1 /**************************************************************************
3 * Copyright © 2007 Red Hat Inc.
4 * Copyright © 2007-2012 Intel Corporation
5 * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA
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:
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.
24 * The above copyright notice and this permission notice (including the
25 * next paragraph) shall be included in all copies or substantial portions
29 **************************************************************************/
31 * Authors: Thomas Hellström <thomas-at-tungstengraphics-dot-com>
32 * Keith Whitwell <keithw-at-tungstengraphics-dot-com>
33 * Eric Anholt <eric@anholt.net>
34 * Dave Airlie <airlied@linux.ie>
42 #include <xf86atomic.h>
50 #include <sys/ioctl.h>
53 #include <sys/types.h>
58 #define ETIME ETIMEDOUT
60 #include "libdrm_lists.h"
61 #include "intel_bufmgr.h"
62 #include "intel_bufmgr_priv.h"
63 #include "intel_chipset.h"
64 #include "intel_aub.h"
77 #define VG_CLEAR(s) VG(memset(&s, 0, sizeof(s)))
79 #define DBG(...) do { \
80 if (bufmgr_gem->bufmgr.debug) \
81 fprintf(stderr, __VA_ARGS__); \
84 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
86 typedef struct _drm_intel_bo_gem drm_intel_bo_gem;
88 struct drm_intel_gem_bo_bucket {
93 typedef struct _drm_intel_bufmgr_gem {
94 drm_intel_bufmgr bufmgr;
100 pthread_mutex_t lock;
102 struct drm_i915_gem_exec_object *exec_objects;
103 struct drm_i915_gem_exec_object2 *exec2_objects;
104 drm_intel_bo **exec_bos;
108 /** Array of lists of cached gem objects of power-of-two sizes */
109 struct drm_intel_gem_bo_bucket cache_bucket[14 * 4];
114 drmMMListHead vma_cache;
115 int vma_count, vma_open, vma_max;
118 int available_fences;
121 unsigned int has_bsd : 1;
122 unsigned int has_blt : 1;
123 unsigned int has_relaxed_fencing : 1;
124 unsigned int has_llc : 1;
125 unsigned int has_wait_timeout : 1;
126 unsigned int bo_reuse : 1;
127 unsigned int no_exec : 1;
132 } drm_intel_bufmgr_gem;
134 #define DRM_INTEL_RELOC_FENCE (1<<0)
136 typedef struct _drm_intel_reloc_target_info {
139 } drm_intel_reloc_target;
141 struct _drm_intel_bo_gem {
149 * Kenel-assigned global name for this object
151 unsigned int global_name;
152 drmMMListHead name_list;
155 * Index of the buffer within the validation list while preparing a
156 * batchbuffer execution.
161 * Current tiling mode
163 uint32_t tiling_mode;
164 uint32_t swizzle_mode;
165 unsigned long stride;
169 /** Array passed to the DRM containing relocation information. */
170 struct drm_i915_gem_relocation_entry *relocs;
172 * Array of info structs corresponding to relocs[i].target_handle etc
174 drm_intel_reloc_target *reloc_target_info;
175 /** Number of entries in relocs */
177 /** Mapped address for the buffer, saved across map/unmap cycles */
179 /** GTT virtual address for the buffer, saved across map/unmap cycles */
182 drmMMListHead vma_list;
188 * Boolean of whether this BO and its children have been included in
189 * the current drm_intel_bufmgr_check_aperture_space() total.
191 bool included_in_check_aperture;
194 * Boolean of whether this buffer has been used as a relocation
195 * target and had its size accounted for, and thus can't have any
196 * further relocations added to it.
198 bool used_as_reloc_target;
201 * Boolean of whether we have encountered an error whilst building the relocation tree.
206 * Boolean of whether this buffer can be re-used
211 * Size in bytes of this buffer and its relocation descendents.
213 * Used to avoid costly tree walking in
214 * drm_intel_bufmgr_check_aperture in the common case.
219 * Number of potential fence registers required by this buffer and its
222 int reloc_tree_fences;
224 /** Flags that we may need to do the SW_FINSIH ioctl on unmap. */
225 bool mapped_cpu_write;
229 drm_intel_aub_annotation *aub_annotations;
230 unsigned aub_annotation_count;
234 drm_intel_gem_estimate_batch_space(drm_intel_bo ** bo_array, int count);
237 drm_intel_gem_compute_batch_space(drm_intel_bo ** bo_array, int count);
240 drm_intel_gem_bo_get_tiling(drm_intel_bo *bo, uint32_t * tiling_mode,
241 uint32_t * swizzle_mode);
244 drm_intel_gem_bo_set_tiling_internal(drm_intel_bo *bo,
245 uint32_t tiling_mode,
248 static void drm_intel_gem_bo_unreference_locked_timed(drm_intel_bo *bo,
251 static void drm_intel_gem_bo_unreference(drm_intel_bo *bo);
253 static void drm_intel_gem_bo_free(drm_intel_bo *bo);
256 drm_intel_gem_bo_tile_size(drm_intel_bufmgr_gem *bufmgr_gem, unsigned long size,
257 uint32_t *tiling_mode)
259 unsigned long min_size, max_size;
262 if (*tiling_mode == I915_TILING_NONE)
265 /* 965+ just need multiples of page size for tiling */
266 if (bufmgr_gem->gen >= 4)
267 return ROUND_UP_TO(size, 4096);
269 /* Older chips need powers of two, of at least 512k or 1M */
270 if (bufmgr_gem->gen == 3) {
271 min_size = 1024*1024;
272 max_size = 128*1024*1024;
275 max_size = 64*1024*1024;
278 if (size > max_size) {
279 *tiling_mode = I915_TILING_NONE;
283 /* Do we need to allocate every page for the fence? */
284 if (bufmgr_gem->has_relaxed_fencing)
285 return ROUND_UP_TO(size, 4096);
287 for (i = min_size; i < size; i <<= 1)
294 * Round a given pitch up to the minimum required for X tiling on a
295 * given chip. We use 512 as the minimum to allow for a later tiling
299 drm_intel_gem_bo_tile_pitch(drm_intel_bufmgr_gem *bufmgr_gem,
300 unsigned long pitch, uint32_t *tiling_mode)
302 unsigned long tile_width;
305 /* If untiled, then just align it so that we can do rendering
306 * to it with the 3D engine.
308 if (*tiling_mode == I915_TILING_NONE)
309 return ALIGN(pitch, 64);
311 if (*tiling_mode == I915_TILING_X
312 || (IS_915(bufmgr_gem->pci_device)
313 && *tiling_mode == I915_TILING_Y))
318 /* 965 is flexible */
319 if (bufmgr_gem->gen >= 4)
320 return ROUND_UP_TO(pitch, tile_width);
322 /* The older hardware has a maximum pitch of 8192 with tiled
323 * surfaces, so fallback to untiled if it's too large.
326 *tiling_mode = I915_TILING_NONE;
327 return ALIGN(pitch, 64);
330 /* Pre-965 needs power of two tile width */
331 for (i = tile_width; i < pitch; i <<= 1)
337 static struct drm_intel_gem_bo_bucket *
338 drm_intel_gem_bo_bucket_for_size(drm_intel_bufmgr_gem *bufmgr_gem,
343 for (i = 0; i < bufmgr_gem->num_buckets; i++) {
344 struct drm_intel_gem_bo_bucket *bucket =
345 &bufmgr_gem->cache_bucket[i];
346 if (bucket->size >= size) {
355 drm_intel_gem_dump_validation_list(drm_intel_bufmgr_gem *bufmgr_gem)
359 for (i = 0; i < bufmgr_gem->exec_count; i++) {
360 drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
361 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
363 if (bo_gem->relocs == NULL) {
364 DBG("%2d: %d (%s)\n", i, bo_gem->gem_handle,
369 for (j = 0; j < bo_gem->reloc_count; j++) {
370 drm_intel_bo *target_bo = bo_gem->reloc_target_info[j].bo;
371 drm_intel_bo_gem *target_gem =
372 (drm_intel_bo_gem *) target_bo;
374 DBG("%2d: %d (%s)@0x%08llx -> "
375 "%d (%s)@0x%08lx + 0x%08x\n",
377 bo_gem->gem_handle, bo_gem->name,
378 (unsigned long long)bo_gem->relocs[j].offset,
379 target_gem->gem_handle,
382 bo_gem->relocs[j].delta);
388 drm_intel_gem_bo_reference(drm_intel_bo *bo)
390 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
392 atomic_inc(&bo_gem->refcount);
396 * Adds the given buffer to the list of buffers to be validated (moved into the
397 * appropriate memory type) with the next batch submission.
399 * If a buffer is validated multiple times in a batch submission, it ends up
400 * with the intersection of the memory type flags and the union of the
404 drm_intel_add_validate_buffer(drm_intel_bo *bo)
406 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
407 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
410 if (bo_gem->validate_index != -1)
413 /* Extend the array of validation entries as necessary. */
414 if (bufmgr_gem->exec_count == bufmgr_gem->exec_size) {
415 int new_size = bufmgr_gem->exec_size * 2;
420 bufmgr_gem->exec_objects =
421 realloc(bufmgr_gem->exec_objects,
422 sizeof(*bufmgr_gem->exec_objects) * new_size);
423 bufmgr_gem->exec_bos =
424 realloc(bufmgr_gem->exec_bos,
425 sizeof(*bufmgr_gem->exec_bos) * new_size);
426 bufmgr_gem->exec_size = new_size;
429 index = bufmgr_gem->exec_count;
430 bo_gem->validate_index = index;
431 /* Fill in array entry */
432 bufmgr_gem->exec_objects[index].handle = bo_gem->gem_handle;
433 bufmgr_gem->exec_objects[index].relocation_count = bo_gem->reloc_count;
434 bufmgr_gem->exec_objects[index].relocs_ptr = (uintptr_t) bo_gem->relocs;
435 bufmgr_gem->exec_objects[index].alignment = 0;
436 bufmgr_gem->exec_objects[index].offset = 0;
437 bufmgr_gem->exec_bos[index] = bo;
438 bufmgr_gem->exec_count++;
442 drm_intel_add_validate_buffer2(drm_intel_bo *bo, int need_fence)
444 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr;
445 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
448 if (bo_gem->validate_index != -1) {
450 bufmgr_gem->exec2_objects[bo_gem->validate_index].flags |=
451 EXEC_OBJECT_NEEDS_FENCE;
455 /* Extend the array of validation entries as necessary. */
456 if (bufmgr_gem->exec_count == bufmgr_gem->exec_size) {
457 int new_size = bufmgr_gem->exec_size * 2;
462 bufmgr_gem->exec2_objects =
463 realloc(bufmgr_gem->exec2_objects,
464 sizeof(*bufmgr_gem->exec2_objects) * new_size);
465 bufmgr_gem->exec_bos =
466 realloc(bufmgr_gem->exec_bos,
467 sizeof(*bufmgr_gem->exec_bos) * new_size);
468 bufmgr_gem->exec_size = new_size;
471 index = bufmgr_gem->exec_count;
472 bo_gem->validate_index = index;
473 /* Fill in array entry */
474 bufmgr_gem->exec2_objects[index].handle = bo_gem->gem_handle;
475 bufmgr_gem->exec2_objects[index].relocation_count = bo_gem->reloc_count;
476 bufmgr_gem->exec2_objects[index].relocs_ptr = (uintptr_t)bo_gem->relocs;
477 bufmgr_gem->exec2_objects[index].alignment = 0;
478 bufmgr_gem->exec2_objects[index].offset = 0;
479 bufmgr_gem->exec_bos[index] = bo;
480 bufmgr_gem->exec2_objects[index].flags = 0;
481 bufmgr_gem->exec2_objects[index].rsvd1 = 0;
482 bufmgr_gem->exec2_objects[index].rsvd2 = 0;
484 bufmgr_gem->exec2_objects[index].flags |=
485 EXEC_OBJECT_NEEDS_FENCE;
487 bufmgr_gem->exec_count++;
490 #define RELOC_BUF_SIZE(x) ((I915_RELOC_HEADER + x * I915_RELOC0_STRIDE) * \
494 drm_intel_bo_gem_set_in_aperture_size(drm_intel_bufmgr_gem *bufmgr_gem,
495 drm_intel_bo_gem *bo_gem)
499 assert(!bo_gem->used_as_reloc_target);
501 /* The older chipsets are far-less flexible in terms of tiling,
502 * and require tiled buffer to be size aligned in the aperture.
503 * This means that in the worst possible case we will need a hole
504 * twice as large as the object in order for it to fit into the
505 * aperture. Optimal packing is for wimps.
507 size = bo_gem->bo.size;
508 if (bufmgr_gem->gen < 4 && bo_gem->tiling_mode != I915_TILING_NONE) {
511 if (bufmgr_gem->has_relaxed_fencing) {
512 if (bufmgr_gem->gen == 3)
513 min_size = 1024*1024;
517 while (min_size < size)
522 /* Account for worst-case alignment. */
526 bo_gem->reloc_tree_size = size;
530 drm_intel_setup_reloc_list(drm_intel_bo *bo)
532 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
533 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
534 unsigned int max_relocs = bufmgr_gem->max_relocs;
536 if (bo->size / 4 < max_relocs)
537 max_relocs = bo->size / 4;
539 bo_gem->relocs = malloc(max_relocs *
540 sizeof(struct drm_i915_gem_relocation_entry));
541 bo_gem->reloc_target_info = malloc(max_relocs *
542 sizeof(drm_intel_reloc_target));
543 if (bo_gem->relocs == NULL || bo_gem->reloc_target_info == NULL) {
544 bo_gem->has_error = true;
546 free (bo_gem->relocs);
547 bo_gem->relocs = NULL;
549 free (bo_gem->reloc_target_info);
550 bo_gem->reloc_target_info = NULL;
559 drm_intel_gem_bo_busy(drm_intel_bo *bo)
561 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
562 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
563 struct drm_i915_gem_busy busy;
567 busy.handle = bo_gem->gem_handle;
569 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_BUSY, &busy);
571 return (ret == 0 && busy.busy);
575 drm_intel_gem_bo_madvise_internal(drm_intel_bufmgr_gem *bufmgr_gem,
576 drm_intel_bo_gem *bo_gem, int state)
578 struct drm_i915_gem_madvise madv;
581 madv.handle = bo_gem->gem_handle;
584 drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv);
586 return madv.retained;
590 drm_intel_gem_bo_madvise(drm_intel_bo *bo, int madv)
592 return drm_intel_gem_bo_madvise_internal
593 ((drm_intel_bufmgr_gem *) bo->bufmgr,
594 (drm_intel_bo_gem *) bo,
598 /* drop the oldest entries that have been purged by the kernel */
600 drm_intel_gem_bo_cache_purge_bucket(drm_intel_bufmgr_gem *bufmgr_gem,
601 struct drm_intel_gem_bo_bucket *bucket)
603 while (!DRMLISTEMPTY(&bucket->head)) {
604 drm_intel_bo_gem *bo_gem;
606 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
607 bucket->head.next, head);
608 if (drm_intel_gem_bo_madvise_internal
609 (bufmgr_gem, bo_gem, I915_MADV_DONTNEED))
612 DRMLISTDEL(&bo_gem->head);
613 drm_intel_gem_bo_free(&bo_gem->bo);
617 static drm_intel_bo *
618 drm_intel_gem_bo_alloc_internal(drm_intel_bufmgr *bufmgr,
622 uint32_t tiling_mode,
623 unsigned long stride)
625 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
626 drm_intel_bo_gem *bo_gem;
627 unsigned int page_size = getpagesize();
629 struct drm_intel_gem_bo_bucket *bucket;
630 bool alloc_from_cache;
631 unsigned long bo_size;
632 bool for_render = false;
634 if (flags & BO_ALLOC_FOR_RENDER)
637 /* Round the allocated size up to a power of two number of pages. */
638 bucket = drm_intel_gem_bo_bucket_for_size(bufmgr_gem, size);
640 /* If we don't have caching at this size, don't actually round the
643 if (bucket == NULL) {
645 if (bo_size < page_size)
648 bo_size = bucket->size;
651 pthread_mutex_lock(&bufmgr_gem->lock);
652 /* Get a buffer out of the cache if available */
654 alloc_from_cache = false;
655 if (bucket != NULL && !DRMLISTEMPTY(&bucket->head)) {
657 /* Allocate new render-target BOs from the tail (MRU)
658 * of the list, as it will likely be hot in the GPU
659 * cache and in the aperture for us.
661 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
662 bucket->head.prev, head);
663 DRMLISTDEL(&bo_gem->head);
664 alloc_from_cache = true;
666 /* For non-render-target BOs (where we're probably
667 * going to map it first thing in order to fill it
668 * with data), check if the last BO in the cache is
669 * unbusy, and only reuse in that case. Otherwise,
670 * allocating a new buffer is probably faster than
671 * waiting for the GPU to finish.
673 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
674 bucket->head.next, head);
675 if (!drm_intel_gem_bo_busy(&bo_gem->bo)) {
676 alloc_from_cache = true;
677 DRMLISTDEL(&bo_gem->head);
681 if (alloc_from_cache) {
682 if (!drm_intel_gem_bo_madvise_internal
683 (bufmgr_gem, bo_gem, I915_MADV_WILLNEED)) {
684 drm_intel_gem_bo_free(&bo_gem->bo);
685 drm_intel_gem_bo_cache_purge_bucket(bufmgr_gem,
690 if (drm_intel_gem_bo_set_tiling_internal(&bo_gem->bo,
693 drm_intel_gem_bo_free(&bo_gem->bo);
698 pthread_mutex_unlock(&bufmgr_gem->lock);
700 if (!alloc_from_cache) {
701 struct drm_i915_gem_create create;
703 bo_gem = calloc(1, sizeof(*bo_gem));
707 bo_gem->bo.size = bo_size;
710 create.size = bo_size;
712 ret = drmIoctl(bufmgr_gem->fd,
713 DRM_IOCTL_I915_GEM_CREATE,
715 bo_gem->gem_handle = create.handle;
716 bo_gem->bo.handle = bo_gem->gem_handle;
721 bo_gem->bo.bufmgr = bufmgr;
723 bo_gem->tiling_mode = I915_TILING_NONE;
724 bo_gem->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
727 if (drm_intel_gem_bo_set_tiling_internal(&bo_gem->bo,
730 drm_intel_gem_bo_free(&bo_gem->bo);
734 DRMINITLISTHEAD(&bo_gem->name_list);
735 DRMINITLISTHEAD(&bo_gem->vma_list);
739 atomic_set(&bo_gem->refcount, 1);
740 bo_gem->validate_index = -1;
741 bo_gem->reloc_tree_fences = 0;
742 bo_gem->used_as_reloc_target = false;
743 bo_gem->has_error = false;
744 bo_gem->reusable = true;
745 bo_gem->aub_annotations = NULL;
746 bo_gem->aub_annotation_count = 0;
748 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
750 DBG("bo_create: buf %d (%s) %ldb\n",
751 bo_gem->gem_handle, bo_gem->name, size);
756 static drm_intel_bo *
757 drm_intel_gem_bo_alloc_for_render(drm_intel_bufmgr *bufmgr,
760 unsigned int alignment)
762 return drm_intel_gem_bo_alloc_internal(bufmgr, name, size,
764 I915_TILING_NONE, 0);
767 static drm_intel_bo *
768 drm_intel_gem_bo_alloc(drm_intel_bufmgr *bufmgr,
771 unsigned int alignment)
773 return drm_intel_gem_bo_alloc_internal(bufmgr, name, size, 0,
774 I915_TILING_NONE, 0);
777 static drm_intel_bo *
778 drm_intel_gem_bo_alloc_tiled(drm_intel_bufmgr *bufmgr, const char *name,
779 int x, int y, int cpp, uint32_t *tiling_mode,
780 unsigned long *pitch, unsigned long flags)
782 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
783 unsigned long size, stride;
787 unsigned long aligned_y, height_alignment;
789 tiling = *tiling_mode;
791 /* If we're tiled, our allocations are in 8 or 32-row blocks,
792 * so failure to align our height means that we won't allocate
795 * If we're untiled, we still have to align to 2 rows high
796 * because the data port accesses 2x2 blocks even if the
797 * bottom row isn't to be rendered, so failure to align means
798 * we could walk off the end of the GTT and fault. This is
799 * documented on 965, and may be the case on older chipsets
800 * too so we try to be careful.
803 height_alignment = 2;
805 if ((bufmgr_gem->gen == 2) && tiling != I915_TILING_NONE)
806 height_alignment = 16;
807 else if (tiling == I915_TILING_X
808 || (IS_915(bufmgr_gem->pci_device)
809 && tiling == I915_TILING_Y))
810 height_alignment = 8;
811 else if (tiling == I915_TILING_Y)
812 height_alignment = 32;
813 aligned_y = ALIGN(y, height_alignment);
816 stride = drm_intel_gem_bo_tile_pitch(bufmgr_gem, stride, tiling_mode);
817 size = stride * aligned_y;
818 size = drm_intel_gem_bo_tile_size(bufmgr_gem, size, tiling_mode);
819 } while (*tiling_mode != tiling);
822 if (tiling == I915_TILING_NONE)
825 return drm_intel_gem_bo_alloc_internal(bufmgr, name, size, flags,
830 * Returns a drm_intel_bo wrapping the given buffer object handle.
832 * This can be used when one application needs to pass a buffer object
836 drm_intel_bo_gem_create_from_name(drm_intel_bufmgr *bufmgr,
840 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
841 drm_intel_bo_gem *bo_gem;
843 struct drm_gem_open open_arg;
844 struct drm_i915_gem_get_tiling get_tiling;
847 /* At the moment most applications only have a few named bo.
848 * For instance, in a DRI client only the render buffers passed
849 * between X and the client are named. And since X returns the
850 * alternating names for the front/back buffer a linear search
851 * provides a sufficiently fast match.
853 for (list = bufmgr_gem->named.next;
854 list != &bufmgr_gem->named;
856 bo_gem = DRMLISTENTRY(drm_intel_bo_gem, list, name_list);
857 if (bo_gem->global_name == handle) {
858 drm_intel_gem_bo_reference(&bo_gem->bo);
863 bo_gem = calloc(1, sizeof(*bo_gem));
868 open_arg.name = handle;
869 ret = drmIoctl(bufmgr_gem->fd,
873 DBG("Couldn't reference %s handle 0x%08x: %s\n",
874 name, handle, strerror(errno));
878 bo_gem->bo.size = open_arg.size;
879 bo_gem->bo.offset = 0;
880 bo_gem->bo.virtual = NULL;
881 bo_gem->bo.bufmgr = bufmgr;
883 atomic_set(&bo_gem->refcount, 1);
884 bo_gem->validate_index = -1;
885 bo_gem->gem_handle = open_arg.handle;
886 bo_gem->bo.handle = open_arg.handle;
887 bo_gem->global_name = handle;
888 bo_gem->reusable = false;
890 VG_CLEAR(get_tiling);
891 get_tiling.handle = bo_gem->gem_handle;
892 ret = drmIoctl(bufmgr_gem->fd,
893 DRM_IOCTL_I915_GEM_GET_TILING,
896 drm_intel_gem_bo_unreference(&bo_gem->bo);
899 bo_gem->tiling_mode = get_tiling.tiling_mode;
900 bo_gem->swizzle_mode = get_tiling.swizzle_mode;
901 /* XXX stride is unknown */
902 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
904 DRMINITLISTHEAD(&bo_gem->vma_list);
905 DRMLISTADDTAIL(&bo_gem->name_list, &bufmgr_gem->named);
906 DBG("bo_create_from_handle: %d (%s)\n", handle, bo_gem->name);
912 drm_intel_gem_bo_free(drm_intel_bo *bo)
914 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
915 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
916 struct drm_gem_close close;
919 DRMLISTDEL(&bo_gem->vma_list);
920 if (bo_gem->mem_virtual) {
921 VG(VALGRIND_FREELIKE_BLOCK(bo_gem->mem_virtual, 0));
922 munmap(bo_gem->mem_virtual, bo_gem->bo.size);
923 bufmgr_gem->vma_count--;
925 if (bo_gem->gtt_virtual) {
926 munmap(bo_gem->gtt_virtual, bo_gem->bo.size);
927 bufmgr_gem->vma_count--;
930 /* Close this object */
932 close.handle = bo_gem->gem_handle;
933 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_GEM_CLOSE, &close);
935 DBG("DRM_IOCTL_GEM_CLOSE %d failed (%s): %s\n",
936 bo_gem->gem_handle, bo_gem->name, strerror(errno));
938 free(bo_gem->aub_annotations);
943 drm_intel_gem_bo_mark_mmaps_incoherent(drm_intel_bo *bo)
946 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
948 if (bo_gem->mem_virtual)
949 VALGRIND_MAKE_MEM_NOACCESS(bo_gem->mem_virtual, bo->size);
951 if (bo_gem->gtt_virtual)
952 VALGRIND_MAKE_MEM_NOACCESS(bo_gem->gtt_virtual, bo->size);
956 /** Frees all cached buffers significantly older than @time. */
958 drm_intel_gem_cleanup_bo_cache(drm_intel_bufmgr_gem *bufmgr_gem, time_t time)
962 if (bufmgr_gem->time == time)
965 for (i = 0; i < bufmgr_gem->num_buckets; i++) {
966 struct drm_intel_gem_bo_bucket *bucket =
967 &bufmgr_gem->cache_bucket[i];
969 while (!DRMLISTEMPTY(&bucket->head)) {
970 drm_intel_bo_gem *bo_gem;
972 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
973 bucket->head.next, head);
974 if (time - bo_gem->free_time <= 1)
977 DRMLISTDEL(&bo_gem->head);
979 drm_intel_gem_bo_free(&bo_gem->bo);
983 bufmgr_gem->time = time;
986 static void drm_intel_gem_bo_purge_vma_cache(drm_intel_bufmgr_gem *bufmgr_gem)
990 DBG("%s: cached=%d, open=%d, limit=%d\n", __FUNCTION__,
991 bufmgr_gem->vma_count, bufmgr_gem->vma_open, bufmgr_gem->vma_max);
993 if (bufmgr_gem->vma_max < 0)
996 /* We may need to evict a few entries in order to create new mmaps */
997 limit = bufmgr_gem->vma_max - 2*bufmgr_gem->vma_open;
1001 while (bufmgr_gem->vma_count > limit) {
1002 drm_intel_bo_gem *bo_gem;
1004 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
1005 bufmgr_gem->vma_cache.next,
1007 assert(bo_gem->map_count == 0);
1008 DRMLISTDELINIT(&bo_gem->vma_list);
1010 if (bo_gem->mem_virtual) {
1011 munmap(bo_gem->mem_virtual, bo_gem->bo.size);
1012 bo_gem->mem_virtual = NULL;
1013 bufmgr_gem->vma_count--;
1015 if (bo_gem->gtt_virtual) {
1016 munmap(bo_gem->gtt_virtual, bo_gem->bo.size);
1017 bo_gem->gtt_virtual = NULL;
1018 bufmgr_gem->vma_count--;
1023 static void drm_intel_gem_bo_close_vma(drm_intel_bufmgr_gem *bufmgr_gem,
1024 drm_intel_bo_gem *bo_gem)
1026 bufmgr_gem->vma_open--;
1027 DRMLISTADDTAIL(&bo_gem->vma_list, &bufmgr_gem->vma_cache);
1028 if (bo_gem->mem_virtual)
1029 bufmgr_gem->vma_count++;
1030 if (bo_gem->gtt_virtual)
1031 bufmgr_gem->vma_count++;
1032 drm_intel_gem_bo_purge_vma_cache(bufmgr_gem);
1035 static void drm_intel_gem_bo_open_vma(drm_intel_bufmgr_gem *bufmgr_gem,
1036 drm_intel_bo_gem *bo_gem)
1038 bufmgr_gem->vma_open++;
1039 DRMLISTDEL(&bo_gem->vma_list);
1040 if (bo_gem->mem_virtual)
1041 bufmgr_gem->vma_count--;
1042 if (bo_gem->gtt_virtual)
1043 bufmgr_gem->vma_count--;
1044 drm_intel_gem_bo_purge_vma_cache(bufmgr_gem);
1048 drm_intel_gem_bo_unreference_final(drm_intel_bo *bo, time_t time)
1050 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1051 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1052 struct drm_intel_gem_bo_bucket *bucket;
1055 /* Unreference all the target buffers */
1056 for (i = 0; i < bo_gem->reloc_count; i++) {
1057 if (bo_gem->reloc_target_info[i].bo != bo) {
1058 drm_intel_gem_bo_unreference_locked_timed(bo_gem->
1059 reloc_target_info[i].bo,
1063 bo_gem->reloc_count = 0;
1064 bo_gem->used_as_reloc_target = false;
1066 DBG("bo_unreference final: %d (%s)\n",
1067 bo_gem->gem_handle, bo_gem->name);
1069 /* release memory associated with this object */
1070 if (bo_gem->reloc_target_info) {
1071 free(bo_gem->reloc_target_info);
1072 bo_gem->reloc_target_info = NULL;
1074 if (bo_gem->relocs) {
1075 free(bo_gem->relocs);
1076 bo_gem->relocs = NULL;
1079 /* Clear any left-over mappings */
1080 if (bo_gem->map_count) {
1081 DBG("bo freed with non-zero map-count %d\n", bo_gem->map_count);
1082 bo_gem->map_count = 0;
1083 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem);
1084 drm_intel_gem_bo_mark_mmaps_incoherent(bo);
1087 DRMLISTDEL(&bo_gem->name_list);
1089 bucket = drm_intel_gem_bo_bucket_for_size(bufmgr_gem, bo->size);
1090 /* Put the buffer into our internal cache for reuse if we can. */
1091 if (bufmgr_gem->bo_reuse && bo_gem->reusable && bucket != NULL &&
1092 drm_intel_gem_bo_madvise_internal(bufmgr_gem, bo_gem,
1093 I915_MADV_DONTNEED)) {
1094 bo_gem->free_time = time;
1096 bo_gem->name = NULL;
1097 bo_gem->validate_index = -1;
1099 DRMLISTADDTAIL(&bo_gem->head, &bucket->head);
1101 drm_intel_gem_bo_free(bo);
1105 static void drm_intel_gem_bo_unreference_locked_timed(drm_intel_bo *bo,
1108 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1110 assert(atomic_read(&bo_gem->refcount) > 0);
1111 if (atomic_dec_and_test(&bo_gem->refcount))
1112 drm_intel_gem_bo_unreference_final(bo, time);
1115 static void drm_intel_gem_bo_unreference(drm_intel_bo *bo)
1117 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1119 assert(atomic_read(&bo_gem->refcount) > 0);
1120 if (atomic_dec_and_test(&bo_gem->refcount)) {
1121 drm_intel_bufmgr_gem *bufmgr_gem =
1122 (drm_intel_bufmgr_gem *) bo->bufmgr;
1123 struct timespec time;
1125 clock_gettime(CLOCK_MONOTONIC, &time);
1127 pthread_mutex_lock(&bufmgr_gem->lock);
1128 drm_intel_gem_bo_unreference_final(bo, time.tv_sec);
1129 drm_intel_gem_cleanup_bo_cache(bufmgr_gem, time.tv_sec);
1130 pthread_mutex_unlock(&bufmgr_gem->lock);
1134 static int drm_intel_gem_bo_map(drm_intel_bo *bo, int write_enable)
1136 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1137 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1138 struct drm_i915_gem_set_domain set_domain;
1141 pthread_mutex_lock(&bufmgr_gem->lock);
1143 if (bo_gem->map_count++ == 0)
1144 drm_intel_gem_bo_open_vma(bufmgr_gem, bo_gem);
1146 if (!bo_gem->mem_virtual) {
1147 struct drm_i915_gem_mmap mmap_arg;
1149 DBG("bo_map: %d (%s), map_count=%d\n",
1150 bo_gem->gem_handle, bo_gem->name, bo_gem->map_count);
1153 mmap_arg.handle = bo_gem->gem_handle;
1154 mmap_arg.offset = 0;
1155 mmap_arg.size = bo->size;
1156 ret = drmIoctl(bufmgr_gem->fd,
1157 DRM_IOCTL_I915_GEM_MMAP,
1161 DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
1162 __FILE__, __LINE__, bo_gem->gem_handle,
1163 bo_gem->name, strerror(errno));
1164 if (--bo_gem->map_count == 0)
1165 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem);
1166 pthread_mutex_unlock(&bufmgr_gem->lock);
1169 VG(VALGRIND_MALLOCLIKE_BLOCK(mmap_arg.addr_ptr, mmap_arg.size, 0, 1));
1170 bo_gem->mem_virtual = (void *)(uintptr_t) mmap_arg.addr_ptr;
1172 DBG("bo_map: %d (%s) -> %p\n", bo_gem->gem_handle, bo_gem->name,
1173 bo_gem->mem_virtual);
1174 bo->virtual = bo_gem->mem_virtual;
1176 VG_CLEAR(set_domain);
1177 set_domain.handle = bo_gem->gem_handle;
1178 set_domain.read_domains = I915_GEM_DOMAIN_CPU;
1180 set_domain.write_domain = I915_GEM_DOMAIN_CPU;
1182 set_domain.write_domain = 0;
1183 ret = drmIoctl(bufmgr_gem->fd,
1184 DRM_IOCTL_I915_GEM_SET_DOMAIN,
1187 DBG("%s:%d: Error setting to CPU domain %d: %s\n",
1188 __FILE__, __LINE__, bo_gem->gem_handle,
1193 bo_gem->mapped_cpu_write = true;
1195 drm_intel_gem_bo_mark_mmaps_incoherent(bo);
1196 VG(VALGRIND_MAKE_MEM_DEFINED(bo_gem->mem_virtual, bo->size));
1197 pthread_mutex_unlock(&bufmgr_gem->lock);
1203 map_gtt(drm_intel_bo *bo)
1205 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1206 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1209 if (bo_gem->map_count++ == 0)
1210 drm_intel_gem_bo_open_vma(bufmgr_gem, bo_gem);
1212 /* Get a mapping of the buffer if we haven't before. */
1213 if (bo_gem->gtt_virtual == NULL) {
1214 struct drm_i915_gem_mmap_gtt mmap_arg;
1216 DBG("bo_map_gtt: mmap %d (%s), map_count=%d\n",
1217 bo_gem->gem_handle, bo_gem->name, bo_gem->map_count);
1220 mmap_arg.handle = bo_gem->gem_handle;
1222 /* Get the fake offset back... */
1223 ret = drmIoctl(bufmgr_gem->fd,
1224 DRM_IOCTL_I915_GEM_MMAP_GTT,
1228 DBG("%s:%d: Error preparing buffer map %d (%s): %s .\n",
1230 bo_gem->gem_handle, bo_gem->name,
1232 if (--bo_gem->map_count == 0)
1233 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem);
1238 bo_gem->gtt_virtual = mmap(0, bo->size, PROT_READ | PROT_WRITE,
1239 MAP_SHARED, bufmgr_gem->fd,
1241 if (bo_gem->gtt_virtual == MAP_FAILED) {
1242 bo_gem->gtt_virtual = NULL;
1244 DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
1246 bo_gem->gem_handle, bo_gem->name,
1248 if (--bo_gem->map_count == 0)
1249 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem);
1254 bo->virtual = bo_gem->gtt_virtual;
1256 DBG("bo_map_gtt: %d (%s) -> %p\n", bo_gem->gem_handle, bo_gem->name,
1257 bo_gem->gtt_virtual);
1262 int drm_intel_gem_bo_map_gtt(drm_intel_bo *bo)
1264 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1265 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1266 struct drm_i915_gem_set_domain set_domain;
1269 pthread_mutex_lock(&bufmgr_gem->lock);
1273 pthread_mutex_unlock(&bufmgr_gem->lock);
1277 /* Now move it to the GTT domain so that the GPU and CPU
1278 * caches are flushed and the GPU isn't actively using the
1281 * The pagefault handler does this domain change for us when
1282 * it has unbound the BO from the GTT, but it's up to us to
1283 * tell it when we're about to use things if we had done
1284 * rendering and it still happens to be bound to the GTT.
1286 VG_CLEAR(set_domain);
1287 set_domain.handle = bo_gem->gem_handle;
1288 set_domain.read_domains = I915_GEM_DOMAIN_GTT;
1289 set_domain.write_domain = I915_GEM_DOMAIN_GTT;
1290 ret = drmIoctl(bufmgr_gem->fd,
1291 DRM_IOCTL_I915_GEM_SET_DOMAIN,
1294 DBG("%s:%d: Error setting domain %d: %s\n",
1295 __FILE__, __LINE__, bo_gem->gem_handle,
1299 drm_intel_gem_bo_mark_mmaps_incoherent(bo);
1300 VG(VALGRIND_MAKE_MEM_DEFINED(bo_gem->gtt_virtual, bo->size));
1301 pthread_mutex_unlock(&bufmgr_gem->lock);
1307 * Performs a mapping of the buffer object like the normal GTT
1308 * mapping, but avoids waiting for the GPU to be done reading from or
1309 * rendering to the buffer.
1311 * This is used in the implementation of GL_ARB_map_buffer_range: The
1312 * user asks to create a buffer, then does a mapping, fills some
1313 * space, runs a drawing command, then asks to map it again without
1314 * synchronizing because it guarantees that it won't write over the
1315 * data that the GPU is busy using (or, more specifically, that if it
1316 * does write over the data, it acknowledges that rendering is
1320 int drm_intel_gem_bo_map_unsynchronized(drm_intel_bo *bo)
1322 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1325 /* If the CPU cache isn't coherent with the GTT, then use a
1326 * regular synchronized mapping. The problem is that we don't
1327 * track where the buffer was last used on the CPU side in
1328 * terms of drm_intel_bo_map vs drm_intel_gem_bo_map_gtt, so
1329 * we would potentially corrupt the buffer even when the user
1330 * does reasonable things.
1332 if (!bufmgr_gem->has_llc)
1333 return drm_intel_gem_bo_map_gtt(bo);
1335 pthread_mutex_lock(&bufmgr_gem->lock);
1337 pthread_mutex_unlock(&bufmgr_gem->lock);
1342 static int drm_intel_gem_bo_unmap(drm_intel_bo *bo)
1344 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1345 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1351 pthread_mutex_lock(&bufmgr_gem->lock);
1353 if (bo_gem->map_count <= 0) {
1354 DBG("attempted to unmap an unmapped bo\n");
1355 pthread_mutex_unlock(&bufmgr_gem->lock);
1356 /* Preserve the old behaviour of just treating this as a
1357 * no-op rather than reporting the error.
1362 if (bo_gem->mapped_cpu_write) {
1363 struct drm_i915_gem_sw_finish sw_finish;
1365 /* Cause a flush to happen if the buffer's pinned for
1366 * scanout, so the results show up in a timely manner.
1367 * Unlike GTT set domains, this only does work if the
1368 * buffer should be scanout-related.
1370 VG_CLEAR(sw_finish);
1371 sw_finish.handle = bo_gem->gem_handle;
1372 ret = drmIoctl(bufmgr_gem->fd,
1373 DRM_IOCTL_I915_GEM_SW_FINISH,
1375 ret = ret == -1 ? -errno : 0;
1377 bo_gem->mapped_cpu_write = false;
1380 /* We need to unmap after every innovation as we cannot track
1381 * an open vma for every bo as that will exhaasut the system
1382 * limits and cause later failures.
1384 if (--bo_gem->map_count == 0) {
1385 drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem);
1386 drm_intel_gem_bo_mark_mmaps_incoherent(bo);
1389 pthread_mutex_unlock(&bufmgr_gem->lock);
1394 int drm_intel_gem_bo_unmap_gtt(drm_intel_bo *bo)
1396 return drm_intel_gem_bo_unmap(bo);
1400 drm_intel_gem_bo_subdata(drm_intel_bo *bo, unsigned long offset,
1401 unsigned long size, const void *data)
1403 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1404 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1405 struct drm_i915_gem_pwrite pwrite;
1409 pwrite.handle = bo_gem->gem_handle;
1410 pwrite.offset = offset;
1412 pwrite.data_ptr = (uint64_t) (uintptr_t) data;
1413 ret = drmIoctl(bufmgr_gem->fd,
1414 DRM_IOCTL_I915_GEM_PWRITE,
1418 DBG("%s:%d: Error writing data to buffer %d: (%d %d) %s .\n",
1419 __FILE__, __LINE__, bo_gem->gem_handle, (int)offset,
1420 (int)size, strerror(errno));
1427 drm_intel_gem_get_pipe_from_crtc_id(drm_intel_bufmgr *bufmgr, int crtc_id)
1429 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
1430 struct drm_i915_get_pipe_from_crtc_id get_pipe_from_crtc_id;
1433 VG_CLEAR(get_pipe_from_crtc_id);
1434 get_pipe_from_crtc_id.crtc_id = crtc_id;
1435 ret = drmIoctl(bufmgr_gem->fd,
1436 DRM_IOCTL_I915_GET_PIPE_FROM_CRTC_ID,
1437 &get_pipe_from_crtc_id);
1439 /* We return -1 here to signal that we don't
1440 * know which pipe is associated with this crtc.
1441 * This lets the caller know that this information
1442 * isn't available; using the wrong pipe for
1443 * vblank waiting can cause the chipset to lock up
1448 return get_pipe_from_crtc_id.pipe;
1452 drm_intel_gem_bo_get_subdata(drm_intel_bo *bo, unsigned long offset,
1453 unsigned long size, void *data)
1455 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1456 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1457 struct drm_i915_gem_pread pread;
1461 pread.handle = bo_gem->gem_handle;
1462 pread.offset = offset;
1464 pread.data_ptr = (uint64_t) (uintptr_t) data;
1465 ret = drmIoctl(bufmgr_gem->fd,
1466 DRM_IOCTL_I915_GEM_PREAD,
1470 DBG("%s:%d: Error reading data from buffer %d: (%d %d) %s .\n",
1471 __FILE__, __LINE__, bo_gem->gem_handle, (int)offset,
1472 (int)size, strerror(errno));
1478 /** Waits for all GPU rendering with the object to have completed. */
1480 drm_intel_gem_bo_wait_rendering(drm_intel_bo *bo)
1482 drm_intel_gem_bo_start_gtt_access(bo, 1);
1486 * Waits on a BO for the given amount of time.
1488 * @bo: buffer object to wait for
1489 * @timeout_ns: amount of time to wait in nanoseconds.
1490 * If value is less than 0, an infinite wait will occur.
1492 * Returns 0 if the wait was successful ie. the last batch referencing the
1493 * object has completed within the allotted time. Otherwise some negative return
1494 * value describes the error. Of particular interest is -ETIME when the wait has
1495 * failed to yield the desired result.
1497 * Similar to drm_intel_gem_bo_wait_rendering except a timeout parameter allows
1498 * the operation to give up after a certain amount of time. Another subtle
1499 * difference is the internal locking semantics are different (this variant does
1500 * not hold the lock for the duration of the wait). This makes the wait subject
1501 * to a larger userspace race window.
1503 * The implementation shall wait until the object is no longer actively
1504 * referenced within a batch buffer at the time of the call. The wait will
1505 * not guarantee that the buffer is re-issued via another thread, or an flinked
1506 * handle. Userspace must make sure this race does not occur if such precision
1509 int drm_intel_gem_bo_wait(drm_intel_bo *bo, int64_t timeout_ns)
1511 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1512 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1513 struct drm_i915_gem_wait wait;
1516 if (!bufmgr_gem->has_wait_timeout) {
1517 DBG("%s:%d: Timed wait is not supported. Falling back to "
1518 "infinite wait\n", __FILE__, __LINE__);
1520 drm_intel_gem_bo_wait_rendering(bo);
1523 return drm_intel_gem_bo_busy(bo) ? -ETIME : 0;
1527 wait.bo_handle = bo_gem->gem_handle;
1528 wait.timeout_ns = timeout_ns;
1530 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_WAIT, &wait);
1538 * Sets the object to the GTT read and possibly write domain, used by the X
1539 * 2D driver in the absence of kernel support to do drm_intel_gem_bo_map_gtt().
1541 * In combination with drm_intel_gem_bo_pin() and manual fence management, we
1542 * can do tiled pixmaps this way.
1545 drm_intel_gem_bo_start_gtt_access(drm_intel_bo *bo, int write_enable)
1547 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1548 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1549 struct drm_i915_gem_set_domain set_domain;
1552 VG_CLEAR(set_domain);
1553 set_domain.handle = bo_gem->gem_handle;
1554 set_domain.read_domains = I915_GEM_DOMAIN_GTT;
1555 set_domain.write_domain = write_enable ? I915_GEM_DOMAIN_GTT : 0;
1556 ret = drmIoctl(bufmgr_gem->fd,
1557 DRM_IOCTL_I915_GEM_SET_DOMAIN,
1560 DBG("%s:%d: Error setting memory domains %d (%08x %08x): %s .\n",
1561 __FILE__, __LINE__, bo_gem->gem_handle,
1562 set_domain.read_domains, set_domain.write_domain,
1568 drm_intel_bufmgr_gem_destroy(drm_intel_bufmgr *bufmgr)
1570 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
1573 free(bufmgr_gem->exec2_objects);
1574 free(bufmgr_gem->exec_objects);
1575 free(bufmgr_gem->exec_bos);
1577 pthread_mutex_destroy(&bufmgr_gem->lock);
1579 /* Free any cached buffer objects we were going to reuse */
1580 for (i = 0; i < bufmgr_gem->num_buckets; i++) {
1581 struct drm_intel_gem_bo_bucket *bucket =
1582 &bufmgr_gem->cache_bucket[i];
1583 drm_intel_bo_gem *bo_gem;
1585 while (!DRMLISTEMPTY(&bucket->head)) {
1586 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
1587 bucket->head.next, head);
1588 DRMLISTDEL(&bo_gem->head);
1590 drm_intel_gem_bo_free(&bo_gem->bo);
1598 * Adds the target buffer to the validation list and adds the relocation
1599 * to the reloc_buffer's relocation list.
1601 * The relocation entry at the given offset must already contain the
1602 * precomputed relocation value, because the kernel will optimize out
1603 * the relocation entry write when the buffer hasn't moved from the
1604 * last known offset in target_bo.
1607 do_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset,
1608 drm_intel_bo *target_bo, uint32_t target_offset,
1609 uint32_t read_domains, uint32_t write_domain,
1612 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1613 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1614 drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *) target_bo;
1615 bool fenced_command;
1617 if (bo_gem->has_error)
1620 if (target_bo_gem->has_error) {
1621 bo_gem->has_error = true;
1625 /* We never use HW fences for rendering on 965+ */
1626 if (bufmgr_gem->gen >= 4)
1629 fenced_command = need_fence;
1630 if (target_bo_gem->tiling_mode == I915_TILING_NONE)
1633 /* Create a new relocation list if needed */
1634 if (bo_gem->relocs == NULL && drm_intel_setup_reloc_list(bo))
1637 /* Check overflow */
1638 assert(bo_gem->reloc_count < bufmgr_gem->max_relocs);
1641 assert(offset <= bo->size - 4);
1642 assert((write_domain & (write_domain - 1)) == 0);
1644 /* Make sure that we're not adding a reloc to something whose size has
1645 * already been accounted for.
1647 assert(!bo_gem->used_as_reloc_target);
1648 if (target_bo_gem != bo_gem) {
1649 target_bo_gem->used_as_reloc_target = true;
1650 bo_gem->reloc_tree_size += target_bo_gem->reloc_tree_size;
1652 /* An object needing a fence is a tiled buffer, so it won't have
1653 * relocs to other buffers.
1656 target_bo_gem->reloc_tree_fences = 1;
1657 bo_gem->reloc_tree_fences += target_bo_gem->reloc_tree_fences;
1659 bo_gem->relocs[bo_gem->reloc_count].offset = offset;
1660 bo_gem->relocs[bo_gem->reloc_count].delta = target_offset;
1661 bo_gem->relocs[bo_gem->reloc_count].target_handle =
1662 target_bo_gem->gem_handle;
1663 bo_gem->relocs[bo_gem->reloc_count].read_domains = read_domains;
1664 bo_gem->relocs[bo_gem->reloc_count].write_domain = write_domain;
1665 bo_gem->relocs[bo_gem->reloc_count].presumed_offset = target_bo->offset;
1667 bo_gem->reloc_target_info[bo_gem->reloc_count].bo = target_bo;
1668 if (target_bo != bo)
1669 drm_intel_gem_bo_reference(target_bo);
1671 bo_gem->reloc_target_info[bo_gem->reloc_count].flags =
1672 DRM_INTEL_RELOC_FENCE;
1674 bo_gem->reloc_target_info[bo_gem->reloc_count].flags = 0;
1676 bo_gem->reloc_count++;
1682 drm_intel_gem_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset,
1683 drm_intel_bo *target_bo, uint32_t target_offset,
1684 uint32_t read_domains, uint32_t write_domain)
1686 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr;
1688 return do_bo_emit_reloc(bo, offset, target_bo, target_offset,
1689 read_domains, write_domain,
1690 !bufmgr_gem->fenced_relocs);
1694 drm_intel_gem_bo_emit_reloc_fence(drm_intel_bo *bo, uint32_t offset,
1695 drm_intel_bo *target_bo,
1696 uint32_t target_offset,
1697 uint32_t read_domains, uint32_t write_domain)
1699 return do_bo_emit_reloc(bo, offset, target_bo, target_offset,
1700 read_domains, write_domain, true);
1704 drm_intel_gem_bo_get_reloc_count(drm_intel_bo *bo)
1706 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1708 return bo_gem->reloc_count;
1712 * Removes existing relocation entries in the BO after "start".
1714 * This allows a user to avoid a two-step process for state setup with
1715 * counting up all the buffer objects and doing a
1716 * drm_intel_bufmgr_check_aperture_space() before emitting any of the
1717 * relocations for the state setup. Instead, save the state of the
1718 * batchbuffer including drm_intel_gem_get_reloc_count(), emit all the
1719 * state, and then check if it still fits in the aperture.
1721 * Any further drm_intel_bufmgr_check_aperture_space() queries
1722 * involving this buffer in the tree are undefined after this call.
1725 drm_intel_gem_bo_clear_relocs(drm_intel_bo *bo, int start)
1727 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1729 struct timespec time;
1731 clock_gettime(CLOCK_MONOTONIC, &time);
1733 assert(bo_gem->reloc_count >= start);
1734 /* Unreference the cleared target buffers */
1735 for (i = start; i < bo_gem->reloc_count; i++) {
1736 if (bo_gem->reloc_target_info[i].bo != bo) {
1737 drm_intel_gem_bo_unreference_locked_timed(bo_gem->
1738 reloc_target_info[i].bo,
1742 bo_gem->reloc_count = start;
1746 * Walk the tree of relocations rooted at BO and accumulate the list of
1747 * validations to be performed and update the relocation buffers with
1748 * index values into the validation list.
1751 drm_intel_gem_bo_process_reloc(drm_intel_bo *bo)
1753 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1756 if (bo_gem->relocs == NULL)
1759 for (i = 0; i < bo_gem->reloc_count; i++) {
1760 drm_intel_bo *target_bo = bo_gem->reloc_target_info[i].bo;
1762 if (target_bo == bo)
1765 drm_intel_gem_bo_mark_mmaps_incoherent(bo);
1767 /* Continue walking the tree depth-first. */
1768 drm_intel_gem_bo_process_reloc(target_bo);
1770 /* Add the target to the validate list */
1771 drm_intel_add_validate_buffer(target_bo);
1776 drm_intel_gem_bo_process_reloc2(drm_intel_bo *bo)
1778 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
1781 if (bo_gem->relocs == NULL)
1784 for (i = 0; i < bo_gem->reloc_count; i++) {
1785 drm_intel_bo *target_bo = bo_gem->reloc_target_info[i].bo;
1788 if (target_bo == bo)
1791 drm_intel_gem_bo_mark_mmaps_incoherent(bo);
1793 /* Continue walking the tree depth-first. */
1794 drm_intel_gem_bo_process_reloc2(target_bo);
1796 need_fence = (bo_gem->reloc_target_info[i].flags &
1797 DRM_INTEL_RELOC_FENCE);
1799 /* Add the target to the validate list */
1800 drm_intel_add_validate_buffer2(target_bo, need_fence);
1806 drm_intel_update_buffer_offsets(drm_intel_bufmgr_gem *bufmgr_gem)
1810 for (i = 0; i < bufmgr_gem->exec_count; i++) {
1811 drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
1812 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1814 /* Update the buffer offset */
1815 if (bufmgr_gem->exec_objects[i].offset != bo->offset) {
1816 DBG("BO %d (%s) migrated: 0x%08lx -> 0x%08llx\n",
1817 bo_gem->gem_handle, bo_gem->name, bo->offset,
1818 (unsigned long long)bufmgr_gem->exec_objects[i].
1820 bo->offset = bufmgr_gem->exec_objects[i].offset;
1826 drm_intel_update_buffer_offsets2 (drm_intel_bufmgr_gem *bufmgr_gem)
1830 for (i = 0; i < bufmgr_gem->exec_count; i++) {
1831 drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
1832 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
1834 /* Update the buffer offset */
1835 if (bufmgr_gem->exec2_objects[i].offset != bo->offset) {
1836 DBG("BO %d (%s) migrated: 0x%08lx -> 0x%08llx\n",
1837 bo_gem->gem_handle, bo_gem->name, bo->offset,
1838 (unsigned long long)bufmgr_gem->exec2_objects[i].offset);
1839 bo->offset = bufmgr_gem->exec2_objects[i].offset;
1845 aub_out(drm_intel_bufmgr_gem *bufmgr_gem, uint32_t data)
1847 fwrite(&data, 1, 4, bufmgr_gem->aub_file);
1851 aub_out_data(drm_intel_bufmgr_gem *bufmgr_gem, void *data, size_t size)
1853 fwrite(data, 1, size, bufmgr_gem->aub_file);
1857 aub_write_bo_data(drm_intel_bo *bo, uint32_t offset, uint32_t size)
1859 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1860 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1864 data = malloc(bo->size);
1865 drm_intel_bo_get_subdata(bo, offset, size, data);
1867 /* Easy mode: write out bo with no relocations */
1868 if (!bo_gem->reloc_count) {
1869 aub_out_data(bufmgr_gem, data, size);
1874 /* Otherwise, handle the relocations while writing. */
1875 for (i = 0; i < size / 4; i++) {
1877 for (r = 0; r < bo_gem->reloc_count; r++) {
1878 struct drm_i915_gem_relocation_entry *reloc;
1879 drm_intel_reloc_target *info;
1881 reloc = &bo_gem->relocs[r];
1882 info = &bo_gem->reloc_target_info[r];
1884 if (reloc->offset == offset + i * 4) {
1885 drm_intel_bo_gem *target_gem;
1888 target_gem = (drm_intel_bo_gem *)info->bo;
1891 val += target_gem->aub_offset;
1893 aub_out(bufmgr_gem, val);
1898 if (r == bo_gem->reloc_count) {
1899 /* no relocation, just the data */
1900 aub_out(bufmgr_gem, data[i]);
1908 aub_bo_get_address(drm_intel_bo *bo)
1910 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1911 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1913 /* Give the object a graphics address in the AUB file. We
1914 * don't just use the GEM object address because we do AUB
1915 * dumping before execution -- we want to successfully log
1916 * when the hardware might hang, and we might even want to aub
1917 * capture for a driver trying to execute on a different
1918 * generation of hardware by disabling the actual kernel exec
1921 bo_gem->aub_offset = bufmgr_gem->aub_offset;
1922 bufmgr_gem->aub_offset += bo->size;
1923 /* XXX: Handle aperture overflow. */
1924 assert(bufmgr_gem->aub_offset < 256 * 1024 * 1024);
1928 aub_write_trace_block(drm_intel_bo *bo, uint32_t type, uint32_t subtype,
1929 uint32_t offset, uint32_t size)
1931 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1932 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1935 CMD_AUB_TRACE_HEADER_BLOCK |
1938 AUB_TRACE_MEMTYPE_GTT | type | AUB_TRACE_OP_DATA_WRITE);
1939 aub_out(bufmgr_gem, subtype);
1940 aub_out(bufmgr_gem, bo_gem->aub_offset + offset);
1941 aub_out(bufmgr_gem, size);
1942 aub_write_bo_data(bo, offset, size);
1946 * Break up large objects into multiple writes. Otherwise a 128kb VBO
1947 * would overflow the 16 bits of size field in the packet header and
1948 * everything goes badly after that.
1951 aub_write_large_trace_block(drm_intel_bo *bo, uint32_t type, uint32_t subtype,
1952 uint32_t offset, uint32_t size)
1954 uint32_t block_size;
1955 uint32_t sub_offset;
1957 for (sub_offset = 0; sub_offset < size; sub_offset += block_size) {
1958 block_size = size - sub_offset;
1960 if (block_size > 8 * 4096)
1961 block_size = 8 * 4096;
1963 aub_write_trace_block(bo, type, subtype, offset + sub_offset,
1969 aub_write_bo(drm_intel_bo *bo)
1971 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1972 uint32_t offset = 0;
1975 aub_bo_get_address(bo);
1977 /* Write out each annotated section separately. */
1978 for (i = 0; i < bo_gem->aub_annotation_count; ++i) {
1979 drm_intel_aub_annotation *annotation =
1980 &bo_gem->aub_annotations[i];
1981 uint32_t ending_offset = annotation->ending_offset;
1982 if (ending_offset > bo->size)
1983 ending_offset = bo->size;
1984 if (ending_offset > offset) {
1985 aub_write_large_trace_block(bo, annotation->type,
1986 annotation->subtype,
1988 ending_offset - offset);
1989 offset = ending_offset;
1993 /* Write out any remaining unannotated data */
1994 if (offset < bo->size) {
1995 aub_write_large_trace_block(bo, AUB_TRACE_TYPE_NOTYPE, 0,
1996 offset, bo->size - offset);
2001 * Make a ringbuffer on fly and dump it
2004 aub_build_dump_ringbuffer(drm_intel_bufmgr_gem *bufmgr_gem,
2005 uint32_t batch_buffer, int ring_flag)
2007 uint32_t ringbuffer[4096];
2008 int ring = AUB_TRACE_TYPE_RING_PRB0; /* The default ring */
2011 if (ring_flag == I915_EXEC_BSD)
2012 ring = AUB_TRACE_TYPE_RING_PRB1;
2014 /* Make a ring buffer to execute our batchbuffer. */
2015 memset(ringbuffer, 0, sizeof(ringbuffer));
2016 ringbuffer[ring_count++] = AUB_MI_BATCH_BUFFER_START;
2017 ringbuffer[ring_count++] = batch_buffer;
2019 /* Write out the ring. This appears to trigger execution of
2020 * the ring in the simulator.
2023 CMD_AUB_TRACE_HEADER_BLOCK |
2026 AUB_TRACE_MEMTYPE_GTT | ring | AUB_TRACE_OP_COMMAND_WRITE);
2027 aub_out(bufmgr_gem, 0); /* general/surface subtype */
2028 aub_out(bufmgr_gem, bufmgr_gem->aub_offset);
2029 aub_out(bufmgr_gem, ring_count * 4);
2031 /* FIXME: Need some flush operations here? */
2032 aub_out_data(bufmgr_gem, ringbuffer, ring_count * 4);
2034 /* Update offset pointer */
2035 bufmgr_gem->aub_offset += 4096;
2039 drm_intel_gem_bo_aub_dump_bmp(drm_intel_bo *bo,
2040 int x1, int y1, int width, int height,
2041 enum aub_dump_bmp_format format,
2042 int pitch, int offset)
2044 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
2045 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
2049 case AUB_DUMP_BMP_FORMAT_8BIT:
2052 case AUB_DUMP_BMP_FORMAT_ARGB_4444:
2055 case AUB_DUMP_BMP_FORMAT_ARGB_0888:
2056 case AUB_DUMP_BMP_FORMAT_ARGB_8888:
2060 printf("Unknown AUB dump format %d\n", format);
2064 if (!bufmgr_gem->aub_file)
2067 aub_out(bufmgr_gem, CMD_AUB_DUMP_BMP | 4);
2068 aub_out(bufmgr_gem, (y1 << 16) | x1);
2073 aub_out(bufmgr_gem, (height << 16) | width);
2074 aub_out(bufmgr_gem, bo_gem->aub_offset + offset);
2076 ((bo_gem->tiling_mode != I915_TILING_NONE) ? (1 << 2) : 0) |
2077 ((bo_gem->tiling_mode == I915_TILING_Y) ? (1 << 3) : 0));
2081 aub_exec(drm_intel_bo *bo, int ring_flag, int used)
2083 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
2084 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2086 bool batch_buffer_needs_annotations;
2088 if (!bufmgr_gem->aub_file)
2091 /* If batch buffer is not annotated, annotate it the best we
2094 batch_buffer_needs_annotations = bo_gem->aub_annotation_count == 0;
2095 if (batch_buffer_needs_annotations) {
2096 drm_intel_aub_annotation annotations[2] = {
2097 { AUB_TRACE_TYPE_BATCH, 0, used },
2098 { AUB_TRACE_TYPE_NOTYPE, 0, bo->size }
2100 drm_intel_bufmgr_gem_set_aub_annotations(bo, annotations, 2);
2103 /* Write out all buffers to AUB memory */
2104 for (i = 0; i < bufmgr_gem->exec_count; i++) {
2105 aub_write_bo(bufmgr_gem->exec_bos[i]);
2108 /* Remove any annotations we added */
2109 if (batch_buffer_needs_annotations)
2110 drm_intel_bufmgr_gem_set_aub_annotations(bo, NULL, 0);
2112 /* Dump ring buffer */
2113 aub_build_dump_ringbuffer(bufmgr_gem, bo_gem->aub_offset, ring_flag);
2115 fflush(bufmgr_gem->aub_file);
2118 * One frame has been dumped. So reset the aub_offset for the next frame.
2120 * FIXME: Can we do this?
2122 bufmgr_gem->aub_offset = 0x10000;
2126 drm_intel_gem_bo_exec(drm_intel_bo *bo, int used,
2127 drm_clip_rect_t * cliprects, int num_cliprects, int DR4)
2129 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
2130 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2131 struct drm_i915_gem_execbuffer execbuf;
2134 if (bo_gem->has_error)
2137 pthread_mutex_lock(&bufmgr_gem->lock);
2138 /* Update indices and set up the validate list. */
2139 drm_intel_gem_bo_process_reloc(bo);
2141 /* Add the batch buffer to the validation list. There are no
2142 * relocations pointing to it.
2144 drm_intel_add_validate_buffer(bo);
2147 execbuf.buffers_ptr = (uintptr_t) bufmgr_gem->exec_objects;
2148 execbuf.buffer_count = bufmgr_gem->exec_count;
2149 execbuf.batch_start_offset = 0;
2150 execbuf.batch_len = used;
2151 execbuf.cliprects_ptr = (uintptr_t) cliprects;
2152 execbuf.num_cliprects = num_cliprects;
2156 ret = drmIoctl(bufmgr_gem->fd,
2157 DRM_IOCTL_I915_GEM_EXECBUFFER,
2161 if (errno == ENOSPC) {
2162 DBG("Execbuffer fails to pin. "
2163 "Estimate: %u. Actual: %u. Available: %u\n",
2164 drm_intel_gem_estimate_batch_space(bufmgr_gem->exec_bos,
2167 drm_intel_gem_compute_batch_space(bufmgr_gem->exec_bos,
2170 (unsigned int)bufmgr_gem->gtt_size);
2173 drm_intel_update_buffer_offsets(bufmgr_gem);
2175 if (bufmgr_gem->bufmgr.debug)
2176 drm_intel_gem_dump_validation_list(bufmgr_gem);
2178 for (i = 0; i < bufmgr_gem->exec_count; i++) {
2179 drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
2180 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2182 /* Disconnect the buffer from the validate list */
2183 bo_gem->validate_index = -1;
2184 bufmgr_gem->exec_bos[i] = NULL;
2186 bufmgr_gem->exec_count = 0;
2187 pthread_mutex_unlock(&bufmgr_gem->lock);
2193 do_exec2(drm_intel_bo *bo, int used, drm_intel_context *ctx,
2194 drm_clip_rect_t *cliprects, int num_cliprects, int DR4,
2197 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr;
2198 struct drm_i915_gem_execbuffer2 execbuf;
2202 switch (flags & 0x7) {
2206 if (!bufmgr_gem->has_blt)
2210 if (!bufmgr_gem->has_bsd)
2213 case I915_EXEC_RENDER:
2214 case I915_EXEC_DEFAULT:
2218 pthread_mutex_lock(&bufmgr_gem->lock);
2219 /* Update indices and set up the validate list. */
2220 drm_intel_gem_bo_process_reloc2(bo);
2222 /* Add the batch buffer to the validation list. There are no relocations
2225 drm_intel_add_validate_buffer2(bo, 0);
2228 execbuf.buffers_ptr = (uintptr_t)bufmgr_gem->exec2_objects;
2229 execbuf.buffer_count = bufmgr_gem->exec_count;
2230 execbuf.batch_start_offset = 0;
2231 execbuf.batch_len = used;
2232 execbuf.cliprects_ptr = (uintptr_t)cliprects;
2233 execbuf.num_cliprects = num_cliprects;
2236 execbuf.flags = flags;
2238 i915_execbuffer2_set_context_id(execbuf, 0);
2240 i915_execbuffer2_set_context_id(execbuf, ctx->ctx_id);
2243 aub_exec(bo, flags, used);
2245 if (bufmgr_gem->no_exec)
2246 goto skip_execution;
2248 ret = drmIoctl(bufmgr_gem->fd,
2249 DRM_IOCTL_I915_GEM_EXECBUFFER2,
2253 if (ret == -ENOSPC) {
2254 DBG("Execbuffer fails to pin. "
2255 "Estimate: %u. Actual: %u. Available: %u\n",
2256 drm_intel_gem_estimate_batch_space(bufmgr_gem->exec_bos,
2257 bufmgr_gem->exec_count),
2258 drm_intel_gem_compute_batch_space(bufmgr_gem->exec_bos,
2259 bufmgr_gem->exec_count),
2260 (unsigned int) bufmgr_gem->gtt_size);
2263 drm_intel_update_buffer_offsets2(bufmgr_gem);
2266 if (bufmgr_gem->bufmgr.debug)
2267 drm_intel_gem_dump_validation_list(bufmgr_gem);
2269 for (i = 0; i < bufmgr_gem->exec_count; i++) {
2270 drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
2271 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
2273 /* Disconnect the buffer from the validate list */
2274 bo_gem->validate_index = -1;
2275 bufmgr_gem->exec_bos[i] = NULL;
2277 bufmgr_gem->exec_count = 0;
2278 pthread_mutex_unlock(&bufmgr_gem->lock);
2284 drm_intel_gem_bo_exec2(drm_intel_bo *bo, int used,
2285 drm_clip_rect_t *cliprects, int num_cliprects,
2288 return do_exec2(bo, used, NULL, cliprects, num_cliprects, DR4,
2293 drm_intel_gem_bo_mrb_exec2(drm_intel_bo *bo, int used,
2294 drm_clip_rect_t *cliprects, int num_cliprects, int DR4,
2297 return do_exec2(bo, used, NULL, cliprects, num_cliprects, DR4,
2302 drm_intel_gem_bo_context_exec(drm_intel_bo *bo, drm_intel_context *ctx,
2303 int used, unsigned int flags)
2305 return do_exec2(bo, used, ctx, NULL, 0, 0, flags);
2309 drm_intel_gem_bo_pin(drm_intel_bo *bo, uint32_t alignment)
2311 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
2312 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2313 struct drm_i915_gem_pin pin;
2317 pin.handle = bo_gem->gem_handle;
2318 pin.alignment = alignment;
2320 ret = drmIoctl(bufmgr_gem->fd,
2321 DRM_IOCTL_I915_GEM_PIN,
2326 bo->offset = pin.offset;
2331 drm_intel_gem_bo_unpin(drm_intel_bo *bo)
2333 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
2334 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2335 struct drm_i915_gem_unpin unpin;
2339 unpin.handle = bo_gem->gem_handle;
2341 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_UNPIN, &unpin);
2349 drm_intel_gem_bo_set_tiling_internal(drm_intel_bo *bo,
2350 uint32_t tiling_mode,
2353 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
2354 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2355 struct drm_i915_gem_set_tiling set_tiling;
2358 if (bo_gem->global_name == 0 &&
2359 tiling_mode == bo_gem->tiling_mode &&
2360 stride == bo_gem->stride)
2363 memset(&set_tiling, 0, sizeof(set_tiling));
2365 /* set_tiling is slightly broken and overwrites the
2366 * input on the error path, so we have to open code
2369 set_tiling.handle = bo_gem->gem_handle;
2370 set_tiling.tiling_mode = tiling_mode;
2371 set_tiling.stride = stride;
2373 ret = ioctl(bufmgr_gem->fd,
2374 DRM_IOCTL_I915_GEM_SET_TILING,
2376 } while (ret == -1 && (errno == EINTR || errno == EAGAIN));
2380 bo_gem->tiling_mode = set_tiling.tiling_mode;
2381 bo_gem->swizzle_mode = set_tiling.swizzle_mode;
2382 bo_gem->stride = set_tiling.stride;
2387 drm_intel_gem_bo_set_tiling(drm_intel_bo *bo, uint32_t * tiling_mode,
2390 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
2391 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2394 /* Linear buffers have no stride. By ensuring that we only ever use
2395 * stride 0 with linear buffers, we simplify our code.
2397 if (*tiling_mode == I915_TILING_NONE)
2400 ret = drm_intel_gem_bo_set_tiling_internal(bo, *tiling_mode, stride);
2402 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
2404 *tiling_mode = bo_gem->tiling_mode;
2409 drm_intel_gem_bo_get_tiling(drm_intel_bo *bo, uint32_t * tiling_mode,
2410 uint32_t * swizzle_mode)
2412 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2414 *tiling_mode = bo_gem->tiling_mode;
2415 *swizzle_mode = bo_gem->swizzle_mode;
2420 drm_intel_bo_gem_create_from_prime(drm_intel_bufmgr *bufmgr, int prime_fd, int size)
2422 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
2425 drm_intel_bo_gem *bo_gem;
2426 struct drm_i915_gem_get_tiling get_tiling;
2428 ret = drmPrimeFDToHandle(bufmgr_gem->fd, prime_fd, &handle);
2430 fprintf(stderr,"ret is %d %d\n", ret, errno);
2434 bo_gem = calloc(1, sizeof(*bo_gem));
2438 bo_gem->bo.size = size;
2439 bo_gem->bo.handle = handle;
2440 bo_gem->bo.bufmgr = bufmgr;
2442 bo_gem->gem_handle = handle;
2444 atomic_set(&bo_gem->refcount, 1);
2446 bo_gem->name = "prime";
2447 bo_gem->validate_index = -1;
2448 bo_gem->reloc_tree_fences = 0;
2449 bo_gem->used_as_reloc_target = false;
2450 bo_gem->has_error = false;
2451 bo_gem->reusable = false;
2453 DRMINITLISTHEAD(&bo_gem->name_list);
2454 DRMINITLISTHEAD(&bo_gem->vma_list);
2456 VG_CLEAR(get_tiling);
2457 get_tiling.handle = bo_gem->gem_handle;
2458 ret = drmIoctl(bufmgr_gem->fd,
2459 DRM_IOCTL_I915_GEM_GET_TILING,
2462 drm_intel_gem_bo_unreference(&bo_gem->bo);
2465 bo_gem->tiling_mode = get_tiling.tiling_mode;
2466 bo_gem->swizzle_mode = get_tiling.swizzle_mode;
2467 /* XXX stride is unknown */
2468 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
2474 drm_intel_bo_gem_export_to_prime(drm_intel_bo *bo, int *prime_fd)
2476 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
2477 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2479 if (drmPrimeHandleToFD(bufmgr_gem->fd, bo_gem->gem_handle,
2480 DRM_CLOEXEC, prime_fd) != 0)
2483 bo_gem->reusable = false;
2489 drm_intel_gem_bo_flink(drm_intel_bo *bo, uint32_t * name)
2491 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
2492 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2495 if (!bo_gem->global_name) {
2496 struct drm_gem_flink flink;
2499 flink.handle = bo_gem->gem_handle;
2501 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_GEM_FLINK, &flink);
2505 bo_gem->global_name = flink.name;
2506 bo_gem->reusable = false;
2508 DRMLISTADDTAIL(&bo_gem->name_list, &bufmgr_gem->named);
2511 *name = bo_gem->global_name;
2516 * Enables unlimited caching of buffer objects for reuse.
2518 * This is potentially very memory expensive, as the cache at each bucket
2519 * size is only bounded by how many buffers of that size we've managed to have
2520 * in flight at once.
2523 drm_intel_bufmgr_gem_enable_reuse(drm_intel_bufmgr *bufmgr)
2525 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
2527 bufmgr_gem->bo_reuse = true;
2531 * Enable use of fenced reloc type.
2533 * New code should enable this to avoid unnecessary fence register
2534 * allocation. If this option is not enabled, all relocs will have fence
2535 * register allocated.
2538 drm_intel_bufmgr_gem_enable_fenced_relocs(drm_intel_bufmgr *bufmgr)
2540 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
2542 if (bufmgr_gem->bufmgr.bo_exec == drm_intel_gem_bo_exec2)
2543 bufmgr_gem->fenced_relocs = true;
2547 * Return the additional aperture space required by the tree of buffer objects
2551 drm_intel_gem_bo_get_aperture_space(drm_intel_bo *bo)
2553 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2557 if (bo == NULL || bo_gem->included_in_check_aperture)
2561 bo_gem->included_in_check_aperture = true;
2563 for (i = 0; i < bo_gem->reloc_count; i++)
2565 drm_intel_gem_bo_get_aperture_space(bo_gem->
2566 reloc_target_info[i].bo);
2572 * Count the number of buffers in this list that need a fence reg
2574 * If the count is greater than the number of available regs, we'll have
2575 * to ask the caller to resubmit a batch with fewer tiled buffers.
2577 * This function over-counts if the same buffer is used multiple times.
2580 drm_intel_gem_total_fences(drm_intel_bo ** bo_array, int count)
2583 unsigned int total = 0;
2585 for (i = 0; i < count; i++) {
2586 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo_array[i];
2591 total += bo_gem->reloc_tree_fences;
2597 * Clear the flag set by drm_intel_gem_bo_get_aperture_space() so we're ready
2598 * for the next drm_intel_bufmgr_check_aperture_space() call.
2601 drm_intel_gem_bo_clear_aperture_space_flag(drm_intel_bo *bo)
2603 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2606 if (bo == NULL || !bo_gem->included_in_check_aperture)
2609 bo_gem->included_in_check_aperture = false;
2611 for (i = 0; i < bo_gem->reloc_count; i++)
2612 drm_intel_gem_bo_clear_aperture_space_flag(bo_gem->
2613 reloc_target_info[i].bo);
2617 * Return a conservative estimate for the amount of aperture required
2618 * for a collection of buffers. This may double-count some buffers.
2621 drm_intel_gem_estimate_batch_space(drm_intel_bo **bo_array, int count)
2624 unsigned int total = 0;
2626 for (i = 0; i < count; i++) {
2627 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo_array[i];
2629 total += bo_gem->reloc_tree_size;
2635 * Return the amount of aperture needed for a collection of buffers.
2636 * This avoids double counting any buffers, at the cost of looking
2637 * at every buffer in the set.
2640 drm_intel_gem_compute_batch_space(drm_intel_bo **bo_array, int count)
2643 unsigned int total = 0;
2645 for (i = 0; i < count; i++) {
2646 total += drm_intel_gem_bo_get_aperture_space(bo_array[i]);
2647 /* For the first buffer object in the array, we get an
2648 * accurate count back for its reloc_tree size (since nothing
2649 * had been flagged as being counted yet). We can save that
2650 * value out as a more conservative reloc_tree_size that
2651 * avoids double-counting target buffers. Since the first
2652 * buffer happens to usually be the batch buffer in our
2653 * callers, this can pull us back from doing the tree
2654 * walk on every new batch emit.
2657 drm_intel_bo_gem *bo_gem =
2658 (drm_intel_bo_gem *) bo_array[i];
2659 bo_gem->reloc_tree_size = total;
2663 for (i = 0; i < count; i++)
2664 drm_intel_gem_bo_clear_aperture_space_flag(bo_array[i]);
2669 * Return -1 if the batchbuffer should be flushed before attempting to
2670 * emit rendering referencing the buffers pointed to by bo_array.
2672 * This is required because if we try to emit a batchbuffer with relocations
2673 * to a tree of buffers that won't simultaneously fit in the aperture,
2674 * the rendering will return an error at a point where the software is not
2675 * prepared to recover from it.
2677 * However, we also want to emit the batchbuffer significantly before we reach
2678 * the limit, as a series of batchbuffers each of which references buffers
2679 * covering almost all of the aperture means that at each emit we end up
2680 * waiting to evict a buffer from the last rendering, and we get synchronous
2681 * performance. By emitting smaller batchbuffers, we eat some CPU overhead to
2682 * get better parallelism.
2685 drm_intel_gem_check_aperture_space(drm_intel_bo **bo_array, int count)
2687 drm_intel_bufmgr_gem *bufmgr_gem =
2688 (drm_intel_bufmgr_gem *) bo_array[0]->bufmgr;
2689 unsigned int total = 0;
2690 unsigned int threshold = bufmgr_gem->gtt_size * 3 / 4;
2693 /* Check for fence reg constraints if necessary */
2694 if (bufmgr_gem->available_fences) {
2695 total_fences = drm_intel_gem_total_fences(bo_array, count);
2696 if (total_fences > bufmgr_gem->available_fences)
2700 total = drm_intel_gem_estimate_batch_space(bo_array, count);
2702 if (total > threshold)
2703 total = drm_intel_gem_compute_batch_space(bo_array, count);
2705 if (total > threshold) {
2706 DBG("check_space: overflowed available aperture, "
2708 total / 1024, (int)bufmgr_gem->gtt_size / 1024);
2711 DBG("drm_check_space: total %dkb vs bufgr %dkb\n", total / 1024,
2712 (int)bufmgr_gem->gtt_size / 1024);
2718 * Disable buffer reuse for objects which are shared with the kernel
2719 * as scanout buffers
2722 drm_intel_gem_bo_disable_reuse(drm_intel_bo *bo)
2724 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2726 bo_gem->reusable = false;
2731 drm_intel_gem_bo_is_reusable(drm_intel_bo *bo)
2733 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2735 return bo_gem->reusable;
2739 _drm_intel_gem_bo_references(drm_intel_bo *bo, drm_intel_bo *target_bo)
2741 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
2744 for (i = 0; i < bo_gem->reloc_count; i++) {
2745 if (bo_gem->reloc_target_info[i].bo == target_bo)
2747 if (bo == bo_gem->reloc_target_info[i].bo)
2749 if (_drm_intel_gem_bo_references(bo_gem->reloc_target_info[i].bo,
2757 /** Return true if target_bo is referenced by bo's relocation tree. */
2759 drm_intel_gem_bo_references(drm_intel_bo *bo, drm_intel_bo *target_bo)
2761 drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *) target_bo;
2763 if (bo == NULL || target_bo == NULL)
2765 if (target_bo_gem->used_as_reloc_target)
2766 return _drm_intel_gem_bo_references(bo, target_bo);
2771 add_bucket(drm_intel_bufmgr_gem *bufmgr_gem, int size)
2773 unsigned int i = bufmgr_gem->num_buckets;
2775 assert(i < ARRAY_SIZE(bufmgr_gem->cache_bucket));
2777 DRMINITLISTHEAD(&bufmgr_gem->cache_bucket[i].head);
2778 bufmgr_gem->cache_bucket[i].size = size;
2779 bufmgr_gem->num_buckets++;
2783 init_cache_buckets(drm_intel_bufmgr_gem *bufmgr_gem)
2785 unsigned long size, cache_max_size = 64 * 1024 * 1024;
2787 /* OK, so power of two buckets was too wasteful of memory.
2788 * Give 3 other sizes between each power of two, to hopefully
2789 * cover things accurately enough. (The alternative is
2790 * probably to just go for exact matching of sizes, and assume
2791 * that for things like composited window resize the tiled
2792 * width/height alignment and rounding of sizes to pages will
2793 * get us useful cache hit rates anyway)
2795 add_bucket(bufmgr_gem, 4096);
2796 add_bucket(bufmgr_gem, 4096 * 2);
2797 add_bucket(bufmgr_gem, 4096 * 3);
2799 /* Initialize the linked lists for BO reuse cache. */
2800 for (size = 4 * 4096; size <= cache_max_size; size *= 2) {
2801 add_bucket(bufmgr_gem, size);
2803 add_bucket(bufmgr_gem, size + size * 1 / 4);
2804 add_bucket(bufmgr_gem, size + size * 2 / 4);
2805 add_bucket(bufmgr_gem, size + size * 3 / 4);
2810 drm_intel_bufmgr_gem_set_vma_cache_size(drm_intel_bufmgr *bufmgr, int limit)
2812 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
2814 bufmgr_gem->vma_max = limit;
2816 drm_intel_gem_bo_purge_vma_cache(bufmgr_gem);
2820 * Get the PCI ID for the device. This can be overridden by setting the
2821 * INTEL_DEVID_OVERRIDE environment variable to the desired ID.
2824 get_pci_device_id(drm_intel_bufmgr_gem *bufmgr_gem)
2826 char *devid_override;
2829 drm_i915_getparam_t gp;
2831 if (geteuid() == getuid()) {
2832 devid_override = getenv("INTEL_DEVID_OVERRIDE");
2833 if (devid_override) {
2834 bufmgr_gem->no_exec = true;
2835 return strtod(devid_override, NULL);
2841 gp.param = I915_PARAM_CHIPSET_ID;
2843 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
2845 fprintf(stderr, "get chip id failed: %d [%d]\n", ret, errno);
2846 fprintf(stderr, "param: %d, val: %d\n", gp.param, *gp.value);
2852 drm_intel_bufmgr_gem_get_devid(drm_intel_bufmgr *bufmgr)
2854 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
2856 return bufmgr_gem->pci_device;
2860 * Sets up AUB dumping.
2862 * This is a trace file format that can be used with the simulator.
2863 * Packets are emitted in a format somewhat like GPU command packets.
2864 * You can set up a GTT and upload your objects into the referenced
2865 * space, then send off batchbuffers and get BMPs out the other end.
2868 drm_intel_bufmgr_gem_set_aub_dump(drm_intel_bufmgr *bufmgr, int enable)
2870 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
2871 int entry = 0x200003;
2873 int gtt_size = 0x10000;
2876 if (bufmgr_gem->aub_file) {
2877 fclose(bufmgr_gem->aub_file);
2878 bufmgr_gem->aub_file = NULL;
2882 if (geteuid() != getuid())
2885 bufmgr_gem->aub_file = fopen("intel.aub", "w+");
2886 if (!bufmgr_gem->aub_file)
2889 /* Start allocating objects from just after the GTT. */
2890 bufmgr_gem->aub_offset = gtt_size;
2892 /* Start with a (required) version packet. */
2893 aub_out(bufmgr_gem, CMD_AUB_HEADER | (13 - 2));
2895 (4 << AUB_HEADER_MAJOR_SHIFT) |
2896 (0 << AUB_HEADER_MINOR_SHIFT));
2897 for (i = 0; i < 8; i++) {
2898 aub_out(bufmgr_gem, 0); /* app name */
2900 aub_out(bufmgr_gem, 0); /* timestamp */
2901 aub_out(bufmgr_gem, 0); /* timestamp */
2902 aub_out(bufmgr_gem, 0); /* comment len */
2904 /* Set up the GTT. The max we can handle is 256M */
2905 aub_out(bufmgr_gem, CMD_AUB_TRACE_HEADER_BLOCK | (5 - 2));
2906 aub_out(bufmgr_gem, AUB_TRACE_MEMTYPE_NONLOCAL | 0 | AUB_TRACE_OP_DATA_WRITE);
2907 aub_out(bufmgr_gem, 0); /* subtype */
2908 aub_out(bufmgr_gem, 0); /* offset */
2909 aub_out(bufmgr_gem, gtt_size); /* size */
2910 for (i = 0x000; i < gtt_size; i += 4, entry += 0x1000) {
2911 aub_out(bufmgr_gem, entry);
2916 drm_intel_gem_context_create(drm_intel_bufmgr *bufmgr)
2918 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
2919 struct drm_i915_gem_context_create create;
2920 drm_intel_context *context = NULL;
2924 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_CONTEXT_CREATE, &create);
2926 DBG("DRM_IOCTL_I915_GEM_CONTEXT_CREATE failed: %s\n",
2931 context = calloc(1, sizeof(*context));
2932 context->ctx_id = create.ctx_id;
2933 context->bufmgr = bufmgr;
2939 drm_intel_gem_context_destroy(drm_intel_context *ctx)
2941 drm_intel_bufmgr_gem *bufmgr_gem;
2942 struct drm_i915_gem_context_destroy destroy;
2950 bufmgr_gem = (drm_intel_bufmgr_gem *)ctx->bufmgr;
2951 destroy.ctx_id = ctx->ctx_id;
2952 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_CONTEXT_DESTROY,
2955 fprintf(stderr, "DRM_IOCTL_I915_GEM_CONTEXT_DESTROY failed: %s\n",
2962 drm_intel_reg_read(drm_intel_bufmgr *bufmgr,
2966 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
2967 struct drm_i915_reg_read reg_read;
2971 reg_read.offset = offset;
2973 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_REG_READ, ®_read);
2975 *result = reg_read.val;
2981 * Annotate the given bo for use in aub dumping.
2983 * \param annotations is an array of drm_intel_aub_annotation objects
2984 * describing the type of data in various sections of the bo. Each
2985 * element of the array specifies the type and subtype of a section of
2986 * the bo, and the past-the-end offset of that section. The elements
2987 * of \c annotations must be sorted so that ending_offset is
2990 * \param count is the number of elements in the \c annotations array.
2991 * If \c count is zero, then \c annotations will not be dereferenced.
2993 * Annotations are copied into a private data structure, so caller may
2994 * re-use the memory pointed to by \c annotations after the call
2997 * Annotations are stored for the lifetime of the bo; to reset to the
2998 * default state (no annotations), call this function with a \c count
3002 drm_intel_bufmgr_gem_set_aub_annotations(drm_intel_bo *bo,
3003 drm_intel_aub_annotation *annotations,
3006 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
3007 unsigned size = sizeof(*annotations) * count;
3008 drm_intel_aub_annotation *new_annotations =
3009 count > 0 ? realloc(bo_gem->aub_annotations, size) : NULL;
3010 if (new_annotations == NULL) {
3011 free(bo_gem->aub_annotations);
3012 bo_gem->aub_annotations = NULL;
3013 bo_gem->aub_annotation_count = 0;
3016 memcpy(new_annotations, annotations, size);
3017 bo_gem->aub_annotations = new_annotations;
3018 bo_gem->aub_annotation_count = count;
3022 * Initializes the GEM buffer manager, which uses the kernel to allocate, map,
3023 * and manage map buffer objections.
3025 * \param fd File descriptor of the opened DRM device.
3028 drm_intel_bufmgr_gem_init(int fd, int batch_size)
3030 drm_intel_bufmgr_gem *bufmgr_gem;
3031 struct drm_i915_gem_get_aperture aperture;
3032 drm_i915_getparam_t gp;
3036 bufmgr_gem = calloc(1, sizeof(*bufmgr_gem));
3037 if (bufmgr_gem == NULL)
3040 bufmgr_gem->fd = fd;
3042 if (pthread_mutex_init(&bufmgr_gem->lock, NULL) != 0) {
3047 ret = drmIoctl(bufmgr_gem->fd,
3048 DRM_IOCTL_I915_GEM_GET_APERTURE,
3052 bufmgr_gem->gtt_size = aperture.aper_available_size;
3054 fprintf(stderr, "DRM_IOCTL_I915_GEM_APERTURE failed: %s\n",
3056 bufmgr_gem->gtt_size = 128 * 1024 * 1024;
3057 fprintf(stderr, "Assuming %dkB available aperture size.\n"
3058 "May lead to reduced performance or incorrect "
3060 (int)bufmgr_gem->gtt_size / 1024);
3063 bufmgr_gem->pci_device = get_pci_device_id(bufmgr_gem);
3065 if (IS_GEN2(bufmgr_gem->pci_device))
3066 bufmgr_gem->gen = 2;
3067 else if (IS_GEN3(bufmgr_gem->pci_device))
3068 bufmgr_gem->gen = 3;
3069 else if (IS_GEN4(bufmgr_gem->pci_device))
3070 bufmgr_gem->gen = 4;
3071 else if (IS_GEN5(bufmgr_gem->pci_device))
3072 bufmgr_gem->gen = 5;
3073 else if (IS_GEN6(bufmgr_gem->pci_device))
3074 bufmgr_gem->gen = 6;
3075 else if (IS_GEN7(bufmgr_gem->pci_device))
3076 bufmgr_gem->gen = 7;
3082 if (IS_GEN3(bufmgr_gem->pci_device) &&
3083 bufmgr_gem->gtt_size > 256*1024*1024) {
3084 /* The unmappable part of gtt on gen 3 (i.e. above 256MB) can't
3085 * be used for tiled blits. To simplify the accounting, just
3086 * substract the unmappable part (fixed to 256MB on all known
3087 * gen3 devices) if the kernel advertises it. */
3088 bufmgr_gem->gtt_size -= 256*1024*1024;
3094 gp.param = I915_PARAM_HAS_EXECBUF2;
3095 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
3099 gp.param = I915_PARAM_HAS_BSD;
3100 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
3101 bufmgr_gem->has_bsd = ret == 0;
3103 gp.param = I915_PARAM_HAS_BLT;
3104 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
3105 bufmgr_gem->has_blt = ret == 0;
3107 gp.param = I915_PARAM_HAS_RELAXED_FENCING;
3108 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
3109 bufmgr_gem->has_relaxed_fencing = ret == 0;
3111 gp.param = I915_PARAM_HAS_WAIT_TIMEOUT;
3112 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
3113 bufmgr_gem->has_wait_timeout = ret == 0;
3115 gp.param = I915_PARAM_HAS_LLC;
3116 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
3118 /* Kernel does not supports HAS_LLC query, fallback to GPU
3119 * generation detection and assume that we have LLC on GEN6/7
3121 bufmgr_gem->has_llc = (IS_GEN6(bufmgr_gem->pci_device) |
3122 IS_GEN7(bufmgr_gem->pci_device));
3124 bufmgr_gem->has_llc = *gp.value;
3126 if (bufmgr_gem->gen < 4) {
3127 gp.param = I915_PARAM_NUM_FENCES_AVAIL;
3128 gp.value = &bufmgr_gem->available_fences;
3129 ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
3131 fprintf(stderr, "get fences failed: %d [%d]\n", ret,
3133 fprintf(stderr, "param: %d, val: %d\n", gp.param,
3135 bufmgr_gem->available_fences = 0;
3137 /* XXX The kernel reports the total number of fences,
3138 * including any that may be pinned.
3140 * We presume that there will be at least one pinned
3141 * fence for the scanout buffer, but there may be more
3142 * than one scanout and the user may be manually
3143 * pinning buffers. Let's move to execbuffer2 and
3144 * thereby forget the insanity of using fences...
3146 bufmgr_gem->available_fences -= 2;
3147 if (bufmgr_gem->available_fences < 0)
3148 bufmgr_gem->available_fences = 0;
3152 /* Let's go with one relocation per every 2 dwords (but round down a bit
3153 * since a power of two will mean an extra page allocation for the reloc
3156 * Every 4 was too few for the blender benchmark.
3158 bufmgr_gem->max_relocs = batch_size / sizeof(uint32_t) / 2 - 2;
3160 bufmgr_gem->bufmgr.bo_alloc = drm_intel_gem_bo_alloc;
3161 bufmgr_gem->bufmgr.bo_alloc_for_render =
3162 drm_intel_gem_bo_alloc_for_render;
3163 bufmgr_gem->bufmgr.bo_alloc_tiled = drm_intel_gem_bo_alloc_tiled;
3164 bufmgr_gem->bufmgr.bo_reference = drm_intel_gem_bo_reference;
3165 bufmgr_gem->bufmgr.bo_unreference = drm_intel_gem_bo_unreference;
3166 bufmgr_gem->bufmgr.bo_map = drm_intel_gem_bo_map;
3167 bufmgr_gem->bufmgr.bo_unmap = drm_intel_gem_bo_unmap;
3168 bufmgr_gem->bufmgr.bo_subdata = drm_intel_gem_bo_subdata;
3169 bufmgr_gem->bufmgr.bo_get_subdata = drm_intel_gem_bo_get_subdata;
3170 bufmgr_gem->bufmgr.bo_wait_rendering = drm_intel_gem_bo_wait_rendering;
3171 bufmgr_gem->bufmgr.bo_emit_reloc = drm_intel_gem_bo_emit_reloc;
3172 bufmgr_gem->bufmgr.bo_emit_reloc_fence = drm_intel_gem_bo_emit_reloc_fence;
3173 bufmgr_gem->bufmgr.bo_pin = drm_intel_gem_bo_pin;
3174 bufmgr_gem->bufmgr.bo_unpin = drm_intel_gem_bo_unpin;
3175 bufmgr_gem->bufmgr.bo_get_tiling = drm_intel_gem_bo_get_tiling;
3176 bufmgr_gem->bufmgr.bo_set_tiling = drm_intel_gem_bo_set_tiling;
3177 bufmgr_gem->bufmgr.bo_flink = drm_intel_gem_bo_flink;
3178 /* Use the new one if available */
3180 bufmgr_gem->bufmgr.bo_exec = drm_intel_gem_bo_exec2;
3181 bufmgr_gem->bufmgr.bo_mrb_exec = drm_intel_gem_bo_mrb_exec2;
3183 bufmgr_gem->bufmgr.bo_exec = drm_intel_gem_bo_exec;
3184 bufmgr_gem->bufmgr.bo_busy = drm_intel_gem_bo_busy;
3185 bufmgr_gem->bufmgr.bo_madvise = drm_intel_gem_bo_madvise;
3186 bufmgr_gem->bufmgr.destroy = drm_intel_bufmgr_gem_destroy;
3187 bufmgr_gem->bufmgr.debug = 0;
3188 bufmgr_gem->bufmgr.check_aperture_space =
3189 drm_intel_gem_check_aperture_space;
3190 bufmgr_gem->bufmgr.bo_disable_reuse = drm_intel_gem_bo_disable_reuse;
3191 bufmgr_gem->bufmgr.bo_is_reusable = drm_intel_gem_bo_is_reusable;
3192 bufmgr_gem->bufmgr.get_pipe_from_crtc_id =
3193 drm_intel_gem_get_pipe_from_crtc_id;
3194 bufmgr_gem->bufmgr.bo_references = drm_intel_gem_bo_references;
3196 DRMINITLISTHEAD(&bufmgr_gem->named);
3197 init_cache_buckets(bufmgr_gem);
3199 DRMINITLISTHEAD(&bufmgr_gem->vma_cache);
3200 bufmgr_gem->vma_max = -1; /* unlimited by default */
3202 return &bufmgr_gem->bufmgr;