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_VERY_LOW: 51 case AMDGPU_CTX_PRIORITY_LOW: 52 case AMDGPU_CTX_PRIORITY_NORMAL: 53 case AMDGPU_CTX_PRIORITY_HIGH: 54 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 55 return true; 56 default: 57 case AMDGPU_CTX_PRIORITY_UNSET: 58 /* UNSET priority is not valid and we don't carry that 59 * around, but set it to NORMAL in the only place this 60 * function is called, amdgpu_ctx_ioctl(). 61 */ 62 return false; 63 } 64 } 65 66 static enum drm_sched_priority 67 amdgpu_ctx_to_drm_sched_prio(int32_t ctx_prio) 68 { 69 switch (ctx_prio) { 70 case AMDGPU_CTX_PRIORITY_UNSET: 71 pr_warn_once("AMD-->DRM context priority value UNSET-->NORMAL"); 72 return DRM_SCHED_PRIORITY_NORMAL; 73 74 case AMDGPU_CTX_PRIORITY_VERY_LOW: 75 return DRM_SCHED_PRIORITY_MIN; 76 77 case AMDGPU_CTX_PRIORITY_LOW: 78 return DRM_SCHED_PRIORITY_MIN; 79 80 case AMDGPU_CTX_PRIORITY_NORMAL: 81 return DRM_SCHED_PRIORITY_NORMAL; 82 83 case AMDGPU_CTX_PRIORITY_HIGH: 84 return DRM_SCHED_PRIORITY_HIGH; 85 86 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 87 return DRM_SCHED_PRIORITY_HIGH; 88 89 /* This should not happen as we sanitized userspace provided priority 90 * already, WARN if this happens. 91 */ 92 default: 93 WARN(1, "Invalid context priority %d\n", ctx_prio); 94 return DRM_SCHED_PRIORITY_NORMAL; 95 } 96 97 } 98 99 static int amdgpu_ctx_priority_permit(struct drm_file *filp, 100 int32_t priority) 101 { 102 /* NORMAL and below are accessible by everyone */ 103 if (priority <= AMDGPU_CTX_PRIORITY_NORMAL) 104 return 0; 105 106 if (capable(CAP_SYS_NICE)) 107 return 0; 108 109 if (drm_is_current_master(filp)) 110 return 0; 111 112 return -EACCES; 113 } 114 115 static enum amdgpu_gfx_pipe_priority amdgpu_ctx_prio_to_gfx_pipe_prio(int32_t prio) 116 { 117 switch (prio) { 118 case AMDGPU_CTX_PRIORITY_HIGH: 119 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 120 return AMDGPU_GFX_PIPE_PRIO_HIGH; 121 default: 122 return AMDGPU_GFX_PIPE_PRIO_NORMAL; 123 } 124 } 125 126 static enum amdgpu_ring_priority_level amdgpu_ctx_sched_prio_to_ring_prio(int32_t prio) 127 { 128 switch (prio) { 129 case AMDGPU_CTX_PRIORITY_HIGH: 130 return AMDGPU_RING_PRIO_1; 131 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 132 return AMDGPU_RING_PRIO_2; 133 default: 134 return AMDGPU_RING_PRIO_0; 135 } 136 } 137 138 static unsigned int amdgpu_ctx_get_hw_prio(struct amdgpu_ctx *ctx, u32 hw_ip) 139 { 140 struct amdgpu_device *adev = ctx->mgr->adev; 141 unsigned int hw_prio; 142 int32_t ctx_prio; 143 144 ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ? 145 ctx->init_priority : ctx->override_priority; 146 147 switch (hw_ip) { 148 case AMDGPU_HW_IP_GFX: 149 case AMDGPU_HW_IP_COMPUTE: 150 hw_prio = amdgpu_ctx_prio_to_gfx_pipe_prio(ctx_prio); 151 break; 152 case AMDGPU_HW_IP_VCE: 153 case AMDGPU_HW_IP_VCN_ENC: 154 hw_prio = amdgpu_ctx_sched_prio_to_ring_prio(ctx_prio); 155 break; 156 default: 157 hw_prio = AMDGPU_RING_PRIO_DEFAULT; 158 break; 159 } 160 161 hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM); 162 if (adev->gpu_sched[hw_ip][hw_prio].num_scheds == 0) 163 hw_prio = AMDGPU_RING_PRIO_DEFAULT; 164 165 return hw_prio; 166 } 167 168 /* Calculate the time spend on the hw */ 169 static ktime_t amdgpu_ctx_fence_time(struct dma_fence *fence) 170 { 171 struct drm_sched_fence *s_fence; 172 173 if (!fence) 174 return ns_to_ktime(0); 175 176 /* When the fence is not even scheduled it can't have spend time */ 177 s_fence = to_drm_sched_fence(fence); 178 if (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &s_fence->scheduled.flags)) 179 return ns_to_ktime(0); 180 181 /* When it is still running account how much already spend */ 182 if (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &s_fence->finished.flags)) 183 return ktime_sub(ktime_get(), s_fence->scheduled.timestamp); 184 185 return ktime_sub(s_fence->finished.timestamp, 186 s_fence->scheduled.timestamp); 187 } 188 189 static ktime_t amdgpu_ctx_entity_time(struct amdgpu_ctx *ctx, 190 struct amdgpu_ctx_entity *centity) 191 { 192 ktime_t res = ns_to_ktime(0); 193 uint32_t i; 194 195 spin_lock(&ctx->ring_lock); 196 for (i = 0; i < amdgpu_sched_jobs; i++) { 197 res = ktime_add(res, amdgpu_ctx_fence_time(centity->fences[i])); 198 } 199 spin_unlock(&ctx->ring_lock); 200 return res; 201 } 202 203 static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip, 204 const u32 ring) 205 { 206 struct drm_gpu_scheduler **scheds = NULL, *sched = NULL; 207 struct amdgpu_device *adev = ctx->mgr->adev; 208 struct amdgpu_ctx_entity *entity; 209 enum drm_sched_priority drm_prio; 210 unsigned int hw_prio, num_scheds; 211 int32_t ctx_prio; 212 int r; 213 214 entity = kzalloc(struct_size(entity, fences, amdgpu_sched_jobs), 215 GFP_KERNEL); 216 if (!entity) 217 return -ENOMEM; 218 219 ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ? 220 ctx->init_priority : ctx->override_priority; 221 entity->hw_ip = hw_ip; 222 entity->sequence = 1; 223 hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip); 224 drm_prio = amdgpu_ctx_to_drm_sched_prio(ctx_prio); 225 226 hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM); 227 228 if (!(adev)->xcp_mgr) { 229 scheds = adev->gpu_sched[hw_ip][hw_prio].sched; 230 num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds; 231 } else { 232 struct amdgpu_fpriv *fpriv; 233 234 fpriv = container_of(ctx->ctx_mgr, struct amdgpu_fpriv, ctx_mgr); 235 r = amdgpu_xcp_select_scheds(adev, hw_ip, hw_prio, fpriv, 236 &num_scheds, &scheds); 237 if (r) 238 goto cleanup_entity; 239 } 240 241 /* disable load balance if the hw engine retains context among dependent jobs */ 242 if (hw_ip == AMDGPU_HW_IP_VCN_ENC || 243 hw_ip == AMDGPU_HW_IP_VCN_DEC || 244 hw_ip == AMDGPU_HW_IP_UVD_ENC || 245 hw_ip == AMDGPU_HW_IP_UVD) { 246 sched = drm_sched_pick_best(scheds, num_scheds); 247 scheds = &sched; 248 num_scheds = 1; 249 } 250 251 r = drm_sched_entity_init(&entity->entity, drm_prio, scheds, num_scheds, 252 &ctx->guilty); 253 if (r) 254 goto error_free_entity; 255 256 /* It's not an error if we fail to install the new entity */ 257 if (cmpxchg(&ctx->entities[hw_ip][ring], NULL, entity)) 258 goto cleanup_entity; 259 260 return 0; 261 262 cleanup_entity: 263 drm_sched_entity_fini(&entity->entity); 264 265 error_free_entity: 266 kfree(entity); 267 268 return r; 269 } 270 271 static ktime_t amdgpu_ctx_fini_entity(struct amdgpu_device *adev, 272 struct amdgpu_ctx_entity *entity) 273 { 274 ktime_t res = ns_to_ktime(0); 275 int i; 276 277 if (!entity) 278 return res; 279 280 for (i = 0; i < amdgpu_sched_jobs; ++i) { 281 res = ktime_add(res, amdgpu_ctx_fence_time(entity->fences[i])); 282 dma_fence_put(entity->fences[i]); 283 } 284 285 amdgpu_xcp_release_sched(adev, entity); 286 287 kfree(entity); 288 return res; 289 } 290 291 static int amdgpu_ctx_get_stable_pstate(struct amdgpu_ctx *ctx, 292 u32 *stable_pstate) 293 { 294 struct amdgpu_device *adev = ctx->mgr->adev; 295 enum amd_dpm_forced_level current_level; 296 297 current_level = amdgpu_dpm_get_performance_level(adev); 298 299 switch (current_level) { 300 case AMD_DPM_FORCED_LEVEL_PROFILE_STANDARD: 301 *stable_pstate = AMDGPU_CTX_STABLE_PSTATE_STANDARD; 302 break; 303 case AMD_DPM_FORCED_LEVEL_PROFILE_MIN_SCLK: 304 *stable_pstate = AMDGPU_CTX_STABLE_PSTATE_MIN_SCLK; 305 break; 306 case AMD_DPM_FORCED_LEVEL_PROFILE_MIN_MCLK: 307 *stable_pstate = AMDGPU_CTX_STABLE_PSTATE_MIN_MCLK; 308 break; 309 case AMD_DPM_FORCED_LEVEL_PROFILE_PEAK: 310 *stable_pstate = AMDGPU_CTX_STABLE_PSTATE_PEAK; 311 break; 312 default: 313 *stable_pstate = AMDGPU_CTX_STABLE_PSTATE_NONE; 314 break; 315 } 316 return 0; 317 } 318 319 static int amdgpu_ctx_init(struct amdgpu_ctx_mgr *mgr, int32_t priority, 320 struct drm_file *filp, struct amdgpu_ctx *ctx) 321 { 322 struct amdgpu_fpriv *fpriv = filp->driver_priv; 323 u32 current_stable_pstate; 324 int r; 325 326 r = amdgpu_ctx_priority_permit(filp, priority); 327 if (r) 328 return r; 329 330 memset(ctx, 0, sizeof(*ctx)); 331 332 kref_init(&ctx->refcount); 333 ctx->mgr = mgr; 334 spin_lock_init(&ctx->ring_lock); 335 336 ctx->reset_counter = atomic_read(&mgr->adev->gpu_reset_counter); 337 ctx->reset_counter_query = ctx->reset_counter; 338 ctx->generation = amdgpu_vm_generation(mgr->adev, &fpriv->vm); 339 ctx->init_priority = priority; 340 ctx->override_priority = AMDGPU_CTX_PRIORITY_UNSET; 341 342 r = amdgpu_ctx_get_stable_pstate(ctx, ¤t_stable_pstate); 343 if (r) 344 return r; 345 346 if (mgr->adev->pm.stable_pstate_ctx) 347 ctx->stable_pstate = mgr->adev->pm.stable_pstate_ctx->stable_pstate; 348 else 349 ctx->stable_pstate = current_stable_pstate; 350 351 ctx->ctx_mgr = &(fpriv->ctx_mgr); 352 return 0; 353 } 354 355 static int amdgpu_ctx_set_stable_pstate(struct amdgpu_ctx *ctx, 356 u32 stable_pstate) 357 { 358 struct amdgpu_device *adev = ctx->mgr->adev; 359 enum amd_dpm_forced_level level; 360 u32 current_stable_pstate; 361 int r; 362 363 mutex_lock(&adev->pm.stable_pstate_ctx_lock); 364 if (adev->pm.stable_pstate_ctx && adev->pm.stable_pstate_ctx != ctx) { 365 r = -EBUSY; 366 goto done; 367 } 368 369 r = amdgpu_ctx_get_stable_pstate(ctx, ¤t_stable_pstate); 370 if (r || (stable_pstate == current_stable_pstate)) 371 goto done; 372 373 switch (stable_pstate) { 374 case AMDGPU_CTX_STABLE_PSTATE_NONE: 375 level = AMD_DPM_FORCED_LEVEL_AUTO; 376 break; 377 case AMDGPU_CTX_STABLE_PSTATE_STANDARD: 378 level = AMD_DPM_FORCED_LEVEL_PROFILE_STANDARD; 379 break; 380 case AMDGPU_CTX_STABLE_PSTATE_MIN_SCLK: 381 level = AMD_DPM_FORCED_LEVEL_PROFILE_MIN_SCLK; 382 break; 383 case AMDGPU_CTX_STABLE_PSTATE_MIN_MCLK: 384 level = AMD_DPM_FORCED_LEVEL_PROFILE_MIN_MCLK; 385 break; 386 case AMDGPU_CTX_STABLE_PSTATE_PEAK: 387 level = AMD_DPM_FORCED_LEVEL_PROFILE_PEAK; 388 break; 389 default: 390 r = -EINVAL; 391 goto done; 392 } 393 394 r = amdgpu_dpm_force_performance_level(adev, level); 395 396 if (level == AMD_DPM_FORCED_LEVEL_AUTO) 397 adev->pm.stable_pstate_ctx = NULL; 398 else 399 adev->pm.stable_pstate_ctx = ctx; 400 done: 401 mutex_unlock(&adev->pm.stable_pstate_ctx_lock); 402 403 return r; 404 } 405 406 static void amdgpu_ctx_fini(struct kref *ref) 407 { 408 struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount); 409 struct amdgpu_ctx_mgr *mgr = ctx->mgr; 410 struct amdgpu_device *adev = mgr->adev; 411 unsigned i, j, idx; 412 413 if (!adev) 414 return; 415 416 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 417 for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) { 418 ktime_t spend; 419 420 spend = amdgpu_ctx_fini_entity(adev, ctx->entities[i][j]); 421 atomic64_add(ktime_to_ns(spend), &mgr->time_spend[i]); 422 } 423 } 424 425 if (drm_dev_enter(adev_to_drm(adev), &idx)) { 426 amdgpu_ctx_set_stable_pstate(ctx, ctx->stable_pstate); 427 drm_dev_exit(idx); 428 } 429 430 kfree(ctx); 431 } 432 433 int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance, 434 u32 ring, struct drm_sched_entity **entity) 435 { 436 int r; 437 struct drm_sched_entity *ctx_entity; 438 439 if (hw_ip >= AMDGPU_HW_IP_NUM) { 440 DRM_ERROR("unknown HW IP type: %d\n", hw_ip); 441 return -EINVAL; 442 } 443 444 /* Right now all IPs have only one instance - multiple rings. */ 445 if (instance != 0) { 446 DRM_DEBUG("invalid ip instance: %d\n", instance); 447 return -EINVAL; 448 } 449 450 if (ring >= amdgpu_ctx_num_entities[hw_ip]) { 451 DRM_DEBUG("invalid ring: %d %d\n", hw_ip, ring); 452 return -EINVAL; 453 } 454 455 if (ctx->entities[hw_ip][ring] == NULL) { 456 r = amdgpu_ctx_init_entity(ctx, hw_ip, ring); 457 if (r) 458 return r; 459 } 460 461 ctx_entity = &ctx->entities[hw_ip][ring]->entity; 462 r = drm_sched_entity_error(ctx_entity); 463 if (r) { 464 DRM_DEBUG("error entity %p\n", ctx_entity); 465 return r; 466 } 467 468 *entity = ctx_entity; 469 return 0; 470 } 471 472 static int amdgpu_ctx_alloc(struct amdgpu_device *adev, 473 struct amdgpu_fpriv *fpriv, 474 struct drm_file *filp, 475 int32_t priority, 476 uint32_t *id) 477 { 478 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr; 479 struct amdgpu_ctx *ctx; 480 int r; 481 482 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); 483 if (!ctx) 484 return -ENOMEM; 485 486 mutex_lock(&mgr->lock); 487 r = idr_alloc(&mgr->ctx_handles, ctx, 1, AMDGPU_VM_MAX_NUM_CTX, GFP_KERNEL); 488 if (r < 0) { 489 mutex_unlock(&mgr->lock); 490 kfree(ctx); 491 return r; 492 } 493 494 *id = (uint32_t)r; 495 r = amdgpu_ctx_init(mgr, priority, filp, ctx); 496 if (r) { 497 idr_remove(&mgr->ctx_handles, *id); 498 *id = 0; 499 kfree(ctx); 500 } 501 mutex_unlock(&mgr->lock); 502 return r; 503 } 504 505 static void amdgpu_ctx_do_release(struct kref *ref) 506 { 507 struct amdgpu_ctx *ctx; 508 u32 i, j; 509 510 ctx = container_of(ref, struct amdgpu_ctx, refcount); 511 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 512 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 513 if (!ctx->entities[i][j]) 514 continue; 515 516 drm_sched_entity_destroy(&ctx->entities[i][j]->entity); 517 } 518 } 519 520 amdgpu_ctx_fini(ref); 521 } 522 523 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id) 524 { 525 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr; 526 struct amdgpu_ctx *ctx; 527 528 mutex_lock(&mgr->lock); 529 ctx = idr_remove(&mgr->ctx_handles, id); 530 if (ctx) 531 kref_put(&ctx->refcount, amdgpu_ctx_do_release); 532 mutex_unlock(&mgr->lock); 533 return ctx ? 0 : -EINVAL; 534 } 535 536 static int amdgpu_ctx_query(struct amdgpu_device *adev, 537 struct amdgpu_fpriv *fpriv, uint32_t id, 538 union drm_amdgpu_ctx_out *out) 539 { 540 struct amdgpu_ctx *ctx; 541 struct amdgpu_ctx_mgr *mgr; 542 unsigned reset_counter; 543 544 if (!fpriv) 545 return -EINVAL; 546 547 mgr = &fpriv->ctx_mgr; 548 mutex_lock(&mgr->lock); 549 ctx = idr_find(&mgr->ctx_handles, id); 550 if (!ctx) { 551 mutex_unlock(&mgr->lock); 552 return -EINVAL; 553 } 554 555 /* TODO: these two are always zero */ 556 out->state.flags = 0x0; 557 out->state.hangs = 0x0; 558 559 /* determine if a GPU reset has occured since the last call */ 560 reset_counter = atomic_read(&adev->gpu_reset_counter); 561 /* TODO: this should ideally return NO, GUILTY, or INNOCENT. */ 562 if (ctx->reset_counter_query == reset_counter) 563 out->state.reset_status = AMDGPU_CTX_NO_RESET; 564 else 565 out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET; 566 ctx->reset_counter_query = reset_counter; 567 568 mutex_unlock(&mgr->lock); 569 return 0; 570 } 571 572 #define AMDGPU_RAS_COUNTE_DELAY_MS 3000 573 574 static int amdgpu_ctx_query2(struct amdgpu_device *adev, 575 struct amdgpu_fpriv *fpriv, uint32_t id, 576 union drm_amdgpu_ctx_out *out) 577 { 578 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 579 struct amdgpu_ctx *ctx; 580 struct amdgpu_ctx_mgr *mgr; 581 582 if (!fpriv) 583 return -EINVAL; 584 585 mgr = &fpriv->ctx_mgr; 586 mutex_lock(&mgr->lock); 587 ctx = idr_find(&mgr->ctx_handles, id); 588 if (!ctx) { 589 mutex_unlock(&mgr->lock); 590 return -EINVAL; 591 } 592 593 out->state.flags = 0x0; 594 out->state.hangs = 0x0; 595 596 if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter)) 597 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET; 598 599 if (ctx->generation != amdgpu_vm_generation(adev, &fpriv->vm)) 600 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST; 601 602 if (atomic_read(&ctx->guilty)) 603 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY; 604 605 if (amdgpu_in_reset(adev)) 606 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET_IN_PROGRESS; 607 608 if (adev->ras_enabled && con) { 609 /* Return the cached values in O(1), 610 * and schedule delayed work to cache 611 * new vaues. 612 */ 613 int ce_count, ue_count; 614 615 ce_count = atomic_read(&con->ras_ce_count); 616 ue_count = atomic_read(&con->ras_ue_count); 617 618 if (ce_count != ctx->ras_counter_ce) { 619 ctx->ras_counter_ce = ce_count; 620 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE; 621 } 622 623 if (ue_count != ctx->ras_counter_ue) { 624 ctx->ras_counter_ue = ue_count; 625 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE; 626 } 627 628 schedule_delayed_work(&con->ras_counte_delay_work, 629 msecs_to_jiffies(AMDGPU_RAS_COUNTE_DELAY_MS)); 630 } 631 632 mutex_unlock(&mgr->lock); 633 return 0; 634 } 635 636 static int amdgpu_ctx_stable_pstate(struct amdgpu_device *adev, 637 struct amdgpu_fpriv *fpriv, uint32_t id, 638 bool set, u32 *stable_pstate) 639 { 640 struct amdgpu_ctx *ctx; 641 struct amdgpu_ctx_mgr *mgr; 642 int r; 643 644 if (!fpriv) 645 return -EINVAL; 646 647 mgr = &fpriv->ctx_mgr; 648 mutex_lock(&mgr->lock); 649 ctx = idr_find(&mgr->ctx_handles, id); 650 if (!ctx) { 651 mutex_unlock(&mgr->lock); 652 return -EINVAL; 653 } 654 655 if (set) 656 r = amdgpu_ctx_set_stable_pstate(ctx, *stable_pstate); 657 else 658 r = amdgpu_ctx_get_stable_pstate(ctx, stable_pstate); 659 660 mutex_unlock(&mgr->lock); 661 return r; 662 } 663 664 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data, 665 struct drm_file *filp) 666 { 667 int r; 668 uint32_t id, stable_pstate; 669 int32_t priority; 670 671 union drm_amdgpu_ctx *args = data; 672 struct amdgpu_device *adev = drm_to_adev(dev); 673 struct amdgpu_fpriv *fpriv = filp->driver_priv; 674 675 id = args->in.ctx_id; 676 priority = args->in.priority; 677 678 /* For backwards compatibility, we need to accept ioctls with garbage 679 * in the priority field. Garbage values in the priority field, result 680 * in the priority being set to NORMAL. 681 */ 682 if (!amdgpu_ctx_priority_is_valid(priority)) 683 priority = AMDGPU_CTX_PRIORITY_NORMAL; 684 685 switch (args->in.op) { 686 case AMDGPU_CTX_OP_ALLOC_CTX: 687 r = amdgpu_ctx_alloc(adev, fpriv, filp, priority, &id); 688 args->out.alloc.ctx_id = id; 689 break; 690 case AMDGPU_CTX_OP_FREE_CTX: 691 r = amdgpu_ctx_free(fpriv, id); 692 break; 693 case AMDGPU_CTX_OP_QUERY_STATE: 694 r = amdgpu_ctx_query(adev, fpriv, id, &args->out); 695 break; 696 case AMDGPU_CTX_OP_QUERY_STATE2: 697 r = amdgpu_ctx_query2(adev, fpriv, id, &args->out); 698 break; 699 case AMDGPU_CTX_OP_GET_STABLE_PSTATE: 700 if (args->in.flags) 701 return -EINVAL; 702 r = amdgpu_ctx_stable_pstate(adev, fpriv, id, false, &stable_pstate); 703 if (!r) 704 args->out.pstate.flags = stable_pstate; 705 break; 706 case AMDGPU_CTX_OP_SET_STABLE_PSTATE: 707 if (args->in.flags & ~AMDGPU_CTX_STABLE_PSTATE_FLAGS_MASK) 708 return -EINVAL; 709 stable_pstate = args->in.flags & AMDGPU_CTX_STABLE_PSTATE_FLAGS_MASK; 710 if (stable_pstate > AMDGPU_CTX_STABLE_PSTATE_PEAK) 711 return -EINVAL; 712 r = amdgpu_ctx_stable_pstate(adev, fpriv, id, true, &stable_pstate); 713 break; 714 default: 715 return -EINVAL; 716 } 717 718 return r; 719 } 720 721 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id) 722 { 723 struct amdgpu_ctx *ctx; 724 struct amdgpu_ctx_mgr *mgr; 725 726 if (!fpriv) 727 return NULL; 728 729 mgr = &fpriv->ctx_mgr; 730 731 mutex_lock(&mgr->lock); 732 ctx = idr_find(&mgr->ctx_handles, id); 733 if (ctx) 734 kref_get(&ctx->refcount); 735 mutex_unlock(&mgr->lock); 736 return ctx; 737 } 738 739 int amdgpu_ctx_put(struct amdgpu_ctx *ctx) 740 { 741 if (ctx == NULL) 742 return -EINVAL; 743 744 kref_put(&ctx->refcount, amdgpu_ctx_do_release); 745 return 0; 746 } 747 748 uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, 749 struct drm_sched_entity *entity, 750 struct dma_fence *fence) 751 { 752 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 753 uint64_t seq = centity->sequence; 754 struct dma_fence *other = NULL; 755 unsigned idx = 0; 756 757 idx = seq & (amdgpu_sched_jobs - 1); 758 other = centity->fences[idx]; 759 WARN_ON(other && !dma_fence_is_signaled(other)); 760 761 dma_fence_get(fence); 762 763 spin_lock(&ctx->ring_lock); 764 centity->fences[idx] = fence; 765 centity->sequence++; 766 spin_unlock(&ctx->ring_lock); 767 768 atomic64_add(ktime_to_ns(amdgpu_ctx_fence_time(other)), 769 &ctx->mgr->time_spend[centity->hw_ip]); 770 771 dma_fence_put(other); 772 return seq; 773 } 774 775 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx, 776 struct drm_sched_entity *entity, 777 uint64_t seq) 778 { 779 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 780 struct dma_fence *fence; 781 782 spin_lock(&ctx->ring_lock); 783 784 if (seq == ~0ull) 785 seq = centity->sequence - 1; 786 787 if (seq >= centity->sequence) { 788 spin_unlock(&ctx->ring_lock); 789 return ERR_PTR(-EINVAL); 790 } 791 792 793 if (seq + amdgpu_sched_jobs < centity->sequence) { 794 spin_unlock(&ctx->ring_lock); 795 return NULL; 796 } 797 798 fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]); 799 spin_unlock(&ctx->ring_lock); 800 801 return fence; 802 } 803 804 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx, 805 struct amdgpu_ctx_entity *aentity, 806 int hw_ip, 807 int32_t priority) 808 { 809 struct amdgpu_device *adev = ctx->mgr->adev; 810 unsigned int hw_prio; 811 struct drm_gpu_scheduler **scheds = NULL; 812 unsigned num_scheds; 813 814 /* set sw priority */ 815 drm_sched_entity_set_priority(&aentity->entity, 816 amdgpu_ctx_to_drm_sched_prio(priority)); 817 818 /* set hw priority */ 819 if (hw_ip == AMDGPU_HW_IP_COMPUTE || hw_ip == AMDGPU_HW_IP_GFX) { 820 hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip); 821 hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX); 822 scheds = adev->gpu_sched[hw_ip][hw_prio].sched; 823 num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds; 824 drm_sched_entity_modify_sched(&aentity->entity, scheds, 825 num_scheds); 826 } 827 } 828 829 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, 830 int32_t priority) 831 { 832 int32_t ctx_prio; 833 unsigned i, j; 834 835 ctx->override_priority = priority; 836 837 ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ? 838 ctx->init_priority : ctx->override_priority; 839 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 840 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 841 if (!ctx->entities[i][j]) 842 continue; 843 844 amdgpu_ctx_set_entity_priority(ctx, ctx->entities[i][j], 845 i, ctx_prio); 846 } 847 } 848 } 849 850 int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx, 851 struct drm_sched_entity *entity) 852 { 853 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 854 struct dma_fence *other; 855 unsigned idx; 856 long r; 857 858 spin_lock(&ctx->ring_lock); 859 idx = centity->sequence & (amdgpu_sched_jobs - 1); 860 other = dma_fence_get(centity->fences[idx]); 861 spin_unlock(&ctx->ring_lock); 862 863 if (!other) 864 return 0; 865 866 r = dma_fence_wait(other, true); 867 if (r < 0 && r != -ERESTARTSYS) 868 DRM_ERROR("Error (%ld) waiting for fence!\n", r); 869 870 dma_fence_put(other); 871 return r; 872 } 873 874 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr, 875 struct amdgpu_device *adev) 876 { 877 unsigned int i; 878 879 mgr->adev = adev; 880 mutex_init(&mgr->lock); 881 idr_init_base(&mgr->ctx_handles, 1); 882 883 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) 884 atomic64_set(&mgr->time_spend[i], 0); 885 } 886 887 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout) 888 { 889 struct amdgpu_ctx *ctx; 890 struct idr *idp; 891 uint32_t id, i, j; 892 893 idp = &mgr->ctx_handles; 894 895 mutex_lock(&mgr->lock); 896 idr_for_each_entry(idp, ctx, id) { 897 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 898 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 899 struct drm_sched_entity *entity; 900 901 if (!ctx->entities[i][j]) 902 continue; 903 904 entity = &ctx->entities[i][j]->entity; 905 timeout = drm_sched_entity_flush(entity, timeout); 906 } 907 } 908 } 909 mutex_unlock(&mgr->lock); 910 return timeout; 911 } 912 913 void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr) 914 { 915 struct amdgpu_ctx *ctx; 916 struct idr *idp; 917 uint32_t id, i, j; 918 919 idp = &mgr->ctx_handles; 920 921 idr_for_each_entry(idp, ctx, id) { 922 if (kref_read(&ctx->refcount) != 1) { 923 DRM_ERROR("ctx %p is still alive\n", ctx); 924 continue; 925 } 926 927 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 928 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 929 struct drm_sched_entity *entity; 930 931 if (!ctx->entities[i][j]) 932 continue; 933 934 entity = &ctx->entities[i][j]->entity; 935 drm_sched_entity_fini(entity); 936 } 937 } 938 } 939 } 940 941 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr) 942 { 943 struct amdgpu_ctx *ctx; 944 struct idr *idp; 945 uint32_t id; 946 947 amdgpu_ctx_mgr_entity_fini(mgr); 948 949 idp = &mgr->ctx_handles; 950 951 idr_for_each_entry(idp, ctx, id) { 952 if (kref_put(&ctx->refcount, amdgpu_ctx_fini) != 1) 953 DRM_ERROR("ctx %p is still alive\n", ctx); 954 } 955 956 idr_destroy(&mgr->ctx_handles); 957 mutex_destroy(&mgr->lock); 958 } 959 960 void amdgpu_ctx_mgr_usage(struct amdgpu_ctx_mgr *mgr, 961 ktime_t usage[AMDGPU_HW_IP_NUM]) 962 { 963 struct amdgpu_ctx *ctx; 964 unsigned int hw_ip, i; 965 uint32_t id; 966 967 /* 968 * This is a little bit racy because it can be that a ctx or a fence are 969 * destroyed just in the moment we try to account them. But that is ok 970 * since exactly that case is explicitely allowed by the interface. 971 */ 972 mutex_lock(&mgr->lock); 973 for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) { 974 uint64_t ns = atomic64_read(&mgr->time_spend[hw_ip]); 975 976 usage[hw_ip] = ns_to_ktime(ns); 977 } 978 979 idr_for_each_entry(&mgr->ctx_handles, ctx, id) { 980 for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) { 981 for (i = 0; i < amdgpu_ctx_num_entities[hw_ip]; ++i) { 982 struct amdgpu_ctx_entity *centity; 983 ktime_t spend; 984 985 centity = ctx->entities[hw_ip][i]; 986 if (!centity) 987 continue; 988 spend = amdgpu_ctx_entity_time(ctx, centity); 989 usage[hw_ip] = ktime_add(usage[hw_ip], spend); 990 } 991 } 992 } 993 mutex_unlock(&mgr->lock); 994 } 995