xref: /openbmc/linux/drivers/vhost/vdpa.c (revision aa6159ab) 
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/iommu.h>
20 #include <linux/uuid.h>
21 #include <linux/vdpa.h>
22 #include <linux/nospec.h>
23 #include <linux/vhost.h>
24 #include <linux/virtio_net.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 };
33 
34 #define VHOST_VDPA_DEV_MAX (1U << MINORBITS)
35 
36 struct vhost_vdpa {
37 	struct vhost_dev vdev;
38 	struct iommu_domain *domain;
39 	struct vhost_virtqueue *vqs;
40 	struct completion completion;
41 	struct vdpa_device *vdpa;
42 	struct device dev;
43 	struct cdev cdev;
44 	atomic_t opened;
45 	int nvqs;
46 	int virtio_id;
47 	int minor;
48 	struct eventfd_ctx *config_ctx;
49 	int in_batch;
50 	struct vdpa_iova_range range;
51 };
52 
53 static DEFINE_IDA(vhost_vdpa_ida);
54 
55 static dev_t vhost_vdpa_major;
56 
57 static void handle_vq_kick(struct vhost_work *work)
58 {
59 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
60 						  poll.work);
61 	struct vhost_vdpa *v = container_of(vq->dev, struct vhost_vdpa, vdev);
62 	const struct vdpa_config_ops *ops = v->vdpa->config;
63 
64 	ops->kick_vq(v->vdpa, vq - v->vqs);
65 }
66 
67 static irqreturn_t vhost_vdpa_virtqueue_cb(void *private)
68 {
69 	struct vhost_virtqueue *vq = private;
70 	struct eventfd_ctx *call_ctx = vq->call_ctx.ctx;
71 
72 	if (call_ctx)
73 		eventfd_signal(call_ctx, 1);
74 
75 	return IRQ_HANDLED;
76 }
77 
78 static irqreturn_t vhost_vdpa_config_cb(void *private)
79 {
80 	struct vhost_vdpa *v = private;
81 	struct eventfd_ctx *config_ctx = v->config_ctx;
82 
83 	if (config_ctx)
84 		eventfd_signal(config_ctx, 1);
85 
86 	return IRQ_HANDLED;
87 }
88 
89 static void vhost_vdpa_setup_vq_irq(struct vhost_vdpa *v, u16 qid)
90 {
91 	struct vhost_virtqueue *vq = &v->vqs[qid];
92 	const struct vdpa_config_ops *ops = v->vdpa->config;
93 	struct vdpa_device *vdpa = v->vdpa;
94 	int ret, irq;
95 
96 	if (!ops->get_vq_irq)
97 		return;
98 
99 	irq = ops->get_vq_irq(vdpa, qid);
100 	irq_bypass_unregister_producer(&vq->call_ctx.producer);
101 	if (!vq->call_ctx.ctx || irq < 0)
102 		return;
103 
104 	vq->call_ctx.producer.token = vq->call_ctx.ctx;
105 	vq->call_ctx.producer.irq = irq;
106 	ret = irq_bypass_register_producer(&vq->call_ctx.producer);
107 	if (unlikely(ret))
108 		dev_info(&v->dev, "vq %u, irq bypass producer (token %p) registration fails, ret =  %d\n",
109 			 qid, vq->call_ctx.producer.token, ret);
110 }
111 
112 static void vhost_vdpa_unsetup_vq_irq(struct vhost_vdpa *v, u16 qid)
113 {
114 	struct vhost_virtqueue *vq = &v->vqs[qid];
115 
116 	irq_bypass_unregister_producer(&vq->call_ctx.producer);
117 }
118 
119 static void vhost_vdpa_reset(struct vhost_vdpa *v)
120 {
121 	struct vdpa_device *vdpa = v->vdpa;
122 
123 	vdpa_reset(vdpa);
124 	v->in_batch = 0;
125 }
126 
127 static long vhost_vdpa_get_device_id(struct vhost_vdpa *v, u8 __user *argp)
128 {
129 	struct vdpa_device *vdpa = v->vdpa;
130 	const struct vdpa_config_ops *ops = vdpa->config;
131 	u32 device_id;
132 
133 	device_id = ops->get_device_id(vdpa);
134 
135 	if (copy_to_user(argp, &device_id, sizeof(device_id)))
136 		return -EFAULT;
137 
138 	return 0;
139 }
140 
141 static long vhost_vdpa_get_status(struct vhost_vdpa *v, u8 __user *statusp)
142 {
143 	struct vdpa_device *vdpa = v->vdpa;
144 	const struct vdpa_config_ops *ops = vdpa->config;
145 	u8 status;
146 
147 	status = ops->get_status(vdpa);
148 
149 	if (copy_to_user(statusp, &status, sizeof(status)))
150 		return -EFAULT;
151 
152 	return 0;
153 }
154 
155 static long vhost_vdpa_set_status(struct vhost_vdpa *v, u8 __user *statusp)
156 {
157 	struct vdpa_device *vdpa = v->vdpa;
158 	const struct vdpa_config_ops *ops = vdpa->config;
159 	u8 status, status_old;
160 	int nvqs = v->nvqs;
161 	u16 i;
162 
163 	if (copy_from_user(&status, statusp, sizeof(status)))
164 		return -EFAULT;
165 
166 	status_old = ops->get_status(vdpa);
167 
168 	/*
169 	 * Userspace shouldn't remove status bits unless reset the
170 	 * status to 0.
171 	 */
172 	if (status != 0 && (ops->get_status(vdpa) & ~status) != 0)
173 		return -EINVAL;
174 
175 	ops->set_status(vdpa, status);
176 
177 	if ((status & VIRTIO_CONFIG_S_DRIVER_OK) && !(status_old & VIRTIO_CONFIG_S_DRIVER_OK))
178 		for (i = 0; i < nvqs; i++)
179 			vhost_vdpa_setup_vq_irq(v, i);
180 
181 	if ((status_old & VIRTIO_CONFIG_S_DRIVER_OK) && !(status & VIRTIO_CONFIG_S_DRIVER_OK))
182 		for (i = 0; i < nvqs; i++)
183 			vhost_vdpa_unsetup_vq_irq(v, i);
184 
185 	return 0;
186 }
187 
188 static int vhost_vdpa_config_validate(struct vhost_vdpa *v,
189 				      struct vhost_vdpa_config *c)
190 {
191 	long size = 0;
192 
193 	switch (v->virtio_id) {
194 	case VIRTIO_ID_NET:
195 		size = sizeof(struct virtio_net_config);
196 		break;
197 	}
198 
199 	if (c->len == 0)
200 		return -EINVAL;
201 
202 	if (c->len > size - c->off)
203 		return -E2BIG;
204 
205 	return 0;
206 }
207 
208 static long vhost_vdpa_get_config(struct vhost_vdpa *v,
209 				  struct vhost_vdpa_config __user *c)
210 {
211 	struct vdpa_device *vdpa = v->vdpa;
212 	struct vhost_vdpa_config config;
213 	unsigned long size = offsetof(struct vhost_vdpa_config, buf);
214 	u8 *buf;
215 
216 	if (copy_from_user(&config, c, size))
217 		return -EFAULT;
218 	if (vhost_vdpa_config_validate(v, &config))
219 		return -EINVAL;
220 	buf = kvzalloc(config.len, GFP_KERNEL);
221 	if (!buf)
222 		return -ENOMEM;
223 
224 	vdpa_get_config(vdpa, config.off, buf, config.len);
225 
226 	if (copy_to_user(c->buf, buf, config.len)) {
227 		kvfree(buf);
228 		return -EFAULT;
229 	}
230 
231 	kvfree(buf);
232 	return 0;
233 }
234 
235 static long vhost_vdpa_set_config(struct vhost_vdpa *v,
236 				  struct vhost_vdpa_config __user *c)
237 {
238 	struct vdpa_device *vdpa = v->vdpa;
239 	const struct vdpa_config_ops *ops = vdpa->config;
240 	struct vhost_vdpa_config config;
241 	unsigned long size = offsetof(struct vhost_vdpa_config, buf);
242 	u8 *buf;
243 
244 	if (copy_from_user(&config, c, size))
245 		return -EFAULT;
246 	if (vhost_vdpa_config_validate(v, &config))
247 		return -EINVAL;
248 	buf = kvzalloc(config.len, GFP_KERNEL);
249 	if (!buf)
250 		return -ENOMEM;
251 
252 	if (copy_from_user(buf, c->buf, config.len)) {
253 		kvfree(buf);
254 		return -EFAULT;
255 	}
256 
257 	ops->set_config(vdpa, config.off, buf, config.len);
258 
259 	kvfree(buf);
260 	return 0;
261 }
262 
263 static long vhost_vdpa_get_features(struct vhost_vdpa *v, u64 __user *featurep)
264 {
265 	struct vdpa_device *vdpa = v->vdpa;
266 	const struct vdpa_config_ops *ops = vdpa->config;
267 	u64 features;
268 
269 	features = ops->get_features(vdpa);
270 
271 	if (copy_to_user(featurep, &features, sizeof(features)))
272 		return -EFAULT;
273 
274 	return 0;
275 }
276 
277 static long vhost_vdpa_set_features(struct vhost_vdpa *v, u64 __user *featurep)
278 {
279 	struct vdpa_device *vdpa = v->vdpa;
280 	const struct vdpa_config_ops *ops = vdpa->config;
281 	u64 features;
282 
283 	/*
284 	 * It's not allowed to change the features after they have
285 	 * been negotiated.
286 	 */
287 	if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_FEATURES_OK)
288 		return -EBUSY;
289 
290 	if (copy_from_user(&features, featurep, sizeof(features)))
291 		return -EFAULT;
292 
293 	if (vdpa_set_features(vdpa, features))
294 		return -EINVAL;
295 
296 	return 0;
297 }
298 
299 static long vhost_vdpa_get_vring_num(struct vhost_vdpa *v, u16 __user *argp)
300 {
301 	struct vdpa_device *vdpa = v->vdpa;
302 	const struct vdpa_config_ops *ops = vdpa->config;
303 	u16 num;
304 
305 	num = ops->get_vq_num_max(vdpa);
306 
307 	if (copy_to_user(argp, &num, sizeof(num)))
308 		return -EFAULT;
309 
310 	return 0;
311 }
312 
313 static void vhost_vdpa_config_put(struct vhost_vdpa *v)
314 {
315 	if (v->config_ctx)
316 		eventfd_ctx_put(v->config_ctx);
317 }
318 
319 static long vhost_vdpa_set_config_call(struct vhost_vdpa *v, u32 __user *argp)
320 {
321 	struct vdpa_callback cb;
322 	int fd;
323 	struct eventfd_ctx *ctx;
324 
325 	cb.callback = vhost_vdpa_config_cb;
326 	cb.private = v->vdpa;
327 	if (copy_from_user(&fd, argp, sizeof(fd)))
328 		return  -EFAULT;
329 
330 	ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
331 	swap(ctx, v->config_ctx);
332 
333 	if (!IS_ERR_OR_NULL(ctx))
334 		eventfd_ctx_put(ctx);
335 
336 	if (IS_ERR(v->config_ctx))
337 		return PTR_ERR(v->config_ctx);
338 
339 	v->vdpa->config->set_config_cb(v->vdpa, &cb);
340 
341 	return 0;
342 }
343 
344 static long vhost_vdpa_get_iova_range(struct vhost_vdpa *v, u32 __user *argp)
345 {
346 	struct vhost_vdpa_iova_range range = {
347 		.first = v->range.first,
348 		.last = v->range.last,
349 	};
350 
351 	if (copy_to_user(argp, &range, sizeof(range)))
352 		return -EFAULT;
353 	return 0;
354 }
355 
356 static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
357 				   void __user *argp)
358 {
359 	struct vdpa_device *vdpa = v->vdpa;
360 	const struct vdpa_config_ops *ops = vdpa->config;
361 	struct vdpa_vq_state vq_state;
362 	struct vdpa_callback cb;
363 	struct vhost_virtqueue *vq;
364 	struct vhost_vring_state s;
365 	u32 idx;
366 	long r;
367 
368 	r = get_user(idx, (u32 __user *)argp);
369 	if (r < 0)
370 		return r;
371 
372 	if (idx >= v->nvqs)
373 		return -ENOBUFS;
374 
375 	idx = array_index_nospec(idx, v->nvqs);
376 	vq = &v->vqs[idx];
377 
378 	switch (cmd) {
379 	case VHOST_VDPA_SET_VRING_ENABLE:
380 		if (copy_from_user(&s, argp, sizeof(s)))
381 			return -EFAULT;
382 		ops->set_vq_ready(vdpa, idx, s.num);
383 		return 0;
384 	case VHOST_GET_VRING_BASE:
385 		r = ops->get_vq_state(v->vdpa, idx, &vq_state);
386 		if (r)
387 			return r;
388 
389 		vq->last_avail_idx = vq_state.avail_index;
390 		break;
391 	}
392 
393 	r = vhost_vring_ioctl(&v->vdev, cmd, argp);
394 	if (r)
395 		return r;
396 
397 	switch (cmd) {
398 	case VHOST_SET_VRING_ADDR:
399 		if (ops->set_vq_address(vdpa, idx,
400 					(u64)(uintptr_t)vq->desc,
401 					(u64)(uintptr_t)vq->avail,
402 					(u64)(uintptr_t)vq->used))
403 			r = -EINVAL;
404 		break;
405 
406 	case VHOST_SET_VRING_BASE:
407 		vq_state.avail_index = vq->last_avail_idx;
408 		if (ops->set_vq_state(vdpa, idx, &vq_state))
409 			r = -EINVAL;
410 		break;
411 
412 	case VHOST_SET_VRING_CALL:
413 		if (vq->call_ctx.ctx) {
414 			cb.callback = vhost_vdpa_virtqueue_cb;
415 			cb.private = vq;
416 		} else {
417 			cb.callback = NULL;
418 			cb.private = NULL;
419 		}
420 		ops->set_vq_cb(vdpa, idx, &cb);
421 		vhost_vdpa_setup_vq_irq(v, idx);
422 		break;
423 
424 	case VHOST_SET_VRING_NUM:
425 		ops->set_vq_num(vdpa, idx, vq->num);
426 		break;
427 	}
428 
429 	return r;
430 }
431 
432 static long vhost_vdpa_unlocked_ioctl(struct file *filep,
433 				      unsigned int cmd, unsigned long arg)
434 {
435 	struct vhost_vdpa *v = filep->private_data;
436 	struct vhost_dev *d = &v->vdev;
437 	void __user *argp = (void __user *)arg;
438 	u64 __user *featurep = argp;
439 	u64 features;
440 	long r = 0;
441 
442 	if (cmd == VHOST_SET_BACKEND_FEATURES) {
443 		if (copy_from_user(&features, featurep, sizeof(features)))
444 			return -EFAULT;
445 		if (features & ~VHOST_VDPA_BACKEND_FEATURES)
446 			return -EOPNOTSUPP;
447 		vhost_set_backend_features(&v->vdev, features);
448 		return 0;
449 	}
450 
451 	mutex_lock(&d->mutex);
452 
453 	switch (cmd) {
454 	case VHOST_VDPA_GET_DEVICE_ID:
455 		r = vhost_vdpa_get_device_id(v, argp);
456 		break;
457 	case VHOST_VDPA_GET_STATUS:
458 		r = vhost_vdpa_get_status(v, argp);
459 		break;
460 	case VHOST_VDPA_SET_STATUS:
461 		r = vhost_vdpa_set_status(v, argp);
462 		break;
463 	case VHOST_VDPA_GET_CONFIG:
464 		r = vhost_vdpa_get_config(v, argp);
465 		break;
466 	case VHOST_VDPA_SET_CONFIG:
467 		r = vhost_vdpa_set_config(v, argp);
468 		break;
469 	case VHOST_GET_FEATURES:
470 		r = vhost_vdpa_get_features(v, argp);
471 		break;
472 	case VHOST_SET_FEATURES:
473 		r = vhost_vdpa_set_features(v, argp);
474 		break;
475 	case VHOST_VDPA_GET_VRING_NUM:
476 		r = vhost_vdpa_get_vring_num(v, argp);
477 		break;
478 	case VHOST_SET_LOG_BASE:
479 	case VHOST_SET_LOG_FD:
480 		r = -ENOIOCTLCMD;
481 		break;
482 	case VHOST_VDPA_SET_CONFIG_CALL:
483 		r = vhost_vdpa_set_config_call(v, argp);
484 		break;
485 	case VHOST_GET_BACKEND_FEATURES:
486 		features = VHOST_VDPA_BACKEND_FEATURES;
487 		if (copy_to_user(featurep, &features, sizeof(features)))
488 			r = -EFAULT;
489 		break;
490 	case VHOST_VDPA_GET_IOVA_RANGE:
491 		r = vhost_vdpa_get_iova_range(v, argp);
492 		break;
493 	default:
494 		r = vhost_dev_ioctl(&v->vdev, cmd, argp);
495 		if (r == -ENOIOCTLCMD)
496 			r = vhost_vdpa_vring_ioctl(v, cmd, argp);
497 		break;
498 	}
499 
500 	mutex_unlock(&d->mutex);
501 	return r;
502 }
503 
504 static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v, u64 start, u64 last)
505 {
506 	struct vhost_dev *dev = &v->vdev;
507 	struct vhost_iotlb *iotlb = dev->iotlb;
508 	struct vhost_iotlb_map *map;
509 	struct page *page;
510 	unsigned long pfn, pinned;
511 
512 	while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
513 		pinned = map->size >> PAGE_SHIFT;
514 		for (pfn = map->addr >> PAGE_SHIFT;
515 		     pinned > 0; pfn++, pinned--) {
516 			page = pfn_to_page(pfn);
517 			if (map->perm & VHOST_ACCESS_WO)
518 				set_page_dirty_lock(page);
519 			unpin_user_page(page);
520 		}
521 		atomic64_sub(map->size >> PAGE_SHIFT, &dev->mm->pinned_vm);
522 		vhost_iotlb_map_free(iotlb, map);
523 	}
524 }
525 
526 static void vhost_vdpa_iotlb_free(struct vhost_vdpa *v)
527 {
528 	struct vhost_dev *dev = &v->vdev;
529 
530 	vhost_vdpa_iotlb_unmap(v, 0ULL, 0ULL - 1);
531 	kfree(dev->iotlb);
532 	dev->iotlb = NULL;
533 }
534 
535 static int perm_to_iommu_flags(u32 perm)
536 {
537 	int flags = 0;
538 
539 	switch (perm) {
540 	case VHOST_ACCESS_WO:
541 		flags |= IOMMU_WRITE;
542 		break;
543 	case VHOST_ACCESS_RO:
544 		flags |= IOMMU_READ;
545 		break;
546 	case VHOST_ACCESS_RW:
547 		flags |= (IOMMU_WRITE | IOMMU_READ);
548 		break;
549 	default:
550 		WARN(1, "invalidate vhost IOTLB permission\n");
551 		break;
552 	}
553 
554 	return flags | IOMMU_CACHE;
555 }
556 
557 static int vhost_vdpa_map(struct vhost_vdpa *v,
558 			  u64 iova, u64 size, u64 pa, u32 perm)
559 {
560 	struct vhost_dev *dev = &v->vdev;
561 	struct vdpa_device *vdpa = v->vdpa;
562 	const struct vdpa_config_ops *ops = vdpa->config;
563 	int r = 0;
564 
565 	r = vhost_iotlb_add_range(dev->iotlb, iova, iova + size - 1,
566 				  pa, perm);
567 	if (r)
568 		return r;
569 
570 	if (ops->dma_map) {
571 		r = ops->dma_map(vdpa, iova, size, pa, perm);
572 	} else if (ops->set_map) {
573 		if (!v->in_batch)
574 			r = ops->set_map(vdpa, dev->iotlb);
575 	} else {
576 		r = iommu_map(v->domain, iova, pa, size,
577 			      perm_to_iommu_flags(perm));
578 	}
579 
580 	if (r)
581 		vhost_iotlb_del_range(dev->iotlb, iova, iova + size - 1);
582 	else
583 		atomic64_add(size >> PAGE_SHIFT, &dev->mm->pinned_vm);
584 
585 	return r;
586 }
587 
588 static void vhost_vdpa_unmap(struct vhost_vdpa *v, u64 iova, u64 size)
589 {
590 	struct vhost_dev *dev = &v->vdev;
591 	struct vdpa_device *vdpa = v->vdpa;
592 	const struct vdpa_config_ops *ops = vdpa->config;
593 
594 	vhost_vdpa_iotlb_unmap(v, iova, iova + size - 1);
595 
596 	if (ops->dma_map) {
597 		ops->dma_unmap(vdpa, iova, size);
598 	} else if (ops->set_map) {
599 		if (!v->in_batch)
600 			ops->set_map(vdpa, dev->iotlb);
601 	} else {
602 		iommu_unmap(v->domain, iova, size);
603 	}
604 }
605 
606 static int vhost_vdpa_process_iotlb_update(struct vhost_vdpa *v,
607 					   struct vhost_iotlb_msg *msg)
608 {
609 	struct vhost_dev *dev = &v->vdev;
610 	struct vhost_iotlb *iotlb = dev->iotlb;
611 	struct page **page_list;
612 	unsigned long list_size = PAGE_SIZE / sizeof(struct page *);
613 	unsigned int gup_flags = FOLL_LONGTERM;
614 	unsigned long npages, cur_base, map_pfn, last_pfn = 0;
615 	unsigned long lock_limit, sz2pin, nchunks, i;
616 	u64 iova = msg->iova;
617 	long pinned;
618 	int ret = 0;
619 
620 	if (msg->iova < v->range.first ||
621 	    msg->iova + msg->size - 1 > v->range.last)
622 		return -EINVAL;
623 
624 	if (vhost_iotlb_itree_first(iotlb, msg->iova,
625 				    msg->iova + msg->size - 1))
626 		return -EEXIST;
627 
628 	/* Limit the use of memory for bookkeeping */
629 	page_list = (struct page **) __get_free_page(GFP_KERNEL);
630 	if (!page_list)
631 		return -ENOMEM;
632 
633 	if (msg->perm & VHOST_ACCESS_WO)
634 		gup_flags |= FOLL_WRITE;
635 
636 	npages = PAGE_ALIGN(msg->size + (iova & ~PAGE_MASK)) >> PAGE_SHIFT;
637 	if (!npages) {
638 		ret = -EINVAL;
639 		goto free;
640 	}
641 
642 	mmap_read_lock(dev->mm);
643 
644 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
645 	if (npages + atomic64_read(&dev->mm->pinned_vm) > lock_limit) {
646 		ret = -ENOMEM;
647 		goto unlock;
648 	}
649 
650 	cur_base = msg->uaddr & PAGE_MASK;
651 	iova &= PAGE_MASK;
652 	nchunks = 0;
653 
654 	while (npages) {
655 		sz2pin = min_t(unsigned long, npages, list_size);
656 		pinned = pin_user_pages(cur_base, sz2pin,
657 					gup_flags, page_list, NULL);
658 		if (sz2pin != pinned) {
659 			if (pinned < 0) {
660 				ret = pinned;
661 			} else {
662 				unpin_user_pages(page_list, pinned);
663 				ret = -ENOMEM;
664 			}
665 			goto out;
666 		}
667 		nchunks++;
668 
669 		if (!last_pfn)
670 			map_pfn = page_to_pfn(page_list[0]);
671 
672 		for (i = 0; i < pinned; i++) {
673 			unsigned long this_pfn = page_to_pfn(page_list[i]);
674 			u64 csize;
675 
676 			if (last_pfn && (this_pfn != last_pfn + 1)) {
677 				/* Pin a contiguous chunk of memory */
678 				csize = (last_pfn - map_pfn + 1) << PAGE_SHIFT;
679 				ret = vhost_vdpa_map(v, iova, csize,
680 						     map_pfn << PAGE_SHIFT,
681 						     msg->perm);
682 				if (ret) {
683 					/*
684 					 * Unpin the pages that are left unmapped
685 					 * from this point on in the current
686 					 * page_list. The remaining outstanding
687 					 * ones which may stride across several
688 					 * chunks will be covered in the common
689 					 * error path subsequently.
690 					 */
691 					unpin_user_pages(&page_list[i],
692 							 pinned - i);
693 					goto out;
694 				}
695 
696 				map_pfn = this_pfn;
697 				iova += csize;
698 				nchunks = 0;
699 			}
700 
701 			last_pfn = this_pfn;
702 		}
703 
704 		cur_base += pinned << PAGE_SHIFT;
705 		npages -= pinned;
706 	}
707 
708 	/* Pin the rest chunk */
709 	ret = vhost_vdpa_map(v, iova, (last_pfn - map_pfn + 1) << PAGE_SHIFT,
710 			     map_pfn << PAGE_SHIFT, msg->perm);
711 out:
712 	if (ret) {
713 		if (nchunks) {
714 			unsigned long pfn;
715 
716 			/*
717 			 * Unpin the outstanding pages which are yet to be
718 			 * mapped but haven't due to vdpa_map() or
719 			 * pin_user_pages() failure.
720 			 *
721 			 * Mapped pages are accounted in vdpa_map(), hence
722 			 * the corresponding unpinning will be handled by
723 			 * vdpa_unmap().
724 			 */
725 			WARN_ON(!last_pfn);
726 			for (pfn = map_pfn; pfn <= last_pfn; pfn++)
727 				unpin_user_page(pfn_to_page(pfn));
728 		}
729 		vhost_vdpa_unmap(v, msg->iova, msg->size);
730 	}
731 unlock:
732 	mmap_read_unlock(dev->mm);
733 free:
734 	free_page((unsigned long)page_list);
735 	return ret;
736 }
737 
738 static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev,
739 					struct vhost_iotlb_msg *msg)
740 {
741 	struct vhost_vdpa *v = container_of(dev, struct vhost_vdpa, vdev);
742 	struct vdpa_device *vdpa = v->vdpa;
743 	const struct vdpa_config_ops *ops = vdpa->config;
744 	int r = 0;
745 
746 	r = vhost_dev_check_owner(dev);
747 	if (r)
748 		return r;
749 
750 	switch (msg->type) {
751 	case VHOST_IOTLB_UPDATE:
752 		r = vhost_vdpa_process_iotlb_update(v, msg);
753 		break;
754 	case VHOST_IOTLB_INVALIDATE:
755 		vhost_vdpa_unmap(v, msg->iova, msg->size);
756 		break;
757 	case VHOST_IOTLB_BATCH_BEGIN:
758 		v->in_batch = true;
759 		break;
760 	case VHOST_IOTLB_BATCH_END:
761 		if (v->in_batch && ops->set_map)
762 			ops->set_map(vdpa, dev->iotlb);
763 		v->in_batch = false;
764 		break;
765 	default:
766 		r = -EINVAL;
767 		break;
768 	}
769 
770 	return r;
771 }
772 
773 static ssize_t vhost_vdpa_chr_write_iter(struct kiocb *iocb,
774 					 struct iov_iter *from)
775 {
776 	struct file *file = iocb->ki_filp;
777 	struct vhost_vdpa *v = file->private_data;
778 	struct vhost_dev *dev = &v->vdev;
779 
780 	return vhost_chr_write_iter(dev, from);
781 }
782 
783 static int vhost_vdpa_alloc_domain(struct vhost_vdpa *v)
784 {
785 	struct vdpa_device *vdpa = v->vdpa;
786 	const struct vdpa_config_ops *ops = vdpa->config;
787 	struct device *dma_dev = vdpa_get_dma_dev(vdpa);
788 	struct bus_type *bus;
789 	int ret;
790 
791 	/* Device want to do DMA by itself */
792 	if (ops->set_map || ops->dma_map)
793 		return 0;
794 
795 	bus = dma_dev->bus;
796 	if (!bus)
797 		return -EFAULT;
798 
799 	if (!iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
800 		return -ENOTSUPP;
801 
802 	v->domain = iommu_domain_alloc(bus);
803 	if (!v->domain)
804 		return -EIO;
805 
806 	ret = iommu_attach_device(v->domain, dma_dev);
807 	if (ret)
808 		goto err_attach;
809 
810 	return 0;
811 
812 err_attach:
813 	iommu_domain_free(v->domain);
814 	return ret;
815 }
816 
817 static void vhost_vdpa_free_domain(struct vhost_vdpa *v)
818 {
819 	struct vdpa_device *vdpa = v->vdpa;
820 	struct device *dma_dev = vdpa_get_dma_dev(vdpa);
821 
822 	if (v->domain) {
823 		iommu_detach_device(v->domain, dma_dev);
824 		iommu_domain_free(v->domain);
825 	}
826 
827 	v->domain = NULL;
828 }
829 
830 static void vhost_vdpa_set_iova_range(struct vhost_vdpa *v)
831 {
832 	struct vdpa_iova_range *range = &v->range;
833 	struct iommu_domain_geometry geo;
834 	struct vdpa_device *vdpa = v->vdpa;
835 	const struct vdpa_config_ops *ops = vdpa->config;
836 
837 	if (ops->get_iova_range) {
838 		*range = ops->get_iova_range(vdpa);
839 	} else if (v->domain &&
840 		   !iommu_domain_get_attr(v->domain,
841 		   DOMAIN_ATTR_GEOMETRY, &geo) &&
842 		   geo.force_aperture) {
843 		range->first = geo.aperture_start;
844 		range->last = geo.aperture_end;
845 	} else {
846 		range->first = 0;
847 		range->last = ULLONG_MAX;
848 	}
849 }
850 
851 static int vhost_vdpa_open(struct inode *inode, struct file *filep)
852 {
853 	struct vhost_vdpa *v;
854 	struct vhost_dev *dev;
855 	struct vhost_virtqueue **vqs;
856 	int nvqs, i, r, opened;
857 
858 	v = container_of(inode->i_cdev, struct vhost_vdpa, cdev);
859 
860 	opened = atomic_cmpxchg(&v->opened, 0, 1);
861 	if (opened)
862 		return -EBUSY;
863 
864 	nvqs = v->nvqs;
865 	vhost_vdpa_reset(v);
866 
867 	vqs = kmalloc_array(nvqs, sizeof(*vqs), GFP_KERNEL);
868 	if (!vqs) {
869 		r = -ENOMEM;
870 		goto err;
871 	}
872 
873 	dev = &v->vdev;
874 	for (i = 0; i < nvqs; i++) {
875 		vqs[i] = &v->vqs[i];
876 		vqs[i]->handle_kick = handle_vq_kick;
877 	}
878 	vhost_dev_init(dev, vqs, nvqs, 0, 0, 0, false,
879 		       vhost_vdpa_process_iotlb_msg);
880 
881 	dev->iotlb = vhost_iotlb_alloc(0, 0);
882 	if (!dev->iotlb) {
883 		r = -ENOMEM;
884 		goto err_init_iotlb;
885 	}
886 
887 	r = vhost_vdpa_alloc_domain(v);
888 	if (r)
889 		goto err_init_iotlb;
890 
891 	vhost_vdpa_set_iova_range(v);
892 
893 	filep->private_data = v;
894 
895 	return 0;
896 
897 err_init_iotlb:
898 	vhost_dev_cleanup(&v->vdev);
899 	kfree(vqs);
900 err:
901 	atomic_dec(&v->opened);
902 	return r;
903 }
904 
905 static void vhost_vdpa_clean_irq(struct vhost_vdpa *v)
906 {
907 	struct vhost_virtqueue *vq;
908 	int i;
909 
910 	for (i = 0; i < v->nvqs; i++) {
911 		vq = &v->vqs[i];
912 		if (vq->call_ctx.producer.irq)
913 			irq_bypass_unregister_producer(&vq->call_ctx.producer);
914 	}
915 }
916 
917 static int vhost_vdpa_release(struct inode *inode, struct file *filep)
918 {
919 	struct vhost_vdpa *v = filep->private_data;
920 	struct vhost_dev *d = &v->vdev;
921 
922 	mutex_lock(&d->mutex);
923 	filep->private_data = NULL;
924 	vhost_vdpa_reset(v);
925 	vhost_dev_stop(&v->vdev);
926 	vhost_vdpa_iotlb_free(v);
927 	vhost_vdpa_free_domain(v);
928 	vhost_vdpa_config_put(v);
929 	vhost_vdpa_clean_irq(v);
930 	vhost_dev_cleanup(&v->vdev);
931 	kfree(v->vdev.vqs);
932 	mutex_unlock(&d->mutex);
933 
934 	atomic_dec(&v->opened);
935 	complete(&v->completion);
936 
937 	return 0;
938 }
939 
940 #ifdef CONFIG_MMU
941 static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
942 {
943 	struct vhost_vdpa *v = vmf->vma->vm_file->private_data;
944 	struct vdpa_device *vdpa = v->vdpa;
945 	const struct vdpa_config_ops *ops = vdpa->config;
946 	struct vdpa_notification_area notify;
947 	struct vm_area_struct *vma = vmf->vma;
948 	u16 index = vma->vm_pgoff;
949 
950 	notify = ops->get_vq_notification(vdpa, index);
951 
952 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
953 	if (remap_pfn_range(vma, vmf->address & PAGE_MASK,
954 			    notify.addr >> PAGE_SHIFT, PAGE_SIZE,
955 			    vma->vm_page_prot))
956 		return VM_FAULT_SIGBUS;
957 
958 	return VM_FAULT_NOPAGE;
959 }
960 
961 static const struct vm_operations_struct vhost_vdpa_vm_ops = {
962 	.fault = vhost_vdpa_fault,
963 };
964 
965 static int vhost_vdpa_mmap(struct file *file, struct vm_area_struct *vma)
966 {
967 	struct vhost_vdpa *v = vma->vm_file->private_data;
968 	struct vdpa_device *vdpa = v->vdpa;
969 	const struct vdpa_config_ops *ops = vdpa->config;
970 	struct vdpa_notification_area notify;
971 	unsigned long index = vma->vm_pgoff;
972 
973 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
974 		return -EINVAL;
975 	if ((vma->vm_flags & VM_SHARED) == 0)
976 		return -EINVAL;
977 	if (vma->vm_flags & VM_READ)
978 		return -EINVAL;
979 	if (index > 65535)
980 		return -EINVAL;
981 	if (!ops->get_vq_notification)
982 		return -ENOTSUPP;
983 
984 	/* To be safe and easily modelled by userspace, We only
985 	 * support the doorbell which sits on the page boundary and
986 	 * does not share the page with other registers.
987 	 */
988 	notify = ops->get_vq_notification(vdpa, index);
989 	if (notify.addr & (PAGE_SIZE - 1))
990 		return -EINVAL;
991 	if (vma->vm_end - vma->vm_start != notify.size)
992 		return -ENOTSUPP;
993 
994 	vma->vm_ops = &vhost_vdpa_vm_ops;
995 	return 0;
996 }
997 #endif /* CONFIG_MMU */
998 
999 static const struct file_operations vhost_vdpa_fops = {
1000 	.owner		= THIS_MODULE,
1001 	.open		= vhost_vdpa_open,
1002 	.release	= vhost_vdpa_release,
1003 	.write_iter	= vhost_vdpa_chr_write_iter,
1004 	.unlocked_ioctl	= vhost_vdpa_unlocked_ioctl,
1005 #ifdef CONFIG_MMU
1006 	.mmap		= vhost_vdpa_mmap,
1007 #endif /* CONFIG_MMU */
1008 	.compat_ioctl	= compat_ptr_ioctl,
1009 };
1010 
1011 static void vhost_vdpa_release_dev(struct device *device)
1012 {
1013 	struct vhost_vdpa *v =
1014 	       container_of(device, struct vhost_vdpa, dev);
1015 
1016 	ida_simple_remove(&vhost_vdpa_ida, v->minor);
1017 	kfree(v->vqs);
1018 	kfree(v);
1019 }
1020 
1021 static int vhost_vdpa_probe(struct vdpa_device *vdpa)
1022 {
1023 	const struct vdpa_config_ops *ops = vdpa->config;
1024 	struct vhost_vdpa *v;
1025 	int minor;
1026 	int r;
1027 
1028 	/* Currently, we only accept the network devices. */
1029 	if (ops->get_device_id(vdpa) != VIRTIO_ID_NET)
1030 		return -ENOTSUPP;
1031 
1032 	v = kzalloc(sizeof(*v), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1033 	if (!v)
1034 		return -ENOMEM;
1035 
1036 	minor = ida_simple_get(&vhost_vdpa_ida, 0,
1037 			       VHOST_VDPA_DEV_MAX, GFP_KERNEL);
1038 	if (minor < 0) {
1039 		kfree(v);
1040 		return minor;
1041 	}
1042 
1043 	atomic_set(&v->opened, 0);
1044 	v->minor = minor;
1045 	v->vdpa = vdpa;
1046 	v->nvqs = vdpa->nvqs;
1047 	v->virtio_id = ops->get_device_id(vdpa);
1048 
1049 	device_initialize(&v->dev);
1050 	v->dev.release = vhost_vdpa_release_dev;
1051 	v->dev.parent = &vdpa->dev;
1052 	v->dev.devt = MKDEV(MAJOR(vhost_vdpa_major), minor);
1053 	v->vqs = kmalloc_array(v->nvqs, sizeof(struct vhost_virtqueue),
1054 			       GFP_KERNEL);
1055 	if (!v->vqs) {
1056 		r = -ENOMEM;
1057 		goto err;
1058 	}
1059 
1060 	r = dev_set_name(&v->dev, "vhost-vdpa-%u", minor);
1061 	if (r)
1062 		goto err;
1063 
1064 	cdev_init(&v->cdev, &vhost_vdpa_fops);
1065 	v->cdev.owner = THIS_MODULE;
1066 
1067 	r = cdev_device_add(&v->cdev, &v->dev);
1068 	if (r)
1069 		goto err;
1070 
1071 	init_completion(&v->completion);
1072 	vdpa_set_drvdata(vdpa, v);
1073 
1074 	return 0;
1075 
1076 err:
1077 	put_device(&v->dev);
1078 	return r;
1079 }
1080 
1081 static void vhost_vdpa_remove(struct vdpa_device *vdpa)
1082 {
1083 	struct vhost_vdpa *v = vdpa_get_drvdata(vdpa);
1084 	int opened;
1085 
1086 	cdev_device_del(&v->cdev, &v->dev);
1087 
1088 	do {
1089 		opened = atomic_cmpxchg(&v->opened, 0, 1);
1090 		if (!opened)
1091 			break;
1092 		wait_for_completion(&v->completion);
1093 	} while (1);
1094 
1095 	put_device(&v->dev);
1096 }
1097 
1098 static struct vdpa_driver vhost_vdpa_driver = {
1099 	.driver = {
1100 		.name	= "vhost_vdpa",
1101 	},
1102 	.probe	= vhost_vdpa_probe,
1103 	.remove	= vhost_vdpa_remove,
1104 };
1105 
1106 static int __init vhost_vdpa_init(void)
1107 {
1108 	int r;
1109 
1110 	r = alloc_chrdev_region(&vhost_vdpa_major, 0, VHOST_VDPA_DEV_MAX,
1111 				"vhost-vdpa");
1112 	if (r)
1113 		goto err_alloc_chrdev;
1114 
1115 	r = vdpa_register_driver(&vhost_vdpa_driver);
1116 	if (r)
1117 		goto err_vdpa_register_driver;
1118 
1119 	return 0;
1120 
1121 err_vdpa_register_driver:
1122 	unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1123 err_alloc_chrdev:
1124 	return r;
1125 }
1126 module_init(vhost_vdpa_init);
1127 
1128 static void __exit vhost_vdpa_exit(void)
1129 {
1130 	vdpa_unregister_driver(&vhost_vdpa_driver);
1131 	unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1132 }
1133 module_exit(vhost_vdpa_exit);
1134 
1135 MODULE_VERSION("0.0.1");
1136 MODULE_LICENSE("GPL v2");
1137 MODULE_AUTHOR("Intel Corporation");
1138 MODULE_DESCRIPTION("vDPA-based vhost backend for virtio");
1139