1 /*
2  * Copyright 2015 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Christian König <[email protected]>
29  */
30 
31 #include <drm/drmP.h>
32 #include "amdgpu.h"
33 #include "amdgpu_trace.h"
34 
35 #define AMDGPU_BO_LIST_MAX_PRIORITY	32u
36 #define AMDGPU_BO_LIST_NUM_BUCKETS	(AMDGPU_BO_LIST_MAX_PRIORITY + 1)
37 
38 static int amdgpu_bo_list_set(struct amdgpu_device *adev,
39 				     struct drm_file *filp,
40 				     struct amdgpu_bo_list *list,
41 				     struct drm_amdgpu_bo_list_entry *info,
42 				     unsigned num_entries);
43 
44 static void amdgpu_bo_list_release_rcu(struct kref *ref)
45 {
46 	unsigned i;
47 	struct amdgpu_bo_list *list = container_of(ref, struct amdgpu_bo_list,
48 						   refcount);
49 
50 	for (i = 0; i < list->num_entries; ++i)
51 		amdgpu_bo_unref(&list->array[i].robj);
52 
53 	mutex_destroy(&list->lock);
54 	kvfree(list->array);
55 	kfree_rcu(list, rhead);
56 }
57 
58 static int amdgpu_bo_list_create(struct amdgpu_device *adev,
59 				 struct drm_file *filp,
60 				 struct drm_amdgpu_bo_list_entry *info,
61 				 unsigned num_entries,
62 				 int *id)
63 {
64 	int r;
65 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
66 	struct amdgpu_bo_list *list;
67 
68 	list = kzalloc(sizeof(struct amdgpu_bo_list), GFP_KERNEL);
69 	if (!list)
70 		return -ENOMEM;
71 
72 	/* initialize bo list*/
73 	mutex_init(&list->lock);
74 	kref_init(&list->refcount);
75 	r = amdgpu_bo_list_set(adev, filp, list, info, num_entries);
76 	if (r) {
77 		kfree(list);
78 		return r;
79 	}
80 
81 	/* idr alloc should be called only after initialization of bo list. */
82 	mutex_lock(&fpriv->bo_list_lock);
83 	r = idr_alloc(&fpriv->bo_list_handles, list, 1, 0, GFP_KERNEL);
84 	mutex_unlock(&fpriv->bo_list_lock);
85 	if (r < 0) {
86 		kfree(list);
87 		return r;
88 	}
89 	*id = r;
90 
91 	return 0;
92 }
93 
94 static void amdgpu_bo_list_destroy(struct amdgpu_fpriv *fpriv, int id)
95 {
96 	struct amdgpu_bo_list *list;
97 
98 	mutex_lock(&fpriv->bo_list_lock);
99 	list = idr_remove(&fpriv->bo_list_handles, id);
100 	mutex_unlock(&fpriv->bo_list_lock);
101 	if (list)
102 		kref_put(&list->refcount, amdgpu_bo_list_release_rcu);
103 }
104 
105 static int amdgpu_bo_list_set(struct amdgpu_device *adev,
106 				     struct drm_file *filp,
107 				     struct amdgpu_bo_list *list,
108 				     struct drm_amdgpu_bo_list_entry *info,
109 				     unsigned num_entries)
110 {
111 	struct amdgpu_bo_list_entry *array;
112 	struct amdgpu_bo *gds_obj = adev->gds.gds_gfx_bo;
113 	struct amdgpu_bo *gws_obj = adev->gds.gws_gfx_bo;
114 	struct amdgpu_bo *oa_obj = adev->gds.oa_gfx_bo;
115 
116 	unsigned last_entry = 0, first_userptr = num_entries;
117 	unsigned i;
118 	int r;
119 	unsigned long total_size = 0;
120 
121 	array = kvmalloc_array(num_entries, sizeof(struct amdgpu_bo_list_entry), GFP_KERNEL);
122 	if (!array)
123 		return -ENOMEM;
124 	memset(array, 0, num_entries * sizeof(struct amdgpu_bo_list_entry));
125 
126 	for (i = 0; i < num_entries; ++i) {
127 		struct amdgpu_bo_list_entry *entry;
128 		struct drm_gem_object *gobj;
129 		struct amdgpu_bo *bo;
130 		struct mm_struct *usermm;
131 
132 		gobj = drm_gem_object_lookup(filp, info[i].bo_handle);
133 		if (!gobj) {
134 			r = -ENOENT;
135 			goto error_free;
136 		}
137 
138 		bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
139 		drm_gem_object_unreference_unlocked(gobj);
140 
141 		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
142 		if (usermm) {
143 			if (usermm != current->mm) {
144 				amdgpu_bo_unref(&bo);
145 				r = -EPERM;
146 				goto error_free;
147 			}
148 			entry = &array[--first_userptr];
149 		} else {
150 			entry = &array[last_entry++];
151 		}
152 
153 		entry->robj = bo;
154 		entry->priority = min(info[i].bo_priority,
155 				      AMDGPU_BO_LIST_MAX_PRIORITY);
156 		entry->tv.bo = &entry->robj->tbo;
157 		entry->tv.shared = !entry->robj->prime_shared_count;
158 
159 		if (entry->robj->prefered_domains == AMDGPU_GEM_DOMAIN_GDS)
160 			gds_obj = entry->robj;
161 		if (entry->robj->prefered_domains == AMDGPU_GEM_DOMAIN_GWS)
162 			gws_obj = entry->robj;
163 		if (entry->robj->prefered_domains == AMDGPU_GEM_DOMAIN_OA)
164 			oa_obj = entry->robj;
165 
166 		total_size += amdgpu_bo_size(entry->robj);
167 		trace_amdgpu_bo_list_set(list, entry->robj);
168 	}
169 
170 	for (i = 0; i < list->num_entries; ++i)
171 		amdgpu_bo_unref(&list->array[i].robj);
172 
173 	kvfree(list->array);
174 
175 	list->gds_obj = gds_obj;
176 	list->gws_obj = gws_obj;
177 	list->oa_obj = oa_obj;
178 	list->first_userptr = first_userptr;
179 	list->array = array;
180 	list->num_entries = num_entries;
181 
182 	trace_amdgpu_cs_bo_status(list->num_entries, total_size);
183 	return 0;
184 
185 error_free:
186 	while (i--)
187 		amdgpu_bo_unref(&array[i].robj);
188 	kvfree(array);
189 	return r;
190 }
191 
192 struct amdgpu_bo_list *
193 amdgpu_bo_list_get(struct amdgpu_fpriv *fpriv, int id)
194 {
195 	struct amdgpu_bo_list *result;
196 
197 	rcu_read_lock();
198 	result = idr_find(&fpriv->bo_list_handles, id);
199 
200 	if (result) {
201 		if (kref_get_unless_zero(&result->refcount))
202 			mutex_lock(&result->lock);
203 		else
204 			result = NULL;
205 	}
206 	rcu_read_unlock();
207 
208 	return result;
209 }
210 
211 void amdgpu_bo_list_get_list(struct amdgpu_bo_list *list,
212 			     struct list_head *validated)
213 {
214 	/* This is based on the bucket sort with O(n) time complexity.
215 	 * An item with priority "i" is added to bucket[i]. The lists are then
216 	 * concatenated in descending order.
217 	 */
218 	struct list_head bucket[AMDGPU_BO_LIST_NUM_BUCKETS];
219 	unsigned i;
220 
221 	for (i = 0; i < AMDGPU_BO_LIST_NUM_BUCKETS; i++)
222 		INIT_LIST_HEAD(&bucket[i]);
223 
224 	/* Since buffers which appear sooner in the relocation list are
225 	 * likely to be used more often than buffers which appear later
226 	 * in the list, the sort mustn't change the ordering of buffers
227 	 * with the same priority, i.e. it must be stable.
228 	 */
229 	for (i = 0; i < list->num_entries; i++) {
230 		unsigned priority = list->array[i].priority;
231 
232 		list_add_tail(&list->array[i].tv.head,
233 			      &bucket[priority]);
234 		list->array[i].user_pages = NULL;
235 	}
236 
237 	/* Connect the sorted buckets in the output list. */
238 	for (i = 0; i < AMDGPU_BO_LIST_NUM_BUCKETS; i++)
239 		list_splice(&bucket[i], validated);
240 }
241 
242 void amdgpu_bo_list_put(struct amdgpu_bo_list *list)
243 {
244 	mutex_unlock(&list->lock);
245 	kref_put(&list->refcount, amdgpu_bo_list_release_rcu);
246 }
247 
248 void amdgpu_bo_list_free(struct amdgpu_bo_list *list)
249 {
250 	unsigned i;
251 
252 	for (i = 0; i < list->num_entries; ++i)
253 		amdgpu_bo_unref(&list->array[i].robj);
254 
255 	mutex_destroy(&list->lock);
256 	kvfree(list->array);
257 	kfree(list);
258 }
259 
260 int amdgpu_bo_list_ioctl(struct drm_device *dev, void *data,
261 				struct drm_file *filp)
262 {
263 	const uint32_t info_size = sizeof(struct drm_amdgpu_bo_list_entry);
264 
265 	struct amdgpu_device *adev = dev->dev_private;
266 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
267 	union drm_amdgpu_bo_list *args = data;
268 	uint32_t handle = args->in.list_handle;
269 	const void __user *uptr = (const void*)(uintptr_t)args->in.bo_info_ptr;
270 
271 	struct drm_amdgpu_bo_list_entry *info;
272 	struct amdgpu_bo_list *list;
273 
274 	int r;
275 
276 	info = kvmalloc_array(args->in.bo_number,
277 			     sizeof(struct drm_amdgpu_bo_list_entry), GFP_KERNEL);
278 	if (!info)
279 		return -ENOMEM;
280 
281 	/* copy the handle array from userspace to a kernel buffer */
282 	r = -EFAULT;
283 	if (likely(info_size == args->in.bo_info_size)) {
284 		unsigned long bytes = args->in.bo_number *
285 			args->in.bo_info_size;
286 
287 		if (copy_from_user(info, uptr, bytes))
288 			goto error_free;
289 
290 	} else {
291 		unsigned long bytes = min(args->in.bo_info_size, info_size);
292 		unsigned i;
293 
294 		memset(info, 0, args->in.bo_number * info_size);
295 		for (i = 0; i < args->in.bo_number; ++i) {
296 			if (copy_from_user(&info[i], uptr, bytes))
297 				goto error_free;
298 
299 			uptr += args->in.bo_info_size;
300 		}
301 	}
302 
303 	switch (args->in.operation) {
304 	case AMDGPU_BO_LIST_OP_CREATE:
305 		r = amdgpu_bo_list_create(adev, filp, info, args->in.bo_number,
306 					  &handle);
307 		if (r)
308 			goto error_free;
309 		break;
310 
311 	case AMDGPU_BO_LIST_OP_DESTROY:
312 		amdgpu_bo_list_destroy(fpriv, handle);
313 		handle = 0;
314 		break;
315 
316 	case AMDGPU_BO_LIST_OP_UPDATE:
317 		r = -ENOENT;
318 		list = amdgpu_bo_list_get(fpriv, handle);
319 		if (!list)
320 			goto error_free;
321 
322 		r = amdgpu_bo_list_set(adev, filp, list, info,
323 					      args->in.bo_number);
324 		amdgpu_bo_list_put(list);
325 		if (r)
326 			goto error_free;
327 
328 		break;
329 
330 	default:
331 		r = -EINVAL;
332 		goto error_free;
333 	}
334 
335 	memset(args, 0, sizeof(*args));
336 	args->out.list_handle = handle;
337 	kvfree(info);
338 
339 	return 0;
340 
341 error_free:
342 	kvfree(info);
343 	return r;
344 }
345