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