projects/linux·blob master · refs

514 lines · 11.6 KiB

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 <[email protected]>
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
38struct amdgpu_sync_entry {
39 struct hlist_node node;
40 struct dma_fence *fence;
41};
42
43static struct kmem_cache *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 */
52void 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 */
65static 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 */
87static 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 */
114static 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 */
133static 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 (dma_fence_is_signaled(e->fence)) {
139 dma_fence_put(e->fence);
140 e->fence = dma_fence_get(f);
141 return true;
142 }
143
144 if (likely(e->fence->context == f->context)) {
145 amdgpu_sync_keep_later(&e->fence, f);
146 return true;
147 }
148 }
149 return false;
150}
151
152/**
153 * amdgpu_sync_fence - remember to sync to this fence
154 *
155 * @sync: sync object to add fence to
156 * @f: fence to sync to
157 * @flags: memory allocation flags to use when allocating sync entry
158 *
159 * Add the fence to the sync object.
160 */
161int amdgpu_sync_fence(struct amdgpu_sync *sync, struct dma_fence *f,
162 gfp_t flags)
163{
164 struct amdgpu_sync_entry *e;
165
166 if (!f)
167 return 0;
168
169 if (amdgpu_sync_add_later(sync, f))
170 return 0;
171
172 e = kmem_cache_alloc(amdgpu_sync_slab, flags);
173 if (!e)
174 return -ENOMEM;
175
176 hash_add(sync->fences, &e->node, f->context);
177 e->fence = dma_fence_get(f);
178 return 0;
179}
180
181/* Determine based on the owner and mode if we should sync to a fence or not */
182static bool amdgpu_sync_test_fence(struct amdgpu_device *adev,
183 enum amdgpu_sync_mode mode,
184 void *owner, struct dma_fence *f)
185{
186 void *fence_owner = amdgpu_sync_get_owner(f);
187
188 /* Always sync to moves, no matter what */
189 if (fence_owner == AMDGPU_FENCE_OWNER_UNDEFINED)
190 return true;
191
192 /* We only want to trigger KFD eviction fences on
193 * evict or move jobs. Skip KFD fences otherwise.
194 */
195 if (fence_owner == AMDGPU_FENCE_OWNER_KFD &&
196 owner != AMDGPU_FENCE_OWNER_UNDEFINED)
197 return false;
198
199 /* Never sync to VM updates either. */
200 if (fence_owner == AMDGPU_FENCE_OWNER_VM &&
201 owner != AMDGPU_FENCE_OWNER_UNDEFINED &&
202 owner != AMDGPU_FENCE_OWNER_KFD)
203 return false;
204
205 /* Ignore fences depending on the sync mode */
206 switch (mode) {
207 case AMDGPU_SYNC_ALWAYS:
208 return true;
209
210 case AMDGPU_SYNC_NE_OWNER:
211 if (amdgpu_sync_same_dev(adev, f) &&
212 fence_owner == owner)
213 return false;
214 break;
215
216 case AMDGPU_SYNC_EQ_OWNER:
217 if (amdgpu_sync_same_dev(adev, f) &&
218 fence_owner != owner)
219 return false;
220 break;
221
222 case AMDGPU_SYNC_EXPLICIT:
223 return false;
224 }
225
226 WARN(debug_evictions && fence_owner == AMDGPU_FENCE_OWNER_KFD,
227 "Adding eviction fence to sync obj");
228 return true;
229}
230
231/**
232 * amdgpu_sync_resv - sync to a reservation object
233 *
234 * @adev: amdgpu device
235 * @sync: sync object to add fences from reservation object to
236 * @resv: reservation object with embedded fence
237 * @mode: how owner affects which fences we sync to
238 * @owner: owner of the planned job submission
239 *
240 * Sync to the fence
241 */
242int amdgpu_sync_resv(struct amdgpu_device *adev, struct amdgpu_sync *sync,
243 struct dma_resv *resv, enum amdgpu_sync_mode mode,
244 void *owner)
245{
246 struct dma_resv_iter cursor;
247 struct dma_fence *f;
248 int r;
249
250 if (resv == NULL)
251 return -EINVAL;
252 /* Implicitly sync only to KERNEL, WRITE and READ */
253 dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_READ, f) {
254 dma_fence_chain_for_each(f, f) {
255 struct dma_fence *tmp = dma_fence_chain_contained(f);
256
257 if (amdgpu_sync_test_fence(adev, mode, owner, tmp)) {
258 r = amdgpu_sync_fence(sync, f, GFP_KERNEL);
259 dma_fence_put(f);
260 if (r)
261 return r;
262 break;
263 }
264 }
265 }
266 return 0;
267}
268
269/**
270 * amdgpu_sync_kfd - sync to KFD fences
271 *
272 * @sync: sync object to add KFD fences to
273 * @resv: reservation object with KFD fences
274 *
275 * Extract all KFD fences and add them to the sync object.
276 */
277int amdgpu_sync_kfd(struct amdgpu_sync *sync, struct dma_resv *resv)
278{
279 struct dma_resv_iter cursor;
280 struct dma_fence *f;
281 int r = 0;
282
283 dma_resv_iter_begin(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP);
284 dma_resv_for_each_fence_unlocked(&cursor, f) {
285 void *fence_owner = amdgpu_sync_get_owner(f);
286
287 if (fence_owner != AMDGPU_FENCE_OWNER_KFD)
288 continue;
289
290 r = amdgpu_sync_fence(sync, f, GFP_KERNEL);
291 if (r)
292 break;
293 }
294 dma_resv_iter_end(&cursor);
295
296 return r;
297}
298
299/* Free the entry back to the slab */
300static void amdgpu_sync_entry_free(struct amdgpu_sync_entry *e)
301{
302 hash_del(&e->node);
303 dma_fence_put(e->fence);
304 kmem_cache_free(amdgpu_sync_slab, e);
305}
306
307/**
308 * amdgpu_sync_peek_fence - get the next fence not signaled yet
309 *
310 * @sync: the sync object
311 * @ring: optional ring to use for test
312 *
313 * Returns the next fence not signaled yet without removing it from the sync
314 * object.
315 */
316struct dma_fence *amdgpu_sync_peek_fence(struct amdgpu_sync *sync,
317 struct amdgpu_ring *ring)
318{
319 struct amdgpu_sync_entry *e;
320 struct hlist_node *tmp;
321 int i;
322
323 hash_for_each_safe(sync->fences, i, tmp, e, node) {
324 struct dma_fence *f = e->fence;
325 struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
326
327 if (dma_fence_is_signaled(f)) {
328 amdgpu_sync_entry_free(e);
329 continue;
330 }
331 if (ring && s_fence) {
332 /* For fences from the same ring it is sufficient
333 * when they are scheduled.
334 */
335 if (s_fence->sched == &ring->sched) {
336 if (dma_fence_is_signaled(&s_fence->scheduled))
337 continue;
338
339 return &s_fence->scheduled;
340 }
341 }
342
343 return f;
344 }
345
346 return NULL;
347}
348
349/**
350 * amdgpu_sync_get_fence - get the next fence from the sync object
351 *
352 * @sync: sync object to use
353 *
354 * Get and removes the next fence from the sync object not signaled yet.
355 */
356struct dma_fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync)
357{
358 struct amdgpu_sync_entry *e;
359 struct hlist_node *tmp;
360 struct dma_fence *f;
361 int i;
362
363 hash_for_each_safe(sync->fences, i, tmp, e, node) {
364
365 f = e->fence;
366
367 hash_del(&e->node);
368 kmem_cache_free(amdgpu_sync_slab, e);
369
370 if (!dma_fence_is_signaled(f))
371 return f;
372
373 dma_fence_put(f);
374 }
375 return NULL;
376}
377
378/**
379 * amdgpu_sync_clone - clone a sync object
380 *
381 * @source: sync object to clone
382 * @clone: pointer to destination sync object
383 *
384 * Adds references to all unsignaled fences in @source to @clone. Also
385 * removes signaled fences from @source while at it.
386 */
387int amdgpu_sync_clone(struct amdgpu_sync *source, struct amdgpu_sync *clone)
388{
389 struct amdgpu_sync_entry *e;
390 struct hlist_node *tmp;
391 struct dma_fence *f;
392 int i, r;
393
394 hash_for_each_safe(source->fences, i, tmp, e, node) {
395 f = e->fence;
396 if (!dma_fence_is_signaled(f)) {
397 r = amdgpu_sync_fence(clone, f, GFP_KERNEL);
398 if (r)
399 return r;
400 } else {
401 amdgpu_sync_entry_free(e);
402 }
403 }
404
405 return 0;
406}
407
408/**
409 * amdgpu_sync_move - move all fences from src to dst
410 *
411 * @src: source of the fences, empty after function
412 * @dst: destination for the fences
413 *
414 * Moves all fences from source to destination. All fences in destination are
415 * freed and source is empty after the function call.
416 */
417void amdgpu_sync_move(struct amdgpu_sync *src, struct amdgpu_sync *dst)
418{
419 unsigned int i;
420
421 amdgpu_sync_free(dst);
422
423 for (i = 0; i < HASH_SIZE(src->fences); ++i)
424 hlist_move_list(&src->fences[i], &dst->fences[i]);
425}
426
427/**
428 * amdgpu_sync_push_to_job - push fences into job
429 * @sync: sync object to get the fences from
430 * @job: job to push the fences into
431 *
432 * Add all unsignaled fences from sync to job.
433 */
434int amdgpu_sync_push_to_job(struct amdgpu_sync *sync, struct amdgpu_job *job)
435{
436 struct amdgpu_sync_entry *e;
437 struct hlist_node *tmp;
438 struct dma_fence *f;
439 int i, r;
440
441 hash_for_each_safe(sync->fences, i, tmp, e, node) {
442 f = e->fence;
443 if (dma_fence_is_signaled(f)) {
444 amdgpu_sync_entry_free(e);
445 continue;
446 }
447
448 dma_fence_get(f);
449 r = drm_sched_job_add_dependency(&job->base, f);
450 if (r) {
451 dma_fence_put(f);
452 return r;
453 }
454 }
455 return 0;
456}
457
458int amdgpu_sync_wait(struct amdgpu_sync *sync, bool intr)
459{
460 struct amdgpu_sync_entry *e;
461 struct hlist_node *tmp;
462 int i, r;
463
464 hash_for_each_safe(sync->fences, i, tmp, e, node) {
465 r = dma_fence_wait(e->fence, intr);
466 if (r)
467 return r;
468
469 amdgpu_sync_entry_free(e);
470 }
471
472 return 0;
473}
474
475/**
476 * amdgpu_sync_free - free the sync object
477 *
478 * @sync: sync object to use
479 *
480 * Free the sync object.
481 */
482void amdgpu_sync_free(struct amdgpu_sync *sync)
483{
484 struct amdgpu_sync_entry *e;
485 struct hlist_node *tmp;
486 unsigned int i;
487
488 hash_for_each_safe(sync->fences, i, tmp, e, node)
489 amdgpu_sync_entry_free(e);
490}
491
492/**
493 * amdgpu_sync_init - init sync object subsystem
494 *
495 * Allocate the slab allocator.
496 */
497int amdgpu_sync_init(void)
498{
499 amdgpu_sync_slab = KMEM_CACHE(amdgpu_sync_entry, SLAB_HWCACHE_ALIGN);
500 if (!amdgpu_sync_slab)
501 return -ENOMEM;
502
503 return 0;
504}
505
506/**
507 * amdgpu_sync_fini - fini sync object subsystem
508 *
509 * Free the slab allocator.
510 */
511void amdgpu_sync_fini(void)
512{
513 kmem_cache_destroy(amdgpu_sync_slab);
514}