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 
69 	if (s_fence) {
70 		struct amdgpu_ring *ring;
71 
72 		ring = container_of(s_fence->sched, struct amdgpu_ring, sched);
73 		return ring->adev == adev;
74 	}
75 
76 	return false;
77 }
78 
79 static bool amdgpu_sync_test_owner(struct fence *f, void *owner)
80 {
81 	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
82 	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
83 	if (s_fence)
84 		return s_fence->owner == owner;
85 	if (a_fence)
86 		return a_fence->owner == owner;
87 	return false;
88 }
89 
90 static void amdgpu_sync_keep_later(struct fence **keep, struct fence *fence)
91 {
92 	if (*keep && fence_is_later(*keep, fence))
93 		return;
94 
95 	fence_put(*keep);
96 	*keep = fence_get(fence);
97 }
98 
99 /**
100  * amdgpu_sync_fence - remember to sync to this fence
101  *
102  * @sync: sync object to add fence to
103  * @fence: fence to sync to
104  *
105  */
106 int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
107 		      struct fence *f)
108 {
109 	struct amdgpu_sync_entry *e;
110 	struct amdgpu_fence *fence;
111 
112 	if (!f)
113 		return 0;
114 
115 	if (amdgpu_sync_same_dev(adev, f) &&
116 	    amdgpu_sync_test_owner(f, AMDGPU_FENCE_OWNER_VM))
117 		amdgpu_sync_keep_later(&sync->last_vm_update, f);
118 
119 	fence = to_amdgpu_fence(f);
120 	if (!fence || fence->ring->adev != adev) {
121 		hash_for_each_possible(sync->fences, e, node, f->context) {
122 			if (unlikely(e->fence->context != f->context))
123 				continue;
124 
125 			amdgpu_sync_keep_later(&e->fence, f);
126 			return 0;
127 		}
128 
129 		e = kmalloc(sizeof(struct amdgpu_sync_entry), GFP_KERNEL);
130 		if (!e)
131 			return -ENOMEM;
132 
133 		hash_add(sync->fences, &e->node, f->context);
134 		e->fence = fence_get(f);
135 		return 0;
136 	}
137 
138 	amdgpu_sync_keep_later(&sync->sync_to[fence->ring->idx], f);
139 
140 	return 0;
141 }
142 
143 static void *amdgpu_sync_get_owner(struct fence *f)
144 {
145 	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
146 	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
147 
148 	if (s_fence)
149 		return s_fence->owner;
150 	else if (a_fence)
151 		return a_fence->owner;
152 	return AMDGPU_FENCE_OWNER_UNDEFINED;
153 }
154 
155 /**
156  * amdgpu_sync_resv - use the semaphores to sync to a reservation object
157  *
158  * @sync: sync object to add fences from reservation object to
159  * @resv: reservation object with embedded fence
160  * @shared: true if we should only sync to the exclusive fence
161  *
162  * Sync to the fence using the semaphore objects
163  */
164 int amdgpu_sync_resv(struct amdgpu_device *adev,
165 		     struct amdgpu_sync *sync,
166 		     struct reservation_object *resv,
167 		     void *owner)
168 {
169 	struct reservation_object_list *flist;
170 	struct fence *f;
171 	void *fence_owner;
172 	unsigned i;
173 	int r = 0;
174 
175 	if (resv == NULL)
176 		return -EINVAL;
177 
178 	/* always sync to the exclusive fence */
179 	f = reservation_object_get_excl(resv);
180 	r = amdgpu_sync_fence(adev, sync, f);
181 
182 	flist = reservation_object_get_list(resv);
183 	if (!flist || r)
184 		return r;
185 
186 	for (i = 0; i < flist->shared_count; ++i) {
187 		f = rcu_dereference_protected(flist->shared[i],
188 					      reservation_object_held(resv));
189 		if (amdgpu_sync_same_dev(adev, f)) {
190 			/* VM updates are only interesting
191 			 * for other VM updates and moves.
192 			 */
193 			fence_owner = amdgpu_sync_get_owner(f);
194 			if ((owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
195 			    (fence_owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
196 			    ((owner == AMDGPU_FENCE_OWNER_VM) !=
197 			     (fence_owner == AMDGPU_FENCE_OWNER_VM)))
198 				continue;
199 
200 			/* Ignore fence from the same owner as
201 			 * long as it isn't undefined.
202 			 */
203 			if (owner != AMDGPU_FENCE_OWNER_UNDEFINED &&
204 			    fence_owner == owner)
205 				continue;
206 		}
207 
208 		r = amdgpu_sync_fence(adev, sync, f);
209 		if (r)
210 			break;
211 	}
212 	return r;
213 }
214 
215 struct fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync)
216 {
217 	struct amdgpu_sync_entry *e;
218 	struct hlist_node *tmp;
219 	struct fence *f;
220 	int i;
221 
222 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
223 
224 		f = e->fence;
225 
226 		hash_del(&e->node);
227 		kfree(e);
228 
229 		if (!fence_is_signaled(f))
230 			return f;
231 
232 		fence_put(f);
233 	}
234 	return NULL;
235 }
236 
237 int amdgpu_sync_wait(struct amdgpu_sync *sync)
238 {
239 	struct amdgpu_sync_entry *e;
240 	struct hlist_node *tmp;
241 	int i, r;
242 
243 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
244 		r = fence_wait(e->fence, false);
245 		if (r)
246 			return r;
247 
248 		hash_del(&e->node);
249 		fence_put(e->fence);
250 		kfree(e);
251 	}
252 
253 	if (amdgpu_enable_semaphores)
254 		return 0;
255 
256 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
257 		struct fence *fence = sync->sync_to[i];
258 		if (!fence)
259 			continue;
260 
261 		r = fence_wait(fence, false);
262 		if (r)
263 			return r;
264 	}
265 
266 	return 0;
267 }
268 
269 /**
270  * amdgpu_sync_rings - sync ring to all registered fences
271  *
272  * @sync: sync object to use
273  * @ring: ring that needs sync
274  *
275  * Ensure that all registered fences are signaled before letting
276  * the ring continue. The caller must hold the ring lock.
277  */
278 int amdgpu_sync_rings(struct amdgpu_sync *sync,
279 		      struct amdgpu_ring *ring)
280 {
281 	struct amdgpu_device *adev = ring->adev;
282 	unsigned count = 0;
283 	int i, r;
284 
285 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
286 		struct amdgpu_ring *other = adev->rings[i];
287 		struct amdgpu_semaphore *semaphore;
288 		struct amdgpu_fence *fence;
289 
290 		if (!sync->sync_to[i])
291 			continue;
292 
293 		fence = to_amdgpu_fence(sync->sync_to[i]);
294 
295 		/* check if we really need to sync */
296 		if (!amdgpu_fence_need_sync(fence, ring))
297 			continue;
298 
299 		/* prevent GPU deadlocks */
300 		if (!other->ready) {
301 			dev_err(adev->dev, "Syncing to a disabled ring!");
302 			return -EINVAL;
303 		}
304 
305 		if (amdgpu_enable_scheduler || !amdgpu_enable_semaphores ||
306 		    (count >= AMDGPU_NUM_SYNCS)) {
307 			/* not enough room, wait manually */
308 			r = fence_wait(&fence->base, false);
309 			if (r)
310 				return r;
311 			continue;
312 		}
313 		r = amdgpu_semaphore_create(adev, &semaphore);
314 		if (r)
315 			return r;
316 
317 		sync->semaphores[count++] = semaphore;
318 
319 		/* allocate enough space for sync command */
320 		r = amdgpu_ring_alloc(other, 16);
321 		if (r)
322 			return r;
323 
324 		/* emit the signal semaphore */
325 		if (!amdgpu_semaphore_emit_signal(other, semaphore)) {
326 			/* signaling wasn't successful wait manually */
327 			amdgpu_ring_undo(other);
328 			r = fence_wait(&fence->base, false);
329 			if (r)
330 				return r;
331 			continue;
332 		}
333 
334 		/* we assume caller has already allocated space on waiters ring */
335 		if (!amdgpu_semaphore_emit_wait(ring, semaphore)) {
336 			/* waiting wasn't successful wait manually */
337 			amdgpu_ring_undo(other);
338 			r = fence_wait(&fence->base, false);
339 			if (r)
340 				return r;
341 			continue;
342 		}
343 
344 		amdgpu_ring_commit(other);
345 		amdgpu_fence_note_sync(fence, ring);
346 	}
347 
348 	return 0;
349 }
350 
351 /**
352  * amdgpu_sync_free - free the sync object
353  *
354  * @adev: amdgpu_device pointer
355  * @sync: sync object to use
356  * @fence: fence to use for the free
357  *
358  * Free the sync object by freeing all semaphores in it.
359  */
360 void amdgpu_sync_free(struct amdgpu_device *adev,
361 		      struct amdgpu_sync *sync,
362 		      struct fence *fence)
363 {
364 	struct amdgpu_sync_entry *e;
365 	struct hlist_node *tmp;
366 	unsigned i;
367 
368 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
369 		hash_del(&e->node);
370 		fence_put(e->fence);
371 		kfree(e);
372 	}
373 
374 	for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
375 		amdgpu_semaphore_free(adev, &sync->semaphores[i], fence);
376 
377 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
378 		fence_put(sync->sync_to[i]);
379 
380 	fence_put(sync->last_vm_update);
381 }
382