xref: /linux/drivers/gpu/drm/i915/gvt/kvmgt.c (revision f2fbc72e6da4f8e01fe5fe3d6871a791e76271c3)
1 /*
2  * KVMGT - the implementation of Intel mediated pass-through framework for KVM
3  *
4  * Copyright(c) 2011-2016 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  *    Eddie Dong <eddie.dong@intel.com>
30  *
31  * Contributors:
32  *    Niu Bing <bing.niu@intel.com>
33  *    Zhi Wang <zhi.a.wang@intel.com>
34  */
35 
36 #include <linux/init.h>
37 #include <linux/mm.h>
38 #include <linux/kthread.h>
39 #include <linux/sched/mm.h>
40 #include <linux/types.h>
41 #include <linux/list.h>
42 #include <linux/rbtree.h>
43 #include <linux/spinlock.h>
44 #include <linux/eventfd.h>
45 #include <linux/mdev.h>
46 #include <linux/debugfs.h>
47 
48 #include <linux/nospec.h>
49 
50 #include <drm/drm_edid.h>
51 
52 #include "i915_drv.h"
53 #include "intel_gvt.h"
54 #include "gvt.h"
55 
56 MODULE_IMPORT_NS(DMA_BUF);
57 MODULE_IMPORT_NS(I915_GVT);
58 
59 /* helper macros copied from vfio-pci */
60 #define VFIO_PCI_OFFSET_SHIFT   40
61 #define VFIO_PCI_OFFSET_TO_INDEX(off)   (off >> VFIO_PCI_OFFSET_SHIFT)
62 #define VFIO_PCI_INDEX_TO_OFFSET(index) ((u64)(index) << VFIO_PCI_OFFSET_SHIFT)
63 #define VFIO_PCI_OFFSET_MASK    (((u64)(1) << VFIO_PCI_OFFSET_SHIFT) - 1)
64 
65 #define EDID_BLOB_OFFSET (PAGE_SIZE/2)
66 
67 #define OPREGION_SIGNATURE "IntelGraphicsMem"
68 
69 struct vfio_region;
70 struct intel_vgpu_regops {
71 	size_t (*rw)(struct intel_vgpu *vgpu, char *buf,
72 			size_t count, loff_t *ppos, bool iswrite);
73 	void (*release)(struct intel_vgpu *vgpu,
74 			struct vfio_region *region);
75 };
76 
77 struct vfio_region {
78 	u32				type;
79 	u32				subtype;
80 	size_t				size;
81 	u32				flags;
82 	const struct intel_vgpu_regops	*ops;
83 	void				*data;
84 };
85 
86 struct vfio_edid_region {
87 	struct vfio_region_gfx_edid vfio_edid_regs;
88 	void *edid_blob;
89 };
90 
91 struct kvmgt_pgfn {
92 	gfn_t gfn;
93 	struct hlist_node hnode;
94 };
95 
96 struct gvt_dma {
97 	struct intel_vgpu *vgpu;
98 	struct rb_node gfn_node;
99 	struct rb_node dma_addr_node;
100 	gfn_t gfn;
101 	dma_addr_t dma_addr;
102 	unsigned long size;
103 	struct kref ref;
104 };
105 
106 #define vfio_dev_to_vgpu(vfio_dev) \
107 	container_of((vfio_dev), struct intel_vgpu, vfio_device)
108 
109 static void kvmgt_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
110 		const u8 *val, int len,
111 		struct kvm_page_track_notifier_node *node);
112 static void kvmgt_page_track_flush_slot(struct kvm *kvm,
113 		struct kvm_memory_slot *slot,
114 		struct kvm_page_track_notifier_node *node);
115 
116 static ssize_t description_show(struct mdev_type *mtype,
117 				struct mdev_type_attribute *attr, char *buf)
118 {
119 	struct intel_vgpu_type *type =
120 		container_of(mtype, struct intel_vgpu_type, type);
121 
122 	return sprintf(buf, "low_gm_size: %dMB\nhigh_gm_size: %dMB\n"
123 		       "fence: %d\nresolution: %s\n"
124 		       "weight: %d\n",
125 		       BYTES_TO_MB(type->conf->low_mm),
126 		       BYTES_TO_MB(type->conf->high_mm),
127 		       type->conf->fence, vgpu_edid_str(type->conf->edid),
128 		       type->conf->weight);
129 }
130 
131 static MDEV_TYPE_ATTR_RO(description);
132 
133 static const struct attribute *gvt_type_attrs[] = {
134 	&mdev_type_attr_description.attr,
135 	NULL,
136 };
137 
138 static void gvt_unpin_guest_page(struct intel_vgpu *vgpu, unsigned long gfn,
139 		unsigned long size)
140 {
141 	vfio_unpin_pages(&vgpu->vfio_device, gfn << PAGE_SHIFT,
142 			 DIV_ROUND_UP(size, PAGE_SIZE));
143 }
144 
145 /* Pin a normal or compound guest page for dma. */
146 static int gvt_pin_guest_page(struct intel_vgpu *vgpu, unsigned long gfn,
147 		unsigned long size, struct page **page)
148 {
149 	int total_pages = DIV_ROUND_UP(size, PAGE_SIZE);
150 	struct page *base_page = NULL;
151 	int npage;
152 	int ret;
153 
154 	/*
155 	 * We pin the pages one-by-one to avoid allocating a big arrary
156 	 * on stack to hold pfns.
157 	 */
158 	for (npage = 0; npage < total_pages; npage++) {
159 		dma_addr_t cur_iova = (gfn + npage) << PAGE_SHIFT;
160 		struct page *cur_page;
161 
162 		ret = vfio_pin_pages(&vgpu->vfio_device, cur_iova, 1,
163 				     IOMMU_READ | IOMMU_WRITE, &cur_page);
164 		if (ret != 1) {
165 			gvt_vgpu_err("vfio_pin_pages failed for iova %pad, ret %d\n",
166 				     &cur_iova, ret);
167 			goto err;
168 		}
169 
170 		if (npage == 0)
171 			base_page = cur_page;
172 		else if (base_page + npage != cur_page) {
173 			gvt_vgpu_err("The pages are not continuous\n");
174 			ret = -EINVAL;
175 			npage++;
176 			goto err;
177 		}
178 	}
179 
180 	*page = base_page;
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->gt->i915->drm.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, 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->gt->i915->drm.dev;
214 
215 	dma_unmap_page(dev, dma_addr, size, 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->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->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->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->gfn_cache);
285 
286 	/* dma_addr_cache maps dma addr to struct gvt_dma. */
287 	parent = NULL;
288 	link = &vgpu->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->dma_addr_cache);
300 
301 	vgpu->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->gfn_cache);
309 	rb_erase(&entry->dma_addr_node, &vgpu->dma_addr_cache);
310 	kfree(entry);
311 	vgpu->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->cache_lock);
321 		node = rb_first(&vgpu->gfn_cache);
322 		if (!node) {
323 			mutex_unlock(&vgpu->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->cache_lock);
330 	}
331 }
332 
333 static void gvt_cache_init(struct intel_vgpu *vgpu)
334 {
335 	vgpu->gfn_cache = RB_ROOT;
336 	vgpu->dma_addr_cache = RB_ROOT;
337 	vgpu->nr_cache_entries = 0;
338 	mutex_init(&vgpu->cache_lock);
339 }
340 
341 static void kvmgt_protect_table_init(struct intel_vgpu *info)
342 {
343 	hash_init(info->ptable);
344 }
345 
346 static void kvmgt_protect_table_destroy(struct intel_vgpu *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 intel_vgpu *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 intel_vgpu *info, gfn_t gfn)
374 {
375 	struct kvmgt_pgfn *p;
376 
377 	p = __kvmgt_protect_table_find(info, gfn);
378 	return !!p;
379 }
380 
381 static void kvmgt_protect_table_add(struct intel_vgpu *info, gfn_t gfn)
382 {
383 	struct kvmgt_pgfn *p;
384 
385 	if (kvmgt_gfn_is_write_protected(info, gfn))
386 		return;
387 
388 	p = kzalloc(sizeof(struct kvmgt_pgfn), GFP_ATOMIC);
389 	if (WARN(!p, "gfn: 0x%llx\n", gfn))
390 		return;
391 
392 	p->gfn = gfn;
393 	hash_add(info->ptable, &p->hnode, gfn);
394 }
395 
396 static void kvmgt_protect_table_del(struct intel_vgpu *info, gfn_t gfn)
397 {
398 	struct kvmgt_pgfn *p;
399 
400 	p = __kvmgt_protect_table_find(info, gfn);
401 	if (p) {
402 		hash_del(&p->hnode);
403 		kfree(p);
404 	}
405 }
406 
407 static size_t intel_vgpu_reg_rw_opregion(struct intel_vgpu *vgpu, char *buf,
408 		size_t count, loff_t *ppos, bool iswrite)
409 {
410 	unsigned int i = VFIO_PCI_OFFSET_TO_INDEX(*ppos) -
411 			VFIO_PCI_NUM_REGIONS;
412 	void *base = vgpu->region[i].data;
413 	loff_t pos = *ppos & VFIO_PCI_OFFSET_MASK;
414 
415 
416 	if (pos >= vgpu->region[i].size || iswrite) {
417 		gvt_vgpu_err("invalid op or offset for Intel vgpu OpRegion\n");
418 		return -EINVAL;
419 	}
420 	count = min(count, (size_t)(vgpu->region[i].size - pos));
421 	memcpy(buf, base + pos, count);
422 
423 	return count;
424 }
425 
426 static void intel_vgpu_reg_release_opregion(struct intel_vgpu *vgpu,
427 		struct vfio_region *region)
428 {
429 }
430 
431 static const struct intel_vgpu_regops intel_vgpu_regops_opregion = {
432 	.rw = intel_vgpu_reg_rw_opregion,
433 	.release = intel_vgpu_reg_release_opregion,
434 };
435 
436 static int handle_edid_regs(struct intel_vgpu *vgpu,
437 			struct vfio_edid_region *region, char *buf,
438 			size_t count, u16 offset, bool is_write)
439 {
440 	struct vfio_region_gfx_edid *regs = &region->vfio_edid_regs;
441 	unsigned int data;
442 
443 	if (offset + count > sizeof(*regs))
444 		return -EINVAL;
445 
446 	if (count != 4)
447 		return -EINVAL;
448 
449 	if (is_write) {
450 		data = *((unsigned int *)buf);
451 		switch (offset) {
452 		case offsetof(struct vfio_region_gfx_edid, link_state):
453 			if (data == VFIO_DEVICE_GFX_LINK_STATE_UP) {
454 				if (!drm_edid_block_valid(
455 					(u8 *)region->edid_blob,
456 					0,
457 					true,
458 					NULL)) {
459 					gvt_vgpu_err("invalid EDID blob\n");
460 					return -EINVAL;
461 				}
462 				intel_vgpu_emulate_hotplug(vgpu, true);
463 			} else if (data == VFIO_DEVICE_GFX_LINK_STATE_DOWN)
464 				intel_vgpu_emulate_hotplug(vgpu, false);
465 			else {
466 				gvt_vgpu_err("invalid EDID link state %d\n",
467 					regs->link_state);
468 				return -EINVAL;
469 			}
470 			regs->link_state = data;
471 			break;
472 		case offsetof(struct vfio_region_gfx_edid, edid_size):
473 			if (data > regs->edid_max_size) {
474 				gvt_vgpu_err("EDID size is bigger than %d!\n",
475 					regs->edid_max_size);
476 				return -EINVAL;
477 			}
478 			regs->edid_size = data;
479 			break;
480 		default:
481 			/* read-only regs */
482 			gvt_vgpu_err("write read-only EDID region at offset %d\n",
483 				offset);
484 			return -EPERM;
485 		}
486 	} else {
487 		memcpy(buf, (char *)regs + offset, count);
488 	}
489 
490 	return count;
491 }
492 
493 static int handle_edid_blob(struct vfio_edid_region *region, char *buf,
494 			size_t count, u16 offset, bool is_write)
495 {
496 	if (offset + count > region->vfio_edid_regs.edid_size)
497 		return -EINVAL;
498 
499 	if (is_write)
500 		memcpy(region->edid_blob + offset, buf, count);
501 	else
502 		memcpy(buf, region->edid_blob + offset, count);
503 
504 	return count;
505 }
506 
507 static size_t intel_vgpu_reg_rw_edid(struct intel_vgpu *vgpu, char *buf,
508 		size_t count, loff_t *ppos, bool iswrite)
509 {
510 	int ret;
511 	unsigned int i = VFIO_PCI_OFFSET_TO_INDEX(*ppos) -
512 			VFIO_PCI_NUM_REGIONS;
513 	struct vfio_edid_region *region = vgpu->region[i].data;
514 	loff_t pos = *ppos & VFIO_PCI_OFFSET_MASK;
515 
516 	if (pos < region->vfio_edid_regs.edid_offset) {
517 		ret = handle_edid_regs(vgpu, region, buf, count, pos, iswrite);
518 	} else {
519 		pos -= EDID_BLOB_OFFSET;
520 		ret = handle_edid_blob(region, buf, count, pos, iswrite);
521 	}
522 
523 	if (ret < 0)
524 		gvt_vgpu_err("failed to access EDID region\n");
525 
526 	return ret;
527 }
528 
529 static void intel_vgpu_reg_release_edid(struct intel_vgpu *vgpu,
530 					struct vfio_region *region)
531 {
532 	kfree(region->data);
533 }
534 
535 static const struct intel_vgpu_regops intel_vgpu_regops_edid = {
536 	.rw = intel_vgpu_reg_rw_edid,
537 	.release = intel_vgpu_reg_release_edid,
538 };
539 
540 static int intel_vgpu_register_reg(struct intel_vgpu *vgpu,
541 		unsigned int type, unsigned int subtype,
542 		const struct intel_vgpu_regops *ops,
543 		size_t size, u32 flags, void *data)
544 {
545 	struct vfio_region *region;
546 
547 	region = krealloc(vgpu->region,
548 			(vgpu->num_regions + 1) * sizeof(*region),
549 			GFP_KERNEL);
550 	if (!region)
551 		return -ENOMEM;
552 
553 	vgpu->region = region;
554 	vgpu->region[vgpu->num_regions].type = type;
555 	vgpu->region[vgpu->num_regions].subtype = subtype;
556 	vgpu->region[vgpu->num_regions].ops = ops;
557 	vgpu->region[vgpu->num_regions].size = size;
558 	vgpu->region[vgpu->num_regions].flags = flags;
559 	vgpu->region[vgpu->num_regions].data = data;
560 	vgpu->num_regions++;
561 	return 0;
562 }
563 
564 int intel_gvt_set_opregion(struct intel_vgpu *vgpu)
565 {
566 	void *base;
567 	int ret;
568 
569 	/* Each vgpu has its own opregion, although VFIO would create another
570 	 * one later. This one is used to expose opregion to VFIO. And the
571 	 * other one created by VFIO later, is used by guest actually.
572 	 */
573 	base = vgpu_opregion(vgpu)->va;
574 	if (!base)
575 		return -ENOMEM;
576 
577 	if (memcmp(base, OPREGION_SIGNATURE, 16)) {
578 		memunmap(base);
579 		return -EINVAL;
580 	}
581 
582 	ret = intel_vgpu_register_reg(vgpu,
583 			PCI_VENDOR_ID_INTEL | VFIO_REGION_TYPE_PCI_VENDOR_TYPE,
584 			VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION,
585 			&intel_vgpu_regops_opregion, OPREGION_SIZE,
586 			VFIO_REGION_INFO_FLAG_READ, base);
587 
588 	return ret;
589 }
590 
591 int intel_gvt_set_edid(struct intel_vgpu *vgpu, int port_num)
592 {
593 	struct intel_vgpu_port *port = intel_vgpu_port(vgpu, port_num);
594 	struct vfio_edid_region *base;
595 	int ret;
596 
597 	base = kzalloc(sizeof(*base), GFP_KERNEL);
598 	if (!base)
599 		return -ENOMEM;
600 
601 	/* TODO: Add multi-port and EDID extension block support */
602 	base->vfio_edid_regs.edid_offset = EDID_BLOB_OFFSET;
603 	base->vfio_edid_regs.edid_max_size = EDID_SIZE;
604 	base->vfio_edid_regs.edid_size = EDID_SIZE;
605 	base->vfio_edid_regs.max_xres = vgpu_edid_xres(port->id);
606 	base->vfio_edid_regs.max_yres = vgpu_edid_yres(port->id);
607 	base->edid_blob = port->edid->edid_block;
608 
609 	ret = intel_vgpu_register_reg(vgpu,
610 			VFIO_REGION_TYPE_GFX,
611 			VFIO_REGION_SUBTYPE_GFX_EDID,
612 			&intel_vgpu_regops_edid, EDID_SIZE,
613 			VFIO_REGION_INFO_FLAG_READ |
614 			VFIO_REGION_INFO_FLAG_WRITE |
615 			VFIO_REGION_INFO_FLAG_CAPS, base);
616 
617 	return ret;
618 }
619 
620 static void intel_vgpu_dma_unmap(struct vfio_device *vfio_dev, u64 iova,
621 				 u64 length)
622 {
623 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
624 	struct gvt_dma *entry;
625 	u64 iov_pfn = iova >> PAGE_SHIFT;
626 	u64 end_iov_pfn = iov_pfn + length / PAGE_SIZE;
627 
628 	mutex_lock(&vgpu->cache_lock);
629 	for (; iov_pfn < end_iov_pfn; iov_pfn++) {
630 		entry = __gvt_cache_find_gfn(vgpu, iov_pfn);
631 		if (!entry)
632 			continue;
633 
634 		gvt_dma_unmap_page(vgpu, entry->gfn, entry->dma_addr,
635 				   entry->size);
636 		__gvt_cache_remove_entry(vgpu, entry);
637 	}
638 	mutex_unlock(&vgpu->cache_lock);
639 }
640 
641 static bool __kvmgt_vgpu_exist(struct intel_vgpu *vgpu)
642 {
643 	struct intel_vgpu *itr;
644 	int id;
645 	bool ret = false;
646 
647 	mutex_lock(&vgpu->gvt->lock);
648 	for_each_active_vgpu(vgpu->gvt, itr, id) {
649 		if (!itr->attached)
650 			continue;
651 
652 		if (vgpu->vfio_device.kvm == itr->vfio_device.kvm) {
653 			ret = true;
654 			goto out;
655 		}
656 	}
657 out:
658 	mutex_unlock(&vgpu->gvt->lock);
659 	return ret;
660 }
661 
662 static int intel_vgpu_open_device(struct vfio_device *vfio_dev)
663 {
664 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
665 
666 	if (vgpu->attached)
667 		return -EEXIST;
668 
669 	if (!vgpu->vfio_device.kvm ||
670 	    vgpu->vfio_device.kvm->mm != current->mm) {
671 		gvt_vgpu_err("KVM is required to use Intel vGPU\n");
672 		return -ESRCH;
673 	}
674 
675 	kvm_get_kvm(vgpu->vfio_device.kvm);
676 
677 	if (__kvmgt_vgpu_exist(vgpu))
678 		return -EEXIST;
679 
680 	vgpu->attached = true;
681 
682 	kvmgt_protect_table_init(vgpu);
683 	gvt_cache_init(vgpu);
684 
685 	vgpu->track_node.track_write = kvmgt_page_track_write;
686 	vgpu->track_node.track_flush_slot = kvmgt_page_track_flush_slot;
687 	kvm_page_track_register_notifier(vgpu->vfio_device.kvm,
688 					 &vgpu->track_node);
689 
690 	debugfs_create_ulong(KVMGT_DEBUGFS_FILENAME, 0444, vgpu->debugfs,
691 			     &vgpu->nr_cache_entries);
692 
693 	intel_gvt_activate_vgpu(vgpu);
694 
695 	atomic_set(&vgpu->released, 0);
696 	return 0;
697 }
698 
699 static void intel_vgpu_release_msi_eventfd_ctx(struct intel_vgpu *vgpu)
700 {
701 	struct eventfd_ctx *trigger;
702 
703 	trigger = vgpu->msi_trigger;
704 	if (trigger) {
705 		eventfd_ctx_put(trigger);
706 		vgpu->msi_trigger = NULL;
707 	}
708 }
709 
710 static void intel_vgpu_close_device(struct vfio_device *vfio_dev)
711 {
712 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
713 
714 	if (!vgpu->attached)
715 		return;
716 
717 	if (atomic_cmpxchg(&vgpu->released, 0, 1))
718 		return;
719 
720 	intel_gvt_release_vgpu(vgpu);
721 
722 	debugfs_remove(debugfs_lookup(KVMGT_DEBUGFS_FILENAME, vgpu->debugfs));
723 
724 	kvm_page_track_unregister_notifier(vgpu->vfio_device.kvm,
725 					   &vgpu->track_node);
726 	kvmgt_protect_table_destroy(vgpu);
727 	gvt_cache_destroy(vgpu);
728 
729 	intel_vgpu_release_msi_eventfd_ctx(vgpu);
730 
731 	vgpu->attached = false;
732 
733 	if (vgpu->vfio_device.kvm)
734 		kvm_put_kvm(vgpu->vfio_device.kvm);
735 }
736 
737 static u64 intel_vgpu_get_bar_addr(struct intel_vgpu *vgpu, int bar)
738 {
739 	u32 start_lo, start_hi;
740 	u32 mem_type;
741 
742 	start_lo = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
743 			PCI_BASE_ADDRESS_MEM_MASK;
744 	mem_type = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
745 			PCI_BASE_ADDRESS_MEM_TYPE_MASK;
746 
747 	switch (mem_type) {
748 	case PCI_BASE_ADDRESS_MEM_TYPE_64:
749 		start_hi = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space
750 						+ bar + 4));
751 		break;
752 	case PCI_BASE_ADDRESS_MEM_TYPE_32:
753 	case PCI_BASE_ADDRESS_MEM_TYPE_1M:
754 		/* 1M mem BAR treated as 32-bit BAR */
755 	default:
756 		/* mem unknown type treated as 32-bit BAR */
757 		start_hi = 0;
758 		break;
759 	}
760 
761 	return ((u64)start_hi << 32) | start_lo;
762 }
763 
764 static int intel_vgpu_bar_rw(struct intel_vgpu *vgpu, int bar, u64 off,
765 			     void *buf, unsigned int count, bool is_write)
766 {
767 	u64 bar_start = intel_vgpu_get_bar_addr(vgpu, bar);
768 	int ret;
769 
770 	if (is_write)
771 		ret = intel_vgpu_emulate_mmio_write(vgpu,
772 					bar_start + off, buf, count);
773 	else
774 		ret = intel_vgpu_emulate_mmio_read(vgpu,
775 					bar_start + off, buf, count);
776 	return ret;
777 }
778 
779 static inline bool intel_vgpu_in_aperture(struct intel_vgpu *vgpu, u64 off)
780 {
781 	return off >= vgpu_aperture_offset(vgpu) &&
782 	       off < vgpu_aperture_offset(vgpu) + vgpu_aperture_sz(vgpu);
783 }
784 
785 static int intel_vgpu_aperture_rw(struct intel_vgpu *vgpu, u64 off,
786 		void *buf, unsigned long count, bool is_write)
787 {
788 	void __iomem *aperture_va;
789 
790 	if (!intel_vgpu_in_aperture(vgpu, off) ||
791 	    !intel_vgpu_in_aperture(vgpu, off + count)) {
792 		gvt_vgpu_err("Invalid aperture offset %llu\n", off);
793 		return -EINVAL;
794 	}
795 
796 	aperture_va = io_mapping_map_wc(&vgpu->gvt->gt->ggtt->iomap,
797 					ALIGN_DOWN(off, PAGE_SIZE),
798 					count + offset_in_page(off));
799 	if (!aperture_va)
800 		return -EIO;
801 
802 	if (is_write)
803 		memcpy_toio(aperture_va + offset_in_page(off), buf, count);
804 	else
805 		memcpy_fromio(buf, aperture_va + offset_in_page(off), count);
806 
807 	io_mapping_unmap(aperture_va);
808 
809 	return 0;
810 }
811 
812 static ssize_t intel_vgpu_rw(struct intel_vgpu *vgpu, char *buf,
813 			size_t count, loff_t *ppos, bool is_write)
814 {
815 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
816 	u64 pos = *ppos & VFIO_PCI_OFFSET_MASK;
817 	int ret = -EINVAL;
818 
819 
820 	if (index >= VFIO_PCI_NUM_REGIONS + vgpu->num_regions) {
821 		gvt_vgpu_err("invalid index: %u\n", index);
822 		return -EINVAL;
823 	}
824 
825 	switch (index) {
826 	case VFIO_PCI_CONFIG_REGION_INDEX:
827 		if (is_write)
828 			ret = intel_vgpu_emulate_cfg_write(vgpu, pos,
829 						buf, count);
830 		else
831 			ret = intel_vgpu_emulate_cfg_read(vgpu, pos,
832 						buf, count);
833 		break;
834 	case VFIO_PCI_BAR0_REGION_INDEX:
835 		ret = intel_vgpu_bar_rw(vgpu, PCI_BASE_ADDRESS_0, pos,
836 					buf, count, is_write);
837 		break;
838 	case VFIO_PCI_BAR2_REGION_INDEX:
839 		ret = intel_vgpu_aperture_rw(vgpu, pos, buf, count, is_write);
840 		break;
841 	case VFIO_PCI_BAR1_REGION_INDEX:
842 	case VFIO_PCI_BAR3_REGION_INDEX:
843 	case VFIO_PCI_BAR4_REGION_INDEX:
844 	case VFIO_PCI_BAR5_REGION_INDEX:
845 	case VFIO_PCI_VGA_REGION_INDEX:
846 	case VFIO_PCI_ROM_REGION_INDEX:
847 		break;
848 	default:
849 		if (index >= VFIO_PCI_NUM_REGIONS + vgpu->num_regions)
850 			return -EINVAL;
851 
852 		index -= VFIO_PCI_NUM_REGIONS;
853 		return vgpu->region[index].ops->rw(vgpu, buf, count,
854 				ppos, is_write);
855 	}
856 
857 	return ret == 0 ? count : ret;
858 }
859 
860 static bool gtt_entry(struct intel_vgpu *vgpu, loff_t *ppos)
861 {
862 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
863 	struct intel_gvt *gvt = vgpu->gvt;
864 	int offset;
865 
866 	/* Only allow MMIO GGTT entry access */
867 	if (index != PCI_BASE_ADDRESS_0)
868 		return false;
869 
870 	offset = (u64)(*ppos & VFIO_PCI_OFFSET_MASK) -
871 		intel_vgpu_get_bar_gpa(vgpu, PCI_BASE_ADDRESS_0);
872 
873 	return (offset >= gvt->device_info.gtt_start_offset &&
874 		offset < gvt->device_info.gtt_start_offset + gvt_ggtt_sz(gvt)) ?
875 			true : false;
876 }
877 
878 static ssize_t intel_vgpu_read(struct vfio_device *vfio_dev, char __user *buf,
879 			size_t count, loff_t *ppos)
880 {
881 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
882 	unsigned int done = 0;
883 	int ret;
884 
885 	while (count) {
886 		size_t filled;
887 
888 		/* Only support GGTT entry 8 bytes read */
889 		if (count >= 8 && !(*ppos % 8) &&
890 			gtt_entry(vgpu, ppos)) {
891 			u64 val;
892 
893 			ret = intel_vgpu_rw(vgpu, (char *)&val, sizeof(val),
894 					ppos, false);
895 			if (ret <= 0)
896 				goto read_err;
897 
898 			if (copy_to_user(buf, &val, sizeof(val)))
899 				goto read_err;
900 
901 			filled = 8;
902 		} else if (count >= 4 && !(*ppos % 4)) {
903 			u32 val;
904 
905 			ret = intel_vgpu_rw(vgpu, (char *)&val, sizeof(val),
906 					ppos, false);
907 			if (ret <= 0)
908 				goto read_err;
909 
910 			if (copy_to_user(buf, &val, sizeof(val)))
911 				goto read_err;
912 
913 			filled = 4;
914 		} else if (count >= 2 && !(*ppos % 2)) {
915 			u16 val;
916 
917 			ret = intel_vgpu_rw(vgpu, (char *)&val, sizeof(val),
918 					ppos, false);
919 			if (ret <= 0)
920 				goto read_err;
921 
922 			if (copy_to_user(buf, &val, sizeof(val)))
923 				goto read_err;
924 
925 			filled = 2;
926 		} else {
927 			u8 val;
928 
929 			ret = intel_vgpu_rw(vgpu, &val, sizeof(val), ppos,
930 					false);
931 			if (ret <= 0)
932 				goto read_err;
933 
934 			if (copy_to_user(buf, &val, sizeof(val)))
935 				goto read_err;
936 
937 			filled = 1;
938 		}
939 
940 		count -= filled;
941 		done += filled;
942 		*ppos += filled;
943 		buf += filled;
944 	}
945 
946 	return done;
947 
948 read_err:
949 	return -EFAULT;
950 }
951 
952 static ssize_t intel_vgpu_write(struct vfio_device *vfio_dev,
953 				const char __user *buf,
954 				size_t count, loff_t *ppos)
955 {
956 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
957 	unsigned int done = 0;
958 	int ret;
959 
960 	while (count) {
961 		size_t filled;
962 
963 		/* Only support GGTT entry 8 bytes write */
964 		if (count >= 8 && !(*ppos % 8) &&
965 			gtt_entry(vgpu, ppos)) {
966 			u64 val;
967 
968 			if (copy_from_user(&val, buf, sizeof(val)))
969 				goto write_err;
970 
971 			ret = intel_vgpu_rw(vgpu, (char *)&val, sizeof(val),
972 					ppos, true);
973 			if (ret <= 0)
974 				goto write_err;
975 
976 			filled = 8;
977 		} else if (count >= 4 && !(*ppos % 4)) {
978 			u32 val;
979 
980 			if (copy_from_user(&val, buf, sizeof(val)))
981 				goto write_err;
982 
983 			ret = intel_vgpu_rw(vgpu, (char *)&val, sizeof(val),
984 					ppos, true);
985 			if (ret <= 0)
986 				goto write_err;
987 
988 			filled = 4;
989 		} else if (count >= 2 && !(*ppos % 2)) {
990 			u16 val;
991 
992 			if (copy_from_user(&val, buf, sizeof(val)))
993 				goto write_err;
994 
995 			ret = intel_vgpu_rw(vgpu, (char *)&val,
996 					sizeof(val), ppos, true);
997 			if (ret <= 0)
998 				goto write_err;
999 
1000 			filled = 2;
1001 		} else {
1002 			u8 val;
1003 
1004 			if (copy_from_user(&val, buf, sizeof(val)))
1005 				goto write_err;
1006 
1007 			ret = intel_vgpu_rw(vgpu, &val, sizeof(val),
1008 					ppos, true);
1009 			if (ret <= 0)
1010 				goto write_err;
1011 
1012 			filled = 1;
1013 		}
1014 
1015 		count -= filled;
1016 		done += filled;
1017 		*ppos += filled;
1018 		buf += filled;
1019 	}
1020 
1021 	return done;
1022 write_err:
1023 	return -EFAULT;
1024 }
1025 
1026 static int intel_vgpu_mmap(struct vfio_device *vfio_dev,
1027 		struct vm_area_struct *vma)
1028 {
1029 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
1030 	unsigned int index;
1031 	u64 virtaddr;
1032 	unsigned long req_size, pgoff, req_start;
1033 	pgprot_t pg_prot;
1034 
1035 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1036 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1037 		return -EINVAL;
1038 
1039 	if (vma->vm_end < vma->vm_start)
1040 		return -EINVAL;
1041 	if ((vma->vm_flags & VM_SHARED) == 0)
1042 		return -EINVAL;
1043 	if (index != VFIO_PCI_BAR2_REGION_INDEX)
1044 		return -EINVAL;
1045 
1046 	pg_prot = vma->vm_page_prot;
1047 	virtaddr = vma->vm_start;
1048 	req_size = vma->vm_end - vma->vm_start;
1049 	pgoff = vma->vm_pgoff &
1050 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1051 	req_start = pgoff << PAGE_SHIFT;
1052 
1053 	if (!intel_vgpu_in_aperture(vgpu, req_start))
1054 		return -EINVAL;
1055 	if (req_start + req_size >
1056 	    vgpu_aperture_offset(vgpu) + vgpu_aperture_sz(vgpu))
1057 		return -EINVAL;
1058 
1059 	pgoff = (gvt_aperture_pa_base(vgpu->gvt) >> PAGE_SHIFT) + pgoff;
1060 
1061 	return remap_pfn_range(vma, virtaddr, pgoff, req_size, pg_prot);
1062 }
1063 
1064 static int intel_vgpu_get_irq_count(struct intel_vgpu *vgpu, int type)
1065 {
1066 	if (type == VFIO_PCI_INTX_IRQ_INDEX || type == VFIO_PCI_MSI_IRQ_INDEX)
1067 		return 1;
1068 
1069 	return 0;
1070 }
1071 
1072 static int intel_vgpu_set_intx_mask(struct intel_vgpu *vgpu,
1073 			unsigned int index, unsigned int start,
1074 			unsigned int count, u32 flags,
1075 			void *data)
1076 {
1077 	return 0;
1078 }
1079 
1080 static int intel_vgpu_set_intx_unmask(struct intel_vgpu *vgpu,
1081 			unsigned int index, unsigned int start,
1082 			unsigned int count, u32 flags, void *data)
1083 {
1084 	return 0;
1085 }
1086 
1087 static int intel_vgpu_set_intx_trigger(struct intel_vgpu *vgpu,
1088 		unsigned int index, unsigned int start, unsigned int count,
1089 		u32 flags, void *data)
1090 {
1091 	return 0;
1092 }
1093 
1094 static int intel_vgpu_set_msi_trigger(struct intel_vgpu *vgpu,
1095 		unsigned int index, unsigned int start, unsigned int count,
1096 		u32 flags, void *data)
1097 {
1098 	struct eventfd_ctx *trigger;
1099 
1100 	if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
1101 		int fd = *(int *)data;
1102 
1103 		trigger = eventfd_ctx_fdget(fd);
1104 		if (IS_ERR(trigger)) {
1105 			gvt_vgpu_err("eventfd_ctx_fdget failed\n");
1106 			return PTR_ERR(trigger);
1107 		}
1108 		vgpu->msi_trigger = trigger;
1109 	} else if ((flags & VFIO_IRQ_SET_DATA_NONE) && !count)
1110 		intel_vgpu_release_msi_eventfd_ctx(vgpu);
1111 
1112 	return 0;
1113 }
1114 
1115 static int intel_vgpu_set_irqs(struct intel_vgpu *vgpu, u32 flags,
1116 		unsigned int index, unsigned int start, unsigned int count,
1117 		void *data)
1118 {
1119 	int (*func)(struct intel_vgpu *vgpu, unsigned int index,
1120 			unsigned int start, unsigned int count, u32 flags,
1121 			void *data) = NULL;
1122 
1123 	switch (index) {
1124 	case VFIO_PCI_INTX_IRQ_INDEX:
1125 		switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
1126 		case VFIO_IRQ_SET_ACTION_MASK:
1127 			func = intel_vgpu_set_intx_mask;
1128 			break;
1129 		case VFIO_IRQ_SET_ACTION_UNMASK:
1130 			func = intel_vgpu_set_intx_unmask;
1131 			break;
1132 		case VFIO_IRQ_SET_ACTION_TRIGGER:
1133 			func = intel_vgpu_set_intx_trigger;
1134 			break;
1135 		}
1136 		break;
1137 	case VFIO_PCI_MSI_IRQ_INDEX:
1138 		switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
1139 		case VFIO_IRQ_SET_ACTION_MASK:
1140 		case VFIO_IRQ_SET_ACTION_UNMASK:
1141 			/* XXX Need masking support exported */
1142 			break;
1143 		case VFIO_IRQ_SET_ACTION_TRIGGER:
1144 			func = intel_vgpu_set_msi_trigger;
1145 			break;
1146 		}
1147 		break;
1148 	}
1149 
1150 	if (!func)
1151 		return -ENOTTY;
1152 
1153 	return func(vgpu, index, start, count, flags, data);
1154 }
1155 
1156 static long intel_vgpu_ioctl(struct vfio_device *vfio_dev, unsigned int cmd,
1157 			     unsigned long arg)
1158 {
1159 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
1160 	unsigned long minsz;
1161 
1162 	gvt_dbg_core("vgpu%d ioctl, cmd: %d\n", vgpu->id, cmd);
1163 
1164 	if (cmd == VFIO_DEVICE_GET_INFO) {
1165 		struct vfio_device_info info;
1166 
1167 		minsz = offsetofend(struct vfio_device_info, num_irqs);
1168 
1169 		if (copy_from_user(&info, (void __user *)arg, minsz))
1170 			return -EFAULT;
1171 
1172 		if (info.argsz < minsz)
1173 			return -EINVAL;
1174 
1175 		info.flags = VFIO_DEVICE_FLAGS_PCI;
1176 		info.flags |= VFIO_DEVICE_FLAGS_RESET;
1177 		info.num_regions = VFIO_PCI_NUM_REGIONS +
1178 				vgpu->num_regions;
1179 		info.num_irqs = VFIO_PCI_NUM_IRQS;
1180 
1181 		return copy_to_user((void __user *)arg, &info, minsz) ?
1182 			-EFAULT : 0;
1183 
1184 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
1185 		struct vfio_region_info info;
1186 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
1187 		unsigned int i;
1188 		int ret;
1189 		struct vfio_region_info_cap_sparse_mmap *sparse = NULL;
1190 		int nr_areas = 1;
1191 		int cap_type_id;
1192 
1193 		minsz = offsetofend(struct vfio_region_info, offset);
1194 
1195 		if (copy_from_user(&info, (void __user *)arg, minsz))
1196 			return -EFAULT;
1197 
1198 		if (info.argsz < minsz)
1199 			return -EINVAL;
1200 
1201 		switch (info.index) {
1202 		case VFIO_PCI_CONFIG_REGION_INDEX:
1203 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1204 			info.size = vgpu->gvt->device_info.cfg_space_size;
1205 			info.flags = VFIO_REGION_INFO_FLAG_READ |
1206 				     VFIO_REGION_INFO_FLAG_WRITE;
1207 			break;
1208 		case VFIO_PCI_BAR0_REGION_INDEX:
1209 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1210 			info.size = vgpu->cfg_space.bar[info.index].size;
1211 			if (!info.size) {
1212 				info.flags = 0;
1213 				break;
1214 			}
1215 
1216 			info.flags = VFIO_REGION_INFO_FLAG_READ |
1217 				     VFIO_REGION_INFO_FLAG_WRITE;
1218 			break;
1219 		case VFIO_PCI_BAR1_REGION_INDEX:
1220 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1221 			info.size = 0;
1222 			info.flags = 0;
1223 			break;
1224 		case VFIO_PCI_BAR2_REGION_INDEX:
1225 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1226 			info.flags = VFIO_REGION_INFO_FLAG_CAPS |
1227 					VFIO_REGION_INFO_FLAG_MMAP |
1228 					VFIO_REGION_INFO_FLAG_READ |
1229 					VFIO_REGION_INFO_FLAG_WRITE;
1230 			info.size = gvt_aperture_sz(vgpu->gvt);
1231 
1232 			sparse = kzalloc(struct_size(sparse, areas, nr_areas),
1233 					 GFP_KERNEL);
1234 			if (!sparse)
1235 				return -ENOMEM;
1236 
1237 			sparse->header.id = VFIO_REGION_INFO_CAP_SPARSE_MMAP;
1238 			sparse->header.version = 1;
1239 			sparse->nr_areas = nr_areas;
1240 			cap_type_id = VFIO_REGION_INFO_CAP_SPARSE_MMAP;
1241 			sparse->areas[0].offset =
1242 					PAGE_ALIGN(vgpu_aperture_offset(vgpu));
1243 			sparse->areas[0].size = vgpu_aperture_sz(vgpu);
1244 			break;
1245 
1246 		case VFIO_PCI_BAR3_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1247 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1248 			info.size = 0;
1249 			info.flags = 0;
1250 
1251 			gvt_dbg_core("get region info bar:%d\n", info.index);
1252 			break;
1253 
1254 		case VFIO_PCI_ROM_REGION_INDEX:
1255 		case VFIO_PCI_VGA_REGION_INDEX:
1256 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1257 			info.size = 0;
1258 			info.flags = 0;
1259 
1260 			gvt_dbg_core("get region info index:%d\n", info.index);
1261 			break;
1262 		default:
1263 			{
1264 				struct vfio_region_info_cap_type cap_type = {
1265 					.header.id = VFIO_REGION_INFO_CAP_TYPE,
1266 					.header.version = 1 };
1267 
1268 				if (info.index >= VFIO_PCI_NUM_REGIONS +
1269 						vgpu->num_regions)
1270 					return -EINVAL;
1271 				info.index =
1272 					array_index_nospec(info.index,
1273 							VFIO_PCI_NUM_REGIONS +
1274 							vgpu->num_regions);
1275 
1276 				i = info.index - VFIO_PCI_NUM_REGIONS;
1277 
1278 				info.offset =
1279 					VFIO_PCI_INDEX_TO_OFFSET(info.index);
1280 				info.size = vgpu->region[i].size;
1281 				info.flags = vgpu->region[i].flags;
1282 
1283 				cap_type.type = vgpu->region[i].type;
1284 				cap_type.subtype = vgpu->region[i].subtype;
1285 
1286 				ret = vfio_info_add_capability(&caps,
1287 							&cap_type.header,
1288 							sizeof(cap_type));
1289 				if (ret)
1290 					return ret;
1291 			}
1292 		}
1293 
1294 		if ((info.flags & VFIO_REGION_INFO_FLAG_CAPS) && sparse) {
1295 			switch (cap_type_id) {
1296 			case VFIO_REGION_INFO_CAP_SPARSE_MMAP:
1297 				ret = vfio_info_add_capability(&caps,
1298 					&sparse->header,
1299 					struct_size(sparse, areas,
1300 						    sparse->nr_areas));
1301 				if (ret) {
1302 					kfree(sparse);
1303 					return ret;
1304 				}
1305 				break;
1306 			default:
1307 				kfree(sparse);
1308 				return -EINVAL;
1309 			}
1310 		}
1311 
1312 		if (caps.size) {
1313 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
1314 			if (info.argsz < sizeof(info) + caps.size) {
1315 				info.argsz = sizeof(info) + caps.size;
1316 				info.cap_offset = 0;
1317 			} else {
1318 				vfio_info_cap_shift(&caps, sizeof(info));
1319 				if (copy_to_user((void __user *)arg +
1320 						  sizeof(info), caps.buf,
1321 						  caps.size)) {
1322 					kfree(caps.buf);
1323 					kfree(sparse);
1324 					return -EFAULT;
1325 				}
1326 				info.cap_offset = sizeof(info);
1327 			}
1328 
1329 			kfree(caps.buf);
1330 		}
1331 
1332 		kfree(sparse);
1333 		return copy_to_user((void __user *)arg, &info, minsz) ?
1334 			-EFAULT : 0;
1335 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
1336 		struct vfio_irq_info info;
1337 
1338 		minsz = offsetofend(struct vfio_irq_info, count);
1339 
1340 		if (copy_from_user(&info, (void __user *)arg, minsz))
1341 			return -EFAULT;
1342 
1343 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
1344 			return -EINVAL;
1345 
1346 		switch (info.index) {
1347 		case VFIO_PCI_INTX_IRQ_INDEX:
1348 		case VFIO_PCI_MSI_IRQ_INDEX:
1349 			break;
1350 		default:
1351 			return -EINVAL;
1352 		}
1353 
1354 		info.flags = VFIO_IRQ_INFO_EVENTFD;
1355 
1356 		info.count = intel_vgpu_get_irq_count(vgpu, info.index);
1357 
1358 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1359 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
1360 				       VFIO_IRQ_INFO_AUTOMASKED);
1361 		else
1362 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
1363 
1364 		return copy_to_user((void __user *)arg, &info, minsz) ?
1365 			-EFAULT : 0;
1366 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
1367 		struct vfio_irq_set hdr;
1368 		u8 *data = NULL;
1369 		int ret = 0;
1370 		size_t data_size = 0;
1371 
1372 		minsz = offsetofend(struct vfio_irq_set, count);
1373 
1374 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1375 			return -EFAULT;
1376 
1377 		if (!(hdr.flags & VFIO_IRQ_SET_DATA_NONE)) {
1378 			int max = intel_vgpu_get_irq_count(vgpu, hdr.index);
1379 
1380 			ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
1381 						VFIO_PCI_NUM_IRQS, &data_size);
1382 			if (ret) {
1383 				gvt_vgpu_err("intel:vfio_set_irqs_validate_and_prepare failed\n");
1384 				return -EINVAL;
1385 			}
1386 			if (data_size) {
1387 				data = memdup_user((void __user *)(arg + minsz),
1388 						   data_size);
1389 				if (IS_ERR(data))
1390 					return PTR_ERR(data);
1391 			}
1392 		}
1393 
1394 		ret = intel_vgpu_set_irqs(vgpu, hdr.flags, hdr.index,
1395 					hdr.start, hdr.count, data);
1396 		kfree(data);
1397 
1398 		return ret;
1399 	} else if (cmd == VFIO_DEVICE_RESET) {
1400 		intel_gvt_reset_vgpu(vgpu);
1401 		return 0;
1402 	} else if (cmd == VFIO_DEVICE_QUERY_GFX_PLANE) {
1403 		struct vfio_device_gfx_plane_info dmabuf;
1404 		int ret = 0;
1405 
1406 		minsz = offsetofend(struct vfio_device_gfx_plane_info,
1407 				    dmabuf_id);
1408 		if (copy_from_user(&dmabuf, (void __user *)arg, minsz))
1409 			return -EFAULT;
1410 		if (dmabuf.argsz < minsz)
1411 			return -EINVAL;
1412 
1413 		ret = intel_vgpu_query_plane(vgpu, &dmabuf);
1414 		if (ret != 0)
1415 			return ret;
1416 
1417 		return copy_to_user((void __user *)arg, &dmabuf, minsz) ?
1418 								-EFAULT : 0;
1419 	} else if (cmd == VFIO_DEVICE_GET_GFX_DMABUF) {
1420 		__u32 dmabuf_id;
1421 
1422 		if (get_user(dmabuf_id, (__u32 __user *)arg))
1423 			return -EFAULT;
1424 		return intel_vgpu_get_dmabuf(vgpu, dmabuf_id);
1425 	}
1426 
1427 	return -ENOTTY;
1428 }
1429 
1430 static ssize_t
1431 vgpu_id_show(struct device *dev, struct device_attribute *attr,
1432 	     char *buf)
1433 {
1434 	struct intel_vgpu *vgpu = dev_get_drvdata(dev);
1435 
1436 	return sprintf(buf, "%d\n", vgpu->id);
1437 }
1438 
1439 static DEVICE_ATTR_RO(vgpu_id);
1440 
1441 static struct attribute *intel_vgpu_attrs[] = {
1442 	&dev_attr_vgpu_id.attr,
1443 	NULL
1444 };
1445 
1446 static const struct attribute_group intel_vgpu_group = {
1447 	.name = "intel_vgpu",
1448 	.attrs = intel_vgpu_attrs,
1449 };
1450 
1451 static const struct attribute_group *intel_vgpu_groups[] = {
1452 	&intel_vgpu_group,
1453 	NULL,
1454 };
1455 
1456 static int intel_vgpu_init_dev(struct vfio_device *vfio_dev)
1457 {
1458 	struct mdev_device *mdev = to_mdev_device(vfio_dev->dev);
1459 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
1460 	struct intel_vgpu_type *type =
1461 		container_of(mdev->type, struct intel_vgpu_type, type);
1462 
1463 	vgpu->gvt = kdev_to_i915(mdev->type->parent->dev)->gvt;
1464 	return intel_gvt_create_vgpu(vgpu, type->conf);
1465 }
1466 
1467 static void intel_vgpu_release_dev(struct vfio_device *vfio_dev)
1468 {
1469 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
1470 
1471 	intel_gvt_destroy_vgpu(vgpu);
1472 	vfio_free_device(vfio_dev);
1473 }
1474 
1475 static const struct vfio_device_ops intel_vgpu_dev_ops = {
1476 	.init		= intel_vgpu_init_dev,
1477 	.release	= intel_vgpu_release_dev,
1478 	.open_device	= intel_vgpu_open_device,
1479 	.close_device	= intel_vgpu_close_device,
1480 	.read		= intel_vgpu_read,
1481 	.write		= intel_vgpu_write,
1482 	.mmap		= intel_vgpu_mmap,
1483 	.ioctl		= intel_vgpu_ioctl,
1484 	.dma_unmap	= intel_vgpu_dma_unmap,
1485 };
1486 
1487 static int intel_vgpu_probe(struct mdev_device *mdev)
1488 {
1489 	struct intel_vgpu *vgpu;
1490 	int ret;
1491 
1492 	vgpu = vfio_alloc_device(intel_vgpu, vfio_device, &mdev->dev,
1493 				 &intel_vgpu_dev_ops);
1494 	if (IS_ERR(vgpu)) {
1495 		gvt_err("failed to create intel vgpu: %ld\n", PTR_ERR(vgpu));
1496 		return PTR_ERR(vgpu);
1497 	}
1498 
1499 	dev_set_drvdata(&mdev->dev, vgpu);
1500 	ret = vfio_register_emulated_iommu_dev(&vgpu->vfio_device);
1501 	if (ret)
1502 		goto out_put_vdev;
1503 
1504 	gvt_dbg_core("intel_vgpu_create succeeded for mdev: %s\n",
1505 		     dev_name(mdev_dev(mdev)));
1506 	return 0;
1507 
1508 out_put_vdev:
1509 	vfio_put_device(&vgpu->vfio_device);
1510 	return ret;
1511 }
1512 
1513 static void intel_vgpu_remove(struct mdev_device *mdev)
1514 {
1515 	struct intel_vgpu *vgpu = dev_get_drvdata(&mdev->dev);
1516 
1517 	if (WARN_ON_ONCE(vgpu->attached))
1518 		return;
1519 
1520 	vfio_unregister_group_dev(&vgpu->vfio_device);
1521 	vfio_put_device(&vgpu->vfio_device);
1522 }
1523 
1524 static unsigned int intel_vgpu_get_available(struct mdev_type *mtype)
1525 {
1526 	struct intel_vgpu_type *type =
1527 		container_of(mtype, struct intel_vgpu_type, type);
1528 	struct intel_gvt *gvt = kdev_to_i915(mtype->parent->dev)->gvt;
1529 	unsigned int low_gm_avail, high_gm_avail, fence_avail;
1530 
1531 	mutex_lock(&gvt->lock);
1532 	low_gm_avail = gvt_aperture_sz(gvt) - HOST_LOW_GM_SIZE -
1533 		gvt->gm.vgpu_allocated_low_gm_size;
1534 	high_gm_avail = gvt_hidden_sz(gvt) - HOST_HIGH_GM_SIZE -
1535 		gvt->gm.vgpu_allocated_high_gm_size;
1536 	fence_avail = gvt_fence_sz(gvt) - HOST_FENCE -
1537 		gvt->fence.vgpu_allocated_fence_num;
1538 	mutex_unlock(&gvt->lock);
1539 
1540 	return min3(low_gm_avail / type->conf->low_mm,
1541 		    high_gm_avail / type->conf->high_mm,
1542 		    fence_avail / type->conf->fence);
1543 }
1544 
1545 static struct mdev_driver intel_vgpu_mdev_driver = {
1546 	.device_api	= VFIO_DEVICE_API_PCI_STRING,
1547 	.driver = {
1548 		.name		= "intel_vgpu_mdev",
1549 		.owner		= THIS_MODULE,
1550 		.dev_groups	= intel_vgpu_groups,
1551 	},
1552 	.probe		= intel_vgpu_probe,
1553 	.remove		= intel_vgpu_remove,
1554 	.get_available	= intel_vgpu_get_available,
1555 	.types_attrs	= gvt_type_attrs,
1556 };
1557 
1558 int intel_gvt_page_track_add(struct intel_vgpu *info, u64 gfn)
1559 {
1560 	struct kvm *kvm = info->vfio_device.kvm;
1561 	struct kvm_memory_slot *slot;
1562 	int idx;
1563 
1564 	if (!info->attached)
1565 		return -ESRCH;
1566 
1567 	idx = srcu_read_lock(&kvm->srcu);
1568 	slot = gfn_to_memslot(kvm, gfn);
1569 	if (!slot) {
1570 		srcu_read_unlock(&kvm->srcu, idx);
1571 		return -EINVAL;
1572 	}
1573 
1574 	write_lock(&kvm->mmu_lock);
1575 
1576 	if (kvmgt_gfn_is_write_protected(info, gfn))
1577 		goto out;
1578 
1579 	kvm_slot_page_track_add_page(kvm, slot, gfn, KVM_PAGE_TRACK_WRITE);
1580 	kvmgt_protect_table_add(info, gfn);
1581 
1582 out:
1583 	write_unlock(&kvm->mmu_lock);
1584 	srcu_read_unlock(&kvm->srcu, idx);
1585 	return 0;
1586 }
1587 
1588 int intel_gvt_page_track_remove(struct intel_vgpu *info, u64 gfn)
1589 {
1590 	struct kvm *kvm = info->vfio_device.kvm;
1591 	struct kvm_memory_slot *slot;
1592 	int idx;
1593 
1594 	if (!info->attached)
1595 		return 0;
1596 
1597 	idx = srcu_read_lock(&kvm->srcu);
1598 	slot = gfn_to_memslot(kvm, gfn);
1599 	if (!slot) {
1600 		srcu_read_unlock(&kvm->srcu, idx);
1601 		return -EINVAL;
1602 	}
1603 
1604 	write_lock(&kvm->mmu_lock);
1605 
1606 	if (!kvmgt_gfn_is_write_protected(info, gfn))
1607 		goto out;
1608 
1609 	kvm_slot_page_track_remove_page(kvm, slot, gfn, KVM_PAGE_TRACK_WRITE);
1610 	kvmgt_protect_table_del(info, gfn);
1611 
1612 out:
1613 	write_unlock(&kvm->mmu_lock);
1614 	srcu_read_unlock(&kvm->srcu, idx);
1615 	return 0;
1616 }
1617 
1618 static void kvmgt_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
1619 		const u8 *val, int len,
1620 		struct kvm_page_track_notifier_node *node)
1621 {
1622 	struct intel_vgpu *info =
1623 		container_of(node, struct intel_vgpu, track_node);
1624 
1625 	if (kvmgt_gfn_is_write_protected(info, gpa_to_gfn(gpa)))
1626 		intel_vgpu_page_track_handler(info, gpa,
1627 						     (void *)val, len);
1628 }
1629 
1630 static void kvmgt_page_track_flush_slot(struct kvm *kvm,
1631 		struct kvm_memory_slot *slot,
1632 		struct kvm_page_track_notifier_node *node)
1633 {
1634 	int i;
1635 	gfn_t gfn;
1636 	struct intel_vgpu *info =
1637 		container_of(node, struct intel_vgpu, track_node);
1638 
1639 	write_lock(&kvm->mmu_lock);
1640 	for (i = 0; i < slot->npages; i++) {
1641 		gfn = slot->base_gfn + i;
1642 		if (kvmgt_gfn_is_write_protected(info, gfn)) {
1643 			kvm_slot_page_track_remove_page(kvm, slot, gfn,
1644 						KVM_PAGE_TRACK_WRITE);
1645 			kvmgt_protect_table_del(info, gfn);
1646 		}
1647 	}
1648 	write_unlock(&kvm->mmu_lock);
1649 }
1650 
1651 void intel_vgpu_detach_regions(struct intel_vgpu *vgpu)
1652 {
1653 	int i;
1654 
1655 	if (!vgpu->region)
1656 		return;
1657 
1658 	for (i = 0; i < vgpu->num_regions; i++)
1659 		if (vgpu->region[i].ops->release)
1660 			vgpu->region[i].ops->release(vgpu,
1661 					&vgpu->region[i]);
1662 	vgpu->num_regions = 0;
1663 	kfree(vgpu->region);
1664 	vgpu->region = NULL;
1665 }
1666 
1667 int intel_gvt_dma_map_guest_page(struct intel_vgpu *vgpu, unsigned long gfn,
1668 		unsigned long size, dma_addr_t *dma_addr)
1669 {
1670 	struct gvt_dma *entry;
1671 	int ret;
1672 
1673 	if (!vgpu->attached)
1674 		return -EINVAL;
1675 
1676 	mutex_lock(&vgpu->cache_lock);
1677 
1678 	entry = __gvt_cache_find_gfn(vgpu, gfn);
1679 	if (!entry) {
1680 		ret = gvt_dma_map_page(vgpu, gfn, dma_addr, size);
1681 		if (ret)
1682 			goto err_unlock;
1683 
1684 		ret = __gvt_cache_add(vgpu, gfn, *dma_addr, size);
1685 		if (ret)
1686 			goto err_unmap;
1687 	} else if (entry->size != size) {
1688 		/* the same gfn with different size: unmap and re-map */
1689 		gvt_dma_unmap_page(vgpu, gfn, entry->dma_addr, entry->size);
1690 		__gvt_cache_remove_entry(vgpu, entry);
1691 
1692 		ret = gvt_dma_map_page(vgpu, gfn, dma_addr, size);
1693 		if (ret)
1694 			goto err_unlock;
1695 
1696 		ret = __gvt_cache_add(vgpu, gfn, *dma_addr, size);
1697 		if (ret)
1698 			goto err_unmap;
1699 	} else {
1700 		kref_get(&entry->ref);
1701 		*dma_addr = entry->dma_addr;
1702 	}
1703 
1704 	mutex_unlock(&vgpu->cache_lock);
1705 	return 0;
1706 
1707 err_unmap:
1708 	gvt_dma_unmap_page(vgpu, gfn, *dma_addr, size);
1709 err_unlock:
1710 	mutex_unlock(&vgpu->cache_lock);
1711 	return ret;
1712 }
1713 
1714 int intel_gvt_dma_pin_guest_page(struct intel_vgpu *vgpu, dma_addr_t dma_addr)
1715 {
1716 	struct gvt_dma *entry;
1717 	int ret = 0;
1718 
1719 	if (!vgpu->attached)
1720 		return -ENODEV;
1721 
1722 	mutex_lock(&vgpu->cache_lock);
1723 	entry = __gvt_cache_find_dma_addr(vgpu, dma_addr);
1724 	if (entry)
1725 		kref_get(&entry->ref);
1726 	else
1727 		ret = -ENOMEM;
1728 	mutex_unlock(&vgpu->cache_lock);
1729 
1730 	return ret;
1731 }
1732 
1733 static void __gvt_dma_release(struct kref *ref)
1734 {
1735 	struct gvt_dma *entry = container_of(ref, typeof(*entry), ref);
1736 
1737 	gvt_dma_unmap_page(entry->vgpu, entry->gfn, entry->dma_addr,
1738 			   entry->size);
1739 	__gvt_cache_remove_entry(entry->vgpu, entry);
1740 }
1741 
1742 void intel_gvt_dma_unmap_guest_page(struct intel_vgpu *vgpu,
1743 		dma_addr_t dma_addr)
1744 {
1745 	struct gvt_dma *entry;
1746 
1747 	if (!vgpu->attached)
1748 		return;
1749 
1750 	mutex_lock(&vgpu->cache_lock);
1751 	entry = __gvt_cache_find_dma_addr(vgpu, dma_addr);
1752 	if (entry)
1753 		kref_put(&entry->ref, __gvt_dma_release);
1754 	mutex_unlock(&vgpu->cache_lock);
1755 }
1756 
1757 static void init_device_info(struct intel_gvt *gvt)
1758 {
1759 	struct intel_gvt_device_info *info = &gvt->device_info;
1760 	struct pci_dev *pdev = to_pci_dev(gvt->gt->i915->drm.dev);
1761 
1762 	info->max_support_vgpus = 8;
1763 	info->cfg_space_size = PCI_CFG_SPACE_EXP_SIZE;
1764 	info->mmio_size = 2 * 1024 * 1024;
1765 	info->mmio_bar = 0;
1766 	info->gtt_start_offset = 8 * 1024 * 1024;
1767 	info->gtt_entry_size = 8;
1768 	info->gtt_entry_size_shift = 3;
1769 	info->gmadr_bytes_in_cmd = 8;
1770 	info->max_surface_size = 36 * 1024 * 1024;
1771 	info->msi_cap_offset = pdev->msi_cap;
1772 }
1773 
1774 static void intel_gvt_test_and_emulate_vblank(struct intel_gvt *gvt)
1775 {
1776 	struct intel_vgpu *vgpu;
1777 	int id;
1778 
1779 	mutex_lock(&gvt->lock);
1780 	idr_for_each_entry((&(gvt)->vgpu_idr), (vgpu), (id)) {
1781 		if (test_and_clear_bit(INTEL_GVT_REQUEST_EMULATE_VBLANK + id,
1782 				       (void *)&gvt->service_request)) {
1783 			if (vgpu->active)
1784 				intel_vgpu_emulate_vblank(vgpu);
1785 		}
1786 	}
1787 	mutex_unlock(&gvt->lock);
1788 }
1789 
1790 static int gvt_service_thread(void *data)
1791 {
1792 	struct intel_gvt *gvt = (struct intel_gvt *)data;
1793 	int ret;
1794 
1795 	gvt_dbg_core("service thread start\n");
1796 
1797 	while (!kthread_should_stop()) {
1798 		ret = wait_event_interruptible(gvt->service_thread_wq,
1799 				kthread_should_stop() || gvt->service_request);
1800 
1801 		if (kthread_should_stop())
1802 			break;
1803 
1804 		if (WARN_ONCE(ret, "service thread is waken up by signal.\n"))
1805 			continue;
1806 
1807 		intel_gvt_test_and_emulate_vblank(gvt);
1808 
1809 		if (test_bit(INTEL_GVT_REQUEST_SCHED,
1810 				(void *)&gvt->service_request) ||
1811 			test_bit(INTEL_GVT_REQUEST_EVENT_SCHED,
1812 					(void *)&gvt->service_request)) {
1813 			intel_gvt_schedule(gvt);
1814 		}
1815 	}
1816 
1817 	return 0;
1818 }
1819 
1820 static void clean_service_thread(struct intel_gvt *gvt)
1821 {
1822 	kthread_stop(gvt->service_thread);
1823 }
1824 
1825 static int init_service_thread(struct intel_gvt *gvt)
1826 {
1827 	init_waitqueue_head(&gvt->service_thread_wq);
1828 
1829 	gvt->service_thread = kthread_run(gvt_service_thread,
1830 			gvt, "gvt_service_thread");
1831 	if (IS_ERR(gvt->service_thread)) {
1832 		gvt_err("fail to start service thread.\n");
1833 		return PTR_ERR(gvt->service_thread);
1834 	}
1835 	return 0;
1836 }
1837 
1838 /**
1839  * intel_gvt_clean_device - clean a GVT device
1840  * @i915: i915 private
1841  *
1842  * This function is called at the driver unloading stage, to free the
1843  * resources owned by a GVT device.
1844  *
1845  */
1846 static void intel_gvt_clean_device(struct drm_i915_private *i915)
1847 {
1848 	struct intel_gvt *gvt = fetch_and_zero(&i915->gvt);
1849 
1850 	if (drm_WARN_ON(&i915->drm, !gvt))
1851 		return;
1852 
1853 	mdev_unregister_parent(&gvt->parent);
1854 	intel_gvt_destroy_idle_vgpu(gvt->idle_vgpu);
1855 	intel_gvt_clean_vgpu_types(gvt);
1856 
1857 	intel_gvt_debugfs_clean(gvt);
1858 	clean_service_thread(gvt);
1859 	intel_gvt_clean_cmd_parser(gvt);
1860 	intel_gvt_clean_sched_policy(gvt);
1861 	intel_gvt_clean_workload_scheduler(gvt);
1862 	intel_gvt_clean_gtt(gvt);
1863 	intel_gvt_free_firmware(gvt);
1864 	intel_gvt_clean_mmio_info(gvt);
1865 	idr_destroy(&gvt->vgpu_idr);
1866 
1867 	kfree(i915->gvt);
1868 }
1869 
1870 /**
1871  * intel_gvt_init_device - initialize a GVT device
1872  * @i915: drm i915 private data
1873  *
1874  * This function is called at the initialization stage, to initialize
1875  * necessary GVT components.
1876  *
1877  * Returns:
1878  * Zero on success, negative error code if failed.
1879  *
1880  */
1881 static int intel_gvt_init_device(struct drm_i915_private *i915)
1882 {
1883 	struct intel_gvt *gvt;
1884 	struct intel_vgpu *vgpu;
1885 	int ret;
1886 
1887 	if (drm_WARN_ON(&i915->drm, i915->gvt))
1888 		return -EEXIST;
1889 
1890 	gvt = kzalloc(sizeof(struct intel_gvt), GFP_KERNEL);
1891 	if (!gvt)
1892 		return -ENOMEM;
1893 
1894 	gvt_dbg_core("init gvt device\n");
1895 
1896 	idr_init_base(&gvt->vgpu_idr, 1);
1897 	spin_lock_init(&gvt->scheduler.mmio_context_lock);
1898 	mutex_init(&gvt->lock);
1899 	mutex_init(&gvt->sched_lock);
1900 	gvt->gt = to_gt(i915);
1901 	i915->gvt = gvt;
1902 
1903 	init_device_info(gvt);
1904 
1905 	ret = intel_gvt_setup_mmio_info(gvt);
1906 	if (ret)
1907 		goto out_clean_idr;
1908 
1909 	intel_gvt_init_engine_mmio_context(gvt);
1910 
1911 	ret = intel_gvt_load_firmware(gvt);
1912 	if (ret)
1913 		goto out_clean_mmio_info;
1914 
1915 	ret = intel_gvt_init_irq(gvt);
1916 	if (ret)
1917 		goto out_free_firmware;
1918 
1919 	ret = intel_gvt_init_gtt(gvt);
1920 	if (ret)
1921 		goto out_free_firmware;
1922 
1923 	ret = intel_gvt_init_workload_scheduler(gvt);
1924 	if (ret)
1925 		goto out_clean_gtt;
1926 
1927 	ret = intel_gvt_init_sched_policy(gvt);
1928 	if (ret)
1929 		goto out_clean_workload_scheduler;
1930 
1931 	ret = intel_gvt_init_cmd_parser(gvt);
1932 	if (ret)
1933 		goto out_clean_sched_policy;
1934 
1935 	ret = init_service_thread(gvt);
1936 	if (ret)
1937 		goto out_clean_cmd_parser;
1938 
1939 	ret = intel_gvt_init_vgpu_types(gvt);
1940 	if (ret)
1941 		goto out_clean_thread;
1942 
1943 	vgpu = intel_gvt_create_idle_vgpu(gvt);
1944 	if (IS_ERR(vgpu)) {
1945 		ret = PTR_ERR(vgpu);
1946 		gvt_err("failed to create idle vgpu\n");
1947 		goto out_clean_types;
1948 	}
1949 	gvt->idle_vgpu = vgpu;
1950 
1951 	intel_gvt_debugfs_init(gvt);
1952 
1953 	ret = mdev_register_parent(&gvt->parent, i915->drm.dev,
1954 				   &intel_vgpu_mdev_driver,
1955 				   gvt->mdev_types, gvt->num_types);
1956 	if (ret)
1957 		goto out_destroy_idle_vgpu;
1958 
1959 	gvt_dbg_core("gvt device initialization is done\n");
1960 	return 0;
1961 
1962 out_destroy_idle_vgpu:
1963 	intel_gvt_destroy_idle_vgpu(gvt->idle_vgpu);
1964 	intel_gvt_debugfs_clean(gvt);
1965 out_clean_types:
1966 	intel_gvt_clean_vgpu_types(gvt);
1967 out_clean_thread:
1968 	clean_service_thread(gvt);
1969 out_clean_cmd_parser:
1970 	intel_gvt_clean_cmd_parser(gvt);
1971 out_clean_sched_policy:
1972 	intel_gvt_clean_sched_policy(gvt);
1973 out_clean_workload_scheduler:
1974 	intel_gvt_clean_workload_scheduler(gvt);
1975 out_clean_gtt:
1976 	intel_gvt_clean_gtt(gvt);
1977 out_free_firmware:
1978 	intel_gvt_free_firmware(gvt);
1979 out_clean_mmio_info:
1980 	intel_gvt_clean_mmio_info(gvt);
1981 out_clean_idr:
1982 	idr_destroy(&gvt->vgpu_idr);
1983 	kfree(gvt);
1984 	i915->gvt = NULL;
1985 	return ret;
1986 }
1987 
1988 static void intel_gvt_pm_resume(struct drm_i915_private *i915)
1989 {
1990 	struct intel_gvt *gvt = i915->gvt;
1991 
1992 	intel_gvt_restore_fence(gvt);
1993 	intel_gvt_restore_mmio(gvt);
1994 	intel_gvt_restore_ggtt(gvt);
1995 }
1996 
1997 static const struct intel_vgpu_ops intel_gvt_vgpu_ops = {
1998 	.init_device	= intel_gvt_init_device,
1999 	.clean_device	= intel_gvt_clean_device,
2000 	.pm_resume	= intel_gvt_pm_resume,
2001 };
2002 
2003 static int __init kvmgt_init(void)
2004 {
2005 	int ret;
2006 
2007 	ret = intel_gvt_set_ops(&intel_gvt_vgpu_ops);
2008 	if (ret)
2009 		return ret;
2010 
2011 	ret = mdev_register_driver(&intel_vgpu_mdev_driver);
2012 	if (ret)
2013 		intel_gvt_clear_ops(&intel_gvt_vgpu_ops);
2014 	return ret;
2015 }
2016 
2017 static void __exit kvmgt_exit(void)
2018 {
2019 	mdev_unregister_driver(&intel_vgpu_mdev_driver);
2020 	intel_gvt_clear_ops(&intel_gvt_vgpu_ops);
2021 }
2022 
2023 module_init(kvmgt_init);
2024 module_exit(kvmgt_exit);
2025 
2026 MODULE_LICENSE("GPL and additional rights");
2027 MODULE_AUTHOR("Intel Corporation");
2028