1 /* 2 * Copyright 2015 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 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 NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 * Authors: monk liu <monk.liu@amd.com> 23 */ 24 25 #include <drm/drm_auth.h> 26 #include <drm/drm_drv.h> 27 #include "amdgpu.h" 28 #include "amdgpu_sched.h" 29 #include "amdgpu_ras.h" 30 #include <linux/nospec.h> 31 32 #define to_amdgpu_ctx_entity(e) \ 33 container_of((e), struct amdgpu_ctx_entity, entity) 34 35 const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM] = { 36 [AMDGPU_HW_IP_GFX] = 1, 37 [AMDGPU_HW_IP_COMPUTE] = 4, 38 [AMDGPU_HW_IP_DMA] = 2, 39 [AMDGPU_HW_IP_UVD] = 1, 40 [AMDGPU_HW_IP_VCE] = 1, 41 [AMDGPU_HW_IP_UVD_ENC] = 1, 42 [AMDGPU_HW_IP_VCN_DEC] = 1, 43 [AMDGPU_HW_IP_VCN_ENC] = 1, 44 [AMDGPU_HW_IP_VCN_JPEG] = 1, 45 }; 46 47 bool amdgpu_ctx_priority_is_valid(int32_t ctx_prio) 48 { 49 switch (ctx_prio) { 50 case AMDGPU_CTX_PRIORITY_UNSET: 51 case AMDGPU_CTX_PRIORITY_VERY_LOW: 52 case AMDGPU_CTX_PRIORITY_LOW: 53 case AMDGPU_CTX_PRIORITY_NORMAL: 54 case AMDGPU_CTX_PRIORITY_HIGH: 55 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 56 return true; 57 default: 58 return false; 59 } 60 } 61 62 static enum drm_sched_priority 63 amdgpu_ctx_to_drm_sched_prio(int32_t ctx_prio) 64 { 65 switch (ctx_prio) { 66 case AMDGPU_CTX_PRIORITY_UNSET: 67 return DRM_SCHED_PRIORITY_UNSET; 68 69 case AMDGPU_CTX_PRIORITY_VERY_LOW: 70 return DRM_SCHED_PRIORITY_MIN; 71 72 case AMDGPU_CTX_PRIORITY_LOW: 73 return DRM_SCHED_PRIORITY_MIN; 74 75 case AMDGPU_CTX_PRIORITY_NORMAL: 76 return DRM_SCHED_PRIORITY_NORMAL; 77 78 case AMDGPU_CTX_PRIORITY_HIGH: 79 return DRM_SCHED_PRIORITY_HIGH; 80 81 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 82 return DRM_SCHED_PRIORITY_HIGH; 83 84 /* This should not happen as we sanitized userspace provided priority 85 * already, WARN if this happens. 86 */ 87 default: 88 WARN(1, "Invalid context priority %d\n", ctx_prio); 89 return DRM_SCHED_PRIORITY_NORMAL; 90 } 91 92 } 93 94 static int amdgpu_ctx_priority_permit(struct drm_file *filp, 95 int32_t priority) 96 { 97 if (!amdgpu_ctx_priority_is_valid(priority)) 98 return -EINVAL; 99 100 /* NORMAL and below are accessible by everyone */ 101 if (priority <= AMDGPU_CTX_PRIORITY_NORMAL) 102 return 0; 103 104 if (capable(CAP_SYS_NICE)) 105 return 0; 106 107 if (drm_is_current_master(filp)) 108 return 0; 109 110 return -EACCES; 111 } 112 113 static enum amdgpu_gfx_pipe_priority amdgpu_ctx_prio_to_compute_prio(int32_t prio) 114 { 115 switch (prio) { 116 case AMDGPU_CTX_PRIORITY_HIGH: 117 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 118 return AMDGPU_GFX_PIPE_PRIO_HIGH; 119 default: 120 return AMDGPU_GFX_PIPE_PRIO_NORMAL; 121 } 122 } 123 124 static enum amdgpu_ring_priority_level amdgpu_ctx_sched_prio_to_ring_prio(int32_t prio) 125 { 126 switch (prio) { 127 case AMDGPU_CTX_PRIORITY_HIGH: 128 return AMDGPU_RING_PRIO_1; 129 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 130 return AMDGPU_RING_PRIO_2; 131 default: 132 return AMDGPU_RING_PRIO_0; 133 } 134 } 135 136 static unsigned int amdgpu_ctx_get_hw_prio(struct amdgpu_ctx *ctx, u32 hw_ip) 137 { 138 struct amdgpu_device *adev = ctx->adev; 139 int32_t ctx_prio; 140 unsigned int hw_prio; 141 142 ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ? 143 ctx->init_priority : ctx->override_priority; 144 145 switch (hw_ip) { 146 case AMDGPU_HW_IP_COMPUTE: 147 hw_prio = amdgpu_ctx_prio_to_compute_prio(ctx_prio); 148 break; 149 case AMDGPU_HW_IP_VCE: 150 case AMDGPU_HW_IP_VCN_ENC: 151 hw_prio = amdgpu_ctx_sched_prio_to_ring_prio(ctx_prio); 152 break; 153 default: 154 hw_prio = AMDGPU_RING_PRIO_DEFAULT; 155 break; 156 } 157 158 hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM); 159 if (adev->gpu_sched[hw_ip][hw_prio].num_scheds == 0) 160 hw_prio = AMDGPU_RING_PRIO_DEFAULT; 161 162 return hw_prio; 163 } 164 165 166 static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip, 167 const u32 ring) 168 { 169 struct amdgpu_device *adev = ctx->adev; 170 struct amdgpu_ctx_entity *entity; 171 struct drm_gpu_scheduler **scheds = NULL, *sched = NULL; 172 unsigned num_scheds = 0; 173 int32_t ctx_prio; 174 unsigned int hw_prio; 175 enum drm_sched_priority drm_prio; 176 int r; 177 178 entity = kzalloc(struct_size(entity, fences, amdgpu_sched_jobs), 179 GFP_KERNEL); 180 if (!entity) 181 return -ENOMEM; 182 183 ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ? 184 ctx->init_priority : ctx->override_priority; 185 entity->sequence = 1; 186 hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip); 187 drm_prio = amdgpu_ctx_to_drm_sched_prio(ctx_prio); 188 189 hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM); 190 scheds = adev->gpu_sched[hw_ip][hw_prio].sched; 191 num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds; 192 193 /* disable load balance if the hw engine retains context among dependent jobs */ 194 if (hw_ip == AMDGPU_HW_IP_VCN_ENC || 195 hw_ip == AMDGPU_HW_IP_VCN_DEC || 196 hw_ip == AMDGPU_HW_IP_UVD_ENC || 197 hw_ip == AMDGPU_HW_IP_UVD) { 198 sched = drm_sched_pick_best(scheds, num_scheds); 199 scheds = &sched; 200 num_scheds = 1; 201 } 202 203 r = drm_sched_entity_init(&entity->entity, drm_prio, scheds, num_scheds, 204 &ctx->guilty); 205 if (r) 206 goto error_free_entity; 207 208 /* It's not an error if we fail to install the new entity */ 209 if (cmpxchg(&ctx->entities[hw_ip][ring], NULL, entity)) 210 goto cleanup_entity; 211 212 return 0; 213 214 cleanup_entity: 215 drm_sched_entity_fini(&entity->entity); 216 217 error_free_entity: 218 kfree(entity); 219 220 return r; 221 } 222 223 static int amdgpu_ctx_init(struct amdgpu_device *adev, 224 int32_t priority, 225 struct drm_file *filp, 226 struct amdgpu_ctx *ctx) 227 { 228 int r; 229 230 r = amdgpu_ctx_priority_permit(filp, priority); 231 if (r) 232 return r; 233 234 memset(ctx, 0, sizeof(*ctx)); 235 236 ctx->adev = adev; 237 238 kref_init(&ctx->refcount); 239 spin_lock_init(&ctx->ring_lock); 240 241 ctx->reset_counter = atomic_read(&adev->gpu_reset_counter); 242 ctx->reset_counter_query = ctx->reset_counter; 243 ctx->vram_lost_counter = atomic_read(&adev->vram_lost_counter); 244 ctx->init_priority = priority; 245 ctx->override_priority = AMDGPU_CTX_PRIORITY_UNSET; 246 ctx->stable_pstate = AMDGPU_CTX_STABLE_PSTATE_NONE; 247 248 return 0; 249 } 250 251 static void amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity) 252 { 253 254 int i; 255 256 if (!entity) 257 return; 258 259 for (i = 0; i < amdgpu_sched_jobs; ++i) 260 dma_fence_put(entity->fences[i]); 261 262 kfree(entity); 263 } 264 265 static int amdgpu_ctx_get_stable_pstate(struct amdgpu_ctx *ctx, 266 u32 *stable_pstate) 267 { 268 struct amdgpu_device *adev = ctx->adev; 269 enum amd_dpm_forced_level current_level; 270 271 current_level = amdgpu_dpm_get_performance_level(adev); 272 273 switch (current_level) { 274 case AMD_DPM_FORCED_LEVEL_PROFILE_STANDARD: 275 *stable_pstate = AMDGPU_CTX_STABLE_PSTATE_STANDARD; 276 break; 277 case AMD_DPM_FORCED_LEVEL_PROFILE_MIN_SCLK: 278 *stable_pstate = AMDGPU_CTX_STABLE_PSTATE_MIN_SCLK; 279 break; 280 case AMD_DPM_FORCED_LEVEL_PROFILE_MIN_MCLK: 281 *stable_pstate = AMDGPU_CTX_STABLE_PSTATE_MIN_MCLK; 282 break; 283 case AMD_DPM_FORCED_LEVEL_PROFILE_PEAK: 284 *stable_pstate = AMDGPU_CTX_STABLE_PSTATE_PEAK; 285 break; 286 default: 287 *stable_pstate = AMDGPU_CTX_STABLE_PSTATE_NONE; 288 break; 289 } 290 return 0; 291 } 292 293 static int amdgpu_ctx_set_stable_pstate(struct amdgpu_ctx *ctx, 294 u32 stable_pstate) 295 { 296 struct amdgpu_device *adev = ctx->adev; 297 enum amd_dpm_forced_level level; 298 int r; 299 300 mutex_lock(&adev->pm.stable_pstate_ctx_lock); 301 if (adev->pm.stable_pstate_ctx && adev->pm.stable_pstate_ctx != ctx) { 302 r = -EBUSY; 303 goto done; 304 } 305 306 switch (stable_pstate) { 307 case AMDGPU_CTX_STABLE_PSTATE_NONE: 308 level = AMD_DPM_FORCED_LEVEL_AUTO; 309 break; 310 case AMDGPU_CTX_STABLE_PSTATE_STANDARD: 311 level = AMD_DPM_FORCED_LEVEL_PROFILE_STANDARD; 312 break; 313 case AMDGPU_CTX_STABLE_PSTATE_MIN_SCLK: 314 level = AMD_DPM_FORCED_LEVEL_PROFILE_MIN_SCLK; 315 break; 316 case AMDGPU_CTX_STABLE_PSTATE_MIN_MCLK: 317 level = AMD_DPM_FORCED_LEVEL_PROFILE_MIN_MCLK; 318 break; 319 case AMDGPU_CTX_STABLE_PSTATE_PEAK: 320 level = AMD_DPM_FORCED_LEVEL_PROFILE_PEAK; 321 break; 322 default: 323 r = -EINVAL; 324 goto done; 325 } 326 327 r = amdgpu_dpm_force_performance_level(adev, level); 328 329 if (level == AMD_DPM_FORCED_LEVEL_AUTO) 330 adev->pm.stable_pstate_ctx = NULL; 331 else 332 adev->pm.stable_pstate_ctx = ctx; 333 done: 334 mutex_unlock(&adev->pm.stable_pstate_ctx_lock); 335 336 return r; 337 } 338 339 static void amdgpu_ctx_fini(struct kref *ref) 340 { 341 struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount); 342 struct amdgpu_device *adev = ctx->adev; 343 unsigned i, j, idx; 344 345 if (!adev) 346 return; 347 348 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 349 for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) { 350 amdgpu_ctx_fini_entity(ctx->entities[i][j]); 351 ctx->entities[i][j] = NULL; 352 } 353 } 354 355 if (drm_dev_enter(&adev->ddev, &idx)) { 356 amdgpu_ctx_set_stable_pstate(ctx, AMDGPU_CTX_STABLE_PSTATE_NONE); 357 drm_dev_exit(idx); 358 } 359 360 kfree(ctx); 361 } 362 363 int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance, 364 u32 ring, struct drm_sched_entity **entity) 365 { 366 int r; 367 368 if (hw_ip >= AMDGPU_HW_IP_NUM) { 369 DRM_ERROR("unknown HW IP type: %d\n", hw_ip); 370 return -EINVAL; 371 } 372 373 /* Right now all IPs have only one instance - multiple rings. */ 374 if (instance != 0) { 375 DRM_DEBUG("invalid ip instance: %d\n", instance); 376 return -EINVAL; 377 } 378 379 if (ring >= amdgpu_ctx_num_entities[hw_ip]) { 380 DRM_DEBUG("invalid ring: %d %d\n", hw_ip, ring); 381 return -EINVAL; 382 } 383 384 if (ctx->entities[hw_ip][ring] == NULL) { 385 r = amdgpu_ctx_init_entity(ctx, hw_ip, ring); 386 if (r) 387 return r; 388 } 389 390 *entity = &ctx->entities[hw_ip][ring]->entity; 391 return 0; 392 } 393 394 static int amdgpu_ctx_alloc(struct amdgpu_device *adev, 395 struct amdgpu_fpriv *fpriv, 396 struct drm_file *filp, 397 int32_t priority, 398 uint32_t *id) 399 { 400 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr; 401 struct amdgpu_ctx *ctx; 402 int r; 403 404 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); 405 if (!ctx) 406 return -ENOMEM; 407 408 mutex_lock(&mgr->lock); 409 r = idr_alloc(&mgr->ctx_handles, ctx, 1, AMDGPU_VM_MAX_NUM_CTX, GFP_KERNEL); 410 if (r < 0) { 411 mutex_unlock(&mgr->lock); 412 kfree(ctx); 413 return r; 414 } 415 416 *id = (uint32_t)r; 417 r = amdgpu_ctx_init(adev, priority, filp, ctx); 418 if (r) { 419 idr_remove(&mgr->ctx_handles, *id); 420 *id = 0; 421 kfree(ctx); 422 } 423 mutex_unlock(&mgr->lock); 424 return r; 425 } 426 427 static void amdgpu_ctx_do_release(struct kref *ref) 428 { 429 struct amdgpu_ctx *ctx; 430 u32 i, j; 431 432 ctx = container_of(ref, struct amdgpu_ctx, refcount); 433 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 434 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 435 if (!ctx->entities[i][j]) 436 continue; 437 438 drm_sched_entity_destroy(&ctx->entities[i][j]->entity); 439 } 440 } 441 442 amdgpu_ctx_fini(ref); 443 } 444 445 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id) 446 { 447 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr; 448 struct amdgpu_ctx *ctx; 449 450 mutex_lock(&mgr->lock); 451 ctx = idr_remove(&mgr->ctx_handles, id); 452 if (ctx) 453 kref_put(&ctx->refcount, amdgpu_ctx_do_release); 454 mutex_unlock(&mgr->lock); 455 return ctx ? 0 : -EINVAL; 456 } 457 458 static int amdgpu_ctx_query(struct amdgpu_device *adev, 459 struct amdgpu_fpriv *fpriv, uint32_t id, 460 union drm_amdgpu_ctx_out *out) 461 { 462 struct amdgpu_ctx *ctx; 463 struct amdgpu_ctx_mgr *mgr; 464 unsigned reset_counter; 465 466 if (!fpriv) 467 return -EINVAL; 468 469 mgr = &fpriv->ctx_mgr; 470 mutex_lock(&mgr->lock); 471 ctx = idr_find(&mgr->ctx_handles, id); 472 if (!ctx) { 473 mutex_unlock(&mgr->lock); 474 return -EINVAL; 475 } 476 477 /* TODO: these two are always zero */ 478 out->state.flags = 0x0; 479 out->state.hangs = 0x0; 480 481 /* determine if a GPU reset has occured since the last call */ 482 reset_counter = atomic_read(&adev->gpu_reset_counter); 483 /* TODO: this should ideally return NO, GUILTY, or INNOCENT. */ 484 if (ctx->reset_counter_query == reset_counter) 485 out->state.reset_status = AMDGPU_CTX_NO_RESET; 486 else 487 out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET; 488 ctx->reset_counter_query = reset_counter; 489 490 mutex_unlock(&mgr->lock); 491 return 0; 492 } 493 494 #define AMDGPU_RAS_COUNTE_DELAY_MS 3000 495 496 static int amdgpu_ctx_query2(struct amdgpu_device *adev, 497 struct amdgpu_fpriv *fpriv, uint32_t id, 498 union drm_amdgpu_ctx_out *out) 499 { 500 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 501 struct amdgpu_ctx *ctx; 502 struct amdgpu_ctx_mgr *mgr; 503 504 if (!fpriv) 505 return -EINVAL; 506 507 mgr = &fpriv->ctx_mgr; 508 mutex_lock(&mgr->lock); 509 ctx = idr_find(&mgr->ctx_handles, id); 510 if (!ctx) { 511 mutex_unlock(&mgr->lock); 512 return -EINVAL; 513 } 514 515 out->state.flags = 0x0; 516 out->state.hangs = 0x0; 517 518 if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter)) 519 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET; 520 521 if (ctx->vram_lost_counter != atomic_read(&adev->vram_lost_counter)) 522 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST; 523 524 if (atomic_read(&ctx->guilty)) 525 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY; 526 527 if (adev->ras_enabled && con) { 528 /* Return the cached values in O(1), 529 * and schedule delayed work to cache 530 * new vaues. 531 */ 532 int ce_count, ue_count; 533 534 ce_count = atomic_read(&con->ras_ce_count); 535 ue_count = atomic_read(&con->ras_ue_count); 536 537 if (ce_count != ctx->ras_counter_ce) { 538 ctx->ras_counter_ce = ce_count; 539 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE; 540 } 541 542 if (ue_count != ctx->ras_counter_ue) { 543 ctx->ras_counter_ue = ue_count; 544 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE; 545 } 546 547 schedule_delayed_work(&con->ras_counte_delay_work, 548 msecs_to_jiffies(AMDGPU_RAS_COUNTE_DELAY_MS)); 549 } 550 551 mutex_unlock(&mgr->lock); 552 return 0; 553 } 554 555 556 557 static int amdgpu_ctx_stable_pstate(struct amdgpu_device *adev, 558 struct amdgpu_fpriv *fpriv, uint32_t id, 559 bool set, u32 *stable_pstate) 560 { 561 struct amdgpu_ctx *ctx; 562 struct amdgpu_ctx_mgr *mgr; 563 int r; 564 565 if (!fpriv) 566 return -EINVAL; 567 568 mgr = &fpriv->ctx_mgr; 569 mutex_lock(&mgr->lock); 570 ctx = idr_find(&mgr->ctx_handles, id); 571 if (!ctx) { 572 mutex_unlock(&mgr->lock); 573 return -EINVAL; 574 } 575 576 if (set) 577 r = amdgpu_ctx_set_stable_pstate(ctx, *stable_pstate); 578 else 579 r = amdgpu_ctx_get_stable_pstate(ctx, stable_pstate); 580 581 mutex_unlock(&mgr->lock); 582 return r; 583 } 584 585 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data, 586 struct drm_file *filp) 587 { 588 int r; 589 uint32_t id, stable_pstate; 590 int32_t priority; 591 592 union drm_amdgpu_ctx *args = data; 593 struct amdgpu_device *adev = drm_to_adev(dev); 594 struct amdgpu_fpriv *fpriv = filp->driver_priv; 595 596 id = args->in.ctx_id; 597 priority = args->in.priority; 598 599 /* For backwards compatibility reasons, we need to accept 600 * ioctls with garbage in the priority field */ 601 if (!amdgpu_ctx_priority_is_valid(priority)) 602 priority = AMDGPU_CTX_PRIORITY_NORMAL; 603 604 switch (args->in.op) { 605 case AMDGPU_CTX_OP_ALLOC_CTX: 606 r = amdgpu_ctx_alloc(adev, fpriv, filp, priority, &id); 607 args->out.alloc.ctx_id = id; 608 break; 609 case AMDGPU_CTX_OP_FREE_CTX: 610 r = amdgpu_ctx_free(fpriv, id); 611 break; 612 case AMDGPU_CTX_OP_QUERY_STATE: 613 r = amdgpu_ctx_query(adev, fpriv, id, &args->out); 614 break; 615 case AMDGPU_CTX_OP_QUERY_STATE2: 616 r = amdgpu_ctx_query2(adev, fpriv, id, &args->out); 617 break; 618 case AMDGPU_CTX_OP_GET_STABLE_PSTATE: 619 if (args->in.flags) 620 return -EINVAL; 621 r = amdgpu_ctx_stable_pstate(adev, fpriv, id, false, &stable_pstate); 622 if (!r) 623 args->out.pstate.flags = stable_pstate; 624 break; 625 case AMDGPU_CTX_OP_SET_STABLE_PSTATE: 626 if (args->in.flags & ~AMDGPU_CTX_STABLE_PSTATE_FLAGS_MASK) 627 return -EINVAL; 628 stable_pstate = args->in.flags & AMDGPU_CTX_STABLE_PSTATE_FLAGS_MASK; 629 if (stable_pstate > AMDGPU_CTX_STABLE_PSTATE_PEAK) 630 return -EINVAL; 631 r = amdgpu_ctx_stable_pstate(adev, fpriv, id, true, &stable_pstate); 632 break; 633 default: 634 return -EINVAL; 635 } 636 637 return r; 638 } 639 640 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id) 641 { 642 struct amdgpu_ctx *ctx; 643 struct amdgpu_ctx_mgr *mgr; 644 645 if (!fpriv) 646 return NULL; 647 648 mgr = &fpriv->ctx_mgr; 649 650 mutex_lock(&mgr->lock); 651 ctx = idr_find(&mgr->ctx_handles, id); 652 if (ctx) 653 kref_get(&ctx->refcount); 654 mutex_unlock(&mgr->lock); 655 return ctx; 656 } 657 658 int amdgpu_ctx_put(struct amdgpu_ctx *ctx) 659 { 660 if (ctx == NULL) 661 return -EINVAL; 662 663 kref_put(&ctx->refcount, amdgpu_ctx_do_release); 664 return 0; 665 } 666 667 void amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, 668 struct drm_sched_entity *entity, 669 struct dma_fence *fence, uint64_t *handle) 670 { 671 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 672 uint64_t seq = centity->sequence; 673 struct dma_fence *other = NULL; 674 unsigned idx = 0; 675 676 idx = seq & (amdgpu_sched_jobs - 1); 677 other = centity->fences[idx]; 678 if (other) 679 BUG_ON(!dma_fence_is_signaled(other)); 680 681 dma_fence_get(fence); 682 683 spin_lock(&ctx->ring_lock); 684 centity->fences[idx] = fence; 685 centity->sequence++; 686 spin_unlock(&ctx->ring_lock); 687 688 dma_fence_put(other); 689 if (handle) 690 *handle = seq; 691 } 692 693 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx, 694 struct drm_sched_entity *entity, 695 uint64_t seq) 696 { 697 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 698 struct dma_fence *fence; 699 700 spin_lock(&ctx->ring_lock); 701 702 if (seq == ~0ull) 703 seq = centity->sequence - 1; 704 705 if (seq >= centity->sequence) { 706 spin_unlock(&ctx->ring_lock); 707 return ERR_PTR(-EINVAL); 708 } 709 710 711 if (seq + amdgpu_sched_jobs < centity->sequence) { 712 spin_unlock(&ctx->ring_lock); 713 return NULL; 714 } 715 716 fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]); 717 spin_unlock(&ctx->ring_lock); 718 719 return fence; 720 } 721 722 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx, 723 struct amdgpu_ctx_entity *aentity, 724 int hw_ip, 725 int32_t priority) 726 { 727 struct amdgpu_device *adev = ctx->adev; 728 unsigned int hw_prio; 729 struct drm_gpu_scheduler **scheds = NULL; 730 unsigned num_scheds; 731 732 /* set sw priority */ 733 drm_sched_entity_set_priority(&aentity->entity, 734 amdgpu_ctx_to_drm_sched_prio(priority)); 735 736 /* set hw priority */ 737 if (hw_ip == AMDGPU_HW_IP_COMPUTE) { 738 hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip); 739 hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX); 740 scheds = adev->gpu_sched[hw_ip][hw_prio].sched; 741 num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds; 742 drm_sched_entity_modify_sched(&aentity->entity, scheds, 743 num_scheds); 744 } 745 } 746 747 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, 748 int32_t priority) 749 { 750 int32_t ctx_prio; 751 unsigned i, j; 752 753 ctx->override_priority = priority; 754 755 ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ? 756 ctx->init_priority : ctx->override_priority; 757 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 758 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 759 if (!ctx->entities[i][j]) 760 continue; 761 762 amdgpu_ctx_set_entity_priority(ctx, ctx->entities[i][j], 763 i, ctx_prio); 764 } 765 } 766 } 767 768 int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx, 769 struct drm_sched_entity *entity) 770 { 771 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 772 struct dma_fence *other; 773 unsigned idx; 774 long r; 775 776 spin_lock(&ctx->ring_lock); 777 idx = centity->sequence & (amdgpu_sched_jobs - 1); 778 other = dma_fence_get(centity->fences[idx]); 779 spin_unlock(&ctx->ring_lock); 780 781 if (!other) 782 return 0; 783 784 r = dma_fence_wait(other, true); 785 if (r < 0 && r != -ERESTARTSYS) 786 DRM_ERROR("Error (%ld) waiting for fence!\n", r); 787 788 dma_fence_put(other); 789 return r; 790 } 791 792 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr) 793 { 794 mutex_init(&mgr->lock); 795 idr_init(&mgr->ctx_handles); 796 } 797 798 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout) 799 { 800 struct amdgpu_ctx *ctx; 801 struct idr *idp; 802 uint32_t id, i, j; 803 804 idp = &mgr->ctx_handles; 805 806 mutex_lock(&mgr->lock); 807 idr_for_each_entry(idp, ctx, id) { 808 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 809 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 810 struct drm_sched_entity *entity; 811 812 if (!ctx->entities[i][j]) 813 continue; 814 815 entity = &ctx->entities[i][j]->entity; 816 timeout = drm_sched_entity_flush(entity, timeout); 817 } 818 } 819 } 820 mutex_unlock(&mgr->lock); 821 return timeout; 822 } 823 824 void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr) 825 { 826 struct amdgpu_ctx *ctx; 827 struct idr *idp; 828 uint32_t id, i, j; 829 830 idp = &mgr->ctx_handles; 831 832 idr_for_each_entry(idp, ctx, id) { 833 if (kref_read(&ctx->refcount) != 1) { 834 DRM_ERROR("ctx %p is still alive\n", ctx); 835 continue; 836 } 837 838 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 839 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 840 struct drm_sched_entity *entity; 841 842 if (!ctx->entities[i][j]) 843 continue; 844 845 entity = &ctx->entities[i][j]->entity; 846 drm_sched_entity_fini(entity); 847 } 848 } 849 } 850 } 851 852 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr) 853 { 854 struct amdgpu_ctx *ctx; 855 struct idr *idp; 856 uint32_t id; 857 858 amdgpu_ctx_mgr_entity_fini(mgr); 859 860 idp = &mgr->ctx_handles; 861 862 idr_for_each_entry(idp, ctx, id) { 863 if (kref_put(&ctx->refcount, amdgpu_ctx_fini) != 1) 864 DRM_ERROR("ctx %p is still alive\n", ctx); 865 } 866 867 idr_destroy(&mgr->ctx_handles); 868 mutex_destroy(&mgr->lock); 869 } 870 871 static void amdgpu_ctx_fence_time(struct amdgpu_ctx *ctx, 872 struct amdgpu_ctx_entity *centity, ktime_t *total, ktime_t *max) 873 { 874 ktime_t now, t1; 875 uint32_t i; 876 877 *total = *max = 0; 878 879 now = ktime_get(); 880 for (i = 0; i < amdgpu_sched_jobs; i++) { 881 struct dma_fence *fence; 882 struct drm_sched_fence *s_fence; 883 884 spin_lock(&ctx->ring_lock); 885 fence = dma_fence_get(centity->fences[i]); 886 spin_unlock(&ctx->ring_lock); 887 if (!fence) 888 continue; 889 s_fence = to_drm_sched_fence(fence); 890 if (!dma_fence_is_signaled(&s_fence->scheduled)) { 891 dma_fence_put(fence); 892 continue; 893 } 894 t1 = s_fence->scheduled.timestamp; 895 if (!ktime_before(t1, now)) { 896 dma_fence_put(fence); 897 continue; 898 } 899 if (dma_fence_is_signaled(&s_fence->finished) && 900 s_fence->finished.timestamp < now) 901 *total += ktime_sub(s_fence->finished.timestamp, t1); 902 else 903 *total += ktime_sub(now, t1); 904 t1 = ktime_sub(now, t1); 905 dma_fence_put(fence); 906 *max = max(t1, *max); 907 } 908 } 909 910 ktime_t amdgpu_ctx_mgr_fence_usage(struct amdgpu_ctx_mgr *mgr, uint32_t hwip, 911 uint32_t idx, uint64_t *elapsed) 912 { 913 struct idr *idp; 914 struct amdgpu_ctx *ctx; 915 uint32_t id; 916 struct amdgpu_ctx_entity *centity; 917 ktime_t total = 0, max = 0; 918 919 if (idx >= AMDGPU_MAX_ENTITY_NUM) 920 return 0; 921 idp = &mgr->ctx_handles; 922 mutex_lock(&mgr->lock); 923 idr_for_each_entry(idp, ctx, id) { 924 ktime_t ttotal, tmax; 925 926 if (!ctx->entities[hwip][idx]) 927 continue; 928 929 centity = ctx->entities[hwip][idx]; 930 amdgpu_ctx_fence_time(ctx, centity, &ttotal, &tmax); 931 932 /* Harmonic mean approximation diverges for very small 933 * values. If ratio < 0.01% ignore 934 */ 935 if (AMDGPU_CTX_FENCE_USAGE_MIN_RATIO(tmax, ttotal)) 936 continue; 937 938 total = ktime_add(total, ttotal); 939 max = ktime_after(tmax, max) ? tmax : max; 940 } 941 942 mutex_unlock(&mgr->lock); 943 if (elapsed) 944 *elapsed = max; 945 946 return total; 947 } 948