1 /* 2 * Copyright 2008 Jrme Glisse 3 * Copyright 2010 Marek Olk <maraeo (at) gmail.com> 4 * Copyright 2015 Advanced Micro Devices, Inc. 5 * All Rights Reserved. 6 * 7 * Permission is hereby granted, free of charge, to any person obtaining 8 * a copy of this software and associated documentation files (the 9 * "Software"), to deal in the Software without restriction, including 10 * without limitation the rights to use, copy, modify, merge, publish, 11 * distribute, sub license, and/or sell copies of the Software, and to 12 * permit persons to whom the Software is furnished to do so, subject to 13 * the following conditions: 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 16 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES 17 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 18 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS 19 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 22 * USE OR OTHER DEALINGS IN THE SOFTWARE. 23 * 24 * The above copyright notice and this permission notice (including the 25 * next paragraph) shall be included in all copies or substantial portions 26 * of the Software. 27 */ 28 29 #include "amdgpu_cs.h" 30 #include "util/os_time.h" 31 #include <inttypes.h> 32 #include <stdio.h> 33 34 #include "amd/common/sid.h" 35 36 DEBUG_GET_ONCE_BOOL_OPTION(noop, "RADEON_NOOP", false) 37 38 /* FENCES */ 39 40 static struct pipe_fence_handle * 41 amdgpu_fence_create(struct amdgpu_ctx *ctx, unsigned ip_type, 42 unsigned ip_instance, unsigned ring) 43 { 44 struct amdgpu_fence *fence = CALLOC_STRUCT(amdgpu_fence); 45 46 fence->reference.count = 1; 47 fence->ws = ctx->ws; 48 fence->ctx = ctx; 49 fence->fence.context = ctx->ctx; 50 fence->fence.ip_type = ip_type; 51 fence->fence.ip_instance = ip_instance; 52 fence->fence.ring = ring; 53 util_queue_fence_init(&fence->submitted); 54 util_queue_fence_reset(&fence->submitted); 55 p_atomic_inc(&ctx->refcount); 56 return (struct pipe_fence_handle *)fence; 57 } 58 59 static struct pipe_fence_handle * 60 amdgpu_fence_import_sync_file(struct radeon_winsys *rws, int fd) 61 { 62 struct amdgpu_winsys *ws = amdgpu_winsys(rws); 63 struct amdgpu_fence *fence = CALLOC_STRUCT(amdgpu_fence); 64 65 if (!fence) 66 return NULL; 67 68 pipe_reference_init(&fence->reference, 1); 69 fence->ws = ws; 70 /* fence->ctx == NULL means that the fence is syncobj-based. */ 71 72 /* Convert sync_file into syncobj. */ 73 int r = amdgpu_cs_create_syncobj(ws->dev, &fence->syncobj); 74 if (r) { 75 FREE(fence); 76 return NULL; 77 } 78 79 r = amdgpu_cs_syncobj_import_sync_file(ws->dev, fence->syncobj, fd); 80 if (r) { 81 amdgpu_cs_destroy_syncobj(ws->dev, fence->syncobj); 82 FREE(fence); 83 return NULL; 84 } 85 86 util_queue_fence_init(&fence->submitted); 87 88 return (struct pipe_fence_handle*)fence; 89 } 90 91 static int amdgpu_fence_export_sync_file(struct radeon_winsys *rws, 92 struct pipe_fence_handle *pfence) 93 { 94 struct amdgpu_winsys *ws = amdgpu_winsys(rws); 95 struct amdgpu_fence *fence = (struct amdgpu_fence*)pfence; 96 97 if (amdgpu_fence_is_syncobj(fence)) { 98 int fd, r; 99 100 /* Convert syncobj into sync_file. */ 101 r = amdgpu_cs_syncobj_export_sync_file(ws->dev, fence->syncobj, &fd); 102 return r ? -1 : fd; 103 } 104 105 util_queue_fence_wait(&fence->submitted); 106 107 /* Convert the amdgpu fence into a fence FD. */ 108 int fd; 109 if (amdgpu_cs_fence_to_handle(ws->dev, &fence->fence, 110 AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD, 111 (uint32_t*)&fd)) 112 return -1; 113 114 return fd; 115 } 116 117 static int amdgpu_export_signalled_sync_file(struct radeon_winsys *rws) 118 { 119 struct amdgpu_winsys *ws = amdgpu_winsys(rws); 120 uint32_t syncobj; 121 int fd = -1; 122 123 int r = amdgpu_cs_create_syncobj2(ws->dev, DRM_SYNCOBJ_CREATE_SIGNALED, 124 &syncobj); 125 if (r) { 126 return -1; 127 } 128 129 r = amdgpu_cs_syncobj_export_sync_file(ws->dev, syncobj, &fd); 130 if (r) { 131 fd = -1; 132 } 133 134 amdgpu_cs_destroy_syncobj(ws->dev, syncobj); 135 return fd; 136 } 137 138 static void amdgpu_fence_submitted(struct pipe_fence_handle *fence, 139 uint64_t seq_no, 140 uint64_t *user_fence_cpu_address) 141 { 142 struct amdgpu_fence *rfence = (struct amdgpu_fence*)fence; 143 144 rfence->fence.fence = seq_no; 145 rfence->user_fence_cpu_address = user_fence_cpu_address; 146 util_queue_fence_signal(&rfence->submitted); 147 } 148 149 static void amdgpu_fence_signalled(struct pipe_fence_handle *fence) 150 { 151 struct amdgpu_fence *rfence = (struct amdgpu_fence*)fence; 152 153 rfence->signalled = true; 154 util_queue_fence_signal(&rfence->submitted); 155 } 156 157 bool amdgpu_fence_wait(struct pipe_fence_handle *fence, uint64_t timeout, 158 bool absolute) 159 { 160 struct amdgpu_fence *rfence = (struct amdgpu_fence*)fence; 161 uint32_t expired; 162 int64_t abs_timeout; 163 uint64_t *user_fence_cpu; 164 int r; 165 166 if (rfence->signalled) 167 return true; 168 169 /* Handle syncobjs. */ 170 if (amdgpu_fence_is_syncobj(rfence)) { 171 /* Absolute timeouts are only be used by BO fences, which aren't 172 * backed by syncobjs. 173 */ 174 assert(!absolute); 175 176 if (amdgpu_cs_syncobj_wait(rfence->ws->dev, &rfence->syncobj, 1, 177 timeout, 0, NULL)) 178 return false; 179 180 rfence->signalled = true; 181 return true; 182 } 183 184 if (absolute) 185 abs_timeout = timeout; 186 else 187 abs_timeout = os_time_get_absolute_timeout(timeout); 188 189 /* The fence might not have a number assigned if its IB is being 190 * submitted in the other thread right now. Wait until the submission 191 * is done. */ 192 if (!util_queue_fence_wait_timeout(&rfence->submitted, abs_timeout)) 193 return false; 194 195 user_fence_cpu = rfence->user_fence_cpu_address; 196 if (user_fence_cpu) { 197 if (*user_fence_cpu >= rfence->fence.fence) { 198 rfence->signalled = true; 199 return true; 200 } 201 202 /* No timeout, just query: no need for the ioctl. */ 203 if (!absolute && !timeout) 204 return false; 205 } 206 207 /* Now use the libdrm query. */ 208 r = amdgpu_cs_query_fence_status(&rfence->fence, 209 abs_timeout, 210 AMDGPU_QUERY_FENCE_TIMEOUT_IS_ABSOLUTE, 211 &expired); 212 if (r) { 213 fprintf(stderr, "amdgpu: amdgpu_cs_query_fence_status failed.\n"); 214 return false; 215 } 216 217 if (expired) { 218 /* This variable can only transition from false to true, so it doesn't 219 * matter if threads race for it. */ 220 rfence->signalled = true; 221 return true; 222 } 223 return false; 224 } 225 226 static bool amdgpu_fence_wait_rel_timeout(struct radeon_winsys *rws, 227 struct pipe_fence_handle *fence, 228 uint64_t timeout) 229 { 230 return amdgpu_fence_wait(fence, timeout, false); 231 } 232 233 static struct pipe_fence_handle * 234 amdgpu_cs_get_next_fence(struct radeon_winsys_cs *rcs) 235 { 236 struct amdgpu_cs *cs = amdgpu_cs(rcs); 237 struct pipe_fence_handle *fence = NULL; 238 239 if (debug_get_option_noop()) 240 return NULL; 241 242 if (cs->next_fence) { 243 amdgpu_fence_reference(&fence, cs->next_fence); 244 return fence; 245 } 246 247 fence = amdgpu_fence_create(cs->ctx, 248 cs->csc->ib[IB_MAIN].ip_type, 249 cs->csc->ib[IB_MAIN].ip_instance, 250 cs->csc->ib[IB_MAIN].ring); 251 if (!fence) 252 return NULL; 253 254 amdgpu_fence_reference(&cs->next_fence, fence); 255 return fence; 256 } 257 258 /* CONTEXTS */ 259 260 static struct radeon_winsys_ctx *amdgpu_ctx_create(struct radeon_winsys *ws) 261 { 262 struct amdgpu_ctx *ctx = CALLOC_STRUCT(amdgpu_ctx); 263 int r; 264 struct amdgpu_bo_alloc_request alloc_buffer = {}; 265 amdgpu_bo_handle buf_handle; 266 267 if (!ctx) 268 return NULL; 269 270 ctx->ws = amdgpu_winsys(ws); 271 ctx->refcount = 1; 272 ctx->initial_num_total_rejected_cs = ctx->ws->num_total_rejected_cs; 273 274 r = amdgpu_cs_ctx_create(ctx->ws->dev, &ctx->ctx); 275 if (r) { 276 fprintf(stderr, "amdgpu: amdgpu_cs_ctx_create failed. (%i)\n", r); 277 goto error_create; 278 } 279 280 alloc_buffer.alloc_size = ctx->ws->info.gart_page_size; 281 alloc_buffer.phys_alignment = ctx->ws->info.gart_page_size; 282 alloc_buffer.preferred_heap = AMDGPU_GEM_DOMAIN_GTT; 283 284 r = amdgpu_bo_alloc(ctx->ws->dev, &alloc_buffer, &buf_handle); 285 if (r) { 286 fprintf(stderr, "amdgpu: amdgpu_bo_alloc failed. (%i)\n", r); 287 goto error_user_fence_alloc; 288 } 289 290 r = amdgpu_bo_cpu_map(buf_handle, (void**)&ctx->user_fence_cpu_address_base); 291 if (r) { 292 fprintf(stderr, "amdgpu: amdgpu_bo_cpu_map failed. (%i)\n", r); 293 goto error_user_fence_map; 294 } 295 296 memset(ctx->user_fence_cpu_address_base, 0, alloc_buffer.alloc_size); 297 ctx->user_fence_bo = buf_handle; 298 299 return (struct radeon_winsys_ctx*)ctx; 300 301 error_user_fence_map: 302 amdgpu_bo_free(buf_handle); 303 error_user_fence_alloc: 304 amdgpu_cs_ctx_free(ctx->ctx); 305 error_create: 306 FREE(ctx); 307 return NULL; 308 } 309 310 static void amdgpu_ctx_destroy(struct radeon_winsys_ctx *rwctx) 311 { 312 amdgpu_ctx_unref((struct amdgpu_ctx*)rwctx); 313 } 314 315 static enum pipe_reset_status 316 amdgpu_ctx_query_reset_status(struct radeon_winsys_ctx *rwctx) 317 { 318 struct amdgpu_ctx *ctx = (struct amdgpu_ctx*)rwctx; 319 uint32_t result, hangs; 320 int r; 321 322 /* Return a failure due to a rejected command submission. */ 323 if (ctx->ws->num_total_rejected_cs > ctx->initial_num_total_rejected_cs) { 324 return ctx->num_rejected_cs ? PIPE_GUILTY_CONTEXT_RESET : 325 PIPE_INNOCENT_CONTEXT_RESET; 326 } 327 328 /* Return a failure due to a GPU hang. */ 329 r = amdgpu_cs_query_reset_state(ctx->ctx, &result, &hangs); 330 if (r) { 331 fprintf(stderr, "amdgpu: amdgpu_cs_query_reset_state failed. (%i)\n", r); 332 return PIPE_NO_RESET; 333 } 334 335 switch (result) { 336 case AMDGPU_CTX_GUILTY_RESET: 337 return PIPE_GUILTY_CONTEXT_RESET; 338 case AMDGPU_CTX_INNOCENT_RESET: 339 return PIPE_INNOCENT_CONTEXT_RESET; 340 case AMDGPU_CTX_UNKNOWN_RESET: 341 return PIPE_UNKNOWN_CONTEXT_RESET; 342 case AMDGPU_CTX_NO_RESET: 343 default: 344 return PIPE_NO_RESET; 345 } 346 } 347 348 /* COMMAND SUBMISSION */ 349 350 static bool amdgpu_cs_has_user_fence(struct amdgpu_cs_context *cs) 351 { 352 return cs->ib[IB_MAIN].ip_type != AMDGPU_HW_IP_UVD && 353 cs->ib[IB_MAIN].ip_type != AMDGPU_HW_IP_VCE && 354 cs->ib[IB_MAIN].ip_type != AMDGPU_HW_IP_VCN_DEC && 355 cs->ib[IB_MAIN].ip_type != AMDGPU_HW_IP_VCN_ENC; 356 } 357 358 static bool amdgpu_cs_has_chaining(struct amdgpu_cs *cs) 359 { 360 return cs->ctx->ws->info.chip_class >= CIK && 361 cs->ring_type == RING_GFX; 362 } 363 364 static unsigned amdgpu_cs_epilog_dws(enum ring_type ring_type) 365 { 366 if (ring_type == RING_GFX) 367 return 4; /* for chaining */ 368 369 return 0; 370 } 371 372 int amdgpu_lookup_buffer(struct amdgpu_cs_context *cs, struct amdgpu_winsys_bo *bo) 373 { 374 unsigned hash = bo->unique_id & (ARRAY_SIZE(cs->buffer_indices_hashlist)-1); 375 int i = cs->buffer_indices_hashlist[hash]; 376 struct amdgpu_cs_buffer *buffers; 377 int num_buffers; 378 379 if (bo->bo) { 380 buffers = cs->real_buffers; 381 num_buffers = cs->num_real_buffers; 382 } else if (!bo->sparse) { 383 buffers = cs->slab_buffers; 384 num_buffers = cs->num_slab_buffers; 385 } else { 386 buffers = cs->sparse_buffers; 387 num_buffers = cs->num_sparse_buffers; 388 } 389 390 /* not found or found */ 391 if (i < 0 || (i < num_buffers && buffers[i].bo == bo)) 392 return i; 393 394 /* Hash collision, look for the BO in the list of buffers linearly. */ 395 for (i = num_buffers - 1; i >= 0; i--) { 396 if (buffers[i].bo == bo) { 397 /* Put this buffer in the hash list. 398 * This will prevent additional hash collisions if there are 399 * several consecutive lookup_buffer calls for the same buffer. 400 * 401 * Example: Assuming buffers A,B,C collide in the hash list, 402 * the following sequence of buffers: 403 * AAAAAAAAAAABBBBBBBBBBBBBBCCCCCCCC 404 * will collide here: ^ and here: ^, 405 * meaning that we should get very few collisions in the end. */ 406 cs->buffer_indices_hashlist[hash] = i; 407 return i; 408 } 409 } 410 return -1; 411 } 412 413 static int 414 amdgpu_do_add_real_buffer(struct amdgpu_cs_context *cs, struct amdgpu_winsys_bo *bo) 415 { 416 struct amdgpu_cs_buffer *buffer; 417 int idx; 418 419 /* New buffer, check if the backing array is large enough. */ 420 if (cs->num_real_buffers >= cs->max_real_buffers) { 421 unsigned new_max = 422 MAX2(cs->max_real_buffers + 16, (unsigned)(cs->max_real_buffers * 1.3)); 423 struct amdgpu_cs_buffer *new_buffers; 424 425 new_buffers = MALLOC(new_max * sizeof(*new_buffers)); 426 427 if (!new_buffers) { 428 fprintf(stderr, "amdgpu_do_add_buffer: allocation failed\n"); 429 FREE(new_buffers); 430 return -1; 431 } 432 433 memcpy(new_buffers, cs->real_buffers, cs->num_real_buffers * sizeof(*new_buffers)); 434 435 FREE(cs->real_buffers); 436 437 cs->max_real_buffers = new_max; 438 cs->real_buffers = new_buffers; 439 } 440 441 idx = cs->num_real_buffers; 442 buffer = &cs->real_buffers[idx]; 443 444 memset(buffer, 0, sizeof(*buffer)); 445 amdgpu_winsys_bo_reference(&buffer->bo, bo); 446 p_atomic_inc(&bo->num_cs_references); 447 cs->num_real_buffers++; 448 449 return idx; 450 } 451 452 static int 453 amdgpu_lookup_or_add_real_buffer(struct amdgpu_cs *acs, struct amdgpu_winsys_bo *bo) 454 { 455 struct amdgpu_cs_context *cs = acs->csc; 456 unsigned hash; 457 int idx = amdgpu_lookup_buffer(cs, bo); 458 459 if (idx >= 0) 460 return idx; 461 462 idx = amdgpu_do_add_real_buffer(cs, bo); 463 464 hash = bo->unique_id & (ARRAY_SIZE(cs->buffer_indices_hashlist)-1); 465 cs->buffer_indices_hashlist[hash] = idx; 466 467 if (bo->initial_domain & RADEON_DOMAIN_VRAM) 468 acs->main.base.used_vram += bo->base.size; 469 else if (bo->initial_domain & RADEON_DOMAIN_GTT) 470 acs->main.base.used_gart += bo->base.size; 471 472 return idx; 473 } 474 475 static int amdgpu_lookup_or_add_slab_buffer(struct amdgpu_cs *acs, 476 struct amdgpu_winsys_bo *bo) 477 { 478 struct amdgpu_cs_context *cs = acs->csc; 479 struct amdgpu_cs_buffer *buffer; 480 unsigned hash; 481 int idx = amdgpu_lookup_buffer(cs, bo); 482 int real_idx; 483 484 if (idx >= 0) 485 return idx; 486 487 real_idx = amdgpu_lookup_or_add_real_buffer(acs, bo->u.slab.real); 488 if (real_idx < 0) 489 return -1; 490 491 /* New buffer, check if the backing array is large enough. */ 492 if (cs->num_slab_buffers >= cs->max_slab_buffers) { 493 unsigned new_max = 494 MAX2(cs->max_slab_buffers + 16, (unsigned)(cs->max_slab_buffers * 1.3)); 495 struct amdgpu_cs_buffer *new_buffers; 496 497 new_buffers = REALLOC(cs->slab_buffers, 498 cs->max_slab_buffers * sizeof(*new_buffers), 499 new_max * sizeof(*new_buffers)); 500 if (!new_buffers) { 501 fprintf(stderr, "amdgpu_lookup_or_add_slab_buffer: allocation failed\n"); 502 return -1; 503 } 504 505 cs->max_slab_buffers = new_max; 506 cs->slab_buffers = new_buffers; 507 } 508 509 idx = cs->num_slab_buffers; 510 buffer = &cs->slab_buffers[idx]; 511 512 memset(buffer, 0, sizeof(*buffer)); 513 amdgpu_winsys_bo_reference(&buffer->bo, bo); 514 buffer->u.slab.real_idx = real_idx; 515 p_atomic_inc(&bo->num_cs_references); 516 cs->num_slab_buffers++; 517 518 hash = bo->unique_id & (ARRAY_SIZE(cs->buffer_indices_hashlist)-1); 519 cs->buffer_indices_hashlist[hash] = idx; 520 521 return idx; 522 } 523 524 static int amdgpu_lookup_or_add_sparse_buffer(struct amdgpu_cs *acs, 525 struct amdgpu_winsys_bo *bo) 526 { 527 struct amdgpu_cs_context *cs = acs->csc; 528 struct amdgpu_cs_buffer *buffer; 529 unsigned hash; 530 int idx = amdgpu_lookup_buffer(cs, bo); 531 532 if (idx >= 0) 533 return idx; 534 535 /* New buffer, check if the backing array is large enough. */ 536 if (cs->num_sparse_buffers >= cs->max_sparse_buffers) { 537 unsigned new_max = 538 MAX2(cs->max_sparse_buffers + 16, (unsigned)(cs->max_sparse_buffers * 1.3)); 539 struct amdgpu_cs_buffer *new_buffers; 540 541 new_buffers = REALLOC(cs->sparse_buffers, 542 cs->max_sparse_buffers * sizeof(*new_buffers), 543 new_max * sizeof(*new_buffers)); 544 if (!new_buffers) { 545 fprintf(stderr, "amdgpu_lookup_or_add_sparse_buffer: allocation failed\n"); 546 return -1; 547 } 548 549 cs->max_sparse_buffers = new_max; 550 cs->sparse_buffers = new_buffers; 551 } 552 553 idx = cs->num_sparse_buffers; 554 buffer = &cs->sparse_buffers[idx]; 555 556 memset(buffer, 0, sizeof(*buffer)); 557 amdgpu_winsys_bo_reference(&buffer->bo, bo); 558 p_atomic_inc(&bo->num_cs_references); 559 cs->num_sparse_buffers++; 560 561 hash = bo->unique_id & (ARRAY_SIZE(cs->buffer_indices_hashlist)-1); 562 cs->buffer_indices_hashlist[hash] = idx; 563 564 /* We delay adding the backing buffers until we really have to. However, 565 * we cannot delay accounting for memory use. 566 */ 567 simple_mtx_lock(&bo->u.sparse.commit_lock); 568 569 list_for_each_entry(struct amdgpu_sparse_backing, backing, &bo->u.sparse.backing, list) { 570 if (bo->initial_domain & RADEON_DOMAIN_VRAM) 571 acs->main.base.used_vram += backing->bo->base.size; 572 else if (bo->initial_domain & RADEON_DOMAIN_GTT) 573 acs->main.base.used_gart += backing->bo->base.size; 574 } 575 576 simple_mtx_unlock(&bo->u.sparse.commit_lock); 577 578 return idx; 579 } 580 581 static unsigned amdgpu_cs_add_buffer(struct radeon_winsys_cs *rcs, 582 struct pb_buffer *buf, 583 enum radeon_bo_usage usage, 584 enum radeon_bo_domain domains, 585 enum radeon_bo_priority priority) 586 { 587 /* Don't use the "domains" parameter. Amdgpu doesn't support changing 588 * the buffer placement during command submission. 589 */ 590 struct amdgpu_cs *acs = amdgpu_cs(rcs); 591 struct amdgpu_cs_context *cs = acs->csc; 592 struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf; 593 struct amdgpu_cs_buffer *buffer; 594 int index; 595 596 /* Fast exit for no-op calls. 597 * This is very effective with suballocators and linear uploaders that 598 * are outside of the winsys. 599 */ 600 if (bo == cs->last_added_bo && 601 (usage & cs->last_added_bo_usage) == usage && 602 (1ull << priority) & cs->last_added_bo_priority_usage) 603 return cs->last_added_bo_index; 604 605 if (!bo->sparse) { 606 if (!bo->bo) { 607 index = amdgpu_lookup_or_add_slab_buffer(acs, bo); 608 if (index < 0) 609 return 0; 610 611 buffer = &cs->slab_buffers[index]; 612 buffer->usage |= usage; 613 614 usage &= ~RADEON_USAGE_SYNCHRONIZED; 615 index = buffer->u.slab.real_idx; 616 } else { 617 index = amdgpu_lookup_or_add_real_buffer(acs, bo); 618 if (index < 0) 619 return 0; 620 } 621 622 buffer = &cs->real_buffers[index]; 623 } else { 624 index = amdgpu_lookup_or_add_sparse_buffer(acs, bo); 625 if (index < 0) 626 return 0; 627 628 buffer = &cs->sparse_buffers[index]; 629 } 630 631 buffer->u.real.priority_usage |= 1ull << priority; 632 buffer->usage |= usage; 633 634 cs->last_added_bo = bo; 635 cs->last_added_bo_index = index; 636 cs->last_added_bo_usage = buffer->usage; 637 cs->last_added_bo_priority_usage = buffer->u.real.priority_usage; 638 return index; 639 } 640 641 static bool amdgpu_ib_new_buffer(struct amdgpu_winsys *ws, struct amdgpu_ib *ib, 642 enum ring_type ring_type) 643 { 644 struct pb_buffer *pb; 645 uint8_t *mapped; 646 unsigned buffer_size; 647 648 /* Always create a buffer that is at least as large as the maximum seen IB 649 * size, aligned to a power of two (and multiplied by 4 to reduce internal 650 * fragmentation if chaining is not available). Limit to 512k dwords, which 651 * is the largest power of two that fits into the size field of the 652 * INDIRECT_BUFFER packet. 653 */ 654 if (amdgpu_cs_has_chaining(amdgpu_cs_from_ib(ib))) 655 buffer_size = 4 *util_next_power_of_two(ib->max_ib_size); 656 else 657 buffer_size = 4 *util_next_power_of_two(4 * ib->max_ib_size); 658 659 buffer_size = MIN2(buffer_size, 4 * 512 * 1024); 660 661 switch (ib->ib_type) { 662 case IB_MAIN: 663 buffer_size = MAX2(buffer_size, 8 * 1024 * 4); 664 break; 665 default: 666 unreachable("unhandled IB type"); 667 } 668 669 pb = ws->base.buffer_create(&ws->base, buffer_size, 670 ws->info.gart_page_size, 671 RADEON_DOMAIN_GTT, 672 RADEON_FLAG_NO_INTERPROCESS_SHARING | 673 (ring_type == RING_GFX || 674 ring_type == RING_COMPUTE || 675 ring_type == RING_DMA ? 676 RADEON_FLAG_READ_ONLY | RADEON_FLAG_GTT_WC : 0)); 677 if (!pb) 678 return false; 679 680 mapped = ws->base.buffer_map(pb, NULL, PIPE_TRANSFER_WRITE); 681 if (!mapped) { 682 pb_reference(&pb, NULL); 683 return false; 684 } 685 686 pb_reference(&ib->big_ib_buffer, pb); 687 pb_reference(&pb, NULL); 688 689 ib->ib_mapped = mapped; 690 ib->used_ib_space = 0; 691 692 return true; 693 } 694 695 static unsigned amdgpu_ib_max_submit_dwords(enum ib_type ib_type) 696 { 697 switch (ib_type) { 698 case IB_MAIN: 699 /* Smaller submits means the GPU gets busy sooner and there is less 700 * waiting for buffers and fences. Proof: 701 * http://www.phoronix.com/scan.php?page=article&item=mesa-111-si&num=1 702 */ 703 return 20 * 1024; 704 default: 705 unreachable("bad ib_type"); 706 } 707 } 708 709 static bool amdgpu_get_new_ib(struct radeon_winsys *ws, struct amdgpu_cs *cs, 710 enum ib_type ib_type) 711 { 712 struct amdgpu_winsys *aws = (struct amdgpu_winsys*)ws; 713 /* Small IBs are better than big IBs, because the GPU goes idle quicker 714 * and there is less waiting for buffers and fences. Proof: 715 * http://www.phoronix.com/scan.php?page=article&item=mesa-111-si&num=1 716 */ 717 struct amdgpu_ib *ib = NULL; 718 struct drm_amdgpu_cs_chunk_ib *info = &cs->csc->ib[ib_type]; 719 unsigned ib_size = 0; 720 721 switch (ib_type) { 722 case IB_MAIN: 723 ib = &cs->main; 724 ib_size = 4 * 1024 * 4; 725 break; 726 default: 727 unreachable("unhandled IB type"); 728 } 729 730 if (!amdgpu_cs_has_chaining(cs)) { 731 ib_size = MAX2(ib_size, 732 4 * MIN2(util_next_power_of_two(ib->max_ib_size), 733 amdgpu_ib_max_submit_dwords(ib_type))); 734 } 735 736 ib->max_ib_size = ib->max_ib_size - ib->max_ib_size / 32; 737 738 ib->base.prev_dw = 0; 739 ib->base.num_prev = 0; 740 ib->base.current.cdw = 0; 741 ib->base.current.buf = NULL; 742 743 /* Allocate a new buffer for IBs if the current buffer is all used. */ 744 if (!ib->big_ib_buffer || 745 ib->used_ib_space + ib_size > ib->big_ib_buffer->size) { 746 if (!amdgpu_ib_new_buffer(aws, ib, cs->ring_type)) 747 return false; 748 } 749 750 info->va_start = amdgpu_winsys_bo(ib->big_ib_buffer)->va + ib->used_ib_space; 751 info->ib_bytes = 0; 752 /* ib_bytes is in dwords and the conversion to bytes will be done before 753 * the CS ioctl. */ 754 ib->ptr_ib_size = &info->ib_bytes; 755 ib->ptr_ib_size_inside_ib = false; 756 757 amdgpu_cs_add_buffer(&cs->main.base, ib->big_ib_buffer, 758 RADEON_USAGE_READ, 0, RADEON_PRIO_IB1); 759 760 ib->base.current.buf = (uint32_t*)(ib->ib_mapped + ib->used_ib_space); 761 762 ib_size = ib->big_ib_buffer->size - ib->used_ib_space; 763 ib->base.current.max_dw = ib_size / 4 - amdgpu_cs_epilog_dws(cs->ring_type); 764 return true; 765 } 766 767 static void amdgpu_set_ib_size(struct amdgpu_ib *ib) 768 { 769 if (ib->ptr_ib_size_inside_ib) { 770 *ib->ptr_ib_size = ib->base.current.cdw | 771 S_3F2_CHAIN(1) | S_3F2_VALID(1); 772 } else { 773 *ib->ptr_ib_size = ib->base.current.cdw; 774 } 775 } 776 777 static void amdgpu_ib_finalize(struct amdgpu_winsys *ws, struct amdgpu_ib *ib) 778 { 779 amdgpu_set_ib_size(ib); 780 ib->used_ib_space += ib->base.current.cdw * 4; 781 ib->used_ib_space = align(ib->used_ib_space, ws->info.ib_start_alignment); 782 ib->max_ib_size = MAX2(ib->max_ib_size, ib->base.prev_dw + ib->base.current.cdw); 783 } 784 785 static bool amdgpu_init_cs_context(struct amdgpu_cs_context *cs, 786 enum ring_type ring_type) 787 { 788 switch (ring_type) { 789 case RING_DMA: 790 cs->ib[IB_MAIN].ip_type = AMDGPU_HW_IP_DMA; 791 break; 792 793 case RING_UVD: 794 cs->ib[IB_MAIN].ip_type = AMDGPU_HW_IP_UVD; 795 break; 796 797 case RING_VCE: 798 cs->ib[IB_MAIN].ip_type = AMDGPU_HW_IP_VCE; 799 break; 800 801 case RING_COMPUTE: 802 cs->ib[IB_MAIN].ip_type = AMDGPU_HW_IP_COMPUTE; 803 break; 804 805 case RING_VCN_DEC: 806 cs->ib[IB_MAIN].ip_type = AMDGPU_HW_IP_VCN_DEC; 807 break; 808 809 case RING_VCN_ENC: 810 cs->ib[IB_MAIN].ip_type = AMDGPU_HW_IP_VCN_ENC; 811 break; 812 813 default: 814 case RING_GFX: 815 cs->ib[IB_MAIN].ip_type = AMDGPU_HW_IP_GFX; 816 break; 817 } 818 819 memset(cs->buffer_indices_hashlist, -1, sizeof(cs->buffer_indices_hashlist)); 820 cs->last_added_bo = NULL; 821 return true; 822 } 823 824 static void amdgpu_cs_context_cleanup(struct amdgpu_cs_context *cs) 825 { 826 unsigned i; 827 828 for (i = 0; i < cs->num_real_buffers; i++) { 829 p_atomic_dec(&cs->real_buffers[i].bo->num_cs_references); 830 amdgpu_winsys_bo_reference(&cs->real_buffers[i].bo, NULL); 831 } 832 for (i = 0; i < cs->num_slab_buffers; i++) { 833 p_atomic_dec(&cs->slab_buffers[i].bo->num_cs_references); 834 amdgpu_winsys_bo_reference(&cs->slab_buffers[i].bo, NULL); 835 } 836 for (i = 0; i < cs->num_sparse_buffers; i++) { 837 p_atomic_dec(&cs->sparse_buffers[i].bo->num_cs_references); 838 amdgpu_winsys_bo_reference(&cs->sparse_buffers[i].bo, NULL); 839 } 840 for (i = 0; i < cs->num_fence_dependencies; i++) 841 amdgpu_fence_reference(&cs->fence_dependencies[i], NULL); 842 843 cs->num_real_buffers = 0; 844 cs->num_slab_buffers = 0; 845 cs->num_sparse_buffers = 0; 846 cs->num_fence_dependencies = 0; 847 amdgpu_fence_reference(&cs->fence, NULL); 848 849 memset(cs->buffer_indices_hashlist, -1, sizeof(cs->buffer_indices_hashlist)); 850 cs->last_added_bo = NULL; 851 } 852 853 static void amdgpu_destroy_cs_context(struct amdgpu_cs_context *cs) 854 { 855 amdgpu_cs_context_cleanup(cs); 856 FREE(cs->flags); 857 FREE(cs->real_buffers); 858 FREE(cs->handles); 859 FREE(cs->slab_buffers); 860 FREE(cs->sparse_buffers); 861 FREE(cs->fence_dependencies); 862 } 863 864 865 static struct radeon_winsys_cs * 866 amdgpu_cs_create(struct radeon_winsys_ctx *rwctx, 867 enum ring_type ring_type, 868 void (*flush)(void *ctx, unsigned flags, 869 struct pipe_fence_handle **fence), 870 void *flush_ctx) 871 { 872 struct amdgpu_ctx *ctx = (struct amdgpu_ctx*)rwctx; 873 struct amdgpu_cs *cs; 874 875 cs = CALLOC_STRUCT(amdgpu_cs); 876 if (!cs) { 877 return NULL; 878 } 879 880 util_queue_fence_init(&cs->flush_completed); 881 882 cs->ctx = ctx; 883 cs->flush_cs = flush; 884 cs->flush_data = flush_ctx; 885 cs->ring_type = ring_type; 886 887 struct amdgpu_cs_fence_info fence_info; 888 fence_info.handle = cs->ctx->user_fence_bo; 889 fence_info.offset = cs->ring_type; 890 amdgpu_cs_chunk_fence_info_to_data(&fence_info, (void*)&cs->fence_chunk); 891 892 cs->main.ib_type = IB_MAIN; 893 894 if (!amdgpu_init_cs_context(&cs->csc1, ring_type)) { 895 FREE(cs); 896 return NULL; 897 } 898 899 if (!amdgpu_init_cs_context(&cs->csc2, ring_type)) { 900 amdgpu_destroy_cs_context(&cs->csc1); 901 FREE(cs); 902 return NULL; 903 } 904 905 /* Set the first submission context as current. */ 906 cs->csc = &cs->csc1; 907 cs->cst = &cs->csc2; 908 909 if (!amdgpu_get_new_ib(&ctx->ws->base, cs, IB_MAIN)) { 910 amdgpu_destroy_cs_context(&cs->csc2); 911 amdgpu_destroy_cs_context(&cs->csc1); 912 FREE(cs); 913 return NULL; 914 } 915 916 p_atomic_inc(&ctx->ws->num_cs); 917 return &cs->main.base; 918 } 919 920 static bool amdgpu_cs_validate(struct radeon_winsys_cs *rcs) 921 { 922 return true; 923 } 924 925 static bool amdgpu_cs_check_space(struct radeon_winsys_cs *rcs, unsigned dw) 926 { 927 struct amdgpu_ib *ib = amdgpu_ib(rcs); 928 struct amdgpu_cs *cs = amdgpu_cs_from_ib(ib); 929 unsigned requested_size = rcs->prev_dw + rcs->current.cdw + dw; 930 uint64_t va; 931 uint32_t *new_ptr_ib_size; 932 933 assert(rcs->current.cdw <= rcs->current.max_dw); 934 935 if (requested_size > amdgpu_ib_max_submit_dwords(ib->ib_type)) 936 return false; 937 938 ib->max_ib_size = MAX2(ib->max_ib_size, requested_size); 939 940 if (rcs->current.max_dw - rcs->current.cdw >= dw) 941 return true; 942 943 if (!amdgpu_cs_has_chaining(cs)) 944 return false; 945 946 /* Allocate a new chunk */ 947 if (rcs->num_prev >= rcs->max_prev) { 948 unsigned new_max_prev = MAX2(1, 2 * rcs->max_prev); 949 struct radeon_winsys_cs_chunk *new_prev; 950 951 new_prev = REALLOC(rcs->prev, 952 sizeof(*new_prev) * rcs->max_prev, 953 sizeof(*new_prev) * new_max_prev); 954 if (!new_prev) 955 return false; 956 957 rcs->prev = new_prev; 958 rcs->max_prev = new_max_prev; 959 } 960 961 if (!amdgpu_ib_new_buffer(cs->ctx->ws, ib, cs->ring_type)) 962 return false; 963 964 assert(ib->used_ib_space == 0); 965 va = amdgpu_winsys_bo(ib->big_ib_buffer)->va; 966 967 /* This space was originally reserved. */ 968 rcs->current.max_dw += 4; 969 assert(ib->used_ib_space + 4 * rcs->current.max_dw <= ib->big_ib_buffer->size); 970 971 /* Pad with NOPs and add INDIRECT_BUFFER packet */ 972 while ((rcs->current.cdw & 7) != 4) 973 radeon_emit(rcs, 0xffff1000); /* type3 nop packet */ 974 975 radeon_emit(rcs, PKT3(ib->ib_type == IB_MAIN ? PKT3_INDIRECT_BUFFER_CIK 976 : PKT3_INDIRECT_BUFFER_CONST, 2, 0)); 977 radeon_emit(rcs, va); 978 radeon_emit(rcs, va >> 32); 979 new_ptr_ib_size = &rcs->current.buf[rcs->current.cdw++]; 980 981 assert((rcs->current.cdw & 7) == 0); 982 assert(rcs->current.cdw <= rcs->current.max_dw); 983 984 amdgpu_set_ib_size(ib); 985 ib->ptr_ib_size = new_ptr_ib_size; 986 ib->ptr_ib_size_inside_ib = true; 987 988 /* Hook up the new chunk */ 989 rcs->prev[rcs->num_prev].buf = rcs->current.buf; 990 rcs->prev[rcs->num_prev].cdw = rcs->current.cdw; 991 rcs->prev[rcs->num_prev].max_dw = rcs->current.cdw; /* no modifications */ 992 rcs->num_prev++; 993 994 ib->base.prev_dw += ib->base.current.cdw; 995 ib->base.current.cdw = 0; 996 997 ib->base.current.buf = (uint32_t*)(ib->ib_mapped + ib->used_ib_space); 998 ib->base.current.max_dw = ib->big_ib_buffer->size / 4 - amdgpu_cs_epilog_dws(cs->ring_type); 999 1000 amdgpu_cs_add_buffer(&cs->main.base, ib->big_ib_buffer, 1001 RADEON_USAGE_READ, 0, RADEON_PRIO_IB1); 1002 1003 return true; 1004 } 1005 1006 static unsigned amdgpu_cs_get_buffer_list(struct radeon_winsys_cs *rcs, 1007 struct radeon_bo_list_item *list) 1008 { 1009 struct amdgpu_cs_context *cs = amdgpu_cs(rcs)->csc; 1010 int i; 1011 1012 if (list) { 1013 for (i = 0; i < cs->num_real_buffers; i++) { 1014 list[i].bo_size = cs->real_buffers[i].bo->base.size; 1015 list[i].vm_address = cs->real_buffers[i].bo->va; 1016 list[i].priority_usage = cs->real_buffers[i].u.real.priority_usage; 1017 } 1018 } 1019 return cs->num_real_buffers; 1020 } 1021 1022 static unsigned add_fence_dependency_entry(struct amdgpu_cs_context *cs) 1023 { 1024 unsigned idx = cs->num_fence_dependencies++; 1025 1026 if (idx >= cs->max_fence_dependencies) { 1027 unsigned size; 1028 const unsigned increment = 8; 1029 1030 cs->max_fence_dependencies = idx + increment; 1031 size = cs->max_fence_dependencies * sizeof(cs->fence_dependencies[0]); 1032 cs->fence_dependencies = realloc(cs->fence_dependencies, size); 1033 /* Clear the newly-allocated elements. */ 1034 memset(cs->fence_dependencies + idx, 0, 1035 increment * sizeof(cs->fence_dependencies[0])); 1036 } 1037 return idx; 1038 } 1039 1040 static bool is_noop_fence_dependency(struct amdgpu_cs *acs, 1041 struct amdgpu_fence *fence) 1042 { 1043 struct amdgpu_cs_context *cs = acs->csc; 1044 1045 if (!amdgpu_fence_is_syncobj(fence) && 1046 fence->ctx == acs->ctx && 1047 fence->fence.ip_type == cs->ib[IB_MAIN].ip_type && 1048 fence->fence.ip_instance == cs->ib[IB_MAIN].ip_instance && 1049 fence->fence.ring == cs->ib[IB_MAIN].ring) 1050 return true; 1051 1052 return amdgpu_fence_wait((void *)fence, 0, false); 1053 } 1054 1055 static void amdgpu_cs_add_fence_dependency(struct radeon_winsys_cs *rws, 1056 struct pipe_fence_handle *pfence) 1057 { 1058 struct amdgpu_cs *acs = amdgpu_cs(rws); 1059 struct amdgpu_cs_context *cs = acs->csc; 1060 struct amdgpu_fence *fence = (struct amdgpu_fence*)pfence; 1061 1062 util_queue_fence_wait(&fence->submitted); 1063 1064 if (is_noop_fence_dependency(acs, fence)) 1065 return; 1066 1067 unsigned idx = add_fence_dependency_entry(cs); 1068 amdgpu_fence_reference(&cs->fence_dependencies[idx], 1069 (struct pipe_fence_handle*)fence); 1070 } 1071 1072 static void amdgpu_add_bo_fence_dependencies(struct amdgpu_cs *acs, 1073 struct amdgpu_cs_buffer *buffer) 1074 { 1075 struct amdgpu_cs_context *cs = acs->csc; 1076 struct amdgpu_winsys_bo *bo = buffer->bo; 1077 unsigned new_num_fences = 0; 1078 1079 for (unsigned j = 0; j < bo->num_fences; ++j) { 1080 struct amdgpu_fence *bo_fence = (void *)bo->fences[j]; 1081 1082 if (is_noop_fence_dependency(acs, bo_fence)) 1083 continue; 1084 1085 amdgpu_fence_reference(&bo->fences[new_num_fences], bo->fences[j]); 1086 new_num_fences++; 1087 1088 if (!(buffer->usage & RADEON_USAGE_SYNCHRONIZED)) 1089 continue; 1090 1091 unsigned idx = add_fence_dependency_entry(cs); 1092 amdgpu_fence_reference(&cs->fence_dependencies[idx], 1093 (struct pipe_fence_handle*)bo_fence); 1094 } 1095 1096 for (unsigned j = new_num_fences; j < bo->num_fences; ++j) 1097 amdgpu_fence_reference(&bo->fences[j], NULL); 1098 1099 bo->num_fences = new_num_fences; 1100 } 1101 1102 /* Add the given list of fences to the buffer's fence list. 1103 * 1104 * Must be called with the winsys bo_fence_lock held. 1105 */ 1106 void amdgpu_add_fences(struct amdgpu_winsys_bo *bo, 1107 unsigned num_fences, 1108 struct pipe_fence_handle **fences) 1109 { 1110 if (bo->num_fences + num_fences > bo->max_fences) { 1111 unsigned new_max_fences = MAX2(bo->num_fences + num_fences, bo->max_fences * 2); 1112 struct pipe_fence_handle **new_fences = 1113 REALLOC(bo->fences, 1114 bo->num_fences * sizeof(*new_fences), 1115 new_max_fences * sizeof(*new_fences)); 1116 if (likely(new_fences)) { 1117 bo->fences = new_fences; 1118 bo->max_fences = new_max_fences; 1119 } else { 1120 unsigned drop; 1121 1122 fprintf(stderr, "amdgpu_add_fences: allocation failure, dropping fence(s)\n"); 1123 if (!bo->num_fences) 1124 return; 1125 1126 bo->num_fences--; /* prefer to keep the most recent fence if possible */ 1127 amdgpu_fence_reference(&bo->fences[bo->num_fences], NULL); 1128 1129 drop = bo->num_fences + num_fences - bo->max_fences; 1130 num_fences -= drop; 1131 fences += drop; 1132 } 1133 } 1134 1135 for (unsigned i = 0; i < num_fences; ++i) { 1136 bo->fences[bo->num_fences] = NULL; 1137 amdgpu_fence_reference(&bo->fences[bo->num_fences], fences[i]); 1138 bo->num_fences++; 1139 } 1140 } 1141 1142 static void amdgpu_add_fence_dependencies_bo_list(struct amdgpu_cs *acs, 1143 struct pipe_fence_handle *fence, 1144 unsigned num_buffers, 1145 struct amdgpu_cs_buffer *buffers) 1146 { 1147 for (unsigned i = 0; i < num_buffers; i++) { 1148 struct amdgpu_cs_buffer *buffer = &buffers[i]; 1149 struct amdgpu_winsys_bo *bo = buffer->bo; 1150 1151 amdgpu_add_bo_fence_dependencies(acs, buffer); 1152 p_atomic_inc(&bo->num_active_ioctls); 1153 amdgpu_add_fences(bo, 1, &fence); 1154 } 1155 } 1156 1157 /* Since the kernel driver doesn't synchronize execution between different 1158 * rings automatically, we have to add fence dependencies manually. 1159 */ 1160 static void amdgpu_add_fence_dependencies_bo_lists(struct amdgpu_cs *acs) 1161 { 1162 struct amdgpu_cs_context *cs = acs->csc; 1163 1164 cs->num_fence_dependencies = 0; 1165 1166 amdgpu_add_fence_dependencies_bo_list(acs, cs->fence, cs->num_real_buffers, cs->real_buffers); 1167 amdgpu_add_fence_dependencies_bo_list(acs, cs->fence, cs->num_slab_buffers, cs->slab_buffers); 1168 amdgpu_add_fence_dependencies_bo_list(acs, cs->fence, cs->num_sparse_buffers, cs->sparse_buffers); 1169 } 1170 1171 /* Add backing of sparse buffers to the buffer list. 1172 * 1173 * This is done late, during submission, to keep the buffer list short before 1174 * submit, and to avoid managing fences for the backing buffers. 1175 */ 1176 static bool amdgpu_add_sparse_backing_buffers(struct amdgpu_cs_context *cs) 1177 { 1178 for (unsigned i = 0; i < cs->num_sparse_buffers; ++i) { 1179 struct amdgpu_cs_buffer *buffer = &cs->sparse_buffers[i]; 1180 struct amdgpu_winsys_bo *bo = buffer->bo; 1181 1182 simple_mtx_lock(&bo->u.sparse.commit_lock); 1183 1184 list_for_each_entry(struct amdgpu_sparse_backing, backing, &bo->u.sparse.backing, list) { 1185 /* We can directly add the buffer here, because we know that each 1186 * backing buffer occurs only once. 1187 */ 1188 int idx = amdgpu_do_add_real_buffer(cs, backing->bo); 1189 if (idx < 0) { 1190 fprintf(stderr, "%s: failed to add buffer\n", __FUNCTION__); 1191 simple_mtx_unlock(&bo->u.sparse.commit_lock); 1192 return false; 1193 } 1194 1195 cs->real_buffers[idx].usage = buffer->usage & ~RADEON_USAGE_SYNCHRONIZED; 1196 cs->real_buffers[idx].u.real.priority_usage = buffer->u.real.priority_usage; 1197 p_atomic_inc(&backing->bo->num_active_ioctls); 1198 } 1199 1200 simple_mtx_unlock(&bo->u.sparse.commit_lock); 1201 } 1202 1203 return true; 1204 } 1205 1206 void amdgpu_cs_submit_ib(void *job, int thread_index) 1207 { 1208 struct amdgpu_cs *acs = (struct amdgpu_cs*)job; 1209 struct amdgpu_winsys *ws = acs->ctx->ws; 1210 struct amdgpu_cs_context *cs = acs->cst; 1211 int i, r; 1212 amdgpu_bo_list_handle bo_list = NULL; 1213 uint64_t seq_no = 0; 1214 bool has_user_fence = amdgpu_cs_has_user_fence(cs); 1215 1216 /* Create the buffer list. 1217 * Use a buffer list containing all allocated buffers if requested. 1218 */ 1219 if (ws->debug_all_bos) { 1220 struct amdgpu_winsys_bo *bo; 1221 amdgpu_bo_handle *handles; 1222 unsigned num = 0; 1223 1224 simple_mtx_lock(&ws->global_bo_list_lock); 1225 1226 handles = malloc(sizeof(handles[0]) * ws->num_buffers); 1227 if (!handles) { 1228 simple_mtx_unlock(&ws->global_bo_list_lock); 1229 amdgpu_cs_context_cleanup(cs); 1230 cs->error_code = -ENOMEM; 1231 return; 1232 } 1233 1234 LIST_FOR_EACH_ENTRY(bo, &ws->global_bo_list, u.real.global_list_item) { 1235 assert(num < ws->num_buffers); 1236 handles[num++] = bo->bo; 1237 } 1238 1239 r = amdgpu_bo_list_create(ws->dev, ws->num_buffers, 1240 handles, NULL, &bo_list); 1241 free(handles); 1242 simple_mtx_unlock(&ws->global_bo_list_lock); 1243 } else { 1244 unsigned num_handles; 1245 1246 if (!amdgpu_add_sparse_backing_buffers(cs)) { 1247 r = -ENOMEM; 1248 goto bo_list_error; 1249 } 1250 1251 if (cs->max_real_submit < cs->num_real_buffers) { 1252 FREE(cs->handles); 1253 FREE(cs->flags); 1254 1255 cs->handles = MALLOC(sizeof(*cs->handles) * cs->num_real_buffers); 1256 cs->flags = MALLOC(sizeof(*cs->flags) * cs->num_real_buffers); 1257 1258 if (!cs->handles || !cs->flags) { 1259 cs->max_real_submit = 0; 1260 r = -ENOMEM; 1261 goto bo_list_error; 1262 } 1263 } 1264 1265 num_handles = 0; 1266 for (i = 0; i < cs->num_real_buffers; ++i) { 1267 struct amdgpu_cs_buffer *buffer = &cs->real_buffers[i]; 1268 1269 if (buffer->bo->is_local) 1270 continue; 1271 1272 assert(buffer->u.real.priority_usage != 0); 1273 1274 cs->handles[num_handles] = buffer->bo->bo; 1275 cs->flags[num_handles] = (util_last_bit64(buffer->u.real.priority_usage) - 1) / 4; 1276 ++num_handles; 1277 } 1278 1279 if (acs->ring_type == RING_GFX) 1280 ws->gfx_bo_list_counter += cs->num_real_buffers; 1281 1282 if (num_handles) { 1283 r = amdgpu_bo_list_create(ws->dev, num_handles, 1284 cs->handles, cs->flags, &bo_list); 1285 } else { 1286 r = 0; 1287 } 1288 } 1289 bo_list_error: 1290 1291 if (r) { 1292 fprintf(stderr, "amdgpu: buffer list creation failed (%d)\n", r); 1293 amdgpu_fence_signalled(cs->fence); 1294 cs->error_code = r; 1295 goto cleanup; 1296 } 1297 1298 if (acs->ctx->num_rejected_cs) { 1299 r = -ECANCELED; 1300 } else { 1301 struct drm_amdgpu_cs_chunk chunks[4]; 1302 unsigned num_chunks = 0; 1303 1304 /* Convert from dwords to bytes. */ 1305 cs->ib[IB_MAIN].ib_bytes *= 4; 1306 1307 /* IB */ 1308 chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_IB; 1309 chunks[num_chunks].length_dw = sizeof(struct drm_amdgpu_cs_chunk_ib) / 4; 1310 chunks[num_chunks].chunk_data = (uintptr_t)&cs->ib[IB_MAIN]; 1311 num_chunks++; 1312 1313 /* Fence */ 1314 if (has_user_fence) { 1315 chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_FENCE; 1316 chunks[num_chunks].length_dw = sizeof(struct drm_amdgpu_cs_chunk_fence) / 4; 1317 chunks[num_chunks].chunk_data = (uintptr_t)&acs->fence_chunk; 1318 num_chunks++; 1319 } 1320 1321 /* Dependencies */ 1322 unsigned num_dependencies = cs->num_fence_dependencies; 1323 unsigned num_syncobj_dependencies = 0; 1324 1325 if (num_dependencies) { 1326 struct drm_amdgpu_cs_chunk_dep *dep_chunk = 1327 alloca(num_dependencies * sizeof(*dep_chunk)); 1328 unsigned num = 0; 1329 1330 for (unsigned i = 0; i < num_dependencies; i++) { 1331 struct amdgpu_fence *fence = 1332 (struct amdgpu_fence*)cs->fence_dependencies[i]; 1333 1334 if (amdgpu_fence_is_syncobj(fence)) { 1335 num_syncobj_dependencies++; 1336 continue; 1337 } 1338 1339 assert(util_queue_fence_is_signalled(&fence->submitted)); 1340 amdgpu_cs_chunk_fence_to_dep(&fence->fence, &dep_chunk[num++]); 1341 } 1342 1343 chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_DEPENDENCIES; 1344 chunks[num_chunks].length_dw = sizeof(dep_chunk[0]) / 4 * num; 1345 chunks[num_chunks].chunk_data = (uintptr_t)dep_chunk; 1346 num_chunks++; 1347 } 1348 1349 /* Syncobj dependencies. */ 1350 if (num_syncobj_dependencies) { 1351 struct drm_amdgpu_cs_chunk_sem *sem_chunk = 1352 alloca(num_syncobj_dependencies * sizeof(sem_chunk[0])); 1353 unsigned num = 0; 1354 1355 for (unsigned i = 0; i < num_dependencies; i++) { 1356 struct amdgpu_fence *fence = 1357 (struct amdgpu_fence*)cs->fence_dependencies[i]; 1358 1359 if (!amdgpu_fence_is_syncobj(fence)) 1360 continue; 1361 1362 assert(util_queue_fence_is_signalled(&fence->submitted)); 1363 sem_chunk[num++].handle = fence->syncobj; 1364 } 1365 1366 chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_SYNCOBJ_IN; 1367 chunks[num_chunks].length_dw = sizeof(sem_chunk[0]) / 4 * num; 1368 chunks[num_chunks].chunk_data = (uintptr_t)sem_chunk; 1369 num_chunks++; 1370 } 1371 1372 assert(num_chunks <= ARRAY_SIZE(chunks)); 1373 1374 r = amdgpu_cs_submit_raw(ws->dev, acs->ctx->ctx, bo_list, 1375 num_chunks, chunks, &seq_no); 1376 } 1377 1378 cs->error_code = r; 1379 if (r) { 1380 if (r == -ENOMEM) 1381 fprintf(stderr, "amdgpu: Not enough memory for command submission.\n"); 1382 else if (r == -ECANCELED) 1383 fprintf(stderr, "amdgpu: The CS has been cancelled because the context is lost.\n"); 1384 else 1385 fprintf(stderr, "amdgpu: The CS has been rejected, " 1386 "see dmesg for more information (%i).\n", r); 1387 1388 amdgpu_fence_signalled(cs->fence); 1389 1390 acs->ctx->num_rejected_cs++; 1391 ws->num_total_rejected_cs++; 1392 } else { 1393 /* Success. */ 1394 uint64_t *user_fence = NULL; 1395 1396 if (has_user_fence) 1397 user_fence = acs->ctx->user_fence_cpu_address_base + acs->ring_type; 1398 amdgpu_fence_submitted(cs->fence, seq_no, user_fence); 1399 } 1400 1401 /* Cleanup. */ 1402 if (bo_list) 1403 amdgpu_bo_list_destroy(bo_list); 1404 1405 cleanup: 1406 for (i = 0; i < cs->num_real_buffers; i++) 1407 p_atomic_dec(&cs->real_buffers[i].bo->num_active_ioctls); 1408 for (i = 0; i < cs->num_slab_buffers; i++) 1409 p_atomic_dec(&cs->slab_buffers[i].bo->num_active_ioctls); 1410 for (i = 0; i < cs->num_sparse_buffers; i++) 1411 p_atomic_dec(&cs->sparse_buffers[i].bo->num_active_ioctls); 1412 1413 amdgpu_cs_context_cleanup(cs); 1414 } 1415 1416 /* Make sure the previous submission is completed. */ 1417 void amdgpu_cs_sync_flush(struct radeon_winsys_cs *rcs) 1418 { 1419 struct amdgpu_cs *cs = amdgpu_cs(rcs); 1420 1421 /* Wait for any pending ioctl of this CS to complete. */ 1422 util_queue_fence_wait(&cs->flush_completed); 1423 } 1424 1425 static int amdgpu_cs_flush(struct radeon_winsys_cs *rcs, 1426 unsigned flags, 1427 struct pipe_fence_handle **fence) 1428 { 1429 struct amdgpu_cs *cs = amdgpu_cs(rcs); 1430 struct amdgpu_winsys *ws = cs->ctx->ws; 1431 int error_code = 0; 1432 1433 rcs->current.max_dw += amdgpu_cs_epilog_dws(cs->ring_type); 1434 1435 switch (cs->ring_type) { 1436 case RING_DMA: 1437 /* pad DMA ring to 8 DWs */ 1438 if (ws->info.chip_class <= SI) { 1439 while (rcs->current.cdw & 7) 1440 radeon_emit(rcs, 0xf0000000); /* NOP packet */ 1441 } else { 1442 while (rcs->current.cdw & 7) 1443 radeon_emit(rcs, 0x00000000); /* NOP packet */ 1444 } 1445 break; 1446 case RING_GFX: 1447 /* pad GFX ring to 8 DWs to meet CP fetch alignment requirements */ 1448 if (ws->info.gfx_ib_pad_with_type2) { 1449 while (rcs->current.cdw & 7) 1450 radeon_emit(rcs, 0x80000000); /* type2 nop packet */ 1451 } else { 1452 while (rcs->current.cdw & 7) 1453 radeon_emit(rcs, 0xffff1000); /* type3 nop packet */ 1454 } 1455 ws->gfx_ib_size_counter += (rcs->prev_dw + rcs->current.cdw) * 4; 1456 break; 1457 case RING_UVD: 1458 while (rcs->current.cdw & 15) 1459 radeon_emit(rcs, 0x80000000); /* type2 nop packet */ 1460 break; 1461 case RING_VCN_DEC: 1462 while (rcs->current.cdw & 15) 1463 radeon_emit(rcs, 0x81ff); /* nop packet */ 1464 break; 1465 default: 1466 break; 1467 } 1468 1469 if (rcs->current.cdw > rcs->current.max_dw) { 1470 fprintf(stderr, "amdgpu: command stream overflowed\n"); 1471 } 1472 1473 /* If the CS is not empty or overflowed.... */ 1474 if (likely(radeon_emitted(&cs->main.base, 0) && 1475 cs->main.base.current.cdw <= cs->main.base.current.max_dw && 1476 !debug_get_option_noop())) { 1477 struct amdgpu_cs_context *cur = cs->csc; 1478 1479 /* Set IB sizes. */ 1480 amdgpu_ib_finalize(ws, &cs->main); 1481 1482 /* Create a fence. */ 1483 amdgpu_fence_reference(&cur->fence, NULL); 1484 if (cs->next_fence) { 1485 /* just move the reference */ 1486 cur->fence = cs->next_fence; 1487 cs->next_fence = NULL; 1488 } else { 1489 cur->fence = amdgpu_fence_create(cs->ctx, 1490 cur->ib[IB_MAIN].ip_type, 1491 cur->ib[IB_MAIN].ip_instance, 1492 cur->ib[IB_MAIN].ring); 1493 } 1494 if (fence) 1495 amdgpu_fence_reference(fence, cur->fence); 1496 1497 amdgpu_cs_sync_flush(rcs); 1498 1499 /* Prepare buffers. 1500 * 1501 * This fence must be held until the submission is queued to ensure 1502 * that the order of fence dependency updates matches the order of 1503 * submissions. 1504 */ 1505 simple_mtx_lock(&ws->bo_fence_lock); 1506 amdgpu_add_fence_dependencies_bo_lists(cs); 1507 1508 /* Swap command streams. "cst" is going to be submitted. */ 1509 cs->csc = cs->cst; 1510 cs->cst = cur; 1511 1512 /* Submit. */ 1513 util_queue_add_job(&ws->cs_queue, cs, &cs->flush_completed, 1514 amdgpu_cs_submit_ib, NULL); 1515 /* The submission has been queued, unlock the fence now. */ 1516 simple_mtx_unlock(&ws->bo_fence_lock); 1517 1518 if (!(flags & PIPE_FLUSH_ASYNC)) { 1519 amdgpu_cs_sync_flush(rcs); 1520 error_code = cur->error_code; 1521 } 1522 } else { 1523 amdgpu_cs_context_cleanup(cs->csc); 1524 } 1525 1526 amdgpu_get_new_ib(&ws->base, cs, IB_MAIN); 1527 1528 cs->main.base.used_gart = 0; 1529 cs->main.base.used_vram = 0; 1530 1531 if (cs->ring_type == RING_GFX) 1532 ws->num_gfx_IBs++; 1533 else if (cs->ring_type == RING_DMA) 1534 ws->num_sdma_IBs++; 1535 1536 return error_code; 1537 } 1538 1539 static void amdgpu_cs_destroy(struct radeon_winsys_cs *rcs) 1540 { 1541 struct amdgpu_cs *cs = amdgpu_cs(rcs); 1542 1543 amdgpu_cs_sync_flush(rcs); 1544 util_queue_fence_destroy(&cs->flush_completed); 1545 p_atomic_dec(&cs->ctx->ws->num_cs); 1546 pb_reference(&cs->main.big_ib_buffer, NULL); 1547 FREE(cs->main.base.prev); 1548 amdgpu_destroy_cs_context(&cs->csc1); 1549 amdgpu_destroy_cs_context(&cs->csc2); 1550 amdgpu_fence_reference(&cs->next_fence, NULL); 1551 FREE(cs); 1552 } 1553 1554 static bool amdgpu_bo_is_referenced(struct radeon_winsys_cs *rcs, 1555 struct pb_buffer *_buf, 1556 enum radeon_bo_usage usage) 1557 { 1558 struct amdgpu_cs *cs = amdgpu_cs(rcs); 1559 struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)_buf; 1560 1561 return amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo, usage); 1562 } 1563 1564 void amdgpu_cs_init_functions(struct amdgpu_winsys *ws) 1565 { 1566 ws->base.ctx_create = amdgpu_ctx_create; 1567 ws->base.ctx_destroy = amdgpu_ctx_destroy; 1568 ws->base.ctx_query_reset_status = amdgpu_ctx_query_reset_status; 1569 ws->base.cs_create = amdgpu_cs_create; 1570 ws->base.cs_destroy = amdgpu_cs_destroy; 1571 ws->base.cs_add_buffer = amdgpu_cs_add_buffer; 1572 ws->base.cs_validate = amdgpu_cs_validate; 1573 ws->base.cs_check_space = amdgpu_cs_check_space; 1574 ws->base.cs_get_buffer_list = amdgpu_cs_get_buffer_list; 1575 ws->base.cs_flush = amdgpu_cs_flush; 1576 ws->base.cs_get_next_fence = amdgpu_cs_get_next_fence; 1577 ws->base.cs_is_buffer_referenced = amdgpu_bo_is_referenced; 1578 ws->base.cs_sync_flush = amdgpu_cs_sync_flush; 1579 ws->base.cs_add_fence_dependency = amdgpu_cs_add_fence_dependency; 1580 ws->base.fence_wait = amdgpu_fence_wait_rel_timeout; 1581 ws->base.fence_reference = amdgpu_fence_reference; 1582 ws->base.fence_import_sync_file = amdgpu_fence_import_sync_file; 1583 ws->base.fence_export_sync_file = amdgpu_fence_export_sync_file; 1584 ws->base.export_signalled_sync_file = amdgpu_export_signalled_sync_file; 1585 } 1586