xref: /linux/drivers/gpu/drm/i915/gvt/kvmgt.c (revision a9aaf1ff88a8cb99a1335c9eb76de637f0cf8c10)
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(gpa_t gpa, const u8 *val, int len,
110 				   struct kvm_page_track_notifier_node *node);
111 static void kvmgt_page_track_remove_region(gfn_t gfn, unsigned long nr_pages,
112 					   struct kvm_page_track_notifier_node *node);
113 
114 static ssize_t intel_vgpu_show_description(struct mdev_type *mtype, char *buf)
115 {
116 	struct intel_vgpu_type *type =
117 		container_of(mtype, struct intel_vgpu_type, type);
118 
119 	return sprintf(buf, "low_gm_size: %dMB\nhigh_gm_size: %dMB\n"
120 		       "fence: %d\nresolution: %s\n"
121 		       "weight: %d\n",
122 		       BYTES_TO_MB(type->conf->low_mm),
123 		       BYTES_TO_MB(type->conf->high_mm),
124 		       type->conf->fence, vgpu_edid_str(type->conf->edid),
125 		       type->conf->weight);
126 }
127 
128 static void gvt_unpin_guest_page(struct intel_vgpu *vgpu, unsigned long gfn,
129 		unsigned long size)
130 {
131 	vfio_unpin_pages(&vgpu->vfio_device, gfn << PAGE_SHIFT,
132 			 DIV_ROUND_UP(size, PAGE_SIZE));
133 }
134 
135 /* Pin a normal or compound guest page for dma. */
136 static int gvt_pin_guest_page(struct intel_vgpu *vgpu, unsigned long gfn,
137 		unsigned long size, struct page **page)
138 {
139 	int total_pages = DIV_ROUND_UP(size, PAGE_SIZE);
140 	struct page *base_page = NULL;
141 	int npage;
142 	int ret;
143 
144 	/*
145 	 * We pin the pages one-by-one to avoid allocating a big arrary
146 	 * on stack to hold pfns.
147 	 */
148 	for (npage = 0; npage < total_pages; npage++) {
149 		dma_addr_t cur_iova = (gfn + npage) << PAGE_SHIFT;
150 		struct page *cur_page;
151 
152 		ret = vfio_pin_pages(&vgpu->vfio_device, cur_iova, 1,
153 				     IOMMU_READ | IOMMU_WRITE, &cur_page);
154 		if (ret != 1) {
155 			gvt_vgpu_err("vfio_pin_pages failed for iova %pad, ret %d\n",
156 				     &cur_iova, ret);
157 			goto err;
158 		}
159 
160 		if (npage == 0)
161 			base_page = cur_page;
162 		else if (page_to_pfn(base_page) + npage != page_to_pfn(cur_page)) {
163 			ret = -EINVAL;
164 			npage++;
165 			goto err;
166 		}
167 	}
168 
169 	*page = base_page;
170 	return 0;
171 err:
172 	if (npage)
173 		gvt_unpin_guest_page(vgpu, gfn, npage * PAGE_SIZE);
174 	return ret;
175 }
176 
177 static int gvt_dma_map_page(struct intel_vgpu *vgpu, unsigned long gfn,
178 		dma_addr_t *dma_addr, unsigned long size)
179 {
180 	struct device *dev = vgpu->gvt->gt->i915->drm.dev;
181 	struct page *page = NULL;
182 	int ret;
183 
184 	ret = gvt_pin_guest_page(vgpu, gfn, size, &page);
185 	if (ret)
186 		return ret;
187 
188 	/* Setup DMA mapping. */
189 	*dma_addr = dma_map_page(dev, page, 0, size, DMA_BIDIRECTIONAL);
190 	if (dma_mapping_error(dev, *dma_addr)) {
191 		gvt_vgpu_err("DMA mapping failed for pfn 0x%lx, ret %d\n",
192 			     page_to_pfn(page), ret);
193 		gvt_unpin_guest_page(vgpu, gfn, size);
194 		return -ENOMEM;
195 	}
196 
197 	return 0;
198 }
199 
200 static void gvt_dma_unmap_page(struct intel_vgpu *vgpu, unsigned long gfn,
201 		dma_addr_t dma_addr, unsigned long size)
202 {
203 	struct device *dev = vgpu->gvt->gt->i915->drm.dev;
204 
205 	dma_unmap_page(dev, dma_addr, size, DMA_BIDIRECTIONAL);
206 	gvt_unpin_guest_page(vgpu, gfn, size);
207 }
208 
209 static struct gvt_dma *__gvt_cache_find_dma_addr(struct intel_vgpu *vgpu,
210 		dma_addr_t dma_addr)
211 {
212 	struct rb_node *node = vgpu->dma_addr_cache.rb_node;
213 	struct gvt_dma *itr;
214 
215 	while (node) {
216 		itr = rb_entry(node, struct gvt_dma, dma_addr_node);
217 
218 		if (dma_addr < itr->dma_addr)
219 			node = node->rb_left;
220 		else if (dma_addr > itr->dma_addr)
221 			node = node->rb_right;
222 		else
223 			return itr;
224 	}
225 	return NULL;
226 }
227 
228 static struct gvt_dma *__gvt_cache_find_gfn(struct intel_vgpu *vgpu, gfn_t gfn)
229 {
230 	struct rb_node *node = vgpu->gfn_cache.rb_node;
231 	struct gvt_dma *itr;
232 
233 	while (node) {
234 		itr = rb_entry(node, struct gvt_dma, gfn_node);
235 
236 		if (gfn < itr->gfn)
237 			node = node->rb_left;
238 		else if (gfn > itr->gfn)
239 			node = node->rb_right;
240 		else
241 			return itr;
242 	}
243 	return NULL;
244 }
245 
246 static int __gvt_cache_add(struct intel_vgpu *vgpu, gfn_t gfn,
247 		dma_addr_t dma_addr, unsigned long size)
248 {
249 	struct gvt_dma *new, *itr;
250 	struct rb_node **link, *parent = NULL;
251 
252 	new = kzalloc(sizeof(struct gvt_dma), GFP_KERNEL);
253 	if (!new)
254 		return -ENOMEM;
255 
256 	new->vgpu = vgpu;
257 	new->gfn = gfn;
258 	new->dma_addr = dma_addr;
259 	new->size = size;
260 	kref_init(&new->ref);
261 
262 	/* gfn_cache maps gfn to struct gvt_dma. */
263 	link = &vgpu->gfn_cache.rb_node;
264 	while (*link) {
265 		parent = *link;
266 		itr = rb_entry(parent, struct gvt_dma, gfn_node);
267 
268 		if (gfn < itr->gfn)
269 			link = &parent->rb_left;
270 		else
271 			link = &parent->rb_right;
272 	}
273 	rb_link_node(&new->gfn_node, parent, link);
274 	rb_insert_color(&new->gfn_node, &vgpu->gfn_cache);
275 
276 	/* dma_addr_cache maps dma addr to struct gvt_dma. */
277 	parent = NULL;
278 	link = &vgpu->dma_addr_cache.rb_node;
279 	while (*link) {
280 		parent = *link;
281 		itr = rb_entry(parent, struct gvt_dma, dma_addr_node);
282 
283 		if (dma_addr < itr->dma_addr)
284 			link = &parent->rb_left;
285 		else
286 			link = &parent->rb_right;
287 	}
288 	rb_link_node(&new->dma_addr_node, parent, link);
289 	rb_insert_color(&new->dma_addr_node, &vgpu->dma_addr_cache);
290 
291 	vgpu->nr_cache_entries++;
292 	return 0;
293 }
294 
295 static void __gvt_cache_remove_entry(struct intel_vgpu *vgpu,
296 				struct gvt_dma *entry)
297 {
298 	rb_erase(&entry->gfn_node, &vgpu->gfn_cache);
299 	rb_erase(&entry->dma_addr_node, &vgpu->dma_addr_cache);
300 	kfree(entry);
301 	vgpu->nr_cache_entries--;
302 }
303 
304 static void gvt_cache_destroy(struct intel_vgpu *vgpu)
305 {
306 	struct gvt_dma *dma;
307 	struct rb_node *node = NULL;
308 
309 	for (;;) {
310 		mutex_lock(&vgpu->cache_lock);
311 		node = rb_first(&vgpu->gfn_cache);
312 		if (!node) {
313 			mutex_unlock(&vgpu->cache_lock);
314 			break;
315 		}
316 		dma = rb_entry(node, struct gvt_dma, gfn_node);
317 		gvt_dma_unmap_page(vgpu, dma->gfn, dma->dma_addr, dma->size);
318 		__gvt_cache_remove_entry(vgpu, dma);
319 		mutex_unlock(&vgpu->cache_lock);
320 	}
321 }
322 
323 static void gvt_cache_init(struct intel_vgpu *vgpu)
324 {
325 	vgpu->gfn_cache = RB_ROOT;
326 	vgpu->dma_addr_cache = RB_ROOT;
327 	vgpu->nr_cache_entries = 0;
328 	mutex_init(&vgpu->cache_lock);
329 }
330 
331 static void kvmgt_protect_table_init(struct intel_vgpu *info)
332 {
333 	hash_init(info->ptable);
334 }
335 
336 static void kvmgt_protect_table_destroy(struct intel_vgpu *info)
337 {
338 	struct kvmgt_pgfn *p;
339 	struct hlist_node *tmp;
340 	int i;
341 
342 	hash_for_each_safe(info->ptable, i, tmp, p, hnode) {
343 		hash_del(&p->hnode);
344 		kfree(p);
345 	}
346 }
347 
348 static struct kvmgt_pgfn *
349 __kvmgt_protect_table_find(struct intel_vgpu *info, gfn_t gfn)
350 {
351 	struct kvmgt_pgfn *p, *res = NULL;
352 
353 	lockdep_assert_held(&info->vgpu_lock);
354 
355 	hash_for_each_possible(info->ptable, p, hnode, gfn) {
356 		if (gfn == p->gfn) {
357 			res = p;
358 			break;
359 		}
360 	}
361 
362 	return res;
363 }
364 
365 static bool kvmgt_gfn_is_write_protected(struct intel_vgpu *info, gfn_t gfn)
366 {
367 	struct kvmgt_pgfn *p;
368 
369 	p = __kvmgt_protect_table_find(info, gfn);
370 	return !!p;
371 }
372 
373 static void kvmgt_protect_table_add(struct intel_vgpu *info, gfn_t gfn)
374 {
375 	struct kvmgt_pgfn *p;
376 
377 	if (kvmgt_gfn_is_write_protected(info, gfn))
378 		return;
379 
380 	p = kzalloc(sizeof(struct kvmgt_pgfn), GFP_ATOMIC);
381 	if (WARN(!p, "gfn: 0x%llx\n", gfn))
382 		return;
383 
384 	p->gfn = gfn;
385 	hash_add(info->ptable, &p->hnode, gfn);
386 }
387 
388 static void kvmgt_protect_table_del(struct intel_vgpu *info, gfn_t gfn)
389 {
390 	struct kvmgt_pgfn *p;
391 
392 	p = __kvmgt_protect_table_find(info, gfn);
393 	if (p) {
394 		hash_del(&p->hnode);
395 		kfree(p);
396 	}
397 }
398 
399 static size_t intel_vgpu_reg_rw_opregion(struct intel_vgpu *vgpu, char *buf,
400 		size_t count, loff_t *ppos, bool iswrite)
401 {
402 	unsigned int i = VFIO_PCI_OFFSET_TO_INDEX(*ppos) -
403 			VFIO_PCI_NUM_REGIONS;
404 	void *base = vgpu->region[i].data;
405 	loff_t pos = *ppos & VFIO_PCI_OFFSET_MASK;
406 
407 
408 	if (pos >= vgpu->region[i].size || iswrite) {
409 		gvt_vgpu_err("invalid op or offset for Intel vgpu OpRegion\n");
410 		return -EINVAL;
411 	}
412 	count = min(count, (size_t)(vgpu->region[i].size - pos));
413 	memcpy(buf, base + pos, count);
414 
415 	return count;
416 }
417 
418 static void intel_vgpu_reg_release_opregion(struct intel_vgpu *vgpu,
419 		struct vfio_region *region)
420 {
421 }
422 
423 static const struct intel_vgpu_regops intel_vgpu_regops_opregion = {
424 	.rw = intel_vgpu_reg_rw_opregion,
425 	.release = intel_vgpu_reg_release_opregion,
426 };
427 
428 static int handle_edid_regs(struct intel_vgpu *vgpu,
429 			struct vfio_edid_region *region, char *buf,
430 			size_t count, u16 offset, bool is_write)
431 {
432 	struct vfio_region_gfx_edid *regs = &region->vfio_edid_regs;
433 	unsigned int data;
434 
435 	if (offset + count > sizeof(*regs))
436 		return -EINVAL;
437 
438 	if (count != 4)
439 		return -EINVAL;
440 
441 	if (is_write) {
442 		data = *((unsigned int *)buf);
443 		switch (offset) {
444 		case offsetof(struct vfio_region_gfx_edid, link_state):
445 			if (data == VFIO_DEVICE_GFX_LINK_STATE_UP) {
446 				if (!drm_edid_block_valid(
447 					(u8 *)region->edid_blob,
448 					0,
449 					true,
450 					NULL)) {
451 					gvt_vgpu_err("invalid EDID blob\n");
452 					return -EINVAL;
453 				}
454 				intel_vgpu_emulate_hotplug(vgpu, true);
455 			} else if (data == VFIO_DEVICE_GFX_LINK_STATE_DOWN)
456 				intel_vgpu_emulate_hotplug(vgpu, false);
457 			else {
458 				gvt_vgpu_err("invalid EDID link state %d\n",
459 					regs->link_state);
460 				return -EINVAL;
461 			}
462 			regs->link_state = data;
463 			break;
464 		case offsetof(struct vfio_region_gfx_edid, edid_size):
465 			if (data > regs->edid_max_size) {
466 				gvt_vgpu_err("EDID size is bigger than %d!\n",
467 					regs->edid_max_size);
468 				return -EINVAL;
469 			}
470 			regs->edid_size = data;
471 			break;
472 		default:
473 			/* read-only regs */
474 			gvt_vgpu_err("write read-only EDID region at offset %d\n",
475 				offset);
476 			return -EPERM;
477 		}
478 	} else {
479 		memcpy(buf, (char *)regs + offset, count);
480 	}
481 
482 	return count;
483 }
484 
485 static int handle_edid_blob(struct vfio_edid_region *region, char *buf,
486 			size_t count, u16 offset, bool is_write)
487 {
488 	if (offset + count > region->vfio_edid_regs.edid_size)
489 		return -EINVAL;
490 
491 	if (is_write)
492 		memcpy(region->edid_blob + offset, buf, count);
493 	else
494 		memcpy(buf, region->edid_blob + offset, count);
495 
496 	return count;
497 }
498 
499 static size_t intel_vgpu_reg_rw_edid(struct intel_vgpu *vgpu, char *buf,
500 		size_t count, loff_t *ppos, bool iswrite)
501 {
502 	int ret;
503 	unsigned int i = VFIO_PCI_OFFSET_TO_INDEX(*ppos) -
504 			VFIO_PCI_NUM_REGIONS;
505 	struct vfio_edid_region *region = vgpu->region[i].data;
506 	loff_t pos = *ppos & VFIO_PCI_OFFSET_MASK;
507 
508 	if (pos < region->vfio_edid_regs.edid_offset) {
509 		ret = handle_edid_regs(vgpu, region, buf, count, pos, iswrite);
510 	} else {
511 		pos -= EDID_BLOB_OFFSET;
512 		ret = handle_edid_blob(region, buf, count, pos, iswrite);
513 	}
514 
515 	if (ret < 0)
516 		gvt_vgpu_err("failed to access EDID region\n");
517 
518 	return ret;
519 }
520 
521 static void intel_vgpu_reg_release_edid(struct intel_vgpu *vgpu,
522 					struct vfio_region *region)
523 {
524 	kfree(region->data);
525 }
526 
527 static const struct intel_vgpu_regops intel_vgpu_regops_edid = {
528 	.rw = intel_vgpu_reg_rw_edid,
529 	.release = intel_vgpu_reg_release_edid,
530 };
531 
532 static int intel_vgpu_register_reg(struct intel_vgpu *vgpu,
533 		unsigned int type, unsigned int subtype,
534 		const struct intel_vgpu_regops *ops,
535 		size_t size, u32 flags, void *data)
536 {
537 	struct vfio_region *region;
538 
539 	region = krealloc(vgpu->region,
540 			(vgpu->num_regions + 1) * sizeof(*region),
541 			GFP_KERNEL);
542 	if (!region)
543 		return -ENOMEM;
544 
545 	vgpu->region = region;
546 	vgpu->region[vgpu->num_regions].type = type;
547 	vgpu->region[vgpu->num_regions].subtype = subtype;
548 	vgpu->region[vgpu->num_regions].ops = ops;
549 	vgpu->region[vgpu->num_regions].size = size;
550 	vgpu->region[vgpu->num_regions].flags = flags;
551 	vgpu->region[vgpu->num_regions].data = data;
552 	vgpu->num_regions++;
553 	return 0;
554 }
555 
556 int intel_gvt_set_opregion(struct intel_vgpu *vgpu)
557 {
558 	void *base;
559 	int ret;
560 
561 	/* Each vgpu has its own opregion, although VFIO would create another
562 	 * one later. This one is used to expose opregion to VFIO. And the
563 	 * other one created by VFIO later, is used by guest actually.
564 	 */
565 	base = vgpu_opregion(vgpu)->va;
566 	if (!base)
567 		return -ENOMEM;
568 
569 	if (memcmp(base, OPREGION_SIGNATURE, 16)) {
570 		memunmap(base);
571 		return -EINVAL;
572 	}
573 
574 	ret = intel_vgpu_register_reg(vgpu,
575 			PCI_VENDOR_ID_INTEL | VFIO_REGION_TYPE_PCI_VENDOR_TYPE,
576 			VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION,
577 			&intel_vgpu_regops_opregion, INTEL_GVT_OPREGION_SIZE,
578 			VFIO_REGION_INFO_FLAG_READ, base);
579 
580 	return ret;
581 }
582 
583 int intel_gvt_set_edid(struct intel_vgpu *vgpu, int port_num)
584 {
585 	struct intel_vgpu_port *port = intel_vgpu_port(vgpu, port_num);
586 	struct vfio_edid_region *base;
587 	int ret;
588 
589 	base = kzalloc(sizeof(*base), GFP_KERNEL);
590 	if (!base)
591 		return -ENOMEM;
592 
593 	/* TODO: Add multi-port and EDID extension block support */
594 	base->vfio_edid_regs.edid_offset = EDID_BLOB_OFFSET;
595 	base->vfio_edid_regs.edid_max_size = EDID_SIZE;
596 	base->vfio_edid_regs.edid_size = EDID_SIZE;
597 	base->vfio_edid_regs.max_xres = vgpu_edid_xres(port->id);
598 	base->vfio_edid_regs.max_yres = vgpu_edid_yres(port->id);
599 	base->edid_blob = port->edid->edid_block;
600 
601 	ret = intel_vgpu_register_reg(vgpu,
602 			VFIO_REGION_TYPE_GFX,
603 			VFIO_REGION_SUBTYPE_GFX_EDID,
604 			&intel_vgpu_regops_edid, EDID_SIZE,
605 			VFIO_REGION_INFO_FLAG_READ |
606 			VFIO_REGION_INFO_FLAG_WRITE |
607 			VFIO_REGION_INFO_FLAG_CAPS, base);
608 
609 	return ret;
610 }
611 
612 static void intel_vgpu_dma_unmap(struct vfio_device *vfio_dev, u64 iova,
613 				 u64 length)
614 {
615 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
616 	struct gvt_dma *entry;
617 	u64 iov_pfn = iova >> PAGE_SHIFT;
618 	u64 end_iov_pfn = iov_pfn + length / PAGE_SIZE;
619 
620 	mutex_lock(&vgpu->cache_lock);
621 	for (; iov_pfn < end_iov_pfn; iov_pfn++) {
622 		entry = __gvt_cache_find_gfn(vgpu, iov_pfn);
623 		if (!entry)
624 			continue;
625 
626 		gvt_dma_unmap_page(vgpu, entry->gfn, entry->dma_addr,
627 				   entry->size);
628 		__gvt_cache_remove_entry(vgpu, entry);
629 	}
630 	mutex_unlock(&vgpu->cache_lock);
631 }
632 
633 static bool __kvmgt_vgpu_exist(struct intel_vgpu *vgpu)
634 {
635 	struct intel_vgpu *itr;
636 	int id;
637 	bool ret = false;
638 
639 	mutex_lock(&vgpu->gvt->lock);
640 	for_each_active_vgpu(vgpu->gvt, itr, id) {
641 		if (!test_bit(INTEL_VGPU_STATUS_ATTACHED, itr->status))
642 			continue;
643 
644 		if (vgpu->vfio_device.kvm == itr->vfio_device.kvm) {
645 			ret = true;
646 			goto out;
647 		}
648 	}
649 out:
650 	mutex_unlock(&vgpu->gvt->lock);
651 	return ret;
652 }
653 
654 static int intel_vgpu_open_device(struct vfio_device *vfio_dev)
655 {
656 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
657 	int ret;
658 
659 	if (__kvmgt_vgpu_exist(vgpu))
660 		return -EEXIST;
661 
662 	vgpu->track_node.track_write = kvmgt_page_track_write;
663 	vgpu->track_node.track_remove_region = kvmgt_page_track_remove_region;
664 	ret = kvm_page_track_register_notifier(vgpu->vfio_device.kvm,
665 					       &vgpu->track_node);
666 	if (ret) {
667 		gvt_vgpu_err("KVM is required to use Intel vGPU\n");
668 		return ret;
669 	}
670 
671 	set_bit(INTEL_VGPU_STATUS_ATTACHED, vgpu->status);
672 
673 	debugfs_create_ulong(KVMGT_DEBUGFS_FILENAME, 0444, vgpu->debugfs,
674 			     &vgpu->nr_cache_entries);
675 
676 	intel_gvt_activate_vgpu(vgpu);
677 
678 	return 0;
679 }
680 
681 static void intel_vgpu_release_msi_eventfd_ctx(struct intel_vgpu *vgpu)
682 {
683 	struct eventfd_ctx *trigger;
684 
685 	trigger = vgpu->msi_trigger;
686 	if (trigger) {
687 		eventfd_ctx_put(trigger);
688 		vgpu->msi_trigger = NULL;
689 	}
690 }
691 
692 static void intel_vgpu_close_device(struct vfio_device *vfio_dev)
693 {
694 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
695 
696 	intel_gvt_release_vgpu(vgpu);
697 
698 	clear_bit(INTEL_VGPU_STATUS_ATTACHED, vgpu->status);
699 
700 	debugfs_lookup_and_remove(KVMGT_DEBUGFS_FILENAME, vgpu->debugfs);
701 
702 	kvm_page_track_unregister_notifier(vgpu->vfio_device.kvm,
703 					   &vgpu->track_node);
704 
705 	kvmgt_protect_table_destroy(vgpu);
706 	gvt_cache_destroy(vgpu);
707 
708 	WARN_ON(vgpu->nr_cache_entries);
709 
710 	vgpu->gfn_cache = RB_ROOT;
711 	vgpu->dma_addr_cache = RB_ROOT;
712 
713 	intel_vgpu_release_msi_eventfd_ctx(vgpu);
714 }
715 
716 static u64 intel_vgpu_get_bar_addr(struct intel_vgpu *vgpu, int bar)
717 {
718 	u32 start_lo, start_hi;
719 	u32 mem_type;
720 
721 	start_lo = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
722 			PCI_BASE_ADDRESS_MEM_MASK;
723 	mem_type = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
724 			PCI_BASE_ADDRESS_MEM_TYPE_MASK;
725 
726 	switch (mem_type) {
727 	case PCI_BASE_ADDRESS_MEM_TYPE_64:
728 		start_hi = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space
729 						+ bar + 4));
730 		break;
731 	case PCI_BASE_ADDRESS_MEM_TYPE_32:
732 	case PCI_BASE_ADDRESS_MEM_TYPE_1M:
733 		/* 1M mem BAR treated as 32-bit BAR */
734 	default:
735 		/* mem unknown type treated as 32-bit BAR */
736 		start_hi = 0;
737 		break;
738 	}
739 
740 	return ((u64)start_hi << 32) | start_lo;
741 }
742 
743 static int intel_vgpu_bar_rw(struct intel_vgpu *vgpu, int bar, u64 off,
744 			     void *buf, unsigned int count, bool is_write)
745 {
746 	u64 bar_start = intel_vgpu_get_bar_addr(vgpu, bar);
747 	int ret;
748 
749 	if (is_write)
750 		ret = intel_vgpu_emulate_mmio_write(vgpu,
751 					bar_start + off, buf, count);
752 	else
753 		ret = intel_vgpu_emulate_mmio_read(vgpu,
754 					bar_start + off, buf, count);
755 	return ret;
756 }
757 
758 static inline bool intel_vgpu_in_aperture(struct intel_vgpu *vgpu, u64 off)
759 {
760 	return off >= vgpu_aperture_offset(vgpu) &&
761 	       off < vgpu_aperture_offset(vgpu) + vgpu_aperture_sz(vgpu);
762 }
763 
764 static int intel_vgpu_aperture_rw(struct intel_vgpu *vgpu, u64 off,
765 		void *buf, unsigned long count, bool is_write)
766 {
767 	void __iomem *aperture_va;
768 
769 	if (!intel_vgpu_in_aperture(vgpu, off) ||
770 	    !intel_vgpu_in_aperture(vgpu, off + count)) {
771 		gvt_vgpu_err("Invalid aperture offset %llu\n", off);
772 		return -EINVAL;
773 	}
774 
775 	aperture_va = io_mapping_map_wc(&vgpu->gvt->gt->ggtt->iomap,
776 					ALIGN_DOWN(off, PAGE_SIZE),
777 					count + offset_in_page(off));
778 	if (!aperture_va)
779 		return -EIO;
780 
781 	if (is_write)
782 		memcpy_toio(aperture_va + offset_in_page(off), buf, count);
783 	else
784 		memcpy_fromio(buf, aperture_va + offset_in_page(off), count);
785 
786 	io_mapping_unmap(aperture_va);
787 
788 	return 0;
789 }
790 
791 static ssize_t intel_vgpu_rw(struct intel_vgpu *vgpu, char *buf,
792 			size_t count, loff_t *ppos, bool is_write)
793 {
794 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
795 	u64 pos = *ppos & VFIO_PCI_OFFSET_MASK;
796 	int ret = -EINVAL;
797 
798 
799 	if (index >= VFIO_PCI_NUM_REGIONS + vgpu->num_regions) {
800 		gvt_vgpu_err("invalid index: %u\n", index);
801 		return -EINVAL;
802 	}
803 
804 	switch (index) {
805 	case VFIO_PCI_CONFIG_REGION_INDEX:
806 		if (is_write)
807 			ret = intel_vgpu_emulate_cfg_write(vgpu, pos,
808 						buf, count);
809 		else
810 			ret = intel_vgpu_emulate_cfg_read(vgpu, pos,
811 						buf, count);
812 		break;
813 	case VFIO_PCI_BAR0_REGION_INDEX:
814 		ret = intel_vgpu_bar_rw(vgpu, PCI_BASE_ADDRESS_0, pos,
815 					buf, count, is_write);
816 		break;
817 	case VFIO_PCI_BAR2_REGION_INDEX:
818 		ret = intel_vgpu_aperture_rw(vgpu, pos, buf, count, is_write);
819 		break;
820 	case VFIO_PCI_BAR1_REGION_INDEX:
821 	case VFIO_PCI_BAR3_REGION_INDEX:
822 	case VFIO_PCI_BAR4_REGION_INDEX:
823 	case VFIO_PCI_BAR5_REGION_INDEX:
824 	case VFIO_PCI_VGA_REGION_INDEX:
825 	case VFIO_PCI_ROM_REGION_INDEX:
826 		break;
827 	default:
828 		if (index >= VFIO_PCI_NUM_REGIONS + vgpu->num_regions)
829 			return -EINVAL;
830 
831 		index -= VFIO_PCI_NUM_REGIONS;
832 		return vgpu->region[index].ops->rw(vgpu, buf, count,
833 				ppos, is_write);
834 	}
835 
836 	return ret == 0 ? count : ret;
837 }
838 
839 static bool gtt_entry(struct intel_vgpu *vgpu, loff_t *ppos)
840 {
841 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
842 	struct intel_gvt *gvt = vgpu->gvt;
843 	int offset;
844 
845 	/* Only allow MMIO GGTT entry access */
846 	if (index != PCI_BASE_ADDRESS_0)
847 		return false;
848 
849 	offset = (u64)(*ppos & VFIO_PCI_OFFSET_MASK) -
850 		intel_vgpu_get_bar_gpa(vgpu, PCI_BASE_ADDRESS_0);
851 
852 	return (offset >= gvt->device_info.gtt_start_offset &&
853 		offset < gvt->device_info.gtt_start_offset + gvt_ggtt_sz(gvt)) ?
854 			true : false;
855 }
856 
857 static ssize_t intel_vgpu_read(struct vfio_device *vfio_dev, char __user *buf,
858 			size_t count, loff_t *ppos)
859 {
860 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
861 	unsigned int done = 0;
862 	int ret;
863 
864 	while (count) {
865 		size_t filled;
866 
867 		/* Only support GGTT entry 8 bytes read */
868 		if (count >= 8 && !(*ppos % 8) &&
869 			gtt_entry(vgpu, ppos)) {
870 			u64 val;
871 
872 			ret = intel_vgpu_rw(vgpu, (char *)&val, sizeof(val),
873 					ppos, false);
874 			if (ret <= 0)
875 				goto read_err;
876 
877 			if (copy_to_user(buf, &val, sizeof(val)))
878 				goto read_err;
879 
880 			filled = 8;
881 		} else if (count >= 4 && !(*ppos % 4)) {
882 			u32 val;
883 
884 			ret = intel_vgpu_rw(vgpu, (char *)&val, sizeof(val),
885 					ppos, false);
886 			if (ret <= 0)
887 				goto read_err;
888 
889 			if (copy_to_user(buf, &val, sizeof(val)))
890 				goto read_err;
891 
892 			filled = 4;
893 		} else if (count >= 2 && !(*ppos % 2)) {
894 			u16 val;
895 
896 			ret = intel_vgpu_rw(vgpu, (char *)&val, sizeof(val),
897 					ppos, false);
898 			if (ret <= 0)
899 				goto read_err;
900 
901 			if (copy_to_user(buf, &val, sizeof(val)))
902 				goto read_err;
903 
904 			filled = 2;
905 		} else {
906 			u8 val;
907 
908 			ret = intel_vgpu_rw(vgpu, &val, sizeof(val), ppos,
909 					false);
910 			if (ret <= 0)
911 				goto read_err;
912 
913 			if (copy_to_user(buf, &val, sizeof(val)))
914 				goto read_err;
915 
916 			filled = 1;
917 		}
918 
919 		count -= filled;
920 		done += filled;
921 		*ppos += filled;
922 		buf += filled;
923 	}
924 
925 	return done;
926 
927 read_err:
928 	return -EFAULT;
929 }
930 
931 static ssize_t intel_vgpu_write(struct vfio_device *vfio_dev,
932 				const char __user *buf,
933 				size_t count, loff_t *ppos)
934 {
935 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
936 	unsigned int done = 0;
937 	int ret;
938 
939 	while (count) {
940 		size_t filled;
941 
942 		/* Only support GGTT entry 8 bytes write */
943 		if (count >= 8 && !(*ppos % 8) &&
944 			gtt_entry(vgpu, ppos)) {
945 			u64 val;
946 
947 			if (copy_from_user(&val, buf, sizeof(val)))
948 				goto write_err;
949 
950 			ret = intel_vgpu_rw(vgpu, (char *)&val, sizeof(val),
951 					ppos, true);
952 			if (ret <= 0)
953 				goto write_err;
954 
955 			filled = 8;
956 		} else if (count >= 4 && !(*ppos % 4)) {
957 			u32 val;
958 
959 			if (copy_from_user(&val, buf, sizeof(val)))
960 				goto write_err;
961 
962 			ret = intel_vgpu_rw(vgpu, (char *)&val, sizeof(val),
963 					ppos, true);
964 			if (ret <= 0)
965 				goto write_err;
966 
967 			filled = 4;
968 		} else if (count >= 2 && !(*ppos % 2)) {
969 			u16 val;
970 
971 			if (copy_from_user(&val, buf, sizeof(val)))
972 				goto write_err;
973 
974 			ret = intel_vgpu_rw(vgpu, (char *)&val,
975 					sizeof(val), ppos, true);
976 			if (ret <= 0)
977 				goto write_err;
978 
979 			filled = 2;
980 		} else {
981 			u8 val;
982 
983 			if (copy_from_user(&val, buf, sizeof(val)))
984 				goto write_err;
985 
986 			ret = intel_vgpu_rw(vgpu, &val, sizeof(val),
987 					ppos, true);
988 			if (ret <= 0)
989 				goto write_err;
990 
991 			filled = 1;
992 		}
993 
994 		count -= filled;
995 		done += filled;
996 		*ppos += filled;
997 		buf += filled;
998 	}
999 
1000 	return done;
1001 write_err:
1002 	return -EFAULT;
1003 }
1004 
1005 static int intel_vgpu_mmap(struct vfio_device *vfio_dev,
1006 		struct vm_area_struct *vma)
1007 {
1008 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
1009 	unsigned int index;
1010 	u64 virtaddr;
1011 	unsigned long req_size, pgoff, req_start;
1012 	pgprot_t pg_prot;
1013 
1014 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1015 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1016 		return -EINVAL;
1017 
1018 	if (vma->vm_end < vma->vm_start)
1019 		return -EINVAL;
1020 	if ((vma->vm_flags & VM_SHARED) == 0)
1021 		return -EINVAL;
1022 	if (index != VFIO_PCI_BAR2_REGION_INDEX)
1023 		return -EINVAL;
1024 
1025 	pg_prot = vma->vm_page_prot;
1026 	virtaddr = vma->vm_start;
1027 	req_size = vma->vm_end - vma->vm_start;
1028 	pgoff = vma->vm_pgoff &
1029 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1030 	req_start = pgoff << PAGE_SHIFT;
1031 
1032 	if (!intel_vgpu_in_aperture(vgpu, req_start))
1033 		return -EINVAL;
1034 	if (req_start + req_size >
1035 	    vgpu_aperture_offset(vgpu) + vgpu_aperture_sz(vgpu))
1036 		return -EINVAL;
1037 
1038 	pgoff = (gvt_aperture_pa_base(vgpu->gvt) >> PAGE_SHIFT) + pgoff;
1039 
1040 	return remap_pfn_range(vma, virtaddr, pgoff, req_size, pg_prot);
1041 }
1042 
1043 static int intel_vgpu_get_irq_count(struct intel_vgpu *vgpu, int type)
1044 {
1045 	if (type == VFIO_PCI_INTX_IRQ_INDEX || type == VFIO_PCI_MSI_IRQ_INDEX)
1046 		return 1;
1047 
1048 	return 0;
1049 }
1050 
1051 static int intel_vgpu_set_intx_mask(struct intel_vgpu *vgpu,
1052 			unsigned int index, unsigned int start,
1053 			unsigned int count, u32 flags,
1054 			void *data)
1055 {
1056 	return 0;
1057 }
1058 
1059 static int intel_vgpu_set_intx_unmask(struct intel_vgpu *vgpu,
1060 			unsigned int index, unsigned int start,
1061 			unsigned int count, u32 flags, void *data)
1062 {
1063 	return 0;
1064 }
1065 
1066 static int intel_vgpu_set_intx_trigger(struct intel_vgpu *vgpu,
1067 		unsigned int index, unsigned int start, unsigned int count,
1068 		u32 flags, void *data)
1069 {
1070 	return 0;
1071 }
1072 
1073 static int intel_vgpu_set_msi_trigger(struct intel_vgpu *vgpu,
1074 		unsigned int index, unsigned int start, unsigned int count,
1075 		u32 flags, void *data)
1076 {
1077 	struct eventfd_ctx *trigger;
1078 
1079 	if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
1080 		int fd = *(int *)data;
1081 
1082 		trigger = eventfd_ctx_fdget(fd);
1083 		if (IS_ERR(trigger)) {
1084 			gvt_vgpu_err("eventfd_ctx_fdget failed\n");
1085 			return PTR_ERR(trigger);
1086 		}
1087 		vgpu->msi_trigger = trigger;
1088 	} else if ((flags & VFIO_IRQ_SET_DATA_NONE) && !count)
1089 		intel_vgpu_release_msi_eventfd_ctx(vgpu);
1090 
1091 	return 0;
1092 }
1093 
1094 static int intel_vgpu_set_irqs(struct intel_vgpu *vgpu, u32 flags,
1095 		unsigned int index, unsigned int start, unsigned int count,
1096 		void *data)
1097 {
1098 	int (*func)(struct intel_vgpu *vgpu, unsigned int index,
1099 			unsigned int start, unsigned int count, u32 flags,
1100 			void *data) = NULL;
1101 
1102 	switch (index) {
1103 	case VFIO_PCI_INTX_IRQ_INDEX:
1104 		switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
1105 		case VFIO_IRQ_SET_ACTION_MASK:
1106 			func = intel_vgpu_set_intx_mask;
1107 			break;
1108 		case VFIO_IRQ_SET_ACTION_UNMASK:
1109 			func = intel_vgpu_set_intx_unmask;
1110 			break;
1111 		case VFIO_IRQ_SET_ACTION_TRIGGER:
1112 			func = intel_vgpu_set_intx_trigger;
1113 			break;
1114 		}
1115 		break;
1116 	case VFIO_PCI_MSI_IRQ_INDEX:
1117 		switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
1118 		case VFIO_IRQ_SET_ACTION_MASK:
1119 		case VFIO_IRQ_SET_ACTION_UNMASK:
1120 			/* XXX Need masking support exported */
1121 			break;
1122 		case VFIO_IRQ_SET_ACTION_TRIGGER:
1123 			func = intel_vgpu_set_msi_trigger;
1124 			break;
1125 		}
1126 		break;
1127 	}
1128 
1129 	if (!func)
1130 		return -ENOTTY;
1131 
1132 	return func(vgpu, index, start, count, flags, data);
1133 }
1134 
1135 static long intel_vgpu_ioctl(struct vfio_device *vfio_dev, unsigned int cmd,
1136 			     unsigned long arg)
1137 {
1138 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
1139 	unsigned long minsz;
1140 
1141 	gvt_dbg_core("vgpu%d ioctl, cmd: %d\n", vgpu->id, cmd);
1142 
1143 	if (cmd == VFIO_DEVICE_GET_INFO) {
1144 		struct vfio_device_info info;
1145 
1146 		minsz = offsetofend(struct vfio_device_info, num_irqs);
1147 
1148 		if (copy_from_user(&info, (void __user *)arg, minsz))
1149 			return -EFAULT;
1150 
1151 		if (info.argsz < minsz)
1152 			return -EINVAL;
1153 
1154 		info.flags = VFIO_DEVICE_FLAGS_PCI;
1155 		info.flags |= VFIO_DEVICE_FLAGS_RESET;
1156 		info.num_regions = VFIO_PCI_NUM_REGIONS +
1157 				vgpu->num_regions;
1158 		info.num_irqs = VFIO_PCI_NUM_IRQS;
1159 
1160 		return copy_to_user((void __user *)arg, &info, minsz) ?
1161 			-EFAULT : 0;
1162 
1163 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
1164 		struct vfio_region_info info;
1165 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
1166 		unsigned int i;
1167 		int ret;
1168 		struct vfio_region_info_cap_sparse_mmap *sparse = NULL;
1169 		int nr_areas = 1;
1170 		int cap_type_id;
1171 
1172 		minsz = offsetofend(struct vfio_region_info, offset);
1173 
1174 		if (copy_from_user(&info, (void __user *)arg, minsz))
1175 			return -EFAULT;
1176 
1177 		if (info.argsz < minsz)
1178 			return -EINVAL;
1179 
1180 		switch (info.index) {
1181 		case VFIO_PCI_CONFIG_REGION_INDEX:
1182 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1183 			info.size = vgpu->gvt->device_info.cfg_space_size;
1184 			info.flags = VFIO_REGION_INFO_FLAG_READ |
1185 				     VFIO_REGION_INFO_FLAG_WRITE;
1186 			break;
1187 		case VFIO_PCI_BAR0_REGION_INDEX:
1188 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1189 			info.size = vgpu->cfg_space.bar[info.index].size;
1190 			if (!info.size) {
1191 				info.flags = 0;
1192 				break;
1193 			}
1194 
1195 			info.flags = VFIO_REGION_INFO_FLAG_READ |
1196 				     VFIO_REGION_INFO_FLAG_WRITE;
1197 			break;
1198 		case VFIO_PCI_BAR1_REGION_INDEX:
1199 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1200 			info.size = 0;
1201 			info.flags = 0;
1202 			break;
1203 		case VFIO_PCI_BAR2_REGION_INDEX:
1204 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1205 			info.flags = VFIO_REGION_INFO_FLAG_CAPS |
1206 					VFIO_REGION_INFO_FLAG_MMAP |
1207 					VFIO_REGION_INFO_FLAG_READ |
1208 					VFIO_REGION_INFO_FLAG_WRITE;
1209 			info.size = gvt_aperture_sz(vgpu->gvt);
1210 
1211 			sparse = kzalloc(struct_size(sparse, areas, nr_areas),
1212 					 GFP_KERNEL);
1213 			if (!sparse)
1214 				return -ENOMEM;
1215 
1216 			sparse->header.id = VFIO_REGION_INFO_CAP_SPARSE_MMAP;
1217 			sparse->header.version = 1;
1218 			sparse->nr_areas = nr_areas;
1219 			cap_type_id = VFIO_REGION_INFO_CAP_SPARSE_MMAP;
1220 			sparse->areas[0].offset =
1221 					PAGE_ALIGN(vgpu_aperture_offset(vgpu));
1222 			sparse->areas[0].size = vgpu_aperture_sz(vgpu);
1223 			break;
1224 
1225 		case VFIO_PCI_BAR3_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1226 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1227 			info.size = 0;
1228 			info.flags = 0;
1229 
1230 			gvt_dbg_core("get region info bar:%d\n", info.index);
1231 			break;
1232 
1233 		case VFIO_PCI_ROM_REGION_INDEX:
1234 		case VFIO_PCI_VGA_REGION_INDEX:
1235 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1236 			info.size = 0;
1237 			info.flags = 0;
1238 
1239 			gvt_dbg_core("get region info index:%d\n", info.index);
1240 			break;
1241 		default:
1242 			{
1243 				struct vfio_region_info_cap_type cap_type = {
1244 					.header.id = VFIO_REGION_INFO_CAP_TYPE,
1245 					.header.version = 1 };
1246 
1247 				if (info.index >= VFIO_PCI_NUM_REGIONS +
1248 						vgpu->num_regions)
1249 					return -EINVAL;
1250 				info.index =
1251 					array_index_nospec(info.index,
1252 							VFIO_PCI_NUM_REGIONS +
1253 							vgpu->num_regions);
1254 
1255 				i = info.index - VFIO_PCI_NUM_REGIONS;
1256 
1257 				info.offset =
1258 					VFIO_PCI_INDEX_TO_OFFSET(info.index);
1259 				info.size = vgpu->region[i].size;
1260 				info.flags = vgpu->region[i].flags;
1261 
1262 				cap_type.type = vgpu->region[i].type;
1263 				cap_type.subtype = vgpu->region[i].subtype;
1264 
1265 				ret = vfio_info_add_capability(&caps,
1266 							&cap_type.header,
1267 							sizeof(cap_type));
1268 				if (ret)
1269 					return ret;
1270 			}
1271 		}
1272 
1273 		if ((info.flags & VFIO_REGION_INFO_FLAG_CAPS) && sparse) {
1274 			switch (cap_type_id) {
1275 			case VFIO_REGION_INFO_CAP_SPARSE_MMAP:
1276 				ret = vfio_info_add_capability(&caps,
1277 					&sparse->header,
1278 					struct_size(sparse, areas,
1279 						    sparse->nr_areas));
1280 				if (ret) {
1281 					kfree(sparse);
1282 					return ret;
1283 				}
1284 				break;
1285 			default:
1286 				kfree(sparse);
1287 				return -EINVAL;
1288 			}
1289 		}
1290 
1291 		if (caps.size) {
1292 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
1293 			if (info.argsz < sizeof(info) + caps.size) {
1294 				info.argsz = sizeof(info) + caps.size;
1295 				info.cap_offset = 0;
1296 			} else {
1297 				vfio_info_cap_shift(&caps, sizeof(info));
1298 				if (copy_to_user((void __user *)arg +
1299 						  sizeof(info), caps.buf,
1300 						  caps.size)) {
1301 					kfree(caps.buf);
1302 					kfree(sparse);
1303 					return -EFAULT;
1304 				}
1305 				info.cap_offset = sizeof(info);
1306 			}
1307 
1308 			kfree(caps.buf);
1309 		}
1310 
1311 		kfree(sparse);
1312 		return copy_to_user((void __user *)arg, &info, minsz) ?
1313 			-EFAULT : 0;
1314 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
1315 		struct vfio_irq_info info;
1316 
1317 		minsz = offsetofend(struct vfio_irq_info, count);
1318 
1319 		if (copy_from_user(&info, (void __user *)arg, minsz))
1320 			return -EFAULT;
1321 
1322 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
1323 			return -EINVAL;
1324 
1325 		switch (info.index) {
1326 		case VFIO_PCI_INTX_IRQ_INDEX:
1327 		case VFIO_PCI_MSI_IRQ_INDEX:
1328 			break;
1329 		default:
1330 			return -EINVAL;
1331 		}
1332 
1333 		info.flags = VFIO_IRQ_INFO_EVENTFD;
1334 
1335 		info.count = intel_vgpu_get_irq_count(vgpu, info.index);
1336 
1337 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1338 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
1339 				       VFIO_IRQ_INFO_AUTOMASKED);
1340 		else
1341 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
1342 
1343 		return copy_to_user((void __user *)arg, &info, minsz) ?
1344 			-EFAULT : 0;
1345 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
1346 		struct vfio_irq_set hdr;
1347 		u8 *data = NULL;
1348 		int ret = 0;
1349 		size_t data_size = 0;
1350 
1351 		minsz = offsetofend(struct vfio_irq_set, count);
1352 
1353 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1354 			return -EFAULT;
1355 
1356 		if (!(hdr.flags & VFIO_IRQ_SET_DATA_NONE)) {
1357 			int max = intel_vgpu_get_irq_count(vgpu, hdr.index);
1358 
1359 			ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
1360 						VFIO_PCI_NUM_IRQS, &data_size);
1361 			if (ret) {
1362 				gvt_vgpu_err("intel:vfio_set_irqs_validate_and_prepare failed\n");
1363 				return -EINVAL;
1364 			}
1365 			if (data_size) {
1366 				data = memdup_user((void __user *)(arg + minsz),
1367 						   data_size);
1368 				if (IS_ERR(data))
1369 					return PTR_ERR(data);
1370 			}
1371 		}
1372 
1373 		ret = intel_vgpu_set_irqs(vgpu, hdr.flags, hdr.index,
1374 					hdr.start, hdr.count, data);
1375 		kfree(data);
1376 
1377 		return ret;
1378 	} else if (cmd == VFIO_DEVICE_RESET) {
1379 		intel_gvt_reset_vgpu(vgpu);
1380 		return 0;
1381 	} else if (cmd == VFIO_DEVICE_QUERY_GFX_PLANE) {
1382 		struct vfio_device_gfx_plane_info dmabuf = {};
1383 		int ret = 0;
1384 
1385 		minsz = offsetofend(struct vfio_device_gfx_plane_info,
1386 				    dmabuf_id);
1387 		if (copy_from_user(&dmabuf, (void __user *)arg, minsz))
1388 			return -EFAULT;
1389 		if (dmabuf.argsz < minsz)
1390 			return -EINVAL;
1391 
1392 		ret = intel_vgpu_query_plane(vgpu, &dmabuf);
1393 		if (ret != 0)
1394 			return ret;
1395 
1396 		return copy_to_user((void __user *)arg, &dmabuf, minsz) ?
1397 								-EFAULT : 0;
1398 	} else if (cmd == VFIO_DEVICE_GET_GFX_DMABUF) {
1399 		__u32 dmabuf_id;
1400 
1401 		if (get_user(dmabuf_id, (__u32 __user *)arg))
1402 			return -EFAULT;
1403 		return intel_vgpu_get_dmabuf(vgpu, dmabuf_id);
1404 	}
1405 
1406 	return -ENOTTY;
1407 }
1408 
1409 static ssize_t
1410 vgpu_id_show(struct device *dev, struct device_attribute *attr,
1411 	     char *buf)
1412 {
1413 	struct intel_vgpu *vgpu = dev_get_drvdata(dev);
1414 
1415 	return sprintf(buf, "%d\n", vgpu->id);
1416 }
1417 
1418 static DEVICE_ATTR_RO(vgpu_id);
1419 
1420 static struct attribute *intel_vgpu_attrs[] = {
1421 	&dev_attr_vgpu_id.attr,
1422 	NULL
1423 };
1424 
1425 static const struct attribute_group intel_vgpu_group = {
1426 	.name = "intel_vgpu",
1427 	.attrs = intel_vgpu_attrs,
1428 };
1429 
1430 static const struct attribute_group *intel_vgpu_groups[] = {
1431 	&intel_vgpu_group,
1432 	NULL,
1433 };
1434 
1435 static int intel_vgpu_init_dev(struct vfio_device *vfio_dev)
1436 {
1437 	struct mdev_device *mdev = to_mdev_device(vfio_dev->dev);
1438 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
1439 	struct intel_vgpu_type *type =
1440 		container_of(mdev->type, struct intel_vgpu_type, type);
1441 	int ret;
1442 
1443 	vgpu->gvt = kdev_to_i915(mdev->type->parent->dev)->gvt;
1444 	ret = intel_gvt_create_vgpu(vgpu, type->conf);
1445 	if (ret)
1446 		return ret;
1447 
1448 	kvmgt_protect_table_init(vgpu);
1449 	gvt_cache_init(vgpu);
1450 
1451 	return 0;
1452 }
1453 
1454 static void intel_vgpu_release_dev(struct vfio_device *vfio_dev)
1455 {
1456 	struct intel_vgpu *vgpu = vfio_dev_to_vgpu(vfio_dev);
1457 
1458 	intel_gvt_destroy_vgpu(vgpu);
1459 }
1460 
1461 static const struct vfio_device_ops intel_vgpu_dev_ops = {
1462 	.init		= intel_vgpu_init_dev,
1463 	.release	= intel_vgpu_release_dev,
1464 	.open_device	= intel_vgpu_open_device,
1465 	.close_device	= intel_vgpu_close_device,
1466 	.read		= intel_vgpu_read,
1467 	.write		= intel_vgpu_write,
1468 	.mmap		= intel_vgpu_mmap,
1469 	.ioctl		= intel_vgpu_ioctl,
1470 	.dma_unmap	= intel_vgpu_dma_unmap,
1471 	.bind_iommufd	= vfio_iommufd_emulated_bind,
1472 	.unbind_iommufd = vfio_iommufd_emulated_unbind,
1473 	.attach_ioas	= vfio_iommufd_emulated_attach_ioas,
1474 	.detach_ioas	= vfio_iommufd_emulated_detach_ioas,
1475 };
1476 
1477 static int intel_vgpu_probe(struct mdev_device *mdev)
1478 {
1479 	struct intel_vgpu *vgpu;
1480 	int ret;
1481 
1482 	vgpu = vfio_alloc_device(intel_vgpu, vfio_device, &mdev->dev,
1483 				 &intel_vgpu_dev_ops);
1484 	if (IS_ERR(vgpu)) {
1485 		gvt_err("failed to create intel vgpu: %ld\n", PTR_ERR(vgpu));
1486 		return PTR_ERR(vgpu);
1487 	}
1488 
1489 	dev_set_drvdata(&mdev->dev, vgpu);
1490 	ret = vfio_register_emulated_iommu_dev(&vgpu->vfio_device);
1491 	if (ret)
1492 		goto out_put_vdev;
1493 
1494 	gvt_dbg_core("intel_vgpu_create succeeded for mdev: %s\n",
1495 		     dev_name(mdev_dev(mdev)));
1496 	return 0;
1497 
1498 out_put_vdev:
1499 	vfio_put_device(&vgpu->vfio_device);
1500 	return ret;
1501 }
1502 
1503 static void intel_vgpu_remove(struct mdev_device *mdev)
1504 {
1505 	struct intel_vgpu *vgpu = dev_get_drvdata(&mdev->dev);
1506 
1507 	vfio_unregister_group_dev(&vgpu->vfio_device);
1508 	vfio_put_device(&vgpu->vfio_device);
1509 }
1510 
1511 static unsigned int intel_vgpu_get_available(struct mdev_type *mtype)
1512 {
1513 	struct intel_vgpu_type *type =
1514 		container_of(mtype, struct intel_vgpu_type, type);
1515 	struct intel_gvt *gvt = kdev_to_i915(mtype->parent->dev)->gvt;
1516 	unsigned int low_gm_avail, high_gm_avail, fence_avail;
1517 
1518 	mutex_lock(&gvt->lock);
1519 	low_gm_avail = gvt_aperture_sz(gvt) - HOST_LOW_GM_SIZE -
1520 		gvt->gm.vgpu_allocated_low_gm_size;
1521 	high_gm_avail = gvt_hidden_sz(gvt) - HOST_HIGH_GM_SIZE -
1522 		gvt->gm.vgpu_allocated_high_gm_size;
1523 	fence_avail = gvt_fence_sz(gvt) - HOST_FENCE -
1524 		gvt->fence.vgpu_allocated_fence_num;
1525 	mutex_unlock(&gvt->lock);
1526 
1527 	return min3(low_gm_avail / type->conf->low_mm,
1528 		    high_gm_avail / type->conf->high_mm,
1529 		    fence_avail / type->conf->fence);
1530 }
1531 
1532 static struct mdev_driver intel_vgpu_mdev_driver = {
1533 	.device_api	= VFIO_DEVICE_API_PCI_STRING,
1534 	.driver = {
1535 		.name		= "intel_vgpu_mdev",
1536 		.owner		= THIS_MODULE,
1537 		.dev_groups	= intel_vgpu_groups,
1538 	},
1539 	.probe			= intel_vgpu_probe,
1540 	.remove			= intel_vgpu_remove,
1541 	.get_available		= intel_vgpu_get_available,
1542 	.show_description	= intel_vgpu_show_description,
1543 };
1544 
1545 int intel_gvt_page_track_add(struct intel_vgpu *info, u64 gfn)
1546 {
1547 	int r;
1548 
1549 	if (!test_bit(INTEL_VGPU_STATUS_ATTACHED, info->status))
1550 		return -ESRCH;
1551 
1552 	if (kvmgt_gfn_is_write_protected(info, gfn))
1553 		return 0;
1554 
1555 	r = kvm_write_track_add_gfn(info->vfio_device.kvm, gfn);
1556 	if (r)
1557 		return r;
1558 
1559 	kvmgt_protect_table_add(info, gfn);
1560 	return 0;
1561 }
1562 
1563 int intel_gvt_page_track_remove(struct intel_vgpu *info, u64 gfn)
1564 {
1565 	int r;
1566 
1567 	if (!test_bit(INTEL_VGPU_STATUS_ATTACHED, info->status))
1568 		return -ESRCH;
1569 
1570 	if (!kvmgt_gfn_is_write_protected(info, gfn))
1571 		return 0;
1572 
1573 	r = kvm_write_track_remove_gfn(info->vfio_device.kvm, gfn);
1574 	if (r)
1575 		return r;
1576 
1577 	kvmgt_protect_table_del(info, gfn);
1578 	return 0;
1579 }
1580 
1581 static void kvmgt_page_track_write(gpa_t gpa, const u8 *val, int len,
1582 				   struct kvm_page_track_notifier_node *node)
1583 {
1584 	struct intel_vgpu *info =
1585 		container_of(node, struct intel_vgpu, track_node);
1586 
1587 	mutex_lock(&info->vgpu_lock);
1588 
1589 	if (kvmgt_gfn_is_write_protected(info, gpa >> PAGE_SHIFT))
1590 		intel_vgpu_page_track_handler(info, gpa,
1591 						     (void *)val, len);
1592 
1593 	mutex_unlock(&info->vgpu_lock);
1594 }
1595 
1596 static void kvmgt_page_track_remove_region(gfn_t gfn, unsigned long nr_pages,
1597 					   struct kvm_page_track_notifier_node *node)
1598 {
1599 	unsigned long i;
1600 	struct intel_vgpu *info =
1601 		container_of(node, struct intel_vgpu, track_node);
1602 
1603 	mutex_lock(&info->vgpu_lock);
1604 
1605 	for (i = 0; i < nr_pages; i++) {
1606 		if (kvmgt_gfn_is_write_protected(info, gfn + i))
1607 			kvmgt_protect_table_del(info, gfn + i);
1608 	}
1609 
1610 	mutex_unlock(&info->vgpu_lock);
1611 }
1612 
1613 void intel_vgpu_detach_regions(struct intel_vgpu *vgpu)
1614 {
1615 	int i;
1616 
1617 	if (!vgpu->region)
1618 		return;
1619 
1620 	for (i = 0; i < vgpu->num_regions; i++)
1621 		if (vgpu->region[i].ops->release)
1622 			vgpu->region[i].ops->release(vgpu,
1623 					&vgpu->region[i]);
1624 	vgpu->num_regions = 0;
1625 	kfree(vgpu->region);
1626 	vgpu->region = NULL;
1627 }
1628 
1629 int intel_gvt_dma_map_guest_page(struct intel_vgpu *vgpu, unsigned long gfn,
1630 		unsigned long size, dma_addr_t *dma_addr)
1631 {
1632 	struct gvt_dma *entry;
1633 	int ret;
1634 
1635 	if (!test_bit(INTEL_VGPU_STATUS_ATTACHED, vgpu->status))
1636 		return -EINVAL;
1637 
1638 	mutex_lock(&vgpu->cache_lock);
1639 
1640 	entry = __gvt_cache_find_gfn(vgpu, gfn);
1641 	if (!entry) {
1642 		ret = gvt_dma_map_page(vgpu, gfn, dma_addr, size);
1643 		if (ret)
1644 			goto err_unlock;
1645 
1646 		ret = __gvt_cache_add(vgpu, gfn, *dma_addr, size);
1647 		if (ret)
1648 			goto err_unmap;
1649 	} else if (entry->size != size) {
1650 		/* the same gfn with different size: unmap and re-map */
1651 		gvt_dma_unmap_page(vgpu, gfn, entry->dma_addr, entry->size);
1652 		__gvt_cache_remove_entry(vgpu, entry);
1653 
1654 		ret = gvt_dma_map_page(vgpu, gfn, dma_addr, size);
1655 		if (ret)
1656 			goto err_unlock;
1657 
1658 		ret = __gvt_cache_add(vgpu, gfn, *dma_addr, size);
1659 		if (ret)
1660 			goto err_unmap;
1661 	} else {
1662 		kref_get(&entry->ref);
1663 		*dma_addr = entry->dma_addr;
1664 	}
1665 
1666 	mutex_unlock(&vgpu->cache_lock);
1667 	return 0;
1668 
1669 err_unmap:
1670 	gvt_dma_unmap_page(vgpu, gfn, *dma_addr, size);
1671 err_unlock:
1672 	mutex_unlock(&vgpu->cache_lock);
1673 	return ret;
1674 }
1675 
1676 int intel_gvt_dma_pin_guest_page(struct intel_vgpu *vgpu, dma_addr_t dma_addr)
1677 {
1678 	struct gvt_dma *entry;
1679 	int ret = 0;
1680 
1681 	if (!test_bit(INTEL_VGPU_STATUS_ATTACHED, vgpu->status))
1682 		return -EINVAL;
1683 
1684 	mutex_lock(&vgpu->cache_lock);
1685 	entry = __gvt_cache_find_dma_addr(vgpu, dma_addr);
1686 	if (entry)
1687 		kref_get(&entry->ref);
1688 	else
1689 		ret = -ENOMEM;
1690 	mutex_unlock(&vgpu->cache_lock);
1691 
1692 	return ret;
1693 }
1694 
1695 static void __gvt_dma_release(struct kref *ref)
1696 {
1697 	struct gvt_dma *entry = container_of(ref, typeof(*entry), ref);
1698 
1699 	gvt_dma_unmap_page(entry->vgpu, entry->gfn, entry->dma_addr,
1700 			   entry->size);
1701 	__gvt_cache_remove_entry(entry->vgpu, entry);
1702 }
1703 
1704 void intel_gvt_dma_unmap_guest_page(struct intel_vgpu *vgpu,
1705 		dma_addr_t dma_addr)
1706 {
1707 	struct gvt_dma *entry;
1708 
1709 	if (!test_bit(INTEL_VGPU_STATUS_ATTACHED, vgpu->status))
1710 		return;
1711 
1712 	mutex_lock(&vgpu->cache_lock);
1713 	entry = __gvt_cache_find_dma_addr(vgpu, dma_addr);
1714 	if (entry)
1715 		kref_put(&entry->ref, __gvt_dma_release);
1716 	mutex_unlock(&vgpu->cache_lock);
1717 }
1718 
1719 static void init_device_info(struct intel_gvt *gvt)
1720 {
1721 	struct intel_gvt_device_info *info = &gvt->device_info;
1722 	struct pci_dev *pdev = to_pci_dev(gvt->gt->i915->drm.dev);
1723 
1724 	info->max_support_vgpus = 8;
1725 	info->cfg_space_size = PCI_CFG_SPACE_EXP_SIZE;
1726 	info->mmio_size = 2 * 1024 * 1024;
1727 	info->mmio_bar = 0;
1728 	info->gtt_start_offset = 8 * 1024 * 1024;
1729 	info->gtt_entry_size = 8;
1730 	info->gtt_entry_size_shift = 3;
1731 	info->gmadr_bytes_in_cmd = 8;
1732 	info->max_surface_size = 36 * 1024 * 1024;
1733 	info->msi_cap_offset = pdev->msi_cap;
1734 }
1735 
1736 static void intel_gvt_test_and_emulate_vblank(struct intel_gvt *gvt)
1737 {
1738 	struct intel_vgpu *vgpu;
1739 	int id;
1740 
1741 	mutex_lock(&gvt->lock);
1742 	idr_for_each_entry((&(gvt)->vgpu_idr), (vgpu), (id)) {
1743 		if (test_and_clear_bit(INTEL_GVT_REQUEST_EMULATE_VBLANK + id,
1744 				       (void *)&gvt->service_request)) {
1745 			if (test_bit(INTEL_VGPU_STATUS_ACTIVE, vgpu->status))
1746 				intel_vgpu_emulate_vblank(vgpu);
1747 		}
1748 	}
1749 	mutex_unlock(&gvt->lock);
1750 }
1751 
1752 static int gvt_service_thread(void *data)
1753 {
1754 	struct intel_gvt *gvt = (struct intel_gvt *)data;
1755 	int ret;
1756 
1757 	gvt_dbg_core("service thread start\n");
1758 
1759 	while (!kthread_should_stop()) {
1760 		ret = wait_event_interruptible(gvt->service_thread_wq,
1761 				kthread_should_stop() || gvt->service_request);
1762 
1763 		if (kthread_should_stop())
1764 			break;
1765 
1766 		if (WARN_ONCE(ret, "service thread is waken up by signal.\n"))
1767 			continue;
1768 
1769 		intel_gvt_test_and_emulate_vblank(gvt);
1770 
1771 		if (test_bit(INTEL_GVT_REQUEST_SCHED,
1772 				(void *)&gvt->service_request) ||
1773 			test_bit(INTEL_GVT_REQUEST_EVENT_SCHED,
1774 					(void *)&gvt->service_request)) {
1775 			intel_gvt_schedule(gvt);
1776 		}
1777 	}
1778 
1779 	return 0;
1780 }
1781 
1782 static void clean_service_thread(struct intel_gvt *gvt)
1783 {
1784 	kthread_stop(gvt->service_thread);
1785 }
1786 
1787 static int init_service_thread(struct intel_gvt *gvt)
1788 {
1789 	init_waitqueue_head(&gvt->service_thread_wq);
1790 
1791 	gvt->service_thread = kthread_run(gvt_service_thread,
1792 			gvt, "gvt_service_thread");
1793 	if (IS_ERR(gvt->service_thread)) {
1794 		gvt_err("fail to start service thread.\n");
1795 		return PTR_ERR(gvt->service_thread);
1796 	}
1797 	return 0;
1798 }
1799 
1800 /**
1801  * intel_gvt_clean_device - clean a GVT device
1802  * @i915: i915 private
1803  *
1804  * This function is called at the driver unloading stage, to free the
1805  * resources owned by a GVT device.
1806  *
1807  */
1808 static void intel_gvt_clean_device(struct drm_i915_private *i915)
1809 {
1810 	struct intel_gvt *gvt = fetch_and_zero(&i915->gvt);
1811 
1812 	if (drm_WARN_ON(&i915->drm, !gvt))
1813 		return;
1814 
1815 	mdev_unregister_parent(&gvt->parent);
1816 	intel_gvt_destroy_idle_vgpu(gvt->idle_vgpu);
1817 	intel_gvt_clean_vgpu_types(gvt);
1818 
1819 	intel_gvt_debugfs_clean(gvt);
1820 	clean_service_thread(gvt);
1821 	intel_gvt_clean_cmd_parser(gvt);
1822 	intel_gvt_clean_sched_policy(gvt);
1823 	intel_gvt_clean_workload_scheduler(gvt);
1824 	intel_gvt_clean_gtt(gvt);
1825 	intel_gvt_free_firmware(gvt);
1826 	intel_gvt_clean_mmio_info(gvt);
1827 	idr_destroy(&gvt->vgpu_idr);
1828 
1829 	kfree(i915->gvt);
1830 }
1831 
1832 /**
1833  * intel_gvt_init_device - initialize a GVT device
1834  * @i915: drm i915 private data
1835  *
1836  * This function is called at the initialization stage, to initialize
1837  * necessary GVT components.
1838  *
1839  * Returns:
1840  * Zero on success, negative error code if failed.
1841  *
1842  */
1843 static int intel_gvt_init_device(struct drm_i915_private *i915)
1844 {
1845 	struct intel_gvt *gvt;
1846 	struct intel_vgpu *vgpu;
1847 	int ret;
1848 
1849 	if (drm_WARN_ON(&i915->drm, i915->gvt))
1850 		return -EEXIST;
1851 
1852 	gvt = kzalloc(sizeof(struct intel_gvt), GFP_KERNEL);
1853 	if (!gvt)
1854 		return -ENOMEM;
1855 
1856 	gvt_dbg_core("init gvt device\n");
1857 
1858 	idr_init_base(&gvt->vgpu_idr, 1);
1859 	spin_lock_init(&gvt->scheduler.mmio_context_lock);
1860 	mutex_init(&gvt->lock);
1861 	mutex_init(&gvt->sched_lock);
1862 	gvt->gt = to_gt(i915);
1863 	i915->gvt = gvt;
1864 
1865 	init_device_info(gvt);
1866 
1867 	ret = intel_gvt_setup_mmio_info(gvt);
1868 	if (ret)
1869 		goto out_clean_idr;
1870 
1871 	intel_gvt_init_engine_mmio_context(gvt);
1872 
1873 	ret = intel_gvt_load_firmware(gvt);
1874 	if (ret)
1875 		goto out_clean_mmio_info;
1876 
1877 	ret = intel_gvt_init_irq(gvt);
1878 	if (ret)
1879 		goto out_free_firmware;
1880 
1881 	ret = intel_gvt_init_gtt(gvt);
1882 	if (ret)
1883 		goto out_free_firmware;
1884 
1885 	ret = intel_gvt_init_workload_scheduler(gvt);
1886 	if (ret)
1887 		goto out_clean_gtt;
1888 
1889 	ret = intel_gvt_init_sched_policy(gvt);
1890 	if (ret)
1891 		goto out_clean_workload_scheduler;
1892 
1893 	ret = intel_gvt_init_cmd_parser(gvt);
1894 	if (ret)
1895 		goto out_clean_sched_policy;
1896 
1897 	ret = init_service_thread(gvt);
1898 	if (ret)
1899 		goto out_clean_cmd_parser;
1900 
1901 	ret = intel_gvt_init_vgpu_types(gvt);
1902 	if (ret)
1903 		goto out_clean_thread;
1904 
1905 	vgpu = intel_gvt_create_idle_vgpu(gvt);
1906 	if (IS_ERR(vgpu)) {
1907 		ret = PTR_ERR(vgpu);
1908 		gvt_err("failed to create idle vgpu\n");
1909 		goto out_clean_types;
1910 	}
1911 	gvt->idle_vgpu = vgpu;
1912 
1913 	intel_gvt_debugfs_init(gvt);
1914 
1915 	ret = mdev_register_parent(&gvt->parent, i915->drm.dev,
1916 				   &intel_vgpu_mdev_driver,
1917 				   gvt->mdev_types, gvt->num_types);
1918 	if (ret)
1919 		goto out_destroy_idle_vgpu;
1920 
1921 	gvt_dbg_core("gvt device initialization is done\n");
1922 	return 0;
1923 
1924 out_destroy_idle_vgpu:
1925 	intel_gvt_destroy_idle_vgpu(gvt->idle_vgpu);
1926 	intel_gvt_debugfs_clean(gvt);
1927 out_clean_types:
1928 	intel_gvt_clean_vgpu_types(gvt);
1929 out_clean_thread:
1930 	clean_service_thread(gvt);
1931 out_clean_cmd_parser:
1932 	intel_gvt_clean_cmd_parser(gvt);
1933 out_clean_sched_policy:
1934 	intel_gvt_clean_sched_policy(gvt);
1935 out_clean_workload_scheduler:
1936 	intel_gvt_clean_workload_scheduler(gvt);
1937 out_clean_gtt:
1938 	intel_gvt_clean_gtt(gvt);
1939 out_free_firmware:
1940 	intel_gvt_free_firmware(gvt);
1941 out_clean_mmio_info:
1942 	intel_gvt_clean_mmio_info(gvt);
1943 out_clean_idr:
1944 	idr_destroy(&gvt->vgpu_idr);
1945 	kfree(gvt);
1946 	i915->gvt = NULL;
1947 	return ret;
1948 }
1949 
1950 static void intel_gvt_pm_resume(struct drm_i915_private *i915)
1951 {
1952 	struct intel_gvt *gvt = i915->gvt;
1953 
1954 	intel_gvt_restore_fence(gvt);
1955 	intel_gvt_restore_mmio(gvt);
1956 	intel_gvt_restore_ggtt(gvt);
1957 }
1958 
1959 static const struct intel_vgpu_ops intel_gvt_vgpu_ops = {
1960 	.init_device	= intel_gvt_init_device,
1961 	.clean_device	= intel_gvt_clean_device,
1962 	.pm_resume	= intel_gvt_pm_resume,
1963 };
1964 
1965 static int __init kvmgt_init(void)
1966 {
1967 	int ret;
1968 
1969 	ret = intel_gvt_set_ops(&intel_gvt_vgpu_ops);
1970 	if (ret)
1971 		return ret;
1972 
1973 	ret = mdev_register_driver(&intel_vgpu_mdev_driver);
1974 	if (ret)
1975 		intel_gvt_clear_ops(&intel_gvt_vgpu_ops);
1976 	return ret;
1977 }
1978 
1979 static void __exit kvmgt_exit(void)
1980 {
1981 	mdev_unregister_driver(&intel_vgpu_mdev_driver);
1982 	intel_gvt_clear_ops(&intel_gvt_vgpu_ops);
1983 }
1984 
1985 module_init(kvmgt_init);
1986 module_exit(kvmgt_exit);
1987 
1988 MODULE_DESCRIPTION("Intel mediated pass-through framework for KVM");
1989 MODULE_LICENSE("GPL and additional rights");
1990 MODULE_AUTHOR("Intel Corporation");
1991