1 /**************************************************************************
3 * Copyright 2006 Tungsten Graphics, Inc., Cedar Park, Texas.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
28 /* Originally a fake version of the buffer manager so that we can
29 * prototype the changes in a driver fairly quickly, has been fleshed
30 * out to a fully functional interim solution.
32 * Basically wraps the old style memory management in the new
33 * programming interface, but is more expressive and avoids many of
34 * the bugs in the old texture manager.
48 #include "intel_bufmgr.h"
49 #include "intel_bufmgr_priv.h"
53 #include "libdrm_macros.h"
54 #include "libdrm_lists.h"
56 #define DBG(...) do { \
57 if (bufmgr_fake->bufmgr.debug) \
58 drmMsg(__VA_ARGS__); \
63 #define BM_NO_BACKING_STORE 0x00000001
64 #define BM_NO_FENCE_SUBDATA 0x00000002
65 #define BM_PINNED 0x00000004
67 /* Wrapper around mm.c's mem_block, which understands that you must
68 * wait for fences to expire before memory can be freed. This is
69 * specific to our use of memcpy for uploads - an upload that was
70 * processed through the command queue wouldn't need to care about
73 #define MAX_RELOCS 4096
75 struct fake_buffer_reloc {
76 /** Buffer object that the relocation points at. */
77 drm_intel_bo *target_buf;
78 /** Offset of the relocation entry within reloc_buf. */
81 * Cached value of the offset when we last performed this relocation.
83 uint32_t last_target_offset;
84 /** Value added to target_buf's offset to get the relocation entry. */
86 /** Cache domains the target buffer is read into. */
87 uint32_t read_domains;
88 /** Cache domain the target buffer will have dirty cachelines in. */
89 uint32_t write_domain;
93 struct block *next, *prev;
94 struct mem_block *mem; /* BM_MEM_AGP */
97 * Marks that the block is currently in the aperture and has yet to be
100 unsigned on_hardware:1;
102 * Marks that the block is currently fenced (being used by rendering)
103 * and can't be freed until @fence is passed.
107 /** Fence cookie for the block. */
108 unsigned fence; /* Split to read_fence, write_fence */
114 typedef struct _bufmgr_fake {
115 drm_intel_bufmgr bufmgr;
117 pthread_mutex_t lock;
119 unsigned long low_offset;
123 struct mem_block *heap;
125 unsigned buf_nr; /* for generating ids */
128 * List of blocks which are currently in the GART but haven't been
131 struct block on_hardware;
133 * List of blocks which are in the GART and have an active fence on
138 * List of blocks which have an expired fence and are ready to be
143 unsigned int last_fence;
146 unsigned need_fence:1;
150 * Driver callback to emit a fence, returning the cookie.
152 * This allows the driver to hook in a replacement for the DRM usage in
155 * Currently, this also requires that a write flush be emitted before
156 * emitting the fence, but this should change.
158 unsigned int (*fence_emit) (void *private);
159 /** Driver callback to wait for a fence cookie to have passed. */
160 void (*fence_wait) (unsigned int fence, void *private);
164 * Driver callback to execute a buffer.
166 * This allows the driver to hook in a replacement for the DRM usage in
169 int (*exec) (drm_intel_bo *bo, unsigned int used, void *priv);
172 /** Driver-supplied argument to driver callbacks */
175 * Pointer to kernel-updated sarea data for the last completed user irq
177 volatile int *last_dispatch;
183 int performed_rendering;
184 } drm_intel_bufmgr_fake;
186 typedef struct _drm_intel_bo_fake {
189 unsigned id; /* debug only */
194 * has the card written to this buffer - we make need to copy it back
196 unsigned card_dirty:1;
197 unsigned int refcount;
198 /* Flags may consist of any of the DRM_BO flags, plus
199 * DRM_BO_NO_BACKING_STORE and BM_NO_FENCE_SUBDATA, which are the
200 * first two driver private flags.
203 /** Cache domains the target buffer is read into. */
204 uint32_t read_domains;
205 /** Cache domain the target buffer will have dirty cachelines in. */
206 uint32_t write_domain;
208 unsigned int alignment;
209 int is_static, validated;
210 unsigned int map_count;
212 /** relocation list */
213 struct fake_buffer_reloc *relocs;
216 * Total size of the target_bos of this buffer.
218 * Used for estimation in check_aperture.
220 unsigned int child_size;
224 void (*invalidate_cb) (drm_intel_bo *bo, void *ptr);
225 void *invalidate_ptr;
228 static int clear_fenced(drm_intel_bufmgr_fake *bufmgr_fake,
229 unsigned int fence_cookie);
231 #define MAXFENCE 0x7fffffff
234 FENCE_LTE(unsigned a, unsigned b)
239 if (a < b && b - a < (1 << 24))
242 if (a > b && MAXFENCE - a + b < (1 << 24))
249 drm_intel_bufmgr_fake_set_fence_callback(drm_intel_bufmgr *bufmgr,
250 unsigned int (*emit) (void *priv),
251 void (*wait) (unsigned int fence,
255 drm_intel_bufmgr_fake *bufmgr_fake = (drm_intel_bufmgr_fake *) bufmgr;
257 bufmgr_fake->fence_emit = emit;
258 bufmgr_fake->fence_wait = wait;
259 bufmgr_fake->fence_priv = priv;
263 _fence_emit_internal(drm_intel_bufmgr_fake *bufmgr_fake)
265 struct drm_i915_irq_emit ie;
268 if (bufmgr_fake->fence_emit != NULL) {
269 seq = bufmgr_fake->fence_emit(bufmgr_fake->fence_priv);
274 ret = drmCommandWriteRead(bufmgr_fake->fd, DRM_I915_IRQ_EMIT,
277 drmMsg("%s: drm_i915_irq_emit: %d\n", __func__, ret);
281 DBG("emit 0x%08x\n", seq);
286 _fence_wait_internal(drm_intel_bufmgr_fake *bufmgr_fake, int seq)
288 struct drm_i915_irq_wait iw;
289 int hw_seq, busy_count = 0;
293 if (bufmgr_fake->fence_wait != NULL) {
294 bufmgr_fake->fence_wait(seq, bufmgr_fake->fence_priv);
295 clear_fenced(bufmgr_fake, seq);
301 DBG("wait 0x%08x\n", iw.irq_seq);
303 /* The kernel IRQ_WAIT implementation is all sorts of broken.
304 * 1) It returns 1 to 0x7fffffff instead of using the full 32-bit
306 * 2) It returns 0 if hw_seq >= seq, not seq - hw_seq < 0 on the 32-bit
308 * 3) It waits if seq < hw_seq, not seq - hw_seq > 0 on the 32-bit
310 * 4) It returns -EBUSY in 3 seconds even if the hardware is still
311 * successfully chewing through buffers.
313 * Assume that in userland we treat sequence numbers as ints, which
314 * makes some of the comparisons convenient, since the sequence
315 * numbers are all positive signed integers.
317 * From this we get several cases we need to handle. Here's a timeline.
318 * 0x2 0x7 0x7ffffff8 0x7ffffffd
320 * ------------------------------------------------------------
322 * A) Normal wait for hw to catch up
325 * ------------------------------------------------------------
326 * seq - hw_seq = 5. If we call IRQ_WAIT, it will wait for hw to
329 * B) Normal wait for a sequence number that's already passed.
332 * ------------------------------------------------------------
333 * seq - hw_seq = -5. If we call IRQ_WAIT, it returns 0 quickly.
335 * C) Hardware has already wrapped around ahead of us
338 * ------------------------------------------------------------
339 * seq - hw_seq = 0x80000000 - 5. If we called IRQ_WAIT, it would wait
340 * for hw_seq >= seq, which may never occur. Thus, we want to catch
341 * this in userland and return 0.
343 * D) We've wrapped around ahead of the hardware.
346 * ------------------------------------------------------------
347 * seq - hw_seq = -(0x80000000 - 5). If we called IRQ_WAIT, it would
348 * return 0 quickly because hw_seq >= seq, even though the hardware
349 * isn't caught up. Thus, we need to catch this early return in
350 * userland and bother the kernel until the hardware really does
353 * E) Hardware might wrap after we test in userland.
356 * ------------------------------------------------------------
357 * seq - hw_seq = 5. If we call IRQ_WAIT, it will likely see seq >=
358 * hw_seq and wait. However, suppose hw_seq wraps before we make it
359 * into the kernel. The kernel sees hw_seq >= seq and waits for 3
360 * seconds then returns -EBUSY. This is case C). We should catch
361 * this and then return successfully.
363 * F) Hardware might take a long time on a buffer.
366 * -------------------------------------------------------------------
367 * seq - hw_seq = 5. If we call IRQ_WAIT, if sequence 2 through 5
368 * take too long, it will return -EBUSY. Batchbuffers in the
369 * gltestperf demo were seen to take up to 7 seconds. We should
370 * catch early -EBUSY return and keep trying.
374 /* Keep a copy of last_dispatch so that if the wait -EBUSYs
375 * because the hardware didn't catch up in 3 seconds, we can
376 * see if it at least made progress and retry.
378 hw_seq = *bufmgr_fake->last_dispatch;
381 if (seq - hw_seq > 0x40000000)
384 ret = drmCommandWrite(bufmgr_fake->fd, DRM_I915_IRQ_WAIT,
387 kernel_lied = (ret == 0) && (seq - *bufmgr_fake->last_dispatch <
392 && (seq - *bufmgr_fake->last_dispatch > 0x40000000))
395 /* Catch case F: Allow up to 15 seconds chewing on one buffer. */
396 if ((ret == -EBUSY) && (hw_seq != *bufmgr_fake->last_dispatch))
400 } while (kernel_lied || ret == -EAGAIN || ret == -EINTR ||
401 (ret == -EBUSY && busy_count < 5));
404 drmMsg("%s:%d: Error waiting for fence: %s.\n", __FILE__,
405 __LINE__, strerror(-ret));
408 clear_fenced(bufmgr_fake, seq);
412 _fence_test(drm_intel_bufmgr_fake *bufmgr_fake, unsigned fence)
414 /* Slight problem with wrap-around:
416 return fence == 0 || FENCE_LTE(fence, bufmgr_fake->last_fence);
420 * Allocate a memory manager block for the buffer.
423 alloc_block(drm_intel_bo *bo)
425 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
426 drm_intel_bufmgr_fake *bufmgr_fake =
427 (drm_intel_bufmgr_fake *) bo->bufmgr;
428 struct block *block = (struct block *)calloc(sizeof *block, 1);
429 unsigned int align_log2 = ffs(bo_fake->alignment) - 1;
435 sz = (bo->size + bo_fake->alignment - 1) & ~(bo_fake->alignment - 1);
437 block->mem = mmAllocMem(bufmgr_fake->heap, sz, align_log2, 0);
443 DRMINITLISTHEAD(block);
445 /* Insert at head or at tail??? */
446 DRMLISTADDTAIL(block, &bufmgr_fake->lru);
448 block->virtual = (uint8_t *) bufmgr_fake->virtual +
449 block->mem->ofs - bufmgr_fake->low_offset;
452 bo_fake->block = block;
457 /* Release the card storage associated with buf:
460 free_block(drm_intel_bufmgr_fake *bufmgr_fake, struct block *block,
463 drm_intel_bo_fake *bo_fake;
464 DBG("free block %p %08x %d %d\n", block, block->mem->ofs,
465 block->on_hardware, block->fenced);
470 bo_fake = (drm_intel_bo_fake *) block->bo;
472 if (bo_fake->flags & (BM_PINNED | BM_NO_BACKING_STORE))
475 if (!skip_dirty_copy && (bo_fake->card_dirty == 1)) {
476 memcpy(bo_fake->backing_store, block->virtual, block->bo->size);
477 bo_fake->card_dirty = 0;
481 if (block->on_hardware) {
483 } else if (block->fenced) {
486 DBG(" - free immediately\n");
489 mmFreeMem(block->mem);
495 alloc_backing_store(drm_intel_bo *bo)
497 drm_intel_bufmgr_fake *bufmgr_fake =
498 (drm_intel_bufmgr_fake *) bo->bufmgr;
499 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
500 assert(!bo_fake->backing_store);
501 assert(!(bo_fake->flags & (BM_PINNED | BM_NO_BACKING_STORE)));
503 bo_fake->backing_store = malloc(bo->size);
505 DBG("alloc_backing - buf %d %p %lu\n", bo_fake->id,
506 bo_fake->backing_store, bo->size);
507 assert(bo_fake->backing_store);
511 free_backing_store(drm_intel_bo *bo)
513 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
515 if (bo_fake->backing_store) {
516 assert(!(bo_fake->flags & (BM_PINNED | BM_NO_BACKING_STORE)));
517 free(bo_fake->backing_store);
518 bo_fake->backing_store = NULL;
523 set_dirty(drm_intel_bo *bo)
525 drm_intel_bufmgr_fake *bufmgr_fake =
526 (drm_intel_bufmgr_fake *) bo->bufmgr;
527 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
529 if (bo_fake->flags & BM_NO_BACKING_STORE
530 && bo_fake->invalidate_cb != NULL)
531 bo_fake->invalidate_cb(bo, bo_fake->invalidate_ptr);
533 assert(!(bo_fake->flags & BM_PINNED));
535 DBG("set_dirty - buf %d\n", bo_fake->id);
540 evict_lru(drm_intel_bufmgr_fake *bufmgr_fake, unsigned int max_fence)
542 struct block *block, *tmp;
544 DBG("%s\n", __func__);
546 DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->lru) {
547 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) block->bo;
549 if (bo_fake != NULL && (bo_fake->flags & BM_NO_FENCE_SUBDATA))
552 if (block->fence && max_fence && !FENCE_LTE(block->fence,
556 set_dirty(&bo_fake->bo);
557 bo_fake->block = NULL;
559 free_block(bufmgr_fake, block, 0);
567 evict_mru(drm_intel_bufmgr_fake *bufmgr_fake)
569 struct block *block, *tmp;
571 DBG("%s\n", __func__);
573 DRMLISTFOREACHSAFEREVERSE(block, tmp, &bufmgr_fake->lru) {
574 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) block->bo;
576 if (bo_fake && (bo_fake->flags & BM_NO_FENCE_SUBDATA))
579 set_dirty(&bo_fake->bo);
580 bo_fake->block = NULL;
582 free_block(bufmgr_fake, block, 0);
590 * Removes all objects from the fenced list older than the given fence.
593 clear_fenced(drm_intel_bufmgr_fake *bufmgr_fake, unsigned int fence_cookie)
595 struct block *block, *tmp;
598 bufmgr_fake->last_fence = fence_cookie;
599 DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->fenced) {
600 assert(block->fenced);
602 if (_fence_test(bufmgr_fake, block->fence)) {
607 DBG("delayed free: offset %x sz %x\n",
608 block->mem->ofs, block->mem->size);
610 mmFreeMem(block->mem);
613 DBG("return to lru: offset %x sz %x\n",
614 block->mem->ofs, block->mem->size);
616 DRMLISTADDTAIL(block, &bufmgr_fake->lru);
621 /* Blocks are ordered by fence, so if one fails, all
622 * from here will fail also:
624 DBG("fence not passed: offset %x sz %x %d %d \n",
625 block->mem->ofs, block->mem->size, block->fence,
626 bufmgr_fake->last_fence);
631 DBG("%s: %d\n", __func__, ret);
636 fence_blocks(drm_intel_bufmgr_fake *bufmgr_fake, unsigned fence)
638 struct block *block, *tmp;
640 DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->on_hardware) {
641 DBG("Fence block %p (sz 0x%x ofs %x buf %p) with fence %d\n",
642 block, block->mem->size, block->mem->ofs, block->bo, fence);
643 block->fence = fence;
645 block->on_hardware = 0;
648 /* Move to tail of pending list here
651 DRMLISTADDTAIL(block, &bufmgr_fake->fenced);
654 assert(DRMLISTEMPTY(&bufmgr_fake->on_hardware));
658 evict_and_alloc_block(drm_intel_bo *bo)
660 drm_intel_bufmgr_fake *bufmgr_fake =
661 (drm_intel_bufmgr_fake *) bo->bufmgr;
662 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
664 assert(bo_fake->block == NULL);
666 /* Search for already free memory:
671 /* If we're not thrashing, allow lru eviction to dig deeper into
672 * recently used textures. We'll probably be thrashing soon:
674 if (!bufmgr_fake->thrashing) {
675 while (evict_lru(bufmgr_fake, 0))
680 /* Keep thrashing counter alive?
682 if (bufmgr_fake->thrashing)
683 bufmgr_fake->thrashing = 20;
685 /* Wait on any already pending fences - here we are waiting for any
686 * freed memory that has been submitted to hardware and fenced to
689 while (!DRMLISTEMPTY(&bufmgr_fake->fenced)) {
690 uint32_t fence = bufmgr_fake->fenced.next->fence;
691 _fence_wait_internal(bufmgr_fake, fence);
697 if (!DRMLISTEMPTY(&bufmgr_fake->on_hardware)) {
698 while (!DRMLISTEMPTY(&bufmgr_fake->fenced)) {
699 uint32_t fence = bufmgr_fake->fenced.next->fence;
700 _fence_wait_internal(bufmgr_fake, fence);
703 if (!bufmgr_fake->thrashing) {
706 bufmgr_fake->thrashing = 20;
712 while (evict_mru(bufmgr_fake))
716 DBG("%s 0x%lx bytes failed\n", __func__, bo->size);
721 /***********************************************************************
726 * Wait for hardware idle by emitting a fence and waiting for it.
729 drm_intel_bufmgr_fake_wait_idle(drm_intel_bufmgr_fake *bufmgr_fake)
733 cookie = _fence_emit_internal(bufmgr_fake);
734 _fence_wait_internal(bufmgr_fake, cookie);
738 * Wait for rendering to a buffer to complete.
740 * It is assumed that the bathcbuffer which performed the rendering included
741 * the necessary flushing.
744 drm_intel_fake_bo_wait_rendering_locked(drm_intel_bo *bo)
746 drm_intel_bufmgr_fake *bufmgr_fake =
747 (drm_intel_bufmgr_fake *) bo->bufmgr;
748 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
750 if (bo_fake->block == NULL || !bo_fake->block->fenced)
753 _fence_wait_internal(bufmgr_fake, bo_fake->block->fence);
757 drm_intel_fake_bo_wait_rendering(drm_intel_bo *bo)
759 drm_intel_bufmgr_fake *bufmgr_fake =
760 (drm_intel_bufmgr_fake *) bo->bufmgr;
762 pthread_mutex_lock(&bufmgr_fake->lock);
763 drm_intel_fake_bo_wait_rendering_locked(bo);
764 pthread_mutex_unlock(&bufmgr_fake->lock);
767 /* Specifically ignore texture memory sharing.
768 * -- just evict everything
769 * -- and wait for idle
772 drm_intel_bufmgr_fake_contended_lock_take(drm_intel_bufmgr *bufmgr)
774 drm_intel_bufmgr_fake *bufmgr_fake = (drm_intel_bufmgr_fake *) bufmgr;
775 struct block *block, *tmp;
777 pthread_mutex_lock(&bufmgr_fake->lock);
779 bufmgr_fake->need_fence = 1;
780 bufmgr_fake->fail = 0;
782 /* Wait for hardware idle. We don't know where acceleration has been
783 * happening, so we'll need to wait anyway before letting anything get
784 * put on the card again.
786 drm_intel_bufmgr_fake_wait_idle(bufmgr_fake);
788 /* Check that we hadn't released the lock without having fenced the last
791 assert(DRMLISTEMPTY(&bufmgr_fake->fenced));
792 assert(DRMLISTEMPTY(&bufmgr_fake->on_hardware));
794 DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->lru) {
795 assert(_fence_test(bufmgr_fake, block->fence));
796 set_dirty(block->bo);
799 pthread_mutex_unlock(&bufmgr_fake->lock);
802 static drm_intel_bo *
803 drm_intel_fake_bo_alloc(drm_intel_bufmgr *bufmgr,
806 unsigned int alignment)
808 drm_intel_bufmgr_fake *bufmgr_fake;
809 drm_intel_bo_fake *bo_fake;
811 bufmgr_fake = (drm_intel_bufmgr_fake *) bufmgr;
815 bo_fake = calloc(1, sizeof(*bo_fake));
819 bo_fake->bo.size = size;
820 bo_fake->bo.offset = -1;
821 bo_fake->bo.virtual = NULL;
822 bo_fake->bo.bufmgr = bufmgr;
823 bo_fake->refcount = 1;
825 /* Alignment must be a power of two */
826 assert((alignment & (alignment - 1)) == 0);
829 bo_fake->alignment = alignment;
830 bo_fake->id = ++bufmgr_fake->buf_nr;
831 bo_fake->name = name;
833 bo_fake->is_static = 0;
835 DBG("drm_bo_alloc: (buf %d: %s, %lu kb)\n", bo_fake->id, bo_fake->name,
836 bo_fake->bo.size / 1024);
841 static drm_intel_bo *
842 drm_intel_fake_bo_alloc_tiled(drm_intel_bufmgr * bufmgr,
844 int x, int y, int cpp,
845 uint32_t *tiling_mode,
846 unsigned long *pitch,
849 unsigned long stride, aligned_y;
851 /* No runtime tiling support for fake. */
852 *tiling_mode = I915_TILING_NONE;
854 /* Align it for being a render target. Shouldn't need anything else. */
856 stride = ROUND_UP_TO(stride, 64);
858 /* 965 subspan loading alignment */
859 aligned_y = ALIGN(y, 2);
863 return drm_intel_fake_bo_alloc(bufmgr, name, stride * aligned_y,
868 drm_intel_bo_fake_alloc_static(drm_intel_bufmgr *bufmgr,
870 unsigned long offset,
871 unsigned long size, void *virtual)
873 drm_intel_bufmgr_fake *bufmgr_fake;
874 drm_intel_bo_fake *bo_fake;
876 bufmgr_fake = (drm_intel_bufmgr_fake *) bufmgr;
880 bo_fake = calloc(1, sizeof(*bo_fake));
884 bo_fake->bo.size = size;
885 bo_fake->bo.offset = offset;
886 bo_fake->bo.virtual = virtual;
887 bo_fake->bo.bufmgr = bufmgr;
888 bo_fake->refcount = 1;
889 bo_fake->id = ++bufmgr_fake->buf_nr;
890 bo_fake->name = name;
891 bo_fake->flags = BM_PINNED;
892 bo_fake->is_static = 1;
894 DBG("drm_bo_alloc_static: (buf %d: %s, %lu kb)\n", bo_fake->id,
895 bo_fake->name, bo_fake->bo.size / 1024);
901 drm_intel_fake_bo_reference(drm_intel_bo *bo)
903 drm_intel_bufmgr_fake *bufmgr_fake =
904 (drm_intel_bufmgr_fake *) bo->bufmgr;
905 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
907 pthread_mutex_lock(&bufmgr_fake->lock);
909 pthread_mutex_unlock(&bufmgr_fake->lock);
913 drm_intel_fake_bo_reference_locked(drm_intel_bo *bo)
915 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
921 drm_intel_fake_bo_unreference_locked(drm_intel_bo *bo)
923 drm_intel_bufmgr_fake *bufmgr_fake =
924 (drm_intel_bufmgr_fake *) bo->bufmgr;
925 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
928 if (--bo_fake->refcount == 0) {
929 assert(bo_fake->map_count == 0);
930 /* No remaining references, so free it */
932 free_block(bufmgr_fake, bo_fake->block, 1);
933 free_backing_store(bo);
935 for (i = 0; i < bo_fake->nr_relocs; i++)
936 drm_intel_fake_bo_unreference_locked(bo_fake->relocs[i].
939 DBG("drm_bo_unreference: free buf %d %s\n", bo_fake->id,
942 free(bo_fake->relocs);
948 drm_intel_fake_bo_unreference(drm_intel_bo *bo)
950 drm_intel_bufmgr_fake *bufmgr_fake =
951 (drm_intel_bufmgr_fake *) bo->bufmgr;
953 pthread_mutex_lock(&bufmgr_fake->lock);
954 drm_intel_fake_bo_unreference_locked(bo);
955 pthread_mutex_unlock(&bufmgr_fake->lock);
959 * Set the buffer as not requiring backing store, and instead get the callback
960 * invoked whenever it would be set dirty.
963 drm_intel_bo_fake_disable_backing_store(drm_intel_bo *bo,
964 void (*invalidate_cb) (drm_intel_bo *bo,
968 drm_intel_bufmgr_fake *bufmgr_fake =
969 (drm_intel_bufmgr_fake *) bo->bufmgr;
970 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
972 pthread_mutex_lock(&bufmgr_fake->lock);
974 if (bo_fake->backing_store)
975 free_backing_store(bo);
977 bo_fake->flags |= BM_NO_BACKING_STORE;
979 DBG("disable_backing_store set buf %d dirty\n", bo_fake->id);
981 bo_fake->invalidate_cb = invalidate_cb;
982 bo_fake->invalidate_ptr = ptr;
984 /* Note that it is invalid right from the start. Also note
985 * invalidate_cb is called with the bufmgr locked, so cannot
986 * itself make bufmgr calls.
988 if (invalidate_cb != NULL)
989 invalidate_cb(bo, ptr);
991 pthread_mutex_unlock(&bufmgr_fake->lock);
995 * Map a buffer into bo->virtual, allocating either card memory space (If
996 * BM_NO_BACKING_STORE or BM_PINNED) or backing store, as necessary.
999 drm_intel_fake_bo_map_locked(drm_intel_bo *bo, int write_enable)
1001 drm_intel_bufmgr_fake *bufmgr_fake =
1002 (drm_intel_bufmgr_fake *) bo->bufmgr;
1003 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
1005 /* Static buffers are always mapped. */
1006 if (bo_fake->is_static) {
1007 if (bo_fake->card_dirty) {
1008 drm_intel_bufmgr_fake_wait_idle(bufmgr_fake);
1009 bo_fake->card_dirty = 0;
1014 /* Allow recursive mapping. Mesa may recursively map buffers with
1015 * nested display loops, and it is used internally in bufmgr_fake
1018 if (bo_fake->map_count++ != 0)
1022 DBG("drm_bo_map: (buf %d: %s, %lu kb)\n", bo_fake->id,
1023 bo_fake->name, bo_fake->bo.size / 1024);
1025 if (bo->virtual != NULL) {
1026 drmMsg("%s: already mapped\n", __func__);
1028 } else if (bo_fake->flags & (BM_NO_BACKING_STORE | BM_PINNED)) {
1030 if (!bo_fake->block && !evict_and_alloc_block(bo)) {
1031 DBG("%s: alloc failed\n", __func__);
1032 bufmgr_fake->fail = 1;
1035 assert(bo_fake->block);
1038 if (!(bo_fake->flags & BM_NO_FENCE_SUBDATA) &&
1039 bo_fake->block->fenced) {
1040 drm_intel_fake_bo_wait_rendering_locked
1044 bo->virtual = bo_fake->block->virtual;
1050 if (bo_fake->backing_store == 0)
1051 alloc_backing_store(bo);
1053 if ((bo_fake->card_dirty == 1) && bo_fake->block) {
1054 if (bo_fake->block->fenced)
1055 drm_intel_fake_bo_wait_rendering_locked
1058 memcpy(bo_fake->backing_store,
1059 bo_fake->block->virtual,
1060 bo_fake->block->bo->size);
1061 bo_fake->card_dirty = 0;
1064 bo->virtual = bo_fake->backing_store;
1072 drm_intel_fake_bo_map(drm_intel_bo *bo, int write_enable)
1074 drm_intel_bufmgr_fake *bufmgr_fake =
1075 (drm_intel_bufmgr_fake *) bo->bufmgr;
1078 pthread_mutex_lock(&bufmgr_fake->lock);
1079 ret = drm_intel_fake_bo_map_locked(bo, write_enable);
1080 pthread_mutex_unlock(&bufmgr_fake->lock);
1086 drm_intel_fake_bo_unmap_locked(drm_intel_bo *bo)
1088 drm_intel_bufmgr_fake *bufmgr_fake =
1089 (drm_intel_bufmgr_fake *) bo->bufmgr;
1090 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
1092 /* Static buffers are always mapped. */
1093 if (bo_fake->is_static)
1096 assert(bo_fake->map_count != 0);
1097 if (--bo_fake->map_count != 0)
1100 DBG("drm_bo_unmap: (buf %d: %s, %lu kb)\n", bo_fake->id, bo_fake->name,
1101 bo_fake->bo.size / 1024);
1108 static int drm_intel_fake_bo_unmap(drm_intel_bo *bo)
1110 drm_intel_bufmgr_fake *bufmgr_fake =
1111 (drm_intel_bufmgr_fake *) bo->bufmgr;
1114 pthread_mutex_lock(&bufmgr_fake->lock);
1115 ret = drm_intel_fake_bo_unmap_locked(bo);
1116 pthread_mutex_unlock(&bufmgr_fake->lock);
1122 drm_intel_fake_bo_subdata(drm_intel_bo *bo, unsigned long offset,
1123 unsigned long size, const void *data)
1127 if (size == 0 || data == NULL)
1130 ret = drm_intel_bo_map(bo, 1);
1133 memcpy((unsigned char *)bo->virtual + offset, data, size);
1134 drm_intel_bo_unmap(bo);
1139 drm_intel_fake_kick_all_locked(drm_intel_bufmgr_fake *bufmgr_fake)
1141 struct block *block, *tmp;
1143 bufmgr_fake->performed_rendering = 0;
1144 /* okay for ever BO that is on the HW kick it off.
1145 seriously not afraid of the POLICE right now */
1146 DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->on_hardware) {
1147 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) block->bo;
1149 block->on_hardware = 0;
1150 free_block(bufmgr_fake, block, 0);
1151 bo_fake->block = NULL;
1152 bo_fake->validated = 0;
1153 if (!(bo_fake->flags & BM_NO_BACKING_STORE))
1160 drm_intel_fake_bo_validate(drm_intel_bo *bo)
1162 drm_intel_bufmgr_fake *bufmgr_fake;
1163 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
1165 bufmgr_fake = (drm_intel_bufmgr_fake *) bo->bufmgr;
1167 DBG("drm_bo_validate: (buf %d: %s, %lu kb)\n", bo_fake->id,
1168 bo_fake->name, bo_fake->bo.size / 1024);
1170 /* Sanity check: Buffers should be unmapped before being validated.
1171 * This is not so much of a problem for bufmgr_fake, but TTM refuses,
1172 * and the problem is harder to debug there.
1174 assert(bo_fake->map_count == 0);
1176 if (bo_fake->is_static) {
1177 /* Add it to the needs-fence list */
1178 bufmgr_fake->need_fence = 1;
1182 /* Allocate the card memory */
1183 if (!bo_fake->block && !evict_and_alloc_block(bo)) {
1184 bufmgr_fake->fail = 1;
1185 DBG("Failed to validate buf %d:%s\n", bo_fake->id,
1190 assert(bo_fake->block);
1191 assert(bo_fake->block->bo == &bo_fake->bo);
1193 bo->offset = bo_fake->block->mem->ofs;
1195 /* Upload the buffer contents if necessary */
1196 if (bo_fake->dirty) {
1197 DBG("Upload dirty buf %d:%s, sz %lu offset 0x%x\n", bo_fake->id,
1198 bo_fake->name, bo->size, bo_fake->block->mem->ofs);
1200 assert(!(bo_fake->flags & (BM_NO_BACKING_STORE | BM_PINNED)));
1202 /* Actually, should be able to just wait for a fence on the
1203 * mmory, hich we would be tracking when we free it. Waiting
1204 * for idle is a sufficiently large hammer for now.
1206 drm_intel_bufmgr_fake_wait_idle(bufmgr_fake);
1208 /* we may never have mapped this BO so it might not have any
1209 * backing store if this happens it should be rare, but 0 the
1210 * card memory in any case */
1211 if (bo_fake->backing_store)
1212 memcpy(bo_fake->block->virtual, bo_fake->backing_store,
1215 memset(bo_fake->block->virtual, 0, bo->size);
1220 bo_fake->block->fenced = 0;
1221 bo_fake->block->on_hardware = 1;
1222 DRMLISTDEL(bo_fake->block);
1223 DRMLISTADDTAIL(bo_fake->block, &bufmgr_fake->on_hardware);
1225 bo_fake->validated = 1;
1226 bufmgr_fake->need_fence = 1;
1232 drm_intel_fake_fence_validated(drm_intel_bufmgr *bufmgr)
1234 drm_intel_bufmgr_fake *bufmgr_fake = (drm_intel_bufmgr_fake *) bufmgr;
1235 unsigned int cookie;
1237 cookie = _fence_emit_internal(bufmgr_fake);
1238 fence_blocks(bufmgr_fake, cookie);
1240 DBG("drm_fence_validated: 0x%08x cookie\n", cookie);
1244 drm_intel_fake_destroy(drm_intel_bufmgr *bufmgr)
1246 drm_intel_bufmgr_fake *bufmgr_fake = (drm_intel_bufmgr_fake *) bufmgr;
1248 pthread_mutex_destroy(&bufmgr_fake->lock);
1249 mmDestroy(bufmgr_fake->heap);
1254 drm_intel_fake_emit_reloc(drm_intel_bo *bo, uint32_t offset,
1255 drm_intel_bo *target_bo, uint32_t target_offset,
1256 uint32_t read_domains, uint32_t write_domain)
1258 drm_intel_bufmgr_fake *bufmgr_fake =
1259 (drm_intel_bufmgr_fake *) bo->bufmgr;
1260 struct fake_buffer_reloc *r;
1261 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
1262 drm_intel_bo_fake *target_fake = (drm_intel_bo_fake *) target_bo;
1265 pthread_mutex_lock(&bufmgr_fake->lock);
1270 if (bo_fake->relocs == NULL) {
1272 malloc(sizeof(struct fake_buffer_reloc) * MAX_RELOCS);
1275 r = &bo_fake->relocs[bo_fake->nr_relocs++];
1277 assert(bo_fake->nr_relocs <= MAX_RELOCS);
1279 drm_intel_fake_bo_reference_locked(target_bo);
1281 if (!target_fake->is_static) {
1282 bo_fake->child_size +=
1283 ALIGN(target_bo->size, target_fake->alignment);
1284 bo_fake->child_size += target_fake->child_size;
1286 r->target_buf = target_bo;
1288 r->last_target_offset = target_bo->offset;
1289 r->delta = target_offset;
1290 r->read_domains = read_domains;
1291 r->write_domain = write_domain;
1293 if (bufmgr_fake->debug) {
1294 /* Check that a conflicting relocation hasn't already been
1297 for (i = 0; i < bo_fake->nr_relocs - 1; i++) {
1298 struct fake_buffer_reloc *r2 = &bo_fake->relocs[i];
1300 assert(r->offset != r2->offset);
1304 pthread_mutex_unlock(&bufmgr_fake->lock);
1310 * Incorporates the validation flags associated with each relocation into
1311 * the combined validation flags for the buffer on this batchbuffer submission.
1314 drm_intel_fake_calculate_domains(drm_intel_bo *bo)
1316 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
1319 for (i = 0; i < bo_fake->nr_relocs; i++) {
1320 struct fake_buffer_reloc *r = &bo_fake->relocs[i];
1321 drm_intel_bo_fake *target_fake =
1322 (drm_intel_bo_fake *) r->target_buf;
1324 /* Do the same for the tree of buffers we depend on */
1325 drm_intel_fake_calculate_domains(r->target_buf);
1327 target_fake->read_domains |= r->read_domains;
1328 target_fake->write_domain |= r->write_domain;
1333 drm_intel_fake_reloc_and_validate_buffer(drm_intel_bo *bo)
1335 drm_intel_bufmgr_fake *bufmgr_fake =
1336 (drm_intel_bufmgr_fake *) bo->bufmgr;
1337 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
1340 assert(bo_fake->map_count == 0);
1342 for (i = 0; i < bo_fake->nr_relocs; i++) {
1343 struct fake_buffer_reloc *r = &bo_fake->relocs[i];
1344 drm_intel_bo_fake *target_fake =
1345 (drm_intel_bo_fake *) r->target_buf;
1346 uint32_t reloc_data;
1348 /* Validate the target buffer if that hasn't been done. */
1349 if (!target_fake->validated) {
1351 drm_intel_fake_reloc_and_validate_buffer(r->target_buf);
1353 if (bo->virtual != NULL)
1354 drm_intel_fake_bo_unmap_locked(bo);
1359 /* Calculate the value of the relocation entry. */
1360 if (r->target_buf->offset != r->last_target_offset) {
1361 reloc_data = r->target_buf->offset + r->delta;
1363 if (bo->virtual == NULL)
1364 drm_intel_fake_bo_map_locked(bo, 1);
1366 *(uint32_t *) ((uint8_t *) bo->virtual + r->offset) =
1369 r->last_target_offset = r->target_buf->offset;
1373 if (bo->virtual != NULL)
1374 drm_intel_fake_bo_unmap_locked(bo);
1376 if (bo_fake->write_domain != 0) {
1377 if (!(bo_fake->flags & (BM_NO_BACKING_STORE | BM_PINNED))) {
1378 if (bo_fake->backing_store == 0)
1379 alloc_backing_store(bo);
1381 bo_fake->card_dirty = 1;
1382 bufmgr_fake->performed_rendering = 1;
1385 return drm_intel_fake_bo_validate(bo);
1389 drm_intel_bo_fake_post_submit(drm_intel_bo *bo)
1391 drm_intel_bufmgr_fake *bufmgr_fake =
1392 (drm_intel_bufmgr_fake *) bo->bufmgr;
1393 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo;
1396 for (i = 0; i < bo_fake->nr_relocs; i++) {
1397 struct fake_buffer_reloc *r = &bo_fake->relocs[i];
1398 drm_intel_bo_fake *target_fake =
1399 (drm_intel_bo_fake *) r->target_buf;
1401 if (target_fake->validated)
1402 drm_intel_bo_fake_post_submit(r->target_buf);
1404 DBG("%s@0x%08x + 0x%08x -> %s@0x%08x + 0x%08x\n",
1405 bo_fake->name, (uint32_t) bo->offset, r->offset,
1406 target_fake->name, (uint32_t) r->target_buf->offset,
1410 assert(bo_fake->map_count == 0);
1411 bo_fake->validated = 0;
1412 bo_fake->read_domains = 0;
1413 bo_fake->write_domain = 0;
1417 drm_intel_bufmgr_fake_set_exec_callback(drm_intel_bufmgr *bufmgr,
1418 int (*exec) (drm_intel_bo *bo,
1423 drm_intel_bufmgr_fake *bufmgr_fake = (drm_intel_bufmgr_fake *) bufmgr;
1425 bufmgr_fake->exec = exec;
1426 bufmgr_fake->exec_priv = priv;
1430 drm_intel_fake_bo_exec(drm_intel_bo *bo, int used,
1431 drm_clip_rect_t * cliprects, int num_cliprects, int DR4)
1433 drm_intel_bufmgr_fake *bufmgr_fake =
1434 (drm_intel_bufmgr_fake *) bo->bufmgr;
1435 drm_intel_bo_fake *batch_fake = (drm_intel_bo_fake *) bo;
1436 struct drm_i915_batchbuffer batch;
1438 int retry_count = 0;
1440 pthread_mutex_lock(&bufmgr_fake->lock);
1442 bufmgr_fake->performed_rendering = 0;
1444 drm_intel_fake_calculate_domains(bo);
1446 batch_fake->read_domains = I915_GEM_DOMAIN_COMMAND;
1448 /* we've ran out of RAM so blow the whole lot away and retry */
1450 ret = drm_intel_fake_reloc_and_validate_buffer(bo);
1451 if (bufmgr_fake->fail == 1) {
1452 if (retry_count == 0) {
1454 drm_intel_fake_kick_all_locked(bufmgr_fake);
1455 bufmgr_fake->fail = 0;
1457 } else /* dump out the memory here */
1458 mmDumpMemInfo(bufmgr_fake->heap);
1463 if (bufmgr_fake->exec != NULL) {
1464 ret = bufmgr_fake->exec(bo, used, bufmgr_fake->exec_priv);
1466 pthread_mutex_unlock(&bufmgr_fake->lock);
1470 batch.start = bo->offset;
1472 batch.cliprects = cliprects;
1473 batch.num_cliprects = num_cliprects;
1478 (bufmgr_fake->fd, DRM_I915_BATCHBUFFER, &batch,
1480 drmMsg("DRM_I915_BATCHBUFFER: %d\n", -errno);
1481 pthread_mutex_unlock(&bufmgr_fake->lock);
1486 drm_intel_fake_fence_validated(bo->bufmgr);
1488 drm_intel_bo_fake_post_submit(bo);
1490 pthread_mutex_unlock(&bufmgr_fake->lock);
1496 * Return an error if the list of BOs will exceed the aperture size.
1498 * This is a rough guess and likely to fail, as during the validate sequence we
1499 * may place a buffer in an inopportune spot early on and then fail to fit
1500 * a set smaller than the aperture.
1503 drm_intel_fake_check_aperture_space(drm_intel_bo ** bo_array, int count)
1505 drm_intel_bufmgr_fake *bufmgr_fake =
1506 (drm_intel_bufmgr_fake *) bo_array[0]->bufmgr;
1507 unsigned int sz = 0;
1510 for (i = 0; i < count; i++) {
1511 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) bo_array[i];
1513 if (bo_fake == NULL)
1516 if (!bo_fake->is_static)
1517 sz += ALIGN(bo_array[i]->size, bo_fake->alignment);
1518 sz += bo_fake->child_size;
1521 if (sz > bufmgr_fake->size) {
1522 DBG("check_space: overflowed bufmgr size, %ukb vs %lukb\n",
1523 sz / 1024, bufmgr_fake->size / 1024);
1527 DBG("drm_check_space: sz %ukb vs bufgr %lukb\n", sz / 1024,
1528 bufmgr_fake->size / 1024);
1533 * Evicts all buffers, waiting for fences to pass and copying contents out
1536 * Used by the X Server on LeaveVT, when the card memory is no longer our
1540 drm_intel_bufmgr_fake_evict_all(drm_intel_bufmgr *bufmgr)
1542 drm_intel_bufmgr_fake *bufmgr_fake = (drm_intel_bufmgr_fake *) bufmgr;
1543 struct block *block, *tmp;
1545 pthread_mutex_lock(&bufmgr_fake->lock);
1547 bufmgr_fake->need_fence = 1;
1548 bufmgr_fake->fail = 0;
1550 /* Wait for hardware idle. We don't know where acceleration has been
1551 * happening, so we'll need to wait anyway before letting anything get
1552 * put on the card again.
1554 drm_intel_bufmgr_fake_wait_idle(bufmgr_fake);
1556 /* Check that we hadn't released the lock without having fenced the last
1559 assert(DRMLISTEMPTY(&bufmgr_fake->fenced));
1560 assert(DRMLISTEMPTY(&bufmgr_fake->on_hardware));
1562 DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->lru) {
1563 drm_intel_bo_fake *bo_fake = (drm_intel_bo_fake *) block->bo;
1564 /* Releases the memory, and memcpys dirty contents out if
1567 free_block(bufmgr_fake, block, 0);
1568 bo_fake->block = NULL;
1571 pthread_mutex_unlock(&bufmgr_fake->lock);
1575 drm_intel_bufmgr_fake_set_last_dispatch(drm_intel_bufmgr *bufmgr,
1576 volatile unsigned int
1579 drm_intel_bufmgr_fake *bufmgr_fake = (drm_intel_bufmgr_fake *) bufmgr;
1581 bufmgr_fake->last_dispatch = (volatile int *)last_dispatch;
1585 drm_intel_bufmgr_fake_init(int fd, unsigned long low_offset,
1586 void *low_virtual, unsigned long size,
1587 volatile unsigned int *last_dispatch)
1589 drm_intel_bufmgr_fake *bufmgr_fake;
1591 bufmgr_fake = calloc(1, sizeof(*bufmgr_fake));
1593 if (pthread_mutex_init(&bufmgr_fake->lock, NULL) != 0) {
1598 /* Initialize allocator */
1599 DRMINITLISTHEAD(&bufmgr_fake->fenced);
1600 DRMINITLISTHEAD(&bufmgr_fake->on_hardware);
1601 DRMINITLISTHEAD(&bufmgr_fake->lru);
1603 bufmgr_fake->low_offset = low_offset;
1604 bufmgr_fake->virtual = low_virtual;
1605 bufmgr_fake->size = size;
1606 bufmgr_fake->heap = mmInit(low_offset, size);
1608 /* Hook in methods */
1609 bufmgr_fake->bufmgr.bo_alloc = drm_intel_fake_bo_alloc;
1610 bufmgr_fake->bufmgr.bo_alloc_for_render = drm_intel_fake_bo_alloc;
1611 bufmgr_fake->bufmgr.bo_alloc_tiled = drm_intel_fake_bo_alloc_tiled;
1612 bufmgr_fake->bufmgr.bo_reference = drm_intel_fake_bo_reference;
1613 bufmgr_fake->bufmgr.bo_unreference = drm_intel_fake_bo_unreference;
1614 bufmgr_fake->bufmgr.bo_map = drm_intel_fake_bo_map;
1615 bufmgr_fake->bufmgr.bo_unmap = drm_intel_fake_bo_unmap;
1616 bufmgr_fake->bufmgr.bo_subdata = drm_intel_fake_bo_subdata;
1617 bufmgr_fake->bufmgr.bo_wait_rendering =
1618 drm_intel_fake_bo_wait_rendering;
1619 bufmgr_fake->bufmgr.bo_emit_reloc = drm_intel_fake_emit_reloc;
1620 bufmgr_fake->bufmgr.destroy = drm_intel_fake_destroy;
1621 bufmgr_fake->bufmgr.bo_exec = drm_intel_fake_bo_exec;
1622 bufmgr_fake->bufmgr.check_aperture_space =
1623 drm_intel_fake_check_aperture_space;
1624 bufmgr_fake->bufmgr.debug = 0;
1626 bufmgr_fake->fd = fd;
1627 bufmgr_fake->last_dispatch = (volatile int *)last_dispatch;
1629 return &bufmgr_fake->bufmgr;