1 /**************************************************************************
3 * Copyright © 2007 Red Hat Inc.
4 * Copyright © 2007 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>
56 #include "libdrm_lists.h"
57 #include "intel_bufmgr.h"
58 #include "intel_bufmgr_priv.h"
59 #include "intel_chipset.h"
64 #define DBG(...) do { \
65 if (bufmgr_gem->bufmgr.debug) \
66 fprintf(stderr, __VA_ARGS__); \
69 typedef struct _drm_intel_bo_gem drm_intel_bo_gem;
71 struct drm_intel_gem_bo_bucket {
76 /* Only cache objects up to 64MB. Bigger than that, and the rounding of the
77 * size makes many operations fail that wouldn't otherwise.
79 #define DRM_INTEL_GEM_BO_BUCKETS 14
80 typedef struct _drm_intel_bufmgr_gem {
81 drm_intel_bufmgr bufmgr;
89 struct drm_i915_gem_exec_object *exec_objects;
90 struct drm_i915_gem_exec_object2 *exec2_objects;
91 drm_intel_bo **exec_bos;
95 /** Array of lists of cached gem objects of power-of-two sizes */
96 struct drm_intel_gem_bo_bucket cache_bucket[DRM_INTEL_GEM_BO_BUCKETS];
104 } drm_intel_bufmgr_gem;
106 #define DRM_INTEL_RELOC_FENCE (1<<0)
108 typedef struct _drm_intel_reloc_target_info {
111 } drm_intel_reloc_target;
113 struct _drm_intel_bo_gem {
121 * Kenel-assigned global name for this object
123 unsigned int global_name;
126 * Index of the buffer within the validation list while preparing a
127 * batchbuffer execution.
132 * Current tiling mode
134 uint32_t tiling_mode;
135 uint32_t swizzle_mode;
139 /** Array passed to the DRM containing relocation information. */
140 struct drm_i915_gem_relocation_entry *relocs;
142 * Array of info structs corresponding to relocs[i].target_handle etc
144 drm_intel_reloc_target *reloc_target_info;
145 /** Number of entries in relocs */
147 /** Mapped address for the buffer, saved across map/unmap cycles */
149 /** GTT virtual address for the buffer, saved across map/unmap cycles */
156 * Boolean of whether this BO and its children have been included in
157 * the current drm_intel_bufmgr_check_aperture_space() total.
159 char included_in_check_aperture;
162 * Boolean of whether this buffer has been used as a relocation
163 * target and had its size accounted for, and thus can't have any
164 * further relocations added to it.
166 char used_as_reloc_target;
169 * Boolean of whether we have encountered an error whilst building the relocation tree.
174 * Boolean of whether this buffer can be re-used
179 * Size in bytes of this buffer and its relocation descendents.
181 * Used to avoid costly tree walking in
182 * drm_intel_bufmgr_check_aperture in the common case.
187 * Number of potential fence registers required by this buffer and its
190 int reloc_tree_fences;
194 drm_intel_gem_estimate_batch_space(drm_intel_bo ** bo_array, int count);
197 drm_intel_gem_compute_batch_space(drm_intel_bo ** bo_array, int count);
200 drm_intel_gem_bo_get_tiling(drm_intel_bo *bo, uint32_t * tiling_mode,
201 uint32_t * swizzle_mode);
204 drm_intel_gem_bo_set_tiling(drm_intel_bo *bo, uint32_t * tiling_mode,
207 static void drm_intel_gem_bo_unreference_locked_timed(drm_intel_bo *bo,
210 static void drm_intel_gem_bo_unreference(drm_intel_bo *bo);
212 static void drm_intel_gem_bo_free(drm_intel_bo *bo);
215 drm_intel_gem_bo_tile_size(drm_intel_bufmgr_gem *bufmgr_gem, unsigned long size,
216 uint32_t *tiling_mode)
218 unsigned long min_size, max_size;
221 if (*tiling_mode == I915_TILING_NONE)
224 /* 965+ just need multiples of page size for tiling */
225 if (bufmgr_gem->gen >= 4)
226 return ROUND_UP_TO(size, 4096);
228 /* Older chips need powers of two, of at least 512k or 1M */
229 if (bufmgr_gem->gen == 3) {
230 min_size = 1024*1024;
231 max_size = 128*1024*1024;
234 max_size = 64*1024*1024;
237 if (size > max_size) {
238 *tiling_mode = I915_TILING_NONE;
242 for (i = min_size; i < size; i <<= 1)
249 * Round a given pitch up to the minimum required for X tiling on a
250 * given chip. We use 512 as the minimum to allow for a later tiling
254 drm_intel_gem_bo_tile_pitch(drm_intel_bufmgr_gem *bufmgr_gem,
255 unsigned long pitch, uint32_t tiling_mode)
257 unsigned long tile_width;
260 /* If untiled, then just align it so that we can do rendering
261 * to it with the 3D engine.
263 if (tiling_mode == I915_TILING_NONE)
264 return ALIGN(pitch, 64);
266 if (tiling_mode == I915_TILING_X)
271 /* 965 is flexible */
272 if (bufmgr_gem->gen >= 4)
273 return ROUND_UP_TO(pitch, tile_width);
275 /* Pre-965 needs power of two tile width */
276 for (i = tile_width; i < pitch; i <<= 1)
282 static struct drm_intel_gem_bo_bucket *
283 drm_intel_gem_bo_bucket_for_size(drm_intel_bufmgr_gem *bufmgr_gem,
288 for (i = 0; i < DRM_INTEL_GEM_BO_BUCKETS; i++) {
289 struct drm_intel_gem_bo_bucket *bucket =
290 &bufmgr_gem->cache_bucket[i];
291 if (bucket->size >= size) {
300 drm_intel_gem_dump_validation_list(drm_intel_bufmgr_gem *bufmgr_gem)
304 for (i = 0; i < bufmgr_gem->exec_count; i++) {
305 drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
306 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
308 if (bo_gem->relocs == NULL) {
309 DBG("%2d: %d (%s)\n", i, bo_gem->gem_handle,
314 for (j = 0; j < bo_gem->reloc_count; j++) {
315 drm_intel_bo *target_bo = bo_gem->reloc_target_info[j].bo;
316 drm_intel_bo_gem *target_gem =
317 (drm_intel_bo_gem *) target_bo;
319 DBG("%2d: %d (%s)@0x%08llx -> "
320 "%d (%s)@0x%08lx + 0x%08x\n",
322 bo_gem->gem_handle, bo_gem->name,
323 (unsigned long long)bo_gem->relocs[j].offset,
324 target_gem->gem_handle,
327 bo_gem->relocs[j].delta);
333 drm_intel_gem_bo_reference(drm_intel_bo *bo)
335 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
337 assert(atomic_read(&bo_gem->refcount) > 0);
338 atomic_inc(&bo_gem->refcount);
342 * Adds the given buffer to the list of buffers to be validated (moved into the
343 * appropriate memory type) with the next batch submission.
345 * If a buffer is validated multiple times in a batch submission, it ends up
346 * with the intersection of the memory type flags and the union of the
350 drm_intel_add_validate_buffer(drm_intel_bo *bo)
352 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
353 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
356 if (bo_gem->validate_index != -1)
359 /* Extend the array of validation entries as necessary. */
360 if (bufmgr_gem->exec_count == bufmgr_gem->exec_size) {
361 int new_size = bufmgr_gem->exec_size * 2;
366 bufmgr_gem->exec_objects =
367 realloc(bufmgr_gem->exec_objects,
368 sizeof(*bufmgr_gem->exec_objects) * new_size);
369 bufmgr_gem->exec_bos =
370 realloc(bufmgr_gem->exec_bos,
371 sizeof(*bufmgr_gem->exec_bos) * new_size);
372 bufmgr_gem->exec_size = new_size;
375 index = bufmgr_gem->exec_count;
376 bo_gem->validate_index = index;
377 /* Fill in array entry */
378 bufmgr_gem->exec_objects[index].handle = bo_gem->gem_handle;
379 bufmgr_gem->exec_objects[index].relocation_count = bo_gem->reloc_count;
380 bufmgr_gem->exec_objects[index].relocs_ptr = (uintptr_t) bo_gem->relocs;
381 bufmgr_gem->exec_objects[index].alignment = 0;
382 bufmgr_gem->exec_objects[index].offset = 0;
383 bufmgr_gem->exec_bos[index] = bo;
384 bufmgr_gem->exec_count++;
388 drm_intel_add_validate_buffer2(drm_intel_bo *bo, int need_fence)
390 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr;
391 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
394 if (bo_gem->validate_index != -1) {
396 bufmgr_gem->exec2_objects[bo_gem->validate_index].flags |=
397 EXEC_OBJECT_NEEDS_FENCE;
401 /* Extend the array of validation entries as necessary. */
402 if (bufmgr_gem->exec_count == bufmgr_gem->exec_size) {
403 int new_size = bufmgr_gem->exec_size * 2;
408 bufmgr_gem->exec2_objects =
409 realloc(bufmgr_gem->exec2_objects,
410 sizeof(*bufmgr_gem->exec2_objects) * new_size);
411 bufmgr_gem->exec_bos =
412 realloc(bufmgr_gem->exec_bos,
413 sizeof(*bufmgr_gem->exec_bos) * new_size);
414 bufmgr_gem->exec_size = new_size;
417 index = bufmgr_gem->exec_count;
418 bo_gem->validate_index = index;
419 /* Fill in array entry */
420 bufmgr_gem->exec2_objects[index].handle = bo_gem->gem_handle;
421 bufmgr_gem->exec2_objects[index].relocation_count = bo_gem->reloc_count;
422 bufmgr_gem->exec2_objects[index].relocs_ptr = (uintptr_t)bo_gem->relocs;
423 bufmgr_gem->exec2_objects[index].alignment = 0;
424 bufmgr_gem->exec2_objects[index].offset = 0;
425 bufmgr_gem->exec_bos[index] = bo;
426 bufmgr_gem->exec2_objects[index].flags = 0;
427 bufmgr_gem->exec2_objects[index].rsvd1 = 0;
428 bufmgr_gem->exec2_objects[index].rsvd2 = 0;
430 bufmgr_gem->exec2_objects[index].flags |=
431 EXEC_OBJECT_NEEDS_FENCE;
433 bufmgr_gem->exec_count++;
436 #define RELOC_BUF_SIZE(x) ((I915_RELOC_HEADER + x * I915_RELOC0_STRIDE) * \
440 drm_intel_bo_gem_set_in_aperture_size(drm_intel_bufmgr_gem *bufmgr_gem,
441 drm_intel_bo_gem *bo_gem)
445 assert(!bo_gem->used_as_reloc_target);
447 /* The older chipsets are far-less flexible in terms of tiling,
448 * and require tiled buffer to be size aligned in the aperture.
449 * This means that in the worst possible case we will need a hole
450 * twice as large as the object in order for it to fit into the
451 * aperture. Optimal packing is for wimps.
453 size = bo_gem->bo.size;
454 if (bufmgr_gem->gen < 4 && bo_gem->tiling_mode != I915_TILING_NONE)
457 bo_gem->reloc_tree_size = size;
461 drm_intel_setup_reloc_list(drm_intel_bo *bo)
463 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
464 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
465 unsigned int max_relocs = bufmgr_gem->max_relocs;
467 if (bo->size / 4 < max_relocs)
468 max_relocs = bo->size / 4;
470 bo_gem->relocs = malloc(max_relocs *
471 sizeof(struct drm_i915_gem_relocation_entry));
472 bo_gem->reloc_target_info = malloc(max_relocs *
473 sizeof(drm_intel_reloc_target));
474 if (bo_gem->relocs == NULL || bo_gem->reloc_target_info == NULL) {
475 bo_gem->has_error = 1;
477 free (bo_gem->relocs);
478 bo_gem->relocs = NULL;
480 free (bo_gem->reloc_target_info);
481 bo_gem->reloc_target_info = NULL;
490 drm_intel_gem_bo_busy(drm_intel_bo *bo)
492 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
493 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
494 struct drm_i915_gem_busy busy;
497 memset(&busy, 0, sizeof(busy));
498 busy.handle = bo_gem->gem_handle;
501 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_BUSY, &busy);
502 } while (ret == -1 && errno == EINTR);
504 return (ret == 0 && busy.busy);
508 drm_intel_gem_bo_madvise_internal(drm_intel_bufmgr_gem *bufmgr_gem,
509 drm_intel_bo_gem *bo_gem, int state)
511 struct drm_i915_gem_madvise madv;
513 madv.handle = bo_gem->gem_handle;
516 ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv);
518 return madv.retained;
522 drm_intel_gem_bo_madvise(drm_intel_bo *bo, int madv)
524 return drm_intel_gem_bo_madvise_internal
525 ((drm_intel_bufmgr_gem *) bo->bufmgr,
526 (drm_intel_bo_gem *) bo,
530 /* drop the oldest entries that have been purged by the kernel */
532 drm_intel_gem_bo_cache_purge_bucket(drm_intel_bufmgr_gem *bufmgr_gem,
533 struct drm_intel_gem_bo_bucket *bucket)
535 while (!DRMLISTEMPTY(&bucket->head)) {
536 drm_intel_bo_gem *bo_gem;
538 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
539 bucket->head.next, head);
540 if (drm_intel_gem_bo_madvise_internal
541 (bufmgr_gem, bo_gem, I915_MADV_DONTNEED))
544 DRMLISTDEL(&bo_gem->head);
545 drm_intel_gem_bo_free(&bo_gem->bo);
549 static drm_intel_bo *
550 drm_intel_gem_bo_alloc_internal(drm_intel_bufmgr *bufmgr,
555 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
556 drm_intel_bo_gem *bo_gem;
557 unsigned int page_size = getpagesize();
559 struct drm_intel_gem_bo_bucket *bucket;
560 int alloc_from_cache;
561 unsigned long bo_size;
564 if (flags & BO_ALLOC_FOR_RENDER)
567 /* Round the allocated size up to a power of two number of pages. */
568 bucket = drm_intel_gem_bo_bucket_for_size(bufmgr_gem, size);
570 /* If we don't have caching at this size, don't actually round the
573 if (bucket == NULL) {
575 if (bo_size < page_size)
578 bo_size = bucket->size;
581 pthread_mutex_lock(&bufmgr_gem->lock);
582 /* Get a buffer out of the cache if available */
584 alloc_from_cache = 0;
585 if (bucket != NULL && !DRMLISTEMPTY(&bucket->head)) {
587 /* Allocate new render-target BOs from the tail (MRU)
588 * of the list, as it will likely be hot in the GPU
589 * cache and in the aperture for us.
591 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
592 bucket->head.prev, head);
593 DRMLISTDEL(&bo_gem->head);
594 alloc_from_cache = 1;
596 /* For non-render-target BOs (where we're probably
597 * going to map it first thing in order to fill it
598 * with data), check if the last BO in the cache is
599 * unbusy, and only reuse in that case. Otherwise,
600 * allocating a new buffer is probably faster than
601 * waiting for the GPU to finish.
603 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
604 bucket->head.next, head);
605 if (!drm_intel_gem_bo_busy(&bo_gem->bo)) {
606 alloc_from_cache = 1;
607 DRMLISTDEL(&bo_gem->head);
611 if (alloc_from_cache) {
612 if (!drm_intel_gem_bo_madvise_internal
613 (bufmgr_gem, bo_gem, I915_MADV_WILLNEED)) {
614 drm_intel_gem_bo_free(&bo_gem->bo);
615 drm_intel_gem_bo_cache_purge_bucket(bufmgr_gem,
621 pthread_mutex_unlock(&bufmgr_gem->lock);
623 if (!alloc_from_cache) {
624 struct drm_i915_gem_create create;
626 bo_gem = calloc(1, sizeof(*bo_gem));
630 bo_gem->bo.size = bo_size;
631 memset(&create, 0, sizeof(create));
632 create.size = bo_size;
635 ret = ioctl(bufmgr_gem->fd,
636 DRM_IOCTL_I915_GEM_CREATE,
638 } while (ret == -1 && errno == EINTR);
639 bo_gem->gem_handle = create.handle;
640 bo_gem->bo.handle = bo_gem->gem_handle;
645 bo_gem->bo.bufmgr = bufmgr;
649 atomic_set(&bo_gem->refcount, 1);
650 bo_gem->validate_index = -1;
651 bo_gem->reloc_tree_fences = 0;
652 bo_gem->used_as_reloc_target = 0;
653 bo_gem->has_error = 0;
654 bo_gem->tiling_mode = I915_TILING_NONE;
655 bo_gem->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
656 bo_gem->reusable = 1;
658 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
660 DBG("bo_create: buf %d (%s) %ldb\n",
661 bo_gem->gem_handle, bo_gem->name, size);
666 static drm_intel_bo *
667 drm_intel_gem_bo_alloc_for_render(drm_intel_bufmgr *bufmgr,
670 unsigned int alignment)
672 return drm_intel_gem_bo_alloc_internal(bufmgr, name, size,
673 BO_ALLOC_FOR_RENDER);
676 static drm_intel_bo *
677 drm_intel_gem_bo_alloc(drm_intel_bufmgr *bufmgr,
680 unsigned int alignment)
682 return drm_intel_gem_bo_alloc_internal(bufmgr, name, size, 0);
685 static drm_intel_bo *
686 drm_intel_gem_bo_alloc_tiled(drm_intel_bufmgr *bufmgr, const char *name,
687 int x, int y, int cpp, uint32_t *tiling_mode,
688 unsigned long *pitch, unsigned long flags)
690 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
692 unsigned long size, stride;
697 unsigned long aligned_y;
699 tiling = *tiling_mode;
701 /* If we're tiled, our allocations are in 8 or 32-row blocks,
702 * so failure to align our height means that we won't allocate
705 * If we're untiled, we still have to align to 2 rows high
706 * because the data port accesses 2x2 blocks even if the
707 * bottom row isn't to be rendered, so failure to align means
708 * we could walk off the end of the GTT and fault. This is
709 * documented on 965, and may be the case on older chipsets
710 * too so we try to be careful.
713 if (tiling == I915_TILING_NONE)
714 aligned_y = ALIGN(y, 2);
715 else if (tiling == I915_TILING_X)
716 aligned_y = ALIGN(y, 8);
717 else if (tiling == I915_TILING_Y)
718 aligned_y = ALIGN(y, 32);
721 stride = drm_intel_gem_bo_tile_pitch(bufmgr_gem, stride, tiling);
722 size = stride * aligned_y;
723 size = drm_intel_gem_bo_tile_size(bufmgr_gem, size, tiling_mode);
724 } while (*tiling_mode != tiling);
726 bo = drm_intel_gem_bo_alloc_internal(bufmgr, name, size, flags);
730 ret = drm_intel_gem_bo_set_tiling(bo, tiling_mode, stride);
732 drm_intel_gem_bo_unreference(bo);
742 * Returns a drm_intel_bo wrapping the given buffer object handle.
744 * This can be used when one application needs to pass a buffer object
748 drm_intel_bo_gem_create_from_name(drm_intel_bufmgr *bufmgr,
752 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
753 drm_intel_bo_gem *bo_gem;
755 struct drm_gem_open open_arg;
756 struct drm_i915_gem_get_tiling get_tiling;
758 bo_gem = calloc(1, sizeof(*bo_gem));
762 memset(&open_arg, 0, sizeof(open_arg));
763 open_arg.name = handle;
765 ret = ioctl(bufmgr_gem->fd,
768 } while (ret == -1 && errno == EINTR);
770 fprintf(stderr, "Couldn't reference %s handle 0x%08x: %s\n",
771 name, handle, strerror(errno));
775 bo_gem->bo.size = open_arg.size;
776 bo_gem->bo.offset = 0;
777 bo_gem->bo.virtual = NULL;
778 bo_gem->bo.bufmgr = bufmgr;
780 atomic_set(&bo_gem->refcount, 1);
781 bo_gem->validate_index = -1;
782 bo_gem->gem_handle = open_arg.handle;
783 bo_gem->global_name = handle;
784 bo_gem->reusable = 0;
786 memset(&get_tiling, 0, sizeof(get_tiling));
787 get_tiling.handle = bo_gem->gem_handle;
788 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_GET_TILING, &get_tiling);
790 drm_intel_gem_bo_unreference(&bo_gem->bo);
793 bo_gem->tiling_mode = get_tiling.tiling_mode;
794 bo_gem->swizzle_mode = get_tiling.swizzle_mode;
795 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
797 DBG("bo_create_from_handle: %d (%s)\n", handle, bo_gem->name);
803 drm_intel_gem_bo_free(drm_intel_bo *bo)
805 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
806 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
807 struct drm_gem_close close;
810 if (bo_gem->mem_virtual)
811 munmap(bo_gem->mem_virtual, bo_gem->bo.size);
812 if (bo_gem->gtt_virtual)
813 munmap(bo_gem->gtt_virtual, bo_gem->bo.size);
815 /* Close this object */
816 memset(&close, 0, sizeof(close));
817 close.handle = bo_gem->gem_handle;
818 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_GEM_CLOSE, &close);
821 "DRM_IOCTL_GEM_CLOSE %d failed (%s): %s\n",
822 bo_gem->gem_handle, bo_gem->name, strerror(errno));
827 /** Frees all cached buffers significantly older than @time. */
829 drm_intel_gem_cleanup_bo_cache(drm_intel_bufmgr_gem *bufmgr_gem, time_t time)
833 for (i = 0; i < DRM_INTEL_GEM_BO_BUCKETS; i++) {
834 struct drm_intel_gem_bo_bucket *bucket =
835 &bufmgr_gem->cache_bucket[i];
837 while (!DRMLISTEMPTY(&bucket->head)) {
838 drm_intel_bo_gem *bo_gem;
840 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
841 bucket->head.next, head);
842 if (time - bo_gem->free_time <= 1)
845 DRMLISTDEL(&bo_gem->head);
847 drm_intel_gem_bo_free(&bo_gem->bo);
853 drm_intel_gem_bo_unreference_final(drm_intel_bo *bo, time_t time)
855 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
856 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
857 struct drm_intel_gem_bo_bucket *bucket;
858 uint32_t tiling_mode;
861 /* Unreference all the target buffers */
862 for (i = 0; i < bo_gem->reloc_count; i++) {
863 drm_intel_gem_bo_unreference_locked_timed(bo_gem->
864 reloc_target_info[i].bo,
867 bo_gem->reloc_count = 0;
868 bo_gem->used_as_reloc_target = 0;
870 DBG("bo_unreference final: %d (%s)\n",
871 bo_gem->gem_handle, bo_gem->name);
873 /* release memory associated with this object */
874 if (bo_gem->reloc_target_info) {
875 free(bo_gem->reloc_target_info);
876 bo_gem->reloc_target_info = NULL;
878 if (bo_gem->relocs) {
879 free(bo_gem->relocs);
880 bo_gem->relocs = NULL;
883 bucket = drm_intel_gem_bo_bucket_for_size(bufmgr_gem, bo->size);
884 /* Put the buffer into our internal cache for reuse if we can. */
885 tiling_mode = I915_TILING_NONE;
886 if (bufmgr_gem->bo_reuse && bo_gem->reusable && bucket != NULL &&
887 drm_intel_gem_bo_set_tiling(bo, &tiling_mode, 0) == 0 &&
888 drm_intel_gem_bo_madvise_internal(bufmgr_gem, bo_gem,
889 I915_MADV_DONTNEED)) {
890 bo_gem->free_time = time;
893 bo_gem->validate_index = -1;
895 DRMLISTADDTAIL(&bo_gem->head, &bucket->head);
897 drm_intel_gem_cleanup_bo_cache(bufmgr_gem, time);
899 drm_intel_gem_bo_free(bo);
903 static void drm_intel_gem_bo_unreference_locked_timed(drm_intel_bo *bo,
906 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
908 assert(atomic_read(&bo_gem->refcount) > 0);
909 if (atomic_dec_and_test(&bo_gem->refcount))
910 drm_intel_gem_bo_unreference_final(bo, time);
913 static void drm_intel_gem_bo_unreference(drm_intel_bo *bo)
915 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
917 assert(atomic_read(&bo_gem->refcount) > 0);
918 if (atomic_dec_and_test(&bo_gem->refcount)) {
919 drm_intel_bufmgr_gem *bufmgr_gem =
920 (drm_intel_bufmgr_gem *) bo->bufmgr;
921 struct timespec time;
923 clock_gettime(CLOCK_MONOTONIC, &time);
925 pthread_mutex_lock(&bufmgr_gem->lock);
926 drm_intel_gem_bo_unreference_final(bo, time.tv_sec);
927 pthread_mutex_unlock(&bufmgr_gem->lock);
931 static int drm_intel_gem_bo_map(drm_intel_bo *bo, int write_enable)
933 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
934 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
935 struct drm_i915_gem_set_domain set_domain;
938 pthread_mutex_lock(&bufmgr_gem->lock);
940 /* Allow recursive mapping. Mesa may recursively map buffers with
941 * nested display loops.
943 if (!bo_gem->mem_virtual) {
944 struct drm_i915_gem_mmap mmap_arg;
946 DBG("bo_map: %d (%s)\n", bo_gem->gem_handle, bo_gem->name);
948 memset(&mmap_arg, 0, sizeof(mmap_arg));
949 mmap_arg.handle = bo_gem->gem_handle;
951 mmap_arg.size = bo->size;
953 ret = ioctl(bufmgr_gem->fd,
954 DRM_IOCTL_I915_GEM_MMAP,
956 } while (ret == -1 && errno == EINTR);
960 "%s:%d: Error mapping buffer %d (%s): %s .\n",
961 __FILE__, __LINE__, bo_gem->gem_handle,
962 bo_gem->name, strerror(errno));
963 pthread_mutex_unlock(&bufmgr_gem->lock);
966 bo_gem->mem_virtual = (void *)(uintptr_t) mmap_arg.addr_ptr;
968 DBG("bo_map: %d (%s) -> %p\n", bo_gem->gem_handle, bo_gem->name,
969 bo_gem->mem_virtual);
970 bo->virtual = bo_gem->mem_virtual;
972 set_domain.handle = bo_gem->gem_handle;
973 set_domain.read_domains = I915_GEM_DOMAIN_CPU;
975 set_domain.write_domain = I915_GEM_DOMAIN_CPU;
977 set_domain.write_domain = 0;
979 ret = ioctl(bufmgr_gem->fd,
980 DRM_IOCTL_I915_GEM_SET_DOMAIN,
982 } while (ret == -1 && errno == EINTR);
985 fprintf(stderr, "%s:%d: Error setting to CPU domain %d: %s\n",
986 __FILE__, __LINE__, bo_gem->gem_handle,
988 pthread_mutex_unlock(&bufmgr_gem->lock);
992 pthread_mutex_unlock(&bufmgr_gem->lock);
997 int drm_intel_gem_bo_map_gtt(drm_intel_bo *bo)
999 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1000 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1001 struct drm_i915_gem_set_domain set_domain;
1004 pthread_mutex_lock(&bufmgr_gem->lock);
1006 /* Get a mapping of the buffer if we haven't before. */
1007 if (bo_gem->gtt_virtual == NULL) {
1008 struct drm_i915_gem_mmap_gtt mmap_arg;
1010 DBG("bo_map_gtt: mmap %d (%s)\n", bo_gem->gem_handle,
1013 memset(&mmap_arg, 0, sizeof(mmap_arg));
1014 mmap_arg.handle = bo_gem->gem_handle;
1016 /* Get the fake offset back... */
1018 ret = ioctl(bufmgr_gem->fd,
1019 DRM_IOCTL_I915_GEM_MMAP_GTT,
1021 } while (ret == -1 && errno == EINTR);
1025 "%s:%d: Error preparing buffer map %d (%s): %s .\n",
1027 bo_gem->gem_handle, bo_gem->name,
1029 pthread_mutex_unlock(&bufmgr_gem->lock);
1034 bo_gem->gtt_virtual = mmap(0, bo->size, PROT_READ | PROT_WRITE,
1035 MAP_SHARED, bufmgr_gem->fd,
1037 if (bo_gem->gtt_virtual == MAP_FAILED) {
1038 bo_gem->gtt_virtual = NULL;
1041 "%s:%d: Error mapping buffer %d (%s): %s .\n",
1043 bo_gem->gem_handle, bo_gem->name,
1045 pthread_mutex_unlock(&bufmgr_gem->lock);
1050 bo->virtual = bo_gem->gtt_virtual;
1052 DBG("bo_map_gtt: %d (%s) -> %p\n", bo_gem->gem_handle, bo_gem->name,
1053 bo_gem->gtt_virtual);
1055 /* Now move it to the GTT domain so that the CPU caches are flushed */
1056 set_domain.handle = bo_gem->gem_handle;
1057 set_domain.read_domains = I915_GEM_DOMAIN_GTT;
1058 set_domain.write_domain = I915_GEM_DOMAIN_GTT;
1060 ret = ioctl(bufmgr_gem->fd,
1061 DRM_IOCTL_I915_GEM_SET_DOMAIN,
1063 } while (ret == -1 && errno == EINTR);
1067 fprintf(stderr, "%s:%d: Error setting domain %d: %s\n",
1068 __FILE__, __LINE__, bo_gem->gem_handle,
1072 pthread_mutex_unlock(&bufmgr_gem->lock);
1077 int drm_intel_gem_bo_unmap_gtt(drm_intel_bo *bo)
1079 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1080 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1086 assert(bo_gem->gtt_virtual != NULL);
1088 pthread_mutex_lock(&bufmgr_gem->lock);
1090 pthread_mutex_unlock(&bufmgr_gem->lock);
1095 static int drm_intel_gem_bo_unmap(drm_intel_bo *bo)
1097 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1098 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1099 struct drm_i915_gem_sw_finish sw_finish;
1105 assert(bo_gem->mem_virtual != NULL);
1107 pthread_mutex_lock(&bufmgr_gem->lock);
1109 /* Cause a flush to happen if the buffer's pinned for scanout, so the
1110 * results show up in a timely manner.
1112 sw_finish.handle = bo_gem->gem_handle;
1114 ret = ioctl(bufmgr_gem->fd,
1115 DRM_IOCTL_I915_GEM_SW_FINISH,
1117 } while (ret == -1 && errno == EINTR);
1118 ret = ret == -1 ? -errno : 0;
1121 pthread_mutex_unlock(&bufmgr_gem->lock);
1127 drm_intel_gem_bo_subdata(drm_intel_bo *bo, unsigned long offset,
1128 unsigned long size, const void *data)
1130 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1131 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1132 struct drm_i915_gem_pwrite pwrite;
1135 memset(&pwrite, 0, sizeof(pwrite));
1136 pwrite.handle = bo_gem->gem_handle;
1137 pwrite.offset = offset;
1139 pwrite.data_ptr = (uint64_t) (uintptr_t) data;
1141 ret = ioctl(bufmgr_gem->fd,
1142 DRM_IOCTL_I915_GEM_PWRITE,
1144 } while (ret == -1 && errno == EINTR);
1148 "%s:%d: Error writing data to buffer %d: (%d %d) %s .\n",
1149 __FILE__, __LINE__, bo_gem->gem_handle, (int)offset,
1150 (int)size, strerror(errno));
1157 drm_intel_gem_get_pipe_from_crtc_id(drm_intel_bufmgr *bufmgr, int crtc_id)
1159 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
1160 struct drm_i915_get_pipe_from_crtc_id get_pipe_from_crtc_id;
1163 get_pipe_from_crtc_id.crtc_id = crtc_id;
1164 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GET_PIPE_FROM_CRTC_ID,
1165 &get_pipe_from_crtc_id);
1167 /* We return -1 here to signal that we don't
1168 * know which pipe is associated with this crtc.
1169 * This lets the caller know that this information
1170 * isn't available; using the wrong pipe for
1171 * vblank waiting can cause the chipset to lock up
1176 return get_pipe_from_crtc_id.pipe;
1180 drm_intel_gem_bo_get_subdata(drm_intel_bo *bo, unsigned long offset,
1181 unsigned long size, void *data)
1183 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1184 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1185 struct drm_i915_gem_pread pread;
1188 memset(&pread, 0, sizeof(pread));
1189 pread.handle = bo_gem->gem_handle;
1190 pread.offset = offset;
1192 pread.data_ptr = (uint64_t) (uintptr_t) data;
1194 ret = ioctl(bufmgr_gem->fd,
1195 DRM_IOCTL_I915_GEM_PREAD,
1197 } while (ret == -1 && errno == EINTR);
1201 "%s:%d: Error reading data from buffer %d: (%d %d) %s .\n",
1202 __FILE__, __LINE__, bo_gem->gem_handle, (int)offset,
1203 (int)size, strerror(errno));
1209 /** Waits for all GPU rendering to the object to have completed. */
1211 drm_intel_gem_bo_wait_rendering(drm_intel_bo *bo)
1213 drm_intel_gem_bo_start_gtt_access(bo, 0);
1217 * Sets the object to the GTT read and possibly write domain, used by the X
1218 * 2D driver in the absence of kernel support to do drm_intel_gem_bo_map_gtt().
1220 * In combination with drm_intel_gem_bo_pin() and manual fence management, we
1221 * can do tiled pixmaps this way.
1224 drm_intel_gem_bo_start_gtt_access(drm_intel_bo *bo, int write_enable)
1226 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1227 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1228 struct drm_i915_gem_set_domain set_domain;
1231 set_domain.handle = bo_gem->gem_handle;
1232 set_domain.read_domains = I915_GEM_DOMAIN_GTT;
1233 set_domain.write_domain = write_enable ? I915_GEM_DOMAIN_GTT : 0;
1235 ret = ioctl(bufmgr_gem->fd,
1236 DRM_IOCTL_I915_GEM_SET_DOMAIN,
1238 } while (ret == -1 && errno == EINTR);
1241 "%s:%d: Error setting memory domains %d (%08x %08x): %s .\n",
1242 __FILE__, __LINE__, bo_gem->gem_handle,
1243 set_domain.read_domains, set_domain.write_domain,
1249 drm_intel_bufmgr_gem_destroy(drm_intel_bufmgr *bufmgr)
1251 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
1254 free(bufmgr_gem->exec2_objects);
1255 free(bufmgr_gem->exec_objects);
1256 free(bufmgr_gem->exec_bos);
1258 pthread_mutex_destroy(&bufmgr_gem->lock);
1260 /* Free any cached buffer objects we were going to reuse */
1261 for (i = 0; i < DRM_INTEL_GEM_BO_BUCKETS; i++) {
1262 struct drm_intel_gem_bo_bucket *bucket =
1263 &bufmgr_gem->cache_bucket[i];
1264 drm_intel_bo_gem *bo_gem;
1266 while (!DRMLISTEMPTY(&bucket->head)) {
1267 bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
1268 bucket->head.next, head);
1269 DRMLISTDEL(&bo_gem->head);
1271 drm_intel_gem_bo_free(&bo_gem->bo);
1279 * Adds the target buffer to the validation list and adds the relocation
1280 * to the reloc_buffer's relocation list.
1282 * The relocation entry at the given offset must already contain the
1283 * precomputed relocation value, because the kernel will optimize out
1284 * the relocation entry write when the buffer hasn't moved from the
1285 * last known offset in target_bo.
1288 do_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset,
1289 drm_intel_bo *target_bo, uint32_t target_offset,
1290 uint32_t read_domains, uint32_t write_domain,
1293 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1294 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1295 drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *) target_bo;
1297 if (bo_gem->has_error)
1300 if (target_bo_gem->has_error) {
1301 bo_gem->has_error = 1;
1305 if (target_bo_gem->tiling_mode == I915_TILING_NONE)
1308 /* We never use HW fences for rendering on 965+ */
1309 if (bufmgr_gem->gen >= 4)
1312 /* Create a new relocation list if needed */
1313 if (bo_gem->relocs == NULL && drm_intel_setup_reloc_list(bo))
1316 /* Check overflow */
1317 assert(bo_gem->reloc_count < bufmgr_gem->max_relocs);
1320 assert(offset <= bo->size - 4);
1321 assert((write_domain & (write_domain - 1)) == 0);
1323 /* Make sure that we're not adding a reloc to something whose size has
1324 * already been accounted for.
1326 assert(!bo_gem->used_as_reloc_target);
1327 if (target_bo_gem != bo_gem) {
1328 target_bo_gem->used_as_reloc_target = 1;
1329 bo_gem->reloc_tree_size += target_bo_gem->reloc_tree_size;
1331 /* An object needing a fence is a tiled buffer, so it won't have
1332 * relocs to other buffers.
1335 target_bo_gem->reloc_tree_fences = 1;
1336 bo_gem->reloc_tree_fences += target_bo_gem->reloc_tree_fences;
1338 /* Flag the target to disallow further relocations in it. */
1340 bo_gem->relocs[bo_gem->reloc_count].offset = offset;
1341 bo_gem->relocs[bo_gem->reloc_count].delta = target_offset;
1342 bo_gem->relocs[bo_gem->reloc_count].target_handle =
1343 target_bo_gem->gem_handle;
1344 bo_gem->relocs[bo_gem->reloc_count].read_domains = read_domains;
1345 bo_gem->relocs[bo_gem->reloc_count].write_domain = write_domain;
1346 bo_gem->relocs[bo_gem->reloc_count].presumed_offset = target_bo->offset;
1348 bo_gem->reloc_target_info[bo_gem->reloc_count].bo = target_bo;
1349 drm_intel_gem_bo_reference(target_bo);
1351 bo_gem->reloc_target_info[bo_gem->reloc_count].flags =
1352 DRM_INTEL_RELOC_FENCE;
1354 bo_gem->reloc_target_info[bo_gem->reloc_count].flags = 0;
1356 bo_gem->reloc_count++;
1362 drm_intel_gem_bo_emit_reloc(drm_intel_bo *bo, uint32_t offset,
1363 drm_intel_bo *target_bo, uint32_t target_offset,
1364 uint32_t read_domains, uint32_t write_domain)
1366 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr;
1368 return do_bo_emit_reloc(bo, offset, target_bo, target_offset,
1369 read_domains, write_domain,
1370 !bufmgr_gem->fenced_relocs);
1374 drm_intel_gem_bo_emit_reloc_fence(drm_intel_bo *bo, uint32_t offset,
1375 drm_intel_bo *target_bo,
1376 uint32_t target_offset,
1377 uint32_t read_domains, uint32_t write_domain)
1379 return do_bo_emit_reloc(bo, offset, target_bo, target_offset,
1380 read_domains, write_domain, 1);
1384 * Walk the tree of relocations rooted at BO and accumulate the list of
1385 * validations to be performed and update the relocation buffers with
1386 * index values into the validation list.
1389 drm_intel_gem_bo_process_reloc(drm_intel_bo *bo)
1391 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1394 if (bo_gem->relocs == NULL)
1397 for (i = 0; i < bo_gem->reloc_count; i++) {
1398 drm_intel_bo *target_bo = bo_gem->reloc_target_info[i].bo;
1400 if (target_bo == bo)
1403 /* Continue walking the tree depth-first. */
1404 drm_intel_gem_bo_process_reloc(target_bo);
1406 /* Add the target to the validate list */
1407 drm_intel_add_validate_buffer(target_bo);
1412 drm_intel_gem_bo_process_reloc2(drm_intel_bo *bo)
1414 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
1417 if (bo_gem->relocs == NULL)
1420 for (i = 0; i < bo_gem->reloc_count; i++) {
1421 drm_intel_bo *target_bo = bo_gem->reloc_target_info[i].bo;
1424 if (target_bo == bo)
1427 /* Continue walking the tree depth-first. */
1428 drm_intel_gem_bo_process_reloc2(target_bo);
1430 need_fence = (bo_gem->reloc_target_info[i].flags &
1431 DRM_INTEL_RELOC_FENCE);
1433 /* Add the target to the validate list */
1434 drm_intel_add_validate_buffer2(target_bo, need_fence);
1440 drm_intel_update_buffer_offsets(drm_intel_bufmgr_gem *bufmgr_gem)
1444 for (i = 0; i < bufmgr_gem->exec_count; i++) {
1445 drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
1446 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1448 /* Update the buffer offset */
1449 if (bufmgr_gem->exec_objects[i].offset != bo->offset) {
1450 DBG("BO %d (%s) migrated: 0x%08lx -> 0x%08llx\n",
1451 bo_gem->gem_handle, bo_gem->name, bo->offset,
1452 (unsigned long long)bufmgr_gem->exec_objects[i].
1454 bo->offset = bufmgr_gem->exec_objects[i].offset;
1460 drm_intel_update_buffer_offsets2 (drm_intel_bufmgr_gem *bufmgr_gem)
1464 for (i = 0; i < bufmgr_gem->exec_count; i++) {
1465 drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
1466 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
1468 /* Update the buffer offset */
1469 if (bufmgr_gem->exec2_objects[i].offset != bo->offset) {
1470 DBG("BO %d (%s) migrated: 0x%08lx -> 0x%08llx\n",
1471 bo_gem->gem_handle, bo_gem->name, bo->offset,
1472 (unsigned long long)bufmgr_gem->exec2_objects[i].offset);
1473 bo->offset = bufmgr_gem->exec2_objects[i].offset;
1479 drm_intel_gem_bo_exec(drm_intel_bo *bo, int used,
1480 drm_clip_rect_t * cliprects, int num_cliprects, int DR4)
1482 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1483 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1484 struct drm_i915_gem_execbuffer execbuf;
1487 if (bo_gem->has_error)
1490 pthread_mutex_lock(&bufmgr_gem->lock);
1491 /* Update indices and set up the validate list. */
1492 drm_intel_gem_bo_process_reloc(bo);
1494 /* Add the batch buffer to the validation list. There are no
1495 * relocations pointing to it.
1497 drm_intel_add_validate_buffer(bo);
1499 execbuf.buffers_ptr = (uintptr_t) bufmgr_gem->exec_objects;
1500 execbuf.buffer_count = bufmgr_gem->exec_count;
1501 execbuf.batch_start_offset = 0;
1502 execbuf.batch_len = used;
1503 execbuf.cliprects_ptr = (uintptr_t) cliprects;
1504 execbuf.num_cliprects = num_cliprects;
1509 ret = ioctl(bufmgr_gem->fd,
1510 DRM_IOCTL_I915_GEM_EXECBUFFER,
1512 } while (ret != 0 && errno == EINTR);
1516 if (errno == ENOSPC) {
1518 "Execbuffer fails to pin. "
1519 "Estimate: %u. Actual: %u. Available: %u\n",
1520 drm_intel_gem_estimate_batch_space(bufmgr_gem->exec_bos,
1523 drm_intel_gem_compute_batch_space(bufmgr_gem->exec_bos,
1526 (unsigned int)bufmgr_gem->gtt_size);
1529 drm_intel_update_buffer_offsets(bufmgr_gem);
1531 if (bufmgr_gem->bufmgr.debug)
1532 drm_intel_gem_dump_validation_list(bufmgr_gem);
1534 for (i = 0; i < bufmgr_gem->exec_count; i++) {
1535 drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
1536 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1538 /* Disconnect the buffer from the validate list */
1539 bo_gem->validate_index = -1;
1540 bufmgr_gem->exec_bos[i] = NULL;
1542 bufmgr_gem->exec_count = 0;
1543 pthread_mutex_unlock(&bufmgr_gem->lock);
1549 drm_intel_gem_bo_mrb_exec2(drm_intel_bo *bo, int used,
1550 drm_clip_rect_t *cliprects, int num_cliprects, int DR4,
1553 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr;
1554 struct drm_i915_gem_execbuffer2 execbuf;
1557 if ((ring_flag != I915_EXEC_RENDER) && (ring_flag != I915_EXEC_BSD))
1560 pthread_mutex_lock(&bufmgr_gem->lock);
1561 /* Update indices and set up the validate list. */
1562 drm_intel_gem_bo_process_reloc2(bo);
1564 /* Add the batch buffer to the validation list. There are no relocations
1567 drm_intel_add_validate_buffer2(bo, 0);
1569 execbuf.buffers_ptr = (uintptr_t)bufmgr_gem->exec2_objects;
1570 execbuf.buffer_count = bufmgr_gem->exec_count;
1571 execbuf.batch_start_offset = 0;
1572 execbuf.batch_len = used;
1573 execbuf.cliprects_ptr = (uintptr_t)cliprects;
1574 execbuf.num_cliprects = num_cliprects;
1577 execbuf.flags = ring_flag;
1582 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_EXECBUFFER2,
1584 } while (ret != 0 && errno == EINTR);
1588 if (ret == -ENOMEM) {
1590 "Execbuffer fails to pin. "
1591 "Estimate: %u. Actual: %u. Available: %u\n",
1592 drm_intel_gem_estimate_batch_space(bufmgr_gem->exec_bos,
1593 bufmgr_gem->exec_count),
1594 drm_intel_gem_compute_batch_space(bufmgr_gem->exec_bos,
1595 bufmgr_gem->exec_count),
1596 (unsigned int) bufmgr_gem->gtt_size);
1599 drm_intel_update_buffer_offsets2(bufmgr_gem);
1601 if (bufmgr_gem->bufmgr.debug)
1602 drm_intel_gem_dump_validation_list(bufmgr_gem);
1604 for (i = 0; i < bufmgr_gem->exec_count; i++) {
1605 drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
1606 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
1608 /* Disconnect the buffer from the validate list */
1609 bo_gem->validate_index = -1;
1610 bufmgr_gem->exec_bos[i] = NULL;
1612 bufmgr_gem->exec_count = 0;
1613 pthread_mutex_unlock(&bufmgr_gem->lock);
1619 drm_intel_gem_bo_exec2(drm_intel_bo *bo, int used,
1620 drm_clip_rect_t *cliprects, int num_cliprects,
1623 return drm_intel_gem_bo_mrb_exec2(bo, used,
1624 cliprects, num_cliprects, DR4,
1629 drm_intel_gem_bo_pin(drm_intel_bo *bo, uint32_t alignment)
1631 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1632 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1633 struct drm_i915_gem_pin pin;
1636 memset(&pin, 0, sizeof(pin));
1637 pin.handle = bo_gem->gem_handle;
1638 pin.alignment = alignment;
1641 ret = ioctl(bufmgr_gem->fd,
1642 DRM_IOCTL_I915_GEM_PIN,
1644 } while (ret == -1 && errno == EINTR);
1649 bo->offset = pin.offset;
1654 drm_intel_gem_bo_unpin(drm_intel_bo *bo)
1656 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1657 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1658 struct drm_i915_gem_unpin unpin;
1661 memset(&unpin, 0, sizeof(unpin));
1662 unpin.handle = bo_gem->gem_handle;
1664 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_UNPIN, &unpin);
1672 drm_intel_gem_bo_set_tiling(drm_intel_bo *bo, uint32_t * tiling_mode,
1675 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1676 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1677 struct drm_i915_gem_set_tiling set_tiling;
1680 if (bo_gem->global_name == 0 && *tiling_mode == bo_gem->tiling_mode)
1683 memset(&set_tiling, 0, sizeof(set_tiling));
1684 set_tiling.handle = bo_gem->gem_handle;
1687 set_tiling.tiling_mode = *tiling_mode;
1688 set_tiling.stride = stride;
1690 ret = ioctl(bufmgr_gem->fd,
1691 DRM_IOCTL_I915_GEM_SET_TILING,
1693 } while (ret == -1 && errno == EINTR);
1695 bo_gem->tiling_mode = set_tiling.tiling_mode;
1696 bo_gem->swizzle_mode = set_tiling.swizzle_mode;
1697 drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
1701 *tiling_mode = bo_gem->tiling_mode;
1706 drm_intel_gem_bo_get_tiling(drm_intel_bo *bo, uint32_t * tiling_mode,
1707 uint32_t * swizzle_mode)
1709 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1711 *tiling_mode = bo_gem->tiling_mode;
1712 *swizzle_mode = bo_gem->swizzle_mode;
1717 drm_intel_gem_bo_flink(drm_intel_bo *bo, uint32_t * name)
1719 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
1720 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1721 struct drm_gem_flink flink;
1724 if (!bo_gem->global_name) {
1725 memset(&flink, 0, sizeof(flink));
1726 flink.handle = bo_gem->gem_handle;
1728 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_GEM_FLINK, &flink);
1731 bo_gem->global_name = flink.name;
1732 bo_gem->reusable = 0;
1735 *name = bo_gem->global_name;
1740 * Enables unlimited caching of buffer objects for reuse.
1742 * This is potentially very memory expensive, as the cache at each bucket
1743 * size is only bounded by how many buffers of that size we've managed to have
1744 * in flight at once.
1747 drm_intel_bufmgr_gem_enable_reuse(drm_intel_bufmgr *bufmgr)
1749 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
1751 bufmgr_gem->bo_reuse = 1;
1755 * Enable use of fenced reloc type.
1757 * New code should enable this to avoid unnecessary fence register
1758 * allocation. If this option is not enabled, all relocs will have fence
1759 * register allocated.
1762 drm_intel_bufmgr_gem_enable_fenced_relocs(drm_intel_bufmgr *bufmgr)
1764 drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
1766 if (bufmgr_gem->bufmgr.bo_exec == drm_intel_gem_bo_exec2)
1767 bufmgr_gem->fenced_relocs = 1;
1771 * Return the additional aperture space required by the tree of buffer objects
1775 drm_intel_gem_bo_get_aperture_space(drm_intel_bo *bo)
1777 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1781 if (bo == NULL || bo_gem->included_in_check_aperture)
1785 bo_gem->included_in_check_aperture = 1;
1787 for (i = 0; i < bo_gem->reloc_count; i++)
1789 drm_intel_gem_bo_get_aperture_space(bo_gem->
1790 reloc_target_info[i].bo);
1796 * Count the number of buffers in this list that need a fence reg
1798 * If the count is greater than the number of available regs, we'll have
1799 * to ask the caller to resubmit a batch with fewer tiled buffers.
1801 * This function over-counts if the same buffer is used multiple times.
1804 drm_intel_gem_total_fences(drm_intel_bo ** bo_array, int count)
1807 unsigned int total = 0;
1809 for (i = 0; i < count; i++) {
1810 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo_array[i];
1815 total += bo_gem->reloc_tree_fences;
1821 * Clear the flag set by drm_intel_gem_bo_get_aperture_space() so we're ready
1822 * for the next drm_intel_bufmgr_check_aperture_space() call.
1825 drm_intel_gem_bo_clear_aperture_space_flag(drm_intel_bo *bo)
1827 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1830 if (bo == NULL || !bo_gem->included_in_check_aperture)
1833 bo_gem->included_in_check_aperture = 0;
1835 for (i = 0; i < bo_gem->reloc_count; i++)
1836 drm_intel_gem_bo_clear_aperture_space_flag(bo_gem->
1837 reloc_target_info[i].bo);
1841 * Return a conservative estimate for the amount of aperture required
1842 * for a collection of buffers. This may double-count some buffers.
1845 drm_intel_gem_estimate_batch_space(drm_intel_bo **bo_array, int count)
1848 unsigned int total = 0;
1850 for (i = 0; i < count; i++) {
1851 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo_array[i];
1853 total += bo_gem->reloc_tree_size;
1859 * Return the amount of aperture needed for a collection of buffers.
1860 * This avoids double counting any buffers, at the cost of looking
1861 * at every buffer in the set.
1864 drm_intel_gem_compute_batch_space(drm_intel_bo **bo_array, int count)
1867 unsigned int total = 0;
1869 for (i = 0; i < count; i++) {
1870 total += drm_intel_gem_bo_get_aperture_space(bo_array[i]);
1871 /* For the first buffer object in the array, we get an
1872 * accurate count back for its reloc_tree size (since nothing
1873 * had been flagged as being counted yet). We can save that
1874 * value out as a more conservative reloc_tree_size that
1875 * avoids double-counting target buffers. Since the first
1876 * buffer happens to usually be the batch buffer in our
1877 * callers, this can pull us back from doing the tree
1878 * walk on every new batch emit.
1881 drm_intel_bo_gem *bo_gem =
1882 (drm_intel_bo_gem *) bo_array[i];
1883 bo_gem->reloc_tree_size = total;
1887 for (i = 0; i < count; i++)
1888 drm_intel_gem_bo_clear_aperture_space_flag(bo_array[i]);
1893 * Return -1 if the batchbuffer should be flushed before attempting to
1894 * emit rendering referencing the buffers pointed to by bo_array.
1896 * This is required because if we try to emit a batchbuffer with relocations
1897 * to a tree of buffers that won't simultaneously fit in the aperture,
1898 * the rendering will return an error at a point where the software is not
1899 * prepared to recover from it.
1901 * However, we also want to emit the batchbuffer significantly before we reach
1902 * the limit, as a series of batchbuffers each of which references buffers
1903 * covering almost all of the aperture means that at each emit we end up
1904 * waiting to evict a buffer from the last rendering, and we get synchronous
1905 * performance. By emitting smaller batchbuffers, we eat some CPU overhead to
1906 * get better parallelism.
1909 drm_intel_gem_check_aperture_space(drm_intel_bo **bo_array, int count)
1911 drm_intel_bufmgr_gem *bufmgr_gem =
1912 (drm_intel_bufmgr_gem *) bo_array[0]->bufmgr;
1913 unsigned int total = 0;
1914 unsigned int threshold = bufmgr_gem->gtt_size * 3 / 4;
1917 /* Check for fence reg constraints if necessary */
1918 if (bufmgr_gem->available_fences) {
1919 total_fences = drm_intel_gem_total_fences(bo_array, count);
1920 if (total_fences > bufmgr_gem->available_fences)
1924 total = drm_intel_gem_estimate_batch_space(bo_array, count);
1926 if (total > threshold)
1927 total = drm_intel_gem_compute_batch_space(bo_array, count);
1929 if (total > threshold) {
1930 DBG("check_space: overflowed available aperture, "
1932 total / 1024, (int)bufmgr_gem->gtt_size / 1024);
1935 DBG("drm_check_space: total %dkb vs bufgr %dkb\n", total / 1024,
1936 (int)bufmgr_gem->gtt_size / 1024);
1942 * Disable buffer reuse for objects which are shared with the kernel
1943 * as scanout buffers
1946 drm_intel_gem_bo_disable_reuse(drm_intel_bo *bo)
1948 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1950 bo_gem->reusable = 0;
1955 drm_intel_gem_bo_is_reusable(drm_intel_bo *bo)
1957 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1959 return bo_gem->reusable;
1963 _drm_intel_gem_bo_references(drm_intel_bo *bo, drm_intel_bo *target_bo)
1965 drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
1968 for (i = 0; i < bo_gem->reloc_count; i++) {
1969 if (bo_gem->reloc_target_info[i].bo == target_bo)
1971 if (_drm_intel_gem_bo_references(bo_gem->reloc_target_info[i].bo,
1979 /** Return true if target_bo is referenced by bo's relocation tree. */
1981 drm_intel_gem_bo_references(drm_intel_bo *bo, drm_intel_bo *target_bo)
1983 drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *) target_bo;
1985 if (bo == NULL || target_bo == NULL)
1987 if (target_bo_gem->used_as_reloc_target)
1988 return _drm_intel_gem_bo_references(bo, target_bo);
1993 * Initializes the GEM buffer manager, which uses the kernel to allocate, map,
1994 * and manage map buffer objections.
1996 * \param fd File descriptor of the opened DRM device.
1999 drm_intel_bufmgr_gem_init(int fd, int batch_size)
2001 drm_intel_bufmgr_gem *bufmgr_gem;
2002 struct drm_i915_gem_get_aperture aperture;
2003 drm_i915_getparam_t gp;
2006 int exec2 = 0, has_bsd = 0;
2008 bufmgr_gem = calloc(1, sizeof(*bufmgr_gem));
2009 if (bufmgr_gem == NULL)
2012 bufmgr_gem->fd = fd;
2014 if (pthread_mutex_init(&bufmgr_gem->lock, NULL) != 0) {
2019 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_GET_APERTURE, &aperture);
2022 bufmgr_gem->gtt_size = aperture.aper_available_size;
2024 fprintf(stderr, "DRM_IOCTL_I915_GEM_APERTURE failed: %s\n",
2026 bufmgr_gem->gtt_size = 128 * 1024 * 1024;
2027 fprintf(stderr, "Assuming %dkB available aperture size.\n"
2028 "May lead to reduced performance or incorrect "
2030 (int)bufmgr_gem->gtt_size / 1024);
2033 gp.param = I915_PARAM_CHIPSET_ID;
2034 gp.value = &bufmgr_gem->pci_device;
2035 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
2037 fprintf(stderr, "get chip id failed: %d [%d]\n", ret, errno);
2038 fprintf(stderr, "param: %d, val: %d\n", gp.param, *gp.value);
2041 if (IS_GEN2(bufmgr_gem))
2042 bufmgr_gem->gen = 2;
2043 else if (IS_GEN3(bufmgr_gem))
2044 bufmgr_gem->gen = 3;
2045 else if (IS_GEN4(bufmgr_gem))
2046 bufmgr_gem->gen = 4;
2048 bufmgr_gem->gen = 6;
2050 gp.param = I915_PARAM_HAS_EXECBUF2;
2051 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
2055 gp.param = I915_PARAM_HAS_BSD;
2056 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
2060 if (bufmgr_gem->gen < 4) {
2061 gp.param = I915_PARAM_NUM_FENCES_AVAIL;
2062 gp.value = &bufmgr_gem->available_fences;
2063 ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
2065 fprintf(stderr, "get fences failed: %d [%d]\n", ret,
2067 fprintf(stderr, "param: %d, val: %d\n", gp.param,
2069 bufmgr_gem->available_fences = 0;
2071 /* XXX The kernel reports the total number of fences,
2072 * including any that may be pinned.
2074 * We presume that there will be at least one pinned
2075 * fence for the scanout buffer, but there may be more
2076 * than one scanout and the user may be manually
2077 * pinning buffers. Let's move to execbuffer2 and
2078 * thereby forget the insanity of using fences...
2080 bufmgr_gem->available_fences -= 2;
2081 if (bufmgr_gem->available_fences < 0)
2082 bufmgr_gem->available_fences = 0;
2086 /* Let's go with one relocation per every 2 dwords (but round down a bit
2087 * since a power of two will mean an extra page allocation for the reloc
2090 * Every 4 was too few for the blender benchmark.
2092 bufmgr_gem->max_relocs = batch_size / sizeof(uint32_t) / 2 - 2;
2094 bufmgr_gem->bufmgr.bo_alloc = drm_intel_gem_bo_alloc;
2095 bufmgr_gem->bufmgr.bo_alloc_for_render =
2096 drm_intel_gem_bo_alloc_for_render;
2097 bufmgr_gem->bufmgr.bo_alloc_tiled = drm_intel_gem_bo_alloc_tiled;
2098 bufmgr_gem->bufmgr.bo_reference = drm_intel_gem_bo_reference;
2099 bufmgr_gem->bufmgr.bo_unreference = drm_intel_gem_bo_unreference;
2100 bufmgr_gem->bufmgr.bo_map = drm_intel_gem_bo_map;
2101 bufmgr_gem->bufmgr.bo_unmap = drm_intel_gem_bo_unmap;
2102 bufmgr_gem->bufmgr.bo_subdata = drm_intel_gem_bo_subdata;
2103 bufmgr_gem->bufmgr.bo_get_subdata = drm_intel_gem_bo_get_subdata;
2104 bufmgr_gem->bufmgr.bo_wait_rendering = drm_intel_gem_bo_wait_rendering;
2105 bufmgr_gem->bufmgr.bo_emit_reloc = drm_intel_gem_bo_emit_reloc;
2106 bufmgr_gem->bufmgr.bo_emit_reloc_fence = drm_intel_gem_bo_emit_reloc_fence;
2107 bufmgr_gem->bufmgr.bo_pin = drm_intel_gem_bo_pin;
2108 bufmgr_gem->bufmgr.bo_unpin = drm_intel_gem_bo_unpin;
2109 bufmgr_gem->bufmgr.bo_get_tiling = drm_intel_gem_bo_get_tiling;
2110 bufmgr_gem->bufmgr.bo_set_tiling = drm_intel_gem_bo_set_tiling;
2111 bufmgr_gem->bufmgr.bo_flink = drm_intel_gem_bo_flink;
2112 /* Use the new one if available */
2114 bufmgr_gem->bufmgr.bo_exec = drm_intel_gem_bo_exec2;
2116 bufmgr_gem->bufmgr.bo_mrb_exec = drm_intel_gem_bo_mrb_exec2;
2118 bufmgr_gem->bufmgr.bo_exec = drm_intel_gem_bo_exec;
2119 bufmgr_gem->bufmgr.bo_busy = drm_intel_gem_bo_busy;
2120 bufmgr_gem->bufmgr.bo_madvise = drm_intel_gem_bo_madvise;
2121 bufmgr_gem->bufmgr.destroy = drm_intel_bufmgr_gem_destroy;
2122 bufmgr_gem->bufmgr.debug = 0;
2123 bufmgr_gem->bufmgr.check_aperture_space =
2124 drm_intel_gem_check_aperture_space;
2125 bufmgr_gem->bufmgr.bo_disable_reuse = drm_intel_gem_bo_disable_reuse;
2126 bufmgr_gem->bufmgr.bo_is_reusable = drm_intel_gem_bo_is_reusable;
2127 bufmgr_gem->bufmgr.get_pipe_from_crtc_id =
2128 drm_intel_gem_get_pipe_from_crtc_id;
2129 bufmgr_gem->bufmgr.bo_references = drm_intel_gem_bo_references;
2131 /* Initialize the linked lists for BO reuse cache. */
2132 for (i = 0, size = 4096; i < DRM_INTEL_GEM_BO_BUCKETS; i++, size *= 2) {
2133 DRMINITLISTHEAD(&bufmgr_gem->cache_bucket[i].head);
2134 bufmgr_gem->cache_bucket[i].size = size;
2137 return &bufmgr_gem->bufmgr;