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 #include <linux/pm_runtime.h> 32 33 #define to_amdgpu_ctx_entity(e) \ 34 container_of((e), struct amdgpu_ctx_entity, entity) 35 36 const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM] = { 37 [AMDGPU_HW_IP_GFX] = 1, 38 [AMDGPU_HW_IP_COMPUTE] = 4, 39 [AMDGPU_HW_IP_DMA] = 2, 40 [AMDGPU_HW_IP_UVD] = 1, 41 [AMDGPU_HW_IP_VCE] = 1, 42 [AMDGPU_HW_IP_UVD_ENC] = 1, 43 [AMDGPU_HW_IP_VCN_DEC] = 1, 44 [AMDGPU_HW_IP_VCN_ENC] = 1, 45 [AMDGPU_HW_IP_VCN_JPEG] = 1, 46 [AMDGPU_HW_IP_VPE] = 1, 47 }; 48 49 bool amdgpu_ctx_priority_is_valid(int32_t ctx_prio) 50 { 51 switch (ctx_prio) { 52 case AMDGPU_CTX_PRIORITY_VERY_LOW: 53 case AMDGPU_CTX_PRIORITY_LOW: 54 case AMDGPU_CTX_PRIORITY_NORMAL: 55 case AMDGPU_CTX_PRIORITY_HIGH: 56 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 57 return true; 58 default: 59 case AMDGPU_CTX_PRIORITY_UNSET: 60 /* UNSET priority is not valid and we don't carry that 61 * around, but set it to NORMAL in the only place this 62 * function is called, amdgpu_ctx_ioctl(). 63 */ 64 return false; 65 } 66 } 67 68 static enum drm_sched_priority 69 amdgpu_ctx_to_drm_sched_prio(int32_t ctx_prio) 70 { 71 switch (ctx_prio) { 72 case AMDGPU_CTX_PRIORITY_UNSET: 73 pr_warn_once("AMD-->DRM context priority value UNSET-->NORMAL"); 74 return DRM_SCHED_PRIORITY_NORMAL; 75 76 case AMDGPU_CTX_PRIORITY_VERY_LOW: 77 return DRM_SCHED_PRIORITY_LOW; 78 79 case AMDGPU_CTX_PRIORITY_LOW: 80 return DRM_SCHED_PRIORITY_LOW; 81 82 case AMDGPU_CTX_PRIORITY_NORMAL: 83 return DRM_SCHED_PRIORITY_NORMAL; 84 85 case AMDGPU_CTX_PRIORITY_HIGH: 86 return DRM_SCHED_PRIORITY_HIGH; 87 88 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 89 return DRM_SCHED_PRIORITY_HIGH; 90 91 /* This should not happen as we sanitized userspace provided priority 92 * already, WARN if this happens. 93 */ 94 default: 95 WARN(1, "Invalid context priority %d\n", ctx_prio); 96 return DRM_SCHED_PRIORITY_NORMAL; 97 } 98 99 } 100 101 static int amdgpu_ctx_priority_permit(struct drm_file *filp, 102 int32_t priority) 103 { 104 /* NORMAL and below are accessible by everyone */ 105 if (priority <= AMDGPU_CTX_PRIORITY_NORMAL) 106 return 0; 107 108 if (capable(CAP_SYS_NICE)) 109 return 0; 110 111 if (drm_is_current_master(filp)) 112 return 0; 113 114 return -EACCES; 115 } 116 117 static enum amdgpu_gfx_pipe_priority amdgpu_ctx_prio_to_gfx_pipe_prio(int32_t prio) 118 { 119 switch (prio) { 120 case AMDGPU_CTX_PRIORITY_HIGH: 121 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 122 return AMDGPU_GFX_PIPE_PRIO_HIGH; 123 default: 124 return AMDGPU_GFX_PIPE_PRIO_NORMAL; 125 } 126 } 127 128 static enum amdgpu_ring_priority_level amdgpu_ctx_sched_prio_to_ring_prio(int32_t prio) 129 { 130 switch (prio) { 131 case AMDGPU_CTX_PRIORITY_HIGH: 132 return AMDGPU_RING_PRIO_1; 133 case AMDGPU_CTX_PRIORITY_VERY_HIGH: 134 return AMDGPU_RING_PRIO_2; 135 default: 136 return AMDGPU_RING_PRIO_0; 137 } 138 } 139 140 static unsigned int amdgpu_ctx_get_hw_prio(struct amdgpu_ctx *ctx, u32 hw_ip) 141 { 142 struct amdgpu_device *adev = ctx->mgr->adev; 143 unsigned int hw_prio; 144 int32_t ctx_prio; 145 146 ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ? 147 ctx->init_priority : ctx->override_priority; 148 149 switch (hw_ip) { 150 case AMDGPU_HW_IP_GFX: 151 case AMDGPU_HW_IP_COMPUTE: 152 hw_prio = amdgpu_ctx_prio_to_gfx_pipe_prio(ctx_prio); 153 break; 154 case AMDGPU_HW_IP_VCE: 155 case AMDGPU_HW_IP_VCN_ENC: 156 hw_prio = amdgpu_ctx_sched_prio_to_ring_prio(ctx_prio); 157 break; 158 default: 159 hw_prio = AMDGPU_RING_PRIO_DEFAULT; 160 break; 161 } 162 163 hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM); 164 if (adev->gpu_sched[hw_ip][hw_prio].num_scheds == 0) 165 hw_prio = AMDGPU_RING_PRIO_DEFAULT; 166 167 return hw_prio; 168 } 169 170 /* Calculate the time spend on the hw */ 171 static ktime_t amdgpu_ctx_fence_time(struct dma_fence *fence) 172 { 173 struct drm_sched_fence *s_fence; 174 175 if (!fence) 176 return ns_to_ktime(0); 177 178 /* When the fence is not even scheduled it can't have spend time */ 179 s_fence = to_drm_sched_fence(fence); 180 if (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &s_fence->scheduled.flags)) 181 return ns_to_ktime(0); 182 183 /* When it is still running account how much already spend */ 184 if (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &s_fence->finished.flags)) 185 return ktime_sub(ktime_get(), s_fence->scheduled.timestamp); 186 187 return ktime_sub(s_fence->finished.timestamp, 188 s_fence->scheduled.timestamp); 189 } 190 191 static ktime_t amdgpu_ctx_entity_time(struct amdgpu_ctx *ctx, 192 struct amdgpu_ctx_entity *centity) 193 { 194 ktime_t res = ns_to_ktime(0); 195 uint32_t i; 196 197 spin_lock(&ctx->ring_lock); 198 for (i = 0; i < amdgpu_sched_jobs; i++) { 199 res = ktime_add(res, amdgpu_ctx_fence_time(centity->fences[i])); 200 } 201 spin_unlock(&ctx->ring_lock); 202 return res; 203 } 204 205 static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip, 206 const u32 ring) 207 { 208 struct drm_gpu_scheduler **scheds = NULL, *sched = NULL; 209 struct amdgpu_device *adev = ctx->mgr->adev; 210 struct amdgpu_ctx_entity *entity; 211 enum drm_sched_priority drm_prio; 212 unsigned int hw_prio, num_scheds; 213 int32_t ctx_prio; 214 int r; 215 216 entity = kzalloc_flex(*entity, fences, amdgpu_sched_jobs); 217 if (!entity) 218 return -ENOMEM; 219 220 ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ? 221 ctx->init_priority : ctx->override_priority; 222 entity->hw_ip = hw_ip; 223 entity->sequence = 1; 224 hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip); 225 drm_prio = amdgpu_ctx_to_drm_sched_prio(ctx_prio); 226 227 hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM); 228 229 if (!(adev)->xcp_mgr) { 230 scheds = adev->gpu_sched[hw_ip][hw_prio].sched; 231 num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds; 232 } else { 233 struct amdgpu_fpriv *fpriv; 234 235 /* TODO: Stop using fpriv here, we only need the xcp_id. */ 236 fpriv = container_of(ctx->mgr, struct amdgpu_fpriv, ctx_mgr); 237 r = amdgpu_xcp_select_scheds(adev, hw_ip, hw_prio, fpriv, 238 &num_scheds, &scheds); 239 if (r) 240 goto error_free_entity; 241 } 242 243 if (num_scheds == 0) { 244 r = -EINVAL; 245 goto error_free_entity; 246 } 247 248 /* disable load balance if the hw engine retains context among dependent jobs */ 249 if (hw_ip == AMDGPU_HW_IP_VCN_ENC || 250 hw_ip == AMDGPU_HW_IP_VCN_DEC || 251 hw_ip == AMDGPU_HW_IP_UVD_ENC || 252 hw_ip == AMDGPU_HW_IP_UVD) { 253 sched = drm_sched_pick_best(scheds, num_scheds); 254 scheds = &sched; 255 num_scheds = 1; 256 } 257 258 r = drm_sched_entity_init(&entity->entity, drm_prio, scheds, num_scheds, 259 NULL); 260 if (r) 261 goto error_free_entity; 262 263 /* It's not an error if we fail to install the new entity */ 264 if (cmpxchg(&ctx->entities[hw_ip][ring], NULL, entity)) 265 goto cleanup_entity; 266 267 return 0; 268 269 cleanup_entity: 270 drm_sched_entity_fini(&entity->entity); 271 272 error_free_entity: 273 kfree(entity); 274 275 return r; 276 } 277 278 static ktime_t amdgpu_ctx_fini_entity(struct amdgpu_device *adev, 279 struct amdgpu_ctx_entity *entity) 280 { 281 ktime_t res = ns_to_ktime(0); 282 int i; 283 284 if (!entity) 285 return res; 286 287 drm_sched_entity_destroy(&entity->entity); 288 289 for (i = 0; i < amdgpu_sched_jobs; ++i) { 290 res = ktime_add(res, amdgpu_ctx_fence_time(entity->fences[i])); 291 dma_fence_put(entity->fences[i]); 292 } 293 294 amdgpu_xcp_release_sched(adev, entity); 295 296 kfree(entity); 297 return res; 298 } 299 300 static u32 amdgpu_get_stable_pstate(struct amdgpu_device *adev) 301 { 302 switch (amdgpu_dpm_get_performance_level(adev)) { 303 case AMD_DPM_FORCED_LEVEL_PROFILE_STANDARD: 304 return AMDGPU_CTX_STABLE_PSTATE_STANDARD; 305 case AMD_DPM_FORCED_LEVEL_PROFILE_MIN_SCLK: 306 return AMDGPU_CTX_STABLE_PSTATE_MIN_SCLK; 307 case AMD_DPM_FORCED_LEVEL_PROFILE_MIN_MCLK: 308 return AMDGPU_CTX_STABLE_PSTATE_MIN_MCLK; 309 case AMD_DPM_FORCED_LEVEL_PROFILE_PEAK: 310 return AMDGPU_CTX_STABLE_PSTATE_PEAK; 311 default: 312 return AMDGPU_CTX_STABLE_PSTATE_NONE; 313 } 314 } 315 316 static int amdgpu_ctx_init(struct amdgpu_ctx_mgr *mgr, int32_t priority, 317 struct drm_file *filp, struct amdgpu_ctx *ctx) 318 { 319 struct amdgpu_fpriv *fpriv = filp->driver_priv; 320 int r; 321 322 r = amdgpu_ctx_priority_permit(filp, priority); 323 if (r) 324 return r; 325 326 memset(ctx, 0, sizeof(*ctx)); 327 328 kref_init(&ctx->refcount); 329 ctx->mgr = mgr; 330 spin_lock_init(&ctx->ring_lock); 331 332 ctx->reset_counter = atomic_read(&mgr->adev->gpu_reset_counter); 333 ctx->reset_counter_query = ctx->reset_counter; 334 ctx->generation = amdgpu_vm_generation(mgr->adev, &fpriv->vm); 335 ctx->init_priority = priority; 336 ctx->override_priority = AMDGPU_CTX_PRIORITY_UNSET; 337 ctx->stable_pstate = AMDGPU_CTX_STABLE_PSTATE_NONE; 338 339 return 0; 340 } 341 342 static int __amdgpu_ctx_set_stable_pstate(struct amdgpu_ctx *ctx, 343 u32 stable_pstate) 344 { 345 struct amdgpu_device *adev = ctx->mgr->adev; 346 enum amd_dpm_forced_level level; 347 struct amdgpu_ctx *current_ctx; 348 u32 current_stable_pstate; 349 int r = 0; 350 351 lockdep_assert_held(&adev->pm.stable_pstate_ctx_lock); 352 353 switch (stable_pstate) { 354 case AMDGPU_CTX_STABLE_PSTATE_NONE: 355 level = AMD_DPM_FORCED_LEVEL_AUTO; 356 break; 357 case AMDGPU_CTX_STABLE_PSTATE_STANDARD: 358 level = AMD_DPM_FORCED_LEVEL_PROFILE_STANDARD; 359 break; 360 case AMDGPU_CTX_STABLE_PSTATE_MIN_SCLK: 361 level = AMD_DPM_FORCED_LEVEL_PROFILE_MIN_SCLK; 362 break; 363 case AMDGPU_CTX_STABLE_PSTATE_MIN_MCLK: 364 level = AMD_DPM_FORCED_LEVEL_PROFILE_MIN_MCLK; 365 break; 366 case AMDGPU_CTX_STABLE_PSTATE_PEAK: 367 level = AMD_DPM_FORCED_LEVEL_PROFILE_PEAK; 368 break; 369 default: 370 return -EINVAL; 371 } 372 373 current_ctx = adev->pm.stable_pstate_ctx; 374 if (current_ctx && current_ctx != ctx) 375 return -EBUSY; 376 377 current_stable_pstate = amdgpu_get_stable_pstate(adev); 378 if (current_stable_pstate == stable_pstate) 379 return 0; 380 381 r = amdgpu_dpm_force_performance_level(adev, level); 382 if (r) 383 return r; 384 385 if (!current_ctx) { 386 adev->pm.stable_pstate_ctx = ctx; 387 /* 388 * Serialized by context taking ownership for the first time 389 * while holding adev->pm.stable_pstate_ctx_lock). 390 */ 391 WRITE_ONCE(ctx->stable_pstate, current_stable_pstate); 392 } 393 394 return 0; 395 } 396 397 static int amdgpu_ctx_set_stable_pstate(struct amdgpu_ctx *ctx, 398 u32 stable_pstate) 399 { 400 struct amdgpu_device *adev = ctx->mgr->adev; 401 int r; 402 403 mutex_lock(&adev->pm.stable_pstate_ctx_lock); 404 r = __amdgpu_ctx_set_stable_pstate(ctx, stable_pstate); 405 mutex_unlock(&adev->pm.stable_pstate_ctx_lock); 406 407 return r; 408 } 409 410 void amdgpu_ctx_fini(struct kref *ref) 411 { 412 struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount); 413 struct amdgpu_ctx_mgr *mgr = ctx->mgr; 414 struct amdgpu_device *adev = mgr->adev; 415 unsigned i, j, idx; 416 417 if (!adev) 418 return; 419 420 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 421 for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) { 422 ktime_t spend; 423 424 spend = amdgpu_ctx_fini_entity(adev, ctx->entities[i][j]); 425 atomic64_add(ktime_to_ns(spend), &mgr->time_spend[i]); 426 } 427 } 428 429 if (drm_dev_enter(adev_to_drm(adev), &idx)) { 430 mutex_lock(&adev->pm.stable_pstate_ctx_lock); 431 if (adev->pm.stable_pstate_ctx == ctx) { 432 __amdgpu_ctx_set_stable_pstate(ctx, ctx->stable_pstate); 433 adev->pm.stable_pstate_ctx = NULL; 434 } 435 mutex_unlock(&adev->pm.stable_pstate_ctx_lock); 436 drm_dev_exit(idx); 437 } 438 439 kfree(ctx); 440 } 441 442 int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance, 443 u32 ring, struct drm_sched_entity **entity) 444 { 445 int r; 446 struct drm_sched_entity *ctx_entity; 447 448 if (hw_ip >= AMDGPU_HW_IP_NUM) { 449 drm_err(adev_to_drm(ctx->mgr->adev), 450 "unknown HW IP type: %d\n", hw_ip); 451 return -EINVAL; 452 } 453 454 /* Right now all IPs have only one instance - multiple rings. */ 455 if (instance != 0) { 456 drm_dbg(adev_to_drm(ctx->mgr->adev), 457 "invalid ip instance: %d\n", instance); 458 return -EINVAL; 459 } 460 461 if (ring >= amdgpu_ctx_num_entities[hw_ip]) { 462 drm_dbg(adev_to_drm(ctx->mgr->adev), 463 "invalid ring: %d %d\n", hw_ip, ring); 464 return -EINVAL; 465 } 466 467 if (ctx->entities[hw_ip][ring] == NULL) { 468 r = amdgpu_ctx_init_entity(ctx, hw_ip, ring); 469 if (r) 470 return r; 471 } 472 473 ctx_entity = &ctx->entities[hw_ip][ring]->entity; 474 r = drm_sched_entity_error(ctx_entity); 475 if (r) { 476 DRM_DEBUG("error entity %p\n", ctx_entity); 477 return r; 478 } 479 480 *entity = ctx_entity; 481 return 0; 482 } 483 484 static int amdgpu_ctx_alloc(struct amdgpu_device *adev, 485 struct amdgpu_fpriv *fpriv, 486 struct drm_file *filp, 487 int32_t priority, 488 uint32_t *id) 489 { 490 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr; 491 struct amdgpu_ctx *ctx; 492 int r; 493 494 ctx = kmalloc_obj(*ctx); 495 if (!ctx) 496 return -ENOMEM; 497 498 r = amdgpu_ctx_init(mgr, priority, filp, ctx); 499 if (r) { 500 kfree(ctx); 501 return r; 502 } 503 504 r = xa_alloc(&mgr->ctx_handles, id, ctx, xa_limit_32b, GFP_KERNEL); 505 if (r) 506 amdgpu_ctx_put(ctx); 507 508 return r; 509 } 510 511 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id) 512 { 513 struct amdgpu_ctx *ctx; 514 515 ctx = xa_erase(&fpriv->ctx_mgr.ctx_handles, id); 516 amdgpu_ctx_put(ctx); 517 518 return ctx ? 0 : -EINVAL; 519 } 520 521 static int amdgpu_ctx_query(struct amdgpu_device *adev, 522 struct amdgpu_fpriv *fpriv, uint32_t id, 523 union drm_amdgpu_ctx_out *out) 524 { 525 struct amdgpu_ctx *ctx; 526 unsigned reset_counter; 527 528 ctx = amdgpu_ctx_get(fpriv, id); 529 if (!ctx) 530 return -EINVAL; 531 532 /* TODO: these two are always zero */ 533 out->state.flags = 0x0; 534 out->state.hangs = 0x0; 535 536 /* determine if a GPU reset has occurred since the last call */ 537 reset_counter = atomic_read(&adev->gpu_reset_counter); 538 /* TODO: this should ideally return NO, GUILTY, or INNOCENT. */ 539 if (ctx->reset_counter_query == reset_counter) 540 out->state.reset_status = AMDGPU_CTX_NO_RESET; 541 else 542 out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET; 543 ctx->reset_counter_query = reset_counter; 544 545 amdgpu_ctx_put(ctx); 546 547 return 0; 548 } 549 550 #define AMDGPU_RAS_COUNTE_DELAY_MS 3000 551 552 static bool amdgpu_ctx_guilty(struct amdgpu_ctx *ctx) 553 { 554 int i, j, r; 555 556 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 557 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 558 struct amdgpu_ctx_entity *ctx_entity; 559 560 ctx_entity = ctx->entities[i][j]; 561 if (!ctx_entity) 562 continue; 563 564 r = drm_sched_entity_error(&ctx_entity->entity); 565 if (r == -ETIME) 566 return true; 567 } 568 } 569 570 return false; 571 } 572 573 static int amdgpu_ctx_query2(struct amdgpu_device *adev, 574 struct amdgpu_fpriv *fpriv, uint32_t id, 575 union drm_amdgpu_ctx_out *out) 576 { 577 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 578 struct amdgpu_ctx *ctx; 579 580 ctx = amdgpu_ctx_get(fpriv, id); 581 if (!ctx) 582 return -EINVAL; 583 584 out->state.flags = 0x0; 585 out->state.hangs = 0x0; 586 587 if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter)) 588 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET; 589 590 if (ctx->generation != amdgpu_vm_generation(adev, &fpriv->vm)) 591 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST; 592 593 if (amdgpu_ctx_guilty(ctx)) 594 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY; 595 596 if (amdgpu_in_reset(adev)) 597 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET_IN_PROGRESS; 598 599 if (adev->ras_enabled && con) { 600 /* Return the cached values in O(1), 601 * and schedule delayed work to cache 602 * new vaues. 603 */ 604 int ce_count, ue_count; 605 606 ce_count = atomic_read(&con->ras_ce_count); 607 ue_count = atomic_read(&con->ras_ue_count); 608 609 if (ce_count != ctx->ras_counter_ce) { 610 ctx->ras_counter_ce = ce_count; 611 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE; 612 } 613 614 if (ue_count != ctx->ras_counter_ue) { 615 ctx->ras_counter_ue = ue_count; 616 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE; 617 } 618 619 schedule_delayed_work(&con->ras_counte_delay_work, 620 msecs_to_jiffies(AMDGPU_RAS_COUNTE_DELAY_MS)); 621 } 622 623 amdgpu_ctx_put(ctx); 624 625 return 0; 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 int r = 0; 634 635 ctx = amdgpu_ctx_get(fpriv, id); 636 if (!ctx) 637 return -EINVAL; 638 639 /* 640 * The get path is odd in this uapi - it will check whether the context 641 * id exist, but otherwise does nothing with it. In other words, the 642 * uapi has historically been implemented as being able to query the 643 * global device state, as long as the caller supplies a random valid 644 * context id. 645 */ 646 647 if (set) { 648 r = pm_runtime_resume_and_get(adev->dev); 649 if (r < 0) 650 return r; 651 r = amdgpu_ctx_set_stable_pstate(ctx, *stable_pstate); 652 pm_runtime_put_autosuspend(adev->dev); 653 } else { 654 *stable_pstate = amdgpu_get_stable_pstate(adev); 655 } 656 657 amdgpu_ctx_put(ctx); 658 return r; 659 } 660 661 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data, 662 struct drm_file *filp) 663 { 664 int r; 665 uint32_t id, stable_pstate; 666 int32_t priority; 667 668 union drm_amdgpu_ctx *args = data; 669 struct amdgpu_device *adev = drm_to_adev(dev); 670 struct amdgpu_fpriv *fpriv = filp->driver_priv; 671 672 id = args->in.ctx_id; 673 priority = args->in.priority; 674 675 /* For backwards compatibility, we need to accept ioctls with garbage 676 * in the priority field. Garbage values in the priority field, result 677 * in the priority being set to NORMAL. 678 */ 679 if (!amdgpu_ctx_priority_is_valid(priority)) 680 priority = AMDGPU_CTX_PRIORITY_NORMAL; 681 682 switch (args->in.op) { 683 case AMDGPU_CTX_OP_ALLOC_CTX: 684 if (args->in.flags) 685 return -EINVAL; 686 r = amdgpu_ctx_alloc(adev, fpriv, filp, priority, &id); 687 args->out.alloc.ctx_id = id; 688 break; 689 case AMDGPU_CTX_OP_FREE_CTX: 690 if (args->in.flags) 691 return -EINVAL; 692 r = amdgpu_ctx_free(fpriv, id); 693 break; 694 case AMDGPU_CTX_OP_QUERY_STATE: 695 if (args->in.flags) 696 return -EINVAL; 697 r = amdgpu_ctx_query(adev, fpriv, id, &args->out); 698 break; 699 case AMDGPU_CTX_OP_QUERY_STATE2: 700 if (args->in.flags) 701 return -EINVAL; 702 r = amdgpu_ctx_query2(adev, fpriv, id, &args->out); 703 break; 704 case AMDGPU_CTX_OP_GET_STABLE_PSTATE: 705 if (args->in.flags) 706 return -EINVAL; 707 r = amdgpu_ctx_stable_pstate(adev, fpriv, id, false, &stable_pstate); 708 if (!r) 709 args->out.pstate.flags = stable_pstate; 710 break; 711 case AMDGPU_CTX_OP_SET_STABLE_PSTATE: 712 if (args->in.flags & ~AMDGPU_CTX_STABLE_PSTATE_FLAGS_MASK) 713 return -EINVAL; 714 stable_pstate = args->in.flags & AMDGPU_CTX_STABLE_PSTATE_FLAGS_MASK; 715 if (stable_pstate > AMDGPU_CTX_STABLE_PSTATE_PEAK) 716 return -EINVAL; 717 r = amdgpu_ctx_stable_pstate(adev, fpriv, id, true, &stable_pstate); 718 break; 719 default: 720 return -EINVAL; 721 } 722 723 return r; 724 } 725 726 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id) 727 { 728 struct amdgpu_ctx *ctx; 729 struct amdgpu_ctx_mgr *mgr; 730 731 if (!fpriv) 732 return NULL; 733 734 mgr = &fpriv->ctx_mgr; 735 736 xa_lock(&mgr->ctx_handles); 737 ctx = xa_load(&mgr->ctx_handles, id); 738 if (ctx) 739 kref_get(&ctx->refcount); 740 xa_unlock(&mgr->ctx_handles); 741 return ctx; 742 } 743 744 uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, 745 struct drm_sched_entity *entity, 746 struct dma_fence *fence) 747 { 748 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 749 uint64_t seq = centity->sequence; 750 struct dma_fence *other = NULL; 751 unsigned idx = 0; 752 753 idx = seq & (amdgpu_sched_jobs - 1); 754 other = centity->fences[idx]; 755 WARN_ON(other && !dma_fence_is_signaled(other)); 756 757 dma_fence_get(fence); 758 759 spin_lock(&ctx->ring_lock); 760 centity->fences[idx] = fence; 761 centity->sequence++; 762 spin_unlock(&ctx->ring_lock); 763 764 atomic64_add(ktime_to_ns(amdgpu_ctx_fence_time(other)), 765 &ctx->mgr->time_spend[centity->hw_ip]); 766 767 dma_fence_put(other); 768 return seq; 769 } 770 771 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx, 772 struct drm_sched_entity *entity, 773 uint64_t seq) 774 { 775 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 776 struct dma_fence *fence; 777 778 spin_lock(&ctx->ring_lock); 779 780 if (seq == ~0ull) 781 seq = centity->sequence - 1; 782 783 if (seq >= centity->sequence) { 784 spin_unlock(&ctx->ring_lock); 785 return ERR_PTR(-EINVAL); 786 } 787 788 789 if (seq + amdgpu_sched_jobs < centity->sequence) { 790 spin_unlock(&ctx->ring_lock); 791 return NULL; 792 } 793 794 fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]); 795 spin_unlock(&ctx->ring_lock); 796 797 return fence; 798 } 799 800 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx, 801 struct amdgpu_ctx_entity *aentity, 802 int hw_ip, 803 int32_t priority) 804 { 805 struct amdgpu_device *adev = ctx->mgr->adev; 806 unsigned int hw_prio; 807 struct drm_gpu_scheduler **scheds = NULL; 808 unsigned num_scheds; 809 810 /* set sw priority */ 811 drm_sched_entity_set_priority(&aentity->entity, 812 amdgpu_ctx_to_drm_sched_prio(priority)); 813 814 /* set hw priority */ 815 if (hw_ip == AMDGPU_HW_IP_COMPUTE || hw_ip == AMDGPU_HW_IP_GFX) { 816 hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip); 817 hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX); 818 scheds = adev->gpu_sched[hw_ip][hw_prio].sched; 819 num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds; 820 drm_sched_entity_modify_sched(&aentity->entity, scheds, 821 num_scheds); 822 } 823 } 824 825 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, 826 int32_t priority) 827 { 828 int32_t ctx_prio; 829 unsigned i, j; 830 831 ctx->override_priority = priority; 832 833 ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ? 834 ctx->init_priority : ctx->override_priority; 835 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 836 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 837 if (!ctx->entities[i][j]) 838 continue; 839 840 amdgpu_ctx_set_entity_priority(ctx, ctx->entities[i][j], 841 i, ctx_prio); 842 } 843 } 844 } 845 846 int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx, 847 struct drm_sched_entity *entity) 848 { 849 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 850 struct dma_fence *other; 851 unsigned idx; 852 long r; 853 854 spin_lock(&ctx->ring_lock); 855 idx = centity->sequence & (amdgpu_sched_jobs - 1); 856 other = dma_fence_get(centity->fences[idx]); 857 spin_unlock(&ctx->ring_lock); 858 859 if (!other) 860 return 0; 861 862 r = dma_fence_wait(other, true); 863 if (r < 0 && r != -ERESTARTSYS) 864 drm_err(adev_to_drm(ctx->mgr->adev), 865 "AMDGPU: Error waiting for fence in ctx %p\n", ctx); 866 867 dma_fence_put(other); 868 return r; 869 } 870 871 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr, 872 struct amdgpu_device *adev) 873 { 874 unsigned int i; 875 876 mgr->adev = adev; 877 xa_init_flags(&mgr->ctx_handles, XA_FLAGS_ALLOC1); 878 879 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) 880 atomic64_set(&mgr->time_spend[i], 0); 881 } 882 883 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout) 884 { 885 struct amdgpu_ctx *ctx; 886 unsigned long id; 887 int i, j; 888 889 xa_lock(&mgr->ctx_handles); 890 xa_for_each(&mgr->ctx_handles, id, ctx) { 891 kref_get(&ctx->refcount); 892 xa_unlock(&mgr->ctx_handles); 893 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 894 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 895 struct drm_sched_entity *entity; 896 897 if (!ctx->entities[i][j]) 898 continue; 899 900 entity = &ctx->entities[i][j]->entity; 901 timeout = drm_sched_entity_flush(entity, timeout); 902 } 903 } 904 amdgpu_ctx_put(ctx); 905 xa_lock(&mgr->ctx_handles); 906 } 907 xa_unlock(&mgr->ctx_handles); 908 return timeout; 909 } 910 911 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr) 912 { 913 struct amdgpu_ctx *ctx; 914 unsigned long id; 915 916 xa_for_each(&mgr->ctx_handles, id, ctx) 917 amdgpu_ctx_put(ctx); 918 xa_destroy(&mgr->ctx_handles); 919 } 920 921 void amdgpu_ctx_mgr_usage(struct amdgpu_ctx_mgr *mgr, 922 ktime_t usage[AMDGPU_HW_IP_NUM]) 923 { 924 struct amdgpu_ctx *ctx; 925 unsigned int hw_ip, i; 926 unsigned long id; 927 928 /* 929 * This is a little bit racy because it can be that a ctx or a fence are 930 * destroyed just in the moment we try to account them. But that is ok 931 * since exactly that case is explicitely allowed by the interface. 932 */ 933 for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) { 934 uint64_t ns = atomic64_read(&mgr->time_spend[hw_ip]); 935 936 usage[hw_ip] = ns_to_ktime(ns); 937 } 938 939 xa_lock(&mgr->ctx_handles); 940 xa_for_each(&mgr->ctx_handles, id, ctx) { 941 for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) { 942 for (i = 0; i < amdgpu_ctx_num_entities[hw_ip]; ++i) { 943 struct amdgpu_ctx_entity *centity; 944 ktime_t spend; 945 946 centity = ctx->entities[hw_ip][i]; 947 if (!centity) 948 continue; 949 spend = amdgpu_ctx_entity_time(ctx, centity); 950 usage[hw_ip] = ktime_add(usage[hw_ip], spend); 951 } 952 } 953 } 954 xa_unlock(&mgr->ctx_handles); 955 } 956