1 /*
2  * Copyright 2014 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 <christian.koenig@amd.com>
29  */
30 
31 #include <drm/drmP.h>
32 #include "amdgpu.h"
33 #include "amdgpu_trace.h"
34 
35 struct amdgpu_sync_entry {
36 	struct hlist_node	node;
37 	struct fence		*fence;
38 };
39 
40 /**
41  * amdgpu_sync_create - zero init sync object
42  *
43  * @sync: sync object to initialize
44  *
45  * Just clear the sync object for now.
46  */
47 void amdgpu_sync_create(struct amdgpu_sync *sync)
48 {
49 	unsigned i;
50 
51 	for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
52 		sync->semaphores[i] = NULL;
53 
54 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
55 		sync->sync_to[i] = NULL;
56 
57 	hash_init(sync->fences);
58 	sync->last_vm_update = NULL;
59 }
60 
61 static bool amdgpu_sync_same_dev(struct amdgpu_device *adev, struct fence *f)
62 {
63 	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
64 	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
65 
66 	if (a_fence)
67 		return a_fence->ring->adev == adev;
68 	if (s_fence)
69 		return (struct amdgpu_device *)s_fence->scheduler->priv == adev;
70 	return false;
71 }
72 
73 static bool amdgpu_sync_test_owner(struct fence *f, void *owner)
74 {
75 	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
76 	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
77 	if (s_fence)
78 		return s_fence->owner == owner;
79 	if (a_fence)
80 		return a_fence->owner == owner;
81 	return false;
82 }
83 
84 /**
85  * amdgpu_sync_fence - remember to sync to this fence
86  *
87  * @sync: sync object to add fence to
88  * @fence: fence to sync to
89  *
90  */
91 int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
92 		      struct fence *f)
93 {
94 	struct amdgpu_sync_entry *e;
95 	struct amdgpu_fence *fence;
96 	struct amdgpu_fence *other;
97 	struct fence *tmp, *later;
98 
99 	if (!f)
100 		return 0;
101 
102 	if (amdgpu_sync_same_dev(adev, f) &&
103 	    amdgpu_sync_test_owner(f, AMDGPU_FENCE_OWNER_VM)) {
104 		if (sync->last_vm_update) {
105 			tmp = sync->last_vm_update;
106 			BUG_ON(f->context != tmp->context);
107 			later = (f->seqno - tmp->seqno <= INT_MAX) ? f : tmp;
108 			sync->last_vm_update = fence_get(later);
109 			fence_put(tmp);
110 		} else
111 			sync->last_vm_update = fence_get(f);
112 	}
113 
114 	fence = to_amdgpu_fence(f);
115 	if (!fence || fence->ring->adev != adev) {
116 		hash_for_each_possible(sync->fences, e, node, f->context) {
117 			struct fence *new;
118 			if (unlikely(e->fence->context != f->context))
119 				continue;
120 			new = fence_get(fence_later(e->fence, f));
121 			if (new) {
122 				fence_put(e->fence);
123 				e->fence = new;
124 			}
125 			return 0;
126 		}
127 
128 		e = kmalloc(sizeof(struct amdgpu_sync_entry), GFP_KERNEL);
129 		if (!e)
130 			return -ENOMEM;
131 
132 		hash_add(sync->fences, &e->node, f->context);
133 		e->fence = fence_get(f);
134 		return 0;
135 	}
136 
137 	other = sync->sync_to[fence->ring->idx];
138 	sync->sync_to[fence->ring->idx] = amdgpu_fence_ref(
139 		amdgpu_fence_later(fence, other));
140 	amdgpu_fence_unref(&other);
141 
142 	return 0;
143 }
144 
145 /**
146  * amdgpu_sync_resv - use the semaphores to sync to a reservation object
147  *
148  * @sync: sync object to add fences from reservation object to
149  * @resv: reservation object with embedded fence
150  * @shared: true if we should only sync to the exclusive fence
151  *
152  * Sync to the fence using the semaphore objects
153  */
154 int amdgpu_sync_resv(struct amdgpu_device *adev,
155 		     struct amdgpu_sync *sync,
156 		     struct reservation_object *resv,
157 		     void *owner)
158 {
159 	struct reservation_object_list *flist;
160 	struct fence *f;
161 	struct amdgpu_fence *fence;
162 	unsigned i;
163 	int r = 0;
164 
165 	if (resv == NULL)
166 		return -EINVAL;
167 
168 	/* always sync to the exclusive fence */
169 	f = reservation_object_get_excl(resv);
170 	r = amdgpu_sync_fence(adev, sync, f);
171 
172 	flist = reservation_object_get_list(resv);
173 	if (!flist || r)
174 		return r;
175 
176 	for (i = 0; i < flist->shared_count; ++i) {
177 		f = rcu_dereference_protected(flist->shared[i],
178 					      reservation_object_held(resv));
179 		fence = f ? to_amdgpu_fence(f) : NULL;
180 		if (fence && fence->ring->adev == adev) {
181 			/* VM updates are only interesting
182 			 * for other VM updates and moves.
183 			 */
184 			if ((owner != AMDGPU_FENCE_OWNER_MOVE) &&
185 			    (fence->owner != AMDGPU_FENCE_OWNER_MOVE) &&
186 			    ((owner == AMDGPU_FENCE_OWNER_VM) !=
187 			     (fence->owner == AMDGPU_FENCE_OWNER_VM)))
188 				continue;
189 
190 			/* Ignore fence from the same owner as
191 			 * long as it isn't undefined.
192 			 */
193 			if (owner != AMDGPU_FENCE_OWNER_UNDEFINED &&
194 			    fence->owner == owner)
195 				continue;
196 		}
197 
198 		r = amdgpu_sync_fence(adev, sync, f);
199 		if (r)
200 			break;
201 	}
202 	return r;
203 }
204 
205 int amdgpu_sync_wait(struct amdgpu_sync *sync)
206 {
207 	struct amdgpu_sync_entry *e;
208 	struct hlist_node *tmp;
209 	int i, r;
210 
211 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
212 		r = fence_wait(e->fence, false);
213 		if (r)
214 			return r;
215 
216 		hash_del(&e->node);
217 		fence_put(e->fence);
218 		kfree(e);
219 	}
220 	return 0;
221 }
222 
223 /**
224  * amdgpu_sync_rings - sync ring to all registered fences
225  *
226  * @sync: sync object to use
227  * @ring: ring that needs sync
228  *
229  * Ensure that all registered fences are signaled before letting
230  * the ring continue. The caller must hold the ring lock.
231  */
232 int amdgpu_sync_rings(struct amdgpu_sync *sync,
233 		      struct amdgpu_ring *ring)
234 {
235 	struct amdgpu_device *adev = ring->adev;
236 	unsigned count = 0;
237 	int i, r;
238 
239 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
240 		struct amdgpu_fence *fence = sync->sync_to[i];
241 		struct amdgpu_semaphore *semaphore;
242 		struct amdgpu_ring *other = adev->rings[i];
243 
244 		/* check if we really need to sync */
245 		if (!amdgpu_fence_need_sync(fence, ring))
246 			continue;
247 
248 		/* prevent GPU deadlocks */
249 		if (!other->ready) {
250 			dev_err(adev->dev, "Syncing to a disabled ring!");
251 			return -EINVAL;
252 		}
253 
254 		if (amdgpu_enable_scheduler || (count >= AMDGPU_NUM_SYNCS)) {
255 			/* not enough room, wait manually */
256 			r = fence_wait(&fence->base, false);
257 			if (r)
258 				return r;
259 			continue;
260 		}
261 		r = amdgpu_semaphore_create(adev, &semaphore);
262 		if (r)
263 			return r;
264 
265 		sync->semaphores[count++] = semaphore;
266 
267 		/* allocate enough space for sync command */
268 		r = amdgpu_ring_alloc(other, 16);
269 		if (r)
270 			return r;
271 
272 		/* emit the signal semaphore */
273 		if (!amdgpu_semaphore_emit_signal(other, semaphore)) {
274 			/* signaling wasn't successful wait manually */
275 			amdgpu_ring_undo(other);
276 			r = fence_wait(&fence->base, false);
277 			if (r)
278 				return r;
279 			continue;
280 		}
281 
282 		/* we assume caller has already allocated space on waiters ring */
283 		if (!amdgpu_semaphore_emit_wait(ring, semaphore)) {
284 			/* waiting wasn't successful wait manually */
285 			amdgpu_ring_undo(other);
286 			r = fence_wait(&fence->base, false);
287 			if (r)
288 				return r;
289 			continue;
290 		}
291 
292 		amdgpu_ring_commit(other);
293 		amdgpu_fence_note_sync(fence, ring);
294 	}
295 
296 	return 0;
297 }
298 
299 /**
300  * amdgpu_sync_free - free the sync object
301  *
302  * @adev: amdgpu_device pointer
303  * @sync: sync object to use
304  * @fence: fence to use for the free
305  *
306  * Free the sync object by freeing all semaphores in it.
307  */
308 void amdgpu_sync_free(struct amdgpu_device *adev,
309 		      struct amdgpu_sync *sync,
310 		      struct fence *fence)
311 {
312 	struct amdgpu_sync_entry *e;
313 	struct hlist_node *tmp;
314 	unsigned i;
315 
316 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
317 		hash_del(&e->node);
318 		fence_put(e->fence);
319 		kfree(e);
320 	}
321 
322 	for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
323 		amdgpu_semaphore_free(adev, &sync->semaphores[i], fence);
324 
325 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
326 		amdgpu_fence_unref(&sync->sync_to[i]);
327 
328 	fence_put(sync->last_vm_update);
329 }
330