amdgpu_sync.c 10.9 KB
Newer Older
1
// SPDX-License-Identifier: MIT
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
/*
 * Copyright 2014 Advanced Micro Devices, Inc.
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the
 * "Software"), to deal in the Software without restriction, including
 * without limitation the rights to use, copy, modify, merge, publish,
 * distribute, sub license, and/or sell copies of the Software, and to
 * permit persons to whom the Software is furnished to do so, subject to
 * the following conditions:
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 *
 * The above copyright notice and this permission notice (including the
 * next paragraph) shall be included in all copies or substantial portions
 * of the Software.
 *
 */
/*
 * Authors:
 *    Christian König <christian.koenig@amd.com>
 */

32 33
#include <linux/dma-fence-chain.h>

34 35
#include "amdgpu.h"
#include "amdgpu_trace.h"
36
#include "amdgpu_amdkfd.h"
37

38 39
struct amdgpu_sync_entry {
	struct hlist_node	node;
40
	struct dma_fence	*fence;
41 42
};

43 44
static struct kmem_cache *amdgpu_sync_slab;

45 46 47 48 49 50 51 52 53
/**
 * amdgpu_sync_create - zero init sync object
 *
 * @sync: sync object to initialize
 *
 * Just clear the sync object for now.
 */
void amdgpu_sync_create(struct amdgpu_sync *sync)
{
54
	hash_init(sync->fences);
55 56
}

57 58 59 60 61 62 63 64
/**
 * amdgpu_sync_same_dev - test if fence belong to us
 *
 * @adev: amdgpu device to use for the test
 * @f: fence to test
 *
 * Test if the fence was issued by us.
 */
65 66
static bool amdgpu_sync_same_dev(struct amdgpu_device *adev,
				 struct dma_fence *f)
67
{
68
	struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
69

70 71 72 73 74 75 76
	if (s_fence) {
		struct amdgpu_ring *ring;

		ring = container_of(s_fence->sched, struct amdgpu_ring, sched);
		return ring->adev == adev;
	}

77 78 79
	return false;
}

80 81 82
/**
 * amdgpu_sync_get_owner - extract the owner of a fence
 *
83
 * @f: fence get the owner from
84 85 86
 *
 * Extract who originally created the fence.
 */
87
static void *amdgpu_sync_get_owner(struct dma_fence *f)
88
{
89 90 91 92 93
	struct drm_sched_fence *s_fence;
	struct amdgpu_amdkfd_fence *kfd_fence;

	if (!f)
		return AMDGPU_FENCE_OWNER_UNDEFINED;
94

95
	s_fence = to_drm_sched_fence(f);
96
	if (s_fence)
97
		return s_fence->owner;
98

99 100 101 102
	kfd_fence = to_amdgpu_amdkfd_fence(f);
	if (kfd_fence)
		return AMDGPU_FENCE_OWNER_KFD;

103
	return AMDGPU_FENCE_OWNER_UNDEFINED;
104 105
}

106 107 108 109 110 111 112 113
/**
 * amdgpu_sync_keep_later - Keep the later fence
 *
 * @keep: existing fence to test
 * @fence: new fence
 *
 * Either keep the existing fence or the new one, depending which one is later.
 */
114 115
static void amdgpu_sync_keep_later(struct dma_fence **keep,
				   struct dma_fence *fence)
116
{
117
	if (*keep && dma_fence_is_later(*keep, fence))
118 119
		return;

120 121
	dma_fence_put(*keep);
	*keep = dma_fence_get(fence);
122 123
}

124 125 126 127 128 129 130 131 132
/**
 * amdgpu_sync_add_later - add the fence to the hash
 *
 * @sync: sync object to add the fence to
 * @f: fence to add
 *
 * Tries to add the fence to an existing hash entry. Returns true when an entry
 * was found, false otherwise.
 */
133
static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct dma_fence *f)
134 135 136 137 138 139 140 141 142 143 144 145 146
{
	struct amdgpu_sync_entry *e;

	hash_for_each_possible(sync->fences, e, node, f->context) {
		if (unlikely(e->fence->context != f->context))
			continue;

		amdgpu_sync_keep_later(&e->fence, f);
		return true;
	}
	return false;
}

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

159 160
	if (!f)
		return 0;
161

162
	if (amdgpu_sync_add_later(sync, f))
163
		return 0;
164

165
	e = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
166 167
	if (!e)
		return -ENOMEM;
168

169
	hash_add(sync->fences, &e->node, f->context);
170
	e->fence = dma_fence_get(f);
171
	return 0;
172 173
}

174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193
/* Determine based on the owner and mode if we should sync to a fence or not */
static bool amdgpu_sync_test_fence(struct amdgpu_device *adev,
				   enum amdgpu_sync_mode mode,
				   void *owner, struct dma_fence *f)
{
	void *fence_owner = amdgpu_sync_get_owner(f);

	/* Always sync to moves, no matter what */
	if (fence_owner == AMDGPU_FENCE_OWNER_UNDEFINED)
		return true;

	/* We only want to trigger KFD eviction fences on
	 * evict or move jobs. Skip KFD fences otherwise.
	 */
	if (fence_owner == AMDGPU_FENCE_OWNER_KFD &&
	    owner != AMDGPU_FENCE_OWNER_UNDEFINED)
		return false;

	/* Never sync to VM updates either. */
	if (fence_owner == AMDGPU_FENCE_OWNER_VM &&
194 195
	    owner != AMDGPU_FENCE_OWNER_UNDEFINED &&
	    owner != AMDGPU_FENCE_OWNER_KFD)
196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
		return false;

	/* Ignore fences depending on the sync mode */
	switch (mode) {
	case AMDGPU_SYNC_ALWAYS:
		return true;

	case AMDGPU_SYNC_NE_OWNER:
		if (amdgpu_sync_same_dev(adev, f) &&
		    fence_owner == owner)
			return false;
		break;

	case AMDGPU_SYNC_EQ_OWNER:
		if (amdgpu_sync_same_dev(adev, f) &&
		    fence_owner != owner)
			return false;
		break;

	case AMDGPU_SYNC_EXPLICIT:
		return false;
	}

	WARN(debug_evictions && fence_owner == AMDGPU_FENCE_OWNER_KFD,
	     "Adding eviction fence to sync obj");
	return true;
}

224
/**
225
 * amdgpu_sync_resv - sync to a reservation object
226
 *
227
 * @adev: amdgpu device
228 229
 * @sync: sync object to add fences from reservation object to
 * @resv: reservation object with embedded fence
230 231
 * @mode: how owner affects which fences we sync to
 * @owner: owner of the planned job submission
232
 *
233
 * Sync to the fence
234
 */
235 236 237
int amdgpu_sync_resv(struct amdgpu_device *adev, struct amdgpu_sync *sync,
		     struct dma_resv *resv, enum amdgpu_sync_mode mode,
		     void *owner)
238
{
239
	struct dma_resv_iter cursor;
240
	struct dma_fence *f;
241
	int r;
242

243 244 245
	if (resv == NULL)
		return -EINVAL;

246
	/* TODO: Use DMA_RESV_USAGE_READ here */
247
	dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP, f) {
248
		dma_fence_chain_for_each(f, f) {
249
			struct dma_fence *tmp = dma_fence_chain_contained(f);
250

251
			if (amdgpu_sync_test_fence(adev, mode, owner, tmp)) {
252 253 254 255 256 257
				r = amdgpu_sync_fence(sync, f);
				dma_fence_put(f);
				if (r)
					return r;
				break;
			}
258
		}
259
	}
260
	return 0;
261 262
}

263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
/**
 * amdgpu_sync_kfd - sync to KFD fences
 *
 * @sync: sync object to add KFD fences to
 * @resv: reservation object with KFD fences
 *
 * Extract all KFD fences and add them to the sync object.
 */
int amdgpu_sync_kfd(struct amdgpu_sync *sync, struct dma_resv *resv)
{
	struct dma_resv_iter cursor;
	struct dma_fence *f;
	int r = 0;

	dma_resv_iter_begin(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP);
	dma_resv_for_each_fence_unlocked(&cursor, f) {
		void *fence_owner = amdgpu_sync_get_owner(f);

		if (fence_owner != AMDGPU_FENCE_OWNER_KFD)
			continue;

		r = amdgpu_sync_fence(sync, f);
		if (r)
			break;
	}
	dma_resv_iter_end(&cursor);

	return r;
}

293 294 295 296 297 298 299 300
/* Free the entry back to the slab */
static void amdgpu_sync_entry_free(struct amdgpu_sync_entry *e)
{
	hash_del(&e->node);
	dma_fence_put(e->fence);
	kmem_cache_free(amdgpu_sync_slab, e);
}

301
/**
302
 * amdgpu_sync_peek_fence - get the next fence not signaled yet
303 304
 *
 * @sync: the sync object
305
 * @ring: optional ring to use for test
306
 *
307 308
 * Returns the next fence not signaled yet without removing it from the sync
 * object.
309
 */
310 311
struct dma_fence *amdgpu_sync_peek_fence(struct amdgpu_sync *sync,
					 struct amdgpu_ring *ring)
312 313 314 315 316 317
{
	struct amdgpu_sync_entry *e;
	struct hlist_node *tmp;
	int i;

	hash_for_each_safe(sync->fences, i, tmp, e, node) {
318
		struct dma_fence *f = e->fence;
319
		struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
320

321
		if (dma_fence_is_signaled(f)) {
322
			amdgpu_sync_entry_free(e);
323 324
			continue;
		}
325 326 327 328
		if (ring && s_fence) {
			/* For fences from the same ring it is sufficient
			 * when they are scheduled.
			 */
329
			if (s_fence->sched == &ring->sched) {
330
				if (dma_fence_is_signaled(&s_fence->scheduled))
331
					continue;
332

333 334
				return &s_fence->scheduled;
			}
335 336
		}

337
		return f;
338 339
	}

340
	return NULL;
341 342
}

343 344 345 346 347 348 349
/**
 * amdgpu_sync_get_fence - get the next fence from the sync object
 *
 * @sync: sync object to use
 *
 * Get and removes the next fence from the sync object not signaled yet.
 */
350
struct dma_fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync)
351 352 353
{
	struct amdgpu_sync_entry *e;
	struct hlist_node *tmp;
354
	struct dma_fence *f;
355
	int i;
356

357 358 359 360 361
	hash_for_each_safe(sync->fences, i, tmp, e, node) {

		f = e->fence;

		hash_del(&e->node);
362
		kmem_cache_free(amdgpu_sync_slab, e);
363

364
		if (!dma_fence_is_signaled(f))
365 366
			return f;

367
		dma_fence_put(f);
368 369 370 371
	}
	return NULL;
}

372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390
/**
 * amdgpu_sync_clone - clone a sync object
 *
 * @source: sync object to clone
 * @clone: pointer to destination sync object
 *
 * Adds references to all unsignaled fences in @source to @clone. Also
 * removes signaled fences from @source while at it.
 */
int amdgpu_sync_clone(struct amdgpu_sync *source, struct amdgpu_sync *clone)
{
	struct amdgpu_sync_entry *e;
	struct hlist_node *tmp;
	struct dma_fence *f;
	int i, r;

	hash_for_each_safe(source->fences, i, tmp, e, node) {
		f = e->fence;
		if (!dma_fence_is_signaled(f)) {
391
			r = amdgpu_sync_fence(clone, f);
392 393 394
			if (r)
				return r;
		} else {
395
			amdgpu_sync_entry_free(e);
396 397 398 399 400 401
		}
	}

	return 0;
}

402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
/**
 * amdgpu_sync_push_to_job - push fences into job
 * @sync: sync object to get the fences from
 * @job: job to push the fences into
 *
 * Add all unsignaled fences from sync to job.
 */
int amdgpu_sync_push_to_job(struct amdgpu_sync *sync, struct amdgpu_job *job)
{
	struct amdgpu_sync_entry *e;
	struct hlist_node *tmp;
	struct dma_fence *f;
	int i, r;

	hash_for_each_safe(sync->fences, i, tmp, e, node) {
		f = e->fence;
		if (dma_fence_is_signaled(f)) {
			amdgpu_sync_entry_free(e);
			continue;
		}

		dma_fence_get(f);
		r = drm_sched_job_add_dependency(&job->base, f);
425 426
		if (r) {
			dma_fence_put(f);
427
			return r;
428
		}
429 430 431 432
	}
	return 0;
}

433 434 435 436 437 438 439 440 441 442 443
int amdgpu_sync_wait(struct amdgpu_sync *sync, bool intr)
{
	struct amdgpu_sync_entry *e;
	struct hlist_node *tmp;
	int i, r;

	hash_for_each_safe(sync->fences, i, tmp, e, node) {
		r = dma_fence_wait(e->fence, intr);
		if (r)
			return r;

444
		amdgpu_sync_entry_free(e);
445 446 447 448 449
	}

	return 0;
}

450 451 452 453 454
/**
 * amdgpu_sync_free - free the sync object
 *
 * @sync: sync object to use
 *
455
 * Free the sync object.
456
 */
457
void amdgpu_sync_free(struct amdgpu_sync *sync)
458
{
459 460
	struct amdgpu_sync_entry *e;
	struct hlist_node *tmp;
461
	unsigned int i;
462

463 464
	hash_for_each_safe(sync->fences, i, tmp, e, node)
		amdgpu_sync_entry_free(e);
465
}
466 467 468 469 470 471 472 473

/**
 * amdgpu_sync_init - init sync object subsystem
 *
 * Allocate the slab allocator.
 */
int amdgpu_sync_init(void)
{
474
	amdgpu_sync_slab = KMEM_CACHE(amdgpu_sync_entry, SLAB_HWCACHE_ALIGN);
475 476 477 478 479 480 481 482 483 484 485 486 487 488 489
	if (!amdgpu_sync_slab)
		return -ENOMEM;

	return 0;
}

/**
 * amdgpu_sync_fini - fini sync object subsystem
 *
 * Free the slab allocator.
 */
void amdgpu_sync_fini(void)
{
	kmem_cache_destroy(amdgpu_sync_slab);
}