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 "amdgpu.h"
32 #include "amdgpu_trace.h"
33 #include "amdgpu_amdkfd.h"
34 
35 struct amdgpu_sync_entry {
36 	struct hlist_node	node;
37 	struct dma_fence	*fence;
38 	bool	explicit;
39 };
40 
41 static struct kmem_cache *amdgpu_sync_slab;
42 
43 /**
44  * amdgpu_sync_create - zero init sync object
45  *
46  * @sync: sync object to initialize
47  *
48  * Just clear the sync object for now.
49  */
50 void amdgpu_sync_create(struct amdgpu_sync *sync)
51 {
52 	hash_init(sync->fences);
53 	sync->last_vm_update = NULL;
54 }
55 
56 /**
57  * amdgpu_sync_same_dev - test if fence belong to us
58  *
59  * @adev: amdgpu device to use for the test
60  * @f: fence to test
61  *
62  * Test if the fence was issued by us.
63  */
64 static bool amdgpu_sync_same_dev(struct amdgpu_device *adev,
65 				 struct dma_fence *f)
66 {
67 	struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
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 /**
80  * amdgpu_sync_get_owner - extract the owner of a fence
81  *
82  * @fence: fence get the owner from
83  *
84  * Extract who originally created the fence.
85  */
86 static void *amdgpu_sync_get_owner(struct dma_fence *f)
87 {
88 	struct drm_sched_fence *s_fence;
89 	struct amdgpu_amdkfd_fence *kfd_fence;
90 
91 	if (!f)
92 		return AMDGPU_FENCE_OWNER_UNDEFINED;
93 
94 	s_fence = to_drm_sched_fence(f);
95 	if (s_fence)
96 		return s_fence->owner;
97 
98 	kfd_fence = to_amdgpu_amdkfd_fence(f);
99 	if (kfd_fence)
100 		return AMDGPU_FENCE_OWNER_KFD;
101 
102 	return AMDGPU_FENCE_OWNER_UNDEFINED;
103 }
104 
105 /**
106  * amdgpu_sync_keep_later - Keep the later fence
107  *
108  * @keep: existing fence to test
109  * @fence: new fence
110  *
111  * Either keep the existing fence or the new one, depending which one is later.
112  */
113 static void amdgpu_sync_keep_later(struct dma_fence **keep,
114 				   struct dma_fence *fence)
115 {
116 	if (*keep && dma_fence_is_later(*keep, fence))
117 		return;
118 
119 	dma_fence_put(*keep);
120 	*keep = dma_fence_get(fence);
121 }
122 
123 /**
124  * amdgpu_sync_add_later - add the fence to the hash
125  *
126  * @sync: sync object to add the fence to
127  * @f: fence to add
128  *
129  * Tries to add the fence to an existing hash entry. Returns true when an entry
130  * was found, false otherwise.
131  */
132 static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct dma_fence *f,
133 				  bool explicit)
134 {
135 	struct amdgpu_sync_entry *e;
136 
137 	hash_for_each_possible(sync->fences, e, node, f->context) {
138 		if (unlikely(e->fence->context != f->context))
139 			continue;
140 
141 		amdgpu_sync_keep_later(&e->fence, f);
142 
143 		/* Preserve eplicit flag to not loose pipe line sync */
144 		e->explicit |= explicit;
145 
146 		return true;
147 	}
148 	return false;
149 }
150 
151 /**
152  * amdgpu_sync_fence - remember to sync to this fence
153  *
154  * @sync: sync object to add fence to
155  * @f: fence to sync to
156  * @explicit: if this is an explicit dependency
157  *
158  * Add the fence to the sync object.
159  */
160 int amdgpu_sync_fence(struct amdgpu_sync *sync, struct dma_fence *f,
161 		      bool explicit)
162 {
163 	struct amdgpu_sync_entry *e;
164 
165 	if (!f)
166 		return 0;
167 
168 	if (amdgpu_sync_add_later(sync, f, explicit))
169 		return 0;
170 
171 	e = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
172 	if (!e)
173 		return -ENOMEM;
174 
175 	e->explicit = explicit;
176 
177 	hash_add(sync->fences, &e->node, f->context);
178 	e->fence = dma_fence_get(f);
179 	return 0;
180 }
181 
182 /**
183  * amdgpu_sync_vm_fence - remember to sync to this VM fence
184  *
185  * @adev: amdgpu device
186  * @sync: sync object to add fence to
187  * @fence: the VM fence to add
188  *
189  * Add the fence to the sync object and remember it as VM update.
190  */
191 int amdgpu_sync_vm_fence(struct amdgpu_sync *sync, struct dma_fence *fence)
192 {
193 	if (!fence)
194 		return 0;
195 
196 	amdgpu_sync_keep_later(&sync->last_vm_update, fence);
197 	return amdgpu_sync_fence(sync, fence, false);
198 }
199 
200 /**
201  * amdgpu_sync_resv - sync to a reservation object
202  *
203  * @sync: sync object to add fences from reservation object to
204  * @resv: reservation object with embedded fence
205  * @explicit_sync: true if we should only sync to the exclusive fence
206  *
207  * Sync to the fence
208  */
209 int amdgpu_sync_resv(struct amdgpu_device *adev,
210 		     struct amdgpu_sync *sync,
211 		     struct dma_resv *resv,
212 		     void *owner, bool explicit_sync)
213 {
214 	struct dma_resv_list *flist;
215 	struct dma_fence *f;
216 	void *fence_owner;
217 	unsigned i;
218 	int r = 0;
219 
220 	if (resv == NULL)
221 		return -EINVAL;
222 
223 	/* always sync to the exclusive fence */
224 	f = dma_resv_get_excl(resv);
225 	r = amdgpu_sync_fence(sync, f, false);
226 
227 	flist = dma_resv_get_list(resv);
228 	if (!flist || r)
229 		return r;
230 
231 	for (i = 0; i < flist->shared_count; ++i) {
232 		f = rcu_dereference_protected(flist->shared[i],
233 					      dma_resv_held(resv));
234 		/* We only want to trigger KFD eviction fences on
235 		 * evict or move jobs. Skip KFD fences otherwise.
236 		 */
237 		fence_owner = amdgpu_sync_get_owner(f);
238 		if (fence_owner == AMDGPU_FENCE_OWNER_KFD &&
239 		    owner != AMDGPU_FENCE_OWNER_UNDEFINED)
240 			continue;
241 
242 		if (amdgpu_sync_same_dev(adev, f)) {
243 			/* VM updates only sync with moves but not with user
244 			 * command submissions or KFD evictions fences
245 			 */
246 			if (owner == AMDGPU_FENCE_OWNER_VM &&
247 			    fence_owner != AMDGPU_FENCE_OWNER_UNDEFINED)
248 				continue;
249 
250 			/* Ignore fence from the same owner and explicit one as
251 			 * long as it isn't undefined.
252 			 */
253 			if (owner != AMDGPU_FENCE_OWNER_UNDEFINED &&
254 			    (fence_owner == owner || explicit_sync))
255 				continue;
256 		}
257 
258 		r = amdgpu_sync_fence(sync, f, false);
259 		if (r)
260 			break;
261 	}
262 	return r;
263 }
264 
265 /**
266  * amdgpu_sync_peek_fence - get the next fence not signaled yet
267  *
268  * @sync: the sync object
269  * @ring: optional ring to use for test
270  *
271  * Returns the next fence not signaled yet without removing it from the sync
272  * object.
273  */
274 struct dma_fence *amdgpu_sync_peek_fence(struct amdgpu_sync *sync,
275 					 struct amdgpu_ring *ring)
276 {
277 	struct amdgpu_sync_entry *e;
278 	struct hlist_node *tmp;
279 	int i;
280 
281 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
282 		struct dma_fence *f = e->fence;
283 		struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
284 
285 		if (dma_fence_is_signaled(f)) {
286 			hash_del(&e->node);
287 			dma_fence_put(f);
288 			kmem_cache_free(amdgpu_sync_slab, e);
289 			continue;
290 		}
291 		if (ring && s_fence) {
292 			/* For fences from the same ring it is sufficient
293 			 * when they are scheduled.
294 			 */
295 			if (s_fence->sched == &ring->sched) {
296 				if (dma_fence_is_signaled(&s_fence->scheduled))
297 					continue;
298 
299 				return &s_fence->scheduled;
300 			}
301 		}
302 
303 		return f;
304 	}
305 
306 	return NULL;
307 }
308 
309 /**
310  * amdgpu_sync_get_fence - get the next fence from the sync object
311  *
312  * @sync: sync object to use
313  * @explicit: true if the next fence is explicit
314  *
315  * Get and removes the next fence from the sync object not signaled yet.
316  */
317 struct dma_fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync, bool *explicit)
318 {
319 	struct amdgpu_sync_entry *e;
320 	struct hlist_node *tmp;
321 	struct dma_fence *f;
322 	int i;
323 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
324 
325 		f = e->fence;
326 		if (explicit)
327 			*explicit = e->explicit;
328 
329 		hash_del(&e->node);
330 		kmem_cache_free(amdgpu_sync_slab, e);
331 
332 		if (!dma_fence_is_signaled(f))
333 			return f;
334 
335 		dma_fence_put(f);
336 	}
337 	return NULL;
338 }
339 
340 /**
341  * amdgpu_sync_clone - clone a sync object
342  *
343  * @source: sync object to clone
344  * @clone: pointer to destination sync object
345  *
346  * Adds references to all unsignaled fences in @source to @clone. Also
347  * removes signaled fences from @source while at it.
348  */
349 int amdgpu_sync_clone(struct amdgpu_sync *source, struct amdgpu_sync *clone)
350 {
351 	struct amdgpu_sync_entry *e;
352 	struct hlist_node *tmp;
353 	struct dma_fence *f;
354 	int i, r;
355 
356 	hash_for_each_safe(source->fences, i, tmp, e, node) {
357 		f = e->fence;
358 		if (!dma_fence_is_signaled(f)) {
359 			r = amdgpu_sync_fence(clone, f, e->explicit);
360 			if (r)
361 				return r;
362 		} else {
363 			hash_del(&e->node);
364 			dma_fence_put(f);
365 			kmem_cache_free(amdgpu_sync_slab, e);
366 		}
367 	}
368 
369 	dma_fence_put(clone->last_vm_update);
370 	clone->last_vm_update = dma_fence_get(source->last_vm_update);
371 
372 	return 0;
373 }
374 
375 int amdgpu_sync_wait(struct amdgpu_sync *sync, bool intr)
376 {
377 	struct amdgpu_sync_entry *e;
378 	struct hlist_node *tmp;
379 	int i, r;
380 
381 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
382 		r = dma_fence_wait(e->fence, intr);
383 		if (r)
384 			return r;
385 
386 		hash_del(&e->node);
387 		dma_fence_put(e->fence);
388 		kmem_cache_free(amdgpu_sync_slab, e);
389 	}
390 
391 	return 0;
392 }
393 
394 /**
395  * amdgpu_sync_free - free the sync object
396  *
397  * @sync: sync object to use
398  *
399  * Free the sync object.
400  */
401 void amdgpu_sync_free(struct amdgpu_sync *sync)
402 {
403 	struct amdgpu_sync_entry *e;
404 	struct hlist_node *tmp;
405 	unsigned i;
406 
407 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
408 		hash_del(&e->node);
409 		dma_fence_put(e->fence);
410 		kmem_cache_free(amdgpu_sync_slab, e);
411 	}
412 
413 	dma_fence_put(sync->last_vm_update);
414 }
415 
416 /**
417  * amdgpu_sync_init - init sync object subsystem
418  *
419  * Allocate the slab allocator.
420  */
421 int amdgpu_sync_init(void)
422 {
423 	amdgpu_sync_slab = kmem_cache_create(
424 		"amdgpu_sync", sizeof(struct amdgpu_sync_entry), 0,
425 		SLAB_HWCACHE_ALIGN, NULL);
426 	if (!amdgpu_sync_slab)
427 		return -ENOMEM;
428 
429 	return 0;
430 }
431 
432 /**
433  * amdgpu_sync_fini - fini sync object subsystem
434  *
435  * Free the slab allocator.
436  */
437 void amdgpu_sync_fini(void)
438 {
439 	kmem_cache_destroy(amdgpu_sync_slab);
440 }
441