xref: /openbmc/linux/drivers/gpu/drm/i915/gvt/kvmgt.c (revision 60772e48)
1 /*
2  * KVMGT - the implementation of Intel mediated pass-through framework for KVM
3  *
4  * Copyright(c) 2014-2016 Intel Corporation. All rights reserved.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice (including the next
14  * paragraph) shall be included in all copies or substantial portions of the
15  * Software.
16  *
17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23  * SOFTWARE.
24  *
25  * Authors:
26  *    Kevin Tian <kevin.tian@intel.com>
27  *    Jike Song <jike.song@intel.com>
28  *    Xiaoguang Chen <xiaoguang.chen@intel.com>
29  */
30 
31 #include <linux/init.h>
32 #include <linux/device.h>
33 #include <linux/mm.h>
34 #include <linux/mmu_context.h>
35 #include <linux/types.h>
36 #include <linux/list.h>
37 #include <linux/rbtree.h>
38 #include <linux/spinlock.h>
39 #include <linux/eventfd.h>
40 #include <linux/uuid.h>
41 #include <linux/kvm_host.h>
42 #include <linux/vfio.h>
43 #include <linux/mdev.h>
44 
45 #include "i915_drv.h"
46 #include "gvt.h"
47 
48 static const struct intel_gvt_ops *intel_gvt_ops;
49 
50 /* helper macros copied from vfio-pci */
51 #define VFIO_PCI_OFFSET_SHIFT   40
52 #define VFIO_PCI_OFFSET_TO_INDEX(off)   (off >> VFIO_PCI_OFFSET_SHIFT)
53 #define VFIO_PCI_INDEX_TO_OFFSET(index) ((u64)(index) << VFIO_PCI_OFFSET_SHIFT)
54 #define VFIO_PCI_OFFSET_MASK    (((u64)(1) << VFIO_PCI_OFFSET_SHIFT) - 1)
55 
56 #define OPREGION_SIGNATURE "IntelGraphicsMem"
57 
58 struct vfio_region;
59 struct intel_vgpu_regops {
60 	size_t (*rw)(struct intel_vgpu *vgpu, char *buf,
61 			size_t count, loff_t *ppos, bool iswrite);
62 	void (*release)(struct intel_vgpu *vgpu,
63 			struct vfio_region *region);
64 };
65 
66 struct vfio_region {
67 	u32				type;
68 	u32				subtype;
69 	size_t				size;
70 	u32				flags;
71 	const struct intel_vgpu_regops	*ops;
72 	void				*data;
73 };
74 
75 struct kvmgt_pgfn {
76 	gfn_t gfn;
77 	struct hlist_node hnode;
78 };
79 
80 struct kvmgt_guest_info {
81 	struct kvm *kvm;
82 	struct intel_vgpu *vgpu;
83 	struct kvm_page_track_notifier_node track_node;
84 #define NR_BKT (1 << 18)
85 	struct hlist_head ptable[NR_BKT];
86 #undef NR_BKT
87 };
88 
89 struct gvt_dma {
90 	struct rb_node node;
91 	gfn_t gfn;
92 	unsigned long iova;
93 };
94 
95 static inline bool handle_valid(unsigned long handle)
96 {
97 	return !!(handle & ~0xff);
98 }
99 
100 static int kvmgt_guest_init(struct mdev_device *mdev);
101 static void intel_vgpu_release_work(struct work_struct *work);
102 static bool kvmgt_guest_exit(struct kvmgt_guest_info *info);
103 
104 static int gvt_dma_map_iova(struct intel_vgpu *vgpu, kvm_pfn_t pfn,
105 		unsigned long *iova)
106 {
107 	struct page *page;
108 	struct device *dev = &vgpu->gvt->dev_priv->drm.pdev->dev;
109 	dma_addr_t daddr;
110 
111 	if (unlikely(!pfn_valid(pfn)))
112 		return -EFAULT;
113 
114 	page = pfn_to_page(pfn);
115 	daddr = dma_map_page(dev, page, 0, PAGE_SIZE,
116 			PCI_DMA_BIDIRECTIONAL);
117 	if (dma_mapping_error(dev, daddr))
118 		return -ENOMEM;
119 
120 	*iova = (unsigned long)(daddr >> PAGE_SHIFT);
121 	return 0;
122 }
123 
124 static void gvt_dma_unmap_iova(struct intel_vgpu *vgpu, unsigned long iova)
125 {
126 	struct device *dev = &vgpu->gvt->dev_priv->drm.pdev->dev;
127 	dma_addr_t daddr;
128 
129 	daddr = (dma_addr_t)(iova << PAGE_SHIFT);
130 	dma_unmap_page(dev, daddr, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
131 }
132 
133 static struct gvt_dma *__gvt_cache_find(struct intel_vgpu *vgpu, gfn_t gfn)
134 {
135 	struct rb_node *node = vgpu->vdev.cache.rb_node;
136 	struct gvt_dma *ret = NULL;
137 
138 	while (node) {
139 		struct gvt_dma *itr = rb_entry(node, struct gvt_dma, node);
140 
141 		if (gfn < itr->gfn)
142 			node = node->rb_left;
143 		else if (gfn > itr->gfn)
144 			node = node->rb_right;
145 		else {
146 			ret = itr;
147 			goto out;
148 		}
149 	}
150 
151 out:
152 	return ret;
153 }
154 
155 static unsigned long gvt_cache_find(struct intel_vgpu *vgpu, gfn_t gfn)
156 {
157 	struct gvt_dma *entry;
158 	unsigned long iova;
159 
160 	mutex_lock(&vgpu->vdev.cache_lock);
161 
162 	entry = __gvt_cache_find(vgpu, gfn);
163 	iova = (entry == NULL) ? INTEL_GVT_INVALID_ADDR : entry->iova;
164 
165 	mutex_unlock(&vgpu->vdev.cache_lock);
166 	return iova;
167 }
168 
169 static void gvt_cache_add(struct intel_vgpu *vgpu, gfn_t gfn,
170 		unsigned long iova)
171 {
172 	struct gvt_dma *new, *itr;
173 	struct rb_node **link = &vgpu->vdev.cache.rb_node, *parent = NULL;
174 
175 	new = kzalloc(sizeof(struct gvt_dma), GFP_KERNEL);
176 	if (!new)
177 		return;
178 
179 	new->gfn = gfn;
180 	new->iova = iova;
181 
182 	mutex_lock(&vgpu->vdev.cache_lock);
183 	while (*link) {
184 		parent = *link;
185 		itr = rb_entry(parent, struct gvt_dma, node);
186 
187 		if (gfn == itr->gfn)
188 			goto out;
189 		else if (gfn < itr->gfn)
190 			link = &parent->rb_left;
191 		else
192 			link = &parent->rb_right;
193 	}
194 
195 	rb_link_node(&new->node, parent, link);
196 	rb_insert_color(&new->node, &vgpu->vdev.cache);
197 	mutex_unlock(&vgpu->vdev.cache_lock);
198 	return;
199 
200 out:
201 	mutex_unlock(&vgpu->vdev.cache_lock);
202 	kfree(new);
203 }
204 
205 static void __gvt_cache_remove_entry(struct intel_vgpu *vgpu,
206 				struct gvt_dma *entry)
207 {
208 	rb_erase(&entry->node, &vgpu->vdev.cache);
209 	kfree(entry);
210 }
211 
212 static void gvt_cache_remove(struct intel_vgpu *vgpu, gfn_t gfn)
213 {
214 	struct device *dev = mdev_dev(vgpu->vdev.mdev);
215 	struct gvt_dma *this;
216 	unsigned long g1;
217 	int rc;
218 
219 	mutex_lock(&vgpu->vdev.cache_lock);
220 	this  = __gvt_cache_find(vgpu, gfn);
221 	if (!this) {
222 		mutex_unlock(&vgpu->vdev.cache_lock);
223 		return;
224 	}
225 
226 	g1 = gfn;
227 	gvt_dma_unmap_iova(vgpu, this->iova);
228 	rc = vfio_unpin_pages(dev, &g1, 1);
229 	WARN_ON(rc != 1);
230 	__gvt_cache_remove_entry(vgpu, this);
231 	mutex_unlock(&vgpu->vdev.cache_lock);
232 }
233 
234 static void gvt_cache_init(struct intel_vgpu *vgpu)
235 {
236 	vgpu->vdev.cache = RB_ROOT;
237 	mutex_init(&vgpu->vdev.cache_lock);
238 }
239 
240 static void gvt_cache_destroy(struct intel_vgpu *vgpu)
241 {
242 	struct gvt_dma *dma;
243 	struct rb_node *node = NULL;
244 	struct device *dev = mdev_dev(vgpu->vdev.mdev);
245 	unsigned long gfn;
246 
247 	for (;;) {
248 		mutex_lock(&vgpu->vdev.cache_lock);
249 		node = rb_first(&vgpu->vdev.cache);
250 		if (!node) {
251 			mutex_unlock(&vgpu->vdev.cache_lock);
252 			break;
253 		}
254 		dma = rb_entry(node, struct gvt_dma, node);
255 		gvt_dma_unmap_iova(vgpu, dma->iova);
256 		gfn = dma->gfn;
257 		__gvt_cache_remove_entry(vgpu, dma);
258 		mutex_unlock(&vgpu->vdev.cache_lock);
259 		vfio_unpin_pages(dev, &gfn, 1);
260 	}
261 }
262 
263 static void kvmgt_protect_table_init(struct kvmgt_guest_info *info)
264 {
265 	hash_init(info->ptable);
266 }
267 
268 static void kvmgt_protect_table_destroy(struct kvmgt_guest_info *info)
269 {
270 	struct kvmgt_pgfn *p;
271 	struct hlist_node *tmp;
272 	int i;
273 
274 	hash_for_each_safe(info->ptable, i, tmp, p, hnode) {
275 		hash_del(&p->hnode);
276 		kfree(p);
277 	}
278 }
279 
280 static struct kvmgt_pgfn *
281 __kvmgt_protect_table_find(struct kvmgt_guest_info *info, gfn_t gfn)
282 {
283 	struct kvmgt_pgfn *p, *res = NULL;
284 
285 	hash_for_each_possible(info->ptable, p, hnode, gfn) {
286 		if (gfn == p->gfn) {
287 			res = p;
288 			break;
289 		}
290 	}
291 
292 	return res;
293 }
294 
295 static bool kvmgt_gfn_is_write_protected(struct kvmgt_guest_info *info,
296 				gfn_t gfn)
297 {
298 	struct kvmgt_pgfn *p;
299 
300 	p = __kvmgt_protect_table_find(info, gfn);
301 	return !!p;
302 }
303 
304 static void kvmgt_protect_table_add(struct kvmgt_guest_info *info, gfn_t gfn)
305 {
306 	struct kvmgt_pgfn *p;
307 
308 	if (kvmgt_gfn_is_write_protected(info, gfn))
309 		return;
310 
311 	p = kzalloc(sizeof(struct kvmgt_pgfn), GFP_ATOMIC);
312 	if (WARN(!p, "gfn: 0x%llx\n", gfn))
313 		return;
314 
315 	p->gfn = gfn;
316 	hash_add(info->ptable, &p->hnode, gfn);
317 }
318 
319 static void kvmgt_protect_table_del(struct kvmgt_guest_info *info,
320 				gfn_t gfn)
321 {
322 	struct kvmgt_pgfn *p;
323 
324 	p = __kvmgt_protect_table_find(info, gfn);
325 	if (p) {
326 		hash_del(&p->hnode);
327 		kfree(p);
328 	}
329 }
330 
331 static size_t intel_vgpu_reg_rw_opregion(struct intel_vgpu *vgpu, char *buf,
332 		size_t count, loff_t *ppos, bool iswrite)
333 {
334 	unsigned int i = VFIO_PCI_OFFSET_TO_INDEX(*ppos) -
335 			VFIO_PCI_NUM_REGIONS;
336 	void *base = vgpu->vdev.region[i].data;
337 	loff_t pos = *ppos & VFIO_PCI_OFFSET_MASK;
338 
339 	if (pos >= vgpu->vdev.region[i].size || iswrite) {
340 		gvt_vgpu_err("invalid op or offset for Intel vgpu OpRegion\n");
341 		return -EINVAL;
342 	}
343 	count = min(count, (size_t)(vgpu->vdev.region[i].size - pos));
344 	memcpy(buf, base + pos, count);
345 
346 	return count;
347 }
348 
349 static void intel_vgpu_reg_release_opregion(struct intel_vgpu *vgpu,
350 		struct vfio_region *region)
351 {
352 }
353 
354 static const struct intel_vgpu_regops intel_vgpu_regops_opregion = {
355 	.rw = intel_vgpu_reg_rw_opregion,
356 	.release = intel_vgpu_reg_release_opregion,
357 };
358 
359 static int intel_vgpu_register_reg(struct intel_vgpu *vgpu,
360 		unsigned int type, unsigned int subtype,
361 		const struct intel_vgpu_regops *ops,
362 		size_t size, u32 flags, void *data)
363 {
364 	struct vfio_region *region;
365 
366 	region = krealloc(vgpu->vdev.region,
367 			(vgpu->vdev.num_regions + 1) * sizeof(*region),
368 			GFP_KERNEL);
369 	if (!region)
370 		return -ENOMEM;
371 
372 	vgpu->vdev.region = region;
373 	vgpu->vdev.region[vgpu->vdev.num_regions].type = type;
374 	vgpu->vdev.region[vgpu->vdev.num_regions].subtype = subtype;
375 	vgpu->vdev.region[vgpu->vdev.num_regions].ops = ops;
376 	vgpu->vdev.region[vgpu->vdev.num_regions].size = size;
377 	vgpu->vdev.region[vgpu->vdev.num_regions].flags = flags;
378 	vgpu->vdev.region[vgpu->vdev.num_regions].data = data;
379 	vgpu->vdev.num_regions++;
380 	return 0;
381 }
382 
383 static int kvmgt_get_vfio_device(void *p_vgpu)
384 {
385 	struct intel_vgpu *vgpu = (struct intel_vgpu *)p_vgpu;
386 
387 	vgpu->vdev.vfio_device = vfio_device_get_from_dev(
388 		mdev_dev(vgpu->vdev.mdev));
389 	if (!vgpu->vdev.vfio_device) {
390 		gvt_vgpu_err("failed to get vfio device\n");
391 		return -ENODEV;
392 	}
393 	return 0;
394 }
395 
396 
397 static int kvmgt_set_opregion(void *p_vgpu)
398 {
399 	struct intel_vgpu *vgpu = (struct intel_vgpu *)p_vgpu;
400 	void *base;
401 	int ret;
402 
403 	/* Each vgpu has its own opregion, although VFIO would create another
404 	 * one later. This one is used to expose opregion to VFIO. And the
405 	 * other one created by VFIO later, is used by guest actually.
406 	 */
407 	base = vgpu_opregion(vgpu)->va;
408 	if (!base)
409 		return -ENOMEM;
410 
411 	if (memcmp(base, OPREGION_SIGNATURE, 16)) {
412 		memunmap(base);
413 		return -EINVAL;
414 	}
415 
416 	ret = intel_vgpu_register_reg(vgpu,
417 			PCI_VENDOR_ID_INTEL | VFIO_REGION_TYPE_PCI_VENDOR_TYPE,
418 			VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION,
419 			&intel_vgpu_regops_opregion, OPREGION_SIZE,
420 			VFIO_REGION_INFO_FLAG_READ, base);
421 
422 	return ret;
423 }
424 
425 static void kvmgt_put_vfio_device(void *vgpu)
426 {
427 	if (WARN_ON(!((struct intel_vgpu *)vgpu)->vdev.vfio_device))
428 		return;
429 
430 	vfio_device_put(((struct intel_vgpu *)vgpu)->vdev.vfio_device);
431 }
432 
433 static int intel_vgpu_create(struct kobject *kobj, struct mdev_device *mdev)
434 {
435 	struct intel_vgpu *vgpu = NULL;
436 	struct intel_vgpu_type *type;
437 	struct device *pdev;
438 	void *gvt;
439 	int ret;
440 
441 	pdev = mdev_parent_dev(mdev);
442 	gvt = kdev_to_i915(pdev)->gvt;
443 
444 	type = intel_gvt_ops->gvt_find_vgpu_type(gvt, kobject_name(kobj));
445 	if (!type) {
446 		gvt_vgpu_err("failed to find type %s to create\n",
447 						kobject_name(kobj));
448 		ret = -EINVAL;
449 		goto out;
450 	}
451 
452 	vgpu = intel_gvt_ops->vgpu_create(gvt, type);
453 	if (IS_ERR_OR_NULL(vgpu)) {
454 		ret = vgpu == NULL ? -EFAULT : PTR_ERR(vgpu);
455 		gvt_vgpu_err("failed to create intel vgpu: %d\n", ret);
456 		goto out;
457 	}
458 
459 	INIT_WORK(&vgpu->vdev.release_work, intel_vgpu_release_work);
460 
461 	vgpu->vdev.mdev = mdev;
462 	mdev_set_drvdata(mdev, vgpu);
463 
464 	gvt_dbg_core("intel_vgpu_create succeeded for mdev: %s\n",
465 		     dev_name(mdev_dev(mdev)));
466 	ret = 0;
467 
468 out:
469 	return ret;
470 }
471 
472 static int intel_vgpu_remove(struct mdev_device *mdev)
473 {
474 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
475 
476 	if (handle_valid(vgpu->handle))
477 		return -EBUSY;
478 
479 	intel_gvt_ops->vgpu_destroy(vgpu);
480 	return 0;
481 }
482 
483 static int intel_vgpu_iommu_notifier(struct notifier_block *nb,
484 				     unsigned long action, void *data)
485 {
486 	struct intel_vgpu *vgpu = container_of(nb,
487 					struct intel_vgpu,
488 					vdev.iommu_notifier);
489 
490 	if (action == VFIO_IOMMU_NOTIFY_DMA_UNMAP) {
491 		struct vfio_iommu_type1_dma_unmap *unmap = data;
492 		unsigned long gfn, end_gfn;
493 
494 		gfn = unmap->iova >> PAGE_SHIFT;
495 		end_gfn = gfn + unmap->size / PAGE_SIZE;
496 
497 		while (gfn < end_gfn)
498 			gvt_cache_remove(vgpu, gfn++);
499 	}
500 
501 	return NOTIFY_OK;
502 }
503 
504 static int intel_vgpu_group_notifier(struct notifier_block *nb,
505 				     unsigned long action, void *data)
506 {
507 	struct intel_vgpu *vgpu = container_of(nb,
508 					struct intel_vgpu,
509 					vdev.group_notifier);
510 
511 	/* the only action we care about */
512 	if (action == VFIO_GROUP_NOTIFY_SET_KVM) {
513 		vgpu->vdev.kvm = data;
514 
515 		if (!data)
516 			schedule_work(&vgpu->vdev.release_work);
517 	}
518 
519 	return NOTIFY_OK;
520 }
521 
522 static int intel_vgpu_open(struct mdev_device *mdev)
523 {
524 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
525 	unsigned long events;
526 	int ret;
527 
528 	vgpu->vdev.iommu_notifier.notifier_call = intel_vgpu_iommu_notifier;
529 	vgpu->vdev.group_notifier.notifier_call = intel_vgpu_group_notifier;
530 
531 	events = VFIO_IOMMU_NOTIFY_DMA_UNMAP;
532 	ret = vfio_register_notifier(mdev_dev(mdev), VFIO_IOMMU_NOTIFY, &events,
533 				&vgpu->vdev.iommu_notifier);
534 	if (ret != 0) {
535 		gvt_vgpu_err("vfio_register_notifier for iommu failed: %d\n",
536 			ret);
537 		goto out;
538 	}
539 
540 	events = VFIO_GROUP_NOTIFY_SET_KVM;
541 	ret = vfio_register_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY, &events,
542 				&vgpu->vdev.group_notifier);
543 	if (ret != 0) {
544 		gvt_vgpu_err("vfio_register_notifier for group failed: %d\n",
545 			ret);
546 		goto undo_iommu;
547 	}
548 
549 	ret = kvmgt_guest_init(mdev);
550 	if (ret)
551 		goto undo_group;
552 
553 	intel_gvt_ops->vgpu_activate(vgpu);
554 
555 	atomic_set(&vgpu->vdev.released, 0);
556 	return ret;
557 
558 undo_group:
559 	vfio_unregister_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY,
560 					&vgpu->vdev.group_notifier);
561 
562 undo_iommu:
563 	vfio_unregister_notifier(mdev_dev(mdev), VFIO_IOMMU_NOTIFY,
564 					&vgpu->vdev.iommu_notifier);
565 out:
566 	return ret;
567 }
568 
569 static void __intel_vgpu_release(struct intel_vgpu *vgpu)
570 {
571 	struct kvmgt_guest_info *info;
572 	int ret;
573 
574 	if (!handle_valid(vgpu->handle))
575 		return;
576 
577 	if (atomic_cmpxchg(&vgpu->vdev.released, 0, 1))
578 		return;
579 
580 	intel_gvt_ops->vgpu_deactivate(vgpu);
581 
582 	ret = vfio_unregister_notifier(mdev_dev(vgpu->vdev.mdev), VFIO_IOMMU_NOTIFY,
583 					&vgpu->vdev.iommu_notifier);
584 	WARN(ret, "vfio_unregister_notifier for iommu failed: %d\n", ret);
585 
586 	ret = vfio_unregister_notifier(mdev_dev(vgpu->vdev.mdev), VFIO_GROUP_NOTIFY,
587 					&vgpu->vdev.group_notifier);
588 	WARN(ret, "vfio_unregister_notifier for group failed: %d\n", ret);
589 
590 	info = (struct kvmgt_guest_info *)vgpu->handle;
591 	kvmgt_guest_exit(info);
592 
593 	vgpu->vdev.kvm = NULL;
594 	vgpu->handle = 0;
595 }
596 
597 static void intel_vgpu_release(struct mdev_device *mdev)
598 {
599 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
600 
601 	__intel_vgpu_release(vgpu);
602 }
603 
604 static void intel_vgpu_release_work(struct work_struct *work)
605 {
606 	struct intel_vgpu *vgpu = container_of(work, struct intel_vgpu,
607 					vdev.release_work);
608 
609 	__intel_vgpu_release(vgpu);
610 }
611 
612 static uint64_t intel_vgpu_get_bar_addr(struct intel_vgpu *vgpu, int bar)
613 {
614 	u32 start_lo, start_hi;
615 	u32 mem_type;
616 
617 	start_lo = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
618 			PCI_BASE_ADDRESS_MEM_MASK;
619 	mem_type = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
620 			PCI_BASE_ADDRESS_MEM_TYPE_MASK;
621 
622 	switch (mem_type) {
623 	case PCI_BASE_ADDRESS_MEM_TYPE_64:
624 		start_hi = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space
625 						+ bar + 4));
626 		break;
627 	case PCI_BASE_ADDRESS_MEM_TYPE_32:
628 	case PCI_BASE_ADDRESS_MEM_TYPE_1M:
629 		/* 1M mem BAR treated as 32-bit BAR */
630 	default:
631 		/* mem unknown type treated as 32-bit BAR */
632 		start_hi = 0;
633 		break;
634 	}
635 
636 	return ((u64)start_hi << 32) | start_lo;
637 }
638 
639 static int intel_vgpu_bar_rw(struct intel_vgpu *vgpu, int bar, uint64_t off,
640 			     void *buf, unsigned int count, bool is_write)
641 {
642 	uint64_t bar_start = intel_vgpu_get_bar_addr(vgpu, bar);
643 	int ret;
644 
645 	if (is_write)
646 		ret = intel_gvt_ops->emulate_mmio_write(vgpu,
647 					bar_start + off, buf, count);
648 	else
649 		ret = intel_gvt_ops->emulate_mmio_read(vgpu,
650 					bar_start + off, buf, count);
651 	return ret;
652 }
653 
654 static inline bool intel_vgpu_in_aperture(struct intel_vgpu *vgpu, uint64_t off)
655 {
656 	return off >= vgpu_aperture_offset(vgpu) &&
657 	       off < vgpu_aperture_offset(vgpu) + vgpu_aperture_sz(vgpu);
658 }
659 
660 static int intel_vgpu_aperture_rw(struct intel_vgpu *vgpu, uint64_t off,
661 		void *buf, unsigned long count, bool is_write)
662 {
663 	void *aperture_va;
664 
665 	if (!intel_vgpu_in_aperture(vgpu, off) ||
666 	    !intel_vgpu_in_aperture(vgpu, off + count)) {
667 		gvt_vgpu_err("Invalid aperture offset %llu\n", off);
668 		return -EINVAL;
669 	}
670 
671 	aperture_va = io_mapping_map_wc(&vgpu->gvt->dev_priv->ggtt.iomap,
672 					ALIGN_DOWN(off, PAGE_SIZE),
673 					count + offset_in_page(off));
674 	if (!aperture_va)
675 		return -EIO;
676 
677 	if (is_write)
678 		memcpy(aperture_va + offset_in_page(off), buf, count);
679 	else
680 		memcpy(buf, aperture_va + offset_in_page(off), count);
681 
682 	io_mapping_unmap(aperture_va);
683 
684 	return 0;
685 }
686 
687 static ssize_t intel_vgpu_rw(struct mdev_device *mdev, char *buf,
688 			size_t count, loff_t *ppos, bool is_write)
689 {
690 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
691 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
692 	uint64_t pos = *ppos & VFIO_PCI_OFFSET_MASK;
693 	int ret = -EINVAL;
694 
695 
696 	if (index >= VFIO_PCI_NUM_REGIONS + vgpu->vdev.num_regions) {
697 		gvt_vgpu_err("invalid index: %u\n", index);
698 		return -EINVAL;
699 	}
700 
701 	switch (index) {
702 	case VFIO_PCI_CONFIG_REGION_INDEX:
703 		if (is_write)
704 			ret = intel_gvt_ops->emulate_cfg_write(vgpu, pos,
705 						buf, count);
706 		else
707 			ret = intel_gvt_ops->emulate_cfg_read(vgpu, pos,
708 						buf, count);
709 		break;
710 	case VFIO_PCI_BAR0_REGION_INDEX:
711 		ret = intel_vgpu_bar_rw(vgpu, PCI_BASE_ADDRESS_0, pos,
712 					buf, count, is_write);
713 		break;
714 	case VFIO_PCI_BAR2_REGION_INDEX:
715 		ret = intel_vgpu_aperture_rw(vgpu, pos, buf, count, is_write);
716 		break;
717 	case VFIO_PCI_BAR1_REGION_INDEX:
718 	case VFIO_PCI_BAR3_REGION_INDEX:
719 	case VFIO_PCI_BAR4_REGION_INDEX:
720 	case VFIO_PCI_BAR5_REGION_INDEX:
721 	case VFIO_PCI_VGA_REGION_INDEX:
722 	case VFIO_PCI_ROM_REGION_INDEX:
723 		break;
724 	default:
725 		if (index >= VFIO_PCI_NUM_REGIONS + vgpu->vdev.num_regions)
726 			return -EINVAL;
727 
728 		index -= VFIO_PCI_NUM_REGIONS;
729 		return vgpu->vdev.region[index].ops->rw(vgpu, buf, count,
730 				ppos, is_write);
731 	}
732 
733 	return ret == 0 ? count : ret;
734 }
735 
736 static bool gtt_entry(struct mdev_device *mdev, loff_t *ppos)
737 {
738 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
739 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
740 	struct intel_gvt *gvt = vgpu->gvt;
741 	int offset;
742 
743 	/* Only allow MMIO GGTT entry access */
744 	if (index != PCI_BASE_ADDRESS_0)
745 		return false;
746 
747 	offset = (u64)(*ppos & VFIO_PCI_OFFSET_MASK) -
748 		intel_vgpu_get_bar_gpa(vgpu, PCI_BASE_ADDRESS_0);
749 
750 	return (offset >= gvt->device_info.gtt_start_offset &&
751 		offset < gvt->device_info.gtt_start_offset + gvt_ggtt_sz(gvt)) ?
752 			true : false;
753 }
754 
755 static ssize_t intel_vgpu_read(struct mdev_device *mdev, char __user *buf,
756 			size_t count, loff_t *ppos)
757 {
758 	unsigned int done = 0;
759 	int ret;
760 
761 	while (count) {
762 		size_t filled;
763 
764 		/* Only support GGTT entry 8 bytes read */
765 		if (count >= 8 && !(*ppos % 8) &&
766 			gtt_entry(mdev, ppos)) {
767 			u64 val;
768 
769 			ret = intel_vgpu_rw(mdev, (char *)&val, sizeof(val),
770 					ppos, false);
771 			if (ret <= 0)
772 				goto read_err;
773 
774 			if (copy_to_user(buf, &val, sizeof(val)))
775 				goto read_err;
776 
777 			filled = 8;
778 		} else if (count >= 4 && !(*ppos % 4)) {
779 			u32 val;
780 
781 			ret = intel_vgpu_rw(mdev, (char *)&val, sizeof(val),
782 					ppos, false);
783 			if (ret <= 0)
784 				goto read_err;
785 
786 			if (copy_to_user(buf, &val, sizeof(val)))
787 				goto read_err;
788 
789 			filled = 4;
790 		} else if (count >= 2 && !(*ppos % 2)) {
791 			u16 val;
792 
793 			ret = intel_vgpu_rw(mdev, (char *)&val, sizeof(val),
794 					ppos, false);
795 			if (ret <= 0)
796 				goto read_err;
797 
798 			if (copy_to_user(buf, &val, sizeof(val)))
799 				goto read_err;
800 
801 			filled = 2;
802 		} else {
803 			u8 val;
804 
805 			ret = intel_vgpu_rw(mdev, &val, sizeof(val), ppos,
806 					false);
807 			if (ret <= 0)
808 				goto read_err;
809 
810 			if (copy_to_user(buf, &val, sizeof(val)))
811 				goto read_err;
812 
813 			filled = 1;
814 		}
815 
816 		count -= filled;
817 		done += filled;
818 		*ppos += filled;
819 		buf += filled;
820 	}
821 
822 	return done;
823 
824 read_err:
825 	return -EFAULT;
826 }
827 
828 static ssize_t intel_vgpu_write(struct mdev_device *mdev,
829 				const char __user *buf,
830 				size_t count, loff_t *ppos)
831 {
832 	unsigned int done = 0;
833 	int ret;
834 
835 	while (count) {
836 		size_t filled;
837 
838 		/* Only support GGTT entry 8 bytes write */
839 		if (count >= 8 && !(*ppos % 8) &&
840 			gtt_entry(mdev, ppos)) {
841 			u64 val;
842 
843 			if (copy_from_user(&val, buf, sizeof(val)))
844 				goto write_err;
845 
846 			ret = intel_vgpu_rw(mdev, (char *)&val, sizeof(val),
847 					ppos, true);
848 			if (ret <= 0)
849 				goto write_err;
850 
851 			filled = 8;
852 		} else if (count >= 4 && !(*ppos % 4)) {
853 			u32 val;
854 
855 			if (copy_from_user(&val, buf, sizeof(val)))
856 				goto write_err;
857 
858 			ret = intel_vgpu_rw(mdev, (char *)&val, sizeof(val),
859 					ppos, true);
860 			if (ret <= 0)
861 				goto write_err;
862 
863 			filled = 4;
864 		} else if (count >= 2 && !(*ppos % 2)) {
865 			u16 val;
866 
867 			if (copy_from_user(&val, buf, sizeof(val)))
868 				goto write_err;
869 
870 			ret = intel_vgpu_rw(mdev, (char *)&val,
871 					sizeof(val), ppos, true);
872 			if (ret <= 0)
873 				goto write_err;
874 
875 			filled = 2;
876 		} else {
877 			u8 val;
878 
879 			if (copy_from_user(&val, buf, sizeof(val)))
880 				goto write_err;
881 
882 			ret = intel_vgpu_rw(mdev, &val, sizeof(val),
883 					ppos, true);
884 			if (ret <= 0)
885 				goto write_err;
886 
887 			filled = 1;
888 		}
889 
890 		count -= filled;
891 		done += filled;
892 		*ppos += filled;
893 		buf += filled;
894 	}
895 
896 	return done;
897 write_err:
898 	return -EFAULT;
899 }
900 
901 static int intel_vgpu_mmap(struct mdev_device *mdev, struct vm_area_struct *vma)
902 {
903 	unsigned int index;
904 	u64 virtaddr;
905 	unsigned long req_size, pgoff = 0;
906 	pgprot_t pg_prot;
907 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
908 
909 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
910 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
911 		return -EINVAL;
912 
913 	if (vma->vm_end < vma->vm_start)
914 		return -EINVAL;
915 	if ((vma->vm_flags & VM_SHARED) == 0)
916 		return -EINVAL;
917 	if (index != VFIO_PCI_BAR2_REGION_INDEX)
918 		return -EINVAL;
919 
920 	pg_prot = vma->vm_page_prot;
921 	virtaddr = vma->vm_start;
922 	req_size = vma->vm_end - vma->vm_start;
923 	pgoff = vgpu_aperture_pa_base(vgpu) >> PAGE_SHIFT;
924 
925 	return remap_pfn_range(vma, virtaddr, pgoff, req_size, pg_prot);
926 }
927 
928 static int intel_vgpu_get_irq_count(struct intel_vgpu *vgpu, int type)
929 {
930 	if (type == VFIO_PCI_INTX_IRQ_INDEX || type == VFIO_PCI_MSI_IRQ_INDEX)
931 		return 1;
932 
933 	return 0;
934 }
935 
936 static int intel_vgpu_set_intx_mask(struct intel_vgpu *vgpu,
937 			unsigned int index, unsigned int start,
938 			unsigned int count, uint32_t flags,
939 			void *data)
940 {
941 	return 0;
942 }
943 
944 static int intel_vgpu_set_intx_unmask(struct intel_vgpu *vgpu,
945 			unsigned int index, unsigned int start,
946 			unsigned int count, uint32_t flags, void *data)
947 {
948 	return 0;
949 }
950 
951 static int intel_vgpu_set_intx_trigger(struct intel_vgpu *vgpu,
952 		unsigned int index, unsigned int start, unsigned int count,
953 		uint32_t flags, void *data)
954 {
955 	return 0;
956 }
957 
958 static int intel_vgpu_set_msi_trigger(struct intel_vgpu *vgpu,
959 		unsigned int index, unsigned int start, unsigned int count,
960 		uint32_t flags, void *data)
961 {
962 	struct eventfd_ctx *trigger;
963 
964 	if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
965 		int fd = *(int *)data;
966 
967 		trigger = eventfd_ctx_fdget(fd);
968 		if (IS_ERR(trigger)) {
969 			gvt_vgpu_err("eventfd_ctx_fdget failed\n");
970 			return PTR_ERR(trigger);
971 		}
972 		vgpu->vdev.msi_trigger = trigger;
973 	}
974 
975 	return 0;
976 }
977 
978 static int intel_vgpu_set_irqs(struct intel_vgpu *vgpu, uint32_t flags,
979 		unsigned int index, unsigned int start, unsigned int count,
980 		void *data)
981 {
982 	int (*func)(struct intel_vgpu *vgpu, unsigned int index,
983 			unsigned int start, unsigned int count, uint32_t flags,
984 			void *data) = NULL;
985 
986 	switch (index) {
987 	case VFIO_PCI_INTX_IRQ_INDEX:
988 		switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
989 		case VFIO_IRQ_SET_ACTION_MASK:
990 			func = intel_vgpu_set_intx_mask;
991 			break;
992 		case VFIO_IRQ_SET_ACTION_UNMASK:
993 			func = intel_vgpu_set_intx_unmask;
994 			break;
995 		case VFIO_IRQ_SET_ACTION_TRIGGER:
996 			func = intel_vgpu_set_intx_trigger;
997 			break;
998 		}
999 		break;
1000 	case VFIO_PCI_MSI_IRQ_INDEX:
1001 		switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
1002 		case VFIO_IRQ_SET_ACTION_MASK:
1003 		case VFIO_IRQ_SET_ACTION_UNMASK:
1004 			/* XXX Need masking support exported */
1005 			break;
1006 		case VFIO_IRQ_SET_ACTION_TRIGGER:
1007 			func = intel_vgpu_set_msi_trigger;
1008 			break;
1009 		}
1010 		break;
1011 	}
1012 
1013 	if (!func)
1014 		return -ENOTTY;
1015 
1016 	return func(vgpu, index, start, count, flags, data);
1017 }
1018 
1019 static long intel_vgpu_ioctl(struct mdev_device *mdev, unsigned int cmd,
1020 			     unsigned long arg)
1021 {
1022 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
1023 	unsigned long minsz;
1024 
1025 	gvt_dbg_core("vgpu%d ioctl, cmd: %d\n", vgpu->id, cmd);
1026 
1027 	if (cmd == VFIO_DEVICE_GET_INFO) {
1028 		struct vfio_device_info info;
1029 
1030 		minsz = offsetofend(struct vfio_device_info, num_irqs);
1031 
1032 		if (copy_from_user(&info, (void __user *)arg, minsz))
1033 			return -EFAULT;
1034 
1035 		if (info.argsz < minsz)
1036 			return -EINVAL;
1037 
1038 		info.flags = VFIO_DEVICE_FLAGS_PCI;
1039 		info.flags |= VFIO_DEVICE_FLAGS_RESET;
1040 		info.num_regions = VFIO_PCI_NUM_REGIONS +
1041 				vgpu->vdev.num_regions;
1042 		info.num_irqs = VFIO_PCI_NUM_IRQS;
1043 
1044 		return copy_to_user((void __user *)arg, &info, minsz) ?
1045 			-EFAULT : 0;
1046 
1047 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
1048 		struct vfio_region_info info;
1049 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
1050 		int i, ret;
1051 		struct vfio_region_info_cap_sparse_mmap *sparse = NULL;
1052 		size_t size;
1053 		int nr_areas = 1;
1054 		int cap_type_id;
1055 
1056 		minsz = offsetofend(struct vfio_region_info, offset);
1057 
1058 		if (copy_from_user(&info, (void __user *)arg, minsz))
1059 			return -EFAULT;
1060 
1061 		if (info.argsz < minsz)
1062 			return -EINVAL;
1063 
1064 		switch (info.index) {
1065 		case VFIO_PCI_CONFIG_REGION_INDEX:
1066 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1067 			info.size = vgpu->gvt->device_info.cfg_space_size;
1068 			info.flags = VFIO_REGION_INFO_FLAG_READ |
1069 				     VFIO_REGION_INFO_FLAG_WRITE;
1070 			break;
1071 		case VFIO_PCI_BAR0_REGION_INDEX:
1072 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1073 			info.size = vgpu->cfg_space.bar[info.index].size;
1074 			if (!info.size) {
1075 				info.flags = 0;
1076 				break;
1077 			}
1078 
1079 			info.flags = VFIO_REGION_INFO_FLAG_READ |
1080 				     VFIO_REGION_INFO_FLAG_WRITE;
1081 			break;
1082 		case VFIO_PCI_BAR1_REGION_INDEX:
1083 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1084 			info.size = 0;
1085 			info.flags = 0;
1086 			break;
1087 		case VFIO_PCI_BAR2_REGION_INDEX:
1088 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1089 			info.flags = VFIO_REGION_INFO_FLAG_CAPS |
1090 					VFIO_REGION_INFO_FLAG_MMAP |
1091 					VFIO_REGION_INFO_FLAG_READ |
1092 					VFIO_REGION_INFO_FLAG_WRITE;
1093 			info.size = gvt_aperture_sz(vgpu->gvt);
1094 
1095 			size = sizeof(*sparse) +
1096 					(nr_areas * sizeof(*sparse->areas));
1097 			sparse = kzalloc(size, GFP_KERNEL);
1098 			if (!sparse)
1099 				return -ENOMEM;
1100 
1101 			sparse->header.id = VFIO_REGION_INFO_CAP_SPARSE_MMAP;
1102 			sparse->header.version = 1;
1103 			sparse->nr_areas = nr_areas;
1104 			cap_type_id = VFIO_REGION_INFO_CAP_SPARSE_MMAP;
1105 			sparse->areas[0].offset =
1106 					PAGE_ALIGN(vgpu_aperture_offset(vgpu));
1107 			sparse->areas[0].size = vgpu_aperture_sz(vgpu);
1108 			break;
1109 
1110 		case VFIO_PCI_BAR3_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1111 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1112 			info.size = 0;
1113 			info.flags = 0;
1114 
1115 			gvt_dbg_core("get region info bar:%d\n", info.index);
1116 			break;
1117 
1118 		case VFIO_PCI_ROM_REGION_INDEX:
1119 		case VFIO_PCI_VGA_REGION_INDEX:
1120 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1121 			info.size = 0;
1122 			info.flags = 0;
1123 
1124 			gvt_dbg_core("get region info index:%d\n", info.index);
1125 			break;
1126 		default:
1127 			{
1128 				struct vfio_region_info_cap_type cap_type = {
1129 					.header.id = VFIO_REGION_INFO_CAP_TYPE,
1130 					.header.version = 1 };
1131 
1132 				if (info.index >= VFIO_PCI_NUM_REGIONS +
1133 						vgpu->vdev.num_regions)
1134 					return -EINVAL;
1135 
1136 				i = info.index - VFIO_PCI_NUM_REGIONS;
1137 
1138 				info.offset =
1139 					VFIO_PCI_INDEX_TO_OFFSET(info.index);
1140 				info.size = vgpu->vdev.region[i].size;
1141 				info.flags = vgpu->vdev.region[i].flags;
1142 
1143 				cap_type.type = vgpu->vdev.region[i].type;
1144 				cap_type.subtype = vgpu->vdev.region[i].subtype;
1145 
1146 				ret = vfio_info_add_capability(&caps,
1147 							&cap_type.header,
1148 							sizeof(cap_type));
1149 				if (ret)
1150 					return ret;
1151 			}
1152 		}
1153 
1154 		if ((info.flags & VFIO_REGION_INFO_FLAG_CAPS) && sparse) {
1155 			switch (cap_type_id) {
1156 			case VFIO_REGION_INFO_CAP_SPARSE_MMAP:
1157 				ret = vfio_info_add_capability(&caps,
1158 					&sparse->header, sizeof(*sparse) +
1159 					(sparse->nr_areas *
1160 						sizeof(*sparse->areas)));
1161 				kfree(sparse);
1162 				if (ret)
1163 					return ret;
1164 				break;
1165 			default:
1166 				return -EINVAL;
1167 			}
1168 		}
1169 
1170 		if (caps.size) {
1171 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
1172 			if (info.argsz < sizeof(info) + caps.size) {
1173 				info.argsz = sizeof(info) + caps.size;
1174 				info.cap_offset = 0;
1175 			} else {
1176 				vfio_info_cap_shift(&caps, sizeof(info));
1177 				if (copy_to_user((void __user *)arg +
1178 						  sizeof(info), caps.buf,
1179 						  caps.size)) {
1180 					kfree(caps.buf);
1181 					return -EFAULT;
1182 				}
1183 				info.cap_offset = sizeof(info);
1184 			}
1185 
1186 			kfree(caps.buf);
1187 		}
1188 
1189 		return copy_to_user((void __user *)arg, &info, minsz) ?
1190 			-EFAULT : 0;
1191 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
1192 		struct vfio_irq_info info;
1193 
1194 		minsz = offsetofend(struct vfio_irq_info, count);
1195 
1196 		if (copy_from_user(&info, (void __user *)arg, minsz))
1197 			return -EFAULT;
1198 
1199 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
1200 			return -EINVAL;
1201 
1202 		switch (info.index) {
1203 		case VFIO_PCI_INTX_IRQ_INDEX:
1204 		case VFIO_PCI_MSI_IRQ_INDEX:
1205 			break;
1206 		default:
1207 			return -EINVAL;
1208 		}
1209 
1210 		info.flags = VFIO_IRQ_INFO_EVENTFD;
1211 
1212 		info.count = intel_vgpu_get_irq_count(vgpu, info.index);
1213 
1214 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1215 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
1216 				       VFIO_IRQ_INFO_AUTOMASKED);
1217 		else
1218 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
1219 
1220 		return copy_to_user((void __user *)arg, &info, minsz) ?
1221 			-EFAULT : 0;
1222 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
1223 		struct vfio_irq_set hdr;
1224 		u8 *data = NULL;
1225 		int ret = 0;
1226 		size_t data_size = 0;
1227 
1228 		minsz = offsetofend(struct vfio_irq_set, count);
1229 
1230 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1231 			return -EFAULT;
1232 
1233 		if (!(hdr.flags & VFIO_IRQ_SET_DATA_NONE)) {
1234 			int max = intel_vgpu_get_irq_count(vgpu, hdr.index);
1235 
1236 			ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
1237 						VFIO_PCI_NUM_IRQS, &data_size);
1238 			if (ret) {
1239 				gvt_vgpu_err("intel:vfio_set_irqs_validate_and_prepare failed\n");
1240 				return -EINVAL;
1241 			}
1242 			if (data_size) {
1243 				data = memdup_user((void __user *)(arg + minsz),
1244 						   data_size);
1245 				if (IS_ERR(data))
1246 					return PTR_ERR(data);
1247 			}
1248 		}
1249 
1250 		ret = intel_vgpu_set_irqs(vgpu, hdr.flags, hdr.index,
1251 					hdr.start, hdr.count, data);
1252 		kfree(data);
1253 
1254 		return ret;
1255 	} else if (cmd == VFIO_DEVICE_RESET) {
1256 		intel_gvt_ops->vgpu_reset(vgpu);
1257 		return 0;
1258 	} else if (cmd == VFIO_DEVICE_QUERY_GFX_PLANE) {
1259 		struct vfio_device_gfx_plane_info dmabuf;
1260 		int ret = 0;
1261 
1262 		minsz = offsetofend(struct vfio_device_gfx_plane_info,
1263 				    dmabuf_id);
1264 		if (copy_from_user(&dmabuf, (void __user *)arg, minsz))
1265 			return -EFAULT;
1266 		if (dmabuf.argsz < minsz)
1267 			return -EINVAL;
1268 
1269 		ret = intel_gvt_ops->vgpu_query_plane(vgpu, &dmabuf);
1270 		if (ret != 0)
1271 			return ret;
1272 
1273 		return copy_to_user((void __user *)arg, &dmabuf, minsz) ?
1274 								-EFAULT : 0;
1275 	} else if (cmd == VFIO_DEVICE_GET_GFX_DMABUF) {
1276 		__u32 dmabuf_id;
1277 		__s32 dmabuf_fd;
1278 
1279 		if (get_user(dmabuf_id, (__u32 __user *)arg))
1280 			return -EFAULT;
1281 
1282 		dmabuf_fd = intel_gvt_ops->vgpu_get_dmabuf(vgpu, dmabuf_id);
1283 		return dmabuf_fd;
1284 
1285 	}
1286 
1287 	return 0;
1288 }
1289 
1290 static ssize_t
1291 vgpu_id_show(struct device *dev, struct device_attribute *attr,
1292 	     char *buf)
1293 {
1294 	struct mdev_device *mdev = mdev_from_dev(dev);
1295 
1296 	if (mdev) {
1297 		struct intel_vgpu *vgpu = (struct intel_vgpu *)
1298 			mdev_get_drvdata(mdev);
1299 		return sprintf(buf, "%d\n", vgpu->id);
1300 	}
1301 	return sprintf(buf, "\n");
1302 }
1303 
1304 static ssize_t
1305 hw_id_show(struct device *dev, struct device_attribute *attr,
1306 	   char *buf)
1307 {
1308 	struct mdev_device *mdev = mdev_from_dev(dev);
1309 
1310 	if (mdev) {
1311 		struct intel_vgpu *vgpu = (struct intel_vgpu *)
1312 			mdev_get_drvdata(mdev);
1313 		return sprintf(buf, "%u\n",
1314 			       vgpu->submission.shadow_ctx->hw_id);
1315 	}
1316 	return sprintf(buf, "\n");
1317 }
1318 
1319 static DEVICE_ATTR_RO(vgpu_id);
1320 static DEVICE_ATTR_RO(hw_id);
1321 
1322 static struct attribute *intel_vgpu_attrs[] = {
1323 	&dev_attr_vgpu_id.attr,
1324 	&dev_attr_hw_id.attr,
1325 	NULL
1326 };
1327 
1328 static const struct attribute_group intel_vgpu_group = {
1329 	.name = "intel_vgpu",
1330 	.attrs = intel_vgpu_attrs,
1331 };
1332 
1333 static const struct attribute_group *intel_vgpu_groups[] = {
1334 	&intel_vgpu_group,
1335 	NULL,
1336 };
1337 
1338 static struct mdev_parent_ops intel_vgpu_ops = {
1339 	.mdev_attr_groups       = intel_vgpu_groups,
1340 	.create			= intel_vgpu_create,
1341 	.remove			= intel_vgpu_remove,
1342 
1343 	.open			= intel_vgpu_open,
1344 	.release		= intel_vgpu_release,
1345 
1346 	.read			= intel_vgpu_read,
1347 	.write			= intel_vgpu_write,
1348 	.mmap			= intel_vgpu_mmap,
1349 	.ioctl			= intel_vgpu_ioctl,
1350 };
1351 
1352 static int kvmgt_host_init(struct device *dev, void *gvt, const void *ops)
1353 {
1354 	struct attribute **kvm_type_attrs;
1355 	struct attribute_group **kvm_vgpu_type_groups;
1356 
1357 	intel_gvt_ops = ops;
1358 	if (!intel_gvt_ops->get_gvt_attrs(&kvm_type_attrs,
1359 			&kvm_vgpu_type_groups))
1360 		return -EFAULT;
1361 	intel_vgpu_ops.supported_type_groups = kvm_vgpu_type_groups;
1362 
1363 	return mdev_register_device(dev, &intel_vgpu_ops);
1364 }
1365 
1366 static void kvmgt_host_exit(struct device *dev, void *gvt)
1367 {
1368 	mdev_unregister_device(dev);
1369 }
1370 
1371 static int kvmgt_write_protect_add(unsigned long handle, u64 gfn)
1372 {
1373 	struct kvmgt_guest_info *info;
1374 	struct kvm *kvm;
1375 	struct kvm_memory_slot *slot;
1376 	int idx;
1377 
1378 	if (!handle_valid(handle))
1379 		return -ESRCH;
1380 
1381 	info = (struct kvmgt_guest_info *)handle;
1382 	kvm = info->kvm;
1383 
1384 	idx = srcu_read_lock(&kvm->srcu);
1385 	slot = gfn_to_memslot(kvm, gfn);
1386 	if (!slot) {
1387 		srcu_read_unlock(&kvm->srcu, idx);
1388 		return -EINVAL;
1389 	}
1390 
1391 	spin_lock(&kvm->mmu_lock);
1392 
1393 	if (kvmgt_gfn_is_write_protected(info, gfn))
1394 		goto out;
1395 
1396 	kvm_slot_page_track_add_page(kvm, slot, gfn, KVM_PAGE_TRACK_WRITE);
1397 	kvmgt_protect_table_add(info, gfn);
1398 
1399 out:
1400 	spin_unlock(&kvm->mmu_lock);
1401 	srcu_read_unlock(&kvm->srcu, idx);
1402 	return 0;
1403 }
1404 
1405 static int kvmgt_write_protect_remove(unsigned long handle, u64 gfn)
1406 {
1407 	struct kvmgt_guest_info *info;
1408 	struct kvm *kvm;
1409 	struct kvm_memory_slot *slot;
1410 	int idx;
1411 
1412 	if (!handle_valid(handle))
1413 		return 0;
1414 
1415 	info = (struct kvmgt_guest_info *)handle;
1416 	kvm = info->kvm;
1417 
1418 	idx = srcu_read_lock(&kvm->srcu);
1419 	slot = gfn_to_memslot(kvm, gfn);
1420 	if (!slot) {
1421 		srcu_read_unlock(&kvm->srcu, idx);
1422 		return -EINVAL;
1423 	}
1424 
1425 	spin_lock(&kvm->mmu_lock);
1426 
1427 	if (!kvmgt_gfn_is_write_protected(info, gfn))
1428 		goto out;
1429 
1430 	kvm_slot_page_track_remove_page(kvm, slot, gfn, KVM_PAGE_TRACK_WRITE);
1431 	kvmgt_protect_table_del(info, gfn);
1432 
1433 out:
1434 	spin_unlock(&kvm->mmu_lock);
1435 	srcu_read_unlock(&kvm->srcu, idx);
1436 	return 0;
1437 }
1438 
1439 static void kvmgt_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1440 		const u8 *val, int len,
1441 		struct kvm_page_track_notifier_node *node)
1442 {
1443 	struct kvmgt_guest_info *info = container_of(node,
1444 					struct kvmgt_guest_info, track_node);
1445 
1446 	if (kvmgt_gfn_is_write_protected(info, gpa_to_gfn(gpa)))
1447 		intel_gvt_ops->write_protect_handler(info->vgpu, gpa,
1448 						     (void *)val, len);
1449 }
1450 
1451 static void kvmgt_page_track_flush_slot(struct kvm *kvm,
1452 		struct kvm_memory_slot *slot,
1453 		struct kvm_page_track_notifier_node *node)
1454 {
1455 	int i;
1456 	gfn_t gfn;
1457 	struct kvmgt_guest_info *info = container_of(node,
1458 					struct kvmgt_guest_info, track_node);
1459 
1460 	spin_lock(&kvm->mmu_lock);
1461 	for (i = 0; i < slot->npages; i++) {
1462 		gfn = slot->base_gfn + i;
1463 		if (kvmgt_gfn_is_write_protected(info, gfn)) {
1464 			kvm_slot_page_track_remove_page(kvm, slot, gfn,
1465 						KVM_PAGE_TRACK_WRITE);
1466 			kvmgt_protect_table_del(info, gfn);
1467 		}
1468 	}
1469 	spin_unlock(&kvm->mmu_lock);
1470 }
1471 
1472 static bool __kvmgt_vgpu_exist(struct intel_vgpu *vgpu, struct kvm *kvm)
1473 {
1474 	struct intel_vgpu *itr;
1475 	struct kvmgt_guest_info *info;
1476 	int id;
1477 	bool ret = false;
1478 
1479 	mutex_lock(&vgpu->gvt->lock);
1480 	for_each_active_vgpu(vgpu->gvt, itr, id) {
1481 		if (!handle_valid(itr->handle))
1482 			continue;
1483 
1484 		info = (struct kvmgt_guest_info *)itr->handle;
1485 		if (kvm && kvm == info->kvm) {
1486 			ret = true;
1487 			goto out;
1488 		}
1489 	}
1490 out:
1491 	mutex_unlock(&vgpu->gvt->lock);
1492 	return ret;
1493 }
1494 
1495 static int kvmgt_guest_init(struct mdev_device *mdev)
1496 {
1497 	struct kvmgt_guest_info *info;
1498 	struct intel_vgpu *vgpu;
1499 	struct kvm *kvm;
1500 
1501 	vgpu = mdev_get_drvdata(mdev);
1502 	if (handle_valid(vgpu->handle))
1503 		return -EEXIST;
1504 
1505 	kvm = vgpu->vdev.kvm;
1506 	if (!kvm || kvm->mm != current->mm) {
1507 		gvt_vgpu_err("KVM is required to use Intel vGPU\n");
1508 		return -ESRCH;
1509 	}
1510 
1511 	if (__kvmgt_vgpu_exist(vgpu, kvm))
1512 		return -EEXIST;
1513 
1514 	info = vzalloc(sizeof(struct kvmgt_guest_info));
1515 	if (!info)
1516 		return -ENOMEM;
1517 
1518 	vgpu->handle = (unsigned long)info;
1519 	info->vgpu = vgpu;
1520 	info->kvm = kvm;
1521 	kvm_get_kvm(info->kvm);
1522 
1523 	kvmgt_protect_table_init(info);
1524 	gvt_cache_init(vgpu);
1525 
1526 	mutex_init(&vgpu->dmabuf_lock);
1527 	init_completion(&vgpu->vblank_done);
1528 
1529 	info->track_node.track_write = kvmgt_page_track_write;
1530 	info->track_node.track_flush_slot = kvmgt_page_track_flush_slot;
1531 	kvm_page_track_register_notifier(kvm, &info->track_node);
1532 
1533 	return 0;
1534 }
1535 
1536 static bool kvmgt_guest_exit(struct kvmgt_guest_info *info)
1537 {
1538 	kvm_page_track_unregister_notifier(info->kvm, &info->track_node);
1539 	kvm_put_kvm(info->kvm);
1540 	kvmgt_protect_table_destroy(info);
1541 	gvt_cache_destroy(info->vgpu);
1542 	vfree(info);
1543 
1544 	return true;
1545 }
1546 
1547 static int kvmgt_attach_vgpu(void *vgpu, unsigned long *handle)
1548 {
1549 	/* nothing to do here */
1550 	return 0;
1551 }
1552 
1553 static void kvmgt_detach_vgpu(unsigned long handle)
1554 {
1555 	/* nothing to do here */
1556 }
1557 
1558 static int kvmgt_inject_msi(unsigned long handle, u32 addr, u16 data)
1559 {
1560 	struct kvmgt_guest_info *info;
1561 	struct intel_vgpu *vgpu;
1562 
1563 	if (!handle_valid(handle))
1564 		return -ESRCH;
1565 
1566 	info = (struct kvmgt_guest_info *)handle;
1567 	vgpu = info->vgpu;
1568 
1569 	if (eventfd_signal(vgpu->vdev.msi_trigger, 1) == 1)
1570 		return 0;
1571 
1572 	return -EFAULT;
1573 }
1574 
1575 static unsigned long kvmgt_gfn_to_pfn(unsigned long handle, unsigned long gfn)
1576 {
1577 	unsigned long iova, pfn;
1578 	struct kvmgt_guest_info *info;
1579 	struct device *dev;
1580 	struct intel_vgpu *vgpu;
1581 	int rc;
1582 
1583 	if (!handle_valid(handle))
1584 		return INTEL_GVT_INVALID_ADDR;
1585 
1586 	info = (struct kvmgt_guest_info *)handle;
1587 	vgpu = info->vgpu;
1588 	iova = gvt_cache_find(info->vgpu, gfn);
1589 	if (iova != INTEL_GVT_INVALID_ADDR)
1590 		return iova;
1591 
1592 	pfn = INTEL_GVT_INVALID_ADDR;
1593 	dev = mdev_dev(info->vgpu->vdev.mdev);
1594 	rc = vfio_pin_pages(dev, &gfn, 1, IOMMU_READ | IOMMU_WRITE, &pfn);
1595 	if (rc != 1) {
1596 		gvt_vgpu_err("vfio_pin_pages failed for gfn 0x%lx: %d\n",
1597 			gfn, rc);
1598 		return INTEL_GVT_INVALID_ADDR;
1599 	}
1600 	/* transfer to host iova for GFX to use DMA */
1601 	rc = gvt_dma_map_iova(info->vgpu, pfn, &iova);
1602 	if (rc) {
1603 		gvt_vgpu_err("gvt_dma_map_iova failed for gfn: 0x%lx\n", gfn);
1604 		vfio_unpin_pages(dev, &gfn, 1);
1605 		return INTEL_GVT_INVALID_ADDR;
1606 	}
1607 
1608 	gvt_cache_add(info->vgpu, gfn, iova);
1609 	return iova;
1610 }
1611 
1612 static int kvmgt_rw_gpa(unsigned long handle, unsigned long gpa,
1613 			void *buf, unsigned long len, bool write)
1614 {
1615 	struct kvmgt_guest_info *info;
1616 	struct kvm *kvm;
1617 	int idx, ret;
1618 	bool kthread = current->mm == NULL;
1619 
1620 	if (!handle_valid(handle))
1621 		return -ESRCH;
1622 
1623 	info = (struct kvmgt_guest_info *)handle;
1624 	kvm = info->kvm;
1625 
1626 	if (kthread)
1627 		use_mm(kvm->mm);
1628 
1629 	idx = srcu_read_lock(&kvm->srcu);
1630 	ret = write ? kvm_write_guest(kvm, gpa, buf, len) :
1631 		      kvm_read_guest(kvm, gpa, buf, len);
1632 	srcu_read_unlock(&kvm->srcu, idx);
1633 
1634 	if (kthread)
1635 		unuse_mm(kvm->mm);
1636 
1637 	return ret;
1638 }
1639 
1640 static int kvmgt_read_gpa(unsigned long handle, unsigned long gpa,
1641 			void *buf, unsigned long len)
1642 {
1643 	return kvmgt_rw_gpa(handle, gpa, buf, len, false);
1644 }
1645 
1646 static int kvmgt_write_gpa(unsigned long handle, unsigned long gpa,
1647 			void *buf, unsigned long len)
1648 {
1649 	return kvmgt_rw_gpa(handle, gpa, buf, len, true);
1650 }
1651 
1652 static unsigned long kvmgt_virt_to_pfn(void *addr)
1653 {
1654 	return PFN_DOWN(__pa(addr));
1655 }
1656 
1657 static bool kvmgt_is_valid_gfn(unsigned long handle, unsigned long gfn)
1658 {
1659 	struct kvmgt_guest_info *info;
1660 	struct kvm *kvm;
1661 
1662 	if (!handle_valid(handle))
1663 		return false;
1664 
1665 	info = (struct kvmgt_guest_info *)handle;
1666 	kvm = info->kvm;
1667 
1668 	return kvm_is_visible_gfn(kvm, gfn);
1669 
1670 }
1671 
1672 struct intel_gvt_mpt kvmgt_mpt = {
1673 	.host_init = kvmgt_host_init,
1674 	.host_exit = kvmgt_host_exit,
1675 	.attach_vgpu = kvmgt_attach_vgpu,
1676 	.detach_vgpu = kvmgt_detach_vgpu,
1677 	.inject_msi = kvmgt_inject_msi,
1678 	.from_virt_to_mfn = kvmgt_virt_to_pfn,
1679 	.set_wp_page = kvmgt_write_protect_add,
1680 	.unset_wp_page = kvmgt_write_protect_remove,
1681 	.read_gpa = kvmgt_read_gpa,
1682 	.write_gpa = kvmgt_write_gpa,
1683 	.gfn_to_mfn = kvmgt_gfn_to_pfn,
1684 	.set_opregion = kvmgt_set_opregion,
1685 	.get_vfio_device = kvmgt_get_vfio_device,
1686 	.put_vfio_device = kvmgt_put_vfio_device,
1687 	.is_valid_gfn = kvmgt_is_valid_gfn,
1688 };
1689 EXPORT_SYMBOL_GPL(kvmgt_mpt);
1690 
1691 static int __init kvmgt_init(void)
1692 {
1693 	return 0;
1694 }
1695 
1696 static void __exit kvmgt_exit(void)
1697 {
1698 }
1699 
1700 module_init(kvmgt_init);
1701 module_exit(kvmgt_exit);
1702 
1703 MODULE_LICENSE("GPL and additional rights");
1704 MODULE_AUTHOR("Intel Corporation");
1705