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 /* Provide additional functionality on top of bufmgr buffers:
29 * - 2d semantics and blit operations
30 * - refcounting of buffers for multiple images in a buffer.
31 * - refcounting of buffer mappings.
32 * - some logic for moving the buffers to the best memory pools for
35 * Most of this is to make it easier to implement the fixed-layout
36 * mipmap tree required by intel hardware in the face of GL's
37 * programming interface where each image can be specifed in random
38 * order and it isn't clear what layout the tree should have until the
42 #include <sys/ioctl.h>
45 #include "intel_context.h"
46 #include "intel_regions.h"
47 #include "intel_blit.h"
48 #include "intel_buffer_objects.h"
49 #include "intel_bufmgr.h"
50 #include "intel_batchbuffer.h"
51 #include "intel_chipset.h"
53 #define FILE_DEBUG_FLAG DEBUG_REGION
55 /* This should be set to the maximum backtrace size desired.
56 * Set it to 0 to disable backtrace debugging.
58 #define DEBUG_BACKTRACE_SIZE 0
60 #if DEBUG_BACKTRACE_SIZE == 0
61 /* Use the standard debug output */
62 #define _DBG(...) DBG(__VA_ARGS__)
64 /* Use backtracing debug output */
65 #define _DBG(...) {debug_backtrace(); DBG(__VA_ARGS__);}
67 /* Backtracing debug support */
73 void *trace
[DEBUG_BACKTRACE_SIZE
];
74 char **strings
= NULL
;
78 traceSize
= backtrace(trace
, DEBUG_BACKTRACE_SIZE
);
79 strings
= backtrace_symbols(trace
, traceSize
);
80 if (strings
== NULL
) {
85 /* Spit out all the strings with a colon separator. Ignore
86 * the first, since we don't really care about the call
87 * to debug_backtrace() itself. Skip until the final "/" in
88 * the trace to avoid really long lines.
90 for (i
= 1; i
< traceSize
; i
++) {
91 char *p
= strings
[i
], *slash
= strings
[i
];
101 /* Free up the memory, and we're done */
109 /* XXX: Thread safety?
112 intel_region_map(struct intel_context
*intel
, struct intel_region
*region
)
114 intelFlush(&intel
->ctx
);
116 _DBG("%s %p\n", __FUNCTION__
, region
);
117 if (!region
->map_refcount
++) {
119 intel_region_cow(intel
, region
);
121 if (region
->tiling
!= I915_TILING_NONE
&&
122 intel
->intelScreen
->kernel_exec_fencing
)
123 drm_intel_gem_bo_map_gtt(region
->buffer
);
125 dri_bo_map(region
->buffer
, GL_TRUE
);
126 region
->map
= region
->buffer
->virtual;
133 intel_region_unmap(struct intel_context
*intel
, struct intel_region
*region
)
135 _DBG("%s %p\n", __FUNCTION__
, region
);
136 if (!--region
->map_refcount
) {
137 if (region
->tiling
!= I915_TILING_NONE
&&
138 intel
->intelScreen
->kernel_exec_fencing
)
139 drm_intel_gem_bo_unmap_gtt(region
->buffer
);
141 dri_bo_unmap(region
->buffer
);
146 static struct intel_region
*
147 intel_region_alloc_internal(struct intel_context
*intel
,
149 GLuint width
, GLuint height
, GLuint pitch
,
152 struct intel_region
*region
;
154 if (buffer
== NULL
) {
155 _DBG("%s <-- NULL\n", __FUNCTION__
);
159 region
= calloc(sizeof(*region
), 1);
161 region
->width
= width
;
162 region
->height
= height
;
163 region
->pitch
= pitch
;
164 region
->refcount
= 1;
165 region
->buffer
= buffer
;
167 /* Default to no tiling */
168 region
->tiling
= I915_TILING_NONE
;
169 region
->bit_6_swizzle
= I915_BIT_6_SWIZZLE_NONE
;
171 _DBG("%s <-- %p\n", __FUNCTION__
, region
);
175 struct intel_region
*
176 intel_region_alloc(struct intel_context
*intel
,
178 GLuint cpp
, GLuint width
, GLuint height
, GLuint pitch
,
179 GLboolean expect_accelerated_upload
)
182 struct intel_region
*region
;
184 /* If we're tiled, our allocations are in 8 or 32-row blocks, so
185 * failure to align our height means that we won't allocate enough pages.
187 * If we're untiled, we still have to align to 2 rows high because the
188 * data port accesses 2x2 blocks even if the bottom row isn't to be
189 * rendered, so failure to align means we could walk off the end of the
192 if (tiling
== I915_TILING_X
)
193 height
= ALIGN(height
, 8);
194 else if (tiling
== I915_TILING_Y
)
195 height
= ALIGN(height
, 32);
197 height
= ALIGN(height
, 2);
199 if (expect_accelerated_upload
) {
200 buffer
= drm_intel_bo_alloc_for_render(intel
->bufmgr
, "region",
201 pitch
* cpp
* height
, 64);
203 buffer
= drm_intel_bo_alloc(intel
->bufmgr
, "region",
204 pitch
* cpp
* height
, 64);
207 region
= intel_region_alloc_internal(intel
, cpp
, width
, height
,
210 if (tiling
!= I915_TILING_NONE
) {
211 assert(((pitch
* cpp
) & 127) == 0);
212 drm_intel_bo_set_tiling(buffer
, &tiling
, pitch
* cpp
);
213 drm_intel_bo_get_tiling(buffer
, ®ion
->tiling
, ®ion
->bit_6_swizzle
);
219 struct intel_region
*
220 intel_region_alloc_for_handle(struct intel_context
*intel
,
222 GLuint width
, GLuint height
, GLuint pitch
,
223 GLuint handle
, const char *name
)
225 struct intel_region
*region
;
229 buffer
= intel_bo_gem_create_from_name(intel
->bufmgr
, name
, handle
);
231 region
= intel_region_alloc_internal(intel
, cpp
,
232 width
, height
, pitch
, buffer
);
236 ret
= dri_bo_get_tiling(region
->buffer
, ®ion
->tiling
,
237 ®ion
->bit_6_swizzle
);
239 fprintf(stderr
, "Couldn't get tiling of buffer %d (%s): %s\n",
240 handle
, name
, strerror(-ret
));
241 intel_region_release(®ion
);
249 intel_region_reference(struct intel_region
**dst
, struct intel_region
*src
)
252 _DBG("%s %p %d\n", __FUNCTION__
, src
, src
->refcount
);
254 assert(*dst
== NULL
);
262 intel_region_release(struct intel_region
**region_handle
)
264 struct intel_region
*region
= *region_handle
;
266 if (region
== NULL
) {
267 _DBG("%s NULL\n", __FUNCTION__
);
271 _DBG("%s %p %d\n", __FUNCTION__
, region
, region
->refcount
- 1);
273 ASSERT(region
->refcount
> 0);
276 if (region
->refcount
== 0) {
277 assert(region
->map_refcount
== 0);
280 region
->pbo
->region
= NULL
;
282 dri_bo_unreference(region
->buffer
);
284 if (region
->classic_map
!= NULL
) {
285 drmUnmap(region
->classic_map
,
286 region
->pitch
* region
->cpp
* region
->height
);
291 *region_handle
= NULL
;
295 * XXX Move this into core Mesa?
298 _mesa_copy_rect(GLubyte
* dst
,
306 GLuint src_pitch
, GLuint src_x
, GLuint src_y
)
314 dst
+= dst_y
* dst_pitch
;
315 src
+= src_y
* dst_pitch
;
318 if (width
== dst_pitch
&& width
== src_pitch
)
319 memcpy(dst
, src
, height
* width
);
321 for (i
= 0; i
< height
; i
++) {
322 memcpy(dst
, src
, width
);
330 /* Upload data to a rectangular sub-region. Lots of choices how to do this:
332 * - memcpy by span to current destination
333 * - upload data as new buffer and blit
335 * Currently always memcpy.
338 intel_region_data(struct intel_context
*intel
,
339 struct intel_region
*dst
,
341 GLuint dstx
, GLuint dsty
,
342 const void *src
, GLuint src_pitch
,
343 GLuint srcx
, GLuint srcy
, GLuint width
, GLuint height
)
345 _DBG("%s\n", __FUNCTION__
);
352 dsty
== 0 && width
== dst
->pitch
&& height
== dst
->height
)
353 intel_region_release_pbo(intel
, dst
);
355 intel_region_cow(intel
, dst
);
358 LOCK_HARDWARE(intel
);
359 _mesa_copy_rect(intel_region_map(intel
, dst
) + dst_offset
,
362 dstx
, dsty
, width
, height
, src
, src_pitch
, srcx
, srcy
);
364 intel_region_unmap(intel
, dst
);
365 UNLOCK_HARDWARE(intel
);
368 /* Copy rectangular sub-regions. Need better logic about when to
369 * push buffers into AGP - will currently do so whenever possible.
372 intel_region_copy(struct intel_context
*intel
,
373 struct intel_region
*dst
,
375 GLuint dstx
, GLuint dsty
,
376 struct intel_region
*src
,
378 GLuint srcx
, GLuint srcy
, GLuint width
, GLuint height
,
381 _DBG("%s\n", __FUNCTION__
);
388 dsty
== 0 && width
== dst
->pitch
&& height
== dst
->height
)
389 intel_region_release_pbo(intel
, dst
);
391 intel_region_cow(intel
, dst
);
394 assert(src
->cpp
== dst
->cpp
);
396 return intelEmitCopyBlit(intel
,
398 src
->pitch
, src
->buffer
, src_offset
, src
->tiling
,
399 dst
->pitch
, dst
->buffer
, dst_offset
, dst
->tiling
,
400 srcx
, srcy
, dstx
, dsty
, width
, height
,
404 /* Attach to a pbo, discarding our data. Effectively zero-copy upload
408 intel_region_attach_pbo(struct intel_context
*intel
,
409 struct intel_region
*region
,
410 struct intel_buffer_object
*pbo
)
414 if (region
->pbo
== pbo
)
417 _DBG("%s %p %p\n", __FUNCTION__
, region
, pbo
);
419 /* If there is already a pbo attached, break the cow tie now.
420 * Don't call intel_region_release_pbo() as that would
421 * unnecessarily allocate a new buffer we would have to immediately
425 region
->pbo
->region
= NULL
;
429 if (region
->buffer
) {
430 dri_bo_unreference(region
->buffer
);
431 region
->buffer
= NULL
;
434 /* make sure pbo has a buffer of its own */
435 buffer
= intel_bufferobj_buffer(intel
, pbo
, INTEL_WRITE_FULL
);
438 region
->pbo
->region
= region
;
439 dri_bo_reference(buffer
);
440 region
->buffer
= buffer
;
444 /* Break the COW tie to the pbo and allocate a new buffer.
445 * The pbo gets to keep the data.
448 intel_region_release_pbo(struct intel_context
*intel
,
449 struct intel_region
*region
)
451 _DBG("%s %p\n", __FUNCTION__
, region
);
452 assert(region
->buffer
== region
->pbo
->buffer
);
453 region
->pbo
->region
= NULL
;
455 dri_bo_unreference(region
->buffer
);
456 region
->buffer
= NULL
;
458 region
->buffer
= dri_bo_alloc(intel
->bufmgr
, "region",
459 region
->pitch
* region
->cpp
* region
->height
,
463 /* Break the COW tie to the pbo. Both the pbo and the region end up
464 * with a copy of the data.
467 intel_region_cow(struct intel_context
*intel
, struct intel_region
*region
)
469 struct intel_buffer_object
*pbo
= region
->pbo
;
472 intel_region_release_pbo(intel
, region
);
474 assert(region
->cpp
* region
->pitch
* region
->height
== pbo
->Base
.Size
);
476 _DBG("%s %p (%d bytes)\n", __FUNCTION__
, region
, pbo
->Base
.Size
);
478 /* Now blit from the texture buffer to the new buffer:
481 LOCK_HARDWARE(intel
);
482 ok
= intelEmitCopyBlit(intel
,
484 region
->pitch
, pbo
->buffer
, 0, region
->tiling
,
485 region
->pitch
, region
->buffer
, 0, region
->tiling
,
487 region
->pitch
, region
->height
,
490 UNLOCK_HARDWARE(intel
);
494 intel_region_buffer(struct intel_context
*intel
,
495 struct intel_region
*region
, GLuint flag
)
498 if (flag
== INTEL_WRITE_PART
)
499 intel_region_cow(intel
, region
);
500 else if (flag
== INTEL_WRITE_FULL
)
501 intel_region_release_pbo(intel
, region
);
504 return region
->buffer
;
507 static struct intel_region
*
508 intel_recreate_static(struct intel_context
*intel
,
510 struct intel_region
*region
,
511 intelRegion
*region_desc
)
513 intelScreenPrivate
*intelScreen
= intel
->intelScreen
;
516 if (region
== NULL
) {
517 region
= calloc(sizeof(*region
), 1);
518 region
->refcount
= 1;
519 _DBG("%s creating new region %p\n", __FUNCTION__
, region
);
522 _DBG("%s %p\n", __FUNCTION__
, region
);
525 if (intel
->ctx
.Visual
.rgbBits
== 24)
528 region
->cpp
= intel
->ctx
.Visual
.rgbBits
/ 8;
529 region
->pitch
= intelScreen
->pitch
;
530 region
->width
= intelScreen
->width
;
531 region
->height
= intelScreen
->height
;
533 if (region
->buffer
!= NULL
) {
534 dri_bo_unreference(region
->buffer
);
535 region
->buffer
= NULL
;
539 assert(region_desc
->bo_handle
!= -1);
540 region
->buffer
= intel_bo_gem_create_from_name(intel
->bufmgr
,
542 region_desc
->bo_handle
);
544 ret
= dri_bo_get_tiling(region
->buffer
, ®ion
->tiling
,
545 ®ion
->bit_6_swizzle
);
547 fprintf(stderr
, "Couldn't get tiling of buffer %d (%s): %s\n",
548 region_desc
->bo_handle
, name
, strerror(-ret
));
549 intel_region_release(®ion
);
553 if (region
->classic_map
!= NULL
) {
554 drmUnmap(region
->classic_map
,
555 region
->pitch
* region
->cpp
* region
->height
);
556 region
->classic_map
= NULL
;
558 ret
= drmMap(intel
->driFd
, region_desc
->handle
,
559 region
->pitch
* region
->cpp
* region
->height
,
560 ®ion
->classic_map
);
562 fprintf(stderr
, "Failed to drmMap %s buffer\n", name
);
567 region
->buffer
= intel_bo_fake_alloc_static(intel
->bufmgr
,
570 region
->pitch
* region
->cpp
*
572 region
->classic_map
);
574 /* The sarea just gives us a boolean for whether it's tiled or not,
575 * instead of which tiling mode it is. Guess.
577 if (region_desc
->tiled
) {
578 if (IS_965(intel
->intelScreen
->deviceID
) &&
579 region_desc
== &intelScreen
->depth
)
580 region
->tiling
= I915_TILING_Y
;
582 region
->tiling
= I915_TILING_X
;
584 region
->tiling
= I915_TILING_NONE
;
587 region
->bit_6_swizzle
= I915_BIT_6_SWIZZLE_NONE
;
590 assert(region
->buffer
!= NULL
);
596 * Create intel_region structs to describe the static front, back, and depth
597 * buffers created by the xserver.
599 * Although FBO's mean we now no longer use these as render targets in
600 * all circumstances, they won't go away until the back and depth
601 * buffers become private, and the front buffer will remain even then.
603 * Note that these don't allocate video memory, just describe
604 * allocations alread made by the X server.
607 intel_recreate_static_regions(struct intel_context
*intel
)
609 intelScreenPrivate
*intelScreen
= intel
->intelScreen
;
611 intel
->front_region
=
612 intel_recreate_static(intel
, "front",
614 &intelScreen
->front
);
617 intel_recreate_static(intel
, "back",
621 /* Still assumes front.cpp == depth.cpp. We can kill this when we move to
624 intel
->depth_region
=
625 intel_recreate_static(intel
, "depth",
627 &intelScreen
->depth
);