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 unsigned int i; 456 int err; 457 458 if (!job) { 459 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 460 return -EINVAL; 461 } 462 463 if (job->job_type) { 464 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 465 return -EINVAL; 466 } 467 468 if (copy_from_user(×tamp, ext, sizeof(timestamp))) 469 return -EFAULT; 470 471 if (timestamp.pad) 472 return -EINVAL; 473 474 job->job_type = V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY; 475 476 job->timestamp_query.queries = kvmalloc_array(timestamp.count, 477 sizeof(struct v3d_timestamp_query), 478 GFP_KERNEL); 479 if (!job->timestamp_query.queries) 480 return -ENOMEM; 481 482 offsets = u64_to_user_ptr(timestamp.offsets); 483 syncs = u64_to_user_ptr(timestamp.syncs); 484 485 for (i = 0; i < timestamp.count; i++) { 486 u32 offset, sync; 487 488 if (copy_from_user(&offset, offsets++, sizeof(offset))) { 489 err = -EFAULT; 490 goto error; 491 } 492 493 job->timestamp_query.queries[i].offset = offset; 494 495 if (copy_from_user(&sync, syncs++, sizeof(sync))) { 496 err = -EFAULT; 497 goto error; 498 } 499 500 job->timestamp_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync); 501 } 502 job->timestamp_query.count = timestamp.count; 503 504 return 0; 505 506 error: 507 v3d_timestamp_query_info_free(&job->timestamp_query, i); 508 return err; 509 } 510 511 static int 512 v3d_get_cpu_reset_timestamp_params(struct drm_file *file_priv, 513 struct drm_v3d_extension __user *ext, 514 struct v3d_cpu_job *job) 515 { 516 u32 __user *syncs; 517 struct drm_v3d_reset_timestamp_query reset; 518 unsigned int i; 519 int err; 520 521 if (!job) { 522 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 523 return -EINVAL; 524 } 525 526 if (job->job_type) { 527 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 528 return -EINVAL; 529 } 530 531 if (copy_from_user(&reset, ext, sizeof(reset))) 532 return -EFAULT; 533 534 job->job_type = V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY; 535 536 job->timestamp_query.queries = kvmalloc_array(reset.count, 537 sizeof(struct v3d_timestamp_query), 538 GFP_KERNEL); 539 if (!job->timestamp_query.queries) 540 return -ENOMEM; 541 542 syncs = u64_to_user_ptr(reset.syncs); 543 544 for (i = 0; i < reset.count; i++) { 545 u32 sync; 546 547 job->timestamp_query.queries[i].offset = reset.offset + 8 * i; 548 549 if (copy_from_user(&sync, syncs++, sizeof(sync))) { 550 err = -EFAULT; 551 goto error; 552 } 553 554 job->timestamp_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync); 555 } 556 job->timestamp_query.count = reset.count; 557 558 return 0; 559 560 error: 561 v3d_timestamp_query_info_free(&job->timestamp_query, i); 562 return err; 563 } 564 565 /* Get data for the copy timestamp query results job submission. */ 566 static int 567 v3d_get_cpu_copy_query_results_params(struct drm_file *file_priv, 568 struct drm_v3d_extension __user *ext, 569 struct v3d_cpu_job *job) 570 { 571 u32 __user *offsets, *syncs; 572 struct drm_v3d_copy_timestamp_query copy; 573 unsigned int i; 574 int err; 575 576 if (!job) { 577 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 578 return -EINVAL; 579 } 580 581 if (job->job_type) { 582 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 583 return -EINVAL; 584 } 585 586 if (copy_from_user(©, ext, sizeof(copy))) 587 return -EFAULT; 588 589 if (copy.pad) 590 return -EINVAL; 591 592 job->job_type = V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY; 593 594 job->timestamp_query.queries = kvmalloc_array(copy.count, 595 sizeof(struct v3d_timestamp_query), 596 GFP_KERNEL); 597 if (!job->timestamp_query.queries) 598 return -ENOMEM; 599 600 offsets = u64_to_user_ptr(copy.offsets); 601 syncs = u64_to_user_ptr(copy.syncs); 602 603 for (i = 0; i < copy.count; i++) { 604 u32 offset, sync; 605 606 if (copy_from_user(&offset, offsets++, sizeof(offset))) { 607 err = -EFAULT; 608 goto error; 609 } 610 611 job->timestamp_query.queries[i].offset = offset; 612 613 if (copy_from_user(&sync, syncs++, sizeof(sync))) { 614 err = -EFAULT; 615 goto error; 616 } 617 618 job->timestamp_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync); 619 } 620 job->timestamp_query.count = copy.count; 621 622 job->copy.do_64bit = copy.do_64bit; 623 job->copy.do_partial = copy.do_partial; 624 job->copy.availability_bit = copy.availability_bit; 625 job->copy.offset = copy.offset; 626 job->copy.stride = copy.stride; 627 628 return 0; 629 630 error: 631 v3d_timestamp_query_info_free(&job->timestamp_query, i); 632 return err; 633 } 634 635 static int 636 v3d_get_cpu_reset_performance_params(struct drm_file *file_priv, 637 struct drm_v3d_extension __user *ext, 638 struct v3d_cpu_job *job) 639 { 640 u32 __user *syncs; 641 u64 __user *kperfmon_ids; 642 struct drm_v3d_reset_performance_query reset; 643 unsigned int i, j; 644 int err; 645 646 if (!job) { 647 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 648 return -EINVAL; 649 } 650 651 if (job->job_type) { 652 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 653 return -EINVAL; 654 } 655 656 if (copy_from_user(&reset, ext, sizeof(reset))) 657 return -EFAULT; 658 659 if (reset.nperfmons > V3D_MAX_PERFMONS) 660 return -EINVAL; 661 662 job->job_type = V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY; 663 664 job->performance_query.queries = kvmalloc_array(reset.count, 665 sizeof(struct v3d_performance_query), 666 GFP_KERNEL); 667 if (!job->performance_query.queries) 668 return -ENOMEM; 669 670 syncs = u64_to_user_ptr(reset.syncs); 671 kperfmon_ids = u64_to_user_ptr(reset.kperfmon_ids); 672 673 for (i = 0; i < reset.count; i++) { 674 u32 sync; 675 u64 ids; 676 u32 __user *ids_pointer; 677 u32 id; 678 679 if (copy_from_user(&sync, syncs++, sizeof(sync))) { 680 err = -EFAULT; 681 goto error; 682 } 683 684 if (copy_from_user(&ids, kperfmon_ids++, sizeof(ids))) { 685 err = -EFAULT; 686 goto error; 687 } 688 689 ids_pointer = u64_to_user_ptr(ids); 690 691 for (j = 0; j < reset.nperfmons; j++) { 692 if (copy_from_user(&id, ids_pointer++, sizeof(id))) { 693 err = -EFAULT; 694 goto error; 695 } 696 697 job->performance_query.queries[i].kperfmon_ids[j] = id; 698 } 699 700 job->performance_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync); 701 } 702 job->performance_query.count = reset.count; 703 job->performance_query.nperfmons = reset.nperfmons; 704 705 return 0; 706 707 error: 708 v3d_performance_query_info_free(&job->performance_query, i); 709 return err; 710 } 711 712 static int 713 v3d_get_cpu_copy_performance_query_params(struct drm_file *file_priv, 714 struct drm_v3d_extension __user *ext, 715 struct v3d_cpu_job *job) 716 { 717 u32 __user *syncs; 718 u64 __user *kperfmon_ids; 719 struct drm_v3d_copy_performance_query copy; 720 unsigned int i, j; 721 int err; 722 723 if (!job) { 724 DRM_DEBUG("CPU job extension was attached to a GPU job.\n"); 725 return -EINVAL; 726 } 727 728 if (job->job_type) { 729 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n"); 730 return -EINVAL; 731 } 732 733 if (copy_from_user(©, ext, sizeof(copy))) 734 return -EFAULT; 735 736 if (copy.pad) 737 return -EINVAL; 738 739 if (copy.nperfmons > V3D_MAX_PERFMONS) 740 return -EINVAL; 741 742 job->job_type = V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY; 743 744 job->performance_query.queries = kvmalloc_array(copy.count, 745 sizeof(struct v3d_performance_query), 746 GFP_KERNEL); 747 if (!job->performance_query.queries) 748 return -ENOMEM; 749 750 syncs = u64_to_user_ptr(copy.syncs); 751 kperfmon_ids = u64_to_user_ptr(copy.kperfmon_ids); 752 753 for (i = 0; i < copy.count; i++) { 754 u32 sync; 755 u64 ids; 756 u32 __user *ids_pointer; 757 u32 id; 758 759 if (copy_from_user(&sync, syncs++, sizeof(sync))) { 760 err = -EFAULT; 761 goto error; 762 } 763 764 if (copy_from_user(&ids, kperfmon_ids++, sizeof(ids))) { 765 err = -EFAULT; 766 goto error; 767 } 768 769 ids_pointer = u64_to_user_ptr(ids); 770 771 for (j = 0; j < copy.nperfmons; j++) { 772 if (copy_from_user(&id, ids_pointer++, sizeof(id))) { 773 err = -EFAULT; 774 goto error; 775 } 776 777 job->performance_query.queries[i].kperfmon_ids[j] = id; 778 } 779 780 job->performance_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync); 781 } 782 job->performance_query.count = copy.count; 783 job->performance_query.nperfmons = copy.nperfmons; 784 job->performance_query.ncounters = copy.ncounters; 785 786 job->copy.do_64bit = copy.do_64bit; 787 job->copy.do_partial = copy.do_partial; 788 job->copy.availability_bit = copy.availability_bit; 789 job->copy.offset = copy.offset; 790 job->copy.stride = copy.stride; 791 792 return 0; 793 794 error: 795 v3d_performance_query_info_free(&job->performance_query, i); 796 return err; 797 } 798 799 /* Whenever userspace sets ioctl extensions, v3d_get_extensions parses data 800 * according to the extension id (name). 801 */ 802 static int 803 v3d_get_extensions(struct drm_file *file_priv, 804 u64 ext_handles, 805 struct v3d_submit_ext *se, 806 struct v3d_cpu_job *job) 807 { 808 struct drm_v3d_extension __user *user_ext; 809 int ret; 810 811 user_ext = u64_to_user_ptr(ext_handles); 812 while (user_ext) { 813 struct drm_v3d_extension ext; 814 815 if (copy_from_user(&ext, user_ext, sizeof(ext))) { 816 DRM_DEBUG("Failed to copy submit extension\n"); 817 return -EFAULT; 818 } 819 820 switch (ext.id) { 821 case DRM_V3D_EXT_ID_MULTI_SYNC: 822 ret = v3d_get_multisync_submit_deps(file_priv, user_ext, se); 823 break; 824 case DRM_V3D_EXT_ID_CPU_INDIRECT_CSD: 825 ret = v3d_get_cpu_indirect_csd_params(file_priv, user_ext, job); 826 break; 827 case DRM_V3D_EXT_ID_CPU_TIMESTAMP_QUERY: 828 ret = v3d_get_cpu_timestamp_query_params(file_priv, user_ext, job); 829 break; 830 case DRM_V3D_EXT_ID_CPU_RESET_TIMESTAMP_QUERY: 831 ret = v3d_get_cpu_reset_timestamp_params(file_priv, user_ext, job); 832 break; 833 case DRM_V3D_EXT_ID_CPU_COPY_TIMESTAMP_QUERY: 834 ret = v3d_get_cpu_copy_query_results_params(file_priv, user_ext, job); 835 break; 836 case DRM_V3D_EXT_ID_CPU_RESET_PERFORMANCE_QUERY: 837 ret = v3d_get_cpu_reset_performance_params(file_priv, user_ext, job); 838 break; 839 case DRM_V3D_EXT_ID_CPU_COPY_PERFORMANCE_QUERY: 840 ret = v3d_get_cpu_copy_performance_query_params(file_priv, user_ext, job); 841 break; 842 default: 843 DRM_DEBUG_DRIVER("Unknown extension id: %d\n", ext.id); 844 return -EINVAL; 845 } 846 847 if (ret) 848 return ret; 849 850 user_ext = u64_to_user_ptr(ext.next); 851 } 852 853 return 0; 854 } 855 856 /** 857 * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D. 858 * @dev: DRM device 859 * @data: ioctl argument 860 * @file_priv: DRM file for this fd 861 * 862 * This is the main entrypoint for userspace to submit a 3D frame to 863 * the GPU. Userspace provides the binner command list (if 864 * applicable), and the kernel sets up the render command list to draw 865 * to the framebuffer described in the ioctl, using the command lists 866 * that the 3D engine's binner will produce. 867 */ 868 int 869 v3d_submit_cl_ioctl(struct drm_device *dev, void *data, 870 struct drm_file *file_priv) 871 { 872 struct v3d_dev *v3d = to_v3d_dev(dev); 873 struct v3d_file_priv *v3d_priv = file_priv->driver_priv; 874 struct drm_v3d_submit_cl *args = data; 875 struct v3d_submit_ext se = {0}; 876 struct v3d_bin_job *bin = NULL; 877 struct v3d_render_job *render = NULL; 878 struct v3d_job *clean_job = NULL; 879 struct v3d_job *last_job; 880 struct ww_acquire_ctx acquire_ctx; 881 int ret = 0; 882 883 trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end); 884 885 if (args->pad) 886 return -EINVAL; 887 888 if (args->flags && 889 args->flags & ~(DRM_V3D_SUBMIT_CL_FLUSH_CACHE | 890 DRM_V3D_SUBMIT_EXTENSION)) { 891 DRM_INFO("invalid flags: %d\n", args->flags); 892 return -EINVAL; 893 } 894 895 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) { 896 ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL); 897 if (ret) { 898 DRM_DEBUG("Failed to get extensions.\n"); 899 return ret; 900 } 901 } 902 903 ret = v3d_job_allocate((void *)&render, sizeof(*render)); 904 if (ret) 905 return ret; 906 907 ret = v3d_job_init(v3d, file_priv, &render->base, 908 v3d_render_job_free, args->in_sync_rcl, &se, V3D_RENDER); 909 if (ret) { 910 v3d_job_deallocate((void *)&render); 911 goto fail; 912 } 913 914 render->start = args->rcl_start; 915 render->end = args->rcl_end; 916 INIT_LIST_HEAD(&render->unref_list); 917 918 if (args->bcl_start != args->bcl_end) { 919 ret = v3d_job_allocate((void *)&bin, sizeof(*bin)); 920 if (ret) 921 goto fail; 922 923 ret = v3d_job_init(v3d, file_priv, &bin->base, 924 v3d_job_free, args->in_sync_bcl, &se, V3D_BIN); 925 if (ret) { 926 v3d_job_deallocate((void *)&bin); 927 goto fail; 928 } 929 930 bin->start = args->bcl_start; 931 bin->end = args->bcl_end; 932 bin->qma = args->qma; 933 bin->qms = args->qms; 934 bin->qts = args->qts; 935 bin->render = render; 936 } 937 938 if (args->flags & DRM_V3D_SUBMIT_CL_FLUSH_CACHE) { 939 ret = v3d_job_allocate((void *)&clean_job, sizeof(*clean_job)); 940 if (ret) 941 goto fail; 942 943 ret = v3d_job_init(v3d, file_priv, clean_job, 944 v3d_job_free, 0, NULL, V3D_CACHE_CLEAN); 945 if (ret) { 946 v3d_job_deallocate((void *)&clean_job); 947 goto fail; 948 } 949 950 last_job = clean_job; 951 } else { 952 last_job = &render->base; 953 } 954 955 ret = v3d_lookup_bos(dev, file_priv, last_job, 956 args->bo_handles, args->bo_handle_count); 957 if (ret) 958 goto fail; 959 960 ret = v3d_lock_bo_reservations(last_job, &acquire_ctx); 961 if (ret) 962 goto fail; 963 964 if (args->perfmon_id) { 965 render->base.perfmon = v3d_perfmon_find(v3d_priv, 966 args->perfmon_id); 967 968 if (!render->base.perfmon) { 969 ret = -ENOENT; 970 goto fail_perfmon; 971 } 972 } 973 974 mutex_lock(&v3d->sched_lock); 975 if (bin) { 976 bin->base.perfmon = render->base.perfmon; 977 v3d_perfmon_get(bin->base.perfmon); 978 v3d_push_job(&bin->base); 979 980 ret = drm_sched_job_add_dependency(&render->base.base, 981 dma_fence_get(bin->base.done_fence)); 982 if (ret) 983 goto fail_unreserve; 984 } 985 986 v3d_push_job(&render->base); 987 988 if (clean_job) { 989 struct dma_fence *render_fence = 990 dma_fence_get(render->base.done_fence); 991 ret = drm_sched_job_add_dependency(&clean_job->base, 992 render_fence); 993 if (ret) 994 goto fail_unreserve; 995 clean_job->perfmon = render->base.perfmon; 996 v3d_perfmon_get(clean_job->perfmon); 997 v3d_push_job(clean_job); 998 } 999 1000 mutex_unlock(&v3d->sched_lock); 1001 1002 v3d_attach_fences_and_unlock_reservation(file_priv, 1003 last_job, 1004 &acquire_ctx, 1005 args->out_sync, 1006 &se, 1007 last_job->done_fence); 1008 1009 v3d_job_put(&bin->base); 1010 v3d_job_put(&render->base); 1011 v3d_job_put(clean_job); 1012 1013 return 0; 1014 1015 fail_unreserve: 1016 mutex_unlock(&v3d->sched_lock); 1017 fail_perfmon: 1018 drm_gem_unlock_reservations(last_job->bo, 1019 last_job->bo_count, &acquire_ctx); 1020 fail: 1021 v3d_job_cleanup((void *)bin); 1022 v3d_job_cleanup((void *)render); 1023 v3d_job_cleanup(clean_job); 1024 v3d_put_multisync_post_deps(&se); 1025 1026 return ret; 1027 } 1028 1029 /** 1030 * v3d_submit_tfu_ioctl() - Submits a TFU (texture formatting) job to the V3D. 1031 * @dev: DRM device 1032 * @data: ioctl argument 1033 * @file_priv: DRM file for this fd 1034 * 1035 * Userspace provides the register setup for the TFU, which we don't 1036 * need to validate since the TFU is behind the MMU. 1037 */ 1038 int 1039 v3d_submit_tfu_ioctl(struct drm_device *dev, void *data, 1040 struct drm_file *file_priv) 1041 { 1042 struct v3d_dev *v3d = to_v3d_dev(dev); 1043 struct drm_v3d_submit_tfu *args = data; 1044 struct v3d_submit_ext se = {0}; 1045 struct v3d_tfu_job *job = NULL; 1046 struct ww_acquire_ctx acquire_ctx; 1047 int ret = 0; 1048 1049 trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia); 1050 1051 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) { 1052 DRM_DEBUG("invalid flags: %d\n", args->flags); 1053 return -EINVAL; 1054 } 1055 1056 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) { 1057 ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL); 1058 if (ret) { 1059 DRM_DEBUG("Failed to get extensions.\n"); 1060 return ret; 1061 } 1062 } 1063 1064 ret = v3d_job_allocate((void *)&job, sizeof(*job)); 1065 if (ret) 1066 return ret; 1067 1068 ret = v3d_job_init(v3d, file_priv, &job->base, 1069 v3d_job_free, args->in_sync, &se, V3D_TFU); 1070 if (ret) { 1071 v3d_job_deallocate((void *)&job); 1072 goto fail; 1073 } 1074 1075 job->base.bo = kcalloc(ARRAY_SIZE(args->bo_handles), 1076 sizeof(*job->base.bo), GFP_KERNEL); 1077 if (!job->base.bo) { 1078 ret = -ENOMEM; 1079 goto fail; 1080 } 1081 1082 job->args = *args; 1083 1084 for (job->base.bo_count = 0; 1085 job->base.bo_count < ARRAY_SIZE(args->bo_handles); 1086 job->base.bo_count++) { 1087 struct drm_gem_object *bo; 1088 1089 if (!args->bo_handles[job->base.bo_count]) 1090 break; 1091 1092 bo = drm_gem_object_lookup(file_priv, args->bo_handles[job->base.bo_count]); 1093 if (!bo) { 1094 DRM_DEBUG("Failed to look up GEM BO %d: %d\n", 1095 job->base.bo_count, 1096 args->bo_handles[job->base.bo_count]); 1097 ret = -ENOENT; 1098 goto fail; 1099 } 1100 job->base.bo[job->base.bo_count] = bo; 1101 } 1102 1103 ret = v3d_lock_bo_reservations(&job->base, &acquire_ctx); 1104 if (ret) 1105 goto fail; 1106 1107 mutex_lock(&v3d->sched_lock); 1108 v3d_push_job(&job->base); 1109 mutex_unlock(&v3d->sched_lock); 1110 1111 v3d_attach_fences_and_unlock_reservation(file_priv, 1112 &job->base, &acquire_ctx, 1113 args->out_sync, 1114 &se, 1115 job->base.done_fence); 1116 1117 v3d_job_put(&job->base); 1118 1119 return 0; 1120 1121 fail: 1122 v3d_job_cleanup((void *)job); 1123 v3d_put_multisync_post_deps(&se); 1124 1125 return ret; 1126 } 1127 1128 /** 1129 * v3d_submit_csd_ioctl() - Submits a CSD (compute shader) job to the V3D. 1130 * @dev: DRM device 1131 * @data: ioctl argument 1132 * @file_priv: DRM file for this fd 1133 * 1134 * Userspace provides the register setup for the CSD, which we don't 1135 * need to validate since the CSD is behind the MMU. 1136 */ 1137 int 1138 v3d_submit_csd_ioctl(struct drm_device *dev, void *data, 1139 struct drm_file *file_priv) 1140 { 1141 struct v3d_dev *v3d = to_v3d_dev(dev); 1142 struct v3d_file_priv *v3d_priv = file_priv->driver_priv; 1143 struct drm_v3d_submit_csd *args = data; 1144 struct v3d_submit_ext se = {0}; 1145 struct v3d_csd_job *job = NULL; 1146 struct v3d_job *clean_job = NULL; 1147 struct ww_acquire_ctx acquire_ctx; 1148 int ret; 1149 1150 trace_v3d_submit_csd_ioctl(&v3d->drm, args->cfg[5], args->cfg[6]); 1151 1152 if (args->pad) 1153 return -EINVAL; 1154 1155 if (!v3d_has_csd(v3d)) { 1156 DRM_DEBUG("Attempting CSD submit on non-CSD hardware\n"); 1157 return -EINVAL; 1158 } 1159 1160 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) { 1161 DRM_INFO("invalid flags: %d\n", args->flags); 1162 return -EINVAL; 1163 } 1164 1165 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) { 1166 ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL); 1167 if (ret) { 1168 DRM_DEBUG("Failed to get extensions.\n"); 1169 return ret; 1170 } 1171 } 1172 1173 ret = v3d_setup_csd_jobs_and_bos(file_priv, v3d, args, 1174 &job, &clean_job, &se, 1175 &acquire_ctx); 1176 if (ret) 1177 goto fail; 1178 1179 if (args->perfmon_id) { 1180 job->base.perfmon = v3d_perfmon_find(v3d_priv, 1181 args->perfmon_id); 1182 if (!job->base.perfmon) { 1183 ret = -ENOENT; 1184 goto fail_perfmon; 1185 } 1186 } 1187 1188 mutex_lock(&v3d->sched_lock); 1189 v3d_push_job(&job->base); 1190 1191 ret = drm_sched_job_add_dependency(&clean_job->base, 1192 dma_fence_get(job->base.done_fence)); 1193 if (ret) 1194 goto fail_unreserve; 1195 1196 v3d_push_job(clean_job); 1197 mutex_unlock(&v3d->sched_lock); 1198 1199 v3d_attach_fences_and_unlock_reservation(file_priv, 1200 clean_job, 1201 &acquire_ctx, 1202 args->out_sync, 1203 &se, 1204 clean_job->done_fence); 1205 1206 v3d_job_put(&job->base); 1207 v3d_job_put(clean_job); 1208 1209 return 0; 1210 1211 fail_unreserve: 1212 mutex_unlock(&v3d->sched_lock); 1213 fail_perfmon: 1214 drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count, 1215 &acquire_ctx); 1216 fail: 1217 v3d_job_cleanup((void *)job); 1218 v3d_job_cleanup(clean_job); 1219 v3d_put_multisync_post_deps(&se); 1220 1221 return ret; 1222 } 1223 1224 static const unsigned int cpu_job_bo_handle_count[] = { 1225 [V3D_CPU_JOB_TYPE_INDIRECT_CSD] = 1, 1226 [V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY] = 1, 1227 [V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY] = 1, 1228 [V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY] = 2, 1229 [V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY] = 0, 1230 [V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY] = 1, 1231 }; 1232 1233 /** 1234 * v3d_submit_cpu_ioctl() - Submits a CPU job to the V3D. 1235 * @dev: DRM device 1236 * @data: ioctl argument 1237 * @file_priv: DRM file for this fd 1238 * 1239 * Userspace specifies the CPU job type and data required to perform its 1240 * operations through the drm_v3d_extension struct. 1241 */ 1242 int 1243 v3d_submit_cpu_ioctl(struct drm_device *dev, void *data, 1244 struct drm_file *file_priv) 1245 { 1246 struct v3d_dev *v3d = to_v3d_dev(dev); 1247 struct drm_v3d_submit_cpu *args = data; 1248 struct v3d_submit_ext se = {0}; 1249 struct v3d_submit_ext *out_se = NULL; 1250 struct v3d_cpu_job *cpu_job = NULL; 1251 struct v3d_csd_job *csd_job = NULL; 1252 struct v3d_job *clean_job = NULL; 1253 struct ww_acquire_ctx acquire_ctx; 1254 int ret; 1255 1256 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) { 1257 DRM_INFO("Invalid flags: %d\n", args->flags); 1258 return -EINVAL; 1259 } 1260 1261 ret = v3d_job_allocate((void *)&cpu_job, sizeof(*cpu_job)); 1262 if (ret) 1263 return ret; 1264 1265 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) { 1266 ret = v3d_get_extensions(file_priv, args->extensions, &se, cpu_job); 1267 if (ret) { 1268 DRM_DEBUG("Failed to get extensions.\n"); 1269 goto fail; 1270 } 1271 } 1272 1273 /* Every CPU job must have a CPU job user extension */ 1274 if (!cpu_job->job_type) { 1275 DRM_DEBUG("CPU job must have a CPU job user extension.\n"); 1276 ret = -EINVAL; 1277 goto fail; 1278 } 1279 1280 if (args->bo_handle_count != cpu_job_bo_handle_count[cpu_job->job_type]) { 1281 DRM_DEBUG("This CPU job was not submitted with the proper number of BOs.\n"); 1282 ret = -EINVAL; 1283 goto fail; 1284 } 1285 1286 trace_v3d_submit_cpu_ioctl(&v3d->drm, cpu_job->job_type); 1287 1288 ret = v3d_job_init(v3d, file_priv, &cpu_job->base, 1289 v3d_job_free, 0, &se, V3D_CPU); 1290 if (ret) { 1291 v3d_job_deallocate((void *)&cpu_job); 1292 goto fail; 1293 } 1294 1295 clean_job = cpu_job->indirect_csd.clean_job; 1296 csd_job = cpu_job->indirect_csd.job; 1297 1298 if (args->bo_handle_count) { 1299 ret = v3d_lookup_bos(dev, file_priv, &cpu_job->base, 1300 args->bo_handles, args->bo_handle_count); 1301 if (ret) 1302 goto fail; 1303 1304 ret = v3d_lock_bo_reservations(&cpu_job->base, &acquire_ctx); 1305 if (ret) 1306 goto fail; 1307 } 1308 1309 mutex_lock(&v3d->sched_lock); 1310 v3d_push_job(&cpu_job->base); 1311 1312 switch (cpu_job->job_type) { 1313 case V3D_CPU_JOB_TYPE_INDIRECT_CSD: 1314 ret = drm_sched_job_add_dependency(&csd_job->base.base, 1315 dma_fence_get(cpu_job->base.done_fence)); 1316 if (ret) 1317 goto fail_unreserve; 1318 1319 v3d_push_job(&csd_job->base); 1320 1321 ret = drm_sched_job_add_dependency(&clean_job->base, 1322 dma_fence_get(csd_job->base.done_fence)); 1323 if (ret) 1324 goto fail_unreserve; 1325 1326 v3d_push_job(clean_job); 1327 1328 break; 1329 default: 1330 break; 1331 } 1332 mutex_unlock(&v3d->sched_lock); 1333 1334 out_se = (cpu_job->job_type == V3D_CPU_JOB_TYPE_INDIRECT_CSD) ? NULL : &se; 1335 1336 v3d_attach_fences_and_unlock_reservation(file_priv, 1337 &cpu_job->base, 1338 &acquire_ctx, 0, 1339 out_se, cpu_job->base.done_fence); 1340 1341 switch (cpu_job->job_type) { 1342 case V3D_CPU_JOB_TYPE_INDIRECT_CSD: 1343 v3d_attach_fences_and_unlock_reservation(file_priv, 1344 clean_job, 1345 &cpu_job->indirect_csd.acquire_ctx, 1346 0, &se, clean_job->done_fence); 1347 break; 1348 default: 1349 break; 1350 } 1351 1352 v3d_job_put(&cpu_job->base); 1353 v3d_job_put(&csd_job->base); 1354 v3d_job_put(clean_job); 1355 1356 return 0; 1357 1358 fail_unreserve: 1359 mutex_unlock(&v3d->sched_lock); 1360 1361 drm_gem_unlock_reservations(cpu_job->base.bo, cpu_job->base.bo_count, 1362 &acquire_ctx); 1363 1364 drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count, 1365 &cpu_job->indirect_csd.acquire_ctx); 1366 1367 fail: 1368 v3d_job_cleanup((void *)cpu_job); 1369 v3d_job_cleanup((void *)csd_job); 1370 v3d_job_cleanup(clean_job); 1371 v3d_put_multisync_post_deps(&se); 1372 kvfree(cpu_job->timestamp_query.queries); 1373 kvfree(cpu_job->performance_query.queries); 1374 1375 return ret; 1376 } 1377