xref: /openbmc/linux/drivers/gpu/drm/i915/gvt/kvmgt.c (revision 9b2433a9)
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/sched/mm.h>
36 #include <linux/types.h>
37 #include <linux/list.h>
38 #include <linux/rbtree.h>
39 #include <linux/spinlock.h>
40 #include <linux/eventfd.h>
41 #include <linux/uuid.h>
42 #include <linux/kvm_host.h>
43 #include <linux/vfio.h>
44 #include <linux/mdev.h>
45 #include <linux/debugfs.h>
46 
47 #include <linux/nospec.h>
48 
49 #include "i915_drv.h"
50 #include "gvt.h"
51 
52 static const struct intel_gvt_ops *intel_gvt_ops;
53 
54 /* helper macros copied from vfio-pci */
55 #define VFIO_PCI_OFFSET_SHIFT   40
56 #define VFIO_PCI_OFFSET_TO_INDEX(off)   (off >> VFIO_PCI_OFFSET_SHIFT)
57 #define VFIO_PCI_INDEX_TO_OFFSET(index) ((u64)(index) << VFIO_PCI_OFFSET_SHIFT)
58 #define VFIO_PCI_OFFSET_MASK    (((u64)(1) << VFIO_PCI_OFFSET_SHIFT) - 1)
59 
60 #define EDID_BLOB_OFFSET (PAGE_SIZE/2)
61 
62 #define OPREGION_SIGNATURE "IntelGraphicsMem"
63 
64 struct vfio_region;
65 struct intel_vgpu_regops {
66 	size_t (*rw)(struct intel_vgpu *vgpu, char *buf,
67 			size_t count, loff_t *ppos, bool iswrite);
68 	void (*release)(struct intel_vgpu *vgpu,
69 			struct vfio_region *region);
70 };
71 
72 struct vfio_region {
73 	u32				type;
74 	u32				subtype;
75 	size_t				size;
76 	u32				flags;
77 	const struct intel_vgpu_regops	*ops;
78 	void				*data;
79 };
80 
81 struct vfio_edid_region {
82 	struct vfio_region_gfx_edid vfio_edid_regs;
83 	void *edid_blob;
84 };
85 
86 struct kvmgt_pgfn {
87 	gfn_t gfn;
88 	struct hlist_node hnode;
89 };
90 
91 struct kvmgt_guest_info {
92 	struct kvm *kvm;
93 	struct intel_vgpu *vgpu;
94 	struct kvm_page_track_notifier_node track_node;
95 #define NR_BKT (1 << 18)
96 	struct hlist_head ptable[NR_BKT];
97 #undef NR_BKT
98 	struct dentry *debugfs_cache_entries;
99 };
100 
101 struct gvt_dma {
102 	struct intel_vgpu *vgpu;
103 	struct rb_node gfn_node;
104 	struct rb_node dma_addr_node;
105 	gfn_t gfn;
106 	dma_addr_t dma_addr;
107 	unsigned long size;
108 	struct kref ref;
109 };
110 
111 static inline bool handle_valid(unsigned long handle)
112 {
113 	return !!(handle & ~0xff);
114 }
115 
116 static int kvmgt_guest_init(struct mdev_device *mdev);
117 static void intel_vgpu_release_work(struct work_struct *work);
118 static bool kvmgt_guest_exit(struct kvmgt_guest_info *info);
119 
120 static void gvt_unpin_guest_page(struct intel_vgpu *vgpu, unsigned long gfn,
121 		unsigned long size)
122 {
123 	int total_pages;
124 	int npage;
125 	int ret;
126 
127 	total_pages = roundup(size, PAGE_SIZE) / PAGE_SIZE;
128 
129 	for (npage = 0; npage < total_pages; npage++) {
130 		unsigned long cur_gfn = gfn + npage;
131 
132 		ret = vfio_unpin_pages(mdev_dev(vgpu->vdev.mdev), &cur_gfn, 1);
133 		WARN_ON(ret != 1);
134 	}
135 }
136 
137 /* Pin a normal or compound guest page for dma. */
138 static int gvt_pin_guest_page(struct intel_vgpu *vgpu, unsigned long gfn,
139 		unsigned long size, struct page **page)
140 {
141 	unsigned long base_pfn = 0;
142 	int total_pages;
143 	int npage;
144 	int ret;
145 
146 	total_pages = roundup(size, PAGE_SIZE) / PAGE_SIZE;
147 	/*
148 	 * We pin the pages one-by-one to avoid allocating a big arrary
149 	 * on stack to hold pfns.
150 	 */
151 	for (npage = 0; npage < total_pages; npage++) {
152 		unsigned long cur_gfn = gfn + npage;
153 		unsigned long pfn;
154 
155 		ret = vfio_pin_pages(mdev_dev(vgpu->vdev.mdev), &cur_gfn, 1,
156 				     IOMMU_READ | IOMMU_WRITE, &pfn);
157 		if (ret != 1) {
158 			gvt_vgpu_err("vfio_pin_pages failed for gfn 0x%lx, ret %d\n",
159 				     cur_gfn, ret);
160 			goto err;
161 		}
162 
163 		if (!pfn_valid(pfn)) {
164 			gvt_vgpu_err("pfn 0x%lx is not mem backed\n", pfn);
165 			npage++;
166 			ret = -EFAULT;
167 			goto err;
168 		}
169 
170 		if (npage == 0)
171 			base_pfn = pfn;
172 		else if (base_pfn + npage != pfn) {
173 			gvt_vgpu_err("The pages are not continuous\n");
174 			ret = -EINVAL;
175 			npage++;
176 			goto err;
177 		}
178 	}
179 
180 	*page = pfn_to_page(base_pfn);
181 	return 0;
182 err:
183 	gvt_unpin_guest_page(vgpu, gfn, npage * PAGE_SIZE);
184 	return ret;
185 }
186 
187 static int gvt_dma_map_page(struct intel_vgpu *vgpu, unsigned long gfn,
188 		dma_addr_t *dma_addr, unsigned long size)
189 {
190 	struct device *dev = &vgpu->gvt->dev_priv->drm.pdev->dev;
191 	struct page *page = NULL;
192 	int ret;
193 
194 	ret = gvt_pin_guest_page(vgpu, gfn, size, &page);
195 	if (ret)
196 		return ret;
197 
198 	/* Setup DMA mapping. */
199 	*dma_addr = dma_map_page(dev, page, 0, size, PCI_DMA_BIDIRECTIONAL);
200 	if (dma_mapping_error(dev, *dma_addr)) {
201 		gvt_vgpu_err("DMA mapping failed for pfn 0x%lx, ret %d\n",
202 			     page_to_pfn(page), ret);
203 		gvt_unpin_guest_page(vgpu, gfn, size);
204 		return -ENOMEM;
205 	}
206 
207 	return 0;
208 }
209 
210 static void gvt_dma_unmap_page(struct intel_vgpu *vgpu, unsigned long gfn,
211 		dma_addr_t dma_addr, unsigned long size)
212 {
213 	struct device *dev = &vgpu->gvt->dev_priv->drm.pdev->dev;
214 
215 	dma_unmap_page(dev, dma_addr, size, PCI_DMA_BIDIRECTIONAL);
216 	gvt_unpin_guest_page(vgpu, gfn, size);
217 }
218 
219 static struct gvt_dma *__gvt_cache_find_dma_addr(struct intel_vgpu *vgpu,
220 		dma_addr_t dma_addr)
221 {
222 	struct rb_node *node = vgpu->vdev.dma_addr_cache.rb_node;
223 	struct gvt_dma *itr;
224 
225 	while (node) {
226 		itr = rb_entry(node, struct gvt_dma, dma_addr_node);
227 
228 		if (dma_addr < itr->dma_addr)
229 			node = node->rb_left;
230 		else if (dma_addr > itr->dma_addr)
231 			node = node->rb_right;
232 		else
233 			return itr;
234 	}
235 	return NULL;
236 }
237 
238 static struct gvt_dma *__gvt_cache_find_gfn(struct intel_vgpu *vgpu, gfn_t gfn)
239 {
240 	struct rb_node *node = vgpu->vdev.gfn_cache.rb_node;
241 	struct gvt_dma *itr;
242 
243 	while (node) {
244 		itr = rb_entry(node, struct gvt_dma, gfn_node);
245 
246 		if (gfn < itr->gfn)
247 			node = node->rb_left;
248 		else if (gfn > itr->gfn)
249 			node = node->rb_right;
250 		else
251 			return itr;
252 	}
253 	return NULL;
254 }
255 
256 static int __gvt_cache_add(struct intel_vgpu *vgpu, gfn_t gfn,
257 		dma_addr_t dma_addr, unsigned long size)
258 {
259 	struct gvt_dma *new, *itr;
260 	struct rb_node **link, *parent = NULL;
261 
262 	new = kzalloc(sizeof(struct gvt_dma), GFP_KERNEL);
263 	if (!new)
264 		return -ENOMEM;
265 
266 	new->vgpu = vgpu;
267 	new->gfn = gfn;
268 	new->dma_addr = dma_addr;
269 	new->size = size;
270 	kref_init(&new->ref);
271 
272 	/* gfn_cache maps gfn to struct gvt_dma. */
273 	link = &vgpu->vdev.gfn_cache.rb_node;
274 	while (*link) {
275 		parent = *link;
276 		itr = rb_entry(parent, struct gvt_dma, gfn_node);
277 
278 		if (gfn < itr->gfn)
279 			link = &parent->rb_left;
280 		else
281 			link = &parent->rb_right;
282 	}
283 	rb_link_node(&new->gfn_node, parent, link);
284 	rb_insert_color(&new->gfn_node, &vgpu->vdev.gfn_cache);
285 
286 	/* dma_addr_cache maps dma addr to struct gvt_dma. */
287 	parent = NULL;
288 	link = &vgpu->vdev.dma_addr_cache.rb_node;
289 	while (*link) {
290 		parent = *link;
291 		itr = rb_entry(parent, struct gvt_dma, dma_addr_node);
292 
293 		if (dma_addr < itr->dma_addr)
294 			link = &parent->rb_left;
295 		else
296 			link = &parent->rb_right;
297 	}
298 	rb_link_node(&new->dma_addr_node, parent, link);
299 	rb_insert_color(&new->dma_addr_node, &vgpu->vdev.dma_addr_cache);
300 
301 	vgpu->vdev.nr_cache_entries++;
302 	return 0;
303 }
304 
305 static void __gvt_cache_remove_entry(struct intel_vgpu *vgpu,
306 				struct gvt_dma *entry)
307 {
308 	rb_erase(&entry->gfn_node, &vgpu->vdev.gfn_cache);
309 	rb_erase(&entry->dma_addr_node, &vgpu->vdev.dma_addr_cache);
310 	kfree(entry);
311 	vgpu->vdev.nr_cache_entries--;
312 }
313 
314 static void gvt_cache_destroy(struct intel_vgpu *vgpu)
315 {
316 	struct gvt_dma *dma;
317 	struct rb_node *node = NULL;
318 
319 	for (;;) {
320 		mutex_lock(&vgpu->vdev.cache_lock);
321 		node = rb_first(&vgpu->vdev.gfn_cache);
322 		if (!node) {
323 			mutex_unlock(&vgpu->vdev.cache_lock);
324 			break;
325 		}
326 		dma = rb_entry(node, struct gvt_dma, gfn_node);
327 		gvt_dma_unmap_page(vgpu, dma->gfn, dma->dma_addr, dma->size);
328 		__gvt_cache_remove_entry(vgpu, dma);
329 		mutex_unlock(&vgpu->vdev.cache_lock);
330 	}
331 }
332 
333 static void gvt_cache_init(struct intel_vgpu *vgpu)
334 {
335 	vgpu->vdev.gfn_cache = RB_ROOT;
336 	vgpu->vdev.dma_addr_cache = RB_ROOT;
337 	vgpu->vdev.nr_cache_entries = 0;
338 	mutex_init(&vgpu->vdev.cache_lock);
339 }
340 
341 static void kvmgt_protect_table_init(struct kvmgt_guest_info *info)
342 {
343 	hash_init(info->ptable);
344 }
345 
346 static void kvmgt_protect_table_destroy(struct kvmgt_guest_info *info)
347 {
348 	struct kvmgt_pgfn *p;
349 	struct hlist_node *tmp;
350 	int i;
351 
352 	hash_for_each_safe(info->ptable, i, tmp, p, hnode) {
353 		hash_del(&p->hnode);
354 		kfree(p);
355 	}
356 }
357 
358 static struct kvmgt_pgfn *
359 __kvmgt_protect_table_find(struct kvmgt_guest_info *info, gfn_t gfn)
360 {
361 	struct kvmgt_pgfn *p, *res = NULL;
362 
363 	hash_for_each_possible(info->ptable, p, hnode, gfn) {
364 		if (gfn == p->gfn) {
365 			res = p;
366 			break;
367 		}
368 	}
369 
370 	return res;
371 }
372 
373 static bool kvmgt_gfn_is_write_protected(struct kvmgt_guest_info *info,
374 				gfn_t gfn)
375 {
376 	struct kvmgt_pgfn *p;
377 
378 	p = __kvmgt_protect_table_find(info, gfn);
379 	return !!p;
380 }
381 
382 static void kvmgt_protect_table_add(struct kvmgt_guest_info *info, gfn_t gfn)
383 {
384 	struct kvmgt_pgfn *p;
385 
386 	if (kvmgt_gfn_is_write_protected(info, gfn))
387 		return;
388 
389 	p = kzalloc(sizeof(struct kvmgt_pgfn), GFP_ATOMIC);
390 	if (WARN(!p, "gfn: 0x%llx\n", gfn))
391 		return;
392 
393 	p->gfn = gfn;
394 	hash_add(info->ptable, &p->hnode, gfn);
395 }
396 
397 static void kvmgt_protect_table_del(struct kvmgt_guest_info *info,
398 				gfn_t gfn)
399 {
400 	struct kvmgt_pgfn *p;
401 
402 	p = __kvmgt_protect_table_find(info, gfn);
403 	if (p) {
404 		hash_del(&p->hnode);
405 		kfree(p);
406 	}
407 }
408 
409 static size_t intel_vgpu_reg_rw_opregion(struct intel_vgpu *vgpu, char *buf,
410 		size_t count, loff_t *ppos, bool iswrite)
411 {
412 	unsigned int i = VFIO_PCI_OFFSET_TO_INDEX(*ppos) -
413 			VFIO_PCI_NUM_REGIONS;
414 	void *base = vgpu->vdev.region[i].data;
415 	loff_t pos = *ppos & VFIO_PCI_OFFSET_MASK;
416 
417 	if (pos >= vgpu->vdev.region[i].size || iswrite) {
418 		gvt_vgpu_err("invalid op or offset for Intel vgpu OpRegion\n");
419 		return -EINVAL;
420 	}
421 	count = min(count, (size_t)(vgpu->vdev.region[i].size - pos));
422 	memcpy(buf, base + pos, count);
423 
424 	return count;
425 }
426 
427 static void intel_vgpu_reg_release_opregion(struct intel_vgpu *vgpu,
428 		struct vfio_region *region)
429 {
430 }
431 
432 static const struct intel_vgpu_regops intel_vgpu_regops_opregion = {
433 	.rw = intel_vgpu_reg_rw_opregion,
434 	.release = intel_vgpu_reg_release_opregion,
435 };
436 
437 static int handle_edid_regs(struct intel_vgpu *vgpu,
438 			struct vfio_edid_region *region, char *buf,
439 			size_t count, u16 offset, bool is_write)
440 {
441 	struct vfio_region_gfx_edid *regs = &region->vfio_edid_regs;
442 	unsigned int data;
443 
444 	if (offset + count > sizeof(*regs))
445 		return -EINVAL;
446 
447 	if (count != 4)
448 		return -EINVAL;
449 
450 	if (is_write) {
451 		data = *((unsigned int *)buf);
452 		switch (offset) {
453 		case offsetof(struct vfio_region_gfx_edid, link_state):
454 			if (data == VFIO_DEVICE_GFX_LINK_STATE_UP) {
455 				if (!drm_edid_block_valid(
456 					(u8 *)region->edid_blob,
457 					0,
458 					true,
459 					NULL)) {
460 					gvt_vgpu_err("invalid EDID blob\n");
461 					return -EINVAL;
462 				}
463 				intel_gvt_ops->emulate_hotplug(vgpu, true);
464 			} else if (data == VFIO_DEVICE_GFX_LINK_STATE_DOWN)
465 				intel_gvt_ops->emulate_hotplug(vgpu, false);
466 			else {
467 				gvt_vgpu_err("invalid EDID link state %d\n",
468 					regs->link_state);
469 				return -EINVAL;
470 			}
471 			regs->link_state = data;
472 			break;
473 		case offsetof(struct vfio_region_gfx_edid, edid_size):
474 			if (data > regs->edid_max_size) {
475 				gvt_vgpu_err("EDID size is bigger than %d!\n",
476 					regs->edid_max_size);
477 				return -EINVAL;
478 			}
479 			regs->edid_size = data;
480 			break;
481 		default:
482 			/* read-only regs */
483 			gvt_vgpu_err("write read-only EDID region at offset %d\n",
484 				offset);
485 			return -EPERM;
486 		}
487 	} else {
488 		memcpy(buf, (char *)regs + offset, count);
489 	}
490 
491 	return count;
492 }
493 
494 static int handle_edid_blob(struct vfio_edid_region *region, char *buf,
495 			size_t count, u16 offset, bool is_write)
496 {
497 	if (offset + count > region->vfio_edid_regs.edid_size)
498 		return -EINVAL;
499 
500 	if (is_write)
501 		memcpy(region->edid_blob + offset, buf, count);
502 	else
503 		memcpy(buf, region->edid_blob + offset, count);
504 
505 	return count;
506 }
507 
508 static size_t intel_vgpu_reg_rw_edid(struct intel_vgpu *vgpu, char *buf,
509 		size_t count, loff_t *ppos, bool iswrite)
510 {
511 	int ret;
512 	unsigned int i = VFIO_PCI_OFFSET_TO_INDEX(*ppos) -
513 			VFIO_PCI_NUM_REGIONS;
514 	struct vfio_edid_region *region =
515 		(struct vfio_edid_region *)vgpu->vdev.region[i].data;
516 	loff_t pos = *ppos & VFIO_PCI_OFFSET_MASK;
517 
518 	if (pos < region->vfio_edid_regs.edid_offset) {
519 		ret = handle_edid_regs(vgpu, region, buf, count, pos, iswrite);
520 	} else {
521 		pos -= EDID_BLOB_OFFSET;
522 		ret = handle_edid_blob(region, buf, count, pos, iswrite);
523 	}
524 
525 	if (ret < 0)
526 		gvt_vgpu_err("failed to access EDID region\n");
527 
528 	return ret;
529 }
530 
531 static void intel_vgpu_reg_release_edid(struct intel_vgpu *vgpu,
532 					struct vfio_region *region)
533 {
534 	kfree(region->data);
535 }
536 
537 static const struct intel_vgpu_regops intel_vgpu_regops_edid = {
538 	.rw = intel_vgpu_reg_rw_edid,
539 	.release = intel_vgpu_reg_release_edid,
540 };
541 
542 static int intel_vgpu_register_reg(struct intel_vgpu *vgpu,
543 		unsigned int type, unsigned int subtype,
544 		const struct intel_vgpu_regops *ops,
545 		size_t size, u32 flags, void *data)
546 {
547 	struct vfio_region *region;
548 
549 	region = krealloc(vgpu->vdev.region,
550 			(vgpu->vdev.num_regions + 1) * sizeof(*region),
551 			GFP_KERNEL);
552 	if (!region)
553 		return -ENOMEM;
554 
555 	vgpu->vdev.region = region;
556 	vgpu->vdev.region[vgpu->vdev.num_regions].type = type;
557 	vgpu->vdev.region[vgpu->vdev.num_regions].subtype = subtype;
558 	vgpu->vdev.region[vgpu->vdev.num_regions].ops = ops;
559 	vgpu->vdev.region[vgpu->vdev.num_regions].size = size;
560 	vgpu->vdev.region[vgpu->vdev.num_regions].flags = flags;
561 	vgpu->vdev.region[vgpu->vdev.num_regions].data = data;
562 	vgpu->vdev.num_regions++;
563 	return 0;
564 }
565 
566 static int kvmgt_get_vfio_device(void *p_vgpu)
567 {
568 	struct intel_vgpu *vgpu = (struct intel_vgpu *)p_vgpu;
569 
570 	vgpu->vdev.vfio_device = vfio_device_get_from_dev(
571 		mdev_dev(vgpu->vdev.mdev));
572 	if (!vgpu->vdev.vfio_device) {
573 		gvt_vgpu_err("failed to get vfio device\n");
574 		return -ENODEV;
575 	}
576 	return 0;
577 }
578 
579 
580 static int kvmgt_set_opregion(void *p_vgpu)
581 {
582 	struct intel_vgpu *vgpu = (struct intel_vgpu *)p_vgpu;
583 	void *base;
584 	int ret;
585 
586 	/* Each vgpu has its own opregion, although VFIO would create another
587 	 * one later. This one is used to expose opregion to VFIO. And the
588 	 * other one created by VFIO later, is used by guest actually.
589 	 */
590 	base = vgpu_opregion(vgpu)->va;
591 	if (!base)
592 		return -ENOMEM;
593 
594 	if (memcmp(base, OPREGION_SIGNATURE, 16)) {
595 		memunmap(base);
596 		return -EINVAL;
597 	}
598 
599 	ret = intel_vgpu_register_reg(vgpu,
600 			PCI_VENDOR_ID_INTEL | VFIO_REGION_TYPE_PCI_VENDOR_TYPE,
601 			VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION,
602 			&intel_vgpu_regops_opregion, OPREGION_SIZE,
603 			VFIO_REGION_INFO_FLAG_READ, base);
604 
605 	return ret;
606 }
607 
608 static int kvmgt_set_edid(void *p_vgpu, int port_num)
609 {
610 	struct intel_vgpu *vgpu = (struct intel_vgpu *)p_vgpu;
611 	struct intel_vgpu_port *port = intel_vgpu_port(vgpu, port_num);
612 	struct vfio_edid_region *base;
613 	int ret;
614 
615 	base = kzalloc(sizeof(*base), GFP_KERNEL);
616 	if (!base)
617 		return -ENOMEM;
618 
619 	/* TODO: Add multi-port and EDID extension block support */
620 	base->vfio_edid_regs.edid_offset = EDID_BLOB_OFFSET;
621 	base->vfio_edid_regs.edid_max_size = EDID_SIZE;
622 	base->vfio_edid_regs.edid_size = EDID_SIZE;
623 	base->vfio_edid_regs.max_xres = vgpu_edid_xres(port->id);
624 	base->vfio_edid_regs.max_yres = vgpu_edid_yres(port->id);
625 	base->edid_blob = port->edid->edid_block;
626 
627 	ret = intel_vgpu_register_reg(vgpu,
628 			VFIO_REGION_TYPE_GFX,
629 			VFIO_REGION_SUBTYPE_GFX_EDID,
630 			&intel_vgpu_regops_edid, EDID_SIZE,
631 			VFIO_REGION_INFO_FLAG_READ |
632 			VFIO_REGION_INFO_FLAG_WRITE |
633 			VFIO_REGION_INFO_FLAG_CAPS, base);
634 
635 	return ret;
636 }
637 
638 static void kvmgt_put_vfio_device(void *vgpu)
639 {
640 	if (WARN_ON(!((struct intel_vgpu *)vgpu)->vdev.vfio_device))
641 		return;
642 
643 	vfio_device_put(((struct intel_vgpu *)vgpu)->vdev.vfio_device);
644 }
645 
646 static int intel_vgpu_create(struct kobject *kobj, struct mdev_device *mdev)
647 {
648 	struct intel_vgpu *vgpu = NULL;
649 	struct intel_vgpu_type *type;
650 	struct device *pdev;
651 	void *gvt;
652 	int ret;
653 
654 	pdev = mdev_parent_dev(mdev);
655 	gvt = kdev_to_i915(pdev)->gvt;
656 
657 	type = intel_gvt_ops->gvt_find_vgpu_type(gvt, kobject_name(kobj));
658 	if (!type) {
659 		gvt_vgpu_err("failed to find type %s to create\n",
660 						kobject_name(kobj));
661 		ret = -EINVAL;
662 		goto out;
663 	}
664 
665 	vgpu = intel_gvt_ops->vgpu_create(gvt, type);
666 	if (IS_ERR_OR_NULL(vgpu)) {
667 		ret = vgpu == NULL ? -EFAULT : PTR_ERR(vgpu);
668 		gvt_err("failed to create intel vgpu: %d\n", ret);
669 		goto out;
670 	}
671 
672 	INIT_WORK(&vgpu->vdev.release_work, intel_vgpu_release_work);
673 
674 	vgpu->vdev.mdev = mdev;
675 	mdev_set_drvdata(mdev, vgpu);
676 
677 	gvt_dbg_core("intel_vgpu_create succeeded for mdev: %s\n",
678 		     dev_name(mdev_dev(mdev)));
679 	ret = 0;
680 
681 out:
682 	return ret;
683 }
684 
685 static int intel_vgpu_remove(struct mdev_device *mdev)
686 {
687 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
688 
689 	if (handle_valid(vgpu->handle))
690 		return -EBUSY;
691 
692 	intel_gvt_ops->vgpu_destroy(vgpu);
693 	return 0;
694 }
695 
696 static int intel_vgpu_iommu_notifier(struct notifier_block *nb,
697 				     unsigned long action, void *data)
698 {
699 	struct intel_vgpu *vgpu = container_of(nb,
700 					struct intel_vgpu,
701 					vdev.iommu_notifier);
702 
703 	if (action == VFIO_IOMMU_NOTIFY_DMA_UNMAP) {
704 		struct vfio_iommu_type1_dma_unmap *unmap = data;
705 		struct gvt_dma *entry;
706 		unsigned long iov_pfn, end_iov_pfn;
707 
708 		iov_pfn = unmap->iova >> PAGE_SHIFT;
709 		end_iov_pfn = iov_pfn + unmap->size / PAGE_SIZE;
710 
711 		mutex_lock(&vgpu->vdev.cache_lock);
712 		for (; iov_pfn < end_iov_pfn; iov_pfn++) {
713 			entry = __gvt_cache_find_gfn(vgpu, iov_pfn);
714 			if (!entry)
715 				continue;
716 
717 			gvt_dma_unmap_page(vgpu, entry->gfn, entry->dma_addr,
718 					   entry->size);
719 			__gvt_cache_remove_entry(vgpu, entry);
720 		}
721 		mutex_unlock(&vgpu->vdev.cache_lock);
722 	}
723 
724 	return NOTIFY_OK;
725 }
726 
727 static int intel_vgpu_group_notifier(struct notifier_block *nb,
728 				     unsigned long action, void *data)
729 {
730 	struct intel_vgpu *vgpu = container_of(nb,
731 					struct intel_vgpu,
732 					vdev.group_notifier);
733 
734 	/* the only action we care about */
735 	if (action == VFIO_GROUP_NOTIFY_SET_KVM) {
736 		vgpu->vdev.kvm = data;
737 
738 		if (!data)
739 			schedule_work(&vgpu->vdev.release_work);
740 	}
741 
742 	return NOTIFY_OK;
743 }
744 
745 static int intel_vgpu_open(struct mdev_device *mdev)
746 {
747 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
748 	unsigned long events;
749 	int ret;
750 
751 	vgpu->vdev.iommu_notifier.notifier_call = intel_vgpu_iommu_notifier;
752 	vgpu->vdev.group_notifier.notifier_call = intel_vgpu_group_notifier;
753 
754 	events = VFIO_IOMMU_NOTIFY_DMA_UNMAP;
755 	ret = vfio_register_notifier(mdev_dev(mdev), VFIO_IOMMU_NOTIFY, &events,
756 				&vgpu->vdev.iommu_notifier);
757 	if (ret != 0) {
758 		gvt_vgpu_err("vfio_register_notifier for iommu failed: %d\n",
759 			ret);
760 		goto out;
761 	}
762 
763 	events = VFIO_GROUP_NOTIFY_SET_KVM;
764 	ret = vfio_register_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY, &events,
765 				&vgpu->vdev.group_notifier);
766 	if (ret != 0) {
767 		gvt_vgpu_err("vfio_register_notifier for group failed: %d\n",
768 			ret);
769 		goto undo_iommu;
770 	}
771 
772 	/* Take a module reference as mdev core doesn't take
773 	 * a reference for vendor driver.
774 	 */
775 	if (!try_module_get(THIS_MODULE))
776 		goto undo_group;
777 
778 	ret = kvmgt_guest_init(mdev);
779 	if (ret)
780 		goto undo_group;
781 
782 	intel_gvt_ops->vgpu_activate(vgpu);
783 
784 	atomic_set(&vgpu->vdev.released, 0);
785 	return ret;
786 
787 undo_group:
788 	vfio_unregister_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY,
789 					&vgpu->vdev.group_notifier);
790 
791 undo_iommu:
792 	vfio_unregister_notifier(mdev_dev(mdev), VFIO_IOMMU_NOTIFY,
793 					&vgpu->vdev.iommu_notifier);
794 out:
795 	return ret;
796 }
797 
798 static void intel_vgpu_release_msi_eventfd_ctx(struct intel_vgpu *vgpu)
799 {
800 	struct eventfd_ctx *trigger;
801 
802 	trigger = vgpu->vdev.msi_trigger;
803 	if (trigger) {
804 		eventfd_ctx_put(trigger);
805 		vgpu->vdev.msi_trigger = NULL;
806 	}
807 }
808 
809 static void __intel_vgpu_release(struct intel_vgpu *vgpu)
810 {
811 	struct kvmgt_guest_info *info;
812 	int ret;
813 
814 	if (!handle_valid(vgpu->handle))
815 		return;
816 
817 	if (atomic_cmpxchg(&vgpu->vdev.released, 0, 1))
818 		return;
819 
820 	intel_gvt_ops->vgpu_release(vgpu);
821 
822 	ret = vfio_unregister_notifier(mdev_dev(vgpu->vdev.mdev), VFIO_IOMMU_NOTIFY,
823 					&vgpu->vdev.iommu_notifier);
824 	WARN(ret, "vfio_unregister_notifier for iommu failed: %d\n", ret);
825 
826 	ret = vfio_unregister_notifier(mdev_dev(vgpu->vdev.mdev), VFIO_GROUP_NOTIFY,
827 					&vgpu->vdev.group_notifier);
828 	WARN(ret, "vfio_unregister_notifier for group failed: %d\n", ret);
829 
830 	/* dereference module reference taken at open */
831 	module_put(THIS_MODULE);
832 
833 	info = (struct kvmgt_guest_info *)vgpu->handle;
834 	kvmgt_guest_exit(info);
835 
836 	intel_vgpu_release_msi_eventfd_ctx(vgpu);
837 
838 	vgpu->vdev.kvm = NULL;
839 	vgpu->handle = 0;
840 }
841 
842 static void intel_vgpu_release(struct mdev_device *mdev)
843 {
844 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
845 
846 	__intel_vgpu_release(vgpu);
847 }
848 
849 static void intel_vgpu_release_work(struct work_struct *work)
850 {
851 	struct intel_vgpu *vgpu = container_of(work, struct intel_vgpu,
852 					vdev.release_work);
853 
854 	__intel_vgpu_release(vgpu);
855 }
856 
857 static u64 intel_vgpu_get_bar_addr(struct intel_vgpu *vgpu, int bar)
858 {
859 	u32 start_lo, start_hi;
860 	u32 mem_type;
861 
862 	start_lo = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
863 			PCI_BASE_ADDRESS_MEM_MASK;
864 	mem_type = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
865 			PCI_BASE_ADDRESS_MEM_TYPE_MASK;
866 
867 	switch (mem_type) {
868 	case PCI_BASE_ADDRESS_MEM_TYPE_64:
869 		start_hi = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space
870 						+ bar + 4));
871 		break;
872 	case PCI_BASE_ADDRESS_MEM_TYPE_32:
873 	case PCI_BASE_ADDRESS_MEM_TYPE_1M:
874 		/* 1M mem BAR treated as 32-bit BAR */
875 	default:
876 		/* mem unknown type treated as 32-bit BAR */
877 		start_hi = 0;
878 		break;
879 	}
880 
881 	return ((u64)start_hi << 32) | start_lo;
882 }
883 
884 static int intel_vgpu_bar_rw(struct intel_vgpu *vgpu, int bar, u64 off,
885 			     void *buf, unsigned int count, bool is_write)
886 {
887 	u64 bar_start = intel_vgpu_get_bar_addr(vgpu, bar);
888 	int ret;
889 
890 	if (is_write)
891 		ret = intel_gvt_ops->emulate_mmio_write(vgpu,
892 					bar_start + off, buf, count);
893 	else
894 		ret = intel_gvt_ops->emulate_mmio_read(vgpu,
895 					bar_start + off, buf, count);
896 	return ret;
897 }
898 
899 static inline bool intel_vgpu_in_aperture(struct intel_vgpu *vgpu, u64 off)
900 {
901 	return off >= vgpu_aperture_offset(vgpu) &&
902 	       off < vgpu_aperture_offset(vgpu) + vgpu_aperture_sz(vgpu);
903 }
904 
905 static int intel_vgpu_aperture_rw(struct intel_vgpu *vgpu, u64 off,
906 		void *buf, unsigned long count, bool is_write)
907 {
908 	void __iomem *aperture_va;
909 
910 	if (!intel_vgpu_in_aperture(vgpu, off) ||
911 	    !intel_vgpu_in_aperture(vgpu, off + count)) {
912 		gvt_vgpu_err("Invalid aperture offset %llu\n", off);
913 		return -EINVAL;
914 	}
915 
916 	aperture_va = io_mapping_map_wc(&vgpu->gvt->dev_priv->ggtt.iomap,
917 					ALIGN_DOWN(off, PAGE_SIZE),
918 					count + offset_in_page(off));
919 	if (!aperture_va)
920 		return -EIO;
921 
922 	if (is_write)
923 		memcpy_toio(aperture_va + offset_in_page(off), buf, count);
924 	else
925 		memcpy_fromio(buf, aperture_va + offset_in_page(off), count);
926 
927 	io_mapping_unmap(aperture_va);
928 
929 	return 0;
930 }
931 
932 static ssize_t intel_vgpu_rw(struct mdev_device *mdev, char *buf,
933 			size_t count, loff_t *ppos, bool is_write)
934 {
935 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
936 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
937 	u64 pos = *ppos & VFIO_PCI_OFFSET_MASK;
938 	int ret = -EINVAL;
939 
940 
941 	if (index >= VFIO_PCI_NUM_REGIONS + vgpu->vdev.num_regions) {
942 		gvt_vgpu_err("invalid index: %u\n", index);
943 		return -EINVAL;
944 	}
945 
946 	switch (index) {
947 	case VFIO_PCI_CONFIG_REGION_INDEX:
948 		if (is_write)
949 			ret = intel_gvt_ops->emulate_cfg_write(vgpu, pos,
950 						buf, count);
951 		else
952 			ret = intel_gvt_ops->emulate_cfg_read(vgpu, pos,
953 						buf, count);
954 		break;
955 	case VFIO_PCI_BAR0_REGION_INDEX:
956 		ret = intel_vgpu_bar_rw(vgpu, PCI_BASE_ADDRESS_0, pos,
957 					buf, count, is_write);
958 		break;
959 	case VFIO_PCI_BAR2_REGION_INDEX:
960 		ret = intel_vgpu_aperture_rw(vgpu, pos, buf, count, is_write);
961 		break;
962 	case VFIO_PCI_BAR1_REGION_INDEX:
963 	case VFIO_PCI_BAR3_REGION_INDEX:
964 	case VFIO_PCI_BAR4_REGION_INDEX:
965 	case VFIO_PCI_BAR5_REGION_INDEX:
966 	case VFIO_PCI_VGA_REGION_INDEX:
967 	case VFIO_PCI_ROM_REGION_INDEX:
968 		break;
969 	default:
970 		if (index >= VFIO_PCI_NUM_REGIONS + vgpu->vdev.num_regions)
971 			return -EINVAL;
972 
973 		index -= VFIO_PCI_NUM_REGIONS;
974 		return vgpu->vdev.region[index].ops->rw(vgpu, buf, count,
975 				ppos, is_write);
976 	}
977 
978 	return ret == 0 ? count : ret;
979 }
980 
981 static bool gtt_entry(struct mdev_device *mdev, loff_t *ppos)
982 {
983 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
984 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
985 	struct intel_gvt *gvt = vgpu->gvt;
986 	int offset;
987 
988 	/* Only allow MMIO GGTT entry access */
989 	if (index != PCI_BASE_ADDRESS_0)
990 		return false;
991 
992 	offset = (u64)(*ppos & VFIO_PCI_OFFSET_MASK) -
993 		intel_vgpu_get_bar_gpa(vgpu, PCI_BASE_ADDRESS_0);
994 
995 	return (offset >= gvt->device_info.gtt_start_offset &&
996 		offset < gvt->device_info.gtt_start_offset + gvt_ggtt_sz(gvt)) ?
997 			true : false;
998 }
999 
1000 static ssize_t intel_vgpu_read(struct mdev_device *mdev, char __user *buf,
1001 			size_t count, loff_t *ppos)
1002 {
1003 	unsigned int done = 0;
1004 	int ret;
1005 
1006 	while (count) {
1007 		size_t filled;
1008 
1009 		/* Only support GGTT entry 8 bytes read */
1010 		if (count >= 8 && !(*ppos % 8) &&
1011 			gtt_entry(mdev, ppos)) {
1012 			u64 val;
1013 
1014 			ret = intel_vgpu_rw(mdev, (char *)&val, sizeof(val),
1015 					ppos, false);
1016 			if (ret <= 0)
1017 				goto read_err;
1018 
1019 			if (copy_to_user(buf, &val, sizeof(val)))
1020 				goto read_err;
1021 
1022 			filled = 8;
1023 		} else if (count >= 4 && !(*ppos % 4)) {
1024 			u32 val;
1025 
1026 			ret = intel_vgpu_rw(mdev, (char *)&val, sizeof(val),
1027 					ppos, false);
1028 			if (ret <= 0)
1029 				goto read_err;
1030 
1031 			if (copy_to_user(buf, &val, sizeof(val)))
1032 				goto read_err;
1033 
1034 			filled = 4;
1035 		} else if (count >= 2 && !(*ppos % 2)) {
1036 			u16 val;
1037 
1038 			ret = intel_vgpu_rw(mdev, (char *)&val, sizeof(val),
1039 					ppos, false);
1040 			if (ret <= 0)
1041 				goto read_err;
1042 
1043 			if (copy_to_user(buf, &val, sizeof(val)))
1044 				goto read_err;
1045 
1046 			filled = 2;
1047 		} else {
1048 			u8 val;
1049 
1050 			ret = intel_vgpu_rw(mdev, &val, sizeof(val), ppos,
1051 					false);
1052 			if (ret <= 0)
1053 				goto read_err;
1054 
1055 			if (copy_to_user(buf, &val, sizeof(val)))
1056 				goto read_err;
1057 
1058 			filled = 1;
1059 		}
1060 
1061 		count -= filled;
1062 		done += filled;
1063 		*ppos += filled;
1064 		buf += filled;
1065 	}
1066 
1067 	return done;
1068 
1069 read_err:
1070 	return -EFAULT;
1071 }
1072 
1073 static ssize_t intel_vgpu_write(struct mdev_device *mdev,
1074 				const char __user *buf,
1075 				size_t count, loff_t *ppos)
1076 {
1077 	unsigned int done = 0;
1078 	int ret;
1079 
1080 	while (count) {
1081 		size_t filled;
1082 
1083 		/* Only support GGTT entry 8 bytes write */
1084 		if (count >= 8 && !(*ppos % 8) &&
1085 			gtt_entry(mdev, ppos)) {
1086 			u64 val;
1087 
1088 			if (copy_from_user(&val, buf, sizeof(val)))
1089 				goto write_err;
1090 
1091 			ret = intel_vgpu_rw(mdev, (char *)&val, sizeof(val),
1092 					ppos, true);
1093 			if (ret <= 0)
1094 				goto write_err;
1095 
1096 			filled = 8;
1097 		} else if (count >= 4 && !(*ppos % 4)) {
1098 			u32 val;
1099 
1100 			if (copy_from_user(&val, buf, sizeof(val)))
1101 				goto write_err;
1102 
1103 			ret = intel_vgpu_rw(mdev, (char *)&val, sizeof(val),
1104 					ppos, true);
1105 			if (ret <= 0)
1106 				goto write_err;
1107 
1108 			filled = 4;
1109 		} else if (count >= 2 && !(*ppos % 2)) {
1110 			u16 val;
1111 
1112 			if (copy_from_user(&val, buf, sizeof(val)))
1113 				goto write_err;
1114 
1115 			ret = intel_vgpu_rw(mdev, (char *)&val,
1116 					sizeof(val), ppos, true);
1117 			if (ret <= 0)
1118 				goto write_err;
1119 
1120 			filled = 2;
1121 		} else {
1122 			u8 val;
1123 
1124 			if (copy_from_user(&val, buf, sizeof(val)))
1125 				goto write_err;
1126 
1127 			ret = intel_vgpu_rw(mdev, &val, sizeof(val),
1128 					ppos, true);
1129 			if (ret <= 0)
1130 				goto write_err;
1131 
1132 			filled = 1;
1133 		}
1134 
1135 		count -= filled;
1136 		done += filled;
1137 		*ppos += filled;
1138 		buf += filled;
1139 	}
1140 
1141 	return done;
1142 write_err:
1143 	return -EFAULT;
1144 }
1145 
1146 static int intel_vgpu_mmap(struct mdev_device *mdev, struct vm_area_struct *vma)
1147 {
1148 	unsigned int index;
1149 	u64 virtaddr;
1150 	unsigned long req_size, pgoff, req_start;
1151 	pgprot_t pg_prot;
1152 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
1153 
1154 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1155 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1156 		return -EINVAL;
1157 
1158 	if (vma->vm_end < vma->vm_start)
1159 		return -EINVAL;
1160 	if ((vma->vm_flags & VM_SHARED) == 0)
1161 		return -EINVAL;
1162 	if (index != VFIO_PCI_BAR2_REGION_INDEX)
1163 		return -EINVAL;
1164 
1165 	pg_prot = vma->vm_page_prot;
1166 	virtaddr = vma->vm_start;
1167 	req_size = vma->vm_end - vma->vm_start;
1168 	pgoff = vma->vm_pgoff &
1169 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1170 	req_start = pgoff << PAGE_SHIFT;
1171 
1172 	if (!intel_vgpu_in_aperture(vgpu, req_start))
1173 		return -EINVAL;
1174 	if (req_start + req_size >
1175 	    vgpu_aperture_offset(vgpu) + vgpu_aperture_sz(vgpu))
1176 		return -EINVAL;
1177 
1178 	pgoff = (gvt_aperture_pa_base(vgpu->gvt) >> PAGE_SHIFT) + pgoff;
1179 
1180 	return remap_pfn_range(vma, virtaddr, pgoff, req_size, pg_prot);
1181 }
1182 
1183 static int intel_vgpu_get_irq_count(struct intel_vgpu *vgpu, int type)
1184 {
1185 	if (type == VFIO_PCI_INTX_IRQ_INDEX || type == VFIO_PCI_MSI_IRQ_INDEX)
1186 		return 1;
1187 
1188 	return 0;
1189 }
1190 
1191 static int intel_vgpu_set_intx_mask(struct intel_vgpu *vgpu,
1192 			unsigned int index, unsigned int start,
1193 			unsigned int count, u32 flags,
1194 			void *data)
1195 {
1196 	return 0;
1197 }
1198 
1199 static int intel_vgpu_set_intx_unmask(struct intel_vgpu *vgpu,
1200 			unsigned int index, unsigned int start,
1201 			unsigned int count, u32 flags, void *data)
1202 {
1203 	return 0;
1204 }
1205 
1206 static int intel_vgpu_set_intx_trigger(struct intel_vgpu *vgpu,
1207 		unsigned int index, unsigned int start, unsigned int count,
1208 		u32 flags, void *data)
1209 {
1210 	return 0;
1211 }
1212 
1213 static int intel_vgpu_set_msi_trigger(struct intel_vgpu *vgpu,
1214 		unsigned int index, unsigned int start, unsigned int count,
1215 		u32 flags, void *data)
1216 {
1217 	struct eventfd_ctx *trigger;
1218 
1219 	if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
1220 		int fd = *(int *)data;
1221 
1222 		trigger = eventfd_ctx_fdget(fd);
1223 		if (IS_ERR(trigger)) {
1224 			gvt_vgpu_err("eventfd_ctx_fdget failed\n");
1225 			return PTR_ERR(trigger);
1226 		}
1227 		vgpu->vdev.msi_trigger = trigger;
1228 	} else if ((flags & VFIO_IRQ_SET_DATA_NONE) && !count)
1229 		intel_vgpu_release_msi_eventfd_ctx(vgpu);
1230 
1231 	return 0;
1232 }
1233 
1234 static int intel_vgpu_set_irqs(struct intel_vgpu *vgpu, u32 flags,
1235 		unsigned int index, unsigned int start, unsigned int count,
1236 		void *data)
1237 {
1238 	int (*func)(struct intel_vgpu *vgpu, unsigned int index,
1239 			unsigned int start, unsigned int count, u32 flags,
1240 			void *data) = NULL;
1241 
1242 	switch (index) {
1243 	case VFIO_PCI_INTX_IRQ_INDEX:
1244 		switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
1245 		case VFIO_IRQ_SET_ACTION_MASK:
1246 			func = intel_vgpu_set_intx_mask;
1247 			break;
1248 		case VFIO_IRQ_SET_ACTION_UNMASK:
1249 			func = intel_vgpu_set_intx_unmask;
1250 			break;
1251 		case VFIO_IRQ_SET_ACTION_TRIGGER:
1252 			func = intel_vgpu_set_intx_trigger;
1253 			break;
1254 		}
1255 		break;
1256 	case VFIO_PCI_MSI_IRQ_INDEX:
1257 		switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
1258 		case VFIO_IRQ_SET_ACTION_MASK:
1259 		case VFIO_IRQ_SET_ACTION_UNMASK:
1260 			/* XXX Need masking support exported */
1261 			break;
1262 		case VFIO_IRQ_SET_ACTION_TRIGGER:
1263 			func = intel_vgpu_set_msi_trigger;
1264 			break;
1265 		}
1266 		break;
1267 	}
1268 
1269 	if (!func)
1270 		return -ENOTTY;
1271 
1272 	return func(vgpu, index, start, count, flags, data);
1273 }
1274 
1275 static long intel_vgpu_ioctl(struct mdev_device *mdev, unsigned int cmd,
1276 			     unsigned long arg)
1277 {
1278 	struct intel_vgpu *vgpu = mdev_get_drvdata(mdev);
1279 	unsigned long minsz;
1280 
1281 	gvt_dbg_core("vgpu%d ioctl, cmd: %d\n", vgpu->id, cmd);
1282 
1283 	if (cmd == VFIO_DEVICE_GET_INFO) {
1284 		struct vfio_device_info info;
1285 
1286 		minsz = offsetofend(struct vfio_device_info, num_irqs);
1287 
1288 		if (copy_from_user(&info, (void __user *)arg, minsz))
1289 			return -EFAULT;
1290 
1291 		if (info.argsz < minsz)
1292 			return -EINVAL;
1293 
1294 		info.flags = VFIO_DEVICE_FLAGS_PCI;
1295 		info.flags |= VFIO_DEVICE_FLAGS_RESET;
1296 		info.num_regions = VFIO_PCI_NUM_REGIONS +
1297 				vgpu->vdev.num_regions;
1298 		info.num_irqs = VFIO_PCI_NUM_IRQS;
1299 
1300 		return copy_to_user((void __user *)arg, &info, minsz) ?
1301 			-EFAULT : 0;
1302 
1303 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
1304 		struct vfio_region_info info;
1305 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
1306 		unsigned int i;
1307 		int ret;
1308 		struct vfio_region_info_cap_sparse_mmap *sparse = NULL;
1309 		int nr_areas = 1;
1310 		int cap_type_id;
1311 
1312 		minsz = offsetofend(struct vfio_region_info, offset);
1313 
1314 		if (copy_from_user(&info, (void __user *)arg, minsz))
1315 			return -EFAULT;
1316 
1317 		if (info.argsz < minsz)
1318 			return -EINVAL;
1319 
1320 		switch (info.index) {
1321 		case VFIO_PCI_CONFIG_REGION_INDEX:
1322 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1323 			info.size = vgpu->gvt->device_info.cfg_space_size;
1324 			info.flags = VFIO_REGION_INFO_FLAG_READ |
1325 				     VFIO_REGION_INFO_FLAG_WRITE;
1326 			break;
1327 		case VFIO_PCI_BAR0_REGION_INDEX:
1328 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1329 			info.size = vgpu->cfg_space.bar[info.index].size;
1330 			if (!info.size) {
1331 				info.flags = 0;
1332 				break;
1333 			}
1334 
1335 			info.flags = VFIO_REGION_INFO_FLAG_READ |
1336 				     VFIO_REGION_INFO_FLAG_WRITE;
1337 			break;
1338 		case VFIO_PCI_BAR1_REGION_INDEX:
1339 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1340 			info.size = 0;
1341 			info.flags = 0;
1342 			break;
1343 		case VFIO_PCI_BAR2_REGION_INDEX:
1344 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1345 			info.flags = VFIO_REGION_INFO_FLAG_CAPS |
1346 					VFIO_REGION_INFO_FLAG_MMAP |
1347 					VFIO_REGION_INFO_FLAG_READ |
1348 					VFIO_REGION_INFO_FLAG_WRITE;
1349 			info.size = gvt_aperture_sz(vgpu->gvt);
1350 
1351 			sparse = kzalloc(struct_size(sparse, areas, nr_areas),
1352 					 GFP_KERNEL);
1353 			if (!sparse)
1354 				return -ENOMEM;
1355 
1356 			sparse->header.id = VFIO_REGION_INFO_CAP_SPARSE_MMAP;
1357 			sparse->header.version = 1;
1358 			sparse->nr_areas = nr_areas;
1359 			cap_type_id = VFIO_REGION_INFO_CAP_SPARSE_MMAP;
1360 			sparse->areas[0].offset =
1361 					PAGE_ALIGN(vgpu_aperture_offset(vgpu));
1362 			sparse->areas[0].size = vgpu_aperture_sz(vgpu);
1363 			break;
1364 
1365 		case VFIO_PCI_BAR3_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1366 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1367 			info.size = 0;
1368 			info.flags = 0;
1369 
1370 			gvt_dbg_core("get region info bar:%d\n", info.index);
1371 			break;
1372 
1373 		case VFIO_PCI_ROM_REGION_INDEX:
1374 		case VFIO_PCI_VGA_REGION_INDEX:
1375 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1376 			info.size = 0;
1377 			info.flags = 0;
1378 
1379 			gvt_dbg_core("get region info index:%d\n", info.index);
1380 			break;
1381 		default:
1382 			{
1383 				struct vfio_region_info_cap_type cap_type = {
1384 					.header.id = VFIO_REGION_INFO_CAP_TYPE,
1385 					.header.version = 1 };
1386 
1387 				if (info.index >= VFIO_PCI_NUM_REGIONS +
1388 						vgpu->vdev.num_regions)
1389 					return -EINVAL;
1390 				info.index =
1391 					array_index_nospec(info.index,
1392 							VFIO_PCI_NUM_REGIONS +
1393 							vgpu->vdev.num_regions);
1394 
1395 				i = info.index - VFIO_PCI_NUM_REGIONS;
1396 
1397 				info.offset =
1398 					VFIO_PCI_INDEX_TO_OFFSET(info.index);
1399 				info.size = vgpu->vdev.region[i].size;
1400 				info.flags = vgpu->vdev.region[i].flags;
1401 
1402 				cap_type.type = vgpu->vdev.region[i].type;
1403 				cap_type.subtype = vgpu->vdev.region[i].subtype;
1404 
1405 				ret = vfio_info_add_capability(&caps,
1406 							&cap_type.header,
1407 							sizeof(cap_type));
1408 				if (ret)
1409 					return ret;
1410 			}
1411 		}
1412 
1413 		if ((info.flags & VFIO_REGION_INFO_FLAG_CAPS) && sparse) {
1414 			switch (cap_type_id) {
1415 			case VFIO_REGION_INFO_CAP_SPARSE_MMAP:
1416 				ret = vfio_info_add_capability(&caps,
1417 					&sparse->header,
1418 					struct_size(sparse, areas,
1419 						    sparse->nr_areas));
1420 				if (ret) {
1421 					kfree(sparse);
1422 					return ret;
1423 				}
1424 				break;
1425 			default:
1426 				kfree(sparse);
1427 				return -EINVAL;
1428 			}
1429 		}
1430 
1431 		if (caps.size) {
1432 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
1433 			if (info.argsz < sizeof(info) + caps.size) {
1434 				info.argsz = sizeof(info) + caps.size;
1435 				info.cap_offset = 0;
1436 			} else {
1437 				vfio_info_cap_shift(&caps, sizeof(info));
1438 				if (copy_to_user((void __user *)arg +
1439 						  sizeof(info), caps.buf,
1440 						  caps.size)) {
1441 					kfree(caps.buf);
1442 					kfree(sparse);
1443 					return -EFAULT;
1444 				}
1445 				info.cap_offset = sizeof(info);
1446 			}
1447 
1448 			kfree(caps.buf);
1449 		}
1450 
1451 		kfree(sparse);
1452 		return copy_to_user((void __user *)arg, &info, minsz) ?
1453 			-EFAULT : 0;
1454 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
1455 		struct vfio_irq_info info;
1456 
1457 		minsz = offsetofend(struct vfio_irq_info, count);
1458 
1459 		if (copy_from_user(&info, (void __user *)arg, minsz))
1460 			return -EFAULT;
1461 
1462 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
1463 			return -EINVAL;
1464 
1465 		switch (info.index) {
1466 		case VFIO_PCI_INTX_IRQ_INDEX:
1467 		case VFIO_PCI_MSI_IRQ_INDEX:
1468 			break;
1469 		default:
1470 			return -EINVAL;
1471 		}
1472 
1473 		info.flags = VFIO_IRQ_INFO_EVENTFD;
1474 
1475 		info.count = intel_vgpu_get_irq_count(vgpu, info.index);
1476 
1477 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1478 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
1479 				       VFIO_IRQ_INFO_AUTOMASKED);
1480 		else
1481 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
1482 
1483 		return copy_to_user((void __user *)arg, &info, minsz) ?
1484 			-EFAULT : 0;
1485 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
1486 		struct vfio_irq_set hdr;
1487 		u8 *data = NULL;
1488 		int ret = 0;
1489 		size_t data_size = 0;
1490 
1491 		minsz = offsetofend(struct vfio_irq_set, count);
1492 
1493 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1494 			return -EFAULT;
1495 
1496 		if (!(hdr.flags & VFIO_IRQ_SET_DATA_NONE)) {
1497 			int max = intel_vgpu_get_irq_count(vgpu, hdr.index);
1498 
1499 			ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
1500 						VFIO_PCI_NUM_IRQS, &data_size);
1501 			if (ret) {
1502 				gvt_vgpu_err("intel:vfio_set_irqs_validate_and_prepare failed\n");
1503 				return -EINVAL;
1504 			}
1505 			if (data_size) {
1506 				data = memdup_user((void __user *)(arg + minsz),
1507 						   data_size);
1508 				if (IS_ERR(data))
1509 					return PTR_ERR(data);
1510 			}
1511 		}
1512 
1513 		ret = intel_vgpu_set_irqs(vgpu, hdr.flags, hdr.index,
1514 					hdr.start, hdr.count, data);
1515 		kfree(data);
1516 
1517 		return ret;
1518 	} else if (cmd == VFIO_DEVICE_RESET) {
1519 		intel_gvt_ops->vgpu_reset(vgpu);
1520 		return 0;
1521 	} else if (cmd == VFIO_DEVICE_QUERY_GFX_PLANE) {
1522 		struct vfio_device_gfx_plane_info dmabuf;
1523 		int ret = 0;
1524 
1525 		minsz = offsetofend(struct vfio_device_gfx_plane_info,
1526 				    dmabuf_id);
1527 		if (copy_from_user(&dmabuf, (void __user *)arg, minsz))
1528 			return -EFAULT;
1529 		if (dmabuf.argsz < minsz)
1530 			return -EINVAL;
1531 
1532 		ret = intel_gvt_ops->vgpu_query_plane(vgpu, &dmabuf);
1533 		if (ret != 0)
1534 			return ret;
1535 
1536 		return copy_to_user((void __user *)arg, &dmabuf, minsz) ?
1537 								-EFAULT : 0;
1538 	} else if (cmd == VFIO_DEVICE_GET_GFX_DMABUF) {
1539 		__u32 dmabuf_id;
1540 		__s32 dmabuf_fd;
1541 
1542 		if (get_user(dmabuf_id, (__u32 __user *)arg))
1543 			return -EFAULT;
1544 
1545 		dmabuf_fd = intel_gvt_ops->vgpu_get_dmabuf(vgpu, dmabuf_id);
1546 		return dmabuf_fd;
1547 
1548 	}
1549 
1550 	return -ENOTTY;
1551 }
1552 
1553 static ssize_t
1554 vgpu_id_show(struct device *dev, struct device_attribute *attr,
1555 	     char *buf)
1556 {
1557 	struct mdev_device *mdev = mdev_from_dev(dev);
1558 
1559 	if (mdev) {
1560 		struct intel_vgpu *vgpu = (struct intel_vgpu *)
1561 			mdev_get_drvdata(mdev);
1562 		return sprintf(buf, "%d\n", vgpu->id);
1563 	}
1564 	return sprintf(buf, "\n");
1565 }
1566 
1567 static DEVICE_ATTR_RO(vgpu_id);
1568 
1569 static struct attribute *intel_vgpu_attrs[] = {
1570 	&dev_attr_vgpu_id.attr,
1571 	NULL
1572 };
1573 
1574 static const struct attribute_group intel_vgpu_group = {
1575 	.name = "intel_vgpu",
1576 	.attrs = intel_vgpu_attrs,
1577 };
1578 
1579 static const struct attribute_group *intel_vgpu_groups[] = {
1580 	&intel_vgpu_group,
1581 	NULL,
1582 };
1583 
1584 static struct mdev_parent_ops intel_vgpu_ops = {
1585 	.mdev_attr_groups       = intel_vgpu_groups,
1586 	.create			= intel_vgpu_create,
1587 	.remove			= intel_vgpu_remove,
1588 
1589 	.open			= intel_vgpu_open,
1590 	.release		= intel_vgpu_release,
1591 
1592 	.read			= intel_vgpu_read,
1593 	.write			= intel_vgpu_write,
1594 	.mmap			= intel_vgpu_mmap,
1595 	.ioctl			= intel_vgpu_ioctl,
1596 };
1597 
1598 static int kvmgt_host_init(struct device *dev, void *gvt, const void *ops)
1599 {
1600 	struct attribute **kvm_type_attrs;
1601 	struct attribute_group **kvm_vgpu_type_groups;
1602 
1603 	intel_gvt_ops = ops;
1604 	if (!intel_gvt_ops->get_gvt_attrs(&kvm_type_attrs,
1605 			&kvm_vgpu_type_groups))
1606 		return -EFAULT;
1607 	intel_vgpu_ops.supported_type_groups = kvm_vgpu_type_groups;
1608 
1609 	return mdev_register_device(dev, &intel_vgpu_ops);
1610 }
1611 
1612 static void kvmgt_host_exit(struct device *dev)
1613 {
1614 	mdev_unregister_device(dev);
1615 }
1616 
1617 static int kvmgt_page_track_add(unsigned long handle, u64 gfn)
1618 {
1619 	struct kvmgt_guest_info *info;
1620 	struct kvm *kvm;
1621 	struct kvm_memory_slot *slot;
1622 	int idx;
1623 
1624 	if (!handle_valid(handle))
1625 		return -ESRCH;
1626 
1627 	info = (struct kvmgt_guest_info *)handle;
1628 	kvm = info->kvm;
1629 
1630 	idx = srcu_read_lock(&kvm->srcu);
1631 	slot = gfn_to_memslot(kvm, gfn);
1632 	if (!slot) {
1633 		srcu_read_unlock(&kvm->srcu, idx);
1634 		return -EINVAL;
1635 	}
1636 
1637 	spin_lock(&kvm->mmu_lock);
1638 
1639 	if (kvmgt_gfn_is_write_protected(info, gfn))
1640 		goto out;
1641 
1642 	kvm_slot_page_track_add_page(kvm, slot, gfn, KVM_PAGE_TRACK_WRITE);
1643 	kvmgt_protect_table_add(info, gfn);
1644 
1645 out:
1646 	spin_unlock(&kvm->mmu_lock);
1647 	srcu_read_unlock(&kvm->srcu, idx);
1648 	return 0;
1649 }
1650 
1651 static int kvmgt_page_track_remove(unsigned long handle, u64 gfn)
1652 {
1653 	struct kvmgt_guest_info *info;
1654 	struct kvm *kvm;
1655 	struct kvm_memory_slot *slot;
1656 	int idx;
1657 
1658 	if (!handle_valid(handle))
1659 		return 0;
1660 
1661 	info = (struct kvmgt_guest_info *)handle;
1662 	kvm = info->kvm;
1663 
1664 	idx = srcu_read_lock(&kvm->srcu);
1665 	slot = gfn_to_memslot(kvm, gfn);
1666 	if (!slot) {
1667 		srcu_read_unlock(&kvm->srcu, idx);
1668 		return -EINVAL;
1669 	}
1670 
1671 	spin_lock(&kvm->mmu_lock);
1672 
1673 	if (!kvmgt_gfn_is_write_protected(info, gfn))
1674 		goto out;
1675 
1676 	kvm_slot_page_track_remove_page(kvm, slot, gfn, KVM_PAGE_TRACK_WRITE);
1677 	kvmgt_protect_table_del(info, gfn);
1678 
1679 out:
1680 	spin_unlock(&kvm->mmu_lock);
1681 	srcu_read_unlock(&kvm->srcu, idx);
1682 	return 0;
1683 }
1684 
1685 static void kvmgt_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1686 		const u8 *val, int len,
1687 		struct kvm_page_track_notifier_node *node)
1688 {
1689 	struct kvmgt_guest_info *info = container_of(node,
1690 					struct kvmgt_guest_info, track_node);
1691 
1692 	if (kvmgt_gfn_is_write_protected(info, gpa_to_gfn(gpa)))
1693 		intel_gvt_ops->write_protect_handler(info->vgpu, gpa,
1694 						     (void *)val, len);
1695 }
1696 
1697 static void kvmgt_page_track_flush_slot(struct kvm *kvm,
1698 		struct kvm_memory_slot *slot,
1699 		struct kvm_page_track_notifier_node *node)
1700 {
1701 	int i;
1702 	gfn_t gfn;
1703 	struct kvmgt_guest_info *info = container_of(node,
1704 					struct kvmgt_guest_info, track_node);
1705 
1706 	spin_lock(&kvm->mmu_lock);
1707 	for (i = 0; i < slot->npages; i++) {
1708 		gfn = slot->base_gfn + i;
1709 		if (kvmgt_gfn_is_write_protected(info, gfn)) {
1710 			kvm_slot_page_track_remove_page(kvm, slot, gfn,
1711 						KVM_PAGE_TRACK_WRITE);
1712 			kvmgt_protect_table_del(info, gfn);
1713 		}
1714 	}
1715 	spin_unlock(&kvm->mmu_lock);
1716 }
1717 
1718 static bool __kvmgt_vgpu_exist(struct intel_vgpu *vgpu, struct kvm *kvm)
1719 {
1720 	struct intel_vgpu *itr;
1721 	struct kvmgt_guest_info *info;
1722 	int id;
1723 	bool ret = false;
1724 
1725 	mutex_lock(&vgpu->gvt->lock);
1726 	for_each_active_vgpu(vgpu->gvt, itr, id) {
1727 		if (!handle_valid(itr->handle))
1728 			continue;
1729 
1730 		info = (struct kvmgt_guest_info *)itr->handle;
1731 		if (kvm && kvm == info->kvm) {
1732 			ret = true;
1733 			goto out;
1734 		}
1735 	}
1736 out:
1737 	mutex_unlock(&vgpu->gvt->lock);
1738 	return ret;
1739 }
1740 
1741 static int kvmgt_guest_init(struct mdev_device *mdev)
1742 {
1743 	struct kvmgt_guest_info *info;
1744 	struct intel_vgpu *vgpu;
1745 	struct kvm *kvm;
1746 
1747 	vgpu = mdev_get_drvdata(mdev);
1748 	if (handle_valid(vgpu->handle))
1749 		return -EEXIST;
1750 
1751 	kvm = vgpu->vdev.kvm;
1752 	if (!kvm || kvm->mm != current->mm) {
1753 		gvt_vgpu_err("KVM is required to use Intel vGPU\n");
1754 		return -ESRCH;
1755 	}
1756 
1757 	if (__kvmgt_vgpu_exist(vgpu, kvm))
1758 		return -EEXIST;
1759 
1760 	info = vzalloc(sizeof(struct kvmgt_guest_info));
1761 	if (!info)
1762 		return -ENOMEM;
1763 
1764 	vgpu->handle = (unsigned long)info;
1765 	info->vgpu = vgpu;
1766 	info->kvm = kvm;
1767 	kvm_get_kvm(info->kvm);
1768 
1769 	kvmgt_protect_table_init(info);
1770 	gvt_cache_init(vgpu);
1771 
1772 	init_completion(&vgpu->vblank_done);
1773 
1774 	info->track_node.track_write = kvmgt_page_track_write;
1775 	info->track_node.track_flush_slot = kvmgt_page_track_flush_slot;
1776 	kvm_page_track_register_notifier(kvm, &info->track_node);
1777 
1778 	info->debugfs_cache_entries = debugfs_create_ulong(
1779 						"kvmgt_nr_cache_entries",
1780 						0444, vgpu->debugfs,
1781 						&vgpu->vdev.nr_cache_entries);
1782 	return 0;
1783 }
1784 
1785 static bool kvmgt_guest_exit(struct kvmgt_guest_info *info)
1786 {
1787 	debugfs_remove(info->debugfs_cache_entries);
1788 
1789 	kvm_page_track_unregister_notifier(info->kvm, &info->track_node);
1790 	kvm_put_kvm(info->kvm);
1791 	kvmgt_protect_table_destroy(info);
1792 	gvt_cache_destroy(info->vgpu);
1793 	vfree(info);
1794 
1795 	return true;
1796 }
1797 
1798 static int kvmgt_attach_vgpu(void *vgpu, unsigned long *handle)
1799 {
1800 	/* nothing to do here */
1801 	return 0;
1802 }
1803 
1804 static void kvmgt_detach_vgpu(void *p_vgpu)
1805 {
1806 	int i;
1807 	struct intel_vgpu *vgpu = (struct intel_vgpu *)p_vgpu;
1808 
1809 	if (!vgpu->vdev.region)
1810 		return;
1811 
1812 	for (i = 0; i < vgpu->vdev.num_regions; i++)
1813 		if (vgpu->vdev.region[i].ops->release)
1814 			vgpu->vdev.region[i].ops->release(vgpu,
1815 					&vgpu->vdev.region[i]);
1816 	vgpu->vdev.num_regions = 0;
1817 	kfree(vgpu->vdev.region);
1818 	vgpu->vdev.region = NULL;
1819 }
1820 
1821 static int kvmgt_inject_msi(unsigned long handle, u32 addr, u16 data)
1822 {
1823 	struct kvmgt_guest_info *info;
1824 	struct intel_vgpu *vgpu;
1825 
1826 	if (!handle_valid(handle))
1827 		return -ESRCH;
1828 
1829 	info = (struct kvmgt_guest_info *)handle;
1830 	vgpu = info->vgpu;
1831 
1832 	/*
1833 	 * When guest is poweroff, msi_trigger is set to NULL, but vgpu's
1834 	 * config and mmio register isn't restored to default during guest
1835 	 * poweroff. If this vgpu is still used in next vm, this vgpu's pipe
1836 	 * may be enabled, then once this vgpu is active, it will get inject
1837 	 * vblank interrupt request. But msi_trigger is null until msi is
1838 	 * enabled by guest. so if msi_trigger is null, success is still
1839 	 * returned and don't inject interrupt into guest.
1840 	 */
1841 	if (vgpu->vdev.msi_trigger == NULL)
1842 		return 0;
1843 
1844 	if (eventfd_signal(vgpu->vdev.msi_trigger, 1) == 1)
1845 		return 0;
1846 
1847 	return -EFAULT;
1848 }
1849 
1850 static unsigned long kvmgt_gfn_to_pfn(unsigned long handle, unsigned long gfn)
1851 {
1852 	struct kvmgt_guest_info *info;
1853 	kvm_pfn_t pfn;
1854 
1855 	if (!handle_valid(handle))
1856 		return INTEL_GVT_INVALID_ADDR;
1857 
1858 	info = (struct kvmgt_guest_info *)handle;
1859 
1860 	pfn = gfn_to_pfn(info->kvm, gfn);
1861 	if (is_error_noslot_pfn(pfn))
1862 		return INTEL_GVT_INVALID_ADDR;
1863 
1864 	return pfn;
1865 }
1866 
1867 static int kvmgt_dma_map_guest_page(unsigned long handle, unsigned long gfn,
1868 		unsigned long size, dma_addr_t *dma_addr)
1869 {
1870 	struct kvmgt_guest_info *info;
1871 	struct intel_vgpu *vgpu;
1872 	struct gvt_dma *entry;
1873 	int ret;
1874 
1875 	if (!handle_valid(handle))
1876 		return -EINVAL;
1877 
1878 	info = (struct kvmgt_guest_info *)handle;
1879 	vgpu = info->vgpu;
1880 
1881 	mutex_lock(&info->vgpu->vdev.cache_lock);
1882 
1883 	entry = __gvt_cache_find_gfn(info->vgpu, gfn);
1884 	if (!entry) {
1885 		ret = gvt_dma_map_page(vgpu, gfn, dma_addr, size);
1886 		if (ret)
1887 			goto err_unlock;
1888 
1889 		ret = __gvt_cache_add(info->vgpu, gfn, *dma_addr, size);
1890 		if (ret)
1891 			goto err_unmap;
1892 	} else if (entry->size != size) {
1893 		/* the same gfn with different size: unmap and re-map */
1894 		gvt_dma_unmap_page(vgpu, gfn, entry->dma_addr, entry->size);
1895 		__gvt_cache_remove_entry(vgpu, entry);
1896 
1897 		ret = gvt_dma_map_page(vgpu, gfn, dma_addr, size);
1898 		if (ret)
1899 			goto err_unlock;
1900 
1901 		ret = __gvt_cache_add(info->vgpu, gfn, *dma_addr, size);
1902 		if (ret)
1903 			goto err_unmap;
1904 	} else {
1905 		kref_get(&entry->ref);
1906 		*dma_addr = entry->dma_addr;
1907 	}
1908 
1909 	mutex_unlock(&info->vgpu->vdev.cache_lock);
1910 	return 0;
1911 
1912 err_unmap:
1913 	gvt_dma_unmap_page(vgpu, gfn, *dma_addr, size);
1914 err_unlock:
1915 	mutex_unlock(&info->vgpu->vdev.cache_lock);
1916 	return ret;
1917 }
1918 
1919 static void __gvt_dma_release(struct kref *ref)
1920 {
1921 	struct gvt_dma *entry = container_of(ref, typeof(*entry), ref);
1922 
1923 	gvt_dma_unmap_page(entry->vgpu, entry->gfn, entry->dma_addr,
1924 			   entry->size);
1925 	__gvt_cache_remove_entry(entry->vgpu, entry);
1926 }
1927 
1928 static void kvmgt_dma_unmap_guest_page(unsigned long handle, dma_addr_t dma_addr)
1929 {
1930 	struct kvmgt_guest_info *info;
1931 	struct gvt_dma *entry;
1932 
1933 	if (!handle_valid(handle))
1934 		return;
1935 
1936 	info = (struct kvmgt_guest_info *)handle;
1937 
1938 	mutex_lock(&info->vgpu->vdev.cache_lock);
1939 	entry = __gvt_cache_find_dma_addr(info->vgpu, dma_addr);
1940 	if (entry)
1941 		kref_put(&entry->ref, __gvt_dma_release);
1942 	mutex_unlock(&info->vgpu->vdev.cache_lock);
1943 }
1944 
1945 static int kvmgt_rw_gpa(unsigned long handle, unsigned long gpa,
1946 			void *buf, unsigned long len, bool write)
1947 {
1948 	struct kvmgt_guest_info *info;
1949 	struct kvm *kvm;
1950 	int idx, ret;
1951 	bool kthread = current->mm == NULL;
1952 
1953 	if (!handle_valid(handle))
1954 		return -ESRCH;
1955 
1956 	info = (struct kvmgt_guest_info *)handle;
1957 	kvm = info->kvm;
1958 
1959 	if (kthread) {
1960 		if (!mmget_not_zero(kvm->mm))
1961 			return -EFAULT;
1962 		use_mm(kvm->mm);
1963 	}
1964 
1965 	idx = srcu_read_lock(&kvm->srcu);
1966 	ret = write ? kvm_write_guest(kvm, gpa, buf, len) :
1967 		      kvm_read_guest(kvm, gpa, buf, len);
1968 	srcu_read_unlock(&kvm->srcu, idx);
1969 
1970 	if (kthread) {
1971 		unuse_mm(kvm->mm);
1972 		mmput(kvm->mm);
1973 	}
1974 
1975 	return ret;
1976 }
1977 
1978 static int kvmgt_read_gpa(unsigned long handle, unsigned long gpa,
1979 			void *buf, unsigned long len)
1980 {
1981 	return kvmgt_rw_gpa(handle, gpa, buf, len, false);
1982 }
1983 
1984 static int kvmgt_write_gpa(unsigned long handle, unsigned long gpa,
1985 			void *buf, unsigned long len)
1986 {
1987 	return kvmgt_rw_gpa(handle, gpa, buf, len, true);
1988 }
1989 
1990 static unsigned long kvmgt_virt_to_pfn(void *addr)
1991 {
1992 	return PFN_DOWN(__pa(addr));
1993 }
1994 
1995 static bool kvmgt_is_valid_gfn(unsigned long handle, unsigned long gfn)
1996 {
1997 	struct kvmgt_guest_info *info;
1998 	struct kvm *kvm;
1999 	int idx;
2000 	bool ret;
2001 
2002 	if (!handle_valid(handle))
2003 		return false;
2004 
2005 	info = (struct kvmgt_guest_info *)handle;
2006 	kvm = info->kvm;
2007 
2008 	idx = srcu_read_lock(&kvm->srcu);
2009 	ret = kvm_is_visible_gfn(kvm, gfn);
2010 	srcu_read_unlock(&kvm->srcu, idx);
2011 
2012 	return ret;
2013 }
2014 
2015 static struct intel_gvt_mpt kvmgt_mpt = {
2016 	.type = INTEL_GVT_HYPERVISOR_KVM,
2017 	.host_init = kvmgt_host_init,
2018 	.host_exit = kvmgt_host_exit,
2019 	.attach_vgpu = kvmgt_attach_vgpu,
2020 	.detach_vgpu = kvmgt_detach_vgpu,
2021 	.inject_msi = kvmgt_inject_msi,
2022 	.from_virt_to_mfn = kvmgt_virt_to_pfn,
2023 	.enable_page_track = kvmgt_page_track_add,
2024 	.disable_page_track = kvmgt_page_track_remove,
2025 	.read_gpa = kvmgt_read_gpa,
2026 	.write_gpa = kvmgt_write_gpa,
2027 	.gfn_to_mfn = kvmgt_gfn_to_pfn,
2028 	.dma_map_guest_page = kvmgt_dma_map_guest_page,
2029 	.dma_unmap_guest_page = kvmgt_dma_unmap_guest_page,
2030 	.set_opregion = kvmgt_set_opregion,
2031 	.set_edid = kvmgt_set_edid,
2032 	.get_vfio_device = kvmgt_get_vfio_device,
2033 	.put_vfio_device = kvmgt_put_vfio_device,
2034 	.is_valid_gfn = kvmgt_is_valid_gfn,
2035 };
2036 
2037 static int __init kvmgt_init(void)
2038 {
2039 	if (intel_gvt_register_hypervisor(&kvmgt_mpt) < 0)
2040 		return -ENODEV;
2041 	return 0;
2042 }
2043 
2044 static void __exit kvmgt_exit(void)
2045 {
2046 	intel_gvt_unregister_hypervisor();
2047 }
2048 
2049 module_init(kvmgt_init);
2050 module_exit(kvmgt_exit);
2051 
2052 MODULE_LICENSE("GPL and additional rights");
2053 MODULE_AUTHOR("Intel Corporation");
2054