1 /************************************************************************** 2 * 3 * Copyright 2007 Red Hat Inc. 4 * Copyright 2007-2012 Intel Corporation 5 * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA 6 * All Rights Reserved. 7 * 8 * Permission is hereby granted, free of charge, to any person obtaining a 9 * copy of this software and associated documentation files (the 10 * "Software"), to deal in the Software without restriction, including 11 * without limitation the rights to use, copy, modify, merge, publish, 12 * distribute, sub license, and/or sell copies of the Software, and to 13 * permit persons to whom the Software is furnished to do so, subject to 14 * the following conditions: 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 19 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 21 * OTHERWISE, 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 **************************************************************************/ 30 /* 31 * Authors: Thomas Hellstrm <thomas-at-tungstengraphics-dot-com> 32 * Keith Whitwell <keithw-at-tungstengraphics-dot-com> 33 * Eric Anholt <eric (at) anholt.net> 34 * Dave Airlie <airlied (at) linux.ie> 35 */ 36 37 #ifdef HAVE_CONFIG_H 38 #include "config.h" 39 #endif 40 41 #include <xf86drm.h> 42 #include <xf86atomic.h> 43 #include <fcntl.h> 44 #include <stdio.h> 45 #include <stdlib.h> 46 #include <string.h> 47 #include <unistd.h> 48 #include <assert.h> 49 #include <pthread.h> 50 #include <sys/ioctl.h> 51 #include <sys/stat.h> 52 #include <sys/types.h> 53 #include <stdbool.h> 54 55 #include "errno.h" 56 #ifndef ETIME 57 #define ETIME ETIMEDOUT 58 #endif 59 #include "libdrm_macros.h" 60 #include "libdrm_lists.h" 61 #include "intel_bufmgr.h" 62 #include "intel_bufmgr_priv.h" 63 #include "intel_chipset.h" 64 #include "string.h" 65 66 #include "i915_drm.h" 67 #include "uthash.h" 68 69 #ifdef HAVE_VALGRIND 70 #include <valgrind.h> 71 #include <memcheck.h> 72 #define VG(x) x 73 #else 74 #define VG(x) 75 #endif 76 77 #define memclear(s) memset(&s, 0, sizeof(s)) 78 79 #define DBG(...) do { \ 80 if (bufmgr_gem->bufmgr.debug) \ 81 fprintf(stderr, __VA_ARGS__); \ 82 } while (0) 83 84 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0])) 85 #define MAX2(A, B) ((A) > (B) ? (A) : (B)) 86 87 /** 88 * upper_32_bits - return bits 32-63 of a number 89 * @n: the number we're accessing 90 * 91 * A basic shift-right of a 64- or 32-bit quantity. Use this to suppress 92 * the "right shift count >= width of type" warning when that quantity is 93 * 32-bits. 94 */ 95 #define upper_32_bits(n) ((__u32)(((n) >> 16) >> 16)) 96 97 /** 98 * lower_32_bits - return bits 0-31 of a number 99 * @n: the number we're accessing 100 */ 101 #define lower_32_bits(n) ((__u32)(n)) 102 103 typedef struct _drm_intel_bo_gem drm_intel_bo_gem; 104 105 struct drm_intel_gem_bo_bucket { 106 drmMMListHead head; 107 unsigned long size; 108 }; 109 110 typedef struct _drm_intel_bufmgr_gem { 111 drm_intel_bufmgr bufmgr; 112 113 atomic_t refcount; 114 115 int fd; 116 117 int max_relocs; 118 119 pthread_mutex_t lock; 120 121 struct drm_i915_gem_exec_object *exec_objects; 122 struct drm_i915_gem_exec_object2 *exec2_objects; 123 drm_intel_bo **exec_bos; 124 int exec_size; 125 int exec_count; 126 127 /** Array of lists of cached gem objects of power-of-two sizes */ 128 struct drm_intel_gem_bo_bucket cache_bucket[14 * 4]; 129 int num_buckets; 130 time_t time; 131 132 drmMMListHead managers; 133 134 drm_intel_bo_gem *name_table; 135 drm_intel_bo_gem *handle_table; 136 137 drmMMListHead vma_cache; 138 int vma_count, vma_open, vma_max; 139 140 uint64_t gtt_size; 141 int available_fences; 142 int pci_device; 143 int gen; 144 unsigned int has_bsd : 1; 145 unsigned int has_blt : 1; 146 unsigned int has_relaxed_fencing : 1; 147 unsigned int has_llc : 1; 148 unsigned int has_wait_timeout : 1; 149 unsigned int bo_reuse : 1; 150 unsigned int no_exec : 1; 151 unsigned int has_vebox : 1; 152 unsigned int has_exec_async : 1; 153 bool fenced_relocs; 154 155 struct { 156 void *ptr; 157 uint32_t handle; 158 } userptr_active; 159 160 } drm_intel_bufmgr_gem; 161 162 #define DRM_INTEL_RELOC_FENCE (1<<0) 163 164 typedef struct _drm_intel_reloc_target_info { 165 drm_intel_bo *bo; 166 int flags; 167 } drm_intel_reloc_target; 168 169 struct _drm_intel_bo_gem { 170 drm_intel_bo bo; 171 172 atomic_t refcount; 173 uint32_t gem_handle; 174 const char *name; 175 176 /** 177 * Kenel-assigned global name for this object 178 * 179 * List contains both flink named and prime fd'd objects 180 */ 181 unsigned int global_name; 182 183 UT_hash_handle handle_hh; 184 UT_hash_handle name_hh; 185 186 /** 187 * Index of the buffer within the validation list while preparing a 188 * batchbuffer execution. 189 */ 190 int validate_index; 191 192 /** 193 * Current tiling mode 194 */ 195 uint32_t tiling_mode; 196 uint32_t swizzle_mode; 197 unsigned long stride; 198 199 unsigned long kflags; 200 201 time_t free_time; 202 203 /** Array passed to the DRM containing relocation information. */ 204 struct drm_i915_gem_relocation_entry *relocs; 205 /** 206 * Array of info structs corresponding to relocs[i].target_handle etc 207 */ 208 drm_intel_reloc_target *reloc_target_info; 209 /** Number of entries in relocs */ 210 int reloc_count; 211 /** Array of BOs that are referenced by this buffer and will be softpinned */ 212 drm_intel_bo **softpin_target; 213 /** Number softpinned BOs that are referenced by this buffer */ 214 int softpin_target_count; 215 /** Maximum amount of softpinned BOs that are referenced by this buffer */ 216 int softpin_target_size; 217 218 /** Mapped address for the buffer, saved across map/unmap cycles */ 219 void *mem_virtual; 220 /** GTT virtual address for the buffer, saved across map/unmap cycles */ 221 void *gtt_virtual; 222 /** WC CPU address for the buffer, saved across map/unmap cycles */ 223 void *wc_virtual; 224 /** 225 * Virtual address of the buffer allocated by user, used for userptr 226 * objects only. 227 */ 228 void *user_virtual; 229 int map_count; 230 drmMMListHead vma_list; 231 232 /** BO cache list */ 233 drmMMListHead head; 234 235 /** 236 * Boolean of whether this BO and its children have been included in 237 * the current drm_intel_bufmgr_check_aperture_space() total. 238 */ 239 bool included_in_check_aperture; 240 241 /** 242 * Boolean of whether this buffer has been used as a relocation 243 * target and had its size accounted for, and thus can't have any 244 * further relocations added to it. 245 */ 246 bool used_as_reloc_target; 247 248 /** 249 * Boolean of whether we have encountered an error whilst building the relocation tree. 250 */ 251 bool has_error; 252 253 /** 254 * Boolean of whether this buffer can be re-used 255 */ 256 bool reusable; 257 258 /** 259 * Boolean of whether the GPU is definitely not accessing the buffer. 260 * 261 * This is only valid when reusable, since non-reusable 262 * buffers are those that have been shared with other 263 * processes, so we don't know their state. 264 */ 265 bool idle; 266 267 /** 268 * Boolean of whether this buffer was allocated with userptr 269 */ 270 bool is_userptr; 271 272 /** 273 * Boolean of whether this buffer can be placed in the full 48-bit 274 * address range on gen8+. 275 * 276 * By default, buffers will be keep in a 32-bit range, unless this 277 * flag is explicitly set. 278 */ 279 bool use_48b_address_range; 280 281 /** 282 * Whether this buffer is softpinned at offset specified by the user 283 */ 284 bool is_softpin; 285 286 /** 287 * Size in bytes of this buffer and its relocation descendents. 288 * 289 * Used to avoid costly tree walking in 290 * drm_intel_bufmgr_check_aperture in the common case. 291 */ 292 int reloc_tree_size; 293 294 /** 295 * Number of potential fence registers required by this buffer and its 296 * relocations. 297 */ 298 int reloc_tree_fences; 299 300 /** Flags that we may need to do the SW_FINISH ioctl on unmap. */ 301 bool mapped_cpu_write; 302 }; 303 304 static unsigned int 305 drm_intel_gem_estimate_batch_space(drm_intel_bo ** bo_array, int count); 306 307 static unsigned int 308 drm_intel_gem_compute_batch_space(drm_intel_bo ** bo_array, int count); 309 310 static int 311 drm_intel_gem_bo_get_tiling(drm_intel_bo *bo, uint32_t * tiling_mode, 312 uint32_t * swizzle_mode); 313 314 static int 315 drm_intel_gem_bo_set_tiling_internal(drm_intel_bo *bo, 316 uint32_t tiling_mode, 317 uint32_t stride); 318 319 static void drm_intel_gem_bo_unreference_locked_timed(drm_intel_bo *bo, 320 time_t time); 321 322 static void drm_intel_gem_bo_unreference(drm_intel_bo *bo); 323 324 static void drm_intel_gem_bo_free(drm_intel_bo *bo); 325 326 static inline drm_intel_bo_gem *to_bo_gem(drm_intel_bo *bo) 327 { 328 return (drm_intel_bo_gem *)bo; 329 } 330 331 static unsigned long 332 drm_intel_gem_bo_tile_size(drm_intel_bufmgr_gem *bufmgr_gem, unsigned long size, 333 uint32_t *tiling_mode) 334 { 335 unsigned long min_size, max_size; 336 unsigned long i; 337 338 if (*tiling_mode == I915_TILING_NONE) 339 return size; 340 341 /* 965+ just need multiples of page size for tiling */ 342 if (bufmgr_gem->gen >= 4) 343 return ROUND_UP_TO(size, 4096); 344 345 /* Older chips need powers of two, of at least 512k or 1M */ 346 if (bufmgr_gem->gen == 3) { 347 min_size = 1024*1024; 348 max_size = 128*1024*1024; 349 } else { 350 min_size = 512*1024; 351 max_size = 64*1024*1024; 352 } 353 354 if (size > max_size) { 355 *tiling_mode = I915_TILING_NONE; 356 return size; 357 } 358 359 /* Do we need to allocate every page for the fence? */ 360 if (bufmgr_gem->has_relaxed_fencing) 361 return ROUND_UP_TO(size, 4096); 362 363 for (i = min_size; i < size; i <<= 1) 364 ; 365 366 return i; 367 } 368 369 /* 370 * Round a given pitch up to the minimum required for X tiling on a 371 * given chip. We use 512 as the minimum to allow for a later tiling 372 * change. 373 */ 374 static unsigned long 375 drm_intel_gem_bo_tile_pitch(drm_intel_bufmgr_gem *bufmgr_gem, 376 unsigned long pitch, uint32_t *tiling_mode) 377 { 378 unsigned long tile_width; 379 unsigned long i; 380 381 /* If untiled, then just align it so that we can do rendering 382 * to it with the 3D engine. 383 */ 384 if (*tiling_mode == I915_TILING_NONE) 385 return ALIGN(pitch, 64); 386 387 if (*tiling_mode == I915_TILING_X 388 || (IS_915(bufmgr_gem->pci_device) 389 && *tiling_mode == I915_TILING_Y)) 390 tile_width = 512; 391 else 392 tile_width = 128; 393 394 /* 965 is flexible */ 395 if (bufmgr_gem->gen >= 4) 396 return ROUND_UP_TO(pitch, tile_width); 397 398 /* The older hardware has a maximum pitch of 8192 with tiled 399 * surfaces, so fallback to untiled if it's too large. 400 */ 401 if (pitch > 8192) { 402 *tiling_mode = I915_TILING_NONE; 403 return ALIGN(pitch, 64); 404 } 405 406 /* Pre-965 needs power of two tile width */ 407 for (i = tile_width; i < pitch; i <<= 1) 408 ; 409 410 return i; 411 } 412 413 static struct drm_intel_gem_bo_bucket * 414 drm_intel_gem_bo_bucket_for_size(drm_intel_bufmgr_gem *bufmgr_gem, 415 unsigned long size) 416 { 417 int i; 418 419 for (i = 0; i < bufmgr_gem->num_buckets; i++) { 420 struct drm_intel_gem_bo_bucket *bucket = 421 &bufmgr_gem->cache_bucket[i]; 422 if (bucket->size >= size) { 423 return bucket; 424 } 425 } 426 427 return NULL; 428 } 429 430 static void 431 drm_intel_gem_dump_validation_list(drm_intel_bufmgr_gem *bufmgr_gem) 432 { 433 int i, j; 434 435 for (i = 0; i < bufmgr_gem->exec_count; i++) { 436 drm_intel_bo *bo = bufmgr_gem->exec_bos[i]; 437 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 438 439 if (bo_gem->relocs == NULL && bo_gem->softpin_target == NULL) { 440 DBG("%2d: %d %s(%s)\n", i, bo_gem->gem_handle, 441 bo_gem->is_softpin ? "*" : "", 442 bo_gem->name); 443 continue; 444 } 445 446 for (j = 0; j < bo_gem->reloc_count; j++) { 447 drm_intel_bo *target_bo = bo_gem->reloc_target_info[j].bo; 448 drm_intel_bo_gem *target_gem = 449 (drm_intel_bo_gem *) target_bo; 450 451 DBG("%2d: %d %s(%s)@0x%08x %08x -> " 452 "%d (%s)@0x%08x %08x + 0x%08x\n", 453 i, 454 bo_gem->gem_handle, 455 bo_gem->is_softpin ? "*" : "", 456 bo_gem->name, 457 upper_32_bits(bo_gem->relocs[j].offset), 458 lower_32_bits(bo_gem->relocs[j].offset), 459 target_gem->gem_handle, 460 target_gem->name, 461 upper_32_bits(target_bo->offset64), 462 lower_32_bits(target_bo->offset64), 463 bo_gem->relocs[j].delta); 464 } 465 466 for (j = 0; j < bo_gem->softpin_target_count; j++) { 467 drm_intel_bo *target_bo = bo_gem->softpin_target[j]; 468 drm_intel_bo_gem *target_gem = 469 (drm_intel_bo_gem *) target_bo; 470 DBG("%2d: %d %s(%s) -> " 471 "%d *(%s)@0x%08x %08x\n", 472 i, 473 bo_gem->gem_handle, 474 bo_gem->is_softpin ? "*" : "", 475 bo_gem->name, 476 target_gem->gem_handle, 477 target_gem->name, 478 upper_32_bits(target_bo->offset64), 479 lower_32_bits(target_bo->offset64)); 480 } 481 } 482 } 483 484 static inline void 485 drm_intel_gem_bo_reference(drm_intel_bo *bo) 486 { 487 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 488 489 atomic_inc(&bo_gem->refcount); 490 } 491 492 /** 493 * Adds the given buffer to the list of buffers to be validated (moved into the 494 * appropriate memory type) with the next batch submission. 495 * 496 * If a buffer is validated multiple times in a batch submission, it ends up 497 * with the intersection of the memory type flags and the union of the 498 * access flags. 499 */ 500 static void 501 drm_intel_add_validate_buffer(drm_intel_bo *bo) 502 { 503 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 504 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 505 int index; 506 507 if (bo_gem->validate_index != -1) 508 return; 509 510 /* Extend the array of validation entries as necessary. */ 511 if (bufmgr_gem->exec_count == bufmgr_gem->exec_size) { 512 int new_size = bufmgr_gem->exec_size * 2; 513 514 if (new_size == 0) 515 new_size = 5; 516 517 bufmgr_gem->exec_objects = 518 realloc(bufmgr_gem->exec_objects, 519 sizeof(*bufmgr_gem->exec_objects) * new_size); 520 bufmgr_gem->exec_bos = 521 realloc(bufmgr_gem->exec_bos, 522 sizeof(*bufmgr_gem->exec_bos) * new_size); 523 bufmgr_gem->exec_size = new_size; 524 } 525 526 index = bufmgr_gem->exec_count; 527 bo_gem->validate_index = index; 528 /* Fill in array entry */ 529 bufmgr_gem->exec_objects[index].handle = bo_gem->gem_handle; 530 bufmgr_gem->exec_objects[index].relocation_count = bo_gem->reloc_count; 531 bufmgr_gem->exec_objects[index].relocs_ptr = (uintptr_t) bo_gem->relocs; 532 bufmgr_gem->exec_objects[index].alignment = bo->align; 533 bufmgr_gem->exec_objects[index].offset = 0; 534 bufmgr_gem->exec_bos[index] = bo; 535 bufmgr_gem->exec_count++; 536 } 537 538 static void 539 drm_intel_add_validate_buffer2(drm_intel_bo *bo, int need_fence) 540 { 541 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr; 542 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo; 543 int index; 544 int flags = 0; 545 546 if (need_fence) 547 flags |= EXEC_OBJECT_NEEDS_FENCE; 548 if (bo_gem->use_48b_address_range) 549 flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS; 550 if (bo_gem->is_softpin) 551 flags |= EXEC_OBJECT_PINNED; 552 553 if (bo_gem->validate_index != -1) { 554 bufmgr_gem->exec2_objects[bo_gem->validate_index].flags |= flags; 555 return; 556 } 557 558 /* Extend the array of validation entries as necessary. */ 559 if (bufmgr_gem->exec_count == bufmgr_gem->exec_size) { 560 int new_size = bufmgr_gem->exec_size * 2; 561 562 if (new_size == 0) 563 new_size = 5; 564 565 bufmgr_gem->exec2_objects = 566 realloc(bufmgr_gem->exec2_objects, 567 sizeof(*bufmgr_gem->exec2_objects) * new_size); 568 bufmgr_gem->exec_bos = 569 realloc(bufmgr_gem->exec_bos, 570 sizeof(*bufmgr_gem->exec_bos) * new_size); 571 bufmgr_gem->exec_size = new_size; 572 } 573 574 index = bufmgr_gem->exec_count; 575 bo_gem->validate_index = index; 576 /* Fill in array entry */ 577 bufmgr_gem->exec2_objects[index].handle = bo_gem->gem_handle; 578 bufmgr_gem->exec2_objects[index].relocation_count = bo_gem->reloc_count; 579 bufmgr_gem->exec2_objects[index].relocs_ptr = (uintptr_t)bo_gem->relocs; 580 bufmgr_gem->exec2_objects[index].alignment = bo->align; 581 bufmgr_gem->exec2_objects[index].offset = bo->offset64; 582 bufmgr_gem->exec2_objects[index].flags = flags | bo_gem->kflags; 583 bufmgr_gem->exec2_objects[index].rsvd1 = 0; 584 bufmgr_gem->exec2_objects[index].rsvd2 = 0; 585 bufmgr_gem->exec_bos[index] = bo; 586 bufmgr_gem->exec_count++; 587 } 588 589 #define RELOC_BUF_SIZE(x) ((I915_RELOC_HEADER + x * I915_RELOC0_STRIDE) * \ 590 sizeof(uint32_t)) 591 592 static void 593 drm_intel_bo_gem_set_in_aperture_size(drm_intel_bufmgr_gem *bufmgr_gem, 594 drm_intel_bo_gem *bo_gem, 595 unsigned int alignment) 596 { 597 unsigned int size; 598 599 assert(!bo_gem->used_as_reloc_target); 600 601 /* The older chipsets are far-less flexible in terms of tiling, 602 * and require tiled buffer to be size aligned in the aperture. 603 * This means that in the worst possible case we will need a hole 604 * twice as large as the object in order for it to fit into the 605 * aperture. Optimal packing is for wimps. 606 */ 607 size = bo_gem->bo.size; 608 if (bufmgr_gem->gen < 4 && bo_gem->tiling_mode != I915_TILING_NONE) { 609 unsigned int min_size; 610 611 if (bufmgr_gem->has_relaxed_fencing) { 612 if (bufmgr_gem->gen == 3) 613 min_size = 1024*1024; 614 else 615 min_size = 512*1024; 616 617 while (min_size < size) 618 min_size *= 2; 619 } else 620 min_size = size; 621 622 /* Account for worst-case alignment. */ 623 alignment = MAX2(alignment, min_size); 624 } 625 626 bo_gem->reloc_tree_size = size + alignment; 627 } 628 629 static int 630 drm_intel_setup_reloc_list(drm_intel_bo *bo) 631 { 632 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 633 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 634 unsigned int max_relocs = bufmgr_gem->max_relocs; 635 636 if (bo->size / 4 < max_relocs) 637 max_relocs = bo->size / 4; 638 639 bo_gem->relocs = malloc(max_relocs * 640 sizeof(struct drm_i915_gem_relocation_entry)); 641 bo_gem->reloc_target_info = malloc(max_relocs * 642 sizeof(drm_intel_reloc_target)); 643 if (bo_gem->relocs == NULL || bo_gem->reloc_target_info == NULL) { 644 bo_gem->has_error = true; 645 646 free (bo_gem->relocs); 647 bo_gem->relocs = NULL; 648 649 free (bo_gem->reloc_target_info); 650 bo_gem->reloc_target_info = NULL; 651 652 return 1; 653 } 654 655 return 0; 656 } 657 658 static int 659 drm_intel_gem_bo_busy(drm_intel_bo *bo) 660 { 661 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 662 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 663 struct drm_i915_gem_busy busy; 664 int ret; 665 666 if (bo_gem->reusable && bo_gem->idle) 667 return false; 668 669 memclear(busy); 670 busy.handle = bo_gem->gem_handle; 671 672 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_BUSY, &busy); 673 if (ret == 0) { 674 bo_gem->idle = !busy.busy; 675 return busy.busy; 676 } else { 677 return false; 678 } 679 return (ret == 0 && busy.busy); 680 } 681 682 static int 683 drm_intel_gem_bo_madvise_internal(drm_intel_bufmgr_gem *bufmgr_gem, 684 drm_intel_bo_gem *bo_gem, int state) 685 { 686 struct drm_i915_gem_madvise madv; 687 688 memclear(madv); 689 madv.handle = bo_gem->gem_handle; 690 madv.madv = state; 691 madv.retained = 1; 692 drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv); 693 694 return madv.retained; 695 } 696 697 static int 698 drm_intel_gem_bo_madvise(drm_intel_bo *bo, int madv) 699 { 700 return drm_intel_gem_bo_madvise_internal 701 ((drm_intel_bufmgr_gem *) bo->bufmgr, 702 (drm_intel_bo_gem *) bo, 703 madv); 704 } 705 706 /* drop the oldest entries that have been purged by the kernel */ 707 static void 708 drm_intel_gem_bo_cache_purge_bucket(drm_intel_bufmgr_gem *bufmgr_gem, 709 struct drm_intel_gem_bo_bucket *bucket) 710 { 711 while (!DRMLISTEMPTY(&bucket->head)) { 712 drm_intel_bo_gem *bo_gem; 713 714 bo_gem = DRMLISTENTRY(drm_intel_bo_gem, 715 bucket->head.next, head); 716 if (drm_intel_gem_bo_madvise_internal 717 (bufmgr_gem, bo_gem, I915_MADV_DONTNEED)) 718 break; 719 720 DRMLISTDEL(&bo_gem->head); 721 drm_intel_gem_bo_free(&bo_gem->bo); 722 } 723 } 724 725 static drm_intel_bo * 726 drm_intel_gem_bo_alloc_internal(drm_intel_bufmgr *bufmgr, 727 const char *name, 728 unsigned long size, 729 unsigned long flags, 730 uint32_t tiling_mode, 731 unsigned long stride, 732 unsigned int alignment) 733 { 734 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr; 735 drm_intel_bo_gem *bo_gem; 736 unsigned int page_size = getpagesize(); 737 int ret; 738 struct drm_intel_gem_bo_bucket *bucket; 739 bool alloc_from_cache; 740 unsigned long bo_size; 741 bool for_render = false; 742 743 if (flags & BO_ALLOC_FOR_RENDER) 744 for_render = true; 745 746 /* Round the allocated size up to a power of two number of pages. */ 747 bucket = drm_intel_gem_bo_bucket_for_size(bufmgr_gem, size); 748 749 /* If we don't have caching at this size, don't actually round the 750 * allocation up. 751 */ 752 if (bucket == NULL) { 753 bo_size = size; 754 if (bo_size < page_size) 755 bo_size = page_size; 756 } else { 757 bo_size = bucket->size; 758 } 759 760 pthread_mutex_lock(&bufmgr_gem->lock); 761 /* Get a buffer out of the cache if available */ 762 retry: 763 alloc_from_cache = false; 764 if (bucket != NULL && !DRMLISTEMPTY(&bucket->head)) { 765 if (for_render) { 766 /* Allocate new render-target BOs from the tail (MRU) 767 * of the list, as it will likely be hot in the GPU 768 * cache and in the aperture for us. 769 */ 770 bo_gem = DRMLISTENTRY(drm_intel_bo_gem, 771 bucket->head.prev, head); 772 DRMLISTDEL(&bo_gem->head); 773 alloc_from_cache = true; 774 bo_gem->bo.align = alignment; 775 } else { 776 assert(alignment == 0); 777 /* For non-render-target BOs (where we're probably 778 * going to map it first thing in order to fill it 779 * with data), check if the last BO in the cache is 780 * unbusy, and only reuse in that case. Otherwise, 781 * allocating a new buffer is probably faster than 782 * waiting for the GPU to finish. 783 */ 784 bo_gem = DRMLISTENTRY(drm_intel_bo_gem, 785 bucket->head.next, head); 786 if (!drm_intel_gem_bo_busy(&bo_gem->bo)) { 787 alloc_from_cache = true; 788 DRMLISTDEL(&bo_gem->head); 789 } 790 } 791 792 if (alloc_from_cache) { 793 if (!drm_intel_gem_bo_madvise_internal 794 (bufmgr_gem, bo_gem, I915_MADV_WILLNEED)) { 795 drm_intel_gem_bo_free(&bo_gem->bo); 796 drm_intel_gem_bo_cache_purge_bucket(bufmgr_gem, 797 bucket); 798 goto retry; 799 } 800 801 if (drm_intel_gem_bo_set_tiling_internal(&bo_gem->bo, 802 tiling_mode, 803 stride)) { 804 drm_intel_gem_bo_free(&bo_gem->bo); 805 goto retry; 806 } 807 } 808 } 809 810 if (!alloc_from_cache) { 811 struct drm_i915_gem_create create; 812 813 bo_gem = calloc(1, sizeof(*bo_gem)); 814 if (!bo_gem) 815 goto err; 816 817 /* drm_intel_gem_bo_free calls DRMLISTDEL() for an uninitialized 818 list (vma_list), so better set the list head here */ 819 DRMINITLISTHEAD(&bo_gem->vma_list); 820 821 bo_gem->bo.size = bo_size; 822 823 memclear(create); 824 create.size = bo_size; 825 826 ret = drmIoctl(bufmgr_gem->fd, 827 DRM_IOCTL_I915_GEM_CREATE, 828 &create); 829 if (ret != 0) { 830 free(bo_gem); 831 goto err; 832 } 833 834 bo_gem->gem_handle = create.handle; 835 bo_gem->bo.handle = bo_gem->gem_handle; 836 bo_gem->bo.bufmgr = bufmgr; 837 bo_gem->bo.align = alignment; 838 839 bo_gem->tiling_mode = I915_TILING_NONE; 840 bo_gem->swizzle_mode = I915_BIT_6_SWIZZLE_NONE; 841 bo_gem->stride = 0; 842 843 if (drm_intel_gem_bo_set_tiling_internal(&bo_gem->bo, 844 tiling_mode, 845 stride)) 846 goto err_free; 847 848 HASH_ADD(handle_hh, bufmgr_gem->handle_table, 849 gem_handle, sizeof(bo_gem->gem_handle), 850 bo_gem); 851 } 852 853 bo_gem->name = name; 854 atomic_set(&bo_gem->refcount, 1); 855 bo_gem->validate_index = -1; 856 bo_gem->reloc_tree_fences = 0; 857 bo_gem->used_as_reloc_target = false; 858 bo_gem->has_error = false; 859 bo_gem->reusable = true; 860 bo_gem->use_48b_address_range = false; 861 862 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem, alignment); 863 pthread_mutex_unlock(&bufmgr_gem->lock); 864 865 DBG("bo_create: buf %d (%s) %ldb\n", 866 bo_gem->gem_handle, bo_gem->name, size); 867 868 return &bo_gem->bo; 869 870 err_free: 871 drm_intel_gem_bo_free(&bo_gem->bo); 872 err: 873 pthread_mutex_unlock(&bufmgr_gem->lock); 874 return NULL; 875 } 876 877 static drm_intel_bo * 878 drm_intel_gem_bo_alloc_for_render(drm_intel_bufmgr *bufmgr, 879 const char *name, 880 unsigned long size, 881 unsigned int alignment) 882 { 883 return drm_intel_gem_bo_alloc_internal(bufmgr, name, size, 884 BO_ALLOC_FOR_RENDER, 885 I915_TILING_NONE, 0, 886 alignment); 887 } 888 889 static drm_intel_bo * 890 drm_intel_gem_bo_alloc(drm_intel_bufmgr *bufmgr, 891 const char *name, 892 unsigned long size, 893 unsigned int alignment) 894 { 895 return drm_intel_gem_bo_alloc_internal(bufmgr, name, size, 0, 896 I915_TILING_NONE, 0, 0); 897 } 898 899 static drm_intel_bo * 900 drm_intel_gem_bo_alloc_tiled(drm_intel_bufmgr *bufmgr, const char *name, 901 int x, int y, int cpp, uint32_t *tiling_mode, 902 unsigned long *pitch, unsigned long flags) 903 { 904 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr; 905 unsigned long size, stride; 906 uint32_t tiling; 907 908 do { 909 unsigned long aligned_y, height_alignment; 910 911 tiling = *tiling_mode; 912 913 /* If we're tiled, our allocations are in 8 or 32-row blocks, 914 * so failure to align our height means that we won't allocate 915 * enough pages. 916 * 917 * If we're untiled, we still have to align to 2 rows high 918 * because the data port accesses 2x2 blocks even if the 919 * bottom row isn't to be rendered, so failure to align means 920 * we could walk off the end of the GTT and fault. This is 921 * documented on 965, and may be the case on older chipsets 922 * too so we try to be careful. 923 */ 924 aligned_y = y; 925 height_alignment = 2; 926 927 if ((bufmgr_gem->gen == 2) && tiling != I915_TILING_NONE) 928 height_alignment = 16; 929 else if (tiling == I915_TILING_X 930 || (IS_915(bufmgr_gem->pci_device) 931 && tiling == I915_TILING_Y)) 932 height_alignment = 8; 933 else if (tiling == I915_TILING_Y) 934 height_alignment = 32; 935 aligned_y = ALIGN(y, height_alignment); 936 937 stride = x * cpp; 938 stride = drm_intel_gem_bo_tile_pitch(bufmgr_gem, stride, tiling_mode); 939 size = stride * aligned_y; 940 size = drm_intel_gem_bo_tile_size(bufmgr_gem, size, tiling_mode); 941 } while (*tiling_mode != tiling); 942 *pitch = stride; 943 944 if (tiling == I915_TILING_NONE) 945 stride = 0; 946 947 return drm_intel_gem_bo_alloc_internal(bufmgr, name, size, flags, 948 tiling, stride, 0); 949 } 950 951 static drm_intel_bo * 952 drm_intel_gem_bo_alloc_userptr(drm_intel_bufmgr *bufmgr, 953 const char *name, 954 void *addr, 955 uint32_t tiling_mode, 956 uint32_t stride, 957 unsigned long size, 958 unsigned long flags) 959 { 960 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr; 961 drm_intel_bo_gem *bo_gem; 962 int ret; 963 struct drm_i915_gem_userptr userptr; 964 965 /* Tiling with userptr surfaces is not supported 966 * on all hardware so refuse it for time being. 967 */ 968 if (tiling_mode != I915_TILING_NONE) 969 return NULL; 970 971 bo_gem = calloc(1, sizeof(*bo_gem)); 972 if (!bo_gem) 973 return NULL; 974 975 atomic_set(&bo_gem->refcount, 1); 976 DRMINITLISTHEAD(&bo_gem->vma_list); 977 978 bo_gem->bo.size = size; 979 980 memclear(userptr); 981 userptr.user_ptr = (__u64)((unsigned long)addr); 982 userptr.user_size = size; 983 userptr.flags = flags; 984 985 ret = drmIoctl(bufmgr_gem->fd, 986 DRM_IOCTL_I915_GEM_USERPTR, 987 &userptr); 988 if (ret != 0) { 989 DBG("bo_create_userptr: " 990 "ioctl failed with user ptr %p size 0x%lx, " 991 "user flags 0x%lx\n", addr, size, flags); 992 free(bo_gem); 993 return NULL; 994 } 995 996 pthread_mutex_lock(&bufmgr_gem->lock); 997 998 bo_gem->gem_handle = userptr.handle; 999 bo_gem->bo.handle = bo_gem->gem_handle; 1000 bo_gem->bo.bufmgr = bufmgr; 1001 bo_gem->is_userptr = true; 1002 bo_gem->bo.virtual = addr; 1003 /* Save the address provided by user */ 1004 bo_gem->user_virtual = addr; 1005 bo_gem->tiling_mode = I915_TILING_NONE; 1006 bo_gem->swizzle_mode = I915_BIT_6_SWIZZLE_NONE; 1007 bo_gem->stride = 0; 1008 1009 HASH_ADD(handle_hh, bufmgr_gem->handle_table, 1010 gem_handle, sizeof(bo_gem->gem_handle), 1011 bo_gem); 1012 1013 bo_gem->name = name; 1014 bo_gem->validate_index = -1; 1015 bo_gem->reloc_tree_fences = 0; 1016 bo_gem->used_as_reloc_target = false; 1017 bo_gem->has_error = false; 1018 bo_gem->reusable = false; 1019 bo_gem->use_48b_address_range = false; 1020 1021 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem, 0); 1022 pthread_mutex_unlock(&bufmgr_gem->lock); 1023 1024 DBG("bo_create_userptr: " 1025 "ptr %p buf %d (%s) size %ldb, stride 0x%x, tile mode %d\n", 1026 addr, bo_gem->gem_handle, bo_gem->name, 1027 size, stride, tiling_mode); 1028 1029 return &bo_gem->bo; 1030 } 1031 1032 static bool 1033 has_userptr(drm_intel_bufmgr_gem *bufmgr_gem) 1034 { 1035 int ret; 1036 void *ptr; 1037 long pgsz; 1038 struct drm_i915_gem_userptr userptr; 1039 1040 pgsz = sysconf(_SC_PAGESIZE); 1041 assert(pgsz > 0); 1042 1043 ret = posix_memalign(&ptr, pgsz, pgsz); 1044 if (ret) { 1045 DBG("Failed to get a page (%ld) for userptr detection!\n", 1046 pgsz); 1047 return false; 1048 } 1049 1050 memclear(userptr); 1051 userptr.user_ptr = (__u64)(unsigned long)ptr; 1052 userptr.user_size = pgsz; 1053 1054 retry: 1055 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_USERPTR, &userptr); 1056 if (ret) { 1057 if (errno == ENODEV && userptr.flags == 0) { 1058 userptr.flags = I915_USERPTR_UNSYNCHRONIZED; 1059 goto retry; 1060 } 1061 free(ptr); 1062 return false; 1063 } 1064 1065 /* We don't release the userptr bo here as we want to keep the 1066 * kernel mm tracking alive for our lifetime. The first time we 1067 * create a userptr object the kernel has to install a mmu_notifer 1068 * which is a heavyweight operation (e.g. it requires taking all 1069 * mm_locks and stop_machine()). 1070 */ 1071 1072 bufmgr_gem->userptr_active.ptr = ptr; 1073 bufmgr_gem->userptr_active.handle = userptr.handle; 1074 1075 return true; 1076 } 1077 1078 static drm_intel_bo * 1079 check_bo_alloc_userptr(drm_intel_bufmgr *bufmgr, 1080 const char *name, 1081 void *addr, 1082 uint32_t tiling_mode, 1083 uint32_t stride, 1084 unsigned long size, 1085 unsigned long flags) 1086 { 1087 if (has_userptr((drm_intel_bufmgr_gem *)bufmgr)) 1088 bufmgr->bo_alloc_userptr = drm_intel_gem_bo_alloc_userptr; 1089 else 1090 bufmgr->bo_alloc_userptr = NULL; 1091 1092 return drm_intel_bo_alloc_userptr(bufmgr, name, addr, 1093 tiling_mode, stride, size, flags); 1094 } 1095 1096 /** 1097 * Returns a drm_intel_bo wrapping the given buffer object handle. 1098 * 1099 * This can be used when one application needs to pass a buffer object 1100 * to another. 1101 */ 1102 drm_intel_bo * 1103 drm_intel_bo_gem_create_from_name(drm_intel_bufmgr *bufmgr, 1104 const char *name, 1105 unsigned int handle) 1106 { 1107 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr; 1108 drm_intel_bo_gem *bo_gem; 1109 int ret; 1110 struct drm_gem_open open_arg; 1111 struct drm_i915_gem_get_tiling get_tiling; 1112 1113 /* At the moment most applications only have a few named bo. 1114 * For instance, in a DRI client only the render buffers passed 1115 * between X and the client are named. And since X returns the 1116 * alternating names for the front/back buffer a linear search 1117 * provides a sufficiently fast match. 1118 */ 1119 pthread_mutex_lock(&bufmgr_gem->lock); 1120 HASH_FIND(name_hh, bufmgr_gem->name_table, 1121 &handle, sizeof(handle), bo_gem); 1122 if (bo_gem) { 1123 drm_intel_gem_bo_reference(&bo_gem->bo); 1124 goto out; 1125 } 1126 1127 memclear(open_arg); 1128 open_arg.name = handle; 1129 ret = drmIoctl(bufmgr_gem->fd, 1130 DRM_IOCTL_GEM_OPEN, 1131 &open_arg); 1132 if (ret != 0) { 1133 DBG("Couldn't reference %s handle 0x%08x: %s\n", 1134 name, handle, strerror(errno)); 1135 bo_gem = NULL; 1136 goto out; 1137 } 1138 /* Now see if someone has used a prime handle to get this 1139 * object from the kernel before by looking through the list 1140 * again for a matching gem_handle 1141 */ 1142 HASH_FIND(handle_hh, bufmgr_gem->handle_table, 1143 &open_arg.handle, sizeof(open_arg.handle), bo_gem); 1144 if (bo_gem) { 1145 drm_intel_gem_bo_reference(&bo_gem->bo); 1146 goto out; 1147 } 1148 1149 bo_gem = calloc(1, sizeof(*bo_gem)); 1150 if (!bo_gem) 1151 goto out; 1152 1153 atomic_set(&bo_gem->refcount, 1); 1154 DRMINITLISTHEAD(&bo_gem->vma_list); 1155 1156 bo_gem->bo.size = open_arg.size; 1157 bo_gem->bo.offset = 0; 1158 bo_gem->bo.offset64 = 0; 1159 bo_gem->bo.virtual = NULL; 1160 bo_gem->bo.bufmgr = bufmgr; 1161 bo_gem->name = name; 1162 bo_gem->validate_index = -1; 1163 bo_gem->gem_handle = open_arg.handle; 1164 bo_gem->bo.handle = open_arg.handle; 1165 bo_gem->global_name = handle; 1166 bo_gem->reusable = false; 1167 bo_gem->use_48b_address_range = false; 1168 1169 HASH_ADD(handle_hh, bufmgr_gem->handle_table, 1170 gem_handle, sizeof(bo_gem->gem_handle), bo_gem); 1171 HASH_ADD(name_hh, bufmgr_gem->name_table, 1172 global_name, sizeof(bo_gem->global_name), bo_gem); 1173 1174 memclear(get_tiling); 1175 get_tiling.handle = bo_gem->gem_handle; 1176 ret = drmIoctl(bufmgr_gem->fd, 1177 DRM_IOCTL_I915_GEM_GET_TILING, 1178 &get_tiling); 1179 if (ret != 0) 1180 goto err_unref; 1181 1182 bo_gem->tiling_mode = get_tiling.tiling_mode; 1183 bo_gem->swizzle_mode = get_tiling.swizzle_mode; 1184 /* XXX stride is unknown */ 1185 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem, 0); 1186 DBG("bo_create_from_handle: %d (%s)\n", handle, bo_gem->name); 1187 1188 out: 1189 pthread_mutex_unlock(&bufmgr_gem->lock); 1190 return &bo_gem->bo; 1191 1192 err_unref: 1193 drm_intel_gem_bo_free(&bo_gem->bo); 1194 pthread_mutex_unlock(&bufmgr_gem->lock); 1195 return NULL; 1196 } 1197 1198 static void 1199 drm_intel_gem_bo_free(drm_intel_bo *bo) 1200 { 1201 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1202 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1203 struct drm_gem_close close; 1204 int ret; 1205 1206 DRMLISTDEL(&bo_gem->vma_list); 1207 if (bo_gem->mem_virtual) { 1208 VG(VALGRIND_FREELIKE_BLOCK(bo_gem->mem_virtual, 0)); 1209 drm_munmap(bo_gem->mem_virtual, bo_gem->bo.size); 1210 bufmgr_gem->vma_count--; 1211 } 1212 if (bo_gem->wc_virtual) { 1213 VG(VALGRIND_FREELIKE_BLOCK(bo_gem->wc_virtual, 0)); 1214 drm_munmap(bo_gem->wc_virtual, bo_gem->bo.size); 1215 bufmgr_gem->vma_count--; 1216 } 1217 if (bo_gem->gtt_virtual) { 1218 drm_munmap(bo_gem->gtt_virtual, bo_gem->bo.size); 1219 bufmgr_gem->vma_count--; 1220 } 1221 1222 if (bo_gem->global_name) 1223 HASH_DELETE(name_hh, bufmgr_gem->name_table, bo_gem); 1224 HASH_DELETE(handle_hh, bufmgr_gem->handle_table, bo_gem); 1225 1226 /* Close this object */ 1227 memclear(close); 1228 close.handle = bo_gem->gem_handle; 1229 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_GEM_CLOSE, &close); 1230 if (ret != 0) { 1231 DBG("DRM_IOCTL_GEM_CLOSE %d failed (%s): %s\n", 1232 bo_gem->gem_handle, bo_gem->name, strerror(errno)); 1233 } 1234 free(bo); 1235 } 1236 1237 static void 1238 drm_intel_gem_bo_mark_mmaps_incoherent(drm_intel_bo *bo) 1239 { 1240 #if HAVE_VALGRIND 1241 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1242 1243 if (bo_gem->mem_virtual) 1244 VALGRIND_MAKE_MEM_NOACCESS(bo_gem->mem_virtual, bo->size); 1245 1246 if (bo_gem->wc_virtual) 1247 VALGRIND_MAKE_MEM_NOACCESS(bo_gem->wc_virtual, bo->size); 1248 1249 if (bo_gem->gtt_virtual) 1250 VALGRIND_MAKE_MEM_NOACCESS(bo_gem->gtt_virtual, bo->size); 1251 #endif 1252 } 1253 1254 /** Frees all cached buffers significantly older than @time. */ 1255 static void 1256 drm_intel_gem_cleanup_bo_cache(drm_intel_bufmgr_gem *bufmgr_gem, time_t time) 1257 { 1258 int i; 1259 1260 if (bufmgr_gem->time == time) 1261 return; 1262 1263 for (i = 0; i < bufmgr_gem->num_buckets; i++) { 1264 struct drm_intel_gem_bo_bucket *bucket = 1265 &bufmgr_gem->cache_bucket[i]; 1266 1267 while (!DRMLISTEMPTY(&bucket->head)) { 1268 drm_intel_bo_gem *bo_gem; 1269 1270 bo_gem = DRMLISTENTRY(drm_intel_bo_gem, 1271 bucket->head.next, head); 1272 if (time - bo_gem->free_time <= 1) 1273 break; 1274 1275 DRMLISTDEL(&bo_gem->head); 1276 1277 drm_intel_gem_bo_free(&bo_gem->bo); 1278 } 1279 } 1280 1281 bufmgr_gem->time = time; 1282 } 1283 1284 static void drm_intel_gem_bo_purge_vma_cache(drm_intel_bufmgr_gem *bufmgr_gem) 1285 { 1286 int limit; 1287 1288 DBG("%s: cached=%d, open=%d, limit=%d\n", __FUNCTION__, 1289 bufmgr_gem->vma_count, bufmgr_gem->vma_open, bufmgr_gem->vma_max); 1290 1291 if (bufmgr_gem->vma_max < 0) 1292 return; 1293 1294 /* We may need to evict a few entries in order to create new mmaps */ 1295 limit = bufmgr_gem->vma_max - 2*bufmgr_gem->vma_open; 1296 if (limit < 0) 1297 limit = 0; 1298 1299 while (bufmgr_gem->vma_count > limit) { 1300 drm_intel_bo_gem *bo_gem; 1301 1302 bo_gem = DRMLISTENTRY(drm_intel_bo_gem, 1303 bufmgr_gem->vma_cache.next, 1304 vma_list); 1305 assert(bo_gem->map_count == 0); 1306 DRMLISTDELINIT(&bo_gem->vma_list); 1307 1308 if (bo_gem->mem_virtual) { 1309 drm_munmap(bo_gem->mem_virtual, bo_gem->bo.size); 1310 bo_gem->mem_virtual = NULL; 1311 bufmgr_gem->vma_count--; 1312 } 1313 if (bo_gem->wc_virtual) { 1314 drm_munmap(bo_gem->wc_virtual, bo_gem->bo.size); 1315 bo_gem->wc_virtual = NULL; 1316 bufmgr_gem->vma_count--; 1317 } 1318 if (bo_gem->gtt_virtual) { 1319 drm_munmap(bo_gem->gtt_virtual, bo_gem->bo.size); 1320 bo_gem->gtt_virtual = NULL; 1321 bufmgr_gem->vma_count--; 1322 } 1323 } 1324 } 1325 1326 static void drm_intel_gem_bo_close_vma(drm_intel_bufmgr_gem *bufmgr_gem, 1327 drm_intel_bo_gem *bo_gem) 1328 { 1329 bufmgr_gem->vma_open--; 1330 DRMLISTADDTAIL(&bo_gem->vma_list, &bufmgr_gem->vma_cache); 1331 if (bo_gem->mem_virtual) 1332 bufmgr_gem->vma_count++; 1333 if (bo_gem->wc_virtual) 1334 bufmgr_gem->vma_count++; 1335 if (bo_gem->gtt_virtual) 1336 bufmgr_gem->vma_count++; 1337 drm_intel_gem_bo_purge_vma_cache(bufmgr_gem); 1338 } 1339 1340 static void drm_intel_gem_bo_open_vma(drm_intel_bufmgr_gem *bufmgr_gem, 1341 drm_intel_bo_gem *bo_gem) 1342 { 1343 bufmgr_gem->vma_open++; 1344 DRMLISTDEL(&bo_gem->vma_list); 1345 if (bo_gem->mem_virtual) 1346 bufmgr_gem->vma_count--; 1347 if (bo_gem->wc_virtual) 1348 bufmgr_gem->vma_count--; 1349 if (bo_gem->gtt_virtual) 1350 bufmgr_gem->vma_count--; 1351 drm_intel_gem_bo_purge_vma_cache(bufmgr_gem); 1352 } 1353 1354 static void 1355 drm_intel_gem_bo_unreference_final(drm_intel_bo *bo, time_t time) 1356 { 1357 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1358 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1359 struct drm_intel_gem_bo_bucket *bucket; 1360 int i; 1361 1362 /* Unreference all the target buffers */ 1363 for (i = 0; i < bo_gem->reloc_count; i++) { 1364 if (bo_gem->reloc_target_info[i].bo != bo) { 1365 drm_intel_gem_bo_unreference_locked_timed(bo_gem-> 1366 reloc_target_info[i].bo, 1367 time); 1368 } 1369 } 1370 for (i = 0; i < bo_gem->softpin_target_count; i++) 1371 drm_intel_gem_bo_unreference_locked_timed(bo_gem->softpin_target[i], 1372 time); 1373 bo_gem->kflags = 0; 1374 bo_gem->reloc_count = 0; 1375 bo_gem->used_as_reloc_target = false; 1376 bo_gem->softpin_target_count = 0; 1377 1378 DBG("bo_unreference final: %d (%s)\n", 1379 bo_gem->gem_handle, bo_gem->name); 1380 1381 /* release memory associated with this object */ 1382 if (bo_gem->reloc_target_info) { 1383 free(bo_gem->reloc_target_info); 1384 bo_gem->reloc_target_info = NULL; 1385 } 1386 if (bo_gem->relocs) { 1387 free(bo_gem->relocs); 1388 bo_gem->relocs = NULL; 1389 } 1390 if (bo_gem->softpin_target) { 1391 free(bo_gem->softpin_target); 1392 bo_gem->softpin_target = NULL; 1393 bo_gem->softpin_target_size = 0; 1394 } 1395 1396 /* Clear any left-over mappings */ 1397 if (bo_gem->map_count) { 1398 DBG("bo freed with non-zero map-count %d\n", bo_gem->map_count); 1399 bo_gem->map_count = 0; 1400 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem); 1401 drm_intel_gem_bo_mark_mmaps_incoherent(bo); 1402 } 1403 1404 bucket = drm_intel_gem_bo_bucket_for_size(bufmgr_gem, bo->size); 1405 /* Put the buffer into our internal cache for reuse if we can. */ 1406 if (bufmgr_gem->bo_reuse && bo_gem->reusable && bucket != NULL && 1407 drm_intel_gem_bo_madvise_internal(bufmgr_gem, bo_gem, 1408 I915_MADV_DONTNEED)) { 1409 bo_gem->free_time = time; 1410 1411 bo_gem->name = NULL; 1412 bo_gem->validate_index = -1; 1413 1414 bo_gem->kflags = 0; 1415 1416 DRMLISTADDTAIL(&bo_gem->head, &bucket->head); 1417 } else { 1418 drm_intel_gem_bo_free(bo); 1419 } 1420 } 1421 1422 static void drm_intel_gem_bo_unreference_locked_timed(drm_intel_bo *bo, 1423 time_t time) 1424 { 1425 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1426 1427 assert(atomic_read(&bo_gem->refcount) > 0); 1428 if (atomic_dec_and_test(&bo_gem->refcount)) 1429 drm_intel_gem_bo_unreference_final(bo, time); 1430 } 1431 1432 static void drm_intel_gem_bo_unreference(drm_intel_bo *bo) 1433 { 1434 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1435 1436 assert(atomic_read(&bo_gem->refcount) > 0); 1437 1438 if (atomic_add_unless(&bo_gem->refcount, -1, 1)) { 1439 drm_intel_bufmgr_gem *bufmgr_gem = 1440 (drm_intel_bufmgr_gem *) bo->bufmgr; 1441 struct timespec time; 1442 1443 clock_gettime(CLOCK_MONOTONIC, &time); 1444 1445 pthread_mutex_lock(&bufmgr_gem->lock); 1446 1447 if (atomic_dec_and_test(&bo_gem->refcount)) { 1448 drm_intel_gem_bo_unreference_final(bo, time.tv_sec); 1449 drm_intel_gem_cleanup_bo_cache(bufmgr_gem, time.tv_sec); 1450 } 1451 1452 pthread_mutex_unlock(&bufmgr_gem->lock); 1453 } 1454 } 1455 1456 static int drm_intel_gem_bo_map(drm_intel_bo *bo, int write_enable) 1457 { 1458 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1459 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1460 struct drm_i915_gem_set_domain set_domain; 1461 int ret; 1462 1463 if (bo_gem->is_userptr) { 1464 /* Return the same user ptr */ 1465 bo->virtual = bo_gem->user_virtual; 1466 return 0; 1467 } 1468 1469 pthread_mutex_lock(&bufmgr_gem->lock); 1470 1471 if (bo_gem->map_count++ == 0) 1472 drm_intel_gem_bo_open_vma(bufmgr_gem, bo_gem); 1473 1474 if (!bo_gem->mem_virtual) { 1475 struct drm_i915_gem_mmap mmap_arg; 1476 1477 DBG("bo_map: %d (%s), map_count=%d\n", 1478 bo_gem->gem_handle, bo_gem->name, bo_gem->map_count); 1479 1480 memclear(mmap_arg); 1481 mmap_arg.handle = bo_gem->gem_handle; 1482 mmap_arg.size = bo->size; 1483 ret = drmIoctl(bufmgr_gem->fd, 1484 DRM_IOCTL_I915_GEM_MMAP, 1485 &mmap_arg); 1486 if (ret != 0) { 1487 ret = -errno; 1488 DBG("%s:%d: Error mapping buffer %d (%s): %s .\n", 1489 __FILE__, __LINE__, bo_gem->gem_handle, 1490 bo_gem->name, strerror(errno)); 1491 if (--bo_gem->map_count == 0) 1492 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem); 1493 pthread_mutex_unlock(&bufmgr_gem->lock); 1494 return ret; 1495 } 1496 VG(VALGRIND_MALLOCLIKE_BLOCK(mmap_arg.addr_ptr, mmap_arg.size, 0, 1)); 1497 bo_gem->mem_virtual = (void *)(uintptr_t) mmap_arg.addr_ptr; 1498 } 1499 DBG("bo_map: %d (%s) -> %p\n", bo_gem->gem_handle, bo_gem->name, 1500 bo_gem->mem_virtual); 1501 bo->virtual = bo_gem->mem_virtual; 1502 1503 memclear(set_domain); 1504 set_domain.handle = bo_gem->gem_handle; 1505 set_domain.read_domains = I915_GEM_DOMAIN_CPU; 1506 if (write_enable) 1507 set_domain.write_domain = I915_GEM_DOMAIN_CPU; 1508 else 1509 set_domain.write_domain = 0; 1510 ret = drmIoctl(bufmgr_gem->fd, 1511 DRM_IOCTL_I915_GEM_SET_DOMAIN, 1512 &set_domain); 1513 if (ret != 0) { 1514 DBG("%s:%d: Error setting to CPU domain %d: %s\n", 1515 __FILE__, __LINE__, bo_gem->gem_handle, 1516 strerror(errno)); 1517 } 1518 1519 if (write_enable) 1520 bo_gem->mapped_cpu_write = true; 1521 1522 drm_intel_gem_bo_mark_mmaps_incoherent(bo); 1523 VG(VALGRIND_MAKE_MEM_DEFINED(bo_gem->mem_virtual, bo->size)); 1524 pthread_mutex_unlock(&bufmgr_gem->lock); 1525 1526 return 0; 1527 } 1528 1529 static int 1530 map_gtt(drm_intel_bo *bo) 1531 { 1532 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1533 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1534 int ret; 1535 1536 if (bo_gem->is_userptr) 1537 return -EINVAL; 1538 1539 if (bo_gem->map_count++ == 0) 1540 drm_intel_gem_bo_open_vma(bufmgr_gem, bo_gem); 1541 1542 /* Get a mapping of the buffer if we haven't before. */ 1543 if (bo_gem->gtt_virtual == NULL) { 1544 struct drm_i915_gem_mmap_gtt mmap_arg; 1545 1546 DBG("bo_map_gtt: mmap %d (%s), map_count=%d\n", 1547 bo_gem->gem_handle, bo_gem->name, bo_gem->map_count); 1548 1549 memclear(mmap_arg); 1550 mmap_arg.handle = bo_gem->gem_handle; 1551 1552 /* Get the fake offset back... */ 1553 ret = drmIoctl(bufmgr_gem->fd, 1554 DRM_IOCTL_I915_GEM_MMAP_GTT, 1555 &mmap_arg); 1556 if (ret != 0) { 1557 ret = -errno; 1558 DBG("%s:%d: Error preparing buffer map %d (%s): %s .\n", 1559 __FILE__, __LINE__, 1560 bo_gem->gem_handle, bo_gem->name, 1561 strerror(errno)); 1562 if (--bo_gem->map_count == 0) 1563 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem); 1564 return ret; 1565 } 1566 1567 /* and mmap it */ 1568 bo_gem->gtt_virtual = drm_mmap(0, bo->size, PROT_READ | PROT_WRITE, 1569 MAP_SHARED, bufmgr_gem->fd, 1570 mmap_arg.offset); 1571 if (bo_gem->gtt_virtual == MAP_FAILED) { 1572 bo_gem->gtt_virtual = NULL; 1573 ret = -errno; 1574 DBG("%s:%d: Error mapping buffer %d (%s): %s .\n", 1575 __FILE__, __LINE__, 1576 bo_gem->gem_handle, bo_gem->name, 1577 strerror(errno)); 1578 if (--bo_gem->map_count == 0) 1579 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem); 1580 return ret; 1581 } 1582 } 1583 1584 bo->virtual = bo_gem->gtt_virtual; 1585 1586 DBG("bo_map_gtt: %d (%s) -> %p\n", bo_gem->gem_handle, bo_gem->name, 1587 bo_gem->gtt_virtual); 1588 1589 return 0; 1590 } 1591 1592 int 1593 drm_intel_gem_bo_map_gtt(drm_intel_bo *bo) 1594 { 1595 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1596 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1597 struct drm_i915_gem_set_domain set_domain; 1598 int ret; 1599 1600 pthread_mutex_lock(&bufmgr_gem->lock); 1601 1602 ret = map_gtt(bo); 1603 if (ret) { 1604 pthread_mutex_unlock(&bufmgr_gem->lock); 1605 return ret; 1606 } 1607 1608 /* Now move it to the GTT domain so that the GPU and CPU 1609 * caches are flushed and the GPU isn't actively using the 1610 * buffer. 1611 * 1612 * The pagefault handler does this domain change for us when 1613 * it has unbound the BO from the GTT, but it's up to us to 1614 * tell it when we're about to use things if we had done 1615 * rendering and it still happens to be bound to the GTT. 1616 */ 1617 memclear(set_domain); 1618 set_domain.handle = bo_gem->gem_handle; 1619 set_domain.read_domains = I915_GEM_DOMAIN_GTT; 1620 set_domain.write_domain = I915_GEM_DOMAIN_GTT; 1621 ret = drmIoctl(bufmgr_gem->fd, 1622 DRM_IOCTL_I915_GEM_SET_DOMAIN, 1623 &set_domain); 1624 if (ret != 0) { 1625 DBG("%s:%d: Error setting domain %d: %s\n", 1626 __FILE__, __LINE__, bo_gem->gem_handle, 1627 strerror(errno)); 1628 } 1629 1630 drm_intel_gem_bo_mark_mmaps_incoherent(bo); 1631 VG(VALGRIND_MAKE_MEM_DEFINED(bo_gem->gtt_virtual, bo->size)); 1632 pthread_mutex_unlock(&bufmgr_gem->lock); 1633 1634 return 0; 1635 } 1636 1637 /** 1638 * Performs a mapping of the buffer object like the normal GTT 1639 * mapping, but avoids waiting for the GPU to be done reading from or 1640 * rendering to the buffer. 1641 * 1642 * This is used in the implementation of GL_ARB_map_buffer_range: The 1643 * user asks to create a buffer, then does a mapping, fills some 1644 * space, runs a drawing command, then asks to map it again without 1645 * synchronizing because it guarantees that it won't write over the 1646 * data that the GPU is busy using (or, more specifically, that if it 1647 * does write over the data, it acknowledges that rendering is 1648 * undefined). 1649 */ 1650 1651 int 1652 drm_intel_gem_bo_map_unsynchronized(drm_intel_bo *bo) 1653 { 1654 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1655 #ifdef HAVE_VALGRIND 1656 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1657 #endif 1658 int ret; 1659 1660 /* If the CPU cache isn't coherent with the GTT, then use a 1661 * regular synchronized mapping. The problem is that we don't 1662 * track where the buffer was last used on the CPU side in 1663 * terms of drm_intel_bo_map vs drm_intel_gem_bo_map_gtt, so 1664 * we would potentially corrupt the buffer even when the user 1665 * does reasonable things. 1666 */ 1667 if (!bufmgr_gem->has_llc) 1668 return drm_intel_gem_bo_map_gtt(bo); 1669 1670 pthread_mutex_lock(&bufmgr_gem->lock); 1671 1672 ret = map_gtt(bo); 1673 if (ret == 0) { 1674 drm_intel_gem_bo_mark_mmaps_incoherent(bo); 1675 VG(VALGRIND_MAKE_MEM_DEFINED(bo_gem->gtt_virtual, bo->size)); 1676 } 1677 1678 pthread_mutex_unlock(&bufmgr_gem->lock); 1679 1680 return ret; 1681 } 1682 1683 static int drm_intel_gem_bo_unmap(drm_intel_bo *bo) 1684 { 1685 drm_intel_bufmgr_gem *bufmgr_gem; 1686 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1687 int ret = 0; 1688 1689 if (bo == NULL) 1690 return 0; 1691 1692 if (bo_gem->is_userptr) 1693 return 0; 1694 1695 bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1696 1697 pthread_mutex_lock(&bufmgr_gem->lock); 1698 1699 if (bo_gem->map_count <= 0) { 1700 DBG("attempted to unmap an unmapped bo\n"); 1701 pthread_mutex_unlock(&bufmgr_gem->lock); 1702 /* Preserve the old behaviour of just treating this as a 1703 * no-op rather than reporting the error. 1704 */ 1705 return 0; 1706 } 1707 1708 if (bo_gem->mapped_cpu_write) { 1709 struct drm_i915_gem_sw_finish sw_finish; 1710 1711 /* Cause a flush to happen if the buffer's pinned for 1712 * scanout, so the results show up in a timely manner. 1713 * Unlike GTT set domains, this only does work if the 1714 * buffer should be scanout-related. 1715 */ 1716 memclear(sw_finish); 1717 sw_finish.handle = bo_gem->gem_handle; 1718 ret = drmIoctl(bufmgr_gem->fd, 1719 DRM_IOCTL_I915_GEM_SW_FINISH, 1720 &sw_finish); 1721 ret = ret == -1 ? -errno : 0; 1722 1723 bo_gem->mapped_cpu_write = false; 1724 } 1725 1726 /* We need to unmap after every innovation as we cannot track 1727 * an open vma for every bo as that will exhaust the system 1728 * limits and cause later failures. 1729 */ 1730 if (--bo_gem->map_count == 0) { 1731 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem); 1732 drm_intel_gem_bo_mark_mmaps_incoherent(bo); 1733 bo->virtual = NULL; 1734 } 1735 pthread_mutex_unlock(&bufmgr_gem->lock); 1736 1737 return ret; 1738 } 1739 1740 int 1741 drm_intel_gem_bo_unmap_gtt(drm_intel_bo *bo) 1742 { 1743 return drm_intel_gem_bo_unmap(bo); 1744 } 1745 1746 static int 1747 drm_intel_gem_bo_subdata(drm_intel_bo *bo, unsigned long offset, 1748 unsigned long size, const void *data) 1749 { 1750 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1751 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1752 struct drm_i915_gem_pwrite pwrite; 1753 int ret; 1754 1755 if (bo_gem->is_userptr) 1756 return -EINVAL; 1757 1758 memclear(pwrite); 1759 pwrite.handle = bo_gem->gem_handle; 1760 pwrite.offset = offset; 1761 pwrite.size = size; 1762 pwrite.data_ptr = (uint64_t) (uintptr_t) data; 1763 ret = drmIoctl(bufmgr_gem->fd, 1764 DRM_IOCTL_I915_GEM_PWRITE, 1765 &pwrite); 1766 if (ret != 0) { 1767 ret = -errno; 1768 DBG("%s:%d: Error writing data to buffer %d: (%d %d) %s .\n", 1769 __FILE__, __LINE__, bo_gem->gem_handle, (int)offset, 1770 (int)size, strerror(errno)); 1771 } 1772 1773 return ret; 1774 } 1775 1776 static int 1777 drm_intel_gem_get_pipe_from_crtc_id(drm_intel_bufmgr *bufmgr, int crtc_id) 1778 { 1779 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr; 1780 struct drm_i915_get_pipe_from_crtc_id get_pipe_from_crtc_id; 1781 int ret; 1782 1783 memclear(get_pipe_from_crtc_id); 1784 get_pipe_from_crtc_id.crtc_id = crtc_id; 1785 ret = drmIoctl(bufmgr_gem->fd, 1786 DRM_IOCTL_I915_GET_PIPE_FROM_CRTC_ID, 1787 &get_pipe_from_crtc_id); 1788 if (ret != 0) { 1789 /* We return -1 here to signal that we don't 1790 * know which pipe is associated with this crtc. 1791 * This lets the caller know that this information 1792 * isn't available; using the wrong pipe for 1793 * vblank waiting can cause the chipset to lock up 1794 */ 1795 return -1; 1796 } 1797 1798 return get_pipe_from_crtc_id.pipe; 1799 } 1800 1801 static int 1802 drm_intel_gem_bo_get_subdata(drm_intel_bo *bo, unsigned long offset, 1803 unsigned long size, void *data) 1804 { 1805 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1806 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1807 struct drm_i915_gem_pread pread; 1808 int ret; 1809 1810 if (bo_gem->is_userptr) 1811 return -EINVAL; 1812 1813 memclear(pread); 1814 pread.handle = bo_gem->gem_handle; 1815 pread.offset = offset; 1816 pread.size = size; 1817 pread.data_ptr = (uint64_t) (uintptr_t) data; 1818 ret = drmIoctl(bufmgr_gem->fd, 1819 DRM_IOCTL_I915_GEM_PREAD, 1820 &pread); 1821 if (ret != 0) { 1822 ret = -errno; 1823 DBG("%s:%d: Error reading data from buffer %d: (%d %d) %s .\n", 1824 __FILE__, __LINE__, bo_gem->gem_handle, (int)offset, 1825 (int)size, strerror(errno)); 1826 } 1827 1828 return ret; 1829 } 1830 1831 /** Waits for all GPU rendering with the object to have completed. */ 1832 static void 1833 drm_intel_gem_bo_wait_rendering(drm_intel_bo *bo) 1834 { 1835 drm_intel_gem_bo_start_gtt_access(bo, 1); 1836 } 1837 1838 /** 1839 * Waits on a BO for the given amount of time. 1840 * 1841 * @bo: buffer object to wait for 1842 * @timeout_ns: amount of time to wait in nanoseconds. 1843 * If value is less than 0, an infinite wait will occur. 1844 * 1845 * Returns 0 if the wait was successful ie. the last batch referencing the 1846 * object has completed within the allotted time. Otherwise some negative return 1847 * value describes the error. Of particular interest is -ETIME when the wait has 1848 * failed to yield the desired result. 1849 * 1850 * Similar to drm_intel_gem_bo_wait_rendering except a timeout parameter allows 1851 * the operation to give up after a certain amount of time. Another subtle 1852 * difference is the internal locking semantics are different (this variant does 1853 * not hold the lock for the duration of the wait). This makes the wait subject 1854 * to a larger userspace race window. 1855 * 1856 * The implementation shall wait until the object is no longer actively 1857 * referenced within a batch buffer at the time of the call. The wait will 1858 * not guarantee that the buffer is re-issued via another thread, or an flinked 1859 * handle. Userspace must make sure this race does not occur if such precision 1860 * is important. 1861 * 1862 * Note that some kernels have broken the inifite wait for negative values 1863 * promise, upgrade to latest stable kernels if this is the case. 1864 */ 1865 int 1866 drm_intel_gem_bo_wait(drm_intel_bo *bo, int64_t timeout_ns) 1867 { 1868 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1869 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1870 struct drm_i915_gem_wait wait; 1871 int ret; 1872 1873 if (!bufmgr_gem->has_wait_timeout) { 1874 DBG("%s:%d: Timed wait is not supported. Falling back to " 1875 "infinite wait\n", __FILE__, __LINE__); 1876 if (timeout_ns) { 1877 drm_intel_gem_bo_wait_rendering(bo); 1878 return 0; 1879 } else { 1880 return drm_intel_gem_bo_busy(bo) ? -ETIME : 0; 1881 } 1882 } 1883 1884 memclear(wait); 1885 wait.bo_handle = bo_gem->gem_handle; 1886 wait.timeout_ns = timeout_ns; 1887 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_WAIT, &wait); 1888 if (ret == -1) 1889 return -errno; 1890 1891 return ret; 1892 } 1893 1894 /** 1895 * Sets the object to the GTT read and possibly write domain, used by the X 1896 * 2D driver in the absence of kernel support to do drm_intel_gem_bo_map_gtt(). 1897 * 1898 * In combination with drm_intel_gem_bo_pin() and manual fence management, we 1899 * can do tiled pixmaps this way. 1900 */ 1901 void 1902 drm_intel_gem_bo_start_gtt_access(drm_intel_bo *bo, int write_enable) 1903 { 1904 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1905 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1906 struct drm_i915_gem_set_domain set_domain; 1907 int ret; 1908 1909 memclear(set_domain); 1910 set_domain.handle = bo_gem->gem_handle; 1911 set_domain.read_domains = I915_GEM_DOMAIN_GTT; 1912 set_domain.write_domain = write_enable ? I915_GEM_DOMAIN_GTT : 0; 1913 ret = drmIoctl(bufmgr_gem->fd, 1914 DRM_IOCTL_I915_GEM_SET_DOMAIN, 1915 &set_domain); 1916 if (ret != 0) { 1917 DBG("%s:%d: Error setting memory domains %d (%08x %08x): %s .\n", 1918 __FILE__, __LINE__, bo_gem->gem_handle, 1919 set_domain.read_domains, set_domain.write_domain, 1920 strerror(errno)); 1921 } 1922 } 1923 1924 static void 1925 drm_intel_bufmgr_gem_destroy(drm_intel_bufmgr *bufmgr) 1926 { 1927 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr; 1928 struct drm_gem_close close_bo; 1929 int i, ret; 1930 1931 free(bufmgr_gem->exec2_objects); 1932 free(bufmgr_gem->exec_objects); 1933 free(bufmgr_gem->exec_bos); 1934 1935 pthread_mutex_destroy(&bufmgr_gem->lock); 1936 1937 /* Free any cached buffer objects we were going to reuse */ 1938 for (i = 0; i < bufmgr_gem->num_buckets; i++) { 1939 struct drm_intel_gem_bo_bucket *bucket = 1940 &bufmgr_gem->cache_bucket[i]; 1941 drm_intel_bo_gem *bo_gem; 1942 1943 while (!DRMLISTEMPTY(&bucket->head)) { 1944 bo_gem = DRMLISTENTRY(drm_intel_bo_gem, 1945 bucket->head.next, head); 1946 DRMLISTDEL(&bo_gem->head); 1947 1948 drm_intel_gem_bo_free(&bo_gem->bo); 1949 } 1950 } 1951 1952 /* Release userptr bo kept hanging around for optimisation. */ 1953 if (bufmgr_gem->userptr_active.ptr) { 1954 memclear(close_bo); 1955 close_bo.handle = bufmgr_gem->userptr_active.handle; 1956 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_GEM_CLOSE, &close_bo); 1957 free(bufmgr_gem->userptr_active.ptr); 1958 if (ret) 1959 fprintf(stderr, 1960 "Failed to release test userptr object! (%d) " 1961 "i915 kernel driver may not be sane!\n", errno); 1962 } 1963 1964 free(bufmgr); 1965 } 1966 1967 /** 1968 * Adds the target buffer to the validation list and adds the relocation 1969 * to the reloc_buffer's relocation list. 1970 * 1971 * The relocation entry at the given offset must already contain the 1972 * precomputed relocation value, because the kernel will optimize out 1973 * the relocation entry write when the buffer hasn't moved from the 1974 * last known offset in target_bo. 1975 */ 1976 static int 1977 do_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset, 1978 drm_intel_bo *target_bo, uint32_t target_offset, 1979 uint32_t read_domains, uint32_t write_domain, 1980 bool need_fence) 1981 { 1982 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 1983 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 1984 drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *) target_bo; 1985 bool fenced_command; 1986 1987 if (bo_gem->has_error) 1988 return -ENOMEM; 1989 1990 if (target_bo_gem->has_error) { 1991 bo_gem->has_error = true; 1992 return -ENOMEM; 1993 } 1994 1995 /* We never use HW fences for rendering on 965+ */ 1996 if (bufmgr_gem->gen >= 4) 1997 need_fence = false; 1998 1999 fenced_command = need_fence; 2000 if (target_bo_gem->tiling_mode == I915_TILING_NONE) 2001 need_fence = false; 2002 2003 /* Create a new relocation list if needed */ 2004 if (bo_gem->relocs == NULL && drm_intel_setup_reloc_list(bo)) 2005 return -ENOMEM; 2006 2007 /* Check overflow */ 2008 assert(bo_gem->reloc_count < bufmgr_gem->max_relocs); 2009 2010 /* Check args */ 2011 assert(offset <= bo->size - 4); 2012 assert((write_domain & (write_domain - 1)) == 0); 2013 2014 /* An object needing a fence is a tiled buffer, so it won't have 2015 * relocs to other buffers. 2016 */ 2017 if (need_fence) { 2018 assert(target_bo_gem->reloc_count == 0); 2019 target_bo_gem->reloc_tree_fences = 1; 2020 } 2021 2022 /* Make sure that we're not adding a reloc to something whose size has 2023 * already been accounted for. 2024 */ 2025 assert(!bo_gem->used_as_reloc_target); 2026 if (target_bo_gem != bo_gem) { 2027 target_bo_gem->used_as_reloc_target = true; 2028 bo_gem->reloc_tree_size += target_bo_gem->reloc_tree_size; 2029 bo_gem->reloc_tree_fences += target_bo_gem->reloc_tree_fences; 2030 } 2031 2032 bo_gem->reloc_target_info[bo_gem->reloc_count].bo = target_bo; 2033 if (target_bo != bo) 2034 drm_intel_gem_bo_reference(target_bo); 2035 if (fenced_command) 2036 bo_gem->reloc_target_info[bo_gem->reloc_count].flags = 2037 DRM_INTEL_RELOC_FENCE; 2038 else 2039 bo_gem->reloc_target_info[bo_gem->reloc_count].flags = 0; 2040 2041 bo_gem->relocs[bo_gem->reloc_count].offset = offset; 2042 bo_gem->relocs[bo_gem->reloc_count].delta = target_offset; 2043 bo_gem->relocs[bo_gem->reloc_count].target_handle = 2044 target_bo_gem->gem_handle; 2045 bo_gem->relocs[bo_gem->reloc_count].read_domains = read_domains; 2046 bo_gem->relocs[bo_gem->reloc_count].write_domain = write_domain; 2047 bo_gem->relocs[bo_gem->reloc_count].presumed_offset = target_bo->offset64; 2048 bo_gem->reloc_count++; 2049 2050 return 0; 2051 } 2052 2053 static void 2054 drm_intel_gem_bo_use_48b_address_range(drm_intel_bo *bo, uint32_t enable) 2055 { 2056 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2057 bo_gem->use_48b_address_range = enable; 2058 } 2059 2060 static int 2061 drm_intel_gem_bo_add_softpin_target(drm_intel_bo *bo, drm_intel_bo *target_bo) 2062 { 2063 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 2064 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2065 drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *) target_bo; 2066 if (bo_gem->has_error) 2067 return -ENOMEM; 2068 2069 if (target_bo_gem->has_error) { 2070 bo_gem->has_error = true; 2071 return -ENOMEM; 2072 } 2073 2074 if (!target_bo_gem->is_softpin) 2075 return -EINVAL; 2076 if (target_bo_gem == bo_gem) 2077 return -EINVAL; 2078 2079 if (bo_gem->softpin_target_count == bo_gem->softpin_target_size) { 2080 int new_size = bo_gem->softpin_target_size * 2; 2081 if (new_size == 0) 2082 new_size = bufmgr_gem->max_relocs; 2083 2084 bo_gem->softpin_target = realloc(bo_gem->softpin_target, new_size * 2085 sizeof(drm_intel_bo *)); 2086 if (!bo_gem->softpin_target) 2087 return -ENOMEM; 2088 2089 bo_gem->softpin_target_size = new_size; 2090 } 2091 bo_gem->softpin_target[bo_gem->softpin_target_count] = target_bo; 2092 drm_intel_gem_bo_reference(target_bo); 2093 bo_gem->softpin_target_count++; 2094 2095 return 0; 2096 } 2097 2098 static int 2099 drm_intel_gem_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset, 2100 drm_intel_bo *target_bo, uint32_t target_offset, 2101 uint32_t read_domains, uint32_t write_domain) 2102 { 2103 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr; 2104 drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *)target_bo; 2105 2106 if (target_bo_gem->is_softpin) 2107 return drm_intel_gem_bo_add_softpin_target(bo, target_bo); 2108 else 2109 return do_bo_emit_reloc(bo, offset, target_bo, target_offset, 2110 read_domains, write_domain, 2111 !bufmgr_gem->fenced_relocs); 2112 } 2113 2114 static int 2115 drm_intel_gem_bo_emit_reloc_fence(drm_intel_bo *bo, uint32_t offset, 2116 drm_intel_bo *target_bo, 2117 uint32_t target_offset, 2118 uint32_t read_domains, uint32_t write_domain) 2119 { 2120 return do_bo_emit_reloc(bo, offset, target_bo, target_offset, 2121 read_domains, write_domain, true); 2122 } 2123 2124 int 2125 drm_intel_gem_bo_get_reloc_count(drm_intel_bo *bo) 2126 { 2127 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2128 2129 return bo_gem->reloc_count; 2130 } 2131 2132 /** 2133 * Removes existing relocation entries in the BO after "start". 2134 * 2135 * This allows a user to avoid a two-step process for state setup with 2136 * counting up all the buffer objects and doing a 2137 * drm_intel_bufmgr_check_aperture_space() before emitting any of the 2138 * relocations for the state setup. Instead, save the state of the 2139 * batchbuffer including drm_intel_gem_get_reloc_count(), emit all the 2140 * state, and then check if it still fits in the aperture. 2141 * 2142 * Any further drm_intel_bufmgr_check_aperture_space() queries 2143 * involving this buffer in the tree are undefined after this call. 2144 * 2145 * This also removes all softpinned targets being referenced by the BO. 2146 */ 2147 void 2148 drm_intel_gem_bo_clear_relocs(drm_intel_bo *bo, int start) 2149 { 2150 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 2151 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2152 int i; 2153 struct timespec time; 2154 2155 clock_gettime(CLOCK_MONOTONIC, &time); 2156 2157 assert(bo_gem->reloc_count >= start); 2158 2159 /* Unreference the cleared target buffers */ 2160 pthread_mutex_lock(&bufmgr_gem->lock); 2161 2162 for (i = start; i < bo_gem->reloc_count; i++) { 2163 drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *) bo_gem->reloc_target_info[i].bo; 2164 if (&target_bo_gem->bo != bo) { 2165 bo_gem->reloc_tree_fences -= target_bo_gem->reloc_tree_fences; 2166 drm_intel_gem_bo_unreference_locked_timed(&target_bo_gem->bo, 2167 time.tv_sec); 2168 } 2169 } 2170 bo_gem->reloc_count = start; 2171 2172 for (i = 0; i < bo_gem->softpin_target_count; i++) { 2173 drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *) bo_gem->softpin_target[i]; 2174 drm_intel_gem_bo_unreference_locked_timed(&target_bo_gem->bo, time.tv_sec); 2175 } 2176 bo_gem->softpin_target_count = 0; 2177 2178 pthread_mutex_unlock(&bufmgr_gem->lock); 2179 2180 } 2181 2182 /** 2183 * Walk the tree of relocations rooted at BO and accumulate the list of 2184 * validations to be performed and update the relocation buffers with 2185 * index values into the validation list. 2186 */ 2187 static void 2188 drm_intel_gem_bo_process_reloc(drm_intel_bo *bo) 2189 { 2190 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2191 int i; 2192 2193 if (bo_gem->relocs == NULL) 2194 return; 2195 2196 for (i = 0; i < bo_gem->reloc_count; i++) { 2197 drm_intel_bo *target_bo = bo_gem->reloc_target_info[i].bo; 2198 2199 if (target_bo == bo) 2200 continue; 2201 2202 drm_intel_gem_bo_mark_mmaps_incoherent(bo); 2203 2204 /* Continue walking the tree depth-first. */ 2205 drm_intel_gem_bo_process_reloc(target_bo); 2206 2207 /* Add the target to the validate list */ 2208 drm_intel_add_validate_buffer(target_bo); 2209 } 2210 } 2211 2212 static void 2213 drm_intel_gem_bo_process_reloc2(drm_intel_bo *bo) 2214 { 2215 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo; 2216 int i; 2217 2218 if (bo_gem->relocs == NULL && bo_gem->softpin_target == NULL) 2219 return; 2220 2221 for (i = 0; i < bo_gem->reloc_count; i++) { 2222 drm_intel_bo *target_bo = bo_gem->reloc_target_info[i].bo; 2223 int need_fence; 2224 2225 if (target_bo == bo) 2226 continue; 2227 2228 drm_intel_gem_bo_mark_mmaps_incoherent(bo); 2229 2230 /* Continue walking the tree depth-first. */ 2231 drm_intel_gem_bo_process_reloc2(target_bo); 2232 2233 need_fence = (bo_gem->reloc_target_info[i].flags & 2234 DRM_INTEL_RELOC_FENCE); 2235 2236 /* Add the target to the validate list */ 2237 drm_intel_add_validate_buffer2(target_bo, need_fence); 2238 } 2239 2240 for (i = 0; i < bo_gem->softpin_target_count; i++) { 2241 drm_intel_bo *target_bo = bo_gem->softpin_target[i]; 2242 2243 if (target_bo == bo) 2244 continue; 2245 2246 drm_intel_gem_bo_mark_mmaps_incoherent(bo); 2247 drm_intel_gem_bo_process_reloc2(target_bo); 2248 drm_intel_add_validate_buffer2(target_bo, false); 2249 } 2250 } 2251 2252 2253 static void 2254 drm_intel_update_buffer_offsets(drm_intel_bufmgr_gem *bufmgr_gem) 2255 { 2256 int i; 2257 2258 for (i = 0; i < bufmgr_gem->exec_count; i++) { 2259 drm_intel_bo *bo = bufmgr_gem->exec_bos[i]; 2260 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2261 2262 /* Update the buffer offset */ 2263 if (bufmgr_gem->exec_objects[i].offset != bo->offset64) { 2264 DBG("BO %d (%s) migrated: 0x%08x %08x -> 0x%08x %08x\n", 2265 bo_gem->gem_handle, bo_gem->name, 2266 upper_32_bits(bo->offset64), 2267 lower_32_bits(bo->offset64), 2268 upper_32_bits(bufmgr_gem->exec_objects[i].offset), 2269 lower_32_bits(bufmgr_gem->exec_objects[i].offset)); 2270 bo->offset64 = bufmgr_gem->exec_objects[i].offset; 2271 bo->offset = bufmgr_gem->exec_objects[i].offset; 2272 } 2273 } 2274 } 2275 2276 static void 2277 drm_intel_update_buffer_offsets2 (drm_intel_bufmgr_gem *bufmgr_gem) 2278 { 2279 int i; 2280 2281 for (i = 0; i < bufmgr_gem->exec_count; i++) { 2282 drm_intel_bo *bo = bufmgr_gem->exec_bos[i]; 2283 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo; 2284 2285 /* Update the buffer offset */ 2286 if (bufmgr_gem->exec2_objects[i].offset != bo->offset64) { 2287 /* If we're seeing softpinned object here it means that the kernel 2288 * has relocated our object... Indicating a programming error 2289 */ 2290 assert(!bo_gem->is_softpin); 2291 DBG("BO %d (%s) migrated: 0x%08x %08x -> 0x%08x %08x\n", 2292 bo_gem->gem_handle, bo_gem->name, 2293 upper_32_bits(bo->offset64), 2294 lower_32_bits(bo->offset64), 2295 upper_32_bits(bufmgr_gem->exec2_objects[i].offset), 2296 lower_32_bits(bufmgr_gem->exec2_objects[i].offset)); 2297 bo->offset64 = bufmgr_gem->exec2_objects[i].offset; 2298 bo->offset = bufmgr_gem->exec2_objects[i].offset; 2299 } 2300 } 2301 } 2302 2303 void 2304 drm_intel_gem_bo_aub_dump_bmp(drm_intel_bo *bo, 2305 int x1, int y1, int width, int height, 2306 enum aub_dump_bmp_format format, 2307 int pitch, int offset) 2308 { 2309 } 2310 2311 static int 2312 drm_intel_gem_bo_exec(drm_intel_bo *bo, int used, 2313 drm_clip_rect_t * cliprects, int num_cliprects, int DR4) 2314 { 2315 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 2316 struct drm_i915_gem_execbuffer execbuf; 2317 int ret, i; 2318 2319 if (to_bo_gem(bo)->has_error) 2320 return -ENOMEM; 2321 2322 pthread_mutex_lock(&bufmgr_gem->lock); 2323 /* Update indices and set up the validate list. */ 2324 drm_intel_gem_bo_process_reloc(bo); 2325 2326 /* Add the batch buffer to the validation list. There are no 2327 * relocations pointing to it. 2328 */ 2329 drm_intel_add_validate_buffer(bo); 2330 2331 memclear(execbuf); 2332 execbuf.buffers_ptr = (uintptr_t) bufmgr_gem->exec_objects; 2333 execbuf.buffer_count = bufmgr_gem->exec_count; 2334 execbuf.batch_start_offset = 0; 2335 execbuf.batch_len = used; 2336 execbuf.cliprects_ptr = (uintptr_t) cliprects; 2337 execbuf.num_cliprects = num_cliprects; 2338 execbuf.DR1 = 0; 2339 execbuf.DR4 = DR4; 2340 2341 ret = drmIoctl(bufmgr_gem->fd, 2342 DRM_IOCTL_I915_GEM_EXECBUFFER, 2343 &execbuf); 2344 if (ret != 0) { 2345 ret = -errno; 2346 if (errno == ENOSPC) { 2347 DBG("Execbuffer fails to pin. " 2348 "Estimate: %u. Actual: %u. Available: %u\n", 2349 drm_intel_gem_estimate_batch_space(bufmgr_gem->exec_bos, 2350 bufmgr_gem-> 2351 exec_count), 2352 drm_intel_gem_compute_batch_space(bufmgr_gem->exec_bos, 2353 bufmgr_gem-> 2354 exec_count), 2355 (unsigned int)bufmgr_gem->gtt_size); 2356 } 2357 } 2358 drm_intel_update_buffer_offsets(bufmgr_gem); 2359 2360 if (bufmgr_gem->bufmgr.debug) 2361 drm_intel_gem_dump_validation_list(bufmgr_gem); 2362 2363 for (i = 0; i < bufmgr_gem->exec_count; i++) { 2364 drm_intel_bo_gem *bo_gem = to_bo_gem(bufmgr_gem->exec_bos[i]); 2365 2366 bo_gem->idle = false; 2367 2368 /* Disconnect the buffer from the validate list */ 2369 bo_gem->validate_index = -1; 2370 bufmgr_gem->exec_bos[i] = NULL; 2371 } 2372 bufmgr_gem->exec_count = 0; 2373 pthread_mutex_unlock(&bufmgr_gem->lock); 2374 2375 return ret; 2376 } 2377 2378 static int 2379 do_exec2(drm_intel_bo *bo, int used, drm_intel_context *ctx, 2380 drm_clip_rect_t *cliprects, int num_cliprects, int DR4, 2381 int in_fence, int *out_fence, 2382 unsigned int flags) 2383 { 2384 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr; 2385 struct drm_i915_gem_execbuffer2 execbuf; 2386 int ret = 0; 2387 int i; 2388 2389 if (to_bo_gem(bo)->has_error) 2390 return -ENOMEM; 2391 2392 switch (flags & 0x7) { 2393 default: 2394 return -EINVAL; 2395 case I915_EXEC_BLT: 2396 if (!bufmgr_gem->has_blt) 2397 return -EINVAL; 2398 break; 2399 case I915_EXEC_BSD: 2400 if (!bufmgr_gem->has_bsd) 2401 return -EINVAL; 2402 break; 2403 case I915_EXEC_VEBOX: 2404 if (!bufmgr_gem->has_vebox) 2405 return -EINVAL; 2406 break; 2407 case I915_EXEC_RENDER: 2408 case I915_EXEC_DEFAULT: 2409 break; 2410 } 2411 2412 pthread_mutex_lock(&bufmgr_gem->lock); 2413 /* Update indices and set up the validate list. */ 2414 drm_intel_gem_bo_process_reloc2(bo); 2415 2416 /* Add the batch buffer to the validation list. There are no relocations 2417 * pointing to it. 2418 */ 2419 drm_intel_add_validate_buffer2(bo, 0); 2420 2421 memclear(execbuf); 2422 execbuf.buffers_ptr = (uintptr_t)bufmgr_gem->exec2_objects; 2423 execbuf.buffer_count = bufmgr_gem->exec_count; 2424 execbuf.batch_start_offset = 0; 2425 execbuf.batch_len = used; 2426 execbuf.cliprects_ptr = (uintptr_t)cliprects; 2427 execbuf.num_cliprects = num_cliprects; 2428 execbuf.DR1 = 0; 2429 execbuf.DR4 = DR4; 2430 execbuf.flags = flags; 2431 if (ctx == NULL) 2432 i915_execbuffer2_set_context_id(execbuf, 0); 2433 else 2434 i915_execbuffer2_set_context_id(execbuf, ctx->ctx_id); 2435 execbuf.rsvd2 = 0; 2436 if (in_fence != -1) { 2437 execbuf.rsvd2 = in_fence; 2438 execbuf.flags |= I915_EXEC_FENCE_IN; 2439 } 2440 if (out_fence != NULL) { 2441 *out_fence = -1; 2442 execbuf.flags |= I915_EXEC_FENCE_OUT; 2443 } 2444 2445 if (bufmgr_gem->no_exec) 2446 goto skip_execution; 2447 2448 ret = drmIoctl(bufmgr_gem->fd, 2449 DRM_IOCTL_I915_GEM_EXECBUFFER2_WR, 2450 &execbuf); 2451 if (ret != 0) { 2452 ret = -errno; 2453 if (ret == -ENOSPC) { 2454 DBG("Execbuffer fails to pin. " 2455 "Estimate: %u. Actual: %u. Available: %u\n", 2456 drm_intel_gem_estimate_batch_space(bufmgr_gem->exec_bos, 2457 bufmgr_gem->exec_count), 2458 drm_intel_gem_compute_batch_space(bufmgr_gem->exec_bos, 2459 bufmgr_gem->exec_count), 2460 (unsigned int) bufmgr_gem->gtt_size); 2461 } 2462 } 2463 drm_intel_update_buffer_offsets2(bufmgr_gem); 2464 2465 if (ret == 0 && out_fence != NULL) 2466 *out_fence = execbuf.rsvd2 >> 32; 2467 2468 skip_execution: 2469 if (bufmgr_gem->bufmgr.debug) 2470 drm_intel_gem_dump_validation_list(bufmgr_gem); 2471 2472 for (i = 0; i < bufmgr_gem->exec_count; i++) { 2473 drm_intel_bo_gem *bo_gem = to_bo_gem(bufmgr_gem->exec_bos[i]); 2474 2475 bo_gem->idle = false; 2476 2477 /* Disconnect the buffer from the validate list */ 2478 bo_gem->validate_index = -1; 2479 bufmgr_gem->exec_bos[i] = NULL; 2480 } 2481 bufmgr_gem->exec_count = 0; 2482 pthread_mutex_unlock(&bufmgr_gem->lock); 2483 2484 return ret; 2485 } 2486 2487 static int 2488 drm_intel_gem_bo_exec2(drm_intel_bo *bo, int used, 2489 drm_clip_rect_t *cliprects, int num_cliprects, 2490 int DR4) 2491 { 2492 return do_exec2(bo, used, NULL, cliprects, num_cliprects, DR4, 2493 -1, NULL, I915_EXEC_RENDER); 2494 } 2495 2496 static int 2497 drm_intel_gem_bo_mrb_exec2(drm_intel_bo *bo, int used, 2498 drm_clip_rect_t *cliprects, int num_cliprects, int DR4, 2499 unsigned int flags) 2500 { 2501 return do_exec2(bo, used, NULL, cliprects, num_cliprects, DR4, 2502 -1, NULL, flags); 2503 } 2504 2505 int 2506 drm_intel_gem_bo_context_exec(drm_intel_bo *bo, drm_intel_context *ctx, 2507 int used, unsigned int flags) 2508 { 2509 return do_exec2(bo, used, ctx, NULL, 0, 0, -1, NULL, flags); 2510 } 2511 2512 int 2513 drm_intel_gem_bo_fence_exec(drm_intel_bo *bo, 2514 drm_intel_context *ctx, 2515 int used, 2516 int in_fence, 2517 int *out_fence, 2518 unsigned int flags) 2519 { 2520 return do_exec2(bo, used, ctx, NULL, 0, 0, in_fence, out_fence, flags); 2521 } 2522 2523 static int 2524 drm_intel_gem_bo_pin(drm_intel_bo *bo, uint32_t alignment) 2525 { 2526 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 2527 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2528 struct drm_i915_gem_pin pin; 2529 int ret; 2530 2531 memclear(pin); 2532 pin.handle = bo_gem->gem_handle; 2533 pin.alignment = alignment; 2534 2535 ret = drmIoctl(bufmgr_gem->fd, 2536 DRM_IOCTL_I915_GEM_PIN, 2537 &pin); 2538 if (ret != 0) 2539 return -errno; 2540 2541 bo->offset64 = pin.offset; 2542 bo->offset = pin.offset; 2543 return 0; 2544 } 2545 2546 static int 2547 drm_intel_gem_bo_unpin(drm_intel_bo *bo) 2548 { 2549 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 2550 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2551 struct drm_i915_gem_unpin unpin; 2552 int ret; 2553 2554 memclear(unpin); 2555 unpin.handle = bo_gem->gem_handle; 2556 2557 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_UNPIN, &unpin); 2558 if (ret != 0) 2559 return -errno; 2560 2561 return 0; 2562 } 2563 2564 static int 2565 drm_intel_gem_bo_set_tiling_internal(drm_intel_bo *bo, 2566 uint32_t tiling_mode, 2567 uint32_t stride) 2568 { 2569 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 2570 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2571 struct drm_i915_gem_set_tiling set_tiling; 2572 int ret; 2573 2574 if (bo_gem->global_name == 0 && 2575 tiling_mode == bo_gem->tiling_mode && 2576 stride == bo_gem->stride) 2577 return 0; 2578 2579 memset(&set_tiling, 0, sizeof(set_tiling)); 2580 do { 2581 /* set_tiling is slightly broken and overwrites the 2582 * input on the error path, so we have to open code 2583 * rmIoctl. 2584 */ 2585 set_tiling.handle = bo_gem->gem_handle; 2586 set_tiling.tiling_mode = tiling_mode; 2587 set_tiling.stride = stride; 2588 2589 ret = ioctl(bufmgr_gem->fd, 2590 DRM_IOCTL_I915_GEM_SET_TILING, 2591 &set_tiling); 2592 } while (ret == -1 && (errno == EINTR || errno == EAGAIN)); 2593 if (ret == -1) 2594 return -errno; 2595 2596 bo_gem->tiling_mode = set_tiling.tiling_mode; 2597 bo_gem->swizzle_mode = set_tiling.swizzle_mode; 2598 bo_gem->stride = set_tiling.stride; 2599 return 0; 2600 } 2601 2602 static int 2603 drm_intel_gem_bo_set_tiling(drm_intel_bo *bo, uint32_t * tiling_mode, 2604 uint32_t stride) 2605 { 2606 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 2607 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2608 int ret; 2609 2610 /* Tiling with userptr surfaces is not supported 2611 * on all hardware so refuse it for time being. 2612 */ 2613 if (bo_gem->is_userptr) 2614 return -EINVAL; 2615 2616 /* Linear buffers have no stride. By ensuring that we only ever use 2617 * stride 0 with linear buffers, we simplify our code. 2618 */ 2619 if (*tiling_mode == I915_TILING_NONE) 2620 stride = 0; 2621 2622 ret = drm_intel_gem_bo_set_tiling_internal(bo, *tiling_mode, stride); 2623 if (ret == 0) 2624 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem, 0); 2625 2626 *tiling_mode = bo_gem->tiling_mode; 2627 return ret; 2628 } 2629 2630 static int 2631 drm_intel_gem_bo_get_tiling(drm_intel_bo *bo, uint32_t * tiling_mode, 2632 uint32_t * swizzle_mode) 2633 { 2634 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2635 2636 *tiling_mode = bo_gem->tiling_mode; 2637 *swizzle_mode = bo_gem->swizzle_mode; 2638 return 0; 2639 } 2640 2641 static int 2642 drm_intel_gem_bo_set_softpin_offset(drm_intel_bo *bo, uint64_t offset) 2643 { 2644 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2645 2646 bo_gem->is_softpin = true; 2647 bo->offset64 = offset; 2648 bo->offset = offset; 2649 return 0; 2650 } 2651 2652 drm_intel_bo * 2653 drm_intel_bo_gem_create_from_prime(drm_intel_bufmgr *bufmgr, int prime_fd, int size) 2654 { 2655 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr; 2656 int ret; 2657 uint32_t handle; 2658 drm_intel_bo_gem *bo_gem; 2659 struct drm_i915_gem_get_tiling get_tiling; 2660 2661 pthread_mutex_lock(&bufmgr_gem->lock); 2662 ret = drmPrimeFDToHandle(bufmgr_gem->fd, prime_fd, &handle); 2663 if (ret) { 2664 DBG("create_from_prime: failed to obtain handle from fd: %s\n", strerror(errno)); 2665 pthread_mutex_unlock(&bufmgr_gem->lock); 2666 return NULL; 2667 } 2668 2669 /* 2670 * See if the kernel has already returned this buffer to us. Just as 2671 * for named buffers, we must not create two bo's pointing at the same 2672 * kernel object 2673 */ 2674 HASH_FIND(handle_hh, bufmgr_gem->handle_table, 2675 &handle, sizeof(handle), bo_gem); 2676 if (bo_gem) { 2677 drm_intel_gem_bo_reference(&bo_gem->bo); 2678 goto out; 2679 } 2680 2681 bo_gem = calloc(1, sizeof(*bo_gem)); 2682 if (!bo_gem) 2683 goto out; 2684 2685 atomic_set(&bo_gem->refcount, 1); 2686 DRMINITLISTHEAD(&bo_gem->vma_list); 2687 2688 /* Determine size of bo. The fd-to-handle ioctl really should 2689 * return the size, but it doesn't. If we have kernel 3.12 or 2690 * later, we can lseek on the prime fd to get the size. Older 2691 * kernels will just fail, in which case we fall back to the 2692 * provided (estimated or guess size). */ 2693 ret = lseek(prime_fd, 0, SEEK_END); 2694 if (ret != -1) 2695 bo_gem->bo.size = ret; 2696 else 2697 bo_gem->bo.size = size; 2698 2699 bo_gem->bo.handle = handle; 2700 bo_gem->bo.bufmgr = bufmgr; 2701 2702 bo_gem->gem_handle = handle; 2703 HASH_ADD(handle_hh, bufmgr_gem->handle_table, 2704 gem_handle, sizeof(bo_gem->gem_handle), bo_gem); 2705 2706 bo_gem->name = "prime"; 2707 bo_gem->validate_index = -1; 2708 bo_gem->reloc_tree_fences = 0; 2709 bo_gem->used_as_reloc_target = false; 2710 bo_gem->has_error = false; 2711 bo_gem->reusable = false; 2712 bo_gem->use_48b_address_range = false; 2713 2714 memclear(get_tiling); 2715 get_tiling.handle = bo_gem->gem_handle; 2716 if (drmIoctl(bufmgr_gem->fd, 2717 DRM_IOCTL_I915_GEM_GET_TILING, 2718 &get_tiling)) 2719 goto err; 2720 2721 bo_gem->tiling_mode = get_tiling.tiling_mode; 2722 bo_gem->swizzle_mode = get_tiling.swizzle_mode; 2723 /* XXX stride is unknown */ 2724 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem, 0); 2725 2726 out: 2727 pthread_mutex_unlock(&bufmgr_gem->lock); 2728 return &bo_gem->bo; 2729 2730 err: 2731 drm_intel_gem_bo_free(&bo_gem->bo); 2732 pthread_mutex_unlock(&bufmgr_gem->lock); 2733 return NULL; 2734 } 2735 2736 int 2737 drm_intel_bo_gem_export_to_prime(drm_intel_bo *bo, int *prime_fd) 2738 { 2739 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 2740 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2741 2742 if (drmPrimeHandleToFD(bufmgr_gem->fd, bo_gem->gem_handle, 2743 DRM_CLOEXEC, prime_fd) != 0) 2744 return -errno; 2745 2746 bo_gem->reusable = false; 2747 2748 return 0; 2749 } 2750 2751 static int 2752 drm_intel_gem_bo_flink(drm_intel_bo *bo, uint32_t * name) 2753 { 2754 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 2755 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2756 2757 if (!bo_gem->global_name) { 2758 struct drm_gem_flink flink; 2759 2760 memclear(flink); 2761 flink.handle = bo_gem->gem_handle; 2762 if (drmIoctl(bufmgr_gem->fd, DRM_IOCTL_GEM_FLINK, &flink)) 2763 return -errno; 2764 2765 pthread_mutex_lock(&bufmgr_gem->lock); 2766 if (!bo_gem->global_name) { 2767 bo_gem->global_name = flink.name; 2768 bo_gem->reusable = false; 2769 2770 HASH_ADD(name_hh, bufmgr_gem->name_table, 2771 global_name, sizeof(bo_gem->global_name), 2772 bo_gem); 2773 } 2774 pthread_mutex_unlock(&bufmgr_gem->lock); 2775 } 2776 2777 *name = bo_gem->global_name; 2778 return 0; 2779 } 2780 2781 /** 2782 * Enables unlimited caching of buffer objects for reuse. 2783 * 2784 * This is potentially very memory expensive, as the cache at each bucket 2785 * size is only bounded by how many buffers of that size we've managed to have 2786 * in flight at once. 2787 */ 2788 void 2789 drm_intel_bufmgr_gem_enable_reuse(drm_intel_bufmgr *bufmgr) 2790 { 2791 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr; 2792 2793 bufmgr_gem->bo_reuse = true; 2794 } 2795 2796 /** 2797 * Disables implicit synchronisation before executing the bo 2798 * 2799 * This will cause rendering corruption unless you correctly manage explicit 2800 * fences for all rendering involving this buffer - including use by others. 2801 * Disabling the implicit serialisation is only required if that serialisation 2802 * is too coarse (for example, you have split the buffer into many 2803 * non-overlapping regions and are sharing the whole buffer between concurrent 2804 * independent command streams). 2805 * 2806 * Note the kernel must advertise support via I915_PARAM_HAS_EXEC_ASYNC, 2807 * which can be checked using drm_intel_bufmgr_can_disable_implicit_sync, 2808 * or subsequent execbufs involving the bo will generate EINVAL. 2809 */ 2810 void 2811 drm_intel_gem_bo_disable_implicit_sync(drm_intel_bo *bo) 2812 { 2813 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2814 2815 bo_gem->kflags |= EXEC_OBJECT_ASYNC; 2816 } 2817 2818 /** 2819 * Enables implicit synchronisation before executing the bo 2820 * 2821 * This is the default behaviour of the kernel, to wait upon prior writes 2822 * completing on the object before rendering with it, or to wait for prior 2823 * reads to complete before writing into the object. 2824 * drm_intel_gem_bo_disable_implicit_sync() can stop this behaviour, telling 2825 * the kernel never to insert a stall before using the object. Then this 2826 * function can be used to restore the implicit sync before subsequent 2827 * rendering. 2828 */ 2829 void 2830 drm_intel_gem_bo_enable_implicit_sync(drm_intel_bo *bo) 2831 { 2832 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2833 2834 bo_gem->kflags &= ~EXEC_OBJECT_ASYNC; 2835 } 2836 2837 /** 2838 * Query whether the kernel supports disabling of its implicit synchronisation 2839 * before execbuf. See drm_intel_gem_bo_disable_implicit_sync() 2840 */ 2841 int 2842 drm_intel_bufmgr_gem_can_disable_implicit_sync(drm_intel_bufmgr *bufmgr) 2843 { 2844 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr; 2845 2846 return bufmgr_gem->has_exec_async; 2847 } 2848 2849 /** 2850 * Enable use of fenced reloc type. 2851 * 2852 * New code should enable this to avoid unnecessary fence register 2853 * allocation. If this option is not enabled, all relocs will have fence 2854 * register allocated. 2855 */ 2856 void 2857 drm_intel_bufmgr_gem_enable_fenced_relocs(drm_intel_bufmgr *bufmgr) 2858 { 2859 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr; 2860 2861 if (bufmgr_gem->bufmgr.bo_exec == drm_intel_gem_bo_exec2) 2862 bufmgr_gem->fenced_relocs = true; 2863 } 2864 2865 /** 2866 * Return the additional aperture space required by the tree of buffer objects 2867 * rooted at bo. 2868 */ 2869 static int 2870 drm_intel_gem_bo_get_aperture_space(drm_intel_bo *bo) 2871 { 2872 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2873 int i; 2874 int total = 0; 2875 2876 if (bo == NULL || bo_gem->included_in_check_aperture) 2877 return 0; 2878 2879 total += bo->size; 2880 bo_gem->included_in_check_aperture = true; 2881 2882 for (i = 0; i < bo_gem->reloc_count; i++) 2883 total += 2884 drm_intel_gem_bo_get_aperture_space(bo_gem-> 2885 reloc_target_info[i].bo); 2886 2887 return total; 2888 } 2889 2890 /** 2891 * Count the number of buffers in this list that need a fence reg 2892 * 2893 * If the count is greater than the number of available regs, we'll have 2894 * to ask the caller to resubmit a batch with fewer tiled buffers. 2895 * 2896 * This function over-counts if the same buffer is used multiple times. 2897 */ 2898 static unsigned int 2899 drm_intel_gem_total_fences(drm_intel_bo ** bo_array, int count) 2900 { 2901 int i; 2902 unsigned int total = 0; 2903 2904 for (i = 0; i < count; i++) { 2905 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo_array[i]; 2906 2907 if (bo_gem == NULL) 2908 continue; 2909 2910 total += bo_gem->reloc_tree_fences; 2911 } 2912 return total; 2913 } 2914 2915 /** 2916 * Clear the flag set by drm_intel_gem_bo_get_aperture_space() so we're ready 2917 * for the next drm_intel_bufmgr_check_aperture_space() call. 2918 */ 2919 static void 2920 drm_intel_gem_bo_clear_aperture_space_flag(drm_intel_bo *bo) 2921 { 2922 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 2923 int i; 2924 2925 if (bo == NULL || !bo_gem->included_in_check_aperture) 2926 return; 2927 2928 bo_gem->included_in_check_aperture = false; 2929 2930 for (i = 0; i < bo_gem->reloc_count; i++) 2931 drm_intel_gem_bo_clear_aperture_space_flag(bo_gem-> 2932 reloc_target_info[i].bo); 2933 } 2934 2935 /** 2936 * Return a conservative estimate for the amount of aperture required 2937 * for a collection of buffers. This may double-count some buffers. 2938 */ 2939 static unsigned int 2940 drm_intel_gem_estimate_batch_space(drm_intel_bo **bo_array, int count) 2941 { 2942 int i; 2943 unsigned int total = 0; 2944 2945 for (i = 0; i < count; i++) { 2946 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo_array[i]; 2947 if (bo_gem != NULL) 2948 total += bo_gem->reloc_tree_size; 2949 } 2950 return total; 2951 } 2952 2953 /** 2954 * Return the amount of aperture needed for a collection of buffers. 2955 * This avoids double counting any buffers, at the cost of looking 2956 * at every buffer in the set. 2957 */ 2958 static unsigned int 2959 drm_intel_gem_compute_batch_space(drm_intel_bo **bo_array, int count) 2960 { 2961 int i; 2962 unsigned int total = 0; 2963 2964 for (i = 0; i < count; i++) { 2965 total += drm_intel_gem_bo_get_aperture_space(bo_array[i]); 2966 /* For the first buffer object in the array, we get an 2967 * accurate count back for its reloc_tree size (since nothing 2968 * had been flagged as being counted yet). We can save that 2969 * value out as a more conservative reloc_tree_size that 2970 * avoids double-counting target buffers. Since the first 2971 * buffer happens to usually be the batch buffer in our 2972 * callers, this can pull us back from doing the tree 2973 * walk on every new batch emit. 2974 */ 2975 if (i == 0) { 2976 drm_intel_bo_gem *bo_gem = 2977 (drm_intel_bo_gem *) bo_array[i]; 2978 bo_gem->reloc_tree_size = total; 2979 } 2980 } 2981 2982 for (i = 0; i < count; i++) 2983 drm_intel_gem_bo_clear_aperture_space_flag(bo_array[i]); 2984 return total; 2985 } 2986 2987 /** 2988 * Return -1 if the batchbuffer should be flushed before attempting to 2989 * emit rendering referencing the buffers pointed to by bo_array. 2990 * 2991 * This is required because if we try to emit a batchbuffer with relocations 2992 * to a tree of buffers that won't simultaneously fit in the aperture, 2993 * the rendering will return an error at a point where the software is not 2994 * prepared to recover from it. 2995 * 2996 * However, we also want to emit the batchbuffer significantly before we reach 2997 * the limit, as a series of batchbuffers each of which references buffers 2998 * covering almost all of the aperture means that at each emit we end up 2999 * waiting to evict a buffer from the last rendering, and we get synchronous 3000 * performance. By emitting smaller batchbuffers, we eat some CPU overhead to 3001 * get better parallelism. 3002 */ 3003 static int 3004 drm_intel_gem_check_aperture_space(drm_intel_bo **bo_array, int count) 3005 { 3006 drm_intel_bufmgr_gem *bufmgr_gem = 3007 (drm_intel_bufmgr_gem *) bo_array[0]->bufmgr; 3008 unsigned int total = 0; 3009 unsigned int threshold = bufmgr_gem->gtt_size * 3 / 4; 3010 int total_fences; 3011 3012 /* Check for fence reg constraints if necessary */ 3013 if (bufmgr_gem->available_fences) { 3014 total_fences = drm_intel_gem_total_fences(bo_array, count); 3015 if (total_fences > bufmgr_gem->available_fences) 3016 return -ENOSPC; 3017 } 3018 3019 total = drm_intel_gem_estimate_batch_space(bo_array, count); 3020 3021 if (total > threshold) 3022 total = drm_intel_gem_compute_batch_space(bo_array, count); 3023 3024 if (total > threshold) { 3025 DBG("check_space: overflowed available aperture, " 3026 "%dkb vs %dkb\n", 3027 total / 1024, (int)bufmgr_gem->gtt_size / 1024); 3028 return -ENOSPC; 3029 } else { 3030 DBG("drm_check_space: total %dkb vs bufgr %dkb\n", total / 1024, 3031 (int)bufmgr_gem->gtt_size / 1024); 3032 return 0; 3033 } 3034 } 3035 3036 /* 3037 * Disable buffer reuse for objects which are shared with the kernel 3038 * as scanout buffers 3039 */ 3040 static int 3041 drm_intel_gem_bo_disable_reuse(drm_intel_bo *bo) 3042 { 3043 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 3044 3045 bo_gem->reusable = false; 3046 return 0; 3047 } 3048 3049 static int 3050 drm_intel_gem_bo_is_reusable(drm_intel_bo *bo) 3051 { 3052 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 3053 3054 return bo_gem->reusable; 3055 } 3056 3057 static int 3058 _drm_intel_gem_bo_references(drm_intel_bo *bo, drm_intel_bo *target_bo) 3059 { 3060 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 3061 int i; 3062 3063 for (i = 0; i < bo_gem->reloc_count; i++) { 3064 if (bo_gem->reloc_target_info[i].bo == target_bo) 3065 return 1; 3066 if (bo == bo_gem->reloc_target_info[i].bo) 3067 continue; 3068 if (_drm_intel_gem_bo_references(bo_gem->reloc_target_info[i].bo, 3069 target_bo)) 3070 return 1; 3071 } 3072 3073 for (i = 0; i< bo_gem->softpin_target_count; i++) { 3074 if (bo_gem->softpin_target[i] == target_bo) 3075 return 1; 3076 if (_drm_intel_gem_bo_references(bo_gem->softpin_target[i], target_bo)) 3077 return 1; 3078 } 3079 3080 return 0; 3081 } 3082 3083 /** Return true if target_bo is referenced by bo's relocation tree. */ 3084 static int 3085 drm_intel_gem_bo_references(drm_intel_bo *bo, drm_intel_bo *target_bo) 3086 { 3087 drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *) target_bo; 3088 3089 if (bo == NULL || target_bo == NULL) 3090 return 0; 3091 if (target_bo_gem->used_as_reloc_target) 3092 return _drm_intel_gem_bo_references(bo, target_bo); 3093 return 0; 3094 } 3095 3096 static void 3097 add_bucket(drm_intel_bufmgr_gem *bufmgr_gem, int size) 3098 { 3099 unsigned int i = bufmgr_gem->num_buckets; 3100 3101 assert(i < ARRAY_SIZE(bufmgr_gem->cache_bucket)); 3102 3103 DRMINITLISTHEAD(&bufmgr_gem->cache_bucket[i].head); 3104 bufmgr_gem->cache_bucket[i].size = size; 3105 bufmgr_gem->num_buckets++; 3106 } 3107 3108 static void 3109 init_cache_buckets(drm_intel_bufmgr_gem *bufmgr_gem) 3110 { 3111 unsigned long size, cache_max_size = 64 * 1024 * 1024; 3112 3113 /* OK, so power of two buckets was too wasteful of memory. 3114 * Give 3 other sizes between each power of two, to hopefully 3115 * cover things accurately enough. (The alternative is 3116 * probably to just go for exact matching of sizes, and assume 3117 * that for things like composited window resize the tiled 3118 * width/height alignment and rounding of sizes to pages will 3119 * get us useful cache hit rates anyway) 3120 */ 3121 add_bucket(bufmgr_gem, 4096); 3122 add_bucket(bufmgr_gem, 4096 * 2); 3123 add_bucket(bufmgr_gem, 4096 * 3); 3124 3125 /* Initialize the linked lists for BO reuse cache. */ 3126 for (size = 4 * 4096; size <= cache_max_size; size *= 2) { 3127 add_bucket(bufmgr_gem, size); 3128 3129 add_bucket(bufmgr_gem, size + size * 1 / 4); 3130 add_bucket(bufmgr_gem, size + size * 2 / 4); 3131 add_bucket(bufmgr_gem, size + size * 3 / 4); 3132 } 3133 } 3134 3135 void 3136 drm_intel_bufmgr_gem_set_vma_cache_size(drm_intel_bufmgr *bufmgr, int limit) 3137 { 3138 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr; 3139 3140 bufmgr_gem->vma_max = limit; 3141 3142 drm_intel_gem_bo_purge_vma_cache(bufmgr_gem); 3143 } 3144 3145 static int 3146 parse_devid_override(const char *devid_override) 3147 { 3148 static const struct { 3149 const char *name; 3150 int pci_id; 3151 } name_map[] = { 3152 { "brw", PCI_CHIP_I965_GM }, 3153 { "g4x", PCI_CHIP_GM45_GM }, 3154 { "ilk", PCI_CHIP_ILD_G }, 3155 { "snb", PCI_CHIP_SANDYBRIDGE_M_GT2_PLUS }, 3156 { "ivb", PCI_CHIP_IVYBRIDGE_S_GT2 }, 3157 { "hsw", PCI_CHIP_HASWELL_CRW_E_GT3 }, 3158 { "byt", PCI_CHIP_VALLEYVIEW_3 }, 3159 { "bdw", 0x1620 | BDW_ULX }, 3160 { "skl", PCI_CHIP_SKYLAKE_DT_GT2 }, 3161 { "kbl", PCI_CHIP_KABYLAKE_DT_GT2 }, 3162 }; 3163 unsigned int i; 3164 3165 for (i = 0; i < ARRAY_SIZE(name_map); i++) { 3166 if (!strcmp(name_map[i].name, devid_override)) 3167 return name_map[i].pci_id; 3168 } 3169 3170 return strtod(devid_override, NULL); 3171 } 3172 3173 /** 3174 * Get the PCI ID for the device. This can be overridden by setting the 3175 * INTEL_DEVID_OVERRIDE environment variable to the desired ID. 3176 */ 3177 static int 3178 get_pci_device_id(drm_intel_bufmgr_gem *bufmgr_gem) 3179 { 3180 char *devid_override; 3181 int devid = 0; 3182 int ret; 3183 drm_i915_getparam_t gp; 3184 3185 if (geteuid() == getuid()) { 3186 devid_override = getenv("INTEL_DEVID_OVERRIDE"); 3187 if (devid_override) { 3188 bufmgr_gem->no_exec = true; 3189 return parse_devid_override(devid_override); 3190 } 3191 } 3192 3193 memclear(gp); 3194 gp.param = I915_PARAM_CHIPSET_ID; 3195 gp.value = &devid; 3196 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3197 if (ret) { 3198 fprintf(stderr, "get chip id failed: %d [%d]\n", ret, errno); 3199 fprintf(stderr, "param: %d, val: %d\n", gp.param, *gp.value); 3200 } 3201 return devid; 3202 } 3203 3204 int 3205 drm_intel_bufmgr_gem_get_devid(drm_intel_bufmgr *bufmgr) 3206 { 3207 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr; 3208 3209 return bufmgr_gem->pci_device; 3210 } 3211 3212 /** 3213 * Sets the AUB filename. 3214 * 3215 * This function has to be called before drm_intel_bufmgr_gem_set_aub_dump() 3216 * for it to have any effect. 3217 */ 3218 void 3219 drm_intel_bufmgr_gem_set_aub_filename(drm_intel_bufmgr *bufmgr, 3220 const char *filename) 3221 { 3222 } 3223 3224 /** 3225 * Sets up AUB dumping. 3226 * 3227 * This is a trace file format that can be used with the simulator. 3228 * Packets are emitted in a format somewhat like GPU command packets. 3229 * You can set up a GTT and upload your objects into the referenced 3230 * space, then send off batchbuffers and get BMPs out the other end. 3231 */ 3232 void 3233 drm_intel_bufmgr_gem_set_aub_dump(drm_intel_bufmgr *bufmgr, int enable) 3234 { 3235 fprintf(stderr, "libdrm aub dumping is deprecated.\n\n" 3236 "Use intel_aubdump from intel-gpu-tools instead. Install intel-gpu-tools,\n" 3237 "then run (for example)\n\n" 3238 "\t$ intel_aubdump --output=trace.aub glxgears -geometry 500x500\n\n" 3239 "See the intel_aubdump man page for more details.\n"); 3240 } 3241 3242 drm_intel_context * 3243 drm_intel_gem_context_create(drm_intel_bufmgr *bufmgr) 3244 { 3245 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr; 3246 struct drm_i915_gem_context_create create; 3247 drm_intel_context *context = NULL; 3248 int ret; 3249 3250 context = calloc(1, sizeof(*context)); 3251 if (!context) 3252 return NULL; 3253 3254 memclear(create); 3255 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_CONTEXT_CREATE, &create); 3256 if (ret != 0) { 3257 DBG("DRM_IOCTL_I915_GEM_CONTEXT_CREATE failed: %s\n", 3258 strerror(errno)); 3259 free(context); 3260 return NULL; 3261 } 3262 3263 context->ctx_id = create.ctx_id; 3264 context->bufmgr = bufmgr; 3265 3266 return context; 3267 } 3268 3269 int 3270 drm_intel_gem_context_get_id(drm_intel_context *ctx, uint32_t *ctx_id) 3271 { 3272 if (ctx == NULL) 3273 return -EINVAL; 3274 3275 *ctx_id = ctx->ctx_id; 3276 3277 return 0; 3278 } 3279 3280 void 3281 drm_intel_gem_context_destroy(drm_intel_context *ctx) 3282 { 3283 drm_intel_bufmgr_gem *bufmgr_gem; 3284 struct drm_i915_gem_context_destroy destroy; 3285 int ret; 3286 3287 if (ctx == NULL) 3288 return; 3289 3290 memclear(destroy); 3291 3292 bufmgr_gem = (drm_intel_bufmgr_gem *)ctx->bufmgr; 3293 destroy.ctx_id = ctx->ctx_id; 3294 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_CONTEXT_DESTROY, 3295 &destroy); 3296 if (ret != 0) 3297 fprintf(stderr, "DRM_IOCTL_I915_GEM_CONTEXT_DESTROY failed: %s\n", 3298 strerror(errno)); 3299 3300 free(ctx); 3301 } 3302 3303 int 3304 drm_intel_get_reset_stats(drm_intel_context *ctx, 3305 uint32_t *reset_count, 3306 uint32_t *active, 3307 uint32_t *pending) 3308 { 3309 drm_intel_bufmgr_gem *bufmgr_gem; 3310 struct drm_i915_reset_stats stats; 3311 int ret; 3312 3313 if (ctx == NULL) 3314 return -EINVAL; 3315 3316 memclear(stats); 3317 3318 bufmgr_gem = (drm_intel_bufmgr_gem *)ctx->bufmgr; 3319 stats.ctx_id = ctx->ctx_id; 3320 ret = drmIoctl(bufmgr_gem->fd, 3321 DRM_IOCTL_I915_GET_RESET_STATS, 3322 &stats); 3323 if (ret == 0) { 3324 if (reset_count != NULL) 3325 *reset_count = stats.reset_count; 3326 3327 if (active != NULL) 3328 *active = stats.batch_active; 3329 3330 if (pending != NULL) 3331 *pending = stats.batch_pending; 3332 } 3333 3334 return ret; 3335 } 3336 3337 int 3338 drm_intel_reg_read(drm_intel_bufmgr *bufmgr, 3339 uint32_t offset, 3340 uint64_t *result) 3341 { 3342 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr; 3343 struct drm_i915_reg_read reg_read; 3344 int ret; 3345 3346 memclear(reg_read); 3347 reg_read.offset = offset; 3348 3349 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_REG_READ, ®_read); 3350 3351 *result = reg_read.val; 3352 return ret; 3353 } 3354 3355 int 3356 drm_intel_get_subslice_total(int fd, unsigned int *subslice_total) 3357 { 3358 drm_i915_getparam_t gp; 3359 int ret; 3360 3361 memclear(gp); 3362 gp.value = (int*)subslice_total; 3363 gp.param = I915_PARAM_SUBSLICE_TOTAL; 3364 ret = drmIoctl(fd, DRM_IOCTL_I915_GETPARAM, &gp); 3365 if (ret) 3366 return -errno; 3367 3368 return 0; 3369 } 3370 3371 int 3372 drm_intel_get_eu_total(int fd, unsigned int *eu_total) 3373 { 3374 drm_i915_getparam_t gp; 3375 int ret; 3376 3377 memclear(gp); 3378 gp.value = (int*)eu_total; 3379 gp.param = I915_PARAM_EU_TOTAL; 3380 ret = drmIoctl(fd, DRM_IOCTL_I915_GETPARAM, &gp); 3381 if (ret) 3382 return -errno; 3383 3384 return 0; 3385 } 3386 3387 int 3388 drm_intel_get_pooled_eu(int fd) 3389 { 3390 drm_i915_getparam_t gp; 3391 int ret = -1; 3392 3393 memclear(gp); 3394 gp.param = I915_PARAM_HAS_POOLED_EU; 3395 gp.value = &ret; 3396 if (drmIoctl(fd, DRM_IOCTL_I915_GETPARAM, &gp)) 3397 return -errno; 3398 3399 return ret; 3400 } 3401 3402 int 3403 drm_intel_get_min_eu_in_pool(int fd) 3404 { 3405 drm_i915_getparam_t gp; 3406 int ret = -1; 3407 3408 memclear(gp); 3409 gp.param = I915_PARAM_MIN_EU_IN_POOL; 3410 gp.value = &ret; 3411 if (drmIoctl(fd, DRM_IOCTL_I915_GETPARAM, &gp)) 3412 return -errno; 3413 3414 return ret; 3415 } 3416 3417 /** 3418 * Annotate the given bo for use in aub dumping. 3419 * 3420 * \param annotations is an array of drm_intel_aub_annotation objects 3421 * describing the type of data in various sections of the bo. Each 3422 * element of the array specifies the type and subtype of a section of 3423 * the bo, and the past-the-end offset of that section. The elements 3424 * of \c annotations must be sorted so that ending_offset is 3425 * increasing. 3426 * 3427 * \param count is the number of elements in the \c annotations array. 3428 * If \c count is zero, then \c annotations will not be dereferenced. 3429 * 3430 * Annotations are copied into a private data structure, so caller may 3431 * re-use the memory pointed to by \c annotations after the call 3432 * returns. 3433 * 3434 * Annotations are stored for the lifetime of the bo; to reset to the 3435 * default state (no annotations), call this function with a \c count 3436 * of zero. 3437 */ 3438 void 3439 drm_intel_bufmgr_gem_set_aub_annotations(drm_intel_bo *bo, 3440 drm_intel_aub_annotation *annotations, 3441 unsigned count) 3442 { 3443 } 3444 3445 static pthread_mutex_t bufmgr_list_mutex = PTHREAD_MUTEX_INITIALIZER; 3446 static drmMMListHead bufmgr_list = { &bufmgr_list, &bufmgr_list }; 3447 3448 static drm_intel_bufmgr_gem * 3449 drm_intel_bufmgr_gem_find(int fd) 3450 { 3451 drm_intel_bufmgr_gem *bufmgr_gem; 3452 3453 DRMLISTFOREACHENTRY(bufmgr_gem, &bufmgr_list, managers) { 3454 if (bufmgr_gem->fd == fd) { 3455 atomic_inc(&bufmgr_gem->refcount); 3456 return bufmgr_gem; 3457 } 3458 } 3459 3460 return NULL; 3461 } 3462 3463 static void 3464 drm_intel_bufmgr_gem_unref(drm_intel_bufmgr *bufmgr) 3465 { 3466 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr; 3467 3468 if (atomic_add_unless(&bufmgr_gem->refcount, -1, 1)) { 3469 pthread_mutex_lock(&bufmgr_list_mutex); 3470 3471 if (atomic_dec_and_test(&bufmgr_gem->refcount)) { 3472 DRMLISTDEL(&bufmgr_gem->managers); 3473 drm_intel_bufmgr_gem_destroy(bufmgr); 3474 } 3475 3476 pthread_mutex_unlock(&bufmgr_list_mutex); 3477 } 3478 } 3479 3480 void *drm_intel_gem_bo_map__gtt(drm_intel_bo *bo) 3481 { 3482 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 3483 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 3484 3485 if (bo_gem->gtt_virtual) 3486 return bo_gem->gtt_virtual; 3487 3488 if (bo_gem->is_userptr) 3489 return NULL; 3490 3491 pthread_mutex_lock(&bufmgr_gem->lock); 3492 if (bo_gem->gtt_virtual == NULL) { 3493 struct drm_i915_gem_mmap_gtt mmap_arg; 3494 void *ptr; 3495 3496 DBG("bo_map_gtt: mmap %d (%s), map_count=%d\n", 3497 bo_gem->gem_handle, bo_gem->name, bo_gem->map_count); 3498 3499 if (bo_gem->map_count++ == 0) 3500 drm_intel_gem_bo_open_vma(bufmgr_gem, bo_gem); 3501 3502 memclear(mmap_arg); 3503 mmap_arg.handle = bo_gem->gem_handle; 3504 3505 /* Get the fake offset back... */ 3506 ptr = MAP_FAILED; 3507 if (drmIoctl(bufmgr_gem->fd, 3508 DRM_IOCTL_I915_GEM_MMAP_GTT, 3509 &mmap_arg) == 0) { 3510 /* and mmap it */ 3511 ptr = drm_mmap(0, bo->size, PROT_READ | PROT_WRITE, 3512 MAP_SHARED, bufmgr_gem->fd, 3513 mmap_arg.offset); 3514 } 3515 if (ptr == MAP_FAILED) { 3516 if (--bo_gem->map_count == 0) 3517 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem); 3518 ptr = NULL; 3519 } 3520 3521 bo_gem->gtt_virtual = ptr; 3522 } 3523 pthread_mutex_unlock(&bufmgr_gem->lock); 3524 3525 return bo_gem->gtt_virtual; 3526 } 3527 3528 void *drm_intel_gem_bo_map__cpu(drm_intel_bo *bo) 3529 { 3530 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 3531 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 3532 3533 if (bo_gem->mem_virtual) 3534 return bo_gem->mem_virtual; 3535 3536 if (bo_gem->is_userptr) { 3537 /* Return the same user ptr */ 3538 return bo_gem->user_virtual; 3539 } 3540 3541 pthread_mutex_lock(&bufmgr_gem->lock); 3542 if (!bo_gem->mem_virtual) { 3543 struct drm_i915_gem_mmap mmap_arg; 3544 3545 if (bo_gem->map_count++ == 0) 3546 drm_intel_gem_bo_open_vma(bufmgr_gem, bo_gem); 3547 3548 DBG("bo_map: %d (%s), map_count=%d\n", 3549 bo_gem->gem_handle, bo_gem->name, bo_gem->map_count); 3550 3551 memclear(mmap_arg); 3552 mmap_arg.handle = bo_gem->gem_handle; 3553 mmap_arg.size = bo->size; 3554 if (drmIoctl(bufmgr_gem->fd, 3555 DRM_IOCTL_I915_GEM_MMAP, 3556 &mmap_arg)) { 3557 DBG("%s:%d: Error mapping buffer %d (%s): %s .\n", 3558 __FILE__, __LINE__, bo_gem->gem_handle, 3559 bo_gem->name, strerror(errno)); 3560 if (--bo_gem->map_count == 0) 3561 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem); 3562 } else { 3563 VG(VALGRIND_MALLOCLIKE_BLOCK(mmap_arg.addr_ptr, mmap_arg.size, 0, 1)); 3564 bo_gem->mem_virtual = (void *)(uintptr_t) mmap_arg.addr_ptr; 3565 } 3566 } 3567 pthread_mutex_unlock(&bufmgr_gem->lock); 3568 3569 return bo_gem->mem_virtual; 3570 } 3571 3572 void *drm_intel_gem_bo_map__wc(drm_intel_bo *bo) 3573 { 3574 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr; 3575 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo; 3576 3577 if (bo_gem->wc_virtual) 3578 return bo_gem->wc_virtual; 3579 3580 if (bo_gem->is_userptr) 3581 return NULL; 3582 3583 pthread_mutex_lock(&bufmgr_gem->lock); 3584 if (!bo_gem->wc_virtual) { 3585 struct drm_i915_gem_mmap mmap_arg; 3586 3587 if (bo_gem->map_count++ == 0) 3588 drm_intel_gem_bo_open_vma(bufmgr_gem, bo_gem); 3589 3590 DBG("bo_map: %d (%s), map_count=%d\n", 3591 bo_gem->gem_handle, bo_gem->name, bo_gem->map_count); 3592 3593 memclear(mmap_arg); 3594 mmap_arg.handle = bo_gem->gem_handle; 3595 mmap_arg.size = bo->size; 3596 mmap_arg.flags = I915_MMAP_WC; 3597 if (drmIoctl(bufmgr_gem->fd, 3598 DRM_IOCTL_I915_GEM_MMAP, 3599 &mmap_arg)) { 3600 DBG("%s:%d: Error mapping buffer %d (%s): %s .\n", 3601 __FILE__, __LINE__, bo_gem->gem_handle, 3602 bo_gem->name, strerror(errno)); 3603 if (--bo_gem->map_count == 0) 3604 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem); 3605 } else { 3606 VG(VALGRIND_MALLOCLIKE_BLOCK(mmap_arg.addr_ptr, mmap_arg.size, 0, 1)); 3607 bo_gem->wc_virtual = (void *)(uintptr_t) mmap_arg.addr_ptr; 3608 } 3609 } 3610 pthread_mutex_unlock(&bufmgr_gem->lock); 3611 3612 return bo_gem->wc_virtual; 3613 } 3614 3615 /** 3616 * Initializes the GEM buffer manager, which uses the kernel to allocate, map, 3617 * and manage map buffer objections. 3618 * 3619 * \param fd File descriptor of the opened DRM device. 3620 */ 3621 drm_intel_bufmgr * 3622 drm_intel_bufmgr_gem_init(int fd, int batch_size) 3623 { 3624 drm_intel_bufmgr_gem *bufmgr_gem; 3625 struct drm_i915_gem_get_aperture aperture; 3626 drm_i915_getparam_t gp; 3627 int ret, tmp; 3628 bool exec2 = false; 3629 3630 pthread_mutex_lock(&bufmgr_list_mutex); 3631 3632 bufmgr_gem = drm_intel_bufmgr_gem_find(fd); 3633 if (bufmgr_gem) 3634 goto exit; 3635 3636 bufmgr_gem = calloc(1, sizeof(*bufmgr_gem)); 3637 if (bufmgr_gem == NULL) 3638 goto exit; 3639 3640 bufmgr_gem->fd = fd; 3641 atomic_set(&bufmgr_gem->refcount, 1); 3642 3643 if (pthread_mutex_init(&bufmgr_gem->lock, NULL) != 0) { 3644 free(bufmgr_gem); 3645 bufmgr_gem = NULL; 3646 goto exit; 3647 } 3648 3649 memclear(aperture); 3650 ret = drmIoctl(bufmgr_gem->fd, 3651 DRM_IOCTL_I915_GEM_GET_APERTURE, 3652 &aperture); 3653 3654 if (ret == 0) 3655 bufmgr_gem->gtt_size = aperture.aper_available_size; 3656 else { 3657 fprintf(stderr, "DRM_IOCTL_I915_GEM_APERTURE failed: %s\n", 3658 strerror(errno)); 3659 bufmgr_gem->gtt_size = 128 * 1024 * 1024; 3660 fprintf(stderr, "Assuming %dkB available aperture size.\n" 3661 "May lead to reduced performance or incorrect " 3662 "rendering.\n", 3663 (int)bufmgr_gem->gtt_size / 1024); 3664 } 3665 3666 bufmgr_gem->pci_device = get_pci_device_id(bufmgr_gem); 3667 3668 if (IS_GEN2(bufmgr_gem->pci_device)) 3669 bufmgr_gem->gen = 2; 3670 else if (IS_GEN3(bufmgr_gem->pci_device)) 3671 bufmgr_gem->gen = 3; 3672 else if (IS_GEN4(bufmgr_gem->pci_device)) 3673 bufmgr_gem->gen = 4; 3674 else if (IS_GEN5(bufmgr_gem->pci_device)) 3675 bufmgr_gem->gen = 5; 3676 else if (IS_GEN6(bufmgr_gem->pci_device)) 3677 bufmgr_gem->gen = 6; 3678 else if (IS_GEN7(bufmgr_gem->pci_device)) 3679 bufmgr_gem->gen = 7; 3680 else if (IS_GEN8(bufmgr_gem->pci_device)) 3681 bufmgr_gem->gen = 8; 3682 else if (IS_GEN9(bufmgr_gem->pci_device)) 3683 bufmgr_gem->gen = 9; 3684 else { 3685 free(bufmgr_gem); 3686 bufmgr_gem = NULL; 3687 goto exit; 3688 } 3689 3690 if (IS_GEN3(bufmgr_gem->pci_device) && 3691 bufmgr_gem->gtt_size > 256*1024*1024) { 3692 /* The unmappable part of gtt on gen 3 (i.e. above 256MB) can't 3693 * be used for tiled blits. To simplify the accounting, just 3694 * subtract the unmappable part (fixed to 256MB on all known 3695 * gen3 devices) if the kernel advertises it. */ 3696 bufmgr_gem->gtt_size -= 256*1024*1024; 3697 } 3698 3699 memclear(gp); 3700 gp.value = &tmp; 3701 3702 gp.param = I915_PARAM_HAS_EXECBUF2; 3703 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3704 if (!ret) 3705 exec2 = true; 3706 3707 gp.param = I915_PARAM_HAS_BSD; 3708 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3709 bufmgr_gem->has_bsd = ret == 0; 3710 3711 gp.param = I915_PARAM_HAS_BLT; 3712 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3713 bufmgr_gem->has_blt = ret == 0; 3714 3715 gp.param = I915_PARAM_HAS_RELAXED_FENCING; 3716 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3717 bufmgr_gem->has_relaxed_fencing = ret == 0; 3718 3719 gp.param = I915_PARAM_HAS_EXEC_ASYNC; 3720 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3721 bufmgr_gem->has_exec_async = ret == 0; 3722 3723 bufmgr_gem->bufmgr.bo_alloc_userptr = check_bo_alloc_userptr; 3724 3725 gp.param = I915_PARAM_HAS_WAIT_TIMEOUT; 3726 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3727 bufmgr_gem->has_wait_timeout = ret == 0; 3728 3729 gp.param = I915_PARAM_HAS_LLC; 3730 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3731 if (ret != 0) { 3732 /* Kernel does not supports HAS_LLC query, fallback to GPU 3733 * generation detection and assume that we have LLC on GEN6/7 3734 */ 3735 bufmgr_gem->has_llc = (IS_GEN6(bufmgr_gem->pci_device) | 3736 IS_GEN7(bufmgr_gem->pci_device)); 3737 } else 3738 bufmgr_gem->has_llc = *gp.value; 3739 3740 gp.param = I915_PARAM_HAS_VEBOX; 3741 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3742 bufmgr_gem->has_vebox = (ret == 0) & (*gp.value > 0); 3743 3744 gp.param = I915_PARAM_HAS_EXEC_SOFTPIN; 3745 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3746 if (ret == 0 && *gp.value > 0) 3747 bufmgr_gem->bufmgr.bo_set_softpin_offset = drm_intel_gem_bo_set_softpin_offset; 3748 3749 if (bufmgr_gem->gen < 4) { 3750 gp.param = I915_PARAM_NUM_FENCES_AVAIL; 3751 gp.value = &bufmgr_gem->available_fences; 3752 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3753 if (ret) { 3754 fprintf(stderr, "get fences failed: %d [%d]\n", ret, 3755 errno); 3756 fprintf(stderr, "param: %d, val: %d\n", gp.param, 3757 *gp.value); 3758 bufmgr_gem->available_fences = 0; 3759 } else { 3760 /* XXX The kernel reports the total number of fences, 3761 * including any that may be pinned. 3762 * 3763 * We presume that there will be at least one pinned 3764 * fence for the scanout buffer, but there may be more 3765 * than one scanout and the user may be manually 3766 * pinning buffers. Let's move to execbuffer2 and 3767 * thereby forget the insanity of using fences... 3768 */ 3769 bufmgr_gem->available_fences -= 2; 3770 if (bufmgr_gem->available_fences < 0) 3771 bufmgr_gem->available_fences = 0; 3772 } 3773 } 3774 3775 if (bufmgr_gem->gen >= 8) { 3776 gp.param = I915_PARAM_HAS_ALIASING_PPGTT; 3777 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp); 3778 if (ret == 0 && *gp.value == 3) 3779 bufmgr_gem->bufmgr.bo_use_48b_address_range = drm_intel_gem_bo_use_48b_address_range; 3780 } 3781 3782 /* Let's go with one relocation per every 2 dwords (but round down a bit 3783 * since a power of two will mean an extra page allocation for the reloc 3784 * buffer). 3785 * 3786 * Every 4 was too few for the blender benchmark. 3787 */ 3788 bufmgr_gem->max_relocs = batch_size / sizeof(uint32_t) / 2 - 2; 3789 3790 bufmgr_gem->bufmgr.bo_alloc = drm_intel_gem_bo_alloc; 3791 bufmgr_gem->bufmgr.bo_alloc_for_render = 3792 drm_intel_gem_bo_alloc_for_render; 3793 bufmgr_gem->bufmgr.bo_alloc_tiled = drm_intel_gem_bo_alloc_tiled; 3794 bufmgr_gem->bufmgr.bo_reference = drm_intel_gem_bo_reference; 3795 bufmgr_gem->bufmgr.bo_unreference = drm_intel_gem_bo_unreference; 3796 bufmgr_gem->bufmgr.bo_map = drm_intel_gem_bo_map; 3797 bufmgr_gem->bufmgr.bo_unmap = drm_intel_gem_bo_unmap; 3798 bufmgr_gem->bufmgr.bo_subdata = drm_intel_gem_bo_subdata; 3799 bufmgr_gem->bufmgr.bo_get_subdata = drm_intel_gem_bo_get_subdata; 3800 bufmgr_gem->bufmgr.bo_wait_rendering = drm_intel_gem_bo_wait_rendering; 3801 bufmgr_gem->bufmgr.bo_emit_reloc = drm_intel_gem_bo_emit_reloc; 3802 bufmgr_gem->bufmgr.bo_emit_reloc_fence = drm_intel_gem_bo_emit_reloc_fence; 3803 bufmgr_gem->bufmgr.bo_pin = drm_intel_gem_bo_pin; 3804 bufmgr_gem->bufmgr.bo_unpin = drm_intel_gem_bo_unpin; 3805 bufmgr_gem->bufmgr.bo_get_tiling = drm_intel_gem_bo_get_tiling; 3806 bufmgr_gem->bufmgr.bo_set_tiling = drm_intel_gem_bo_set_tiling; 3807 bufmgr_gem->bufmgr.bo_flink = drm_intel_gem_bo_flink; 3808 /* Use the new one if available */ 3809 if (exec2) { 3810 bufmgr_gem->bufmgr.bo_exec = drm_intel_gem_bo_exec2; 3811 bufmgr_gem->bufmgr.bo_mrb_exec = drm_intel_gem_bo_mrb_exec2; 3812 } else 3813 bufmgr_gem->bufmgr.bo_exec = drm_intel_gem_bo_exec; 3814 bufmgr_gem->bufmgr.bo_busy = drm_intel_gem_bo_busy; 3815 bufmgr_gem->bufmgr.bo_madvise = drm_intel_gem_bo_madvise; 3816 bufmgr_gem->bufmgr.destroy = drm_intel_bufmgr_gem_unref; 3817 bufmgr_gem->bufmgr.debug = 0; 3818 bufmgr_gem->bufmgr.check_aperture_space = 3819 drm_intel_gem_check_aperture_space; 3820 bufmgr_gem->bufmgr.bo_disable_reuse = drm_intel_gem_bo_disable_reuse; 3821 bufmgr_gem->bufmgr.bo_is_reusable = drm_intel_gem_bo_is_reusable; 3822 bufmgr_gem->bufmgr.get_pipe_from_crtc_id = 3823 drm_intel_gem_get_pipe_from_crtc_id; 3824 bufmgr_gem->bufmgr.bo_references = drm_intel_gem_bo_references; 3825 3826 init_cache_buckets(bufmgr_gem); 3827 3828 DRMINITLISTHEAD(&bufmgr_gem->vma_cache); 3829 bufmgr_gem->vma_max = -1; /* unlimited by default */ 3830 3831 DRMLISTADD(&bufmgr_gem->managers, &bufmgr_list); 3832 3833 exit: 3834 pthread_mutex_unlock(&bufmgr_list_mutex); 3835 3836 return bufmgr_gem != NULL ? &bufmgr_gem->bufmgr : NULL; 3837 } 3838