1 /* 2 * Copyright 2015 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 */ 23 24 #include <assert.h> 25 #include <stdbool.h> 26 #include <string.h> 27 #include <unistd.h> 28 #include <fcntl.h> 29 30 #include "anv_private.h" 31 32 #include "genxml/gen8_pack.h" 33 34 #include "util/debug.h" 35 36 /** \file anv_batch_chain.c 37 * 38 * This file contains functions related to anv_cmd_buffer as a data 39 * structure. This involves everything required to create and destroy 40 * the actual batch buffers as well as link them together and handle 41 * relocations and surface state. It specifically does *not* contain any 42 * handling of actual vkCmd calls beyond vkCmdExecuteCommands. 43 */ 44 45 /*-----------------------------------------------------------------------* 46 * Functions related to anv_reloc_list 47 *-----------------------------------------------------------------------*/ 48 49 static VkResult 50 anv_reloc_list_init_clone(struct anv_reloc_list *list, 51 const VkAllocationCallbacks *alloc, 52 const struct anv_reloc_list *other_list) 53 { 54 if (other_list) { 55 list->num_relocs = other_list->num_relocs; 56 list->array_length = other_list->array_length; 57 } else { 58 list->num_relocs = 0; 59 list->array_length = 256; 60 } 61 62 list->relocs = 63 vk_alloc(alloc, list->array_length * sizeof(*list->relocs), 8, 64 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); 65 66 if (list->relocs == NULL) 67 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 68 69 list->reloc_bos = 70 vk_alloc(alloc, list->array_length * sizeof(*list->reloc_bos), 8, 71 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); 72 73 if (list->reloc_bos == NULL) { 74 vk_free(alloc, list->relocs); 75 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 76 } 77 78 if (other_list) { 79 memcpy(list->relocs, other_list->relocs, 80 list->array_length * sizeof(*list->relocs)); 81 memcpy(list->reloc_bos, other_list->reloc_bos, 82 list->array_length * sizeof(*list->reloc_bos)); 83 } 84 85 return VK_SUCCESS; 86 } 87 88 VkResult 89 anv_reloc_list_init(struct anv_reloc_list *list, 90 const VkAllocationCallbacks *alloc) 91 { 92 return anv_reloc_list_init_clone(list, alloc, NULL); 93 } 94 95 void 96 anv_reloc_list_finish(struct anv_reloc_list *list, 97 const VkAllocationCallbacks *alloc) 98 { 99 vk_free(alloc, list->relocs); 100 vk_free(alloc, list->reloc_bos); 101 } 102 103 static VkResult 104 anv_reloc_list_grow(struct anv_reloc_list *list, 105 const VkAllocationCallbacks *alloc, 106 size_t num_additional_relocs) 107 { 108 if (list->num_relocs + num_additional_relocs <= list->array_length) 109 return VK_SUCCESS; 110 111 size_t new_length = list->array_length * 2; 112 while (new_length < list->num_relocs + num_additional_relocs) 113 new_length *= 2; 114 115 struct drm_i915_gem_relocation_entry *new_relocs = 116 vk_alloc(alloc, new_length * sizeof(*list->relocs), 8, 117 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); 118 if (new_relocs == NULL) 119 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 120 121 struct anv_bo **new_reloc_bos = 122 vk_alloc(alloc, new_length * sizeof(*list->reloc_bos), 8, 123 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); 124 if (new_reloc_bos == NULL) { 125 vk_free(alloc, new_relocs); 126 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 127 } 128 129 memcpy(new_relocs, list->relocs, list->num_relocs * sizeof(*list->relocs)); 130 memcpy(new_reloc_bos, list->reloc_bos, 131 list->num_relocs * sizeof(*list->reloc_bos)); 132 133 vk_free(alloc, list->relocs); 134 vk_free(alloc, list->reloc_bos); 135 136 list->array_length = new_length; 137 list->relocs = new_relocs; 138 list->reloc_bos = new_reloc_bos; 139 140 return VK_SUCCESS; 141 } 142 143 uint64_t 144 anv_reloc_list_add(struct anv_reloc_list *list, 145 const VkAllocationCallbacks *alloc, 146 uint32_t offset, struct anv_bo *target_bo, uint32_t delta) 147 { 148 struct drm_i915_gem_relocation_entry *entry; 149 int index; 150 151 const uint32_t domain = 152 target_bo->is_winsys_bo ? I915_GEM_DOMAIN_RENDER : 0; 153 154 anv_reloc_list_grow(list, alloc, 1); 155 /* TODO: Handle failure */ 156 157 /* XXX: Can we use I915_EXEC_HANDLE_LUT? */ 158 index = list->num_relocs++; 159 list->reloc_bos[index] = target_bo; 160 entry = &list->relocs[index]; 161 entry->target_handle = target_bo->gem_handle; 162 entry->delta = delta; 163 entry->offset = offset; 164 entry->presumed_offset = target_bo->offset; 165 entry->read_domains = domain; 166 entry->write_domain = domain; 167 VG(VALGRIND_CHECK_MEM_IS_DEFINED(entry, sizeof(*entry))); 168 169 return target_bo->offset + delta; 170 } 171 172 static void 173 anv_reloc_list_append(struct anv_reloc_list *list, 174 const VkAllocationCallbacks *alloc, 175 struct anv_reloc_list *other, uint32_t offset) 176 { 177 anv_reloc_list_grow(list, alloc, other->num_relocs); 178 /* TODO: Handle failure */ 179 180 memcpy(&list->relocs[list->num_relocs], &other->relocs[0], 181 other->num_relocs * sizeof(other->relocs[0])); 182 memcpy(&list->reloc_bos[list->num_relocs], &other->reloc_bos[0], 183 other->num_relocs * sizeof(other->reloc_bos[0])); 184 185 for (uint32_t i = 0; i < other->num_relocs; i++) 186 list->relocs[i + list->num_relocs].offset += offset; 187 188 list->num_relocs += other->num_relocs; 189 } 190 191 /*-----------------------------------------------------------------------* 192 * Functions related to anv_batch 193 *-----------------------------------------------------------------------*/ 194 195 void * 196 anv_batch_emit_dwords(struct anv_batch *batch, int num_dwords) 197 { 198 if (batch->next + num_dwords * 4 > batch->end) 199 batch->extend_cb(batch, batch->user_data); 200 201 void *p = batch->next; 202 203 batch->next += num_dwords * 4; 204 assert(batch->next <= batch->end); 205 206 return p; 207 } 208 209 uint64_t 210 anv_batch_emit_reloc(struct anv_batch *batch, 211 void *location, struct anv_bo *bo, uint32_t delta) 212 { 213 return anv_reloc_list_add(batch->relocs, batch->alloc, 214 location - batch->start, bo, delta); 215 } 216 217 void 218 anv_batch_emit_batch(struct anv_batch *batch, struct anv_batch *other) 219 { 220 uint32_t size, offset; 221 222 size = other->next - other->start; 223 assert(size % 4 == 0); 224 225 if (batch->next + size > batch->end) 226 batch->extend_cb(batch, batch->user_data); 227 228 assert(batch->next + size <= batch->end); 229 230 VG(VALGRIND_CHECK_MEM_IS_DEFINED(other->start, size)); 231 memcpy(batch->next, other->start, size); 232 233 offset = batch->next - batch->start; 234 anv_reloc_list_append(batch->relocs, batch->alloc, 235 other->relocs, offset); 236 237 batch->next += size; 238 } 239 240 /*-----------------------------------------------------------------------* 241 * Functions related to anv_batch_bo 242 *-----------------------------------------------------------------------*/ 243 244 static VkResult 245 anv_batch_bo_create(struct anv_cmd_buffer *cmd_buffer, 246 struct anv_batch_bo **bbo_out) 247 { 248 VkResult result; 249 250 struct anv_batch_bo *bbo = vk_alloc(&cmd_buffer->pool->alloc, sizeof(*bbo), 251 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); 252 if (bbo == NULL) 253 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 254 255 result = anv_bo_pool_alloc(&cmd_buffer->device->batch_bo_pool, &bbo->bo, 256 ANV_CMD_BUFFER_BATCH_SIZE); 257 if (result != VK_SUCCESS) 258 goto fail_alloc; 259 260 result = anv_reloc_list_init(&bbo->relocs, &cmd_buffer->pool->alloc); 261 if (result != VK_SUCCESS) 262 goto fail_bo_alloc; 263 264 *bbo_out = bbo; 265 266 return VK_SUCCESS; 267 268 fail_bo_alloc: 269 anv_bo_pool_free(&cmd_buffer->device->batch_bo_pool, &bbo->bo); 270 fail_alloc: 271 vk_free(&cmd_buffer->pool->alloc, bbo); 272 273 return result; 274 } 275 276 static VkResult 277 anv_batch_bo_clone(struct anv_cmd_buffer *cmd_buffer, 278 const struct anv_batch_bo *other_bbo, 279 struct anv_batch_bo **bbo_out) 280 { 281 VkResult result; 282 283 struct anv_batch_bo *bbo = vk_alloc(&cmd_buffer->pool->alloc, sizeof(*bbo), 284 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); 285 if (bbo == NULL) 286 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 287 288 result = anv_bo_pool_alloc(&cmd_buffer->device->batch_bo_pool, &bbo->bo, 289 other_bbo->bo.size); 290 if (result != VK_SUCCESS) 291 goto fail_alloc; 292 293 result = anv_reloc_list_init_clone(&bbo->relocs, &cmd_buffer->pool->alloc, 294 &other_bbo->relocs); 295 if (result != VK_SUCCESS) 296 goto fail_bo_alloc; 297 298 bbo->length = other_bbo->length; 299 memcpy(bbo->bo.map, other_bbo->bo.map, other_bbo->length); 300 301 *bbo_out = bbo; 302 303 return VK_SUCCESS; 304 305 fail_bo_alloc: 306 anv_bo_pool_free(&cmd_buffer->device->batch_bo_pool, &bbo->bo); 307 fail_alloc: 308 vk_free(&cmd_buffer->pool->alloc, bbo); 309 310 return result; 311 } 312 313 static void 314 anv_batch_bo_start(struct anv_batch_bo *bbo, struct anv_batch *batch, 315 size_t batch_padding) 316 { 317 batch->next = batch->start = bbo->bo.map; 318 batch->end = bbo->bo.map + bbo->bo.size - batch_padding; 319 batch->relocs = &bbo->relocs; 320 bbo->relocs.num_relocs = 0; 321 } 322 323 static void 324 anv_batch_bo_continue(struct anv_batch_bo *bbo, struct anv_batch *batch, 325 size_t batch_padding) 326 { 327 batch->start = bbo->bo.map; 328 batch->next = bbo->bo.map + bbo->length; 329 batch->end = bbo->bo.map + bbo->bo.size - batch_padding; 330 batch->relocs = &bbo->relocs; 331 } 332 333 static void 334 anv_batch_bo_finish(struct anv_batch_bo *bbo, struct anv_batch *batch) 335 { 336 assert(batch->start == bbo->bo.map); 337 bbo->length = batch->next - batch->start; 338 VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch->start, bbo->length)); 339 } 340 341 static VkResult 342 anv_batch_bo_grow(struct anv_cmd_buffer *cmd_buffer, struct anv_batch_bo *bbo, 343 struct anv_batch *batch, size_t aditional, 344 size_t batch_padding) 345 { 346 assert(batch->start == bbo->bo.map); 347 bbo->length = batch->next - batch->start; 348 349 size_t new_size = bbo->bo.size; 350 while (new_size <= bbo->length + aditional + batch_padding) 351 new_size *= 2; 352 353 if (new_size == bbo->bo.size) 354 return VK_SUCCESS; 355 356 struct anv_bo new_bo; 357 VkResult result = anv_bo_pool_alloc(&cmd_buffer->device->batch_bo_pool, 358 &new_bo, new_size); 359 if (result != VK_SUCCESS) 360 return result; 361 362 memcpy(new_bo.map, bbo->bo.map, bbo->length); 363 364 anv_bo_pool_free(&cmd_buffer->device->batch_bo_pool, &bbo->bo); 365 366 bbo->bo = new_bo; 367 anv_batch_bo_continue(bbo, batch, batch_padding); 368 369 return VK_SUCCESS; 370 } 371 372 static void 373 anv_batch_bo_destroy(struct anv_batch_bo *bbo, 374 struct anv_cmd_buffer *cmd_buffer) 375 { 376 anv_reloc_list_finish(&bbo->relocs, &cmd_buffer->pool->alloc); 377 anv_bo_pool_free(&cmd_buffer->device->batch_bo_pool, &bbo->bo); 378 vk_free(&cmd_buffer->pool->alloc, bbo); 379 } 380 381 static VkResult 382 anv_batch_bo_list_clone(const struct list_head *list, 383 struct anv_cmd_buffer *cmd_buffer, 384 struct list_head *new_list) 385 { 386 VkResult result = VK_SUCCESS; 387 388 list_inithead(new_list); 389 390 struct anv_batch_bo *prev_bbo = NULL; 391 list_for_each_entry(struct anv_batch_bo, bbo, list, link) { 392 struct anv_batch_bo *new_bbo = NULL; 393 result = anv_batch_bo_clone(cmd_buffer, bbo, &new_bbo); 394 if (result != VK_SUCCESS) 395 break; 396 list_addtail(&new_bbo->link, new_list); 397 398 if (prev_bbo) { 399 /* As we clone this list of batch_bo's, they chain one to the 400 * other using MI_BATCH_BUFFER_START commands. We need to fix up 401 * those relocations as we go. Fortunately, this is pretty easy 402 * as it will always be the last relocation in the list. 403 */ 404 uint32_t last_idx = prev_bbo->relocs.num_relocs - 1; 405 assert(prev_bbo->relocs.reloc_bos[last_idx] == &bbo->bo); 406 prev_bbo->relocs.reloc_bos[last_idx] = &new_bbo->bo; 407 } 408 409 prev_bbo = new_bbo; 410 } 411 412 if (result != VK_SUCCESS) { 413 list_for_each_entry_safe(struct anv_batch_bo, bbo, new_list, link) 414 anv_batch_bo_destroy(bbo, cmd_buffer); 415 } 416 417 return result; 418 } 419 420 /*-----------------------------------------------------------------------* 421 * Functions related to anv_batch_bo 422 *-----------------------------------------------------------------------*/ 423 424 static inline struct anv_batch_bo * 425 anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer *cmd_buffer) 426 { 427 return LIST_ENTRY(struct anv_batch_bo, cmd_buffer->batch_bos.prev, link); 428 } 429 430 struct anv_address 431 anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer *cmd_buffer) 432 { 433 return (struct anv_address) { 434 .bo = &cmd_buffer->device->surface_state_block_pool.bo, 435 .offset = *(int32_t *)u_vector_head(&cmd_buffer->bt_blocks), 436 }; 437 } 438 439 static void 440 emit_batch_buffer_start(struct anv_cmd_buffer *cmd_buffer, 441 struct anv_bo *bo, uint32_t offset) 442 { 443 /* In gen8+ the address field grew to two dwords to accomodate 48 bit 444 * offsets. The high 16 bits are in the last dword, so we can use the gen8 445 * version in either case, as long as we set the instruction length in the 446 * header accordingly. This means that we always emit three dwords here 447 * and all the padding and adjustment we do in this file works for all 448 * gens. 449 */ 450 451 #define GEN7_MI_BATCH_BUFFER_START_length 2 452 #define GEN7_MI_BATCH_BUFFER_START_length_bias 2 453 454 const uint32_t gen7_length = 455 GEN7_MI_BATCH_BUFFER_START_length - GEN7_MI_BATCH_BUFFER_START_length_bias; 456 const uint32_t gen8_length = 457 GEN8_MI_BATCH_BUFFER_START_length - GEN8_MI_BATCH_BUFFER_START_length_bias; 458 459 anv_batch_emit(&cmd_buffer->batch, GEN8_MI_BATCH_BUFFER_START, bbs) { 460 bbs.DWordLength = cmd_buffer->device->info.gen < 8 ? 461 gen7_length : gen8_length; 462 bbs._2ndLevelBatchBuffer = _1stlevelbatch; 463 bbs.AddressSpaceIndicator = ASI_PPGTT; 464 bbs.BatchBufferStartAddress = (struct anv_address) { bo, offset }; 465 } 466 } 467 468 static void 469 cmd_buffer_chain_to_batch_bo(struct anv_cmd_buffer *cmd_buffer, 470 struct anv_batch_bo *bbo) 471 { 472 struct anv_batch *batch = &cmd_buffer->batch; 473 struct anv_batch_bo *current_bbo = 474 anv_cmd_buffer_current_batch_bo(cmd_buffer); 475 476 /* We set the end of the batch a little short so we would be sure we 477 * have room for the chaining command. Since we're about to emit the 478 * chaining command, let's set it back where it should go. 479 */ 480 batch->end += GEN8_MI_BATCH_BUFFER_START_length * 4; 481 assert(batch->end == current_bbo->bo.map + current_bbo->bo.size); 482 483 emit_batch_buffer_start(cmd_buffer, &bbo->bo, 0); 484 485 anv_batch_bo_finish(current_bbo, batch); 486 } 487 488 static VkResult 489 anv_cmd_buffer_chain_batch(struct anv_batch *batch, void *_data) 490 { 491 struct anv_cmd_buffer *cmd_buffer = _data; 492 struct anv_batch_bo *new_bbo; 493 494 VkResult result = anv_batch_bo_create(cmd_buffer, &new_bbo); 495 if (result != VK_SUCCESS) 496 return result; 497 498 struct anv_batch_bo **seen_bbo = u_vector_add(&cmd_buffer->seen_bbos); 499 if (seen_bbo == NULL) { 500 anv_batch_bo_destroy(new_bbo, cmd_buffer); 501 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 502 } 503 *seen_bbo = new_bbo; 504 505 cmd_buffer_chain_to_batch_bo(cmd_buffer, new_bbo); 506 507 list_addtail(&new_bbo->link, &cmd_buffer->batch_bos); 508 509 anv_batch_bo_start(new_bbo, batch, GEN8_MI_BATCH_BUFFER_START_length * 4); 510 511 return VK_SUCCESS; 512 } 513 514 static VkResult 515 anv_cmd_buffer_grow_batch(struct anv_batch *batch, void *_data) 516 { 517 struct anv_cmd_buffer *cmd_buffer = _data; 518 struct anv_batch_bo *bbo = anv_cmd_buffer_current_batch_bo(cmd_buffer); 519 520 anv_batch_bo_grow(cmd_buffer, bbo, &cmd_buffer->batch, 4096, 521 GEN8_MI_BATCH_BUFFER_START_length * 4); 522 523 return VK_SUCCESS; 524 } 525 526 /** Allocate a binding table 527 * 528 * This function allocates a binding table. This is a bit more complicated 529 * than one would think due to a combination of Vulkan driver design and some 530 * unfortunate hardware restrictions. 531 * 532 * The 3DSTATE_BINDING_TABLE_POINTERS_* packets only have a 16-bit field for 533 * the binding table pointer which means that all binding tables need to live 534 * in the bottom 64k of surface state base address. The way the GL driver has 535 * classically dealt with this restriction is to emit all surface states 536 * on-the-fly into the batch and have a batch buffer smaller than 64k. This 537 * isn't really an option in Vulkan for a couple of reasons: 538 * 539 * 1) In Vulkan, we have growing (or chaining) batches so surface states have 540 * to live in their own buffer and we have to be able to re-emit 541 * STATE_BASE_ADDRESS as needed which requires a full pipeline stall. In 542 * order to avoid emitting STATE_BASE_ADDRESS any more often than needed 543 * (it's not that hard to hit 64k of just binding tables), we allocate 544 * surface state objects up-front when VkImageView is created. In order 545 * for this to work, surface state objects need to be allocated from a 546 * global buffer. 547 * 548 * 2) We tried to design the surface state system in such a way that it's 549 * already ready for bindless texturing. The way bindless texturing works 550 * on our hardware is that you have a big pool of surface state objects 551 * (with its own state base address) and the bindless handles are simply 552 * offsets into that pool. With the architecture we chose, we already 553 * have that pool and it's exactly the same pool that we use for regular 554 * surface states so we should already be ready for bindless. 555 * 556 * 3) For render targets, we need to be able to fill out the surface states 557 * later in vkBeginRenderPass so that we can assign clear colors 558 * correctly. One way to do this would be to just create the surface 559 * state data and then repeatedly copy it into the surface state BO every 560 * time we have to re-emit STATE_BASE_ADDRESS. While this works, it's 561 * rather annoying and just being able to allocate them up-front and 562 * re-use them for the entire render pass. 563 * 564 * While none of these are technically blockers for emitting state on the fly 565 * like we do in GL, the ability to have a single surface state pool is 566 * simplifies things greatly. Unfortunately, it comes at a cost... 567 * 568 * Because of the 64k limitation of 3DSTATE_BINDING_TABLE_POINTERS_*, we can't 569 * place the binding tables just anywhere in surface state base address. 570 * Because 64k isn't a whole lot of space, we can't simply restrict the 571 * surface state buffer to 64k, we have to be more clever. The solution we've 572 * chosen is to have a block pool with a maximum size of 2G that starts at 573 * zero and grows in both directions. All surface states are allocated from 574 * the top of the pool (positive offsets) and we allocate blocks (< 64k) of 575 * binding tables from the bottom of the pool (negative offsets). Every time 576 * we allocate a new binding table block, we set surface state base address to 577 * point to the bottom of the binding table block. This way all of the 578 * binding tables in the block are in the bottom 64k of surface state base 579 * address. When we fill out the binding table, we add the distance between 580 * the bottom of our binding table block and zero of the block pool to the 581 * surface state offsets so that they are correct relative to out new surface 582 * state base address at the bottom of the binding table block. 583 * 584 * \see adjust_relocations_from_block_pool() 585 * \see adjust_relocations_too_block_pool() 586 * 587 * \param[in] entries The number of surface state entries the binding 588 * table should be able to hold. 589 * 590 * \param[out] state_offset The offset surface surface state base address 591 * where the surface states live. This must be 592 * added to the surface state offset when it is 593 * written into the binding table entry. 594 * 595 * \return An anv_state representing the binding table 596 */ 597 struct anv_state 598 anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer *cmd_buffer, 599 uint32_t entries, uint32_t *state_offset) 600 { 601 struct anv_block_pool *block_pool = 602 &cmd_buffer->device->surface_state_block_pool; 603 int32_t *bt_block = u_vector_head(&cmd_buffer->bt_blocks); 604 struct anv_state state; 605 606 state.alloc_size = align_u32(entries * 4, 32); 607 608 if (cmd_buffer->bt_next + state.alloc_size > block_pool->block_size) 609 return (struct anv_state) { 0 }; 610 611 state.offset = cmd_buffer->bt_next; 612 state.map = block_pool->map + *bt_block + state.offset; 613 614 cmd_buffer->bt_next += state.alloc_size; 615 616 assert(*bt_block < 0); 617 *state_offset = -(*bt_block); 618 619 return state; 620 } 621 622 struct anv_state 623 anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer *cmd_buffer) 624 { 625 struct isl_device *isl_dev = &cmd_buffer->device->isl_dev; 626 return anv_state_stream_alloc(&cmd_buffer->surface_state_stream, 627 isl_dev->ss.size, isl_dev->ss.align); 628 } 629 630 struct anv_state 631 anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer *cmd_buffer, 632 uint32_t size, uint32_t alignment) 633 { 634 return anv_state_stream_alloc(&cmd_buffer->dynamic_state_stream, 635 size, alignment); 636 } 637 638 VkResult 639 anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer *cmd_buffer) 640 { 641 struct anv_block_pool *block_pool = 642 &cmd_buffer->device->surface_state_block_pool; 643 644 int32_t *offset = u_vector_add(&cmd_buffer->bt_blocks); 645 if (offset == NULL) 646 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 647 648 *offset = anv_block_pool_alloc_back(block_pool); 649 cmd_buffer->bt_next = 0; 650 651 return VK_SUCCESS; 652 } 653 654 VkResult 655 anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer) 656 { 657 struct anv_batch_bo *batch_bo; 658 VkResult result; 659 660 list_inithead(&cmd_buffer->batch_bos); 661 662 result = anv_batch_bo_create(cmd_buffer, &batch_bo); 663 if (result != VK_SUCCESS) 664 return result; 665 666 list_addtail(&batch_bo->link, &cmd_buffer->batch_bos); 667 668 cmd_buffer->batch.alloc = &cmd_buffer->pool->alloc; 669 cmd_buffer->batch.user_data = cmd_buffer; 670 671 if (cmd_buffer->device->can_chain_batches) { 672 cmd_buffer->batch.extend_cb = anv_cmd_buffer_chain_batch; 673 } else { 674 cmd_buffer->batch.extend_cb = anv_cmd_buffer_grow_batch; 675 } 676 677 anv_batch_bo_start(batch_bo, &cmd_buffer->batch, 678 GEN8_MI_BATCH_BUFFER_START_length * 4); 679 680 int success = u_vector_init(&cmd_buffer->seen_bbos, 681 sizeof(struct anv_bo *), 682 8 * sizeof(struct anv_bo *)); 683 if (!success) 684 goto fail_batch_bo; 685 686 *(struct anv_batch_bo **)u_vector_add(&cmd_buffer->seen_bbos) = batch_bo; 687 688 success = u_vector_init(&cmd_buffer->bt_blocks, sizeof(int32_t), 689 8 * sizeof(int32_t)); 690 if (!success) 691 goto fail_seen_bbos; 692 693 result = anv_reloc_list_init(&cmd_buffer->surface_relocs, 694 &cmd_buffer->pool->alloc); 695 if (result != VK_SUCCESS) 696 goto fail_bt_blocks; 697 cmd_buffer->last_ss_pool_center = 0; 698 699 anv_cmd_buffer_new_binding_table_block(cmd_buffer); 700 701 return VK_SUCCESS; 702 703 fail_bt_blocks: 704 u_vector_finish(&cmd_buffer->bt_blocks); 705 fail_seen_bbos: 706 u_vector_finish(&cmd_buffer->seen_bbos); 707 fail_batch_bo: 708 anv_batch_bo_destroy(batch_bo, cmd_buffer); 709 710 return result; 711 } 712 713 void 714 anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer) 715 { 716 int32_t *bt_block; 717 u_vector_foreach(bt_block, &cmd_buffer->bt_blocks) { 718 anv_block_pool_free(&cmd_buffer->device->surface_state_block_pool, 719 *bt_block); 720 } 721 u_vector_finish(&cmd_buffer->bt_blocks); 722 723 anv_reloc_list_finish(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc); 724 725 u_vector_finish(&cmd_buffer->seen_bbos); 726 727 /* Destroy all of the batch buffers */ 728 list_for_each_entry_safe(struct anv_batch_bo, bbo, 729 &cmd_buffer->batch_bos, link) { 730 anv_batch_bo_destroy(bbo, cmd_buffer); 731 } 732 } 733 734 void 735 anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer) 736 { 737 /* Delete all but the first batch bo */ 738 assert(!list_empty(&cmd_buffer->batch_bos)); 739 while (cmd_buffer->batch_bos.next != cmd_buffer->batch_bos.prev) { 740 struct anv_batch_bo *bbo = anv_cmd_buffer_current_batch_bo(cmd_buffer); 741 list_del(&bbo->link); 742 anv_batch_bo_destroy(bbo, cmd_buffer); 743 } 744 assert(!list_empty(&cmd_buffer->batch_bos)); 745 746 anv_batch_bo_start(anv_cmd_buffer_current_batch_bo(cmd_buffer), 747 &cmd_buffer->batch, 748 GEN8_MI_BATCH_BUFFER_START_length * 4); 749 750 while (u_vector_length(&cmd_buffer->bt_blocks) > 1) { 751 int32_t *bt_block = u_vector_remove(&cmd_buffer->bt_blocks); 752 anv_block_pool_free(&cmd_buffer->device->surface_state_block_pool, 753 *bt_block); 754 } 755 assert(u_vector_length(&cmd_buffer->bt_blocks) == 1); 756 cmd_buffer->bt_next = 0; 757 758 cmd_buffer->surface_relocs.num_relocs = 0; 759 cmd_buffer->last_ss_pool_center = 0; 760 761 /* Reset the list of seen buffers */ 762 cmd_buffer->seen_bbos.head = 0; 763 cmd_buffer->seen_bbos.tail = 0; 764 765 *(struct anv_batch_bo **)u_vector_add(&cmd_buffer->seen_bbos) = 766 anv_cmd_buffer_current_batch_bo(cmd_buffer); 767 } 768 769 void 770 anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer *cmd_buffer) 771 { 772 struct anv_batch_bo *batch_bo = anv_cmd_buffer_current_batch_bo(cmd_buffer); 773 774 if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) { 775 /* When we start a batch buffer, we subtract a certain amount of 776 * padding from the end to ensure that we always have room to emit a 777 * BATCH_BUFFER_START to chain to the next BO. We need to remove 778 * that padding before we end the batch; otherwise, we may end up 779 * with our BATCH_BUFFER_END in another BO. 780 */ 781 cmd_buffer->batch.end += GEN8_MI_BATCH_BUFFER_START_length * 4; 782 assert(cmd_buffer->batch.end == batch_bo->bo.map + batch_bo->bo.size); 783 784 anv_batch_emit(&cmd_buffer->batch, GEN8_MI_BATCH_BUFFER_END, bbe); 785 786 /* Round batch up to an even number of dwords. */ 787 if ((cmd_buffer->batch.next - cmd_buffer->batch.start) & 4) 788 anv_batch_emit(&cmd_buffer->batch, GEN8_MI_NOOP, noop); 789 790 cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_PRIMARY; 791 } 792 793 anv_batch_bo_finish(batch_bo, &cmd_buffer->batch); 794 795 if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY) { 796 /* If this is a secondary command buffer, we need to determine the 797 * mode in which it will be executed with vkExecuteCommands. We 798 * determine this statically here so that this stays in sync with the 799 * actual ExecuteCommands implementation. 800 */ 801 if (!cmd_buffer->device->can_chain_batches) { 802 cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT; 803 } else if ((cmd_buffer->batch_bos.next == cmd_buffer->batch_bos.prev) && 804 (batch_bo->length < ANV_CMD_BUFFER_BATCH_SIZE / 2)) { 805 /* If the secondary has exactly one batch buffer in its list *and* 806 * that batch buffer is less than half of the maximum size, we're 807 * probably better of simply copying it into our batch. 808 */ 809 cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_EMIT; 810 } else if (!(cmd_buffer->usage_flags & 811 VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT)) { 812 cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_CHAIN; 813 814 /* When we chain, we need to add an MI_BATCH_BUFFER_START command 815 * with its relocation. In order to handle this we'll increment here 816 * so we can unconditionally decrement right before adding the 817 * MI_BATCH_BUFFER_START command. 818 */ 819 batch_bo->relocs.num_relocs++; 820 cmd_buffer->batch.next += GEN8_MI_BATCH_BUFFER_START_length * 4; 821 } else { 822 cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN; 823 } 824 } 825 } 826 827 static inline VkResult 828 anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer *cmd_buffer, 829 struct list_head *list) 830 { 831 list_for_each_entry(struct anv_batch_bo, bbo, list, link) { 832 struct anv_batch_bo **bbo_ptr = u_vector_add(&cmd_buffer->seen_bbos); 833 if (bbo_ptr == NULL) 834 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 835 836 *bbo_ptr = bbo; 837 } 838 839 return VK_SUCCESS; 840 } 841 842 void 843 anv_cmd_buffer_add_secondary(struct anv_cmd_buffer *primary, 844 struct anv_cmd_buffer *secondary) 845 { 846 switch (secondary->exec_mode) { 847 case ANV_CMD_BUFFER_EXEC_MODE_EMIT: 848 anv_batch_emit_batch(&primary->batch, &secondary->batch); 849 break; 850 case ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT: { 851 struct anv_batch_bo *bbo = anv_cmd_buffer_current_batch_bo(primary); 852 unsigned length = secondary->batch.end - secondary->batch.start; 853 anv_batch_bo_grow(primary, bbo, &primary->batch, length, 854 GEN8_MI_BATCH_BUFFER_START_length * 4); 855 anv_batch_emit_batch(&primary->batch, &secondary->batch); 856 break; 857 } 858 case ANV_CMD_BUFFER_EXEC_MODE_CHAIN: { 859 struct anv_batch_bo *first_bbo = 860 list_first_entry(&secondary->batch_bos, struct anv_batch_bo, link); 861 struct anv_batch_bo *last_bbo = 862 list_last_entry(&secondary->batch_bos, struct anv_batch_bo, link); 863 864 emit_batch_buffer_start(primary, &first_bbo->bo, 0); 865 866 struct anv_batch_bo *this_bbo = anv_cmd_buffer_current_batch_bo(primary); 867 assert(primary->batch.start == this_bbo->bo.map); 868 uint32_t offset = primary->batch.next - primary->batch.start; 869 const uint32_t inst_size = GEN8_MI_BATCH_BUFFER_START_length * 4; 870 871 /* Roll back the previous MI_BATCH_BUFFER_START and its relocation so we 872 * can emit a new command and relocation for the current splice. In 873 * order to handle the initial-use case, we incremented next and 874 * num_relocs in end_batch_buffer() so we can alyways just subtract 875 * here. 876 */ 877 last_bbo->relocs.num_relocs--; 878 secondary->batch.next -= inst_size; 879 emit_batch_buffer_start(secondary, &this_bbo->bo, offset); 880 anv_cmd_buffer_add_seen_bbos(primary, &secondary->batch_bos); 881 882 /* After patching up the secondary buffer, we need to clflush the 883 * modified instruction in case we're on a !llc platform. We use a 884 * little loop to handle the case where the instruction crosses a cache 885 * line boundary. 886 */ 887 if (!primary->device->info.has_llc) { 888 void *inst = secondary->batch.next - inst_size; 889 void *p = (void *) (((uintptr_t) inst) & ~CACHELINE_MASK); 890 __builtin_ia32_mfence(); 891 while (p < secondary->batch.next) { 892 __builtin_ia32_clflush(p); 893 p += CACHELINE_SIZE; 894 } 895 } 896 break; 897 } 898 case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN: { 899 struct list_head copy_list; 900 VkResult result = anv_batch_bo_list_clone(&secondary->batch_bos, 901 secondary, 902 ©_list); 903 if (result != VK_SUCCESS) 904 return; /* FIXME */ 905 906 anv_cmd_buffer_add_seen_bbos(primary, ©_list); 907 908 struct anv_batch_bo *first_bbo = 909 list_first_entry(©_list, struct anv_batch_bo, link); 910 struct anv_batch_bo *last_bbo = 911 list_last_entry(©_list, struct anv_batch_bo, link); 912 913 cmd_buffer_chain_to_batch_bo(primary, first_bbo); 914 915 list_splicetail(©_list, &primary->batch_bos); 916 917 anv_batch_bo_continue(last_bbo, &primary->batch, 918 GEN8_MI_BATCH_BUFFER_START_length * 4); 919 break; 920 } 921 default: 922 assert(!"Invalid execution mode"); 923 } 924 925 anv_reloc_list_append(&primary->surface_relocs, &primary->pool->alloc, 926 &secondary->surface_relocs, 0); 927 } 928 929 struct anv_execbuf { 930 struct drm_i915_gem_execbuffer2 execbuf; 931 932 struct drm_i915_gem_exec_object2 * objects; 933 uint32_t bo_count; 934 struct anv_bo ** bos; 935 936 /* Allocated length of the 'objects' and 'bos' arrays */ 937 uint32_t array_length; 938 }; 939 940 static void 941 anv_execbuf_init(struct anv_execbuf *exec) 942 { 943 memset(exec, 0, sizeof(*exec)); 944 } 945 946 static void 947 anv_execbuf_finish(struct anv_execbuf *exec, 948 const VkAllocationCallbacks *alloc) 949 { 950 vk_free(alloc, exec->objects); 951 vk_free(alloc, exec->bos); 952 } 953 954 static VkResult 955 anv_execbuf_add_bo(struct anv_execbuf *exec, 956 struct anv_bo *bo, 957 struct anv_reloc_list *relocs, 958 const VkAllocationCallbacks *alloc) 959 { 960 struct drm_i915_gem_exec_object2 *obj = NULL; 961 962 if (bo->index < exec->bo_count && exec->bos[bo->index] == bo) 963 obj = &exec->objects[bo->index]; 964 965 if (obj == NULL) { 966 /* We've never seen this one before. Add it to the list and assign 967 * an id that we can use later. 968 */ 969 if (exec->bo_count >= exec->array_length) { 970 uint32_t new_len = exec->objects ? exec->array_length * 2 : 64; 971 972 struct drm_i915_gem_exec_object2 *new_objects = 973 vk_alloc(alloc, new_len * sizeof(*new_objects), 974 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND); 975 if (new_objects == NULL) 976 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 977 978 struct anv_bo **new_bos = 979 vk_alloc(alloc, new_len * sizeof(*new_bos), 980 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND); 981 if (new_bos == NULL) { 982 vk_free(alloc, new_objects); 983 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 984 } 985 986 if (exec->objects) { 987 memcpy(new_objects, exec->objects, 988 exec->bo_count * sizeof(*new_objects)); 989 memcpy(new_bos, exec->bos, 990 exec->bo_count * sizeof(*new_bos)); 991 } 992 993 vk_free(alloc, exec->objects); 994 vk_free(alloc, exec->bos); 995 996 exec->objects = new_objects; 997 exec->bos = new_bos; 998 exec->array_length = new_len; 999 } 1000 1001 assert(exec->bo_count < exec->array_length); 1002 1003 bo->index = exec->bo_count++; 1004 obj = &exec->objects[bo->index]; 1005 exec->bos[bo->index] = bo; 1006 1007 obj->handle = bo->gem_handle; 1008 obj->relocation_count = 0; 1009 obj->relocs_ptr = 0; 1010 obj->alignment = 0; 1011 obj->offset = bo->offset; 1012 obj->flags = bo->is_winsys_bo ? EXEC_OBJECT_WRITE : 0; 1013 obj->rsvd1 = 0; 1014 obj->rsvd2 = 0; 1015 } 1016 1017 if (relocs != NULL && obj->relocation_count == 0) { 1018 /* This is the first time we've ever seen a list of relocations for 1019 * this BO. Go ahead and set the relocations and then walk the list 1020 * of relocations and add them all. 1021 */ 1022 obj->relocation_count = relocs->num_relocs; 1023 obj->relocs_ptr = (uintptr_t) relocs->relocs; 1024 1025 for (size_t i = 0; i < relocs->num_relocs; i++) { 1026 /* A quick sanity check on relocations */ 1027 assert(relocs->relocs[i].offset < bo->size); 1028 anv_execbuf_add_bo(exec, relocs->reloc_bos[i], NULL, alloc); 1029 } 1030 } 1031 1032 return VK_SUCCESS; 1033 } 1034 1035 static void 1036 anv_cmd_buffer_process_relocs(struct anv_cmd_buffer *cmd_buffer, 1037 struct anv_reloc_list *list) 1038 { 1039 for (size_t i = 0; i < list->num_relocs; i++) 1040 list->relocs[i].target_handle = list->reloc_bos[i]->index; 1041 } 1042 1043 static void 1044 write_reloc(const struct anv_device *device, void *p, uint64_t v, bool flush) 1045 { 1046 unsigned reloc_size = 0; 1047 if (device->info.gen >= 8) { 1048 /* From the Broadwell PRM Vol. 2a, MI_LOAD_REGISTER_MEM::MemoryAddress: 1049 * 1050 * "This field specifies the address of the memory location where the 1051 * register value specified in the DWord above will read from. The 1052 * address specifies the DWord location of the data. Range = 1053 * GraphicsVirtualAddress[63:2] for a DWord register GraphicsAddress 1054 * [63:48] are ignored by the HW and assumed to be in correct 1055 * canonical form [63:48] == [47]." 1056 */ 1057 const int shift = 63 - 47; 1058 reloc_size = sizeof(uint64_t); 1059 *(uint64_t *)p = (((int64_t)v) << shift) >> shift; 1060 } else { 1061 reloc_size = sizeof(uint32_t); 1062 *(uint32_t *)p = v; 1063 } 1064 1065 if (flush && !device->info.has_llc) 1066 anv_clflush_range(p, reloc_size); 1067 } 1068 1069 static void 1070 adjust_relocations_from_state_pool(struct anv_block_pool *pool, 1071 struct anv_reloc_list *relocs, 1072 uint32_t last_pool_center_bo_offset) 1073 { 1074 assert(last_pool_center_bo_offset <= pool->center_bo_offset); 1075 uint32_t delta = pool->center_bo_offset - last_pool_center_bo_offset; 1076 1077 for (size_t i = 0; i < relocs->num_relocs; i++) { 1078 /* All of the relocations from this block pool to other BO's should 1079 * have been emitted relative to the surface block pool center. We 1080 * need to add the center offset to make them relative to the 1081 * beginning of the actual GEM bo. 1082 */ 1083 relocs->relocs[i].offset += delta; 1084 } 1085 } 1086 1087 static void 1088 adjust_relocations_to_state_pool(struct anv_block_pool *pool, 1089 struct anv_bo *from_bo, 1090 struct anv_reloc_list *relocs, 1091 uint32_t last_pool_center_bo_offset) 1092 { 1093 assert(last_pool_center_bo_offset <= pool->center_bo_offset); 1094 uint32_t delta = pool->center_bo_offset - last_pool_center_bo_offset; 1095 1096 /* When we initially emit relocations into a block pool, we don't 1097 * actually know what the final center_bo_offset will be so we just emit 1098 * it as if center_bo_offset == 0. Now that we know what the center 1099 * offset is, we need to walk the list of relocations and adjust any 1100 * relocations that point to the pool bo with the correct offset. 1101 */ 1102 for (size_t i = 0; i < relocs->num_relocs; i++) { 1103 if (relocs->reloc_bos[i] == &pool->bo) { 1104 /* Adjust the delta value in the relocation to correctly 1105 * correspond to the new delta. Initially, this value may have 1106 * been negative (if treated as unsigned), but we trust in 1107 * uint32_t roll-over to fix that for us at this point. 1108 */ 1109 relocs->relocs[i].delta += delta; 1110 1111 /* Since the delta has changed, we need to update the actual 1112 * relocated value with the new presumed value. This function 1113 * should only be called on batch buffers, so we know it isn't in 1114 * use by the GPU at the moment. 1115 */ 1116 assert(relocs->relocs[i].offset < from_bo->size); 1117 write_reloc(pool->device, from_bo->map + relocs->relocs[i].offset, 1118 relocs->relocs[i].presumed_offset + 1119 relocs->relocs[i].delta, false); 1120 } 1121 } 1122 } 1123 1124 static void 1125 anv_reloc_list_apply(struct anv_device *device, 1126 struct anv_reloc_list *list, 1127 struct anv_bo *bo, 1128 bool always_relocate) 1129 { 1130 for (size_t i = 0; i < list->num_relocs; i++) { 1131 struct anv_bo *target_bo = list->reloc_bos[i]; 1132 if (list->relocs[i].presumed_offset == target_bo->offset && 1133 !always_relocate) 1134 continue; 1135 1136 void *p = bo->map + list->relocs[i].offset; 1137 write_reloc(device, p, target_bo->offset + list->relocs[i].delta, true); 1138 list->relocs[i].presumed_offset = target_bo->offset; 1139 } 1140 } 1141 1142 /** 1143 * This function applies the relocation for a command buffer and writes the 1144 * actual addresses into the buffers as per what we were told by the kernel on 1145 * the previous execbuf2 call. This should be safe to do because, for each 1146 * relocated address, we have two cases: 1147 * 1148 * 1) The target BO is inactive (as seen by the kernel). In this case, it is 1149 * not in use by the GPU so updating the address is 100% ok. It won't be 1150 * in-use by the GPU (from our context) again until the next execbuf2 1151 * happens. If the kernel decides to move it in the next execbuf2, it 1152 * will have to do the relocations itself, but that's ok because it should 1153 * have all of the information needed to do so. 1154 * 1155 * 2) The target BO is active (as seen by the kernel). In this case, it 1156 * hasn't moved since the last execbuffer2 call because GTT shuffling 1157 * *only* happens when the BO is idle. (From our perspective, it only 1158 * happens inside the execbuffer2 ioctl, but the shuffling may be 1159 * triggered by another ioctl, with full-ppgtt this is limited to only 1160 * execbuffer2 ioctls on the same context, or memory pressure.) Since the 1161 * target BO hasn't moved, our anv_bo::offset exactly matches the BO's GTT 1162 * address and the relocated value we are writing into the BO will be the 1163 * same as the value that is already there. 1164 * 1165 * There is also a possibility that the target BO is active but the exact 1166 * RENDER_SURFACE_STATE object we are writing the relocation into isn't in 1167 * use. In this case, the address currently in the RENDER_SURFACE_STATE 1168 * may be stale but it's still safe to write the relocation because that 1169 * particular RENDER_SURFACE_STATE object isn't in-use by the GPU and 1170 * won't be until the next execbuf2 call. 1171 * 1172 * By doing relocations on the CPU, we can tell the kernel that it doesn't 1173 * need to bother. We want to do this because the surface state buffer is 1174 * used by every command buffer so, if the kernel does the relocations, it 1175 * will always be busy and the kernel will always stall. This is also 1176 * probably the fastest mechanism for doing relocations since the kernel would 1177 * have to make a full copy of all the relocations lists. 1178 */ 1179 static bool 1180 relocate_cmd_buffer(struct anv_cmd_buffer *cmd_buffer, 1181 struct anv_execbuf *exec) 1182 { 1183 static int userspace_relocs = -1; 1184 if (userspace_relocs < 0) 1185 userspace_relocs = env_var_as_boolean("ANV_USERSPACE_RELOCS", true); 1186 if (!userspace_relocs) 1187 return false; 1188 1189 /* First, we have to check to see whether or not we can even do the 1190 * relocation. New buffers which have never been submitted to the kernel 1191 * don't have a valid offset so we need to let the kernel do relocations so 1192 * that we can get offsets for them. On future execbuf2 calls, those 1193 * buffers will have offsets and we will be able to skip relocating. 1194 * Invalid offsets are indicated by anv_bo::offset == (uint64_t)-1. 1195 */ 1196 for (uint32_t i = 0; i < exec->bo_count; i++) { 1197 if (exec->bos[i]->offset == (uint64_t)-1) 1198 return false; 1199 } 1200 1201 /* Since surface states are shared between command buffers and we don't 1202 * know what order they will be submitted to the kernel, we don't know 1203 * what address is actually written in the surface state object at any 1204 * given time. The only option is to always relocate them. 1205 */ 1206 anv_reloc_list_apply(cmd_buffer->device, &cmd_buffer->surface_relocs, 1207 &cmd_buffer->device->surface_state_block_pool.bo, 1208 true /* always relocate surface states */); 1209 1210 /* Since we own all of the batch buffers, we know what values are stored 1211 * in the relocated addresses and only have to update them if the offsets 1212 * have changed. 1213 */ 1214 struct anv_batch_bo **bbo; 1215 u_vector_foreach(bbo, &cmd_buffer->seen_bbos) { 1216 anv_reloc_list_apply(cmd_buffer->device, 1217 &(*bbo)->relocs, &(*bbo)->bo, false); 1218 } 1219 1220 for (uint32_t i = 0; i < exec->bo_count; i++) 1221 exec->objects[i].offset = exec->bos[i]->offset; 1222 1223 return true; 1224 } 1225 1226 VkResult 1227 anv_cmd_buffer_execbuf(struct anv_device *device, 1228 struct anv_cmd_buffer *cmd_buffer) 1229 { 1230 struct anv_batch *batch = &cmd_buffer->batch; 1231 struct anv_block_pool *ss_pool = 1232 &cmd_buffer->device->surface_state_block_pool; 1233 1234 struct anv_execbuf execbuf; 1235 anv_execbuf_init(&execbuf); 1236 1237 adjust_relocations_from_state_pool(ss_pool, &cmd_buffer->surface_relocs, 1238 cmd_buffer->last_ss_pool_center); 1239 anv_execbuf_add_bo(&execbuf, &ss_pool->bo, &cmd_buffer->surface_relocs, 1240 &cmd_buffer->pool->alloc); 1241 1242 /* First, we walk over all of the bos we've seen and add them and their 1243 * relocations to the validate list. 1244 */ 1245 struct anv_batch_bo **bbo; 1246 u_vector_foreach(bbo, &cmd_buffer->seen_bbos) { 1247 adjust_relocations_to_state_pool(ss_pool, &(*bbo)->bo, &(*bbo)->relocs, 1248 cmd_buffer->last_ss_pool_center); 1249 1250 anv_execbuf_add_bo(&execbuf, &(*bbo)->bo, &(*bbo)->relocs, 1251 &cmd_buffer->pool->alloc); 1252 } 1253 1254 /* Now that we've adjusted all of the surface state relocations, we need to 1255 * record the surface state pool center so future executions of the command 1256 * buffer can adjust correctly. 1257 */ 1258 cmd_buffer->last_ss_pool_center = ss_pool->center_bo_offset; 1259 1260 struct anv_batch_bo *first_batch_bo = 1261 list_first_entry(&cmd_buffer->batch_bos, struct anv_batch_bo, link); 1262 1263 /* The kernel requires that the last entry in the validation list be the 1264 * batch buffer to execute. We can simply swap the element 1265 * corresponding to the first batch_bo in the chain with the last 1266 * element in the list. 1267 */ 1268 if (first_batch_bo->bo.index != execbuf.bo_count - 1) { 1269 uint32_t idx = first_batch_bo->bo.index; 1270 uint32_t last_idx = execbuf.bo_count - 1; 1271 1272 struct drm_i915_gem_exec_object2 tmp_obj = execbuf.objects[idx]; 1273 assert(execbuf.bos[idx] == &first_batch_bo->bo); 1274 1275 execbuf.objects[idx] = execbuf.objects[last_idx]; 1276 execbuf.bos[idx] = execbuf.bos[last_idx]; 1277 execbuf.bos[idx]->index = idx; 1278 1279 execbuf.objects[last_idx] = tmp_obj; 1280 execbuf.bos[last_idx] = &first_batch_bo->bo; 1281 first_batch_bo->bo.index = last_idx; 1282 } 1283 1284 /* Now we go through and fixup all of the relocation lists to point to 1285 * the correct indices in the object array. We have to do this after we 1286 * reorder the list above as some of the indices may have changed. 1287 */ 1288 u_vector_foreach(bbo, &cmd_buffer->seen_bbos) 1289 anv_cmd_buffer_process_relocs(cmd_buffer, &(*bbo)->relocs); 1290 1291 anv_cmd_buffer_process_relocs(cmd_buffer, &cmd_buffer->surface_relocs); 1292 1293 if (!cmd_buffer->device->info.has_llc) { 1294 __builtin_ia32_mfence(); 1295 u_vector_foreach(bbo, &cmd_buffer->seen_bbos) { 1296 for (uint32_t i = 0; i < (*bbo)->length; i += CACHELINE_SIZE) 1297 __builtin_ia32_clflush((*bbo)->bo.map + i); 1298 } 1299 } 1300 1301 execbuf.execbuf = (struct drm_i915_gem_execbuffer2) { 1302 .buffers_ptr = (uintptr_t) execbuf.objects, 1303 .buffer_count = execbuf.bo_count, 1304 .batch_start_offset = 0, 1305 .batch_len = batch->next - batch->start, 1306 .cliprects_ptr = 0, 1307 .num_cliprects = 0, 1308 .DR1 = 0, 1309 .DR4 = 0, 1310 .flags = I915_EXEC_HANDLE_LUT | I915_EXEC_RENDER | 1311 I915_EXEC_CONSTANTS_REL_GENERAL, 1312 .rsvd1 = cmd_buffer->device->context_id, 1313 .rsvd2 = 0, 1314 }; 1315 1316 if (relocate_cmd_buffer(cmd_buffer, &execbuf)) { 1317 /* If we were able to successfully relocate everything, tell the kernel 1318 * that it can skip doing relocations. The requirement for using 1319 * NO_RELOC is: 1320 * 1321 * 1) The addresses written in the objects must match the corresponding 1322 * reloc.presumed_offset which in turn must match the corresponding 1323 * execobject.offset. 1324 * 1325 * 2) To avoid stalling, execobject.offset should match the current 1326 * address of that object within the active context. 1327 * 1328 * In order to satisfy all of the invariants that make userspace 1329 * relocations to be safe (see relocate_cmd_buffer()), we need to 1330 * further ensure that the addresses we use match those used by the 1331 * kernel for the most recent execbuf2. 1332 * 1333 * The kernel may still choose to do relocations anyway if something has 1334 * moved in the GTT. In this case, the relocation list still needs to be 1335 * valid. All relocations on the batch buffers are already valid and 1336 * kept up-to-date. For surface state relocations, by applying the 1337 * relocations in relocate_cmd_buffer, we ensured that the address in 1338 * the RENDER_SURFACE_STATE matches presumed_offset, so it should be 1339 * safe for the kernel to relocate them as needed. 1340 */ 1341 execbuf.execbuf.flags |= I915_EXEC_NO_RELOC; 1342 } else { 1343 /* In the case where we fall back to doing kernel relocations, we need 1344 * to ensure that the relocation list is valid. All relocations on the 1345 * batch buffers are already valid and kept up-to-date. Since surface 1346 * states are shared between command buffers and we don't know what 1347 * order they will be submitted to the kernel, we don't know what 1348 * address is actually written in the surface state object at any given 1349 * time. The only option is to set a bogus presumed offset and let the 1350 * kernel relocate them. 1351 */ 1352 for (size_t i = 0; i < cmd_buffer->surface_relocs.num_relocs; i++) 1353 cmd_buffer->surface_relocs.relocs[i].presumed_offset = -1; 1354 } 1355 1356 VkResult result = anv_device_execbuf(device, &execbuf.execbuf, execbuf.bos); 1357 1358 anv_execbuf_finish(&execbuf, &cmd_buffer->pool->alloc); 1359 1360 return result; 1361 } 1362