xref: /linux/drivers/vfio/vfio_main.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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/fs.h>
17 #include <linux/idr.h>
18 #include <linux/iommu.h>
19 #if IS_ENABLED(CONFIG_KVM)
20 #include <linux/kvm_host.h>
21 #endif
22 #include <linux/list.h>
23 #include <linux/miscdevice.h>
24 #include <linux/module.h>
25 #include <linux/mount.h>
26 #include <linux/mutex.h>
27 #include <linux/pci.h>
28 #include <linux/pseudo_fs.h>
29 #include <linux/rwsem.h>
30 #include <linux/sched.h>
31 #include <linux/slab.h>
32 #include <linux/stat.h>
33 #include <linux/string.h>
34 #include <linux/uaccess.h>
35 #include <linux/vfio.h>
36 #include <linux/wait.h>
37 #include <linux/sched/signal.h>
38 #include <linux/pm_runtime.h>
39 #include <linux/interval_tree.h>
40 #include <linux/iova_bitmap.h>
41 #include <linux/iommufd.h>
42 #include "vfio.h"
43 
44 #define DRIVER_VERSION	"0.3"
45 #define DRIVER_AUTHOR	"Alex Williamson <alex.williamson@redhat.com>"
46 #define DRIVER_DESC	"VFIO - User Level meta-driver"
47 
48 #define VFIO_MAGIC 0x5646494f /* "VFIO" */
49 
50 static struct vfio {
51 	struct class			*device_class;
52 	struct ida			device_ida;
53 	struct vfsmount			*vfs_mount;
54 	int				fs_count;
55 } vfio;
56 
57 #ifdef CONFIG_VFIO_NOIOMMU
58 bool vfio_noiommu __read_mostly;
59 module_param_named(enable_unsafe_noiommu_mode,
60 		   vfio_noiommu, bool, S_IRUGO | S_IWUSR);
61 MODULE_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)");
62 #endif
63 
64 static DEFINE_XARRAY(vfio_device_set_xa);
65 
66 int vfio_assign_device_set(struct vfio_device *device, void *set_id)
67 {
68 	unsigned long idx = (unsigned long)set_id;
69 	struct vfio_device_set *new_dev_set;
70 	struct vfio_device_set *dev_set;
71 
72 	if (WARN_ON(!set_id))
73 		return -EINVAL;
74 
75 	/*
76 	 * Atomically acquire a singleton object in the xarray for this set_id
77 	 */
78 	xa_lock(&vfio_device_set_xa);
79 	dev_set = xa_load(&vfio_device_set_xa, idx);
80 	if (dev_set)
81 		goto found_get_ref;
82 	xa_unlock(&vfio_device_set_xa);
83 
84 	new_dev_set = kzalloc(sizeof(*new_dev_set), GFP_KERNEL);
85 	if (!new_dev_set)
86 		return -ENOMEM;
87 	mutex_init(&new_dev_set->lock);
88 	INIT_LIST_HEAD(&new_dev_set->device_list);
89 	new_dev_set->set_id = set_id;
90 
91 	xa_lock(&vfio_device_set_xa);
92 	dev_set = __xa_cmpxchg(&vfio_device_set_xa, idx, NULL, new_dev_set,
93 			       GFP_KERNEL);
94 	if (!dev_set) {
95 		dev_set = new_dev_set;
96 		goto found_get_ref;
97 	}
98 
99 	kfree(new_dev_set);
100 	if (xa_is_err(dev_set)) {
101 		xa_unlock(&vfio_device_set_xa);
102 		return xa_err(dev_set);
103 	}
104 
105 found_get_ref:
106 	dev_set->device_count++;
107 	xa_unlock(&vfio_device_set_xa);
108 	mutex_lock(&dev_set->lock);
109 	device->dev_set = dev_set;
110 	list_add_tail(&device->dev_set_list, &dev_set->device_list);
111 	mutex_unlock(&dev_set->lock);
112 	return 0;
113 }
114 EXPORT_SYMBOL_GPL(vfio_assign_device_set);
115 
116 static void vfio_release_device_set(struct vfio_device *device)
117 {
118 	struct vfio_device_set *dev_set = device->dev_set;
119 
120 	if (!dev_set)
121 		return;
122 
123 	mutex_lock(&dev_set->lock);
124 	list_del(&device->dev_set_list);
125 	mutex_unlock(&dev_set->lock);
126 
127 	xa_lock(&vfio_device_set_xa);
128 	if (!--dev_set->device_count) {
129 		__xa_erase(&vfio_device_set_xa,
130 			   (unsigned long)dev_set->set_id);
131 		mutex_destroy(&dev_set->lock);
132 		kfree(dev_set);
133 	}
134 	xa_unlock(&vfio_device_set_xa);
135 }
136 
137 unsigned int vfio_device_set_open_count(struct vfio_device_set *dev_set)
138 {
139 	struct vfio_device *cur;
140 	unsigned int open_count = 0;
141 
142 	lockdep_assert_held(&dev_set->lock);
143 
144 	list_for_each_entry(cur, &dev_set->device_list, dev_set_list)
145 		open_count += cur->open_count;
146 	return open_count;
147 }
148 EXPORT_SYMBOL_GPL(vfio_device_set_open_count);
149 
150 struct vfio_device *
151 vfio_find_device_in_devset(struct vfio_device_set *dev_set,
152 			   struct device *dev)
153 {
154 	struct vfio_device *cur;
155 
156 	lockdep_assert_held(&dev_set->lock);
157 
158 	list_for_each_entry(cur, &dev_set->device_list, dev_set_list)
159 		if (cur->dev == dev)
160 			return cur;
161 	return NULL;
162 }
163 EXPORT_SYMBOL_GPL(vfio_find_device_in_devset);
164 
165 /*
166  * Device objects - create, release, get, put, search
167  */
168 /* Device reference always implies a group reference */
169 void vfio_device_put_registration(struct vfio_device *device)
170 {
171 	if (refcount_dec_and_test(&device->refcount))
172 		complete(&device->comp);
173 }
174 
175 bool vfio_device_try_get_registration(struct vfio_device *device)
176 {
177 	return refcount_inc_not_zero(&device->refcount);
178 }
179 
180 /*
181  * VFIO driver API
182  */
183 /* Release helper called by vfio_put_device() */
184 static void vfio_device_release(struct device *dev)
185 {
186 	struct vfio_device *device =
187 			container_of(dev, struct vfio_device, device);
188 
189 	vfio_release_device_set(device);
190 	ida_free(&vfio.device_ida, device->index);
191 
192 	if (device->ops->release)
193 		device->ops->release(device);
194 
195 	iput(device->inode);
196 	simple_release_fs(&vfio.vfs_mount, &vfio.fs_count);
197 	kvfree(device);
198 }
199 
200 static int vfio_init_device(struct vfio_device *device, struct device *dev,
201 			    const struct vfio_device_ops *ops);
202 
203 /*
204  * Allocate and initialize vfio_device so it can be registered to vfio
205  * core.
206  *
207  * Drivers should use the wrapper vfio_alloc_device() for allocation.
208  * @size is the size of the structure to be allocated, including any
209  * private data used by the driver.
210  *
211  * Driver may provide an @init callback to cover device private data.
212  *
213  * Use vfio_put_device() to release the structure after success return.
214  */
215 struct vfio_device *_vfio_alloc_device(size_t size, struct device *dev,
216 				       const struct vfio_device_ops *ops)
217 {
218 	struct vfio_device *device;
219 	int ret;
220 
221 	if (WARN_ON(size < sizeof(struct vfio_device)))
222 		return ERR_PTR(-EINVAL);
223 
224 	device = kvzalloc(size, GFP_KERNEL);
225 	if (!device)
226 		return ERR_PTR(-ENOMEM);
227 
228 	ret = vfio_init_device(device, dev, ops);
229 	if (ret)
230 		goto out_free;
231 	return device;
232 
233 out_free:
234 	kvfree(device);
235 	return ERR_PTR(ret);
236 }
237 EXPORT_SYMBOL_GPL(_vfio_alloc_device);
238 
239 static int vfio_fs_init_fs_context(struct fs_context *fc)
240 {
241 	return init_pseudo(fc, VFIO_MAGIC) ? 0 : -ENOMEM;
242 }
243 
244 static struct file_system_type vfio_fs_type = {
245 	.name = "vfio",
246 	.owner = THIS_MODULE,
247 	.init_fs_context = vfio_fs_init_fs_context,
248 	.kill_sb = kill_anon_super,
249 };
250 
251 static struct inode *vfio_fs_inode_new(void)
252 {
253 	struct inode *inode;
254 	int ret;
255 
256 	ret = simple_pin_fs(&vfio_fs_type, &vfio.vfs_mount, &vfio.fs_count);
257 	if (ret)
258 		return ERR_PTR(ret);
259 
260 	inode = alloc_anon_inode(vfio.vfs_mount->mnt_sb);
261 	if (IS_ERR(inode))
262 		simple_release_fs(&vfio.vfs_mount, &vfio.fs_count);
263 
264 	return inode;
265 }
266 
267 /*
268  * Initialize a vfio_device so it can be registered to vfio core.
269  */
270 static int vfio_init_device(struct vfio_device *device, struct device *dev,
271 			    const struct vfio_device_ops *ops)
272 {
273 	int ret;
274 
275 	ret = ida_alloc_max(&vfio.device_ida, MINORMASK, GFP_KERNEL);
276 	if (ret < 0) {
277 		dev_dbg(dev, "Error to alloc index\n");
278 		return ret;
279 	}
280 
281 	device->index = ret;
282 	init_completion(&device->comp);
283 	device->dev = dev;
284 	device->ops = ops;
285 	device->inode = vfio_fs_inode_new();
286 	if (IS_ERR(device->inode)) {
287 		ret = PTR_ERR(device->inode);
288 		goto out_inode;
289 	}
290 
291 	if (ops->init) {
292 		ret = ops->init(device);
293 		if (ret)
294 			goto out_uninit;
295 	}
296 
297 	device_initialize(&device->device);
298 	device->device.release = vfio_device_release;
299 	device->device.class = vfio.device_class;
300 	device->device.parent = device->dev;
301 	return 0;
302 
303 out_uninit:
304 	iput(device->inode);
305 	simple_release_fs(&vfio.vfs_mount, &vfio.fs_count);
306 out_inode:
307 	vfio_release_device_set(device);
308 	ida_free(&vfio.device_ida, device->index);
309 	return ret;
310 }
311 
312 static int __vfio_register_dev(struct vfio_device *device,
313 			       enum vfio_group_type type)
314 {
315 	int ret;
316 
317 	if (WARN_ON(IS_ENABLED(CONFIG_IOMMUFD) &&
318 		    (!device->ops->bind_iommufd ||
319 		     !device->ops->unbind_iommufd ||
320 		     !device->ops->attach_ioas ||
321 		     !device->ops->detach_ioas)))
322 		return -EINVAL;
323 
324 	/*
325 	 * If the driver doesn't specify a set then the device is added to a
326 	 * singleton set just for itself.
327 	 */
328 	if (!device->dev_set)
329 		vfio_assign_device_set(device, device);
330 
331 	ret = dev_set_name(&device->device, "vfio%d", device->index);
332 	if (ret)
333 		return ret;
334 
335 	ret = vfio_device_set_group(device, type);
336 	if (ret)
337 		return ret;
338 
339 	/*
340 	 * VFIO always sets IOMMU_CACHE because we offer no way for userspace to
341 	 * restore cache coherency. It has to be checked here because it is only
342 	 * valid for cases where we are using iommu groups.
343 	 */
344 	if (type == VFIO_IOMMU && !vfio_device_is_noiommu(device) &&
345 	    !device_iommu_capable(device->dev, IOMMU_CAP_CACHE_COHERENCY)) {
346 		ret = -EINVAL;
347 		goto err_out;
348 	}
349 
350 	ret = vfio_device_add(device);
351 	if (ret)
352 		goto err_out;
353 
354 	/* Refcounting can't start until the driver calls register */
355 	refcount_set(&device->refcount, 1);
356 
357 	vfio_device_group_register(device);
358 	vfio_device_debugfs_init(device);
359 
360 	return 0;
361 err_out:
362 	vfio_device_remove_group(device);
363 	return ret;
364 }
365 
366 int vfio_register_group_dev(struct vfio_device *device)
367 {
368 	return __vfio_register_dev(device, VFIO_IOMMU);
369 }
370 EXPORT_SYMBOL_GPL(vfio_register_group_dev);
371 
372 /*
373  * Register a virtual device without IOMMU backing.  The user of this
374  * device must not be able to directly trigger unmediated DMA.
375  */
376 int vfio_register_emulated_iommu_dev(struct vfio_device *device)
377 {
378 	return __vfio_register_dev(device, VFIO_EMULATED_IOMMU);
379 }
380 EXPORT_SYMBOL_GPL(vfio_register_emulated_iommu_dev);
381 
382 /*
383  * Decrement the device reference count and wait for the device to be
384  * removed.  Open file descriptors for the device... */
385 void vfio_unregister_group_dev(struct vfio_device *device)
386 {
387 	unsigned int i = 0;
388 	bool interrupted = false;
389 	long rc;
390 
391 	/*
392 	 * Prevent new device opened by userspace via the
393 	 * VFIO_GROUP_GET_DEVICE_FD in the group path.
394 	 */
395 	vfio_device_group_unregister(device);
396 
397 	/*
398 	 * Balances vfio_device_add() in register path, also prevents
399 	 * new device opened by userspace in the cdev path.
400 	 */
401 	vfio_device_del(device);
402 
403 	vfio_device_put_registration(device);
404 	rc = try_wait_for_completion(&device->comp);
405 	while (rc <= 0) {
406 		if (device->ops->request)
407 			device->ops->request(device, i++);
408 
409 		if (interrupted) {
410 			rc = wait_for_completion_timeout(&device->comp,
411 							 HZ * 10);
412 		} else {
413 			rc = wait_for_completion_interruptible_timeout(
414 				&device->comp, HZ * 10);
415 			if (rc < 0) {
416 				interrupted = true;
417 				dev_warn(device->dev,
418 					 "Device is currently in use, task"
419 					 " \"%s\" (%d) "
420 					 "blocked until device is released",
421 					 current->comm, task_pid_nr(current));
422 			}
423 		}
424 	}
425 
426 	vfio_device_debugfs_exit(device);
427 	/* Balances vfio_device_set_group in register path */
428 	vfio_device_remove_group(device);
429 }
430 EXPORT_SYMBOL_GPL(vfio_unregister_group_dev);
431 
432 #if IS_ENABLED(CONFIG_KVM)
433 void vfio_device_get_kvm_safe(struct vfio_device *device, struct kvm *kvm)
434 {
435 	void (*pfn)(struct kvm *kvm);
436 	bool (*fn)(struct kvm *kvm);
437 	bool ret;
438 
439 	lockdep_assert_held(&device->dev_set->lock);
440 
441 	if (!kvm)
442 		return;
443 
444 	pfn = symbol_get(kvm_put_kvm);
445 	if (WARN_ON(!pfn))
446 		return;
447 
448 	fn = symbol_get(kvm_get_kvm_safe);
449 	if (WARN_ON(!fn)) {
450 		symbol_put(kvm_put_kvm);
451 		return;
452 	}
453 
454 	ret = fn(kvm);
455 	symbol_put(kvm_get_kvm_safe);
456 	if (!ret) {
457 		symbol_put(kvm_put_kvm);
458 		return;
459 	}
460 
461 	device->put_kvm = pfn;
462 	device->kvm = kvm;
463 }
464 
465 void vfio_device_put_kvm(struct vfio_device *device)
466 {
467 	lockdep_assert_held(&device->dev_set->lock);
468 
469 	if (!device->kvm)
470 		return;
471 
472 	if (WARN_ON(!device->put_kvm))
473 		goto clear;
474 
475 	device->put_kvm(device->kvm);
476 	device->put_kvm = NULL;
477 	symbol_put(kvm_put_kvm);
478 
479 clear:
480 	device->kvm = NULL;
481 }
482 #endif
483 
484 /* true if the vfio_device has open_device() called but not close_device() */
485 static bool vfio_assert_device_open(struct vfio_device *device)
486 {
487 	return !WARN_ON_ONCE(!READ_ONCE(device->open_count));
488 }
489 
490 struct vfio_device_file *
491 vfio_allocate_device_file(struct vfio_device *device)
492 {
493 	struct vfio_device_file *df;
494 
495 	df = kzalloc(sizeof(*df), GFP_KERNEL_ACCOUNT);
496 	if (!df)
497 		return ERR_PTR(-ENOMEM);
498 
499 	df->device = device;
500 	spin_lock_init(&df->kvm_ref_lock);
501 
502 	return df;
503 }
504 
505 static int vfio_df_device_first_open(struct vfio_device_file *df)
506 {
507 	struct vfio_device *device = df->device;
508 	struct iommufd_ctx *iommufd = df->iommufd;
509 	int ret;
510 
511 	lockdep_assert_held(&device->dev_set->lock);
512 
513 	if (!try_module_get(device->dev->driver->owner))
514 		return -ENODEV;
515 
516 	if (iommufd)
517 		ret = vfio_df_iommufd_bind(df);
518 	else
519 		ret = vfio_device_group_use_iommu(device);
520 	if (ret)
521 		goto err_module_put;
522 
523 	if (device->ops->open_device) {
524 		ret = device->ops->open_device(device);
525 		if (ret)
526 			goto err_unuse_iommu;
527 	}
528 	return 0;
529 
530 err_unuse_iommu:
531 	if (iommufd)
532 		vfio_df_iommufd_unbind(df);
533 	else
534 		vfio_device_group_unuse_iommu(device);
535 err_module_put:
536 	module_put(device->dev->driver->owner);
537 	return ret;
538 }
539 
540 static void vfio_df_device_last_close(struct vfio_device_file *df)
541 {
542 	struct vfio_device *device = df->device;
543 	struct iommufd_ctx *iommufd = df->iommufd;
544 
545 	lockdep_assert_held(&device->dev_set->lock);
546 
547 	if (device->ops->close_device)
548 		device->ops->close_device(device);
549 	if (iommufd)
550 		vfio_df_iommufd_unbind(df);
551 	else
552 		vfio_device_group_unuse_iommu(device);
553 	module_put(device->dev->driver->owner);
554 }
555 
556 int vfio_df_open(struct vfio_device_file *df)
557 {
558 	struct vfio_device *device = df->device;
559 	int ret = 0;
560 
561 	lockdep_assert_held(&device->dev_set->lock);
562 
563 	/*
564 	 * Only the group path allows the device to be opened multiple
565 	 * times.  The device cdev path doesn't have a secure way for it.
566 	 */
567 	if (device->open_count != 0 && !df->group)
568 		return -EINVAL;
569 
570 	device->open_count++;
571 	if (device->open_count == 1) {
572 		ret = vfio_df_device_first_open(df);
573 		if (ret)
574 			device->open_count--;
575 	}
576 
577 	return ret;
578 }
579 
580 void vfio_df_close(struct vfio_device_file *df)
581 {
582 	struct vfio_device *device = df->device;
583 
584 	lockdep_assert_held(&device->dev_set->lock);
585 
586 	vfio_assert_device_open(device);
587 	if (device->open_count == 1)
588 		vfio_df_device_last_close(df);
589 	device->open_count--;
590 }
591 
592 /*
593  * Wrapper around pm_runtime_resume_and_get().
594  * Return error code on failure or 0 on success.
595  */
596 static inline int vfio_device_pm_runtime_get(struct vfio_device *device)
597 {
598 	struct device *dev = device->dev;
599 
600 	if (dev->driver && dev->driver->pm) {
601 		int ret;
602 
603 		ret = pm_runtime_resume_and_get(dev);
604 		if (ret) {
605 			dev_info_ratelimited(dev,
606 				"vfio: runtime resume failed %d\n", ret);
607 			return -EIO;
608 		}
609 	}
610 
611 	return 0;
612 }
613 
614 /*
615  * Wrapper around pm_runtime_put().
616  */
617 static inline void vfio_device_pm_runtime_put(struct vfio_device *device)
618 {
619 	struct device *dev = device->dev;
620 
621 	if (dev->driver && dev->driver->pm)
622 		pm_runtime_put(dev);
623 }
624 
625 /*
626  * VFIO Device fd
627  */
628 static int vfio_device_fops_release(struct inode *inode, struct file *filep)
629 {
630 	struct vfio_device_file *df = filep->private_data;
631 	struct vfio_device *device = df->device;
632 
633 	if (df->group)
634 		vfio_df_group_close(df);
635 	else
636 		vfio_df_unbind_iommufd(df);
637 
638 	vfio_device_put_registration(device);
639 
640 	kfree(df);
641 
642 	return 0;
643 }
644 
645 /*
646  * vfio_mig_get_next_state - Compute the next step in the FSM
647  * @cur_fsm - The current state the device is in
648  * @new_fsm - The target state to reach
649  * @next_fsm - Pointer to the next step to get to new_fsm
650  *
651  * Return 0 upon success, otherwise -errno
652  * Upon success the next step in the state progression between cur_fsm and
653  * new_fsm will be set in next_fsm.
654  *
655  * This breaks down requests for combination transitions into smaller steps and
656  * returns the next step to get to new_fsm. The function may need to be called
657  * multiple times before reaching new_fsm.
658  *
659  */
660 int vfio_mig_get_next_state(struct vfio_device *device,
661 			    enum vfio_device_mig_state cur_fsm,
662 			    enum vfio_device_mig_state new_fsm,
663 			    enum vfio_device_mig_state *next_fsm)
664 {
665 	enum { VFIO_DEVICE_NUM_STATES = VFIO_DEVICE_STATE_PRE_COPY_P2P + 1 };
666 	/*
667 	 * The coding in this table requires the driver to implement the
668 	 * following FSM arcs:
669 	 *         RESUMING -> STOP
670 	 *         STOP -> RESUMING
671 	 *         STOP -> STOP_COPY
672 	 *         STOP_COPY -> STOP
673 	 *
674 	 * If P2P is supported then the driver must also implement these FSM
675 	 * arcs:
676 	 *         RUNNING -> RUNNING_P2P
677 	 *         RUNNING_P2P -> RUNNING
678 	 *         RUNNING_P2P -> STOP
679 	 *         STOP -> RUNNING_P2P
680 	 *
681 	 * If precopy is supported then the driver must support these additional
682 	 * FSM arcs:
683 	 *         RUNNING -> PRE_COPY
684 	 *         PRE_COPY -> RUNNING
685 	 *         PRE_COPY -> STOP_COPY
686 	 * However, if precopy and P2P are supported together then the driver
687 	 * must support these additional arcs beyond the P2P arcs above:
688 	 *         PRE_COPY -> RUNNING
689 	 *         PRE_COPY -> PRE_COPY_P2P
690 	 *         PRE_COPY_P2P -> PRE_COPY
691 	 *         PRE_COPY_P2P -> RUNNING_P2P
692 	 *         PRE_COPY_P2P -> STOP_COPY
693 	 *         RUNNING -> PRE_COPY
694 	 *         RUNNING_P2P -> PRE_COPY_P2P
695 	 *
696 	 * Without P2P and precopy the driver must implement:
697 	 *         RUNNING -> STOP
698 	 *         STOP -> RUNNING
699 	 *
700 	 * The coding will step through multiple states for some combination
701 	 * transitions; if all optional features are supported, this means the
702 	 * following ones:
703 	 *         PRE_COPY -> PRE_COPY_P2P -> STOP_COPY
704 	 *         PRE_COPY -> RUNNING -> RUNNING_P2P
705 	 *         PRE_COPY -> RUNNING -> RUNNING_P2P -> STOP
706 	 *         PRE_COPY -> RUNNING -> RUNNING_P2P -> STOP -> RESUMING
707 	 *         PRE_COPY_P2P -> RUNNING_P2P -> RUNNING
708 	 *         PRE_COPY_P2P -> RUNNING_P2P -> STOP
709 	 *         PRE_COPY_P2P -> RUNNING_P2P -> STOP -> RESUMING
710 	 *         RESUMING -> STOP -> RUNNING_P2P
711 	 *         RESUMING -> STOP -> RUNNING_P2P -> PRE_COPY_P2P
712 	 *         RESUMING -> STOP -> RUNNING_P2P -> RUNNING
713 	 *         RESUMING -> STOP -> RUNNING_P2P -> RUNNING -> PRE_COPY
714 	 *         RESUMING -> STOP -> STOP_COPY
715 	 *         RUNNING -> RUNNING_P2P -> PRE_COPY_P2P
716 	 *         RUNNING -> RUNNING_P2P -> STOP
717 	 *         RUNNING -> RUNNING_P2P -> STOP -> RESUMING
718 	 *         RUNNING -> RUNNING_P2P -> STOP -> STOP_COPY
719 	 *         RUNNING_P2P -> RUNNING -> PRE_COPY
720 	 *         RUNNING_P2P -> STOP -> RESUMING
721 	 *         RUNNING_P2P -> STOP -> STOP_COPY
722 	 *         STOP -> RUNNING_P2P -> PRE_COPY_P2P
723 	 *         STOP -> RUNNING_P2P -> RUNNING
724 	 *         STOP -> RUNNING_P2P -> RUNNING -> PRE_COPY
725 	 *         STOP_COPY -> STOP -> RESUMING
726 	 *         STOP_COPY -> STOP -> RUNNING_P2P
727 	 *         STOP_COPY -> STOP -> RUNNING_P2P -> RUNNING
728 	 *
729 	 *  The following transitions are blocked:
730 	 *         STOP_COPY -> PRE_COPY
731 	 *         STOP_COPY -> PRE_COPY_P2P
732 	 */
733 	static const u8 vfio_from_fsm_table[VFIO_DEVICE_NUM_STATES][VFIO_DEVICE_NUM_STATES] = {
734 		[VFIO_DEVICE_STATE_STOP] = {
735 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
736 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING_P2P,
737 			[VFIO_DEVICE_STATE_PRE_COPY] = VFIO_DEVICE_STATE_RUNNING_P2P,
738 			[VFIO_DEVICE_STATE_PRE_COPY_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
739 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP_COPY,
740 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RESUMING,
741 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
742 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
743 		},
744 		[VFIO_DEVICE_STATE_RUNNING] = {
745 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_RUNNING_P2P,
746 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING,
747 			[VFIO_DEVICE_STATE_PRE_COPY] = VFIO_DEVICE_STATE_PRE_COPY,
748 			[VFIO_DEVICE_STATE_PRE_COPY_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
749 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_RUNNING_P2P,
750 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RUNNING_P2P,
751 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
752 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
753 		},
754 		[VFIO_DEVICE_STATE_PRE_COPY] = {
755 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_RUNNING,
756 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING,
757 			[VFIO_DEVICE_STATE_PRE_COPY] = VFIO_DEVICE_STATE_PRE_COPY,
758 			[VFIO_DEVICE_STATE_PRE_COPY_P2P] = VFIO_DEVICE_STATE_PRE_COPY_P2P,
759 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_PRE_COPY_P2P,
760 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RUNNING,
761 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING,
762 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
763 		},
764 		[VFIO_DEVICE_STATE_PRE_COPY_P2P] = {
765 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_RUNNING_P2P,
766 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING_P2P,
767 			[VFIO_DEVICE_STATE_PRE_COPY] = VFIO_DEVICE_STATE_PRE_COPY,
768 			[VFIO_DEVICE_STATE_PRE_COPY_P2P] = VFIO_DEVICE_STATE_PRE_COPY_P2P,
769 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP_COPY,
770 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RUNNING_P2P,
771 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
772 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
773 		},
774 		[VFIO_DEVICE_STATE_STOP_COPY] = {
775 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
776 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_STOP,
777 			[VFIO_DEVICE_STATE_PRE_COPY] = VFIO_DEVICE_STATE_ERROR,
778 			[VFIO_DEVICE_STATE_PRE_COPY_P2P] = VFIO_DEVICE_STATE_ERROR,
779 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP_COPY,
780 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_STOP,
781 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_STOP,
782 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
783 		},
784 		[VFIO_DEVICE_STATE_RESUMING] = {
785 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
786 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_STOP,
787 			[VFIO_DEVICE_STATE_PRE_COPY] = VFIO_DEVICE_STATE_STOP,
788 			[VFIO_DEVICE_STATE_PRE_COPY_P2P] = VFIO_DEVICE_STATE_STOP,
789 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP,
790 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RESUMING,
791 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_STOP,
792 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
793 		},
794 		[VFIO_DEVICE_STATE_RUNNING_P2P] = {
795 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
796 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING,
797 			[VFIO_DEVICE_STATE_PRE_COPY] = VFIO_DEVICE_STATE_RUNNING,
798 			[VFIO_DEVICE_STATE_PRE_COPY_P2P] = VFIO_DEVICE_STATE_PRE_COPY_P2P,
799 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP,
800 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_STOP,
801 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
802 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
803 		},
804 		[VFIO_DEVICE_STATE_ERROR] = {
805 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_ERROR,
806 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_ERROR,
807 			[VFIO_DEVICE_STATE_PRE_COPY] = VFIO_DEVICE_STATE_ERROR,
808 			[VFIO_DEVICE_STATE_PRE_COPY_P2P] = VFIO_DEVICE_STATE_ERROR,
809 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_ERROR,
810 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_ERROR,
811 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_ERROR,
812 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
813 		},
814 	};
815 
816 	static const unsigned int state_flags_table[VFIO_DEVICE_NUM_STATES] = {
817 		[VFIO_DEVICE_STATE_STOP] = VFIO_MIGRATION_STOP_COPY,
818 		[VFIO_DEVICE_STATE_RUNNING] = VFIO_MIGRATION_STOP_COPY,
819 		[VFIO_DEVICE_STATE_PRE_COPY] =
820 			VFIO_MIGRATION_STOP_COPY | VFIO_MIGRATION_PRE_COPY,
821 		[VFIO_DEVICE_STATE_PRE_COPY_P2P] = VFIO_MIGRATION_STOP_COPY |
822 						   VFIO_MIGRATION_P2P |
823 						   VFIO_MIGRATION_PRE_COPY,
824 		[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_MIGRATION_STOP_COPY,
825 		[VFIO_DEVICE_STATE_RESUMING] = VFIO_MIGRATION_STOP_COPY,
826 		[VFIO_DEVICE_STATE_RUNNING_P2P] =
827 			VFIO_MIGRATION_STOP_COPY | VFIO_MIGRATION_P2P,
828 		[VFIO_DEVICE_STATE_ERROR] = ~0U,
829 	};
830 
831 	if (WARN_ON(cur_fsm >= ARRAY_SIZE(vfio_from_fsm_table) ||
832 		    (state_flags_table[cur_fsm] & device->migration_flags) !=
833 			state_flags_table[cur_fsm]))
834 		return -EINVAL;
835 
836 	if (new_fsm >= ARRAY_SIZE(vfio_from_fsm_table) ||
837 	   (state_flags_table[new_fsm] & device->migration_flags) !=
838 			state_flags_table[new_fsm])
839 		return -EINVAL;
840 
841 	/*
842 	 * Arcs touching optional and unsupported states are skipped over. The
843 	 * driver will instead see an arc from the original state to the next
844 	 * logical state, as per the above comment.
845 	 */
846 	*next_fsm = vfio_from_fsm_table[cur_fsm][new_fsm];
847 	while ((state_flags_table[*next_fsm] & device->migration_flags) !=
848 			state_flags_table[*next_fsm])
849 		*next_fsm = vfio_from_fsm_table[*next_fsm][new_fsm];
850 
851 	return (*next_fsm != VFIO_DEVICE_STATE_ERROR) ? 0 : -EINVAL;
852 }
853 EXPORT_SYMBOL_GPL(vfio_mig_get_next_state);
854 
855 /*
856  * Convert the drivers's struct file into a FD number and return it to userspace
857  */
858 static int vfio_ioct_mig_return_fd(struct file *filp, void __user *arg,
859 				   struct vfio_device_feature_mig_state *mig)
860 {
861 	int ret;
862 	int fd;
863 
864 	fd = get_unused_fd_flags(O_CLOEXEC);
865 	if (fd < 0) {
866 		ret = fd;
867 		goto out_fput;
868 	}
869 
870 	mig->data_fd = fd;
871 	if (copy_to_user(arg, mig, sizeof(*mig))) {
872 		ret = -EFAULT;
873 		goto out_put_unused;
874 	}
875 	fd_install(fd, filp);
876 	return 0;
877 
878 out_put_unused:
879 	put_unused_fd(fd);
880 out_fput:
881 	fput(filp);
882 	return ret;
883 }
884 
885 static int
886 vfio_ioctl_device_feature_mig_device_state(struct vfio_device *device,
887 					   u32 flags, void __user *arg,
888 					   size_t argsz)
889 {
890 	size_t minsz =
891 		offsetofend(struct vfio_device_feature_mig_state, data_fd);
892 	struct vfio_device_feature_mig_state mig;
893 	struct file *filp = NULL;
894 	int ret;
895 
896 	if (!device->mig_ops)
897 		return -ENOTTY;
898 
899 	ret = vfio_check_feature(flags, argsz,
900 				 VFIO_DEVICE_FEATURE_SET |
901 				 VFIO_DEVICE_FEATURE_GET,
902 				 sizeof(mig));
903 	if (ret != 1)
904 		return ret;
905 
906 	if (copy_from_user(&mig, arg, minsz))
907 		return -EFAULT;
908 
909 	if (flags & VFIO_DEVICE_FEATURE_GET) {
910 		enum vfio_device_mig_state curr_state;
911 
912 		ret = device->mig_ops->migration_get_state(device,
913 							   &curr_state);
914 		if (ret)
915 			return ret;
916 		mig.device_state = curr_state;
917 		goto out_copy;
918 	}
919 
920 	/* Handle the VFIO_DEVICE_FEATURE_SET */
921 	filp = device->mig_ops->migration_set_state(device, mig.device_state);
922 	if (IS_ERR(filp) || !filp)
923 		goto out_copy;
924 
925 	return vfio_ioct_mig_return_fd(filp, arg, &mig);
926 out_copy:
927 	mig.data_fd = -1;
928 	if (copy_to_user(arg, &mig, sizeof(mig)))
929 		return -EFAULT;
930 	if (IS_ERR(filp))
931 		return PTR_ERR(filp);
932 	return 0;
933 }
934 
935 static int
936 vfio_ioctl_device_feature_migration_data_size(struct vfio_device *device,
937 					      u32 flags, void __user *arg,
938 					      size_t argsz)
939 {
940 	struct vfio_device_feature_mig_data_size data_size = {};
941 	unsigned long stop_copy_length;
942 	int ret;
943 
944 	if (!device->mig_ops)
945 		return -ENOTTY;
946 
947 	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_GET,
948 				 sizeof(data_size));
949 	if (ret != 1)
950 		return ret;
951 
952 	ret = device->mig_ops->migration_get_data_size(device, &stop_copy_length);
953 	if (ret)
954 		return ret;
955 
956 	data_size.stop_copy_length = stop_copy_length;
957 	if (copy_to_user(arg, &data_size, sizeof(data_size)))
958 		return -EFAULT;
959 
960 	return 0;
961 }
962 
963 static int vfio_ioctl_device_feature_migration(struct vfio_device *device,
964 					       u32 flags, void __user *arg,
965 					       size_t argsz)
966 {
967 	struct vfio_device_feature_migration mig = {
968 		.flags = device->migration_flags,
969 	};
970 	int ret;
971 
972 	if (!device->mig_ops)
973 		return -ENOTTY;
974 
975 	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_GET,
976 				 sizeof(mig));
977 	if (ret != 1)
978 		return ret;
979 	if (copy_to_user(arg, &mig, sizeof(mig)))
980 		return -EFAULT;
981 	return 0;
982 }
983 
984 void vfio_combine_iova_ranges(struct rb_root_cached *root, u32 cur_nodes,
985 			      u32 req_nodes)
986 {
987 	struct interval_tree_node *prev, *curr, *comb_start, *comb_end;
988 	unsigned long min_gap, curr_gap;
989 
990 	/* Special shortcut when a single range is required */
991 	if (req_nodes == 1) {
992 		unsigned long last;
993 
994 		comb_start = interval_tree_iter_first(root, 0, ULONG_MAX);
995 
996 		/* Empty list */
997 		if (WARN_ON_ONCE(!comb_start))
998 			return;
999 
1000 		curr = comb_start;
1001 		while (curr) {
1002 			last = curr->last;
1003 			prev = curr;
1004 			curr = interval_tree_iter_next(curr, 0, ULONG_MAX);
1005 			if (prev != comb_start)
1006 				interval_tree_remove(prev, root);
1007 		}
1008 		comb_start->last = last;
1009 		return;
1010 	}
1011 
1012 	/* Combine ranges which have the smallest gap */
1013 	while (cur_nodes > req_nodes) {
1014 		prev = NULL;
1015 		min_gap = ULONG_MAX;
1016 		curr = interval_tree_iter_first(root, 0, ULONG_MAX);
1017 		while (curr) {
1018 			if (prev) {
1019 				curr_gap = curr->start - prev->last;
1020 				if (curr_gap < min_gap) {
1021 					min_gap = curr_gap;
1022 					comb_start = prev;
1023 					comb_end = curr;
1024 				}
1025 			}
1026 			prev = curr;
1027 			curr = interval_tree_iter_next(curr, 0, ULONG_MAX);
1028 		}
1029 
1030 		/* Empty list or no nodes to combine */
1031 		if (WARN_ON_ONCE(min_gap == ULONG_MAX))
1032 			break;
1033 
1034 		comb_start->last = comb_end->last;
1035 		interval_tree_remove(comb_end, root);
1036 		cur_nodes--;
1037 	}
1038 }
1039 EXPORT_SYMBOL_GPL(vfio_combine_iova_ranges);
1040 
1041 /* Ranges should fit into a single kernel page */
1042 #define LOG_MAX_RANGES \
1043 	(PAGE_SIZE / sizeof(struct vfio_device_feature_dma_logging_range))
1044 
1045 static int
1046 vfio_ioctl_device_feature_logging_start(struct vfio_device *device,
1047 					u32 flags, void __user *arg,
1048 					size_t argsz)
1049 {
1050 	size_t minsz =
1051 		offsetofend(struct vfio_device_feature_dma_logging_control,
1052 			    ranges);
1053 	struct vfio_device_feature_dma_logging_range __user *ranges;
1054 	struct vfio_device_feature_dma_logging_control control;
1055 	struct vfio_device_feature_dma_logging_range range;
1056 	struct rb_root_cached root = RB_ROOT_CACHED;
1057 	struct interval_tree_node *nodes;
1058 	u64 iova_end;
1059 	u32 nnodes;
1060 	int i, ret;
1061 
1062 	if (!device->log_ops)
1063 		return -ENOTTY;
1064 
1065 	ret = vfio_check_feature(flags, argsz,
1066 				 VFIO_DEVICE_FEATURE_SET,
1067 				 sizeof(control));
1068 	if (ret != 1)
1069 		return ret;
1070 
1071 	if (copy_from_user(&control, arg, minsz))
1072 		return -EFAULT;
1073 
1074 	nnodes = control.num_ranges;
1075 	if (!nnodes)
1076 		return -EINVAL;
1077 
1078 	if (nnodes > LOG_MAX_RANGES)
1079 		return -E2BIG;
1080 
1081 	ranges = u64_to_user_ptr(control.ranges);
1082 	nodes = kmalloc_array(nnodes, sizeof(struct interval_tree_node),
1083 			      GFP_KERNEL);
1084 	if (!nodes)
1085 		return -ENOMEM;
1086 
1087 	for (i = 0; i < nnodes; i++) {
1088 		if (copy_from_user(&range, &ranges[i], sizeof(range))) {
1089 			ret = -EFAULT;
1090 			goto end;
1091 		}
1092 		if (!IS_ALIGNED(range.iova, control.page_size) ||
1093 		    !IS_ALIGNED(range.length, control.page_size)) {
1094 			ret = -EINVAL;
1095 			goto end;
1096 		}
1097 
1098 		if (check_add_overflow(range.iova, range.length, &iova_end) ||
1099 		    iova_end > ULONG_MAX) {
1100 			ret = -EOVERFLOW;
1101 			goto end;
1102 		}
1103 
1104 		nodes[i].start = range.iova;
1105 		nodes[i].last = range.iova + range.length - 1;
1106 		if (interval_tree_iter_first(&root, nodes[i].start,
1107 					     nodes[i].last)) {
1108 			/* Range overlapping */
1109 			ret = -EINVAL;
1110 			goto end;
1111 		}
1112 		interval_tree_insert(nodes + i, &root);
1113 	}
1114 
1115 	ret = device->log_ops->log_start(device, &root, nnodes,
1116 					 &control.page_size);
1117 	if (ret)
1118 		goto end;
1119 
1120 	if (copy_to_user(arg, &control, sizeof(control))) {
1121 		ret = -EFAULT;
1122 		device->log_ops->log_stop(device);
1123 	}
1124 
1125 end:
1126 	kfree(nodes);
1127 	return ret;
1128 }
1129 
1130 static int
1131 vfio_ioctl_device_feature_logging_stop(struct vfio_device *device,
1132 				       u32 flags, void __user *arg,
1133 				       size_t argsz)
1134 {
1135 	int ret;
1136 
1137 	if (!device->log_ops)
1138 		return -ENOTTY;
1139 
1140 	ret = vfio_check_feature(flags, argsz,
1141 				 VFIO_DEVICE_FEATURE_SET, 0);
1142 	if (ret != 1)
1143 		return ret;
1144 
1145 	return device->log_ops->log_stop(device);
1146 }
1147 
1148 static int vfio_device_log_read_and_clear(struct iova_bitmap *iter,
1149 					  unsigned long iova, size_t length,
1150 					  void *opaque)
1151 {
1152 	struct vfio_device *device = opaque;
1153 
1154 	return device->log_ops->log_read_and_clear(device, iova, length, iter);
1155 }
1156 
1157 static int
1158 vfio_ioctl_device_feature_logging_report(struct vfio_device *device,
1159 					 u32 flags, void __user *arg,
1160 					 size_t argsz)
1161 {
1162 	size_t minsz =
1163 		offsetofend(struct vfio_device_feature_dma_logging_report,
1164 			    bitmap);
1165 	struct vfio_device_feature_dma_logging_report report;
1166 	struct iova_bitmap *iter;
1167 	u64 iova_end;
1168 	int ret;
1169 
1170 	if (!device->log_ops)
1171 		return -ENOTTY;
1172 
1173 	ret = vfio_check_feature(flags, argsz,
1174 				 VFIO_DEVICE_FEATURE_GET,
1175 				 sizeof(report));
1176 	if (ret != 1)
1177 		return ret;
1178 
1179 	if (copy_from_user(&report, arg, minsz))
1180 		return -EFAULT;
1181 
1182 	if (report.page_size < SZ_4K || !is_power_of_2(report.page_size))
1183 		return -EINVAL;
1184 
1185 	if (check_add_overflow(report.iova, report.length, &iova_end) ||
1186 	    iova_end > ULONG_MAX)
1187 		return -EOVERFLOW;
1188 
1189 	iter = iova_bitmap_alloc(report.iova, report.length,
1190 				 report.page_size,
1191 				 u64_to_user_ptr(report.bitmap));
1192 	if (IS_ERR(iter))
1193 		return PTR_ERR(iter);
1194 
1195 	ret = iova_bitmap_for_each(iter, device,
1196 				   vfio_device_log_read_and_clear);
1197 
1198 	iova_bitmap_free(iter);
1199 	return ret;
1200 }
1201 
1202 static int vfio_ioctl_device_feature(struct vfio_device *device,
1203 				     struct vfio_device_feature __user *arg)
1204 {
1205 	size_t minsz = offsetofend(struct vfio_device_feature, flags);
1206 	struct vfio_device_feature feature;
1207 
1208 	if (copy_from_user(&feature, arg, minsz))
1209 		return -EFAULT;
1210 
1211 	if (feature.argsz < minsz)
1212 		return -EINVAL;
1213 
1214 	/* Check unknown flags */
1215 	if (feature.flags &
1216 	    ~(VFIO_DEVICE_FEATURE_MASK | VFIO_DEVICE_FEATURE_SET |
1217 	      VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_PROBE))
1218 		return -EINVAL;
1219 
1220 	/* GET & SET are mutually exclusive except with PROBE */
1221 	if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1222 	    (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1223 	    (feature.flags & VFIO_DEVICE_FEATURE_GET))
1224 		return -EINVAL;
1225 
1226 	switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1227 	case VFIO_DEVICE_FEATURE_MIGRATION:
1228 		return vfio_ioctl_device_feature_migration(
1229 			device, feature.flags, arg->data,
1230 			feature.argsz - minsz);
1231 	case VFIO_DEVICE_FEATURE_MIG_DEVICE_STATE:
1232 		return vfio_ioctl_device_feature_mig_device_state(
1233 			device, feature.flags, arg->data,
1234 			feature.argsz - minsz);
1235 	case VFIO_DEVICE_FEATURE_DMA_LOGGING_START:
1236 		return vfio_ioctl_device_feature_logging_start(
1237 			device, feature.flags, arg->data,
1238 			feature.argsz - minsz);
1239 	case VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP:
1240 		return vfio_ioctl_device_feature_logging_stop(
1241 			device, feature.flags, arg->data,
1242 			feature.argsz - minsz);
1243 	case VFIO_DEVICE_FEATURE_DMA_LOGGING_REPORT:
1244 		return vfio_ioctl_device_feature_logging_report(
1245 			device, feature.flags, arg->data,
1246 			feature.argsz - minsz);
1247 	case VFIO_DEVICE_FEATURE_MIG_DATA_SIZE:
1248 		return vfio_ioctl_device_feature_migration_data_size(
1249 			device, feature.flags, arg->data,
1250 			feature.argsz - minsz);
1251 	default:
1252 		if (unlikely(!device->ops->device_feature))
1253 			return -EINVAL;
1254 		return device->ops->device_feature(device, feature.flags,
1255 						   arg->data,
1256 						   feature.argsz - minsz);
1257 	}
1258 }
1259 
1260 static long vfio_device_fops_unl_ioctl(struct file *filep,
1261 				       unsigned int cmd, unsigned long arg)
1262 {
1263 	struct vfio_device_file *df = filep->private_data;
1264 	struct vfio_device *device = df->device;
1265 	void __user *uptr = (void __user *)arg;
1266 	int ret;
1267 
1268 	if (cmd == VFIO_DEVICE_BIND_IOMMUFD)
1269 		return vfio_df_ioctl_bind_iommufd(df, uptr);
1270 
1271 	/* Paired with smp_store_release() following vfio_df_open() */
1272 	if (!smp_load_acquire(&df->access_granted))
1273 		return -EINVAL;
1274 
1275 	ret = vfio_device_pm_runtime_get(device);
1276 	if (ret)
1277 		return ret;
1278 
1279 	/* cdev only ioctls */
1280 	if (IS_ENABLED(CONFIG_VFIO_DEVICE_CDEV) && !df->group) {
1281 		switch (cmd) {
1282 		case VFIO_DEVICE_ATTACH_IOMMUFD_PT:
1283 			ret = vfio_df_ioctl_attach_pt(df, uptr);
1284 			goto out;
1285 
1286 		case VFIO_DEVICE_DETACH_IOMMUFD_PT:
1287 			ret = vfio_df_ioctl_detach_pt(df, uptr);
1288 			goto out;
1289 		}
1290 	}
1291 
1292 	switch (cmd) {
1293 	case VFIO_DEVICE_FEATURE:
1294 		ret = vfio_ioctl_device_feature(device, uptr);
1295 		break;
1296 
1297 	default:
1298 		if (unlikely(!device->ops->ioctl))
1299 			ret = -EINVAL;
1300 		else
1301 			ret = device->ops->ioctl(device, cmd, arg);
1302 		break;
1303 	}
1304 out:
1305 	vfio_device_pm_runtime_put(device);
1306 	return ret;
1307 }
1308 
1309 static ssize_t vfio_device_fops_read(struct file *filep, char __user *buf,
1310 				     size_t count, loff_t *ppos)
1311 {
1312 	struct vfio_device_file *df = filep->private_data;
1313 	struct vfio_device *device = df->device;
1314 
1315 	/* Paired with smp_store_release() following vfio_df_open() */
1316 	if (!smp_load_acquire(&df->access_granted))
1317 		return -EINVAL;
1318 
1319 	if (unlikely(!device->ops->read))
1320 		return -EINVAL;
1321 
1322 	return device->ops->read(device, buf, count, ppos);
1323 }
1324 
1325 static ssize_t vfio_device_fops_write(struct file *filep,
1326 				      const char __user *buf,
1327 				      size_t count, loff_t *ppos)
1328 {
1329 	struct vfio_device_file *df = filep->private_data;
1330 	struct vfio_device *device = df->device;
1331 
1332 	/* Paired with smp_store_release() following vfio_df_open() */
1333 	if (!smp_load_acquire(&df->access_granted))
1334 		return -EINVAL;
1335 
1336 	if (unlikely(!device->ops->write))
1337 		return -EINVAL;
1338 
1339 	return device->ops->write(device, buf, count, ppos);
1340 }
1341 
1342 static int vfio_device_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1343 {
1344 	struct vfio_device_file *df = filep->private_data;
1345 	struct vfio_device *device = df->device;
1346 
1347 	/* Paired with smp_store_release() following vfio_df_open() */
1348 	if (!smp_load_acquire(&df->access_granted))
1349 		return -EINVAL;
1350 
1351 	if (unlikely(!device->ops->mmap))
1352 		return -EINVAL;
1353 
1354 	return device->ops->mmap(device, vma);
1355 }
1356 
1357 const struct file_operations vfio_device_fops = {
1358 	.owner		= THIS_MODULE,
1359 	.open		= vfio_device_fops_cdev_open,
1360 	.release	= vfio_device_fops_release,
1361 	.read		= vfio_device_fops_read,
1362 	.write		= vfio_device_fops_write,
1363 	.unlocked_ioctl	= vfio_device_fops_unl_ioctl,
1364 	.compat_ioctl	= compat_ptr_ioctl,
1365 	.mmap		= vfio_device_fops_mmap,
1366 };
1367 
1368 static struct vfio_device *vfio_device_from_file(struct file *file)
1369 {
1370 	struct vfio_device_file *df = file->private_data;
1371 
1372 	if (file->f_op != &vfio_device_fops)
1373 		return NULL;
1374 	return df->device;
1375 }
1376 
1377 /**
1378  * vfio_file_is_valid - True if the file is valid vfio file
1379  * @file: VFIO group file or VFIO device file
1380  */
1381 bool vfio_file_is_valid(struct file *file)
1382 {
1383 	return vfio_group_from_file(file) ||
1384 	       vfio_device_from_file(file);
1385 }
1386 EXPORT_SYMBOL_GPL(vfio_file_is_valid);
1387 
1388 /**
1389  * vfio_file_enforced_coherent - True if the DMA associated with the VFIO file
1390  *        is always CPU cache coherent
1391  * @file: VFIO group file or VFIO device file
1392  *
1393  * Enforced coherency means that the IOMMU ignores things like the PCIe no-snoop
1394  * bit in DMA transactions. A return of false indicates that the user has
1395  * rights to access additional instructions such as wbinvd on x86.
1396  */
1397 bool vfio_file_enforced_coherent(struct file *file)
1398 {
1399 	struct vfio_device *device;
1400 	struct vfio_group *group;
1401 
1402 	group = vfio_group_from_file(file);
1403 	if (group)
1404 		return vfio_group_enforced_coherent(group);
1405 
1406 	device = vfio_device_from_file(file);
1407 	if (device)
1408 		return device_iommu_capable(device->dev,
1409 					    IOMMU_CAP_ENFORCE_CACHE_COHERENCY);
1410 
1411 	return true;
1412 }
1413 EXPORT_SYMBOL_GPL(vfio_file_enforced_coherent);
1414 
1415 static void vfio_device_file_set_kvm(struct file *file, struct kvm *kvm)
1416 {
1417 	struct vfio_device_file *df = file->private_data;
1418 
1419 	/*
1420 	 * The kvm is first recorded in the vfio_device_file, and will
1421 	 * be propagated to vfio_device::kvm when the file is bound to
1422 	 * iommufd successfully in the vfio device cdev path.
1423 	 */
1424 	spin_lock(&df->kvm_ref_lock);
1425 	df->kvm = kvm;
1426 	spin_unlock(&df->kvm_ref_lock);
1427 }
1428 
1429 /**
1430  * vfio_file_set_kvm - Link a kvm with VFIO drivers
1431  * @file: VFIO group file or VFIO device file
1432  * @kvm: KVM to link
1433  *
1434  * When a VFIO device is first opened the KVM will be available in
1435  * device->kvm if one was associated with the file.
1436  */
1437 void vfio_file_set_kvm(struct file *file, struct kvm *kvm)
1438 {
1439 	struct vfio_group *group;
1440 
1441 	group = vfio_group_from_file(file);
1442 	if (group)
1443 		vfio_group_set_kvm(group, kvm);
1444 
1445 	if (vfio_device_from_file(file))
1446 		vfio_device_file_set_kvm(file, kvm);
1447 }
1448 EXPORT_SYMBOL_GPL(vfio_file_set_kvm);
1449 
1450 /*
1451  * Sub-module support
1452  */
1453 /*
1454  * Helper for managing a buffer of info chain capabilities, allocate or
1455  * reallocate a buffer with additional @size, filling in @id and @version
1456  * of the capability.  A pointer to the new capability is returned.
1457  *
1458  * NB. The chain is based at the head of the buffer, so new entries are
1459  * added to the tail, vfio_info_cap_shift() should be called to fixup the
1460  * next offsets prior to copying to the user buffer.
1461  */
1462 struct vfio_info_cap_header *vfio_info_cap_add(struct vfio_info_cap *caps,
1463 					       size_t size, u16 id, u16 version)
1464 {
1465 	void *buf;
1466 	struct vfio_info_cap_header *header, *tmp;
1467 
1468 	/* Ensure that the next capability struct will be aligned */
1469 	size = ALIGN(size, sizeof(u64));
1470 
1471 	buf = krealloc(caps->buf, caps->size + size, GFP_KERNEL);
1472 	if (!buf) {
1473 		kfree(caps->buf);
1474 		caps->buf = NULL;
1475 		caps->size = 0;
1476 		return ERR_PTR(-ENOMEM);
1477 	}
1478 
1479 	caps->buf = buf;
1480 	header = buf + caps->size;
1481 
1482 	/* Eventually copied to user buffer, zero */
1483 	memset(header, 0, size);
1484 
1485 	header->id = id;
1486 	header->version = version;
1487 
1488 	/* Add to the end of the capability chain */
1489 	for (tmp = buf; tmp->next; tmp = buf + tmp->next)
1490 		; /* nothing */
1491 
1492 	tmp->next = caps->size;
1493 	caps->size += size;
1494 
1495 	return header;
1496 }
1497 EXPORT_SYMBOL_GPL(vfio_info_cap_add);
1498 
1499 void vfio_info_cap_shift(struct vfio_info_cap *caps, size_t offset)
1500 {
1501 	struct vfio_info_cap_header *tmp;
1502 	void *buf = (void *)caps->buf;
1503 
1504 	/* Capability structs should start with proper alignment */
1505 	WARN_ON(!IS_ALIGNED(offset, sizeof(u64)));
1506 
1507 	for (tmp = buf; tmp->next; tmp = buf + tmp->next - offset)
1508 		tmp->next += offset;
1509 }
1510 EXPORT_SYMBOL(vfio_info_cap_shift);
1511 
1512 int vfio_info_add_capability(struct vfio_info_cap *caps,
1513 			     struct vfio_info_cap_header *cap, size_t size)
1514 {
1515 	struct vfio_info_cap_header *header;
1516 
1517 	header = vfio_info_cap_add(caps, size, cap->id, cap->version);
1518 	if (IS_ERR(header))
1519 		return PTR_ERR(header);
1520 
1521 	memcpy(header + 1, cap + 1, size - sizeof(*header));
1522 
1523 	return 0;
1524 }
1525 EXPORT_SYMBOL(vfio_info_add_capability);
1526 
1527 int vfio_set_irqs_validate_and_prepare(struct vfio_irq_set *hdr, int num_irqs,
1528 				       int max_irq_type, size_t *data_size)
1529 {
1530 	unsigned long minsz;
1531 	size_t size;
1532 
1533 	minsz = offsetofend(struct vfio_irq_set, count);
1534 
1535 	if ((hdr->argsz < minsz) || (hdr->index >= max_irq_type) ||
1536 	    (hdr->count >= (U32_MAX - hdr->start)) ||
1537 	    (hdr->flags & ~(VFIO_IRQ_SET_DATA_TYPE_MASK |
1538 				VFIO_IRQ_SET_ACTION_TYPE_MASK)))
1539 		return -EINVAL;
1540 
1541 	if (data_size)
1542 		*data_size = 0;
1543 
1544 	if (hdr->start >= num_irqs || hdr->start + hdr->count > num_irqs)
1545 		return -EINVAL;
1546 
1547 	switch (hdr->flags & VFIO_IRQ_SET_DATA_TYPE_MASK) {
1548 	case VFIO_IRQ_SET_DATA_NONE:
1549 		size = 0;
1550 		break;
1551 	case VFIO_IRQ_SET_DATA_BOOL:
1552 		size = sizeof(uint8_t);
1553 		break;
1554 	case VFIO_IRQ_SET_DATA_EVENTFD:
1555 		size = sizeof(int32_t);
1556 		break;
1557 	default:
1558 		return -EINVAL;
1559 	}
1560 
1561 	if (size) {
1562 		if (hdr->argsz - minsz < hdr->count * size)
1563 			return -EINVAL;
1564 
1565 		if (!data_size)
1566 			return -EINVAL;
1567 
1568 		*data_size = hdr->count * size;
1569 	}
1570 
1571 	return 0;
1572 }
1573 EXPORT_SYMBOL(vfio_set_irqs_validate_and_prepare);
1574 
1575 /*
1576  * Pin contiguous user pages and return their associated host pages for local
1577  * domain only.
1578  * @device [in]  : device
1579  * @iova [in]    : starting IOVA of user pages to be pinned.
1580  * @npage [in]   : count of pages to be pinned.  This count should not
1581  *		   be greater than VFIO_PIN_PAGES_MAX_ENTRIES.
1582  * @prot [in]    : protection flags
1583  * @pages[out]   : array of host pages
1584  * Return error or number of pages pinned.
1585  *
1586  * A driver may only call this function if the vfio_device was created
1587  * by vfio_register_emulated_iommu_dev() due to vfio_device_container_pin_pages().
1588  */
1589 int vfio_pin_pages(struct vfio_device *device, dma_addr_t iova,
1590 		   int npage, int prot, struct page **pages)
1591 {
1592 	/* group->container cannot change while a vfio device is open */
1593 	if (!pages || !npage || WARN_ON(!vfio_assert_device_open(device)))
1594 		return -EINVAL;
1595 	if (!device->ops->dma_unmap)
1596 		return -EINVAL;
1597 	if (vfio_device_has_container(device))
1598 		return vfio_device_container_pin_pages(device, iova,
1599 						       npage, prot, pages);
1600 	if (device->iommufd_access) {
1601 		int ret;
1602 
1603 		if (iova > ULONG_MAX)
1604 			return -EINVAL;
1605 		/*
1606 		 * VFIO ignores the sub page offset, npages is from the start of
1607 		 * a PAGE_SIZE chunk of IOVA. The caller is expected to recover
1608 		 * the sub page offset by doing:
1609 		 *     pages[0] + (iova % PAGE_SIZE)
1610 		 */
1611 		ret = iommufd_access_pin_pages(
1612 			device->iommufd_access, ALIGN_DOWN(iova, PAGE_SIZE),
1613 			npage * PAGE_SIZE, pages,
1614 			(prot & IOMMU_WRITE) ? IOMMUFD_ACCESS_RW_WRITE : 0);
1615 		if (ret)
1616 			return ret;
1617 		return npage;
1618 	}
1619 	return -EINVAL;
1620 }
1621 EXPORT_SYMBOL(vfio_pin_pages);
1622 
1623 /*
1624  * Unpin contiguous host pages for local domain only.
1625  * @device [in]  : device
1626  * @iova [in]    : starting address of user pages to be unpinned.
1627  * @npage [in]   : count of pages to be unpinned.  This count should not
1628  *                 be greater than VFIO_PIN_PAGES_MAX_ENTRIES.
1629  */
1630 void vfio_unpin_pages(struct vfio_device *device, dma_addr_t iova, int npage)
1631 {
1632 	if (WARN_ON(!vfio_assert_device_open(device)))
1633 		return;
1634 	if (WARN_ON(!device->ops->dma_unmap))
1635 		return;
1636 
1637 	if (vfio_device_has_container(device)) {
1638 		vfio_device_container_unpin_pages(device, iova, npage);
1639 		return;
1640 	}
1641 	if (device->iommufd_access) {
1642 		if (WARN_ON(iova > ULONG_MAX))
1643 			return;
1644 		iommufd_access_unpin_pages(device->iommufd_access,
1645 					   ALIGN_DOWN(iova, PAGE_SIZE),
1646 					   npage * PAGE_SIZE);
1647 		return;
1648 	}
1649 }
1650 EXPORT_SYMBOL(vfio_unpin_pages);
1651 
1652 /*
1653  * This interface allows the CPUs to perform some sort of virtual DMA on
1654  * behalf of the device.
1655  *
1656  * CPUs read/write from/into a range of IOVAs pointing to user space memory
1657  * into/from a kernel buffer.
1658  *
1659  * As the read/write of user space memory is conducted via the CPUs and is
1660  * not a real device DMA, it is not necessary to pin the user space memory.
1661  *
1662  * @device [in]		: VFIO device
1663  * @iova [in]		: base IOVA of a user space buffer
1664  * @data [in]		: pointer to kernel buffer
1665  * @len [in]		: kernel buffer length
1666  * @write		: indicate read or write
1667  * Return error code on failure or 0 on success.
1668  */
1669 int vfio_dma_rw(struct vfio_device *device, dma_addr_t iova, void *data,
1670 		size_t len, bool write)
1671 {
1672 	if (!data || len <= 0 || !vfio_assert_device_open(device))
1673 		return -EINVAL;
1674 
1675 	if (vfio_device_has_container(device))
1676 		return vfio_device_container_dma_rw(device, iova,
1677 						    data, len, write);
1678 
1679 	if (device->iommufd_access) {
1680 		unsigned int flags = 0;
1681 
1682 		if (iova > ULONG_MAX)
1683 			return -EINVAL;
1684 
1685 		/* VFIO historically tries to auto-detect a kthread */
1686 		if (!current->mm)
1687 			flags |= IOMMUFD_ACCESS_RW_KTHREAD;
1688 		if (write)
1689 			flags |= IOMMUFD_ACCESS_RW_WRITE;
1690 		return iommufd_access_rw(device->iommufd_access, iova, data,
1691 					 len, flags);
1692 	}
1693 	return -EINVAL;
1694 }
1695 EXPORT_SYMBOL(vfio_dma_rw);
1696 
1697 /*
1698  * Module/class support
1699  */
1700 static int __init vfio_init(void)
1701 {
1702 	int ret;
1703 
1704 	ida_init(&vfio.device_ida);
1705 
1706 	ret = vfio_group_init();
1707 	if (ret)
1708 		return ret;
1709 
1710 	ret = vfio_virqfd_init();
1711 	if (ret)
1712 		goto err_virqfd;
1713 
1714 	/* /sys/class/vfio-dev/vfioX */
1715 	vfio.device_class = class_create("vfio-dev");
1716 	if (IS_ERR(vfio.device_class)) {
1717 		ret = PTR_ERR(vfio.device_class);
1718 		goto err_dev_class;
1719 	}
1720 
1721 	ret = vfio_cdev_init(vfio.device_class);
1722 	if (ret)
1723 		goto err_alloc_dev_chrdev;
1724 
1725 	vfio_debugfs_create_root();
1726 	pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n");
1727 	return 0;
1728 
1729 err_alloc_dev_chrdev:
1730 	class_destroy(vfio.device_class);
1731 	vfio.device_class = NULL;
1732 err_dev_class:
1733 	vfio_virqfd_exit();
1734 err_virqfd:
1735 	vfio_group_cleanup();
1736 	return ret;
1737 }
1738 
1739 static void __exit vfio_cleanup(void)
1740 {
1741 	vfio_debugfs_remove_root();
1742 	ida_destroy(&vfio.device_ida);
1743 	vfio_cdev_cleanup();
1744 	class_destroy(vfio.device_class);
1745 	vfio.device_class = NULL;
1746 	vfio_virqfd_exit();
1747 	vfio_group_cleanup();
1748 	xa_destroy(&vfio_device_set_xa);
1749 }
1750 
1751 module_init(vfio_init);
1752 module_exit(vfio_cleanup);
1753 
1754 MODULE_IMPORT_NS(IOMMUFD);
1755 MODULE_VERSION(DRIVER_VERSION);
1756 MODULE_LICENSE("GPL v2");
1757 MODULE_AUTHOR(DRIVER_AUTHOR);
1758 MODULE_DESCRIPTION(DRIVER_DESC);
1759 MODULE_SOFTDEP("post: vfio_iommu_type1 vfio_iommu_spapr_tce");
1760