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 * @mode: how owner affects which fences we sync to 206 * @owner: owner of the planned job submission 207 * 208 * Sync to the fence 209 */ 210 int amdgpu_sync_resv(struct amdgpu_device *adev, struct amdgpu_sync *sync, 211 struct dma_resv *resv, enum amdgpu_sync_mode mode, 212 void *owner) 213 { 214 struct dma_resv_list *flist; 215 struct dma_fence *f; 216 unsigned i; 217 int r = 0; 218 219 if (resv == NULL) 220 return -EINVAL; 221 222 /* always sync to the exclusive fence */ 223 f = dma_resv_get_excl(resv); 224 r = amdgpu_sync_fence(sync, f, false); 225 226 flist = dma_resv_get_list(resv); 227 if (!flist || r) 228 return r; 229 230 for (i = 0; i < flist->shared_count; ++i) { 231 void *fence_owner; 232 233 f = rcu_dereference_protected(flist->shared[i], 234 dma_resv_held(resv)); 235 236 fence_owner = amdgpu_sync_get_owner(f); 237 238 /* Always sync to moves, no matter what */ 239 if (fence_owner == AMDGPU_FENCE_OWNER_UNDEFINED) { 240 r = amdgpu_sync_fence(sync, f, false); 241 if (r) 242 break; 243 } 244 245 /* We only want to trigger KFD eviction fences on 246 * evict or move jobs. Skip KFD fences otherwise. 247 */ 248 if (fence_owner == AMDGPU_FENCE_OWNER_KFD && 249 owner != AMDGPU_FENCE_OWNER_UNDEFINED) 250 continue; 251 252 /* Ignore fences depending on the sync mode */ 253 switch (mode) { 254 case AMDGPU_SYNC_ALWAYS: 255 break; 256 257 case AMDGPU_SYNC_NE_OWNER: 258 if (amdgpu_sync_same_dev(adev, f) && 259 fence_owner == owner) 260 continue; 261 break; 262 263 case AMDGPU_SYNC_EQ_OWNER: 264 if (amdgpu_sync_same_dev(adev, f) && 265 fence_owner != owner) 266 continue; 267 break; 268 269 case AMDGPU_SYNC_EXPLICIT: 270 continue; 271 } 272 273 r = amdgpu_sync_fence(sync, f, false); 274 if (r) 275 break; 276 } 277 return r; 278 } 279 280 /** 281 * amdgpu_sync_peek_fence - get the next fence not signaled yet 282 * 283 * @sync: the sync object 284 * @ring: optional ring to use for test 285 * 286 * Returns the next fence not signaled yet without removing it from the sync 287 * object. 288 */ 289 struct dma_fence *amdgpu_sync_peek_fence(struct amdgpu_sync *sync, 290 struct amdgpu_ring *ring) 291 { 292 struct amdgpu_sync_entry *e; 293 struct hlist_node *tmp; 294 int i; 295 296 hash_for_each_safe(sync->fences, i, tmp, e, node) { 297 struct dma_fence *f = e->fence; 298 struct drm_sched_fence *s_fence = to_drm_sched_fence(f); 299 300 if (dma_fence_is_signaled(f)) { 301 hash_del(&e->node); 302 dma_fence_put(f); 303 kmem_cache_free(amdgpu_sync_slab, e); 304 continue; 305 } 306 if (ring && s_fence) { 307 /* For fences from the same ring it is sufficient 308 * when they are scheduled. 309 */ 310 if (s_fence->sched == &ring->sched) { 311 if (dma_fence_is_signaled(&s_fence->scheduled)) 312 continue; 313 314 return &s_fence->scheduled; 315 } 316 } 317 318 return f; 319 } 320 321 return NULL; 322 } 323 324 /** 325 * amdgpu_sync_get_fence - get the next fence from the sync object 326 * 327 * @sync: sync object to use 328 * @explicit: true if the next fence is explicit 329 * 330 * Get and removes the next fence from the sync object not signaled yet. 331 */ 332 struct dma_fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync, bool *explicit) 333 { 334 struct amdgpu_sync_entry *e; 335 struct hlist_node *tmp; 336 struct dma_fence *f; 337 int i; 338 hash_for_each_safe(sync->fences, i, tmp, e, node) { 339 340 f = e->fence; 341 if (explicit) 342 *explicit = e->explicit; 343 344 hash_del(&e->node); 345 kmem_cache_free(amdgpu_sync_slab, e); 346 347 if (!dma_fence_is_signaled(f)) 348 return f; 349 350 dma_fence_put(f); 351 } 352 return NULL; 353 } 354 355 /** 356 * amdgpu_sync_clone - clone a sync object 357 * 358 * @source: sync object to clone 359 * @clone: pointer to destination sync object 360 * 361 * Adds references to all unsignaled fences in @source to @clone. Also 362 * removes signaled fences from @source while at it. 363 */ 364 int amdgpu_sync_clone(struct amdgpu_sync *source, struct amdgpu_sync *clone) 365 { 366 struct amdgpu_sync_entry *e; 367 struct hlist_node *tmp; 368 struct dma_fence *f; 369 int i, r; 370 371 hash_for_each_safe(source->fences, i, tmp, e, node) { 372 f = e->fence; 373 if (!dma_fence_is_signaled(f)) { 374 r = amdgpu_sync_fence(clone, f, e->explicit); 375 if (r) 376 return r; 377 } else { 378 hash_del(&e->node); 379 dma_fence_put(f); 380 kmem_cache_free(amdgpu_sync_slab, e); 381 } 382 } 383 384 dma_fence_put(clone->last_vm_update); 385 clone->last_vm_update = dma_fence_get(source->last_vm_update); 386 387 return 0; 388 } 389 390 int amdgpu_sync_wait(struct amdgpu_sync *sync, bool intr) 391 { 392 struct amdgpu_sync_entry *e; 393 struct hlist_node *tmp; 394 int i, r; 395 396 hash_for_each_safe(sync->fences, i, tmp, e, node) { 397 r = dma_fence_wait(e->fence, intr); 398 if (r) 399 return r; 400 401 hash_del(&e->node); 402 dma_fence_put(e->fence); 403 kmem_cache_free(amdgpu_sync_slab, e); 404 } 405 406 return 0; 407 } 408 409 /** 410 * amdgpu_sync_free - free the sync object 411 * 412 * @sync: sync object to use 413 * 414 * Free the sync object. 415 */ 416 void amdgpu_sync_free(struct amdgpu_sync *sync) 417 { 418 struct amdgpu_sync_entry *e; 419 struct hlist_node *tmp; 420 unsigned i; 421 422 hash_for_each_safe(sync->fences, i, tmp, e, node) { 423 hash_del(&e->node); 424 dma_fence_put(e->fence); 425 kmem_cache_free(amdgpu_sync_slab, e); 426 } 427 428 dma_fence_put(sync->last_vm_update); 429 } 430 431 /** 432 * amdgpu_sync_init - init sync object subsystem 433 * 434 * Allocate the slab allocator. 435 */ 436 int amdgpu_sync_init(void) 437 { 438 amdgpu_sync_slab = kmem_cache_create( 439 "amdgpu_sync", sizeof(struct amdgpu_sync_entry), 0, 440 SLAB_HWCACHE_ALIGN, NULL); 441 if (!amdgpu_sync_slab) 442 return -ENOMEM; 443 444 return 0; 445 } 446 447 /** 448 * amdgpu_sync_fini - fini sync object subsystem 449 * 450 * Free the slab allocator. 451 */ 452 void amdgpu_sync_fini(void) 453 { 454 kmem_cache_destroy(amdgpu_sync_slab); 455 } 456