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 static int amdgpu_ctx_priority_permit(struct drm_file *filp, 47 enum drm_sched_priority priority) 48 { 49 if (priority < 0 || priority >= DRM_SCHED_PRIORITY_MAX) 50 return -EINVAL; 51 52 /* NORMAL and below are accessible by everyone */ 53 if (priority <= DRM_SCHED_PRIORITY_NORMAL) 54 return 0; 55 56 if (capable(CAP_SYS_NICE)) 57 return 0; 58 59 if (drm_is_current_master(filp)) 60 return 0; 61 62 return -EACCES; 63 } 64 65 static enum gfx_pipe_priority amdgpu_ctx_sched_prio_to_compute_prio(enum drm_sched_priority prio) 66 { 67 switch (prio) { 68 case DRM_SCHED_PRIORITY_HIGH_HW: 69 case DRM_SCHED_PRIORITY_KERNEL: 70 return AMDGPU_GFX_PIPE_PRIO_HIGH; 71 default: 72 return AMDGPU_GFX_PIPE_PRIO_NORMAL; 73 } 74 } 75 76 static unsigned int amdgpu_ctx_prio_sched_to_hw(struct amdgpu_device *adev, 77 enum drm_sched_priority prio, 78 u32 hw_ip) 79 { 80 unsigned int hw_prio; 81 82 hw_prio = (hw_ip == AMDGPU_HW_IP_COMPUTE) ? 83 amdgpu_ctx_sched_prio_to_compute_prio(prio) : 84 AMDGPU_RING_PRIO_DEFAULT; 85 hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM); 86 if (adev->gpu_sched[hw_ip][hw_prio].num_scheds == 0) 87 hw_prio = AMDGPU_RING_PRIO_DEFAULT; 88 89 return hw_prio; 90 } 91 92 static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip, 93 const u32 ring) 94 { 95 struct amdgpu_device *adev = ctx->adev; 96 struct amdgpu_ctx_entity *entity; 97 struct drm_gpu_scheduler **scheds = NULL, *sched = NULL; 98 unsigned num_scheds = 0; 99 unsigned int hw_prio; 100 enum drm_sched_priority priority; 101 int r; 102 103 entity = kcalloc(1, offsetof(typeof(*entity), fences[amdgpu_sched_jobs]), 104 GFP_KERNEL); 105 if (!entity) 106 return -ENOMEM; 107 108 entity->sequence = 1; 109 priority = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ? 110 ctx->init_priority : ctx->override_priority; 111 hw_prio = amdgpu_ctx_prio_sched_to_hw(adev, priority, hw_ip); 112 113 hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM); 114 scheds = adev->gpu_sched[hw_ip][hw_prio].sched; 115 num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds; 116 117 if (hw_ip == AMDGPU_HW_IP_VCN_ENC || hw_ip == AMDGPU_HW_IP_VCN_DEC) { 118 sched = drm_sched_pick_best(scheds, num_scheds); 119 scheds = &sched; 120 num_scheds = 1; 121 } 122 123 r = drm_sched_entity_init(&entity->entity, priority, scheds, num_scheds, 124 &ctx->guilty); 125 if (r) 126 goto error_free_entity; 127 128 ctx->entities[hw_ip][ring] = entity; 129 return 0; 130 131 error_free_entity: 132 kfree(entity); 133 134 return r; 135 } 136 137 static int amdgpu_ctx_init(struct amdgpu_device *adev, 138 enum drm_sched_priority priority, 139 struct drm_file *filp, 140 struct amdgpu_ctx *ctx) 141 { 142 int r; 143 144 r = amdgpu_ctx_priority_permit(filp, priority); 145 if (r) 146 return r; 147 148 memset(ctx, 0, sizeof(*ctx)); 149 150 ctx->adev = adev; 151 152 kref_init(&ctx->refcount); 153 spin_lock_init(&ctx->ring_lock); 154 mutex_init(&ctx->lock); 155 156 ctx->reset_counter = atomic_read(&adev->gpu_reset_counter); 157 ctx->reset_counter_query = ctx->reset_counter; 158 ctx->vram_lost_counter = atomic_read(&adev->vram_lost_counter); 159 ctx->init_priority = priority; 160 ctx->override_priority = DRM_SCHED_PRIORITY_UNSET; 161 162 return 0; 163 } 164 165 static void amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity) 166 { 167 168 int i; 169 170 if (!entity) 171 return; 172 173 for (i = 0; i < amdgpu_sched_jobs; ++i) 174 dma_fence_put(entity->fences[i]); 175 176 kfree(entity); 177 } 178 179 static void amdgpu_ctx_fini(struct kref *ref) 180 { 181 struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount); 182 struct amdgpu_device *adev = ctx->adev; 183 unsigned i, j; 184 185 if (!adev) 186 return; 187 188 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 189 for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) { 190 amdgpu_ctx_fini_entity(ctx->entities[i][j]); 191 ctx->entities[i][j] = NULL; 192 } 193 } 194 195 mutex_destroy(&ctx->lock); 196 kfree(ctx); 197 } 198 199 int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance, 200 u32 ring, struct drm_sched_entity **entity) 201 { 202 int r; 203 204 if (hw_ip >= AMDGPU_HW_IP_NUM) { 205 DRM_ERROR("unknown HW IP type: %d\n", hw_ip); 206 return -EINVAL; 207 } 208 209 /* Right now all IPs have only one instance - multiple rings. */ 210 if (instance != 0) { 211 DRM_DEBUG("invalid ip instance: %d\n", instance); 212 return -EINVAL; 213 } 214 215 if (ring >= amdgpu_ctx_num_entities[hw_ip]) { 216 DRM_DEBUG("invalid ring: %d %d\n", hw_ip, ring); 217 return -EINVAL; 218 } 219 220 if (ctx->entities[hw_ip][ring] == NULL) { 221 r = amdgpu_ctx_init_entity(ctx, hw_ip, ring); 222 if (r) 223 return r; 224 } 225 226 *entity = &ctx->entities[hw_ip][ring]->entity; 227 return 0; 228 } 229 230 static int amdgpu_ctx_alloc(struct amdgpu_device *adev, 231 struct amdgpu_fpriv *fpriv, 232 struct drm_file *filp, 233 enum drm_sched_priority priority, 234 uint32_t *id) 235 { 236 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr; 237 struct amdgpu_ctx *ctx; 238 int r; 239 240 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); 241 if (!ctx) 242 return -ENOMEM; 243 244 mutex_lock(&mgr->lock); 245 r = idr_alloc(&mgr->ctx_handles, ctx, 1, AMDGPU_VM_MAX_NUM_CTX, GFP_KERNEL); 246 if (r < 0) { 247 mutex_unlock(&mgr->lock); 248 kfree(ctx); 249 return r; 250 } 251 252 *id = (uint32_t)r; 253 r = amdgpu_ctx_init(adev, priority, filp, ctx); 254 if (r) { 255 idr_remove(&mgr->ctx_handles, *id); 256 *id = 0; 257 kfree(ctx); 258 } 259 mutex_unlock(&mgr->lock); 260 return r; 261 } 262 263 static void amdgpu_ctx_do_release(struct kref *ref) 264 { 265 struct amdgpu_ctx *ctx; 266 u32 i, j; 267 268 ctx = container_of(ref, struct amdgpu_ctx, refcount); 269 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 270 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 271 if (!ctx->entities[i][j]) 272 continue; 273 274 drm_sched_entity_destroy(&ctx->entities[i][j]->entity); 275 } 276 } 277 278 amdgpu_ctx_fini(ref); 279 } 280 281 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id) 282 { 283 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr; 284 struct amdgpu_ctx *ctx; 285 286 mutex_lock(&mgr->lock); 287 ctx = idr_remove(&mgr->ctx_handles, id); 288 if (ctx) 289 kref_put(&ctx->refcount, amdgpu_ctx_do_release); 290 mutex_unlock(&mgr->lock); 291 return ctx ? 0 : -EINVAL; 292 } 293 294 static int amdgpu_ctx_query(struct amdgpu_device *adev, 295 struct amdgpu_fpriv *fpriv, uint32_t id, 296 union drm_amdgpu_ctx_out *out) 297 { 298 struct amdgpu_ctx *ctx; 299 struct amdgpu_ctx_mgr *mgr; 300 unsigned reset_counter; 301 302 if (!fpriv) 303 return -EINVAL; 304 305 mgr = &fpriv->ctx_mgr; 306 mutex_lock(&mgr->lock); 307 ctx = idr_find(&mgr->ctx_handles, id); 308 if (!ctx) { 309 mutex_unlock(&mgr->lock); 310 return -EINVAL; 311 } 312 313 /* TODO: these two are always zero */ 314 out->state.flags = 0x0; 315 out->state.hangs = 0x0; 316 317 /* determine if a GPU reset has occured since the last call */ 318 reset_counter = atomic_read(&adev->gpu_reset_counter); 319 /* TODO: this should ideally return NO, GUILTY, or INNOCENT. */ 320 if (ctx->reset_counter_query == reset_counter) 321 out->state.reset_status = AMDGPU_CTX_NO_RESET; 322 else 323 out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET; 324 ctx->reset_counter_query = reset_counter; 325 326 mutex_unlock(&mgr->lock); 327 return 0; 328 } 329 330 static int amdgpu_ctx_query2(struct amdgpu_device *adev, 331 struct amdgpu_fpriv *fpriv, uint32_t id, 332 union drm_amdgpu_ctx_out *out) 333 { 334 struct amdgpu_ctx *ctx; 335 struct amdgpu_ctx_mgr *mgr; 336 unsigned long ras_counter; 337 338 if (!fpriv) 339 return -EINVAL; 340 341 mgr = &fpriv->ctx_mgr; 342 mutex_lock(&mgr->lock); 343 ctx = idr_find(&mgr->ctx_handles, id); 344 if (!ctx) { 345 mutex_unlock(&mgr->lock); 346 return -EINVAL; 347 } 348 349 out->state.flags = 0x0; 350 out->state.hangs = 0x0; 351 352 if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter)) 353 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET; 354 355 if (ctx->vram_lost_counter != atomic_read(&adev->vram_lost_counter)) 356 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST; 357 358 if (atomic_read(&ctx->guilty)) 359 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY; 360 361 /*query ue count*/ 362 ras_counter = amdgpu_ras_query_error_count(adev, false); 363 /*ras counter is monotonic increasing*/ 364 if (ras_counter != ctx->ras_counter_ue) { 365 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE; 366 ctx->ras_counter_ue = ras_counter; 367 } 368 369 /*query ce count*/ 370 ras_counter = amdgpu_ras_query_error_count(adev, true); 371 if (ras_counter != ctx->ras_counter_ce) { 372 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE; 373 ctx->ras_counter_ce = ras_counter; 374 } 375 376 mutex_unlock(&mgr->lock); 377 return 0; 378 } 379 380 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data, 381 struct drm_file *filp) 382 { 383 int r; 384 uint32_t id; 385 enum drm_sched_priority priority; 386 387 union drm_amdgpu_ctx *args = data; 388 struct amdgpu_device *adev = dev->dev_private; 389 struct amdgpu_fpriv *fpriv = filp->driver_priv; 390 391 r = 0; 392 id = args->in.ctx_id; 393 priority = amdgpu_to_sched_priority(args->in.priority); 394 395 /* For backwards compatibility reasons, we need to accept 396 * ioctls with garbage in the priority field */ 397 if (priority == DRM_SCHED_PRIORITY_INVALID) 398 priority = DRM_SCHED_PRIORITY_NORMAL; 399 400 switch (args->in.op) { 401 case AMDGPU_CTX_OP_ALLOC_CTX: 402 r = amdgpu_ctx_alloc(adev, fpriv, filp, priority, &id); 403 args->out.alloc.ctx_id = id; 404 break; 405 case AMDGPU_CTX_OP_FREE_CTX: 406 r = amdgpu_ctx_free(fpriv, id); 407 break; 408 case AMDGPU_CTX_OP_QUERY_STATE: 409 r = amdgpu_ctx_query(adev, fpriv, id, &args->out); 410 break; 411 case AMDGPU_CTX_OP_QUERY_STATE2: 412 r = amdgpu_ctx_query2(adev, fpriv, id, &args->out); 413 break; 414 default: 415 return -EINVAL; 416 } 417 418 return r; 419 } 420 421 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id) 422 { 423 struct amdgpu_ctx *ctx; 424 struct amdgpu_ctx_mgr *mgr; 425 426 if (!fpriv) 427 return NULL; 428 429 mgr = &fpriv->ctx_mgr; 430 431 mutex_lock(&mgr->lock); 432 ctx = idr_find(&mgr->ctx_handles, id); 433 if (ctx) 434 kref_get(&ctx->refcount); 435 mutex_unlock(&mgr->lock); 436 return ctx; 437 } 438 439 int amdgpu_ctx_put(struct amdgpu_ctx *ctx) 440 { 441 if (ctx == NULL) 442 return -EINVAL; 443 444 kref_put(&ctx->refcount, amdgpu_ctx_do_release); 445 return 0; 446 } 447 448 void amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, 449 struct drm_sched_entity *entity, 450 struct dma_fence *fence, uint64_t* handle) 451 { 452 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 453 uint64_t seq = centity->sequence; 454 struct dma_fence *other = NULL; 455 unsigned idx = 0; 456 457 idx = seq & (amdgpu_sched_jobs - 1); 458 other = centity->fences[idx]; 459 if (other) 460 BUG_ON(!dma_fence_is_signaled(other)); 461 462 dma_fence_get(fence); 463 464 spin_lock(&ctx->ring_lock); 465 centity->fences[idx] = fence; 466 centity->sequence++; 467 spin_unlock(&ctx->ring_lock); 468 469 dma_fence_put(other); 470 if (handle) 471 *handle = seq; 472 } 473 474 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx, 475 struct drm_sched_entity *entity, 476 uint64_t seq) 477 { 478 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 479 struct dma_fence *fence; 480 481 spin_lock(&ctx->ring_lock); 482 483 if (seq == ~0ull) 484 seq = centity->sequence - 1; 485 486 if (seq >= centity->sequence) { 487 spin_unlock(&ctx->ring_lock); 488 return ERR_PTR(-EINVAL); 489 } 490 491 492 if (seq + amdgpu_sched_jobs < centity->sequence) { 493 spin_unlock(&ctx->ring_lock); 494 return NULL; 495 } 496 497 fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]); 498 spin_unlock(&ctx->ring_lock); 499 500 return fence; 501 } 502 503 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx, 504 struct amdgpu_ctx_entity *aentity, 505 int hw_ip, 506 enum drm_sched_priority priority) 507 { 508 struct amdgpu_device *adev = ctx->adev; 509 unsigned int hw_prio; 510 struct drm_gpu_scheduler **scheds = NULL; 511 unsigned num_scheds; 512 513 /* set sw priority */ 514 drm_sched_entity_set_priority(&aentity->entity, priority); 515 516 /* set hw priority */ 517 if (hw_ip == AMDGPU_HW_IP_COMPUTE) { 518 hw_prio = amdgpu_ctx_prio_sched_to_hw(adev, priority, 519 AMDGPU_HW_IP_COMPUTE); 520 hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX); 521 scheds = adev->gpu_sched[hw_ip][hw_prio].sched; 522 num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds; 523 drm_sched_entity_modify_sched(&aentity->entity, scheds, 524 num_scheds); 525 } 526 } 527 528 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, 529 enum drm_sched_priority priority) 530 { 531 enum drm_sched_priority ctx_prio; 532 unsigned i, j; 533 534 ctx->override_priority = priority; 535 536 ctx_prio = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ? 537 ctx->init_priority : ctx->override_priority; 538 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 539 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 540 if (!ctx->entities[i][j]) 541 continue; 542 543 amdgpu_ctx_set_entity_priority(ctx, ctx->entities[i][j], 544 i, ctx_prio); 545 } 546 } 547 } 548 549 int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx, 550 struct drm_sched_entity *entity) 551 { 552 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); 553 struct dma_fence *other; 554 unsigned idx; 555 long r; 556 557 spin_lock(&ctx->ring_lock); 558 idx = centity->sequence & (amdgpu_sched_jobs - 1); 559 other = dma_fence_get(centity->fences[idx]); 560 spin_unlock(&ctx->ring_lock); 561 562 if (!other) 563 return 0; 564 565 r = dma_fence_wait(other, true); 566 if (r < 0 && r != -ERESTARTSYS) 567 DRM_ERROR("Error (%ld) waiting for fence!\n", r); 568 569 dma_fence_put(other); 570 return r; 571 } 572 573 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr) 574 { 575 mutex_init(&mgr->lock); 576 idr_init(&mgr->ctx_handles); 577 } 578 579 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout) 580 { 581 struct amdgpu_ctx *ctx; 582 struct idr *idp; 583 uint32_t id, i, j; 584 585 idp = &mgr->ctx_handles; 586 587 mutex_lock(&mgr->lock); 588 idr_for_each_entry(idp, ctx, id) { 589 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 590 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 591 struct drm_sched_entity *entity; 592 593 if (!ctx->entities[i][j]) 594 continue; 595 596 entity = &ctx->entities[i][j]->entity; 597 timeout = drm_sched_entity_flush(entity, timeout); 598 } 599 } 600 } 601 mutex_unlock(&mgr->lock); 602 return timeout; 603 } 604 605 void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr) 606 { 607 struct amdgpu_ctx *ctx; 608 struct idr *idp; 609 uint32_t id, i, j; 610 611 idp = &mgr->ctx_handles; 612 613 idr_for_each_entry(idp, ctx, id) { 614 if (kref_read(&ctx->refcount) != 1) { 615 DRM_ERROR("ctx %p is still alive\n", ctx); 616 continue; 617 } 618 619 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) { 620 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) { 621 struct drm_sched_entity *entity; 622 623 if (!ctx->entities[i][j]) 624 continue; 625 626 entity = &ctx->entities[i][j]->entity; 627 drm_sched_entity_fini(entity); 628 } 629 } 630 } 631 } 632 633 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr) 634 { 635 struct amdgpu_ctx *ctx; 636 struct idr *idp; 637 uint32_t id; 638 639 amdgpu_ctx_mgr_entity_fini(mgr); 640 641 idp = &mgr->ctx_handles; 642 643 idr_for_each_entry(idp, ctx, id) { 644 if (kref_put(&ctx->refcount, amdgpu_ctx_fini) != 1) 645 DRM_ERROR("ctx %p is still alive\n", ctx); 646 } 647 648 idr_destroy(&mgr->ctx_handles); 649 mutex_destroy(&mgr->lock); 650 } 651