2354	init_waitqueue_head(&vfio.release_q);
2355
2356	ret = misc_register(&vfio_dev);
2357	if (ret) {
2358		pr_err("vfio: misc device register failed\n");
2359		return ret;
2360	}
2361
2362	/* /dev/vfio/$GROUP */
2363	vfio.class = class_create(THIS_MODULE, "vfio");
2364	if (IS_ERR(vfio.class)) {
2365		ret = PTR_ERR(vfio.class);
2366		goto err_class;
2367	}
2368
2369	vfio.class->devnode = vfio_devnode;
2370
2371	ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK + 1, "vfio");
2372	if (ret)
2373		goto err_alloc_chrdev;
2374
2375	cdev_init(&vfio.group_cdev, &vfio_group_fops);
2376	ret = cdev_add(&vfio.group_cdev, vfio.group_devt, MINORMASK + 1);
2377	if (ret)
2378		goto err_cdev_add;
2379
2380	pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n");
2381
2382#ifdef CONFIG_VFIO_NOIOMMU
2383	vfio_register_iommu_driver(&vfio_noiommu_ops);
2384#endif
2385	return 0;
2386
2387err_cdev_add:
2388	unregister_chrdev_region(vfio.group_devt, MINORMASK + 1);
2389err_alloc_chrdev:
2390	class_destroy(vfio.class);
2391	vfio.class = NULL;
2392err_class:
2393	misc_deregister(&vfio_dev);
2394	return ret;
2395}
2396
2397static void __exit vfio_cleanup(void)
2398{
2399	WARN_ON(!list_empty(&vfio.group_list));
2400
2401#ifdef CONFIG_VFIO_NOIOMMU
2402	vfio_unregister_iommu_driver(&vfio_noiommu_ops);
2403#endif
2404	idr_destroy(&vfio.group_idr);
2405	cdev_del(&vfio.group_cdev);
2406	unregister_chrdev_region(vfio.group_devt, MINORMASK + 1);
2407	class_destroy(vfio.class);
2408	vfio.class = NULL;
2409	misc_deregister(&vfio_dev);
2410}
2411
2412module_init(vfio_init);
2413module_exit(vfio_cleanup);
2414
2415MODULE_VERSION(DRIVER_VERSION);
2416MODULE_LICENSE("GPL v2");
2417MODULE_AUTHOR(DRIVER_AUTHOR);
2418MODULE_DESCRIPTION(DRIVER_DESC);
2419MODULE_ALIAS_MISCDEV(VFIO_MINOR);
2420MODULE_ALIAS("devname:vfio/vfio");
2421MODULE_SOFTDEP("post: vfio_iommu_type1 vfio_iommu_spapr_tce");