1 // SPDX-License-Identifier: GPL-2.0+ 2 /* 3 * Copyright (C) 2014-2018 Broadcom 4 * Copyright (C) 2023 Raspberry Pi 5 */ 6 7 #include <drm/drm_syncobj.h> 8 9 #include "v3d_drv.h" 10 #include "v3d_regs.h" 11 #include "v3d_trace.h" 12 13 /* Takes the reservation lock on all the BOs being referenced, so that 14 * at queue submit time we can update the reservations. 15 * 16 * We don't lock the RCL the tile alloc/state BOs, or overflow memory 17 * (all of which are on exec->unref_list). They're entirely private 18 * to v3d, so we don't attach dma-buf fences to them. 19 */ 20 static int 21 v3d_lock_bo_reservations(struct v3d_job *job, 22 struct ww_acquire_ctx *acquire_ctx) 23 { 24 int i, ret; 25 26 ret = drm_gem_lock_reservations(job->bo, job->bo_count, acquire_ctx); 27 if (ret) 28 return ret; 29 30 for (i = 0; i < job->bo_count; i++) { 31 ret = dma_resv_reserve_fences(job->bo[i]->resv, 1); 32 if (ret) 33 goto fail; 34 35 ret = drm_sched_job_add_implicit_dependencies(&job->base, 36 job->bo[i], true); 37 if (ret) 38 goto fail; 39 } 40 41 return 0; 42 43 fail: 44 drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx); 45 return ret; 46 } 47 48 /** 49 * v3d_lookup_bos() - Sets up job->bo[] with the GEM objects 50 * referenced by the job. 51 * @dev: DRM device 52 * @file_priv: DRM file for this fd 53 * @job: V3D job being set up 54 * @bo_handles: GEM handles 55 * @bo_count: Number of GEM handles passed in 56 * 57 * The command validator needs to reference BOs by their index within 58 * the submitted job's BO list. This does the validation of the job's 59 * BO list and reference counting for the lifetime of the job. 60 * 61 * Note that this function doesn't need to unreference the BOs on 62 * failure, because that will happen at v3d_exec_cleanup() time. 63 */ 64 static int 65 v3d_lookup_bos(struct drm_device *dev, 66 struct drm_file *file_priv, 67 struct v3d_job *job, 68 u64 bo_handles, 69 u32 bo_count) 70 { 71 job->bo_count = bo_count; 72 73 if (!job->bo_count) { 74 /* See comment on bo_index for why we have to check 75 * this. 76 */ 77 DRM_DEBUG("Rendering requires BOs\n"); 78 return -EINVAL; 79 } 80 81 return drm_gem_objects_lookup(file_priv, 82 (void __user *)(uintptr_t)bo_handles, 83 job->bo_count, &job->bo); 84 } 85 86 static void 87 v3d_job_free(struct kref *ref) 88 { 89 struct v3d_job *job = container_of(ref, struct v3d_job, refcount); 90 int i; 91 92 if (job->bo) { 93 for (i = 0; i < job->bo_count; i++) 94 drm_gem_object_put(job->bo[i]); 95 kvfree(job->bo); 96 } 97 98 dma_fence_put(job->irq_fence); 99 dma_fence_put(job->done_fence); 100 101 if (job->perfmon) 102 v3d_perfmon_put(job->perfmon); 103 104 kfree(job); 105 } 106 107 static void 108 v3d_render_job_free(struct kref *ref) 109 { 110 struct v3d_render_job *job = container_of(ref, struct v3d_render_job, 111 base.refcount); 112 struct v3d_bo *bo, *save; 113 114 list_for_each_entry_safe(bo, save, &job->unref_list, unref_head) { 115 drm_gem_object_put(&bo->base.base); 116 } 117 118 v3d_job_free(ref); 119 } 120 121 void v3d_job_cleanup(struct v3d_job *job) 122 { 123 if (!job) 124 return; 125 126 drm_sched_job_cleanup(&job->base); 127 v3d_job_put(job); 128 } 129 130 void v3d_job_put(struct v3d_job *job) 131 { 132 if (!job) 133 return; 134 135 kref_put(&job->refcount, job->free); 136 } 137 138 static int 139 v3d_job_allocate(void **container, size_t size) 140 { 141 *container = kcalloc(1, size, GFP_KERNEL); 142 if (!*container) { 143 DRM_ERROR("Cannot allocate memory for V3D job.\n"); 144 return -ENOMEM; 145 } 146 147 return 0; 148 } 149 150 static void 151 v3d_job_deallocate(void **container) 152 { 153 kfree(*container); 154 *container = NULL; 155 } 156 157 static int 158 v3d_job_init(struct v3d_dev *v3d, struct drm_file *file_priv, 159 struct v3d_job *job, void (*free)(struct kref *ref), 160 u32 in_sync, struct v3d_submit_ext *se, enum v3d_queue queue) 161 { 162 struct v3d_file_priv *v3d_priv = file_priv->driver_priv; 163 bool has_multisync = se && (se->flags & DRM_V3D_EXT_ID_MULTI_SYNC); 164 int ret, i; 165 166 job->v3d = v3d; 167 job->free = free; 168 job->file = file_priv; 169 170 ret = drm_sched_job_init(&job->base, &v3d_priv->sched_entity[queue], 171 1, v3d_priv); 172 if (ret) 173 return ret; 174 175 if (has_multisync) { 176 if (se->in_sync_count && se->wait_stage == queue) { 177 struct drm_v3d_sem __user *handle = u64_to_user_ptr(se->in_syncs); 178 179 for (i = 0; i < se->in_sync_count; i++) { 180 struct drm_v3d_sem in; 181 182 if (copy_from_user(&in, handle++, sizeof(in))) { 183 ret = -EFAULT; 184 DRM_DEBUG("Failed to copy wait dep handle.\n"); 185 goto fail_deps; 186 } 187 ret = drm_sched_job_add_syncobj_dependency(&job->base, file_priv, in.handle, 0); 188 189 // TODO: Investigate why this was filtered out for the IOCTL. 190 if (ret && ret != -ENOENT) 191 goto fail_deps; 192 } 193 } 194 } else { 195 ret = drm_sched_job_add_syncobj_dependency(&job->base, file_priv, in_sync, 0); 196 197 // TODO: Investigate why this was filtered out for the IOCTL. 198 if (ret && ret != -ENOENT) 199 goto fail_deps; 200 } 201 202 kref_init(&job->refcount); 203 204 return 0; 205 206 fail_deps: 207 drm_sched_job_cleanup(&job->base); 208 return ret; 209 } 210 211 static void 212 v3d_push_job(struct v3d_job *job) 213 { 214 drm_sched_job_arm(&job->base); 215 216 job->done_fence = dma_fence_get(&job->base.s_fence->finished); 217 218 /* put by scheduler job completion */ 219 kref_get(&job->refcount); 220 221 drm_sched_entity_push_job(&job->base); 222 } 223 224 static void 225 v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv, 226 struct v3d_job *job, 227 struct ww_acquire_ctx *acquire_ctx, 228 u32 out_sync, 229 struct v3d_submit_ext *se, 230 struct dma_fence *done_fence) 231 { 232 struct drm_syncobj *sync_out; 233 bool has_multisync = se && (se->flags & DRM_V3D_EXT_ID_MULTI_SYNC); 234 int i; 235 236 for (i = 0; i < job->bo_count; i++) { 237 /* XXX: Use shared fences for read-only objects. */ 238 dma_resv_add_fence(job->bo[i]->resv, job->done_fence, 239 DMA_RESV_USAGE_WRITE); 240 } 241 242 drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx); 243 244 /* Update the return sync object for the job */ 245 /* If it only supports a single signal semaphore*/ 246 if (!has_multisync) { 247 sync_out = drm_syncobj_find(file_priv, out_sync); 248 if (sync_out) { 249 drm_syncobj_replace_fence(sync_out, done_fence); 250 drm_syncobj_put(sync_out); 251 } 252 return; 253 } 254 255 /* If multiple semaphores extension is supported */ 256 if (se->out_sync_count) { 257 for (i = 0; i < se->out_sync_count; i++) { 258 drm_syncobj_replace_fence(se->out_syncs[i].syncobj, 259 done_fence); 260 drm_syncobj_put(se->out_syncs[i].syncobj); 261 } 262 kvfree(se->out_syncs); 263 } 264 } 265 266 static int 267 v3d_setup_csd_jobs_and_bos(struct drm_file *file_priv, 268 struct v3d_dev *v3d, 269 struct drm_v3d_submit_csd *args, 270 struct v3d_csd_job **job, 271 struct v3d_job **clean_job, 272 struct v3d_submit_ext *se, 273 struct ww_acquire_ctx *acquire_ctx) 274 { 275 int ret; 276 277 ret = v3d_job_allocate((void *)job, sizeof(**job)); 278 if (ret) 279 return ret; 280 281 ret = v3d_job_init(v3d, file_priv, &(*job)->base, 282 v3d_job_free, args->in_sync, se, V3D_CSD); 283 if (ret) { 284 v3d_job_deallocate((void *)job); 285 return ret; 286 } 287 288 ret = v3d_job_allocate((void *)clean_job, sizeof(**clean_job)); 289 if (ret) 290 return ret; 291 292 ret = v3d_job_init(v3d, file_priv, *clean_job, 293 v3d_job_free, 0, NULL, V3D_CACHE_CLEAN); 294 if (ret) { 295 v3d_job_deallocate((void *)clean_job); 296 return ret; 297 } 298 299 (*job)->args = *args; 300 301 ret = v3d_lookup_bos(&v3d->drm, file_priv, *clean_job, 302 args->bo_handles, args->bo_handle_count); 303 if (ret) 304 return ret; 305 306 return v3d_lock_bo_reservations(*clean_job, acquire_ctx); 307 } 308 309 static void 310 v3d_put_multisync_post_deps(struct v3d_submit_ext *se) 311 { 312 unsigned int i; 313 314 if (!(se && se->out_sync_count)) 315 return; 316 317 for (i = 0; i < se->out_sync_count; i++) 318 drm_syncobj_put(se->out_syncs[i].syncobj); 319 kvfree(se->out_syncs); 320 } 321 322 static int 323 v3d_get_multisync_post_deps(struct drm_file *file_priv, 324 struct v3d_submit_ext *se, 325 u32 count, u64 handles) 326 { 327 struct drm_v3d_sem __user *post_deps; 328 int i, ret; 329 330 if (!count) 331 return 0; 332 333 se->out_syncs = (struct v3d_submit_outsync *) 334 kvmalloc_array(count, 335 sizeof(struct v3d_submit_outsync), 336 GFP_KERNEL); 337 if (!se->out_syncs) 338 return -ENOMEM; 339 340 post_deps = u64_to_user_ptr(handles); 341 342 for (i = 0; i < count; i++) { 343 struct drm_v3d_sem out; 344 345 if (copy_from_user(&out, post_deps++, sizeof(out))) { 346 ret = -EFAULT; 347 DRM_DEBUG("Failed to copy post dep handles\n"); 348 goto fail; 349 } 350 351 se->out_syncs[i].syncobj = drm_syncobj_find(file_priv, 352 out.handle); 353 if (!se->out_syncs[i].syncobj) { 354 ret = -EINVAL; 355 goto fail; 356 } 357 } 358 se->out_sync_count = count; 359 360 return 0; 361 362 fail: 363 for (i--; i >= 0; i--) 364 drm_syncobj_put(se->out_syncs[i].syncobj); 365 kvfree(se->out_syncs); 366 367 return ret; 368 } 369 370 /* Get data for multiple binary semaphores synchronization. Parse syncobj 371 * to be signaled when job completes (out_sync). 372 */ 373 static int 374 v3d_get_multisync_submit_deps(struct drm_file *file_priv, 375 struct drm_v3d_extension __user *ext, 376 struct v3d_submit_ext *se) 377 { 378 struct drm_v3d_multi_sync multisync; 379 int ret; 380 381 if (se->in_sync_count || se->out_sync_count) { 382 DRM_DEBUG("Two multisync extensions were added to the same job."); 383 return -EINVAL; 384 } 385 386 if (copy_from_user(&multisync, ext, sizeof(multisync))) 387 return -EFAULT; 388 389 if (multisync.pad) 390 return -EINVAL; 391 392 ret = v3d_get_multisync_post_deps(file_priv, se, multisync.out_sync_count, 393 multisync.out_syncs); 394 if (ret) 395 return ret; 396 397 se->in_sync_count = multisync.in_sync_count; 398 se->in_syncs = multisync.in_syncs; 399 se->flags |= DRM_V3D_EXT_ID_MULTI_SYNC; 400 se->wait_stage = multisync.wait_stage; 401 402 return 0; 403 } 404 405 /* Get data for the indirect CSD job submission. */ 406 static int 407 v3d_get_cpu_indirect_csd_params(struct drm_file *file_priv, 408 struct drm_v3d_extension __user *ext, 409 struct v3d_cpu_job *job) 410 { 411 struct v3d_file_priv *v3d_priv = file_priv->driver_priv; 412 struct v3d_dev *v3d = v3d_priv->v3d; 413 struct drm_v3d_indirect_csd indirect_csd; 414 struct v3d_indirect_csd_info *info = &job->indirect_csd; 415 416 if (!job) { 417 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 418 return -EINVAL; 419 } 420 421 if (job->job_type) { 422 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 423 return -EINVAL; 424 } 425 426 if (copy_from_user(&indirect_csd, ext, sizeof(indirect_csd))) 427 return -EFAULT; 428 429 if (!v3d_has_csd(v3d)) { 430 DRM_DEBUG("Attempting CSD submit on non-CSD hardware.\n"); 431 return -EINVAL; 432 } 433 434 job->job_type = V3D_CPU_JOB_TYPE_INDIRECT_CSD; 435 info->offset = indirect_csd.offset; 436 info->wg_size = indirect_csd.wg_size; 437 memcpy(&info->wg_uniform_offsets, &indirect_csd.wg_uniform_offsets, 438 sizeof(indirect_csd.wg_uniform_offsets)); 439 440 info->indirect = drm_gem_object_lookup(file_priv, indirect_csd.indirect); 441 442 return v3d_setup_csd_jobs_and_bos(file_priv, v3d, &indirect_csd.submit, 443 &info->job, &info->clean_job, 444 NULL, &info->acquire_ctx); 445 } 446 447 /* Get data for the query timestamp job submission. */ 448 static int 449 v3d_get_cpu_timestamp_query_params(struct drm_file *file_priv, 450 struct drm_v3d_extension __user *ext, 451 struct v3d_cpu_job *job) 452 { 453 u32 __user *offsets, *syncs; 454 struct drm_v3d_timestamp_query timestamp; 455 456 if (!job) { 457 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 458 return -EINVAL; 459 } 460 461 if (job->job_type) { 462 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 463 return -EINVAL; 464 } 465 466 if (copy_from_user(×tamp, ext, sizeof(timestamp))) 467 return -EFAULT; 468 469 if (timestamp.pad) 470 return -EINVAL; 471 472 job->job_type = V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY; 473 474 job->timestamp_query.queries = kvmalloc_array(timestamp.count, 475 sizeof(struct v3d_timestamp_query), 476 GFP_KERNEL); 477 if (!job->timestamp_query.queries) 478 return -ENOMEM; 479 480 offsets = u64_to_user_ptr(timestamp.offsets); 481 syncs = u64_to_user_ptr(timestamp.syncs); 482 483 for (int i = 0; i < timestamp.count; i++) { 484 u32 offset, sync; 485 486 if (copy_from_user(&offset, offsets++, sizeof(offset))) { 487 kvfree(job->timestamp_query.queries); 488 return -EFAULT; 489 } 490 491 job->timestamp_query.queries[i].offset = offset; 492 493 if (copy_from_user(&sync, syncs++, sizeof(sync))) { 494 kvfree(job->timestamp_query.queries); 495 return -EFAULT; 496 } 497 498 job->timestamp_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync); 499 } 500 job->timestamp_query.count = timestamp.count; 501 502 return 0; 503 } 504 505 static int 506 v3d_get_cpu_reset_timestamp_params(struct drm_file *file_priv, 507 struct drm_v3d_extension __user *ext, 508 struct v3d_cpu_job *job) 509 { 510 u32 __user *syncs; 511 struct drm_v3d_reset_timestamp_query reset; 512 513 if (!job) { 514 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 515 return -EINVAL; 516 } 517 518 if (job->job_type) { 519 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 520 return -EINVAL; 521 } 522 523 if (copy_from_user(&reset, ext, sizeof(reset))) 524 return -EFAULT; 525 526 job->job_type = V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY; 527 528 job->timestamp_query.queries = kvmalloc_array(reset.count, 529 sizeof(struct v3d_timestamp_query), 530 GFP_KERNEL); 531 if (!job->timestamp_query.queries) 532 return -ENOMEM; 533 534 syncs = u64_to_user_ptr(reset.syncs); 535 536 for (int i = 0; i < reset.count; i++) { 537 u32 sync; 538 539 job->timestamp_query.queries[i].offset = reset.offset + 8 * i; 540 541 if (copy_from_user(&sync, syncs++, sizeof(sync))) { 542 kvfree(job->timestamp_query.queries); 543 return -EFAULT; 544 } 545 546 job->timestamp_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync); 547 } 548 job->timestamp_query.count = reset.count; 549 550 return 0; 551 } 552 553 /* Get data for the copy timestamp query results job submission. */ 554 static int 555 v3d_get_cpu_copy_query_results_params(struct drm_file *file_priv, 556 struct drm_v3d_extension __user *ext, 557 struct v3d_cpu_job *job) 558 { 559 u32 __user *offsets, *syncs; 560 struct drm_v3d_copy_timestamp_query copy; 561 int i; 562 563 if (!job) { 564 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 565 return -EINVAL; 566 } 567 568 if (job->job_type) { 569 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 570 return -EINVAL; 571 } 572 573 if (copy_from_user(©, ext, sizeof(copy))) 574 return -EFAULT; 575 576 if (copy.pad) 577 return -EINVAL; 578 579 job->job_type = V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY; 580 581 job->timestamp_query.queries = kvmalloc_array(copy.count, 582 sizeof(struct v3d_timestamp_query), 583 GFP_KERNEL); 584 if (!job->timestamp_query.queries) 585 return -ENOMEM; 586 587 offsets = u64_to_user_ptr(copy.offsets); 588 syncs = u64_to_user_ptr(copy.syncs); 589 590 for (i = 0; i < copy.count; i++) { 591 u32 offset, sync; 592 593 if (copy_from_user(&offset, offsets++, sizeof(offset))) { 594 kvfree(job->timestamp_query.queries); 595 return -EFAULT; 596 } 597 598 job->timestamp_query.queries[i].offset = offset; 599 600 if (copy_from_user(&sync, syncs++, sizeof(sync))) { 601 kvfree(job->timestamp_query.queries); 602 return -EFAULT; 603 } 604 605 job->timestamp_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync); 606 } 607 job->timestamp_query.count = copy.count; 608 609 job->copy.do_64bit = copy.do_64bit; 610 job->copy.do_partial = copy.do_partial; 611 job->copy.availability_bit = copy.availability_bit; 612 job->copy.offset = copy.offset; 613 job->copy.stride = copy.stride; 614 615 return 0; 616 } 617 618 static int 619 v3d_get_cpu_reset_performance_params(struct drm_file *file_priv, 620 struct drm_v3d_extension __user *ext, 621 struct v3d_cpu_job *job) 622 { 623 u32 __user *syncs; 624 u64 __user *kperfmon_ids; 625 struct drm_v3d_reset_performance_query reset; 626 627 if (!job) { 628 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 629 return -EINVAL; 630 } 631 632 if (job->job_type) { 633 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 634 return -EINVAL; 635 } 636 637 if (copy_from_user(&reset, ext, sizeof(reset))) 638 return -EFAULT; 639 640 if (reset.nperfmons > V3D_MAX_PERFMONS) 641 return -EINVAL; 642 643 job->job_type = V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY; 644 645 job->performance_query.queries = kvmalloc_array(reset.count, 646 sizeof(struct v3d_performance_query), 647 GFP_KERNEL); 648 if (!job->performance_query.queries) 649 return -ENOMEM; 650 651 syncs = u64_to_user_ptr(reset.syncs); 652 kperfmon_ids = u64_to_user_ptr(reset.kperfmon_ids); 653 654 for (int i = 0; i < reset.count; i++) { 655 u32 sync; 656 u64 ids; 657 u32 __user *ids_pointer; 658 u32 id; 659 660 if (copy_from_user(&sync, syncs++, sizeof(sync))) { 661 kvfree(job->performance_query.queries); 662 return -EFAULT; 663 } 664 665 job->performance_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync); 666 667 if (copy_from_user(&ids, kperfmon_ids++, sizeof(ids))) { 668 kvfree(job->performance_query.queries); 669 return -EFAULT; 670 } 671 672 ids_pointer = u64_to_user_ptr(ids); 673 674 for (int j = 0; j < reset.nperfmons; j++) { 675 if (copy_from_user(&id, ids_pointer++, sizeof(id))) { 676 kvfree(job->performance_query.queries); 677 return -EFAULT; 678 } 679 680 job->performance_query.queries[i].kperfmon_ids[j] = id; 681 } 682 } 683 job->performance_query.count = reset.count; 684 job->performance_query.nperfmons = reset.nperfmons; 685 686 return 0; 687 } 688 689 static int 690 v3d_get_cpu_copy_performance_query_params(struct drm_file *file_priv, 691 struct drm_v3d_extension __user *ext, 692 struct v3d_cpu_job *job) 693 { 694 u32 __user *syncs; 695 u64 __user *kperfmon_ids; 696 struct drm_v3d_copy_performance_query copy; 697 698 if (!job) { 699 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 700 return -EINVAL; 701 } 702 703 if (job->job_type) { 704 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 705 return -EINVAL; 706 } 707 708 if (copy_from_user(©, ext, sizeof(copy))) 709 return -EFAULT; 710 711 if (copy.pad) 712 return -EINVAL; 713 714 if (copy.nperfmons > V3D_MAX_PERFMONS) 715 return -EINVAL; 716 717 job->job_type = V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY; 718 719 job->performance_query.queries = kvmalloc_array(copy.count, 720 sizeof(struct v3d_performance_query), 721 GFP_KERNEL); 722 if (!job->performance_query.queries) 723 return -ENOMEM; 724 725 syncs = u64_to_user_ptr(copy.syncs); 726 kperfmon_ids = u64_to_user_ptr(copy.kperfmon_ids); 727 728 for (int i = 0; i < copy.count; i++) { 729 u32 sync; 730 u64 ids; 731 u32 __user *ids_pointer; 732 u32 id; 733 734 if (copy_from_user(&sync, syncs++, sizeof(sync))) { 735 kvfree(job->performance_query.queries); 736 return -EFAULT; 737 } 738 739 job->performance_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync); 740 741 if (copy_from_user(&ids, kperfmon_ids++, sizeof(ids))) { 742 kvfree(job->performance_query.queries); 743 return -EFAULT; 744 } 745 746 ids_pointer = u64_to_user_ptr(ids); 747 748 for (int j = 0; j < copy.nperfmons; j++) { 749 if (copy_from_user(&id, ids_pointer++, sizeof(id))) { 750 kvfree(job->performance_query.queries); 751 return -EFAULT; 752 } 753 754 job->performance_query.queries[i].kperfmon_ids[j] = id; 755 } 756 } 757 job->performance_query.count = copy.count; 758 job->performance_query.nperfmons = copy.nperfmons; 759 job->performance_query.ncounters = copy.ncounters; 760 761 job->copy.do_64bit = copy.do_64bit; 762 job->copy.do_partial = copy.do_partial; 763 job->copy.availability_bit = copy.availability_bit; 764 job->copy.offset = copy.offset; 765 job->copy.stride = copy.stride; 766 767 return 0; 768 } 769 770 /* Whenever userspace sets ioctl extensions, v3d_get_extensions parses data 771 * according to the extension id (name). 772 */ 773 static int 774 v3d_get_extensions(struct drm_file *file_priv, 775 u64 ext_handles, 776 struct v3d_submit_ext *se, 777 struct v3d_cpu_job *job) 778 { 779 struct drm_v3d_extension __user *user_ext; 780 int ret; 781 782 user_ext = u64_to_user_ptr(ext_handles); 783 while (user_ext) { 784 struct drm_v3d_extension ext; 785 786 if (copy_from_user(&ext, user_ext, sizeof(ext))) { 787 DRM_DEBUG("Failed to copy submit extension\n"); 788 return -EFAULT; 789 } 790 791 switch (ext.id) { 792 case DRM_V3D_EXT_ID_MULTI_SYNC: 793 ret = v3d_get_multisync_submit_deps(file_priv, user_ext, se); 794 break; 795 case DRM_V3D_EXT_ID_CPU_INDIRECT_CSD: 796 ret = v3d_get_cpu_indirect_csd_params(file_priv, user_ext, job); 797 break; 798 case DRM_V3D_EXT_ID_CPU_TIMESTAMP_QUERY: 799 ret = v3d_get_cpu_timestamp_query_params(file_priv, user_ext, job); 800 break; 801 case DRM_V3D_EXT_ID_CPU_RESET_TIMESTAMP_QUERY: 802 ret = v3d_get_cpu_reset_timestamp_params(file_priv, user_ext, job); 803 break; 804 case DRM_V3D_EXT_ID_CPU_COPY_TIMESTAMP_QUERY: 805 ret = v3d_get_cpu_copy_query_results_params(file_priv, user_ext, job); 806 break; 807 case DRM_V3D_EXT_ID_CPU_RESET_PERFORMANCE_QUERY: 808 ret = v3d_get_cpu_reset_performance_params(file_priv, user_ext, job); 809 break; 810 case DRM_V3D_EXT_ID_CPU_COPY_PERFORMANCE_QUERY: 811 ret = v3d_get_cpu_copy_performance_query_params(file_priv, user_ext, job); 812 break; 813 default: 814 DRM_DEBUG_DRIVER("Unknown extension id: %d\n", ext.id); 815 return -EINVAL; 816 } 817 818 if (ret) 819 return ret; 820 821 user_ext = u64_to_user_ptr(ext.next); 822 } 823 824 return 0; 825 } 826 827 /** 828 * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D. 829 * @dev: DRM device 830 * @data: ioctl argument 831 * @file_priv: DRM file for this fd 832 * 833 * This is the main entrypoint for userspace to submit a 3D frame to 834 * the GPU. Userspace provides the binner command list (if 835 * applicable), and the kernel sets up the render command list to draw 836 * to the framebuffer described in the ioctl, using the command lists 837 * that the 3D engine's binner will produce. 838 */ 839 int 840 v3d_submit_cl_ioctl(struct drm_device *dev, void *data, 841 struct drm_file *file_priv) 842 { 843 struct v3d_dev *v3d = to_v3d_dev(dev); 844 struct v3d_file_priv *v3d_priv = file_priv->driver_priv; 845 struct drm_v3d_submit_cl *args = data; 846 struct v3d_submit_ext se = {0}; 847 struct v3d_bin_job *bin = NULL; 848 struct v3d_render_job *render = NULL; 849 struct v3d_job *clean_job = NULL; 850 struct v3d_job *last_job; 851 struct ww_acquire_ctx acquire_ctx; 852 int ret = 0; 853 854 trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end); 855 856 if (args->pad) 857 return -EINVAL; 858 859 if (args->flags && 860 args->flags & ~(DRM_V3D_SUBMIT_CL_FLUSH_CACHE | 861 DRM_V3D_SUBMIT_EXTENSION)) { 862 DRM_INFO("invalid flags: %d\n", args->flags); 863 return -EINVAL; 864 } 865 866 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) { 867 ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL); 868 if (ret) { 869 DRM_DEBUG("Failed to get extensions.\n"); 870 return ret; 871 } 872 } 873 874 ret = v3d_job_allocate((void *)&render, sizeof(*render)); 875 if (ret) 876 return ret; 877 878 ret = v3d_job_init(v3d, file_priv, &render->base, 879 v3d_render_job_free, args->in_sync_rcl, &se, V3D_RENDER); 880 if (ret) { 881 v3d_job_deallocate((void *)&render); 882 goto fail; 883 } 884 885 render->start = args->rcl_start; 886 render->end = args->rcl_end; 887 INIT_LIST_HEAD(&render->unref_list); 888 889 if (args->bcl_start != args->bcl_end) { 890 ret = v3d_job_allocate((void *)&bin, sizeof(*bin)); 891 if (ret) 892 goto fail; 893 894 ret = v3d_job_init(v3d, file_priv, &bin->base, 895 v3d_job_free, args->in_sync_bcl, &se, V3D_BIN); 896 if (ret) { 897 v3d_job_deallocate((void *)&bin); 898 goto fail; 899 } 900 901 bin->start = args->bcl_start; 902 bin->end = args->bcl_end; 903 bin->qma = args->qma; 904 bin->qms = args->qms; 905 bin->qts = args->qts; 906 bin->render = render; 907 } 908 909 if (args->flags & DRM_V3D_SUBMIT_CL_FLUSH_CACHE) { 910 ret = v3d_job_allocate((void *)&clean_job, sizeof(*clean_job)); 911 if (ret) 912 goto fail; 913 914 ret = v3d_job_init(v3d, file_priv, clean_job, 915 v3d_job_free, 0, NULL, V3D_CACHE_CLEAN); 916 if (ret) { 917 v3d_job_deallocate((void *)&clean_job); 918 goto fail; 919 } 920 921 last_job = clean_job; 922 } else { 923 last_job = &render->base; 924 } 925 926 ret = v3d_lookup_bos(dev, file_priv, last_job, 927 args->bo_handles, args->bo_handle_count); 928 if (ret) 929 goto fail; 930 931 ret = v3d_lock_bo_reservations(last_job, &acquire_ctx); 932 if (ret) 933 goto fail; 934 935 if (args->perfmon_id) { 936 render->base.perfmon = v3d_perfmon_find(v3d_priv, 937 args->perfmon_id); 938 939 if (!render->base.perfmon) { 940 ret = -ENOENT; 941 goto fail_perfmon; 942 } 943 } 944 945 mutex_lock(&v3d->sched_lock); 946 if (bin) { 947 bin->base.perfmon = render->base.perfmon; 948 v3d_perfmon_get(bin->base.perfmon); 949 v3d_push_job(&bin->base); 950 951 ret = drm_sched_job_add_dependency(&render->base.base, 952 dma_fence_get(bin->base.done_fence)); 953 if (ret) 954 goto fail_unreserve; 955 } 956 957 v3d_push_job(&render->base); 958 959 if (clean_job) { 960 struct dma_fence *render_fence = 961 dma_fence_get(render->base.done_fence); 962 ret = drm_sched_job_add_dependency(&clean_job->base, 963 render_fence); 964 if (ret) 965 goto fail_unreserve; 966 clean_job->perfmon = render->base.perfmon; 967 v3d_perfmon_get(clean_job->perfmon); 968 v3d_push_job(clean_job); 969 } 970 971 mutex_unlock(&v3d->sched_lock); 972 973 v3d_attach_fences_and_unlock_reservation(file_priv, 974 last_job, 975 &acquire_ctx, 976 args->out_sync, 977 &se, 978 last_job->done_fence); 979 980 v3d_job_put(&bin->base); 981 v3d_job_put(&render->base); 982 v3d_job_put(clean_job); 983 984 return 0; 985 986 fail_unreserve: 987 mutex_unlock(&v3d->sched_lock); 988 fail_perfmon: 989 drm_gem_unlock_reservations(last_job->bo, 990 last_job->bo_count, &acquire_ctx); 991 fail: 992 v3d_job_cleanup((void *)bin); 993 v3d_job_cleanup((void *)render); 994 v3d_job_cleanup(clean_job); 995 v3d_put_multisync_post_deps(&se); 996 997 return ret; 998 } 999 1000 /** 1001 * v3d_submit_tfu_ioctl() - Submits a TFU (texture formatting) job to the V3D. 1002 * @dev: DRM device 1003 * @data: ioctl argument 1004 * @file_priv: DRM file for this fd 1005 * 1006 * Userspace provides the register setup for the TFU, which we don't 1007 * need to validate since the TFU is behind the MMU. 1008 */ 1009 int 1010 v3d_submit_tfu_ioctl(struct drm_device *dev, void *data, 1011 struct drm_file *file_priv) 1012 { 1013 struct v3d_dev *v3d = to_v3d_dev(dev); 1014 struct drm_v3d_submit_tfu *args = data; 1015 struct v3d_submit_ext se = {0}; 1016 struct v3d_tfu_job *job = NULL; 1017 struct ww_acquire_ctx acquire_ctx; 1018 int ret = 0; 1019 1020 trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia); 1021 1022 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) { 1023 DRM_DEBUG("invalid flags: %d\n", args->flags); 1024 return -EINVAL; 1025 } 1026 1027 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) { 1028 ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL); 1029 if (ret) { 1030 DRM_DEBUG("Failed to get extensions.\n"); 1031 return ret; 1032 } 1033 } 1034 1035 ret = v3d_job_allocate((void *)&job, sizeof(*job)); 1036 if (ret) 1037 return ret; 1038 1039 ret = v3d_job_init(v3d, file_priv, &job->base, 1040 v3d_job_free, args->in_sync, &se, V3D_TFU); 1041 if (ret) { 1042 v3d_job_deallocate((void *)&job); 1043 goto fail; 1044 } 1045 1046 job->base.bo = kcalloc(ARRAY_SIZE(args->bo_handles), 1047 sizeof(*job->base.bo), GFP_KERNEL); 1048 if (!job->base.bo) { 1049 ret = -ENOMEM; 1050 goto fail; 1051 } 1052 1053 job->args = *args; 1054 1055 for (job->base.bo_count = 0; 1056 job->base.bo_count < ARRAY_SIZE(args->bo_handles); 1057 job->base.bo_count++) { 1058 struct drm_gem_object *bo; 1059 1060 if (!args->bo_handles[job->base.bo_count]) 1061 break; 1062 1063 bo = drm_gem_object_lookup(file_priv, args->bo_handles[job->base.bo_count]); 1064 if (!bo) { 1065 DRM_DEBUG("Failed to look up GEM BO %d: %d\n", 1066 job->base.bo_count, 1067 args->bo_handles[job->base.bo_count]); 1068 ret = -ENOENT; 1069 goto fail; 1070 } 1071 job->base.bo[job->base.bo_count] = bo; 1072 } 1073 1074 ret = v3d_lock_bo_reservations(&job->base, &acquire_ctx); 1075 if (ret) 1076 goto fail; 1077 1078 mutex_lock(&v3d->sched_lock); 1079 v3d_push_job(&job->base); 1080 mutex_unlock(&v3d->sched_lock); 1081 1082 v3d_attach_fences_and_unlock_reservation(file_priv, 1083 &job->base, &acquire_ctx, 1084 args->out_sync, 1085 &se, 1086 job->base.done_fence); 1087 1088 v3d_job_put(&job->base); 1089 1090 return 0; 1091 1092 fail: 1093 v3d_job_cleanup((void *)job); 1094 v3d_put_multisync_post_deps(&se); 1095 1096 return ret; 1097 } 1098 1099 /** 1100 * v3d_submit_csd_ioctl() - Submits a CSD (compute shader) job to the V3D. 1101 * @dev: DRM device 1102 * @data: ioctl argument 1103 * @file_priv: DRM file for this fd 1104 * 1105 * Userspace provides the register setup for the CSD, which we don't 1106 * need to validate since the CSD is behind the MMU. 1107 */ 1108 int 1109 v3d_submit_csd_ioctl(struct drm_device *dev, void *data, 1110 struct drm_file *file_priv) 1111 { 1112 struct v3d_dev *v3d = to_v3d_dev(dev); 1113 struct v3d_file_priv *v3d_priv = file_priv->driver_priv; 1114 struct drm_v3d_submit_csd *args = data; 1115 struct v3d_submit_ext se = {0}; 1116 struct v3d_csd_job *job = NULL; 1117 struct v3d_job *clean_job = NULL; 1118 struct ww_acquire_ctx acquire_ctx; 1119 int ret; 1120 1121 trace_v3d_submit_csd_ioctl(&v3d->drm, args->cfg[5], args->cfg[6]); 1122 1123 if (args->pad) 1124 return -EINVAL; 1125 1126 if (!v3d_has_csd(v3d)) { 1127 DRM_DEBUG("Attempting CSD submit on non-CSD hardware\n"); 1128 return -EINVAL; 1129 } 1130 1131 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) { 1132 DRM_INFO("invalid flags: %d\n", args->flags); 1133 return -EINVAL; 1134 } 1135 1136 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) { 1137 ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL); 1138 if (ret) { 1139 DRM_DEBUG("Failed to get extensions.\n"); 1140 return ret; 1141 } 1142 } 1143 1144 ret = v3d_setup_csd_jobs_and_bos(file_priv, v3d, args, 1145 &job, &clean_job, &se, 1146 &acquire_ctx); 1147 if (ret) 1148 goto fail; 1149 1150 if (args->perfmon_id) { 1151 job->base.perfmon = v3d_perfmon_find(v3d_priv, 1152 args->perfmon_id); 1153 if (!job->base.perfmon) { 1154 ret = -ENOENT; 1155 goto fail_perfmon; 1156 } 1157 } 1158 1159 mutex_lock(&v3d->sched_lock); 1160 v3d_push_job(&job->base); 1161 1162 ret = drm_sched_job_add_dependency(&clean_job->base, 1163 dma_fence_get(job->base.done_fence)); 1164 if (ret) 1165 goto fail_unreserve; 1166 1167 v3d_push_job(clean_job); 1168 mutex_unlock(&v3d->sched_lock); 1169 1170 v3d_attach_fences_and_unlock_reservation(file_priv, 1171 clean_job, 1172 &acquire_ctx, 1173 args->out_sync, 1174 &se, 1175 clean_job->done_fence); 1176 1177 v3d_job_put(&job->base); 1178 v3d_job_put(clean_job); 1179 1180 return 0; 1181 1182 fail_unreserve: 1183 mutex_unlock(&v3d->sched_lock); 1184 fail_perfmon: 1185 drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count, 1186 &acquire_ctx); 1187 fail: 1188 v3d_job_cleanup((void *)job); 1189 v3d_job_cleanup(clean_job); 1190 v3d_put_multisync_post_deps(&se); 1191 1192 return ret; 1193 } 1194 1195 static const unsigned int cpu_job_bo_handle_count[] = { 1196 [V3D_CPU_JOB_TYPE_INDIRECT_CSD] = 1, 1197 [V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY] = 1, 1198 [V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY] = 1, 1199 [V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY] = 2, 1200 [V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY] = 0, 1201 [V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY] = 1, 1202 }; 1203 1204 /** 1205 * v3d_submit_cpu_ioctl() - Submits a CPU job to the V3D. 1206 * @dev: DRM device 1207 * @data: ioctl argument 1208 * @file_priv: DRM file for this fd 1209 * 1210 * Userspace specifies the CPU job type and data required to perform its 1211 * operations through the drm_v3d_extension struct. 1212 */ 1213 int 1214 v3d_submit_cpu_ioctl(struct drm_device *dev, void *data, 1215 struct drm_file *file_priv) 1216 { 1217 struct v3d_dev *v3d = to_v3d_dev(dev); 1218 struct drm_v3d_submit_cpu *args = data; 1219 struct v3d_submit_ext se = {0}; 1220 struct v3d_submit_ext *out_se = NULL; 1221 struct v3d_cpu_job *cpu_job = NULL; 1222 struct v3d_csd_job *csd_job = NULL; 1223 struct v3d_job *clean_job = NULL; 1224 struct ww_acquire_ctx acquire_ctx; 1225 int ret; 1226 1227 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) { 1228 DRM_INFO("Invalid flags: %d\n", args->flags); 1229 return -EINVAL; 1230 } 1231 1232 ret = v3d_job_allocate((void *)&cpu_job, sizeof(*cpu_job)); 1233 if (ret) 1234 return ret; 1235 1236 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) { 1237 ret = v3d_get_extensions(file_priv, args->extensions, &se, cpu_job); 1238 if (ret) { 1239 DRM_DEBUG("Failed to get extensions.\n"); 1240 goto fail; 1241 } 1242 } 1243 1244 /* Every CPU job must have a CPU job user extension */ 1245 if (!cpu_job->job_type) { 1246 DRM_DEBUG("CPU job must have a CPU job user extension.\n"); 1247 ret = -EINVAL; 1248 goto fail; 1249 } 1250 1251 if (args->bo_handle_count != cpu_job_bo_handle_count[cpu_job->job_type]) { 1252 DRM_DEBUG("This CPU job was not submitted with the proper number of BOs.\n"); 1253 ret = -EINVAL; 1254 goto fail; 1255 } 1256 1257 trace_v3d_submit_cpu_ioctl(&v3d->drm, cpu_job->job_type); 1258 1259 ret = v3d_job_init(v3d, file_priv, &cpu_job->base, 1260 v3d_job_free, 0, &se, V3D_CPU); 1261 if (ret) { 1262 v3d_job_deallocate((void *)&cpu_job); 1263 goto fail; 1264 } 1265 1266 clean_job = cpu_job->indirect_csd.clean_job; 1267 csd_job = cpu_job->indirect_csd.job; 1268 1269 if (args->bo_handle_count) { 1270 ret = v3d_lookup_bos(dev, file_priv, &cpu_job->base, 1271 args->bo_handles, args->bo_handle_count); 1272 if (ret) 1273 goto fail; 1274 1275 ret = v3d_lock_bo_reservations(&cpu_job->base, &acquire_ctx); 1276 if (ret) 1277 goto fail; 1278 } 1279 1280 mutex_lock(&v3d->sched_lock); 1281 v3d_push_job(&cpu_job->base); 1282 1283 switch (cpu_job->job_type) { 1284 case V3D_CPU_JOB_TYPE_INDIRECT_CSD: 1285 ret = drm_sched_job_add_dependency(&csd_job->base.base, 1286 dma_fence_get(cpu_job->base.done_fence)); 1287 if (ret) 1288 goto fail_unreserve; 1289 1290 v3d_push_job(&csd_job->base); 1291 1292 ret = drm_sched_job_add_dependency(&clean_job->base, 1293 dma_fence_get(csd_job->base.done_fence)); 1294 if (ret) 1295 goto fail_unreserve; 1296 1297 v3d_push_job(clean_job); 1298 1299 break; 1300 default: 1301 break; 1302 } 1303 mutex_unlock(&v3d->sched_lock); 1304 1305 out_se = (cpu_job->job_type == V3D_CPU_JOB_TYPE_INDIRECT_CSD) ? NULL : &se; 1306 1307 v3d_attach_fences_and_unlock_reservation(file_priv, 1308 &cpu_job->base, 1309 &acquire_ctx, 0, 1310 out_se, cpu_job->base.done_fence); 1311 1312 switch (cpu_job->job_type) { 1313 case V3D_CPU_JOB_TYPE_INDIRECT_CSD: 1314 v3d_attach_fences_and_unlock_reservation(file_priv, 1315 clean_job, 1316 &cpu_job->indirect_csd.acquire_ctx, 1317 0, &se, clean_job->done_fence); 1318 break; 1319 default: 1320 break; 1321 } 1322 1323 v3d_job_put(&cpu_job->base); 1324 v3d_job_put(&csd_job->base); 1325 v3d_job_put(clean_job); 1326 1327 return 0; 1328 1329 fail_unreserve: 1330 mutex_unlock(&v3d->sched_lock); 1331 1332 drm_gem_unlock_reservations(cpu_job->base.bo, cpu_job->base.bo_count, 1333 &acquire_ctx); 1334 1335 drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count, 1336 &cpu_job->indirect_csd.acquire_ctx); 1337 1338 fail: 1339 v3d_job_cleanup((void *)cpu_job); 1340 v3d_job_cleanup((void *)csd_job); 1341 v3d_job_cleanup(clean_job); 1342 v3d_put_multisync_post_deps(&se); 1343 kvfree(cpu_job->timestamp_query.queries); 1344 kvfree(cpu_job->performance_query.queries); 1345 1346 return ret; 1347 } 1348