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