Merge branch 'mesa_7_6_branch' into mesa_7_7_branch
[mesa.git] / src / mesa / drivers / dri / intel / intel_buffer_objects.c
1 /**************************************************************************
2 *
3 * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
4 * All Rights Reserved.
5 *
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:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
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.
25 *
26 **************************************************************************/
27
28
29 #include "main/imports.h"
30 #include "main/mtypes.h"
31 #include "main/macros.h"
32 #include "main/bufferobj.h"
33
34 #include "intel_context.h"
35 #include "intel_blit.h"
36 #include "intel_buffer_objects.h"
37 #include "intel_batchbuffer.h"
38 #include "intel_regions.h"
39
40 static GLboolean
41 intel_bufferobj_unmap(GLcontext * ctx,
42 GLenum target, struct gl_buffer_object *obj);
43
44 /** Allocates a new dri_bo to store the data for the buffer object. */
45 static void
46 intel_bufferobj_alloc_buffer(struct intel_context *intel,
47 struct intel_buffer_object *intel_obj)
48 {
49 intel_obj->buffer = dri_bo_alloc(intel->bufmgr, "bufferobj",
50 intel_obj->Base.Size, 64);
51 }
52
53 /**
54 * There is some duplication between mesa's bufferobjects and our
55 * bufmgr buffers. Both have an integer handle and a hashtable to
56 * lookup an opaque structure. It would be nice if the handles and
57 * internal structure where somehow shared.
58 */
59 static struct gl_buffer_object *
60 intel_bufferobj_alloc(GLcontext * ctx, GLuint name, GLenum target)
61 {
62 struct intel_buffer_object *obj = CALLOC_STRUCT(intel_buffer_object);
63
64 _mesa_initialize_buffer_object(&obj->Base, name, target);
65
66 obj->buffer = NULL;
67
68 return &obj->Base;
69 }
70
71 /* Break the COW tie to the region. The region gets to keep the data.
72 */
73 void
74 intel_bufferobj_release_region(struct intel_context *intel,
75 struct intel_buffer_object *intel_obj)
76 {
77 assert(intel_obj->region->buffer == intel_obj->buffer);
78 intel_obj->region->pbo = NULL;
79 intel_obj->region = NULL;
80
81 dri_bo_unreference(intel_obj->buffer);
82 intel_obj->buffer = NULL;
83 }
84
85 /* Break the COW tie to the region. Both the pbo and the region end
86 * up with a copy of the data.
87 */
88 void
89 intel_bufferobj_cow(struct intel_context *intel,
90 struct intel_buffer_object *intel_obj)
91 {
92 assert(intel_obj->region);
93 intel_region_cow(intel, intel_obj->region);
94 }
95
96
97 /**
98 * Deallocate/free a vertex/pixel buffer object.
99 * Called via glDeleteBuffersARB().
100 */
101 static void
102 intel_bufferobj_free(GLcontext * ctx, struct gl_buffer_object *obj)
103 {
104 struct intel_context *intel = intel_context(ctx);
105 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
106
107 assert(intel_obj);
108
109 /* Buffer objects are automatically unmapped when deleting according
110 * to the spec, but Mesa doesn't do UnmapBuffer for us at context destroy
111 * (though it does if you call glDeleteBuffers)
112 */
113 if (obj->Pointer)
114 intel_bufferobj_unmap(ctx, 0, obj);
115
116 _mesa_free(intel_obj->sys_buffer);
117 if (intel_obj->region) {
118 intel_bufferobj_release_region(intel, intel_obj);
119 }
120 else if (intel_obj->buffer) {
121 dri_bo_unreference(intel_obj->buffer);
122 }
123
124 _mesa_free(intel_obj);
125 }
126
127
128
129 /**
130 * Allocate space for and store data in a buffer object. Any data that was
131 * previously stored in the buffer object is lost. If data is NULL,
132 * memory will be allocated, but no copy will occur.
133 * Called via ctx->Driver.BufferData().
134 * \return GL_TRUE for success, GL_FALSE if out of memory
135 */
136 static GLboolean
137 intel_bufferobj_data(GLcontext * ctx,
138 GLenum target,
139 GLsizeiptrARB size,
140 const GLvoid * data,
141 GLenum usage, struct gl_buffer_object *obj)
142 {
143 struct intel_context *intel = intel_context(ctx);
144 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
145
146 intel_obj->Base.Size = size;
147 intel_obj->Base.Usage = usage;
148
149 assert(!obj->Pointer); /* Mesa should have unmapped it */
150
151 if (intel_obj->region)
152 intel_bufferobj_release_region(intel, intel_obj);
153
154 if (intel_obj->buffer != NULL) {
155 dri_bo_unreference(intel_obj->buffer);
156 intel_obj->buffer = NULL;
157 }
158 _mesa_free(intel_obj->sys_buffer);
159 intel_obj->sys_buffer = NULL;
160
161 if (size != 0) {
162 #ifdef I915
163 /* On pre-965, stick VBOs in system memory, as we're always doing swtnl
164 * with their contents anyway.
165 */
166 if (target == GL_ARRAY_BUFFER || target == GL_ELEMENT_ARRAY_BUFFER) {
167 intel_obj->sys_buffer = _mesa_malloc(size);
168 if (intel_obj->sys_buffer != NULL) {
169 if (data != NULL)
170 memcpy(intel_obj->sys_buffer, data, size);
171 return GL_TRUE;
172 }
173 }
174 #endif
175 intel_bufferobj_alloc_buffer(intel, intel_obj);
176 if (!intel_obj->buffer)
177 return GL_FALSE;
178
179 if (data != NULL)
180 dri_bo_subdata(intel_obj->buffer, 0, size, data);
181 }
182
183 return GL_TRUE;
184 }
185
186
187 /**
188 * Replace data in a subrange of buffer object. If the data range
189 * specified by size + offset extends beyond the end of the buffer or
190 * if data is NULL, no copy is performed.
191 * Called via glBufferSubDataARB().
192 */
193 static void
194 intel_bufferobj_subdata(GLcontext * ctx,
195 GLenum target,
196 GLintptrARB offset,
197 GLsizeiptrARB size,
198 const GLvoid * data, struct gl_buffer_object *obj)
199 {
200 struct intel_context *intel = intel_context(ctx);
201 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
202
203 assert(intel_obj);
204
205 if (intel_obj->region)
206 intel_bufferobj_cow(intel, intel_obj);
207
208 if (intel_obj->sys_buffer)
209 memcpy((char *)intel_obj->sys_buffer + offset, data, size);
210 else {
211 /* Flush any existing batchbuffer that might reference this data. */
212 if (drm_intel_bo_busy(intel_obj->buffer) ||
213 drm_intel_bo_references(intel->batch->buf, intel_obj->buffer)) {
214 drm_intel_bo *temp_bo;
215
216 temp_bo = drm_intel_bo_alloc(intel->bufmgr, "subdata temp", size, 64);
217
218 drm_intel_bo_subdata(temp_bo, 0, size, data);
219
220 intel_emit_linear_blit(intel,
221 intel_obj->buffer, offset,
222 temp_bo, 0,
223 size);
224
225 drm_intel_bo_unreference(temp_bo);
226 } else {
227 dri_bo_subdata(intel_obj->buffer, offset, size, data);
228 }
229 }
230 }
231
232
233 /**
234 * Called via glGetBufferSubDataARB().
235 */
236 static void
237 intel_bufferobj_get_subdata(GLcontext * ctx,
238 GLenum target,
239 GLintptrARB offset,
240 GLsizeiptrARB size,
241 GLvoid * data, struct gl_buffer_object *obj)
242 {
243 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
244
245 assert(intel_obj);
246 if (intel_obj->sys_buffer)
247 memcpy(data, (char *)intel_obj->sys_buffer + offset, size);
248 else
249 dri_bo_get_subdata(intel_obj->buffer, offset, size, data);
250 }
251
252
253
254 /**
255 * Called via glMapBufferARB().
256 */
257 static void *
258 intel_bufferobj_map(GLcontext * ctx,
259 GLenum target,
260 GLenum access, struct gl_buffer_object *obj)
261 {
262 struct intel_context *intel = intel_context(ctx);
263 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
264 GLboolean read_only = (access == GL_READ_ONLY_ARB);
265 GLboolean write_only = (access == GL_WRITE_ONLY_ARB);
266
267 assert(intel_obj);
268
269 if (intel_obj->sys_buffer) {
270 obj->Pointer = intel_obj->sys_buffer;
271 obj->Length = obj->Size;
272 obj->Offset = 0;
273 return obj->Pointer;
274 }
275
276 /* Flush any existing batchbuffer that might reference this data. */
277 if (drm_intel_bo_references(intel->batch->buf, intel_obj->buffer))
278 intelFlush(ctx);
279
280 if (intel_obj->region)
281 intel_bufferobj_cow(intel, intel_obj);
282
283 if (intel_obj->buffer == NULL) {
284 obj->Pointer = NULL;
285 return NULL;
286 }
287
288 if (write_only && intel->intelScreen->kernel_exec_fencing) {
289 drm_intel_gem_bo_map_gtt(intel_obj->buffer);
290 intel_obj->mapped_gtt = GL_TRUE;
291 } else {
292 drm_intel_bo_map(intel_obj->buffer, !read_only);
293 intel_obj->mapped_gtt = GL_FALSE;
294 }
295
296 obj->Pointer = intel_obj->buffer->virtual;
297 obj->Length = obj->Size;
298 obj->Offset = 0;
299
300 return obj->Pointer;
301 }
302
303 /**
304 * Called via glMapBufferRange().
305 *
306 * The goal of this extension is to allow apps to accumulate their rendering
307 * at the same time as they accumulate their buffer object. Without it,
308 * you'd end up blocking on execution of rendering every time you mapped
309 * the buffer to put new data in.
310 *
311 * We support it in 3 ways: If unsynchronized, then don't bother
312 * flushing the batchbuffer before mapping the buffer, which can save blocking
313 * in many cases. If we would still block, and they allow the whole buffer
314 * to be invalidated, then just allocate a new buffer to replace the old one.
315 * If not, and we'd block, and they allow the subrange of the buffer to be
316 * invalidated, then we can make a new little BO, let them write into that,
317 * and blit it into the real BO at unmap time.
318 */
319 static void *
320 intel_bufferobj_map_range(GLcontext * ctx,
321 GLenum target, GLintptr offset, GLsizeiptr length,
322 GLbitfield access, struct gl_buffer_object *obj)
323 {
324 struct intel_context *intel = intel_context(ctx);
325 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
326
327 assert(intel_obj);
328
329 /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
330 * internally uses our functions directly.
331 */
332 obj->Offset = offset;
333 obj->Length = length;
334 obj->AccessFlags = access;
335
336 if (intel_obj->sys_buffer) {
337 obj->Pointer = intel_obj->sys_buffer + offset;
338 return obj->Pointer;
339 }
340
341 if (intel_obj->region)
342 intel_bufferobj_cow(intel, intel_obj);
343
344 /* If the mapping is synchronized with other GL operations, flush
345 * the batchbuffer so that GEM knows about the buffer access for later
346 * syncing.
347 */
348 if (!(access & GL_MAP_UNSYNCHRONIZED_BIT) &&
349 drm_intel_bo_references(intel->batch->buf, intel_obj->buffer))
350 intelFlush(ctx);
351
352 if (intel_obj->buffer == NULL) {
353 obj->Pointer = NULL;
354 return NULL;
355 }
356
357 /* If the user doesn't care about existing buffer contents and mapping
358 * would cause us to block, then throw out the old buffer.
359 */
360 if (!(access & GL_MAP_UNSYNCHRONIZED_BIT) &&
361 (access & GL_MAP_INVALIDATE_BUFFER_BIT) &&
362 drm_intel_bo_busy(intel_obj->buffer)) {
363 drm_intel_bo_unreference(intel_obj->buffer);
364 intel_obj->buffer = dri_bo_alloc(intel->bufmgr, "bufferobj",
365 intel_obj->Base.Size, 64);
366 }
367
368 /* If the user is mapping a range of an active buffer object but
369 * doesn't require the current contents of that range, make a new
370 * BO, and we'll copy what they put in there out at unmap or
371 * FlushRange time.
372 */
373 if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
374 drm_intel_bo_busy(intel_obj->buffer)) {
375 if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
376 intel_obj->range_map_buffer = _mesa_malloc(length);
377 obj->Pointer = intel_obj->range_map_buffer;
378 } else {
379 intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
380 "range map",
381 length, 64);
382 if (!(access & GL_MAP_READ_BIT) &&
383 intel->intelScreen->kernel_exec_fencing) {
384 drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
385 intel_obj->mapped_gtt = GL_TRUE;
386 } else {
387 drm_intel_bo_map(intel_obj->range_map_bo,
388 (access & GL_MAP_WRITE_BIT) != 0);
389 intel_obj->mapped_gtt = GL_FALSE;
390 }
391 obj->Pointer = intel_obj->range_map_bo->virtual;
392 }
393 return obj->Pointer;
394 }
395
396 if (!(access & GL_MAP_READ_BIT) &&
397 intel->intelScreen->kernel_exec_fencing) {
398 drm_intel_gem_bo_map_gtt(intel_obj->buffer);
399 intel_obj->mapped_gtt = GL_TRUE;
400 } else {
401 drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
402 intel_obj->mapped_gtt = GL_FALSE;
403 }
404
405 obj->Pointer = intel_obj->buffer->virtual + offset;
406 return obj->Pointer;
407 }
408
409 /* Ideally we'd use a BO to avoid taking up cache space for the temporary
410 * data, but FlushMappedBufferRange may be followed by further writes to
411 * the pointer, so we would have to re-map after emitting our blit, which
412 * would defeat the point.
413 */
414 static void
415 intel_bufferobj_flush_mapped_range(GLcontext *ctx, GLenum target,
416 GLintptr offset, GLsizeiptr length,
417 struct gl_buffer_object *obj)
418 {
419 struct intel_context *intel = intel_context(ctx);
420 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
421 drm_intel_bo *temp_bo;
422
423 /* Unless we're in the range map using a temporary system buffer,
424 * there's no work to do.
425 */
426 if (intel_obj->range_map_buffer == NULL)
427 return;
428
429 temp_bo = drm_intel_bo_alloc(intel->bufmgr, "range map flush", length, 64);
430
431 drm_intel_bo_subdata(temp_bo, 0, length, intel_obj->range_map_buffer);
432
433 intel_emit_linear_blit(intel,
434 intel_obj->buffer, obj->Offset + offset,
435 temp_bo, 0,
436 length);
437
438 drm_intel_bo_unreference(temp_bo);
439 }
440
441
442 /**
443 * Called via glUnmapBuffer().
444 */
445 static GLboolean
446 intel_bufferobj_unmap(GLcontext * ctx,
447 GLenum target, struct gl_buffer_object *obj)
448 {
449 struct intel_context *intel = intel_context(ctx);
450 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
451
452 assert(intel_obj);
453 assert(obj->Pointer);
454 if (intel_obj->sys_buffer != NULL) {
455 /* always keep the mapping around. */
456 } else if (intel_obj->range_map_buffer != NULL) {
457 /* Since we've emitted some blits to buffers that will (likely) be used
458 * in rendering operations in other cache domains in this batch, emit a
459 * flush. Once again, we wish for a domain tracker in libdrm to cover
460 * usage inside of a batchbuffer.
461 */
462 intel_batchbuffer_emit_mi_flush(intel->batch);
463 free(intel_obj->range_map_buffer);
464 intel_obj->range_map_buffer = NULL;
465 } else if (intel_obj->range_map_bo != NULL) {
466 if (intel_obj->mapped_gtt) {
467 drm_intel_gem_bo_unmap_gtt(intel_obj->range_map_bo);
468 } else {
469 drm_intel_bo_unmap(intel_obj->range_map_bo);
470 }
471
472 intel_emit_linear_blit(intel,
473 intel_obj->buffer, obj->Offset,
474 intel_obj->range_map_bo, 0,
475 obj->Length);
476
477 /* Since we've emitted some blits to buffers that will (likely) be used
478 * in rendering operations in other cache domains in this batch, emit a
479 * flush. Once again, we wish for a domain tracker in libdrm to cover
480 * usage inside of a batchbuffer.
481 */
482 intel_batchbuffer_emit_mi_flush(intel->batch);
483
484 drm_intel_bo_unreference(intel_obj->range_map_bo);
485 intel_obj->range_map_bo = NULL;
486 } else if (intel_obj->buffer != NULL) {
487 if (intel_obj->mapped_gtt) {
488 drm_intel_gem_bo_unmap_gtt(intel_obj->buffer);
489 } else {
490 drm_intel_bo_unmap(intel_obj->buffer);
491 }
492 }
493 obj->Pointer = NULL;
494 obj->Offset = 0;
495 obj->Length = 0;
496
497 return GL_TRUE;
498 }
499
500 dri_bo *
501 intel_bufferobj_buffer(struct intel_context *intel,
502 struct intel_buffer_object *intel_obj, GLuint flag)
503 {
504 if (intel_obj->region) {
505 if (flag == INTEL_WRITE_PART)
506 intel_bufferobj_cow(intel, intel_obj);
507 else if (flag == INTEL_WRITE_FULL) {
508 intel_bufferobj_release_region(intel, intel_obj);
509 intel_bufferobj_alloc_buffer(intel, intel_obj);
510 }
511 }
512
513 if (intel_obj->buffer == NULL) {
514 void *sys_buffer = intel_obj->sys_buffer;
515
516 /* only one of buffer and sys_buffer could be non-NULL */
517 intel_bufferobj_alloc_buffer(intel, intel_obj);
518 intel_obj->sys_buffer = NULL;
519
520 intel_bufferobj_subdata(&intel->ctx,
521 GL_ARRAY_BUFFER_ARB,
522 0,
523 intel_obj->Base.Size,
524 sys_buffer,
525 &intel_obj->Base);
526 _mesa_free(sys_buffer);
527 intel_obj->sys_buffer = NULL;
528 }
529
530 return intel_obj->buffer;
531 }
532
533 static void
534 intel_bufferobj_copy_subdata(GLcontext *ctx,
535 struct gl_buffer_object *src,
536 struct gl_buffer_object *dst,
537 GLintptr read_offset, GLintptr write_offset,
538 GLsizeiptr size)
539 {
540 struct intel_context *intel = intel_context(ctx);
541 struct intel_buffer_object *intel_src = intel_buffer_object(src);
542 struct intel_buffer_object *intel_dst = intel_buffer_object(dst);
543 drm_intel_bo *src_bo, *dst_bo;
544
545 if (size == 0)
546 return;
547
548 /* If we're in system memory, just map and memcpy. */
549 if (intel_src->sys_buffer || intel_dst->sys_buffer) {
550 /* The same buffer may be used, but note that regions copied may
551 * not overlap.
552 */
553 if (src == dst) {
554 char *ptr = intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
555 GL_READ_WRITE, dst);
556 memcpy(ptr + write_offset, ptr + read_offset, size);
557 intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
558 } else {
559 const char *src_ptr;
560 char *dst_ptr;
561
562 src_ptr = intel_bufferobj_map(ctx, GL_COPY_READ_BUFFER,
563 GL_READ_ONLY, src);
564 dst_ptr = intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
565 GL_WRITE_ONLY, dst);
566
567 memcpy(dst_ptr + write_offset, src_ptr + read_offset, size);
568
569 intel_bufferobj_unmap(ctx, GL_COPY_READ_BUFFER, src);
570 intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
571 }
572 }
573
574 /* Otherwise, we have real BOs, so blit them. */
575
576 dst_bo = intel_bufferobj_buffer(intel, intel_dst, INTEL_WRITE_PART);
577 src_bo = intel_bufferobj_buffer(intel, intel_src, INTEL_READ);
578
579 intel_emit_linear_blit(intel,
580 dst_bo, write_offset,
581 src_bo, read_offset, size);
582
583 /* Since we've emitted some blits to buffers that will (likely) be used
584 * in rendering operations in other cache domains in this batch, emit a
585 * flush. Once again, we wish for a domain tracker in libdrm to cover
586 * usage inside of a batchbuffer.
587 */
588 intel_batchbuffer_emit_mi_flush(intel->batch);
589 }
590
591 void
592 intelInitBufferObjectFuncs(struct dd_function_table *functions)
593 {
594 functions->NewBufferObject = intel_bufferobj_alloc;
595 functions->DeleteBuffer = intel_bufferobj_free;
596 functions->BufferData = intel_bufferobj_data;
597 functions->BufferSubData = intel_bufferobj_subdata;
598 functions->GetBufferSubData = intel_bufferobj_get_subdata;
599 functions->MapBuffer = intel_bufferobj_map;
600 functions->MapBufferRange = intel_bufferobj_map_range;
601 functions->FlushMappedBufferRange = intel_bufferobj_flush_mapped_range;
602 functions->UnmapBuffer = intel_bufferobj_unmap;
603 functions->CopyBufferSubData = intel_bufferobj_copy_subdata;
604 }