xref: /openbsd-src/sys/dev/pci/drm/amd/amdgpu/amdgpu_sync.c (revision 68dd5bb1859285b71cb62a10bf107b8ad54064d9)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright 2014 Advanced Micro Devices, Inc.
4  * 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
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
18  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
19  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
20  * USE OR OTHER DEALINGS IN THE SOFTWARE.
21  *
22  * The above copyright notice and this permission notice (including the
23  * next paragraph) shall be included in all copies or substantial portions
24  * of the Software.
25  *
26  */
27 /*
28  * Authors:
29  *    Christian König <christian.koenig@amd.com>
30  */
31 
32 #include <linux/dma-fence-chain.h>
33 
34 #include "amdgpu.h"
35 #include "amdgpu_trace.h"
36 #include "amdgpu_amdkfd.h"
37 
38 struct amdgpu_sync_entry {
39 	struct hlist_node	node;
40 	struct dma_fence	*fence;
41 };
42 
43 static struct pool amdgpu_sync_slab;
44 
45 /**
46  * amdgpu_sync_create - zero init sync object
47  *
48  * @sync: sync object to initialize
49  *
50  * Just clear the sync object for now.
51  */
52 void amdgpu_sync_create(struct amdgpu_sync *sync)
53 {
54 	hash_init(sync->fences);
55 }
56 
57 /**
58  * amdgpu_sync_same_dev - test if fence belong to us
59  *
60  * @adev: amdgpu device to use for the test
61  * @f: fence to test
62  *
63  * Test if the fence was issued by us.
64  */
65 static bool amdgpu_sync_same_dev(struct amdgpu_device *adev,
66 				 struct dma_fence *f)
67 {
68 	struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
69 
70 	if (s_fence) {
71 		struct amdgpu_ring *ring;
72 
73 		ring = container_of(s_fence->sched, struct amdgpu_ring, sched);
74 		return ring->adev == adev;
75 	}
76 
77 	return false;
78 }
79 
80 /**
81  * amdgpu_sync_get_owner - extract the owner of a fence
82  *
83  * @f: fence get the owner from
84  *
85  * Extract who originally created the fence.
86  */
87 static void *amdgpu_sync_get_owner(struct dma_fence *f)
88 {
89 	struct drm_sched_fence *s_fence;
90 	struct amdgpu_amdkfd_fence *kfd_fence;
91 
92 	if (!f)
93 		return AMDGPU_FENCE_OWNER_UNDEFINED;
94 
95 	s_fence = to_drm_sched_fence(f);
96 	if (s_fence)
97 		return s_fence->owner;
98 
99 	kfd_fence = to_amdgpu_amdkfd_fence(f);
100 	if (kfd_fence)
101 		return AMDGPU_FENCE_OWNER_KFD;
102 
103 	return AMDGPU_FENCE_OWNER_UNDEFINED;
104 }
105 
106 /**
107  * amdgpu_sync_keep_later - Keep the later fence
108  *
109  * @keep: existing fence to test
110  * @fence: new fence
111  *
112  * Either keep the existing fence or the new one, depending which one is later.
113  */
114 static void amdgpu_sync_keep_later(struct dma_fence **keep,
115 				   struct dma_fence *fence)
116 {
117 	if (*keep && dma_fence_is_later(*keep, fence))
118 		return;
119 
120 	dma_fence_put(*keep);
121 	*keep = dma_fence_get(fence);
122 }
123 
124 /**
125  * amdgpu_sync_add_later - add the fence to the hash
126  *
127  * @sync: sync object to add the fence to
128  * @f: fence to add
129  *
130  * Tries to add the fence to an existing hash entry. Returns true when an entry
131  * was found, false otherwise.
132  */
133 static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct dma_fence *f)
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 		return true;
143 	}
144 	return false;
145 }
146 
147 /**
148  * amdgpu_sync_fence - remember to sync to this fence
149  *
150  * @sync: sync object to add fence to
151  * @f: fence to sync to
152  *
153  * Add the fence to the sync object.
154  */
155 int amdgpu_sync_fence(struct amdgpu_sync *sync, struct dma_fence *f)
156 {
157 	struct amdgpu_sync_entry *e;
158 
159 	if (!f)
160 		return 0;
161 
162 	if (amdgpu_sync_add_later(sync, f))
163 		return 0;
164 
165 #ifdef __linux__
166 	e = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
167 #else
168 	e = pool_get(&amdgpu_sync_slab, PR_WAITOK);
169 #endif
170 	if (!e)
171 		return -ENOMEM;
172 
173 	hash_add(sync->fences, &e->node, f->context);
174 	e->fence = dma_fence_get(f);
175 	return 0;
176 }
177 
178 /* Determine based on the owner and mode if we should sync to a fence or not */
179 static bool amdgpu_sync_test_fence(struct amdgpu_device *adev,
180 				   enum amdgpu_sync_mode mode,
181 				   void *owner, struct dma_fence *f)
182 {
183 	void *fence_owner = amdgpu_sync_get_owner(f);
184 
185 	/* Always sync to moves, no matter what */
186 	if (fence_owner == AMDGPU_FENCE_OWNER_UNDEFINED)
187 		return true;
188 
189 	/* We only want to trigger KFD eviction fences on
190 	 * evict or move jobs. Skip KFD fences otherwise.
191 	 */
192 	if (fence_owner == AMDGPU_FENCE_OWNER_KFD &&
193 	    owner != AMDGPU_FENCE_OWNER_UNDEFINED)
194 		return false;
195 
196 	/* Never sync to VM updates either. */
197 	if (fence_owner == AMDGPU_FENCE_OWNER_VM &&
198 	    owner != AMDGPU_FENCE_OWNER_UNDEFINED)
199 		return false;
200 
201 	/* Ignore fences depending on the sync mode */
202 	switch (mode) {
203 	case AMDGPU_SYNC_ALWAYS:
204 		return true;
205 
206 	case AMDGPU_SYNC_NE_OWNER:
207 		if (amdgpu_sync_same_dev(adev, f) &&
208 		    fence_owner == owner)
209 			return false;
210 		break;
211 
212 	case AMDGPU_SYNC_EQ_OWNER:
213 		if (amdgpu_sync_same_dev(adev, f) &&
214 		    fence_owner != owner)
215 			return false;
216 		break;
217 
218 	case AMDGPU_SYNC_EXPLICIT:
219 		return false;
220 	}
221 
222 	WARN(debug_evictions && fence_owner == AMDGPU_FENCE_OWNER_KFD,
223 	     "Adding eviction fence to sync obj");
224 	return true;
225 }
226 
227 /**
228  * amdgpu_sync_resv - sync to a reservation object
229  *
230  * @adev: amdgpu device
231  * @sync: sync object to add fences from reservation object to
232  * @resv: reservation object with embedded fence
233  * @mode: how owner affects which fences we sync to
234  * @owner: owner of the planned job submission
235  *
236  * Sync to the fence
237  */
238 int amdgpu_sync_resv(struct amdgpu_device *adev, struct amdgpu_sync *sync,
239 		     struct dma_resv *resv, enum amdgpu_sync_mode mode,
240 		     void *owner)
241 {
242 	struct dma_resv_iter cursor;
243 	struct dma_fence *f;
244 	int r;
245 
246 	if (resv == NULL)
247 		return -EINVAL;
248 
249 	/* TODO: Use DMA_RESV_USAGE_READ here */
250 	dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP, f) {
251 		dma_fence_chain_for_each(f, f) {
252 			struct dma_fence *tmp = dma_fence_chain_contained(f);
253 
254 			if (amdgpu_sync_test_fence(adev, mode, owner, tmp)) {
255 				r = amdgpu_sync_fence(sync, f);
256 				dma_fence_put(f);
257 				if (r)
258 					return r;
259 				break;
260 			}
261 		}
262 	}
263 	return 0;
264 }
265 
266 /* Free the entry back to the slab */
267 static void amdgpu_sync_entry_free(struct amdgpu_sync_entry *e)
268 {
269 	hash_del(&e->node);
270 	dma_fence_put(e->fence);
271 #ifdef __linux__
272 	kmem_cache_free(amdgpu_sync_slab, e);
273 #else
274 	pool_put(&amdgpu_sync_slab, e);
275 #endif
276 }
277 
278 /**
279  * amdgpu_sync_peek_fence - get the next fence not signaled yet
280  *
281  * @sync: the sync object
282  * @ring: optional ring to use for test
283  *
284  * Returns the next fence not signaled yet without removing it from the sync
285  * object.
286  */
287 struct dma_fence *amdgpu_sync_peek_fence(struct amdgpu_sync *sync,
288 					 struct amdgpu_ring *ring)
289 {
290 	struct amdgpu_sync_entry *e;
291 	struct hlist_node *tmp;
292 	int i;
293 
294 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
295 		struct dma_fence *f = e->fence;
296 		struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
297 
298 		if (dma_fence_is_signaled(f)) {
299 			amdgpu_sync_entry_free(e);
300 			continue;
301 		}
302 		if (ring && s_fence) {
303 			/* For fences from the same ring it is sufficient
304 			 * when they are scheduled.
305 			 */
306 			if (s_fence->sched == &ring->sched) {
307 				if (dma_fence_is_signaled(&s_fence->scheduled))
308 					continue;
309 
310 				return &s_fence->scheduled;
311 			}
312 		}
313 
314 		return f;
315 	}
316 
317 	return NULL;
318 }
319 
320 /**
321  * amdgpu_sync_get_fence - get the next fence from the sync object
322  *
323  * @sync: sync object to use
324  *
325  * Get and removes the next fence from the sync object not signaled yet.
326  */
327 struct dma_fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync)
328 {
329 	struct amdgpu_sync_entry *e;
330 	struct hlist_node *tmp;
331 	struct dma_fence *f;
332 	int i;
333 
334 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
335 
336 		f = e->fence;
337 
338 		hash_del(&e->node);
339 #ifdef __linux__
340 		kmem_cache_free(amdgpu_sync_slab, e);
341 #else
342 		pool_put(&amdgpu_sync_slab, e);
343 #endif
344 
345 		if (!dma_fence_is_signaled(f))
346 			return f;
347 
348 		dma_fence_put(f);
349 	}
350 	return NULL;
351 }
352 
353 /**
354  * amdgpu_sync_clone - clone a sync object
355  *
356  * @source: sync object to clone
357  * @clone: pointer to destination sync object
358  *
359  * Adds references to all unsignaled fences in @source to @clone. Also
360  * removes signaled fences from @source while at it.
361  */
362 int amdgpu_sync_clone(struct amdgpu_sync *source, struct amdgpu_sync *clone)
363 {
364 	struct amdgpu_sync_entry *e;
365 	struct hlist_node *tmp;
366 	struct dma_fence *f;
367 	int i, r;
368 
369 	hash_for_each_safe(source->fences, i, tmp, e, node) {
370 		f = e->fence;
371 		if (!dma_fence_is_signaled(f)) {
372 			r = amdgpu_sync_fence(clone, f);
373 			if (r)
374 				return r;
375 		} else {
376 			amdgpu_sync_entry_free(e);
377 		}
378 	}
379 
380 	return 0;
381 }
382 
383 /**
384  * amdgpu_sync_push_to_job - push fences into job
385  * @sync: sync object to get the fences from
386  * @job: job to push the fences into
387  *
388  * Add all unsignaled fences from sync to job.
389  */
390 int amdgpu_sync_push_to_job(struct amdgpu_sync *sync, struct amdgpu_job *job)
391 {
392 	struct amdgpu_sync_entry *e;
393 	struct hlist_node *tmp;
394 	struct dma_fence *f;
395 	int i, r;
396 
397 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
398 		f = e->fence;
399 		if (dma_fence_is_signaled(f)) {
400 			amdgpu_sync_entry_free(e);
401 			continue;
402 		}
403 
404 		dma_fence_get(f);
405 		r = drm_sched_job_add_dependency(&job->base, f);
406 		if (r) {
407 			dma_fence_put(f);
408 			return r;
409 		}
410 	}
411 	return 0;
412 }
413 
414 int amdgpu_sync_wait(struct amdgpu_sync *sync, bool intr)
415 {
416 	struct amdgpu_sync_entry *e;
417 	struct hlist_node *tmp;
418 	int i, r;
419 
420 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
421 		r = dma_fence_wait(e->fence, intr);
422 		if (r)
423 			return r;
424 
425 		amdgpu_sync_entry_free(e);
426 	}
427 
428 	return 0;
429 }
430 
431 /**
432  * amdgpu_sync_free - free the sync object
433  *
434  * @sync: sync object to use
435  *
436  * Free the sync object.
437  */
438 void amdgpu_sync_free(struct amdgpu_sync *sync)
439 {
440 	struct amdgpu_sync_entry *e;
441 	struct hlist_node *tmp;
442 	unsigned int i;
443 
444 	hash_for_each_safe(sync->fences, i, tmp, e, node)
445 		amdgpu_sync_entry_free(e);
446 }
447 
448 /**
449  * amdgpu_sync_init - init sync object subsystem
450  *
451  * Allocate the slab allocator.
452  */
453 int amdgpu_sync_init(void)
454 {
455 #ifdef __linux__
456 	amdgpu_sync_slab = kmem_cache_create(
457 		"amdgpu_sync", sizeof(struct amdgpu_sync_entry), 0,
458 		SLAB_HWCACHE_ALIGN, NULL);
459 	if (!amdgpu_sync_slab)
460 		return -ENOMEM;
461 #else
462 	pool_init(&amdgpu_sync_slab, sizeof(struct amdgpu_sync_entry),
463 	    CACHELINESIZE, IPL_TTY, 0, "amdgpu_sync", NULL);
464 #endif
465 
466 	return 0;
467 }
468 
469 /**
470  * amdgpu_sync_fini - fini sync object subsystem
471  *
472  * Free the slab allocator.
473  */
474 void amdgpu_sync_fini(void)
475 {
476 #ifdef __linux__
477 	kmem_cache_destroy(amdgpu_sync_slab);
478 #else
479 	pool_destroy(&amdgpu_sync_slab);
480 #endif
481 }
482