xref: /linux/drivers/vhost/vdpa.c (revision f0cf7ffc) 
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2018-2020 Intel Corporation.
4  * Copyright (C) 2020 Red Hat, Inc.
5  *
6  * Author: Tiwei Bie <tiwei.bie@intel.com>
7  *         Jason Wang <jasowang@redhat.com>
8  *
9  * Thanks Michael S. Tsirkin for the valuable comments and
10  * suggestions.  And thanks to Cunming Liang and Zhihong Wang for all
11  * their supports.
12  */
13 
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/cdev.h>
17 #include <linux/device.h>
18 #include <linux/mm.h>
19 #include <linux/slab.h>
20 #include <linux/iommu.h>
21 #include <linux/uuid.h>
22 #include <linux/vdpa.h>
23 #include <linux/nospec.h>
24 #include <linux/vhost.h>
25 
26 #include "vhost.h"
27 
28 enum {
29 	VHOST_VDPA_BACKEND_FEATURES =
30 	(1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2) |
31 	(1ULL << VHOST_BACKEND_F_IOTLB_BATCH) |
32 	(1ULL << VHOST_BACKEND_F_IOTLB_ASID),
33 };
34 
35 #define VHOST_VDPA_DEV_MAX (1U << MINORBITS)
36 
37 #define VHOST_VDPA_IOTLB_BUCKETS 16
38 
39 struct vhost_vdpa_as {
40 	struct hlist_node hash_link;
41 	struct vhost_iotlb iotlb;
42 	u32 id;
43 };
44 
45 struct vhost_vdpa {
46 	struct vhost_dev vdev;
47 	struct iommu_domain *domain;
48 	struct vhost_virtqueue *vqs;
49 	struct completion completion;
50 	struct vdpa_device *vdpa;
51 	struct hlist_head as[VHOST_VDPA_IOTLB_BUCKETS];
52 	struct device dev;
53 	struct cdev cdev;
54 	atomic_t opened;
55 	u32 nvqs;
56 	int virtio_id;
57 	int minor;
58 	struct eventfd_ctx *config_ctx;
59 	int in_batch;
60 	struct vdpa_iova_range range;
61 	u32 batch_asid;
62 	bool suspended;
63 };
64 
65 static DEFINE_IDA(vhost_vdpa_ida);
66 
67 static dev_t vhost_vdpa_major;
68 
69 static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v,
70 				   struct vhost_iotlb *iotlb, u64 start,
71 				   u64 last, u32 asid);
72 
73 static inline u32 iotlb_to_asid(struct vhost_iotlb *iotlb)
74 {
75 	struct vhost_vdpa_as *as = container_of(iotlb, struct
76 						vhost_vdpa_as, iotlb);
77 	return as->id;
78 }
79 
80 static struct vhost_vdpa_as *asid_to_as(struct vhost_vdpa *v, u32 asid)
81 {
82 	struct hlist_head *head = &v->as[asid % VHOST_VDPA_IOTLB_BUCKETS];
83 	struct vhost_vdpa_as *as;
84 
85 	hlist_for_each_entry(as, head, hash_link)
86 		if (as->id == asid)
87 			return as;
88 
89 	return NULL;
90 }
91 
92 static struct vhost_iotlb *asid_to_iotlb(struct vhost_vdpa *v, u32 asid)
93 {
94 	struct vhost_vdpa_as *as = asid_to_as(v, asid);
95 
96 	if (!as)
97 		return NULL;
98 
99 	return &as->iotlb;
100 }
101 
102 static struct vhost_vdpa_as *vhost_vdpa_alloc_as(struct vhost_vdpa *v, u32 asid)
103 {
104 	struct hlist_head *head = &v->as[asid % VHOST_VDPA_IOTLB_BUCKETS];
105 	struct vhost_vdpa_as *as;
106 
107 	if (asid_to_as(v, asid))
108 		return NULL;
109 
110 	if (asid >= v->vdpa->nas)
111 		return NULL;
112 
113 	as = kmalloc(sizeof(*as), GFP_KERNEL);
114 	if (!as)
115 		return NULL;
116 
117 	vhost_iotlb_init(&as->iotlb, 0, 0);
118 	as->id = asid;
119 	hlist_add_head(&as->hash_link, head);
120 
121 	return as;
122 }
123 
124 static struct vhost_vdpa_as *vhost_vdpa_find_alloc_as(struct vhost_vdpa *v,
125 						      u32 asid)
126 {
127 	struct vhost_vdpa_as *as = asid_to_as(v, asid);
128 
129 	if (as)
130 		return as;
131 
132 	return vhost_vdpa_alloc_as(v, asid);
133 }
134 
135 static void vhost_vdpa_reset_map(struct vhost_vdpa *v, u32 asid)
136 {
137 	struct vdpa_device *vdpa = v->vdpa;
138 	const struct vdpa_config_ops *ops = vdpa->config;
139 
140 	if (ops->reset_map)
141 		ops->reset_map(vdpa, asid);
142 }
143 
144 static int vhost_vdpa_remove_as(struct vhost_vdpa *v, u32 asid)
145 {
146 	struct vhost_vdpa_as *as = asid_to_as(v, asid);
147 
148 	if (!as)
149 		return -EINVAL;
150 
151 	hlist_del(&as->hash_link);
152 	vhost_vdpa_iotlb_unmap(v, &as->iotlb, 0ULL, 0ULL - 1, asid);
153 	/*
154 	 * Devices with vendor specific IOMMU may need to restore
155 	 * iotlb to the initial or default state, which cannot be
156 	 * cleaned up in the all range unmap call above. Give them
157 	 * a chance to clean up or reset the map to the desired
158 	 * state.
159 	 */
160 	vhost_vdpa_reset_map(v, asid);
161 	kfree(as);
162 
163 	return 0;
164 }
165 
166 static void handle_vq_kick(struct vhost_work *work)
167 {
168 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
169 						  poll.work);
170 	struct vhost_vdpa *v = container_of(vq->dev, struct vhost_vdpa, vdev);
171 	const struct vdpa_config_ops *ops = v->vdpa->config;
172 
173 	ops->kick_vq(v->vdpa, vq - v->vqs);
174 }
175 
176 static irqreturn_t vhost_vdpa_virtqueue_cb(void *private)
177 {
178 	struct vhost_virtqueue *vq = private;
179 	struct eventfd_ctx *call_ctx = vq->call_ctx.ctx;
180 
181 	if (call_ctx)
182 		eventfd_signal(call_ctx);
183 
184 	return IRQ_HANDLED;
185 }
186 
187 static irqreturn_t vhost_vdpa_config_cb(void *private)
188 {
189 	struct vhost_vdpa *v = private;
190 	struct eventfd_ctx *config_ctx = v->config_ctx;
191 
192 	if (config_ctx)
193 		eventfd_signal(config_ctx);
194 
195 	return IRQ_HANDLED;
196 }
197 
198 static void vhost_vdpa_setup_vq_irq(struct vhost_vdpa *v, u16 qid)
199 {
200 	struct vhost_virtqueue *vq = &v->vqs[qid];
201 	const struct vdpa_config_ops *ops = v->vdpa->config;
202 	struct vdpa_device *vdpa = v->vdpa;
203 	int ret, irq;
204 
205 	if (!ops->get_vq_irq)
206 		return;
207 
208 	irq = ops->get_vq_irq(vdpa, qid);
209 	if (irq < 0)
210 		return;
211 
212 	irq_bypass_unregister_producer(&vq->call_ctx.producer);
213 	if (!vq->call_ctx.ctx)
214 		return;
215 
216 	vq->call_ctx.producer.token = vq->call_ctx.ctx;
217 	vq->call_ctx.producer.irq = irq;
218 	ret = irq_bypass_register_producer(&vq->call_ctx.producer);
219 	if (unlikely(ret))
220 		dev_info(&v->dev, "vq %u, irq bypass producer (token %p) registration fails, ret =  %d\n",
221 			 qid, vq->call_ctx.producer.token, ret);
222 }
223 
224 static void vhost_vdpa_unsetup_vq_irq(struct vhost_vdpa *v, u16 qid)
225 {
226 	struct vhost_virtqueue *vq = &v->vqs[qid];
227 
228 	irq_bypass_unregister_producer(&vq->call_ctx.producer);
229 }
230 
231 static int _compat_vdpa_reset(struct vhost_vdpa *v)
232 {
233 	struct vdpa_device *vdpa = v->vdpa;
234 	u32 flags = 0;
235 
236 	v->suspended = false;
237 
238 	if (v->vdev.vqs) {
239 		flags |= !vhost_backend_has_feature(v->vdev.vqs[0],
240 						    VHOST_BACKEND_F_IOTLB_PERSIST) ?
241 			 VDPA_RESET_F_CLEAN_MAP : 0;
242 	}
243 
244 	return vdpa_reset(vdpa, flags);
245 }
246 
247 static int vhost_vdpa_reset(struct vhost_vdpa *v)
248 {
249 	v->in_batch = 0;
250 	return _compat_vdpa_reset(v);
251 }
252 
253 static long vhost_vdpa_bind_mm(struct vhost_vdpa *v)
254 {
255 	struct vdpa_device *vdpa = v->vdpa;
256 	const struct vdpa_config_ops *ops = vdpa->config;
257 
258 	if (!vdpa->use_va || !ops->bind_mm)
259 		return 0;
260 
261 	return ops->bind_mm(vdpa, v->vdev.mm);
262 }
263 
264 static void vhost_vdpa_unbind_mm(struct vhost_vdpa *v)
265 {
266 	struct vdpa_device *vdpa = v->vdpa;
267 	const struct vdpa_config_ops *ops = vdpa->config;
268 
269 	if (!vdpa->use_va || !ops->unbind_mm)
270 		return;
271 
272 	ops->unbind_mm(vdpa);
273 }
274 
275 static long vhost_vdpa_get_device_id(struct vhost_vdpa *v, u8 __user *argp)
276 {
277 	struct vdpa_device *vdpa = v->vdpa;
278 	const struct vdpa_config_ops *ops = vdpa->config;
279 	u32 device_id;
280 
281 	device_id = ops->get_device_id(vdpa);
282 
283 	if (copy_to_user(argp, &device_id, sizeof(device_id)))
284 		return -EFAULT;
285 
286 	return 0;
287 }
288 
289 static long vhost_vdpa_get_status(struct vhost_vdpa *v, u8 __user *statusp)
290 {
291 	struct vdpa_device *vdpa = v->vdpa;
292 	const struct vdpa_config_ops *ops = vdpa->config;
293 	u8 status;
294 
295 	status = ops->get_status(vdpa);
296 
297 	if (copy_to_user(statusp, &status, sizeof(status)))
298 		return -EFAULT;
299 
300 	return 0;
301 }
302 
303 static long vhost_vdpa_set_status(struct vhost_vdpa *v, u8 __user *statusp)
304 {
305 	struct vdpa_device *vdpa = v->vdpa;
306 	const struct vdpa_config_ops *ops = vdpa->config;
307 	u8 status, status_old;
308 	u32 nvqs = v->nvqs;
309 	int ret;
310 	u16 i;
311 
312 	if (copy_from_user(&status, statusp, sizeof(status)))
313 		return -EFAULT;
314 
315 	status_old = ops->get_status(vdpa);
316 
317 	/*
318 	 * Userspace shouldn't remove status bits unless reset the
319 	 * status to 0.
320 	 */
321 	if (status != 0 && (status_old & ~status) != 0)
322 		return -EINVAL;
323 
324 	if ((status_old & VIRTIO_CONFIG_S_DRIVER_OK) && !(status & VIRTIO_CONFIG_S_DRIVER_OK))
325 		for (i = 0; i < nvqs; i++)
326 			vhost_vdpa_unsetup_vq_irq(v, i);
327 
328 	if (status == 0) {
329 		ret = _compat_vdpa_reset(v);
330 		if (ret)
331 			return ret;
332 	} else
333 		vdpa_set_status(vdpa, status);
334 
335 	if ((status & VIRTIO_CONFIG_S_DRIVER_OK) && !(status_old & VIRTIO_CONFIG_S_DRIVER_OK))
336 		for (i = 0; i < nvqs; i++)
337 			vhost_vdpa_setup_vq_irq(v, i);
338 
339 	return 0;
340 }
341 
342 static int vhost_vdpa_config_validate(struct vhost_vdpa *v,
343 				      struct vhost_vdpa_config *c)
344 {
345 	struct vdpa_device *vdpa = v->vdpa;
346 	size_t size = vdpa->config->get_config_size(vdpa);
347 
348 	if (c->len == 0 || c->off > size)
349 		return -EINVAL;
350 
351 	if (c->len > size - c->off)
352 		return -E2BIG;
353 
354 	return 0;
355 }
356 
357 static long vhost_vdpa_get_config(struct vhost_vdpa *v,
358 				  struct vhost_vdpa_config __user *c)
359 {
360 	struct vdpa_device *vdpa = v->vdpa;
361 	struct vhost_vdpa_config config;
362 	unsigned long size = offsetof(struct vhost_vdpa_config, buf);
363 	u8 *buf;
364 
365 	if (copy_from_user(&config, c, size))
366 		return -EFAULT;
367 	if (vhost_vdpa_config_validate(v, &config))
368 		return -EINVAL;
369 	buf = kvzalloc(config.len, GFP_KERNEL);
370 	if (!buf)
371 		return -ENOMEM;
372 
373 	vdpa_get_config(vdpa, config.off, buf, config.len);
374 
375 	if (copy_to_user(c->buf, buf, config.len)) {
376 		kvfree(buf);
377 		return -EFAULT;
378 	}
379 
380 	kvfree(buf);
381 	return 0;
382 }
383 
384 static long vhost_vdpa_set_config(struct vhost_vdpa *v,
385 				  struct vhost_vdpa_config __user *c)
386 {
387 	struct vdpa_device *vdpa = v->vdpa;
388 	struct vhost_vdpa_config config;
389 	unsigned long size = offsetof(struct vhost_vdpa_config, buf);
390 	u8 *buf;
391 
392 	if (copy_from_user(&config, c, size))
393 		return -EFAULT;
394 	if (vhost_vdpa_config_validate(v, &config))
395 		return -EINVAL;
396 
397 	buf = vmemdup_user(c->buf, config.len);
398 	if (IS_ERR(buf))
399 		return PTR_ERR(buf);
400 
401 	vdpa_set_config(vdpa, config.off, buf, config.len);
402 
403 	kvfree(buf);
404 	return 0;
405 }
406 
407 static bool vhost_vdpa_can_suspend(const struct vhost_vdpa *v)
408 {
409 	struct vdpa_device *vdpa = v->vdpa;
410 	const struct vdpa_config_ops *ops = vdpa->config;
411 
412 	return ops->suspend;
413 }
414 
415 static bool vhost_vdpa_can_resume(const struct vhost_vdpa *v)
416 {
417 	struct vdpa_device *vdpa = v->vdpa;
418 	const struct vdpa_config_ops *ops = vdpa->config;
419 
420 	return ops->resume;
421 }
422 
423 static bool vhost_vdpa_has_desc_group(const struct vhost_vdpa *v)
424 {
425 	struct vdpa_device *vdpa = v->vdpa;
426 	const struct vdpa_config_ops *ops = vdpa->config;
427 
428 	return ops->get_vq_desc_group;
429 }
430 
431 static long vhost_vdpa_get_features(struct vhost_vdpa *v, u64 __user *featurep)
432 {
433 	struct vdpa_device *vdpa = v->vdpa;
434 	const struct vdpa_config_ops *ops = vdpa->config;
435 	u64 features;
436 
437 	features = ops->get_device_features(vdpa);
438 
439 	if (copy_to_user(featurep, &features, sizeof(features)))
440 		return -EFAULT;
441 
442 	return 0;
443 }
444 
445 static u64 vhost_vdpa_get_backend_features(const struct vhost_vdpa *v)
446 {
447 	struct vdpa_device *vdpa = v->vdpa;
448 	const struct vdpa_config_ops *ops = vdpa->config;
449 
450 	if (!ops->get_backend_features)
451 		return 0;
452 	else
453 		return ops->get_backend_features(vdpa);
454 }
455 
456 static bool vhost_vdpa_has_persistent_map(const struct vhost_vdpa *v)
457 {
458 	struct vdpa_device *vdpa = v->vdpa;
459 	const struct vdpa_config_ops *ops = vdpa->config;
460 
461 	return (!ops->set_map && !ops->dma_map) || ops->reset_map ||
462 	       vhost_vdpa_get_backend_features(v) & BIT_ULL(VHOST_BACKEND_F_IOTLB_PERSIST);
463 }
464 
465 static long vhost_vdpa_set_features(struct vhost_vdpa *v, u64 __user *featurep)
466 {
467 	struct vdpa_device *vdpa = v->vdpa;
468 	const struct vdpa_config_ops *ops = vdpa->config;
469 	struct vhost_dev *d = &v->vdev;
470 	u64 actual_features;
471 	u64 features;
472 	int i;
473 
474 	/*
475 	 * It's not allowed to change the features after they have
476 	 * been negotiated.
477 	 */
478 	if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_FEATURES_OK)
479 		return -EBUSY;
480 
481 	if (copy_from_user(&features, featurep, sizeof(features)))
482 		return -EFAULT;
483 
484 	if (vdpa_set_features(vdpa, features))
485 		return -EINVAL;
486 
487 	/* let the vqs know what has been configured */
488 	actual_features = ops->get_driver_features(vdpa);
489 	for (i = 0; i < d->nvqs; ++i) {
490 		struct vhost_virtqueue *vq = d->vqs[i];
491 
492 		mutex_lock(&vq->mutex);
493 		vq->acked_features = actual_features;
494 		mutex_unlock(&vq->mutex);
495 	}
496 
497 	return 0;
498 }
499 
500 static long vhost_vdpa_get_vring_num(struct vhost_vdpa *v, u16 __user *argp)
501 {
502 	struct vdpa_device *vdpa = v->vdpa;
503 	const struct vdpa_config_ops *ops = vdpa->config;
504 	u16 num;
505 
506 	num = ops->get_vq_num_max(vdpa);
507 
508 	if (copy_to_user(argp, &num, sizeof(num)))
509 		return -EFAULT;
510 
511 	return 0;
512 }
513 
514 static void vhost_vdpa_config_put(struct vhost_vdpa *v)
515 {
516 	if (v->config_ctx) {
517 		eventfd_ctx_put(v->config_ctx);
518 		v->config_ctx = NULL;
519 	}
520 }
521 
522 static long vhost_vdpa_set_config_call(struct vhost_vdpa *v, u32 __user *argp)
523 {
524 	struct vdpa_callback cb;
525 	int fd;
526 	struct eventfd_ctx *ctx;
527 
528 	cb.callback = vhost_vdpa_config_cb;
529 	cb.private = v;
530 	if (copy_from_user(&fd, argp, sizeof(fd)))
531 		return  -EFAULT;
532 
533 	ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
534 	swap(ctx, v->config_ctx);
535 
536 	if (!IS_ERR_OR_NULL(ctx))
537 		eventfd_ctx_put(ctx);
538 
539 	if (IS_ERR(v->config_ctx)) {
540 		long ret = PTR_ERR(v->config_ctx);
541 
542 		v->config_ctx = NULL;
543 		return ret;
544 	}
545 
546 	v->vdpa->config->set_config_cb(v->vdpa, &cb);
547 
548 	return 0;
549 }
550 
551 static long vhost_vdpa_get_iova_range(struct vhost_vdpa *v, u32 __user *argp)
552 {
553 	struct vhost_vdpa_iova_range range = {
554 		.first = v->range.first,
555 		.last = v->range.last,
556 	};
557 
558 	if (copy_to_user(argp, &range, sizeof(range)))
559 		return -EFAULT;
560 	return 0;
561 }
562 
563 static long vhost_vdpa_get_config_size(struct vhost_vdpa *v, u32 __user *argp)
564 {
565 	struct vdpa_device *vdpa = v->vdpa;
566 	const struct vdpa_config_ops *ops = vdpa->config;
567 	u32 size;
568 
569 	size = ops->get_config_size(vdpa);
570 
571 	if (copy_to_user(argp, &size, sizeof(size)))
572 		return -EFAULT;
573 
574 	return 0;
575 }
576 
577 static long vhost_vdpa_get_vqs_count(struct vhost_vdpa *v, u32 __user *argp)
578 {
579 	struct vdpa_device *vdpa = v->vdpa;
580 
581 	if (copy_to_user(argp, &vdpa->nvqs, sizeof(vdpa->nvqs)))
582 		return -EFAULT;
583 
584 	return 0;
585 }
586 
587 /* After a successful return of ioctl the device must not process more
588  * virtqueue descriptors. The device can answer to read or writes of config
589  * fields as if it were not suspended. In particular, writing to "queue_enable"
590  * with a value of 1 will not make the device start processing buffers.
591  */
592 static long vhost_vdpa_suspend(struct vhost_vdpa *v)
593 {
594 	struct vdpa_device *vdpa = v->vdpa;
595 	const struct vdpa_config_ops *ops = vdpa->config;
596 	int ret;
597 
598 	if (!(ops->get_status(vdpa) & VIRTIO_CONFIG_S_DRIVER_OK))
599 		return 0;
600 
601 	if (!ops->suspend)
602 		return -EOPNOTSUPP;
603 
604 	ret = ops->suspend(vdpa);
605 	if (!ret)
606 		v->suspended = true;
607 
608 	return ret;
609 }
610 
611 /* After a successful return of this ioctl the device resumes processing
612  * virtqueue descriptors. The device becomes fully operational the same way it
613  * was before it was suspended.
614  */
615 static long vhost_vdpa_resume(struct vhost_vdpa *v)
616 {
617 	struct vdpa_device *vdpa = v->vdpa;
618 	const struct vdpa_config_ops *ops = vdpa->config;
619 	int ret;
620 
621 	if (!(ops->get_status(vdpa) & VIRTIO_CONFIG_S_DRIVER_OK))
622 		return 0;
623 
624 	if (!ops->resume)
625 		return -EOPNOTSUPP;
626 
627 	ret = ops->resume(vdpa);
628 	if (!ret)
629 		v->suspended = false;
630 
631 	return ret;
632 }
633 
634 static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
635 				   void __user *argp)
636 {
637 	struct vdpa_device *vdpa = v->vdpa;
638 	const struct vdpa_config_ops *ops = vdpa->config;
639 	struct vdpa_vq_state vq_state;
640 	struct vdpa_callback cb;
641 	struct vhost_virtqueue *vq;
642 	struct vhost_vring_state s;
643 	u32 idx;
644 	long r;
645 
646 	r = get_user(idx, (u32 __user *)argp);
647 	if (r < 0)
648 		return r;
649 
650 	if (idx >= v->nvqs)
651 		return -ENOBUFS;
652 
653 	idx = array_index_nospec(idx, v->nvqs);
654 	vq = &v->vqs[idx];
655 
656 	switch (cmd) {
657 	case VHOST_VDPA_SET_VRING_ENABLE:
658 		if (copy_from_user(&s, argp, sizeof(s)))
659 			return -EFAULT;
660 		ops->set_vq_ready(vdpa, idx, s.num);
661 		return 0;
662 	case VHOST_VDPA_GET_VRING_GROUP:
663 		if (!ops->get_vq_group)
664 			return -EOPNOTSUPP;
665 		s.index = idx;
666 		s.num = ops->get_vq_group(vdpa, idx);
667 		if (s.num >= vdpa->ngroups)
668 			return -EIO;
669 		else if (copy_to_user(argp, &s, sizeof(s)))
670 			return -EFAULT;
671 		return 0;
672 	case VHOST_VDPA_GET_VRING_DESC_GROUP:
673 		if (!vhost_vdpa_has_desc_group(v))
674 			return -EOPNOTSUPP;
675 		s.index = idx;
676 		s.num = ops->get_vq_desc_group(vdpa, idx);
677 		if (s.num >= vdpa->ngroups)
678 			return -EIO;
679 		else if (copy_to_user(argp, &s, sizeof(s)))
680 			return -EFAULT;
681 		return 0;
682 	case VHOST_VDPA_SET_GROUP_ASID:
683 		if (copy_from_user(&s, argp, sizeof(s)))
684 			return -EFAULT;
685 		if (s.num >= vdpa->nas)
686 			return -EINVAL;
687 		if (!ops->set_group_asid)
688 			return -EOPNOTSUPP;
689 		return ops->set_group_asid(vdpa, idx, s.num);
690 	case VHOST_VDPA_GET_VRING_SIZE:
691 		if (!ops->get_vq_size)
692 			return -EOPNOTSUPP;
693 		s.index = idx;
694 		s.num = ops->get_vq_size(vdpa, idx);
695 		if (copy_to_user(argp, &s, sizeof(s)))
696 			return -EFAULT;
697 		return 0;
698 	case VHOST_GET_VRING_BASE:
699 		r = ops->get_vq_state(v->vdpa, idx, &vq_state);
700 		if (r)
701 			return r;
702 
703 		if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
704 			vq->last_avail_idx = vq_state.packed.last_avail_idx |
705 					     (vq_state.packed.last_avail_counter << 15);
706 			vq->last_used_idx = vq_state.packed.last_used_idx |
707 					    (vq_state.packed.last_used_counter << 15);
708 		} else {
709 			vq->last_avail_idx = vq_state.split.avail_index;
710 		}
711 		break;
712 	}
713 
714 	r = vhost_vring_ioctl(&v->vdev, cmd, argp);
715 	if (r)
716 		return r;
717 
718 	switch (cmd) {
719 	case VHOST_SET_VRING_ADDR:
720 		if ((ops->get_status(vdpa) & VIRTIO_CONFIG_S_DRIVER_OK) && !v->suspended)
721 			return -EINVAL;
722 
723 		if (ops->set_vq_address(vdpa, idx,
724 					(u64)(uintptr_t)vq->desc,
725 					(u64)(uintptr_t)vq->avail,
726 					(u64)(uintptr_t)vq->used))
727 			r = -EINVAL;
728 		break;
729 
730 	case VHOST_SET_VRING_BASE:
731 		if ((ops->get_status(vdpa) & VIRTIO_CONFIG_S_DRIVER_OK) && !v->suspended)
732 			return -EINVAL;
733 
734 		if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
735 			vq_state.packed.last_avail_idx = vq->last_avail_idx & 0x7fff;
736 			vq_state.packed.last_avail_counter = !!(vq->last_avail_idx & 0x8000);
737 			vq_state.packed.last_used_idx = vq->last_used_idx & 0x7fff;
738 			vq_state.packed.last_used_counter = !!(vq->last_used_idx & 0x8000);
739 		} else {
740 			vq_state.split.avail_index = vq->last_avail_idx;
741 		}
742 		r = ops->set_vq_state(vdpa, idx, &vq_state);
743 		break;
744 
745 	case VHOST_SET_VRING_CALL:
746 		if (vq->call_ctx.ctx) {
747 			cb.callback = vhost_vdpa_virtqueue_cb;
748 			cb.private = vq;
749 			cb.trigger = vq->call_ctx.ctx;
750 		} else {
751 			cb.callback = NULL;
752 			cb.private = NULL;
753 			cb.trigger = NULL;
754 		}
755 		ops->set_vq_cb(vdpa, idx, &cb);
756 		vhost_vdpa_setup_vq_irq(v, idx);
757 		break;
758 
759 	case VHOST_SET_VRING_NUM:
760 		ops->set_vq_num(vdpa, idx, vq->num);
761 		break;
762 	}
763 
764 	return r;
765 }
766 
767 static long vhost_vdpa_unlocked_ioctl(struct file *filep,
768 				      unsigned int cmd, unsigned long arg)
769 {
770 	struct vhost_vdpa *v = filep->private_data;
771 	struct vhost_dev *d = &v->vdev;
772 	void __user *argp = (void __user *)arg;
773 	u64 __user *featurep = argp;
774 	u64 features;
775 	long r = 0;
776 
777 	if (cmd == VHOST_SET_BACKEND_FEATURES) {
778 		if (copy_from_user(&features, featurep, sizeof(features)))
779 			return -EFAULT;
780 		if (features & ~(VHOST_VDPA_BACKEND_FEATURES |
781 				 BIT_ULL(VHOST_BACKEND_F_DESC_ASID) |
782 				 BIT_ULL(VHOST_BACKEND_F_IOTLB_PERSIST) |
783 				 BIT_ULL(VHOST_BACKEND_F_SUSPEND) |
784 				 BIT_ULL(VHOST_BACKEND_F_RESUME) |
785 				 BIT_ULL(VHOST_BACKEND_F_ENABLE_AFTER_DRIVER_OK)))
786 			return -EOPNOTSUPP;
787 		if ((features & BIT_ULL(VHOST_BACKEND_F_SUSPEND)) &&
788 		     !vhost_vdpa_can_suspend(v))
789 			return -EOPNOTSUPP;
790 		if ((features & BIT_ULL(VHOST_BACKEND_F_RESUME)) &&
791 		     !vhost_vdpa_can_resume(v))
792 			return -EOPNOTSUPP;
793 		if ((features & BIT_ULL(VHOST_BACKEND_F_DESC_ASID)) &&
794 		    !(features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID)))
795 			return -EINVAL;
796 		if ((features & BIT_ULL(VHOST_BACKEND_F_DESC_ASID)) &&
797 		     !vhost_vdpa_has_desc_group(v))
798 			return -EOPNOTSUPP;
799 		if ((features & BIT_ULL(VHOST_BACKEND_F_IOTLB_PERSIST)) &&
800 		     !vhost_vdpa_has_persistent_map(v))
801 			return -EOPNOTSUPP;
802 		vhost_set_backend_features(&v->vdev, features);
803 		return 0;
804 	}
805 
806 	mutex_lock(&d->mutex);
807 
808 	switch (cmd) {
809 	case VHOST_VDPA_GET_DEVICE_ID:
810 		r = vhost_vdpa_get_device_id(v, argp);
811 		break;
812 	case VHOST_VDPA_GET_STATUS:
813 		r = vhost_vdpa_get_status(v, argp);
814 		break;
815 	case VHOST_VDPA_SET_STATUS:
816 		r = vhost_vdpa_set_status(v, argp);
817 		break;
818 	case VHOST_VDPA_GET_CONFIG:
819 		r = vhost_vdpa_get_config(v, argp);
820 		break;
821 	case VHOST_VDPA_SET_CONFIG:
822 		r = vhost_vdpa_set_config(v, argp);
823 		break;
824 	case VHOST_GET_FEATURES:
825 		r = vhost_vdpa_get_features(v, argp);
826 		break;
827 	case VHOST_SET_FEATURES:
828 		r = vhost_vdpa_set_features(v, argp);
829 		break;
830 	case VHOST_VDPA_GET_VRING_NUM:
831 		r = vhost_vdpa_get_vring_num(v, argp);
832 		break;
833 	case VHOST_VDPA_GET_GROUP_NUM:
834 		if (copy_to_user(argp, &v->vdpa->ngroups,
835 				 sizeof(v->vdpa->ngroups)))
836 			r = -EFAULT;
837 		break;
838 	case VHOST_VDPA_GET_AS_NUM:
839 		if (copy_to_user(argp, &v->vdpa->nas, sizeof(v->vdpa->nas)))
840 			r = -EFAULT;
841 		break;
842 	case VHOST_SET_LOG_BASE:
843 	case VHOST_SET_LOG_FD:
844 		r = -ENOIOCTLCMD;
845 		break;
846 	case VHOST_VDPA_SET_CONFIG_CALL:
847 		r = vhost_vdpa_set_config_call(v, argp);
848 		break;
849 	case VHOST_GET_BACKEND_FEATURES:
850 		features = VHOST_VDPA_BACKEND_FEATURES;
851 		if (vhost_vdpa_can_suspend(v))
852 			features |= BIT_ULL(VHOST_BACKEND_F_SUSPEND);
853 		if (vhost_vdpa_can_resume(v))
854 			features |= BIT_ULL(VHOST_BACKEND_F_RESUME);
855 		if (vhost_vdpa_has_desc_group(v))
856 			features |= BIT_ULL(VHOST_BACKEND_F_DESC_ASID);
857 		if (vhost_vdpa_has_persistent_map(v))
858 			features |= BIT_ULL(VHOST_BACKEND_F_IOTLB_PERSIST);
859 		features |= vhost_vdpa_get_backend_features(v);
860 		if (copy_to_user(featurep, &features, sizeof(features)))
861 			r = -EFAULT;
862 		break;
863 	case VHOST_VDPA_GET_IOVA_RANGE:
864 		r = vhost_vdpa_get_iova_range(v, argp);
865 		break;
866 	case VHOST_VDPA_GET_CONFIG_SIZE:
867 		r = vhost_vdpa_get_config_size(v, argp);
868 		break;
869 	case VHOST_VDPA_GET_VQS_COUNT:
870 		r = vhost_vdpa_get_vqs_count(v, argp);
871 		break;
872 	case VHOST_VDPA_SUSPEND:
873 		r = vhost_vdpa_suspend(v);
874 		break;
875 	case VHOST_VDPA_RESUME:
876 		r = vhost_vdpa_resume(v);
877 		break;
878 	default:
879 		r = vhost_dev_ioctl(&v->vdev, cmd, argp);
880 		if (r == -ENOIOCTLCMD)
881 			r = vhost_vdpa_vring_ioctl(v, cmd, argp);
882 		break;
883 	}
884 
885 	if (r)
886 		goto out;
887 
888 	switch (cmd) {
889 	case VHOST_SET_OWNER:
890 		r = vhost_vdpa_bind_mm(v);
891 		if (r)
892 			vhost_dev_reset_owner(d, NULL);
893 		break;
894 	}
895 out:
896 	mutex_unlock(&d->mutex);
897 	return r;
898 }
899 static void vhost_vdpa_general_unmap(struct vhost_vdpa *v,
900 				     struct vhost_iotlb_map *map, u32 asid)
901 {
902 	struct vdpa_device *vdpa = v->vdpa;
903 	const struct vdpa_config_ops *ops = vdpa->config;
904 	if (ops->dma_map) {
905 		ops->dma_unmap(vdpa, asid, map->start, map->size);
906 	} else if (ops->set_map == NULL) {
907 		iommu_unmap(v->domain, map->start, map->size);
908 	}
909 }
910 
911 static void vhost_vdpa_pa_unmap(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
912 				u64 start, u64 last, u32 asid)
913 {
914 	struct vhost_dev *dev = &v->vdev;
915 	struct vhost_iotlb_map *map;
916 	struct page *page;
917 	unsigned long pfn, pinned;
918 
919 	while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
920 		pinned = PFN_DOWN(map->size);
921 		for (pfn = PFN_DOWN(map->addr);
922 		     pinned > 0; pfn++, pinned--) {
923 			page = pfn_to_page(pfn);
924 			if (map->perm & VHOST_ACCESS_WO)
925 				set_page_dirty_lock(page);
926 			unpin_user_page(page);
927 		}
928 		atomic64_sub(PFN_DOWN(map->size), &dev->mm->pinned_vm);
929 		vhost_vdpa_general_unmap(v, map, asid);
930 		vhost_iotlb_map_free(iotlb, map);
931 	}
932 }
933 
934 static void vhost_vdpa_va_unmap(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
935 				u64 start, u64 last, u32 asid)
936 {
937 	struct vhost_iotlb_map *map;
938 	struct vdpa_map_file *map_file;
939 
940 	while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
941 		map_file = (struct vdpa_map_file *)map->opaque;
942 		fput(map_file->file);
943 		kfree(map_file);
944 		vhost_vdpa_general_unmap(v, map, asid);
945 		vhost_iotlb_map_free(iotlb, map);
946 	}
947 }
948 
949 static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v,
950 				   struct vhost_iotlb *iotlb, u64 start,
951 				   u64 last, u32 asid)
952 {
953 	struct vdpa_device *vdpa = v->vdpa;
954 
955 	if (vdpa->use_va)
956 		return vhost_vdpa_va_unmap(v, iotlb, start, last, asid);
957 
958 	return vhost_vdpa_pa_unmap(v, iotlb, start, last, asid);
959 }
960 
961 static int perm_to_iommu_flags(u32 perm)
962 {
963 	int flags = 0;
964 
965 	switch (perm) {
966 	case VHOST_ACCESS_WO:
967 		flags |= IOMMU_WRITE;
968 		break;
969 	case VHOST_ACCESS_RO:
970 		flags |= IOMMU_READ;
971 		break;
972 	case VHOST_ACCESS_RW:
973 		flags |= (IOMMU_WRITE | IOMMU_READ);
974 		break;
975 	default:
976 		WARN(1, "invalidate vhost IOTLB permission\n");
977 		break;
978 	}
979 
980 	return flags | IOMMU_CACHE;
981 }
982 
983 static int vhost_vdpa_map(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
984 			  u64 iova, u64 size, u64 pa, u32 perm, void *opaque)
985 {
986 	struct vhost_dev *dev = &v->vdev;
987 	struct vdpa_device *vdpa = v->vdpa;
988 	const struct vdpa_config_ops *ops = vdpa->config;
989 	u32 asid = iotlb_to_asid(iotlb);
990 	int r = 0;
991 
992 	r = vhost_iotlb_add_range_ctx(iotlb, iova, iova + size - 1,
993 				      pa, perm, opaque);
994 	if (r)
995 		return r;
996 
997 	if (ops->dma_map) {
998 		r = ops->dma_map(vdpa, asid, iova, size, pa, perm, opaque);
999 	} else if (ops->set_map) {
1000 		if (!v->in_batch)
1001 			r = ops->set_map(vdpa, asid, iotlb);
1002 	} else {
1003 		r = iommu_map(v->domain, iova, pa, size,
1004 			      perm_to_iommu_flags(perm),
1005 			      GFP_KERNEL_ACCOUNT);
1006 	}
1007 	if (r) {
1008 		vhost_iotlb_del_range(iotlb, iova, iova + size - 1);
1009 		return r;
1010 	}
1011 
1012 	if (!vdpa->use_va)
1013 		atomic64_add(PFN_DOWN(size), &dev->mm->pinned_vm);
1014 
1015 	return 0;
1016 }
1017 
1018 static void vhost_vdpa_unmap(struct vhost_vdpa *v,
1019 			     struct vhost_iotlb *iotlb,
1020 			     u64 iova, u64 size)
1021 {
1022 	struct vdpa_device *vdpa = v->vdpa;
1023 	const struct vdpa_config_ops *ops = vdpa->config;
1024 	u32 asid = iotlb_to_asid(iotlb);
1025 
1026 	vhost_vdpa_iotlb_unmap(v, iotlb, iova, iova + size - 1, asid);
1027 
1028 	if (ops->set_map) {
1029 		if (!v->in_batch)
1030 			ops->set_map(vdpa, asid, iotlb);
1031 	}
1032 
1033 }
1034 
1035 static int vhost_vdpa_va_map(struct vhost_vdpa *v,
1036 			     struct vhost_iotlb *iotlb,
1037 			     u64 iova, u64 size, u64 uaddr, u32 perm)
1038 {
1039 	struct vhost_dev *dev = &v->vdev;
1040 	u64 offset, map_size, map_iova = iova;
1041 	struct vdpa_map_file *map_file;
1042 	struct vm_area_struct *vma;
1043 	int ret = 0;
1044 
1045 	mmap_read_lock(dev->mm);
1046 
1047 	while (size) {
1048 		vma = find_vma(dev->mm, uaddr);
1049 		if (!vma) {
1050 			ret = -EINVAL;
1051 			break;
1052 		}
1053 		map_size = min(size, vma->vm_end - uaddr);
1054 		if (!(vma->vm_file && (vma->vm_flags & VM_SHARED) &&
1055 			!(vma->vm_flags & (VM_IO | VM_PFNMAP))))
1056 			goto next;
1057 
1058 		map_file = kzalloc(sizeof(*map_file), GFP_KERNEL);
1059 		if (!map_file) {
1060 			ret = -ENOMEM;
1061 			break;
1062 		}
1063 		offset = (vma->vm_pgoff << PAGE_SHIFT) + uaddr - vma->vm_start;
1064 		map_file->offset = offset;
1065 		map_file->file = get_file(vma->vm_file);
1066 		ret = vhost_vdpa_map(v, iotlb, map_iova, map_size, uaddr,
1067 				     perm, map_file);
1068 		if (ret) {
1069 			fput(map_file->file);
1070 			kfree(map_file);
1071 			break;
1072 		}
1073 next:
1074 		size -= map_size;
1075 		uaddr += map_size;
1076 		map_iova += map_size;
1077 	}
1078 	if (ret)
1079 		vhost_vdpa_unmap(v, iotlb, iova, map_iova - iova);
1080 
1081 	mmap_read_unlock(dev->mm);
1082 
1083 	return ret;
1084 }
1085 
1086 static int vhost_vdpa_pa_map(struct vhost_vdpa *v,
1087 			     struct vhost_iotlb *iotlb,
1088 			     u64 iova, u64 size, u64 uaddr, u32 perm)
1089 {
1090 	struct vhost_dev *dev = &v->vdev;
1091 	struct page **page_list;
1092 	unsigned long list_size = PAGE_SIZE / sizeof(struct page *);
1093 	unsigned int gup_flags = FOLL_LONGTERM;
1094 	unsigned long npages, cur_base, map_pfn, last_pfn = 0;
1095 	unsigned long lock_limit, sz2pin, nchunks, i;
1096 	u64 start = iova;
1097 	long pinned;
1098 	int ret = 0;
1099 
1100 	/* Limit the use of memory for bookkeeping */
1101 	page_list = (struct page **) __get_free_page(GFP_KERNEL);
1102 	if (!page_list)
1103 		return -ENOMEM;
1104 
1105 	if (perm & VHOST_ACCESS_WO)
1106 		gup_flags |= FOLL_WRITE;
1107 
1108 	npages = PFN_UP(size + (iova & ~PAGE_MASK));
1109 	if (!npages) {
1110 		ret = -EINVAL;
1111 		goto free;
1112 	}
1113 
1114 	mmap_read_lock(dev->mm);
1115 
1116 	lock_limit = PFN_DOWN(rlimit(RLIMIT_MEMLOCK));
1117 	if (npages + atomic64_read(&dev->mm->pinned_vm) > lock_limit) {
1118 		ret = -ENOMEM;
1119 		goto unlock;
1120 	}
1121 
1122 	cur_base = uaddr & PAGE_MASK;
1123 	iova &= PAGE_MASK;
1124 	nchunks = 0;
1125 
1126 	while (npages) {
1127 		sz2pin = min_t(unsigned long, npages, list_size);
1128 		pinned = pin_user_pages(cur_base, sz2pin,
1129 					gup_flags, page_list);
1130 		if (sz2pin != pinned) {
1131 			if (pinned < 0) {
1132 				ret = pinned;
1133 			} else {
1134 				unpin_user_pages(page_list, pinned);
1135 				ret = -ENOMEM;
1136 			}
1137 			goto out;
1138 		}
1139 		nchunks++;
1140 
1141 		if (!last_pfn)
1142 			map_pfn = page_to_pfn(page_list[0]);
1143 
1144 		for (i = 0; i < pinned; i++) {
1145 			unsigned long this_pfn = page_to_pfn(page_list[i]);
1146 			u64 csize;
1147 
1148 			if (last_pfn && (this_pfn != last_pfn + 1)) {
1149 				/* Pin a contiguous chunk of memory */
1150 				csize = PFN_PHYS(last_pfn - map_pfn + 1);
1151 				ret = vhost_vdpa_map(v, iotlb, iova, csize,
1152 						     PFN_PHYS(map_pfn),
1153 						     perm, NULL);
1154 				if (ret) {
1155 					/*
1156 					 * Unpin the pages that are left unmapped
1157 					 * from this point on in the current
1158 					 * page_list. The remaining outstanding
1159 					 * ones which may stride across several
1160 					 * chunks will be covered in the common
1161 					 * error path subsequently.
1162 					 */
1163 					unpin_user_pages(&page_list[i],
1164 							 pinned - i);
1165 					goto out;
1166 				}
1167 
1168 				map_pfn = this_pfn;
1169 				iova += csize;
1170 				nchunks = 0;
1171 			}
1172 
1173 			last_pfn = this_pfn;
1174 		}
1175 
1176 		cur_base += PFN_PHYS(pinned);
1177 		npages -= pinned;
1178 	}
1179 
1180 	/* Pin the rest chunk */
1181 	ret = vhost_vdpa_map(v, iotlb, iova, PFN_PHYS(last_pfn - map_pfn + 1),
1182 			     PFN_PHYS(map_pfn), perm, NULL);
1183 out:
1184 	if (ret) {
1185 		if (nchunks) {
1186 			unsigned long pfn;
1187 
1188 			/*
1189 			 * Unpin the outstanding pages which are yet to be
1190 			 * mapped but haven't due to vdpa_map() or
1191 			 * pin_user_pages() failure.
1192 			 *
1193 			 * Mapped pages are accounted in vdpa_map(), hence
1194 			 * the corresponding unpinning will be handled by
1195 			 * vdpa_unmap().
1196 			 */
1197 			WARN_ON(!last_pfn);
1198 			for (pfn = map_pfn; pfn <= last_pfn; pfn++)
1199 				unpin_user_page(pfn_to_page(pfn));
1200 		}
1201 		vhost_vdpa_unmap(v, iotlb, start, size);
1202 	}
1203 unlock:
1204 	mmap_read_unlock(dev->mm);
1205 free:
1206 	free_page((unsigned long)page_list);
1207 	return ret;
1208 
1209 }
1210 
1211 static int vhost_vdpa_process_iotlb_update(struct vhost_vdpa *v,
1212 					   struct vhost_iotlb *iotlb,
1213 					   struct vhost_iotlb_msg *msg)
1214 {
1215 	struct vdpa_device *vdpa = v->vdpa;
1216 
1217 	if (msg->iova < v->range.first || !msg->size ||
1218 	    msg->iova > U64_MAX - msg->size + 1 ||
1219 	    msg->iova + msg->size - 1 > v->range.last)
1220 		return -EINVAL;
1221 
1222 	if (vhost_iotlb_itree_first(iotlb, msg->iova,
1223 				    msg->iova + msg->size - 1))
1224 		return -EEXIST;
1225 
1226 	if (vdpa->use_va)
1227 		return vhost_vdpa_va_map(v, iotlb, msg->iova, msg->size,
1228 					 msg->uaddr, msg->perm);
1229 
1230 	return vhost_vdpa_pa_map(v, iotlb, msg->iova, msg->size, msg->uaddr,
1231 				 msg->perm);
1232 }
1233 
1234 static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1235 					struct vhost_iotlb_msg *msg)
1236 {
1237 	struct vhost_vdpa *v = container_of(dev, struct vhost_vdpa, vdev);
1238 	struct vdpa_device *vdpa = v->vdpa;
1239 	const struct vdpa_config_ops *ops = vdpa->config;
1240 	struct vhost_iotlb *iotlb = NULL;
1241 	struct vhost_vdpa_as *as = NULL;
1242 	int r = 0;
1243 
1244 	mutex_lock(&dev->mutex);
1245 
1246 	r = vhost_dev_check_owner(dev);
1247 	if (r)
1248 		goto unlock;
1249 
1250 	if (msg->type == VHOST_IOTLB_UPDATE ||
1251 	    msg->type == VHOST_IOTLB_BATCH_BEGIN) {
1252 		as = vhost_vdpa_find_alloc_as(v, asid);
1253 		if (!as) {
1254 			dev_err(&v->dev, "can't find and alloc asid %d\n",
1255 				asid);
1256 			r = -EINVAL;
1257 			goto unlock;
1258 		}
1259 		iotlb = &as->iotlb;
1260 	} else
1261 		iotlb = asid_to_iotlb(v, asid);
1262 
1263 	if ((v->in_batch && v->batch_asid != asid) || !iotlb) {
1264 		if (v->in_batch && v->batch_asid != asid) {
1265 			dev_info(&v->dev, "batch id %d asid %d\n",
1266 				 v->batch_asid, asid);
1267 		}
1268 		if (!iotlb)
1269 			dev_err(&v->dev, "no iotlb for asid %d\n", asid);
1270 		r = -EINVAL;
1271 		goto unlock;
1272 	}
1273 
1274 	switch (msg->type) {
1275 	case VHOST_IOTLB_UPDATE:
1276 		r = vhost_vdpa_process_iotlb_update(v, iotlb, msg);
1277 		break;
1278 	case VHOST_IOTLB_INVALIDATE:
1279 		vhost_vdpa_unmap(v, iotlb, msg->iova, msg->size);
1280 		break;
1281 	case VHOST_IOTLB_BATCH_BEGIN:
1282 		v->batch_asid = asid;
1283 		v->in_batch = true;
1284 		break;
1285 	case VHOST_IOTLB_BATCH_END:
1286 		if (v->in_batch && ops->set_map)
1287 			ops->set_map(vdpa, asid, iotlb);
1288 		v->in_batch = false;
1289 		break;
1290 	default:
1291 		r = -EINVAL;
1292 		break;
1293 	}
1294 unlock:
1295 	mutex_unlock(&dev->mutex);
1296 
1297 	return r;
1298 }
1299 
1300 static ssize_t vhost_vdpa_chr_write_iter(struct kiocb *iocb,
1301 					 struct iov_iter *from)
1302 {
1303 	struct file *file = iocb->ki_filp;
1304 	struct vhost_vdpa *v = file->private_data;
1305 	struct vhost_dev *dev = &v->vdev;
1306 
1307 	return vhost_chr_write_iter(dev, from);
1308 }
1309 
1310 static int vhost_vdpa_alloc_domain(struct vhost_vdpa *v)
1311 {
1312 	struct vdpa_device *vdpa = v->vdpa;
1313 	const struct vdpa_config_ops *ops = vdpa->config;
1314 	struct device *dma_dev = vdpa_get_dma_dev(vdpa);
1315 	const struct bus_type *bus;
1316 	int ret;
1317 
1318 	/* Device want to do DMA by itself */
1319 	if (ops->set_map || ops->dma_map)
1320 		return 0;
1321 
1322 	bus = dma_dev->bus;
1323 	if (!bus)
1324 		return -EFAULT;
1325 
1326 	if (!device_iommu_capable(dma_dev, IOMMU_CAP_CACHE_COHERENCY)) {
1327 		dev_warn_once(&v->dev,
1328 			      "Failed to allocate domain, device is not IOMMU cache coherent capable\n");
1329 		return -ENOTSUPP;
1330 	}
1331 
1332 	v->domain = iommu_domain_alloc(bus);
1333 	if (!v->domain)
1334 		return -EIO;
1335 
1336 	ret = iommu_attach_device(v->domain, dma_dev);
1337 	if (ret)
1338 		goto err_attach;
1339 
1340 	return 0;
1341 
1342 err_attach:
1343 	iommu_domain_free(v->domain);
1344 	v->domain = NULL;
1345 	return ret;
1346 }
1347 
1348 static void vhost_vdpa_free_domain(struct vhost_vdpa *v)
1349 {
1350 	struct vdpa_device *vdpa = v->vdpa;
1351 	struct device *dma_dev = vdpa_get_dma_dev(vdpa);
1352 
1353 	if (v->domain) {
1354 		iommu_detach_device(v->domain, dma_dev);
1355 		iommu_domain_free(v->domain);
1356 	}
1357 
1358 	v->domain = NULL;
1359 }
1360 
1361 static void vhost_vdpa_set_iova_range(struct vhost_vdpa *v)
1362 {
1363 	struct vdpa_iova_range *range = &v->range;
1364 	struct vdpa_device *vdpa = v->vdpa;
1365 	const struct vdpa_config_ops *ops = vdpa->config;
1366 
1367 	if (ops->get_iova_range) {
1368 		*range = ops->get_iova_range(vdpa);
1369 	} else if (v->domain && v->domain->geometry.force_aperture) {
1370 		range->first = v->domain->geometry.aperture_start;
1371 		range->last = v->domain->geometry.aperture_end;
1372 	} else {
1373 		range->first = 0;
1374 		range->last = ULLONG_MAX;
1375 	}
1376 }
1377 
1378 static void vhost_vdpa_cleanup(struct vhost_vdpa *v)
1379 {
1380 	struct vhost_vdpa_as *as;
1381 	u32 asid;
1382 
1383 	for (asid = 0; asid < v->vdpa->nas; asid++) {
1384 		as = asid_to_as(v, asid);
1385 		if (as)
1386 			vhost_vdpa_remove_as(v, asid);
1387 	}
1388 
1389 	vhost_vdpa_free_domain(v);
1390 	vhost_dev_cleanup(&v->vdev);
1391 	kfree(v->vdev.vqs);
1392 	v->vdev.vqs = NULL;
1393 }
1394 
1395 static int vhost_vdpa_open(struct inode *inode, struct file *filep)
1396 {
1397 	struct vhost_vdpa *v;
1398 	struct vhost_dev *dev;
1399 	struct vhost_virtqueue **vqs;
1400 	int r, opened;
1401 	u32 i, nvqs;
1402 
1403 	v = container_of(inode->i_cdev, struct vhost_vdpa, cdev);
1404 
1405 	opened = atomic_cmpxchg(&v->opened, 0, 1);
1406 	if (opened)
1407 		return -EBUSY;
1408 
1409 	nvqs = v->nvqs;
1410 	r = vhost_vdpa_reset(v);
1411 	if (r)
1412 		goto err;
1413 
1414 	vqs = kmalloc_array(nvqs, sizeof(*vqs), GFP_KERNEL);
1415 	if (!vqs) {
1416 		r = -ENOMEM;
1417 		goto err;
1418 	}
1419 
1420 	dev = &v->vdev;
1421 	for (i = 0; i < nvqs; i++) {
1422 		vqs[i] = &v->vqs[i];
1423 		vqs[i]->handle_kick = handle_vq_kick;
1424 	}
1425 	vhost_dev_init(dev, vqs, nvqs, 0, 0, 0, false,
1426 		       vhost_vdpa_process_iotlb_msg);
1427 
1428 	r = vhost_vdpa_alloc_domain(v);
1429 	if (r)
1430 		goto err_alloc_domain;
1431 
1432 	vhost_vdpa_set_iova_range(v);
1433 
1434 	filep->private_data = v;
1435 
1436 	return 0;
1437 
1438 err_alloc_domain:
1439 	vhost_vdpa_cleanup(v);
1440 err:
1441 	atomic_dec(&v->opened);
1442 	return r;
1443 }
1444 
1445 static void vhost_vdpa_clean_irq(struct vhost_vdpa *v)
1446 {
1447 	u32 i;
1448 
1449 	for (i = 0; i < v->nvqs; i++)
1450 		vhost_vdpa_unsetup_vq_irq(v, i);
1451 }
1452 
1453 static int vhost_vdpa_release(struct inode *inode, struct file *filep)
1454 {
1455 	struct vhost_vdpa *v = filep->private_data;
1456 	struct vhost_dev *d = &v->vdev;
1457 
1458 	mutex_lock(&d->mutex);
1459 	filep->private_data = NULL;
1460 	vhost_vdpa_clean_irq(v);
1461 	vhost_vdpa_reset(v);
1462 	vhost_dev_stop(&v->vdev);
1463 	vhost_vdpa_unbind_mm(v);
1464 	vhost_vdpa_config_put(v);
1465 	vhost_vdpa_cleanup(v);
1466 	mutex_unlock(&d->mutex);
1467 
1468 	atomic_dec(&v->opened);
1469 	complete(&v->completion);
1470 
1471 	return 0;
1472 }
1473 
1474 #ifdef CONFIG_MMU
1475 static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
1476 {
1477 	struct vhost_vdpa *v = vmf->vma->vm_file->private_data;
1478 	struct vdpa_device *vdpa = v->vdpa;
1479 	const struct vdpa_config_ops *ops = vdpa->config;
1480 	struct vdpa_notification_area notify;
1481 	struct vm_area_struct *vma = vmf->vma;
1482 	u16 index = vma->vm_pgoff;
1483 
1484 	notify = ops->get_vq_notification(vdpa, index);
1485 
1486 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1487 	if (remap_pfn_range(vma, vmf->address & PAGE_MASK,
1488 			    PFN_DOWN(notify.addr), PAGE_SIZE,
1489 			    vma->vm_page_prot))
1490 		return VM_FAULT_SIGBUS;
1491 
1492 	return VM_FAULT_NOPAGE;
1493 }
1494 
1495 static const struct vm_operations_struct vhost_vdpa_vm_ops = {
1496 	.fault = vhost_vdpa_fault,
1497 };
1498 
1499 static int vhost_vdpa_mmap(struct file *file, struct vm_area_struct *vma)
1500 {
1501 	struct vhost_vdpa *v = vma->vm_file->private_data;
1502 	struct vdpa_device *vdpa = v->vdpa;
1503 	const struct vdpa_config_ops *ops = vdpa->config;
1504 	struct vdpa_notification_area notify;
1505 	unsigned long index = vma->vm_pgoff;
1506 
1507 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1508 		return -EINVAL;
1509 	if ((vma->vm_flags & VM_SHARED) == 0)
1510 		return -EINVAL;
1511 	if (vma->vm_flags & VM_READ)
1512 		return -EINVAL;
1513 	if (index > 65535)
1514 		return -EINVAL;
1515 	if (!ops->get_vq_notification)
1516 		return -ENOTSUPP;
1517 
1518 	/* To be safe and easily modelled by userspace, We only
1519 	 * support the doorbell which sits on the page boundary and
1520 	 * does not share the page with other registers.
1521 	 */
1522 	notify = ops->get_vq_notification(vdpa, index);
1523 	if (notify.addr & (PAGE_SIZE - 1))
1524 		return -EINVAL;
1525 	if (vma->vm_end - vma->vm_start != notify.size)
1526 		return -ENOTSUPP;
1527 
1528 	vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
1529 	vma->vm_ops = &vhost_vdpa_vm_ops;
1530 	return 0;
1531 }
1532 #endif /* CONFIG_MMU */
1533 
1534 static const struct file_operations vhost_vdpa_fops = {
1535 	.owner		= THIS_MODULE,
1536 	.open		= vhost_vdpa_open,
1537 	.release	= vhost_vdpa_release,
1538 	.write_iter	= vhost_vdpa_chr_write_iter,
1539 	.unlocked_ioctl	= vhost_vdpa_unlocked_ioctl,
1540 #ifdef CONFIG_MMU
1541 	.mmap		= vhost_vdpa_mmap,
1542 #endif /* CONFIG_MMU */
1543 	.compat_ioctl	= compat_ptr_ioctl,
1544 };
1545 
1546 static void vhost_vdpa_release_dev(struct device *device)
1547 {
1548 	struct vhost_vdpa *v =
1549 	       container_of(device, struct vhost_vdpa, dev);
1550 
1551 	ida_simple_remove(&vhost_vdpa_ida, v->minor);
1552 	kfree(v->vqs);
1553 	kfree(v);
1554 }
1555 
1556 static int vhost_vdpa_probe(struct vdpa_device *vdpa)
1557 {
1558 	const struct vdpa_config_ops *ops = vdpa->config;
1559 	struct vhost_vdpa *v;
1560 	int minor;
1561 	int i, r;
1562 
1563 	/* We can't support platform IOMMU device with more than 1
1564 	 * group or as
1565 	 */
1566 	if (!ops->set_map && !ops->dma_map &&
1567 	    (vdpa->ngroups > 1 || vdpa->nas > 1))
1568 		return -EOPNOTSUPP;
1569 
1570 	v = kzalloc(sizeof(*v), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1571 	if (!v)
1572 		return -ENOMEM;
1573 
1574 	minor = ida_simple_get(&vhost_vdpa_ida, 0,
1575 			       VHOST_VDPA_DEV_MAX, GFP_KERNEL);
1576 	if (minor < 0) {
1577 		kfree(v);
1578 		return minor;
1579 	}
1580 
1581 	atomic_set(&v->opened, 0);
1582 	v->minor = minor;
1583 	v->vdpa = vdpa;
1584 	v->nvqs = vdpa->nvqs;
1585 	v->virtio_id = ops->get_device_id(vdpa);
1586 
1587 	device_initialize(&v->dev);
1588 	v->dev.release = vhost_vdpa_release_dev;
1589 	v->dev.parent = &vdpa->dev;
1590 	v->dev.devt = MKDEV(MAJOR(vhost_vdpa_major), minor);
1591 	v->vqs = kmalloc_array(v->nvqs, sizeof(struct vhost_virtqueue),
1592 			       GFP_KERNEL);
1593 	if (!v->vqs) {
1594 		r = -ENOMEM;
1595 		goto err;
1596 	}
1597 
1598 	r = dev_set_name(&v->dev, "vhost-vdpa-%u", minor);
1599 	if (r)
1600 		goto err;
1601 
1602 	cdev_init(&v->cdev, &vhost_vdpa_fops);
1603 	v->cdev.owner = THIS_MODULE;
1604 
1605 	r = cdev_device_add(&v->cdev, &v->dev);
1606 	if (r)
1607 		goto err;
1608 
1609 	init_completion(&v->completion);
1610 	vdpa_set_drvdata(vdpa, v);
1611 
1612 	for (i = 0; i < VHOST_VDPA_IOTLB_BUCKETS; i++)
1613 		INIT_HLIST_HEAD(&v->as[i]);
1614 
1615 	return 0;
1616 
1617 err:
1618 	put_device(&v->dev);
1619 	return r;
1620 }
1621 
1622 static void vhost_vdpa_remove(struct vdpa_device *vdpa)
1623 {
1624 	struct vhost_vdpa *v = vdpa_get_drvdata(vdpa);
1625 	int opened;
1626 
1627 	cdev_device_del(&v->cdev, &v->dev);
1628 
1629 	do {
1630 		opened = atomic_cmpxchg(&v->opened, 0, 1);
1631 		if (!opened)
1632 			break;
1633 		wait_for_completion(&v->completion);
1634 	} while (1);
1635 
1636 	put_device(&v->dev);
1637 }
1638 
1639 static struct vdpa_driver vhost_vdpa_driver = {
1640 	.driver = {
1641 		.name	= "vhost_vdpa",
1642 	},
1643 	.probe	= vhost_vdpa_probe,
1644 	.remove	= vhost_vdpa_remove,
1645 };
1646 
1647 static int __init vhost_vdpa_init(void)
1648 {
1649 	int r;
1650 
1651 	r = alloc_chrdev_region(&vhost_vdpa_major, 0, VHOST_VDPA_DEV_MAX,
1652 				"vhost-vdpa");
1653 	if (r)
1654 		goto err_alloc_chrdev;
1655 
1656 	r = vdpa_register_driver(&vhost_vdpa_driver);
1657 	if (r)
1658 		goto err_vdpa_register_driver;
1659 
1660 	return 0;
1661 
1662 err_vdpa_register_driver:
1663 	unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1664 err_alloc_chrdev:
1665 	return r;
1666 }
1667 module_init(vhost_vdpa_init);
1668 
1669 static void __exit vhost_vdpa_exit(void)
1670 {
1671 	vdpa_unregister_driver(&vhost_vdpa_driver);
1672 	unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1673 }
1674 module_exit(vhost_vdpa_exit);
1675 
1676 MODULE_VERSION("0.0.1");
1677 MODULE_LICENSE("GPL v2");
1678 MODULE_AUTHOR("Intel Corporation");
1679 MODULE_DESCRIPTION("vDPA-based vhost backend for virtio");
1680