Revert "i965: prevent potentially null pointer access"
[mesa.git] / src / mesa / drivers / dri / i965 / intel_screen.c
1 /*
2 * Copyright 2003 VMware, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sublicense, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the
14 * next paragraph) shall be included in all copies or substantial portions
15 * of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
19 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
20 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
21 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
22 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
23 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
24 */
25
26 #include <drm_fourcc.h>
27 #include <errno.h>
28 #include <time.h>
29 #include <unistd.h>
30 #include "main/context.h"
31 #include "main/framebuffer.h"
32 #include "main/renderbuffer.h"
33 #include "main/texobj.h"
34 #include "main/hash.h"
35 #include "main/fbobject.h"
36 #include "main/version.h"
37 #include "swrast/s_renderbuffer.h"
38 #include "util/ralloc.h"
39 #include "util/disk_cache.h"
40 #include "brw_defines.h"
41 #include "brw_state.h"
42 #include "compiler/nir/nir.h"
43
44 #include "utils.h"
45 #include "util/disk_cache.h"
46 #include "util/xmlpool.h"
47
48 static const __DRIconfigOptionsExtension brw_config_options = {
49 .base = { __DRI_CONFIG_OPTIONS, 1 },
50 .xml =
51 DRI_CONF_BEGIN
52 DRI_CONF_SECTION_PERFORMANCE
53 /* Options correspond to DRI_CONF_BO_REUSE_DISABLED,
54 * DRI_CONF_BO_REUSE_ALL
55 */
56 DRI_CONF_OPT_BEGIN_V(bo_reuse, enum, 1, "0:1")
57 DRI_CONF_DESC_BEGIN(en, "Buffer object reuse")
58 DRI_CONF_ENUM(0, "Disable buffer object reuse")
59 DRI_CONF_ENUM(1, "Enable reuse of all sizes of buffer objects")
60 DRI_CONF_DESC_END
61 DRI_CONF_OPT_END
62 DRI_CONF_MESA_NO_ERROR("false")
63 DRI_CONF_SECTION_END
64
65 DRI_CONF_SECTION_QUALITY
66 DRI_CONF_PRECISE_TRIG("false")
67
68 DRI_CONF_OPT_BEGIN(clamp_max_samples, int, -1)
69 DRI_CONF_DESC(en, "Clamp the value of GL_MAX_SAMPLES to the "
70 "given integer. If negative, then do not clamp.")
71 DRI_CONF_OPT_END
72 DRI_CONF_SECTION_END
73
74 DRI_CONF_SECTION_DEBUG
75 DRI_CONF_NO_RAST("false")
76 DRI_CONF_ALWAYS_FLUSH_BATCH("false")
77 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
78 DRI_CONF_DISABLE_THROTTLING("false")
79 DRI_CONF_FORCE_GLSL_EXTENSIONS_WARN("false")
80 DRI_CONF_FORCE_GLSL_VERSION(0)
81 DRI_CONF_DISABLE_GLSL_LINE_CONTINUATIONS("false")
82 DRI_CONF_DISABLE_BLEND_FUNC_EXTENDED("false")
83 DRI_CONF_DUAL_COLOR_BLEND_BY_LOCATION("false")
84 DRI_CONF_ALLOW_GLSL_EXTENSION_DIRECTIVE_MIDSHADER("false")
85 DRI_CONF_ALLOW_GLSL_BUILTIN_VARIABLE_REDECLARATION("false")
86 DRI_CONF_ALLOW_GLSL_CROSS_STAGE_INTERPOLATION_MISMATCH("false")
87 DRI_CONF_ALLOW_HIGHER_COMPAT_VERSION("false")
88 DRI_CONF_FORCE_GLSL_ABS_SQRT("false")
89
90 DRI_CONF_OPT_BEGIN_B(shader_precompile, "true")
91 DRI_CONF_DESC(en, "Perform code generation at shader link time.")
92 DRI_CONF_OPT_END
93 DRI_CONF_SECTION_END
94
95 DRI_CONF_SECTION_MISCELLANEOUS
96 DRI_CONF_GLSL_ZERO_INIT("false")
97 DRI_CONF_ALLOW_RGB10_CONFIGS("false")
98 DRI_CONF_SECTION_END
99 DRI_CONF_END
100 };
101
102 #include "intel_batchbuffer.h"
103 #include "intel_buffers.h"
104 #include "brw_bufmgr.h"
105 #include "intel_fbo.h"
106 #include "intel_mipmap_tree.h"
107 #include "intel_screen.h"
108 #include "intel_tex.h"
109 #include "intel_image.h"
110
111 #include "brw_context.h"
112
113 #include "i915_drm.h"
114
115 /**
116 * For debugging purposes, this returns a time in seconds.
117 */
118 double
119 get_time(void)
120 {
121 struct timespec tp;
122
123 clock_gettime(CLOCK_MONOTONIC, &tp);
124
125 return tp.tv_sec + tp.tv_nsec / 1000000000.0;
126 }
127
128 static const __DRItexBufferExtension intelTexBufferExtension = {
129 .base = { __DRI_TEX_BUFFER, 3 },
130
131 .setTexBuffer = intelSetTexBuffer,
132 .setTexBuffer2 = intelSetTexBuffer2,
133 .releaseTexBuffer = NULL,
134 };
135
136 static void
137 intel_dri2_flush_with_flags(__DRIcontext *cPriv,
138 __DRIdrawable *dPriv,
139 unsigned flags,
140 enum __DRI2throttleReason reason)
141 {
142 struct brw_context *brw = cPriv->driverPrivate;
143
144 if (!brw)
145 return;
146
147 struct gl_context *ctx = &brw->ctx;
148
149 FLUSH_VERTICES(ctx, 0);
150
151 if (flags & __DRI2_FLUSH_DRAWABLE)
152 intel_resolve_for_dri2_flush(brw, dPriv);
153
154 if (reason == __DRI2_THROTTLE_SWAPBUFFER)
155 brw->need_swap_throttle = true;
156 if (reason == __DRI2_THROTTLE_FLUSHFRONT)
157 brw->need_flush_throttle = true;
158
159 intel_batchbuffer_flush(brw);
160 }
161
162 /**
163 * Provides compatibility with loaders that only support the older (version
164 * 1-3) flush interface.
165 *
166 * That includes libGL up to Mesa 9.0, and the X Server at least up to 1.13.
167 */
168 static void
169 intel_dri2_flush(__DRIdrawable *drawable)
170 {
171 intel_dri2_flush_with_flags(drawable->driContextPriv, drawable,
172 __DRI2_FLUSH_DRAWABLE,
173 __DRI2_THROTTLE_SWAPBUFFER);
174 }
175
176 static const struct __DRI2flushExtensionRec intelFlushExtension = {
177 .base = { __DRI2_FLUSH, 4 },
178
179 .flush = intel_dri2_flush,
180 .invalidate = dri2InvalidateDrawable,
181 .flush_with_flags = intel_dri2_flush_with_flags,
182 };
183
184 static const struct intel_image_format intel_image_formats[] = {
185 { __DRI_IMAGE_FOURCC_ARGB2101010, __DRI_IMAGE_COMPONENTS_RGBA, 1,
186 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ARGB2101010, 4 } } },
187
188 { __DRI_IMAGE_FOURCC_XRGB2101010, __DRI_IMAGE_COMPONENTS_RGB, 1,
189 { { 0, 0, 0, __DRI_IMAGE_FORMAT_XRGB2101010, 4 } } },
190
191 { __DRI_IMAGE_FOURCC_ARGB8888, __DRI_IMAGE_COMPONENTS_RGBA, 1,
192 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ARGB8888, 4 } } },
193
194 { __DRI_IMAGE_FOURCC_ABGR8888, __DRI_IMAGE_COMPONENTS_RGBA, 1,
195 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ABGR8888, 4 } } },
196
197 { __DRI_IMAGE_FOURCC_SARGB8888, __DRI_IMAGE_COMPONENTS_RGBA, 1,
198 { { 0, 0, 0, __DRI_IMAGE_FORMAT_SARGB8, 4 } } },
199
200 { __DRI_IMAGE_FOURCC_XRGB8888, __DRI_IMAGE_COMPONENTS_RGB, 1,
201 { { 0, 0, 0, __DRI_IMAGE_FORMAT_XRGB8888, 4 }, } },
202
203 { __DRI_IMAGE_FOURCC_XBGR8888, __DRI_IMAGE_COMPONENTS_RGB, 1,
204 { { 0, 0, 0, __DRI_IMAGE_FORMAT_XBGR8888, 4 }, } },
205
206 { __DRI_IMAGE_FOURCC_ARGB1555, __DRI_IMAGE_COMPONENTS_RGBA, 1,
207 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ARGB1555, 2 } } },
208
209 { __DRI_IMAGE_FOURCC_RGB565, __DRI_IMAGE_COMPONENTS_RGB, 1,
210 { { 0, 0, 0, __DRI_IMAGE_FORMAT_RGB565, 2 } } },
211
212 { __DRI_IMAGE_FOURCC_R8, __DRI_IMAGE_COMPONENTS_R, 1,
213 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 }, } },
214
215 { __DRI_IMAGE_FOURCC_R16, __DRI_IMAGE_COMPONENTS_R, 1,
216 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R16, 1 }, } },
217
218 { __DRI_IMAGE_FOURCC_GR88, __DRI_IMAGE_COMPONENTS_RG, 1,
219 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR88, 2 }, } },
220
221 { __DRI_IMAGE_FOURCC_GR1616, __DRI_IMAGE_COMPONENTS_RG, 1,
222 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR1616, 2 }, } },
223
224 { __DRI_IMAGE_FOURCC_YUV410, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
225 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
226 { 1, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 },
227 { 2, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 } } },
228
229 { __DRI_IMAGE_FOURCC_YUV411, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
230 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
231 { 1, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 },
232 { 2, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
233
234 { __DRI_IMAGE_FOURCC_YUV420, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
235 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
236 { 1, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 },
237 { 2, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 } } },
238
239 { __DRI_IMAGE_FOURCC_YUV422, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
240 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
241 { 1, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 },
242 { 2, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
243
244 { __DRI_IMAGE_FOURCC_YUV444, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
245 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
246 { 1, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
247 { 2, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
248
249 { __DRI_IMAGE_FOURCC_YVU410, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
250 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
251 { 2, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 },
252 { 1, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 } } },
253
254 { __DRI_IMAGE_FOURCC_YVU411, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
255 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
256 { 2, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 },
257 { 1, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
258
259 { __DRI_IMAGE_FOURCC_YVU420, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
260 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
261 { 2, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 },
262 { 1, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 } } },
263
264 { __DRI_IMAGE_FOURCC_YVU422, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
265 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
266 { 2, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 },
267 { 1, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
268
269 { __DRI_IMAGE_FOURCC_YVU444, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
270 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
271 { 2, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
272 { 1, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
273
274 { __DRI_IMAGE_FOURCC_NV12, __DRI_IMAGE_COMPONENTS_Y_UV, 2,
275 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
276 { 1, 1, 1, __DRI_IMAGE_FORMAT_GR88, 2 } } },
277
278 { __DRI_IMAGE_FOURCC_NV16, __DRI_IMAGE_COMPONENTS_Y_UV, 2,
279 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
280 { 1, 1, 0, __DRI_IMAGE_FORMAT_GR88, 2 } } },
281
282 /* For YUYV and UYVY buffers, we set up two overlapping DRI images
283 * and treat them as planar buffers in the compositors.
284 * Plane 0 is GR88 and samples YU or YV pairs and places Y into
285 * the R component, while plane 1 is ARGB/ABGR and samples YUYV/UYVY
286 * clusters and places pairs and places U into the G component and
287 * V into A. This lets the texture sampler interpolate the Y
288 * components correctly when sampling from plane 0, and interpolate
289 * U and V correctly when sampling from plane 1. */
290 { __DRI_IMAGE_FOURCC_YUYV, __DRI_IMAGE_COMPONENTS_Y_XUXV, 2,
291 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR88, 2 },
292 { 0, 1, 0, __DRI_IMAGE_FORMAT_ARGB8888, 4 } } },
293 { __DRI_IMAGE_FOURCC_UYVY, __DRI_IMAGE_COMPONENTS_Y_UXVX, 2,
294 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR88, 2 },
295 { 0, 1, 0, __DRI_IMAGE_FORMAT_ABGR8888, 4 } } }
296 };
297
298 static const struct {
299 uint64_t modifier;
300 unsigned since_gen;
301 } supported_modifiers[] = {
302 { .modifier = DRM_FORMAT_MOD_LINEAR , .since_gen = 1 },
303 { .modifier = I915_FORMAT_MOD_X_TILED , .since_gen = 1 },
304 { .modifier = I915_FORMAT_MOD_Y_TILED , .since_gen = 6 },
305 { .modifier = I915_FORMAT_MOD_Y_TILED_CCS , .since_gen = 9 },
306 };
307
308 static bool
309 modifier_is_supported(const struct gen_device_info *devinfo,
310 const struct intel_image_format *fmt, int dri_format,
311 uint64_t modifier)
312 {
313 const struct isl_drm_modifier_info *modinfo =
314 isl_drm_modifier_get_info(modifier);
315 int i;
316
317 /* ISL had better know about the modifier */
318 if (!modinfo)
319 return false;
320
321 if (modinfo->aux_usage == ISL_AUX_USAGE_CCS_E) {
322 /* If INTEL_DEBUG=norbc is set, don't support any CCS_E modifiers */
323 if (unlikely(INTEL_DEBUG & DEBUG_NO_RBC))
324 return false;
325
326 /* CCS_E is not supported for planar images */
327 if (fmt && fmt->nplanes > 1)
328 return false;
329
330 if (fmt) {
331 assert(dri_format == 0);
332 dri_format = fmt->planes[0].dri_format;
333 }
334
335 mesa_format format = driImageFormatToGLFormat(dri_format);
336 format = _mesa_get_srgb_format_linear(format);
337 if (!isl_format_supports_ccs_e(devinfo,
338 brw_isl_format_for_mesa_format(format)))
339 return false;
340 }
341
342 for (i = 0; i < ARRAY_SIZE(supported_modifiers); i++) {
343 if (supported_modifiers[i].modifier != modifier)
344 continue;
345
346 return supported_modifiers[i].since_gen <= devinfo->gen;
347 }
348
349 return false;
350 }
351
352 static uint64_t
353 tiling_to_modifier(uint32_t tiling)
354 {
355 static const uint64_t map[] = {
356 [I915_TILING_NONE] = DRM_FORMAT_MOD_LINEAR,
357 [I915_TILING_X] = I915_FORMAT_MOD_X_TILED,
358 [I915_TILING_Y] = I915_FORMAT_MOD_Y_TILED,
359 };
360
361 assert(tiling < ARRAY_SIZE(map));
362
363 return map[tiling];
364 }
365
366 static void
367 intel_image_warn_if_unaligned(__DRIimage *image, const char *func)
368 {
369 uint32_t tiling, swizzle;
370 brw_bo_get_tiling(image->bo, &tiling, &swizzle);
371
372 if (tiling != I915_TILING_NONE && (image->offset & 0xfff)) {
373 _mesa_warning(NULL, "%s: offset 0x%08x not on tile boundary",
374 func, image->offset);
375 }
376 }
377
378 static const struct intel_image_format *
379 intel_image_format_lookup(int fourcc)
380 {
381 for (unsigned i = 0; i < ARRAY_SIZE(intel_image_formats); i++) {
382 if (intel_image_formats[i].fourcc == fourcc)
383 return &intel_image_formats[i];
384 }
385
386 return NULL;
387 }
388
389 static boolean intel_lookup_fourcc(int dri_format, int *fourcc)
390 {
391 for (unsigned i = 0; i < ARRAY_SIZE(intel_image_formats); i++) {
392 if (intel_image_formats[i].planes[0].dri_format == dri_format) {
393 *fourcc = intel_image_formats[i].fourcc;
394 return true;
395 }
396 }
397 return false;
398 }
399
400 static __DRIimage *
401 intel_allocate_image(struct intel_screen *screen, int dri_format,
402 void *loaderPrivate)
403 {
404 __DRIimage *image;
405
406 image = calloc(1, sizeof *image);
407 if (image == NULL)
408 return NULL;
409
410 image->screen = screen;
411 image->dri_format = dri_format;
412 image->offset = 0;
413
414 image->format = driImageFormatToGLFormat(dri_format);
415 if (dri_format != __DRI_IMAGE_FORMAT_NONE &&
416 image->format == MESA_FORMAT_NONE) {
417 free(image);
418 return NULL;
419 }
420
421 image->internal_format = _mesa_get_format_base_format(image->format);
422 image->data = loaderPrivate;
423
424 return image;
425 }
426
427 /**
428 * Sets up a DRIImage structure to point to a slice out of a miptree.
429 */
430 static void
431 intel_setup_image_from_mipmap_tree(struct brw_context *brw, __DRIimage *image,
432 struct intel_mipmap_tree *mt, GLuint level,
433 GLuint zoffset)
434 {
435 intel_miptree_make_shareable(brw, mt);
436
437 intel_miptree_check_level_layer(mt, level, zoffset);
438
439 image->width = minify(mt->surf.phys_level0_sa.width,
440 level - mt->first_level);
441 image->height = minify(mt->surf.phys_level0_sa.height,
442 level - mt->first_level);
443 image->pitch = mt->surf.row_pitch;
444
445 image->offset = intel_miptree_get_tile_offsets(mt, level, zoffset,
446 &image->tile_x,
447 &image->tile_y);
448
449 brw_bo_unreference(image->bo);
450 image->bo = mt->bo;
451 brw_bo_reference(mt->bo);
452 }
453
454 static __DRIimage *
455 intel_create_image_from_name(__DRIscreen *dri_screen,
456 int width, int height, int format,
457 int name, int pitch, void *loaderPrivate)
458 {
459 struct intel_screen *screen = dri_screen->driverPrivate;
460 __DRIimage *image;
461 int cpp;
462
463 image = intel_allocate_image(screen, format, loaderPrivate);
464 if (image == NULL)
465 return NULL;
466
467 if (image->format == MESA_FORMAT_NONE)
468 cpp = 1;
469 else
470 cpp = _mesa_get_format_bytes(image->format);
471
472 image->width = width;
473 image->height = height;
474 image->pitch = pitch * cpp;
475 image->bo = brw_bo_gem_create_from_name(screen->bufmgr, "image",
476 name);
477 if (!image->bo) {
478 free(image);
479 return NULL;
480 }
481 image->modifier = tiling_to_modifier(image->bo->tiling_mode);
482
483 return image;
484 }
485
486 static __DRIimage *
487 intel_create_image_from_renderbuffer(__DRIcontext *context,
488 int renderbuffer, void *loaderPrivate)
489 {
490 __DRIimage *image;
491 struct brw_context *brw = context->driverPrivate;
492 struct gl_context *ctx = &brw->ctx;
493 struct gl_renderbuffer *rb;
494 struct intel_renderbuffer *irb;
495
496 rb = _mesa_lookup_renderbuffer(ctx, renderbuffer);
497 if (!rb) {
498 _mesa_error(ctx, GL_INVALID_OPERATION, "glRenderbufferExternalMESA");
499 return NULL;
500 }
501
502 irb = intel_renderbuffer(rb);
503 intel_miptree_make_shareable(brw, irb->mt);
504 image = calloc(1, sizeof *image);
505 if (image == NULL)
506 return NULL;
507
508 image->internal_format = rb->InternalFormat;
509 image->format = rb->Format;
510 image->modifier = tiling_to_modifier(
511 isl_tiling_to_i915_tiling(irb->mt->surf.tiling));
512 image->offset = 0;
513 image->data = loaderPrivate;
514 brw_bo_unreference(image->bo);
515 image->bo = irb->mt->bo;
516 brw_bo_reference(irb->mt->bo);
517 image->width = rb->Width;
518 image->height = rb->Height;
519 image->pitch = irb->mt->surf.row_pitch;
520 image->dri_format = driGLFormatToImageFormat(image->format);
521 image->has_depthstencil = irb->mt->stencil_mt? true : false;
522
523 rb->NeedsFinishRenderTexture = true;
524 return image;
525 }
526
527 static __DRIimage *
528 intel_create_image_from_texture(__DRIcontext *context, int target,
529 unsigned texture, int zoffset,
530 int level,
531 unsigned *error,
532 void *loaderPrivate)
533 {
534 __DRIimage *image;
535 struct brw_context *brw = context->driverPrivate;
536 struct gl_texture_object *obj;
537 struct intel_texture_object *iobj;
538 GLuint face = 0;
539
540 obj = _mesa_lookup_texture(&brw->ctx, texture);
541 if (!obj || obj->Target != target) {
542 *error = __DRI_IMAGE_ERROR_BAD_PARAMETER;
543 return NULL;
544 }
545
546 if (target == GL_TEXTURE_CUBE_MAP)
547 face = zoffset;
548
549 _mesa_test_texobj_completeness(&brw->ctx, obj);
550 iobj = intel_texture_object(obj);
551 if (!obj->_BaseComplete || (level > 0 && !obj->_MipmapComplete)) {
552 *error = __DRI_IMAGE_ERROR_BAD_PARAMETER;
553 return NULL;
554 }
555
556 if (level < obj->BaseLevel || level > obj->_MaxLevel) {
557 *error = __DRI_IMAGE_ERROR_BAD_MATCH;
558 return NULL;
559 }
560
561 if (target == GL_TEXTURE_3D && obj->Image[face][level]->Depth < zoffset) {
562 *error = __DRI_IMAGE_ERROR_BAD_MATCH;
563 return NULL;
564 }
565 image = calloc(1, sizeof *image);
566 if (image == NULL) {
567 *error = __DRI_IMAGE_ERROR_BAD_ALLOC;
568 return NULL;
569 }
570
571 image->internal_format = obj->Image[face][level]->InternalFormat;
572 image->format = obj->Image[face][level]->TexFormat;
573 image->modifier = tiling_to_modifier(
574 isl_tiling_to_i915_tiling(iobj->mt->surf.tiling));
575 image->data = loaderPrivate;
576 intel_setup_image_from_mipmap_tree(brw, image, iobj->mt, level, zoffset);
577 image->dri_format = driGLFormatToImageFormat(image->format);
578 image->has_depthstencil = iobj->mt->stencil_mt? true : false;
579 if (image->dri_format == MESA_FORMAT_NONE) {
580 *error = __DRI_IMAGE_ERROR_BAD_PARAMETER;
581 free(image);
582 return NULL;
583 }
584
585 *error = __DRI_IMAGE_ERROR_SUCCESS;
586 return image;
587 }
588
589 static void
590 intel_destroy_image(__DRIimage *image)
591 {
592 brw_bo_unreference(image->bo);
593 free(image);
594 }
595
596 enum modifier_priority {
597 MODIFIER_PRIORITY_INVALID = 0,
598 MODIFIER_PRIORITY_LINEAR,
599 MODIFIER_PRIORITY_X,
600 MODIFIER_PRIORITY_Y,
601 MODIFIER_PRIORITY_Y_CCS,
602 };
603
604 const uint64_t priority_to_modifier[] = {
605 [MODIFIER_PRIORITY_INVALID] = DRM_FORMAT_MOD_INVALID,
606 [MODIFIER_PRIORITY_LINEAR] = DRM_FORMAT_MOD_LINEAR,
607 [MODIFIER_PRIORITY_X] = I915_FORMAT_MOD_X_TILED,
608 [MODIFIER_PRIORITY_Y] = I915_FORMAT_MOD_Y_TILED,
609 [MODIFIER_PRIORITY_Y_CCS] = I915_FORMAT_MOD_Y_TILED_CCS,
610 };
611
612 static uint64_t
613 select_best_modifier(struct gen_device_info *devinfo,
614 int dri_format,
615 const uint64_t *modifiers,
616 const unsigned count)
617 {
618 enum modifier_priority prio = MODIFIER_PRIORITY_INVALID;
619
620 for (int i = 0; i < count; i++) {
621 if (!modifier_is_supported(devinfo, NULL, dri_format, modifiers[i]))
622 continue;
623
624 switch (modifiers[i]) {
625 case I915_FORMAT_MOD_Y_TILED_CCS:
626 prio = MAX2(prio, MODIFIER_PRIORITY_Y_CCS);
627 break;
628 case I915_FORMAT_MOD_Y_TILED:
629 prio = MAX2(prio, MODIFIER_PRIORITY_Y);
630 break;
631 case I915_FORMAT_MOD_X_TILED:
632 prio = MAX2(prio, MODIFIER_PRIORITY_X);
633 break;
634 case DRM_FORMAT_MOD_LINEAR:
635 prio = MAX2(prio, MODIFIER_PRIORITY_LINEAR);
636 break;
637 case DRM_FORMAT_MOD_INVALID:
638 default:
639 break;
640 }
641 }
642
643 return priority_to_modifier[prio];
644 }
645
646 static __DRIimage *
647 intel_create_image_common(__DRIscreen *dri_screen,
648 int width, int height, int format,
649 unsigned int use,
650 const uint64_t *modifiers,
651 unsigned count,
652 void *loaderPrivate)
653 {
654 __DRIimage *image;
655 struct intel_screen *screen = dri_screen->driverPrivate;
656 uint64_t modifier = DRM_FORMAT_MOD_INVALID;
657 bool ok;
658
659 /* Callers of this may specify a modifier, or a dri usage, but not both. The
660 * newer modifier interface deprecates the older usage flags newer modifier
661 * interface deprecates the older usage flags.
662 */
663 assert(!(use && count));
664
665 if (use & __DRI_IMAGE_USE_CURSOR) {
666 if (width != 64 || height != 64)
667 return NULL;
668 modifier = DRM_FORMAT_MOD_LINEAR;
669 }
670
671 if (use & __DRI_IMAGE_USE_LINEAR)
672 modifier = DRM_FORMAT_MOD_LINEAR;
673
674 if (modifier == DRM_FORMAT_MOD_INVALID) {
675 if (modifiers) {
676 /* User requested specific modifiers */
677 modifier = select_best_modifier(&screen->devinfo, format,
678 modifiers, count);
679 if (modifier == DRM_FORMAT_MOD_INVALID)
680 return NULL;
681 } else {
682 /* Historically, X-tiled was the default, and so lack of modifier means
683 * X-tiled.
684 */
685 modifier = I915_FORMAT_MOD_X_TILED;
686 }
687 }
688
689 image = intel_allocate_image(screen, format, loaderPrivate);
690 if (image == NULL)
691 return NULL;
692
693 const struct isl_drm_modifier_info *mod_info =
694 isl_drm_modifier_get_info(modifier);
695
696 struct isl_surf surf;
697 ok = isl_surf_init(&screen->isl_dev, &surf,
698 .dim = ISL_SURF_DIM_2D,
699 .format = brw_isl_format_for_mesa_format(image->format),
700 .width = width,
701 .height = height,
702 .depth = 1,
703 .levels = 1,
704 .array_len = 1,
705 .samples = 1,
706 .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT |
707 ISL_SURF_USAGE_TEXTURE_BIT |
708 ISL_SURF_USAGE_STORAGE_BIT,
709 .tiling_flags = (1 << mod_info->tiling));
710 assert(ok);
711 if (!ok) {
712 free(image);
713 return NULL;
714 }
715
716 struct isl_surf aux_surf;
717 if (mod_info->aux_usage == ISL_AUX_USAGE_CCS_E) {
718 ok = isl_surf_get_ccs_surf(&screen->isl_dev, &surf, &aux_surf, 0);
719 if (!ok) {
720 free(image);
721 return NULL;
722 }
723 } else {
724 assert(mod_info->aux_usage == ISL_AUX_USAGE_NONE);
725 aux_surf.size = 0;
726 }
727
728 /* We request that the bufmgr zero the buffer for us for two reasons:
729 *
730 * 1) If a buffer gets re-used from the pool, we don't want to leak random
731 * garbage from our process to some other.
732 *
733 * 2) For images with CCS_E, we want to ensure that the CCS starts off in
734 * a valid state. A CCS value of 0 indicates that the given block is
735 * in the pass-through state which is what we want.
736 */
737 image->bo = brw_bo_alloc_tiled(screen->bufmgr, "image",
738 surf.size + aux_surf.size,
739 isl_tiling_to_i915_tiling(mod_info->tiling),
740 surf.row_pitch, BO_ALLOC_ZEROED);
741 if (image->bo == NULL) {
742 free(image);
743 return NULL;
744 }
745 image->width = width;
746 image->height = height;
747 image->pitch = surf.row_pitch;
748 image->modifier = modifier;
749
750 if (aux_surf.size) {
751 image->aux_offset = surf.size;
752 image->aux_pitch = aux_surf.row_pitch;
753 }
754
755 return image;
756 }
757
758 static __DRIimage *
759 intel_create_image(__DRIscreen *dri_screen,
760 int width, int height, int format,
761 unsigned int use,
762 void *loaderPrivate)
763 {
764 return intel_create_image_common(dri_screen, width, height, format, use, NULL, 0,
765 loaderPrivate);
766 }
767
768 static void *
769 intel_map_image(__DRIcontext *context, __DRIimage *image,
770 int x0, int y0, int width, int height,
771 unsigned int flags, int *stride, void **map_info)
772 {
773 struct brw_context *brw = NULL;
774 struct brw_bo *bo = NULL;
775 void *raw_data = NULL;
776 GLuint pix_w = 1;
777 GLuint pix_h = 1;
778 GLint pix_bytes = 1;
779
780 if (!context || !image || !stride || !map_info || *map_info)
781 return NULL;
782
783 if (x0 < 0 || x0 >= image->width || width > image->width - x0)
784 return NULL;
785
786 if (y0 < 0 || y0 >= image->height || height > image->height - y0)
787 return NULL;
788
789 if (flags & MAP_INTERNAL_MASK)
790 return NULL;
791
792 brw = context->driverPrivate;
793 bo = image->bo;
794
795 assert(brw);
796 assert(bo);
797
798 /* DRI flags and GL_MAP.*_BIT flags are the same, so just pass them on. */
799 raw_data = brw_bo_map(brw, bo, flags);
800 if (!raw_data)
801 return NULL;
802
803 _mesa_get_format_block_size(image->format, &pix_w, &pix_h);
804 pix_bytes = _mesa_get_format_bytes(image->format);
805
806 assert(pix_w);
807 assert(pix_h);
808 assert(pix_bytes > 0);
809
810 raw_data += (x0 / pix_w) * pix_bytes + (y0 / pix_h) * image->pitch;
811
812 brw_bo_reference(bo);
813
814 *stride = image->pitch;
815 *map_info = bo;
816
817 return raw_data;
818 }
819
820 static void
821 intel_unmap_image(__DRIcontext *context, __DRIimage *image, void *map_info)
822 {
823 struct brw_bo *bo = map_info;
824
825 brw_bo_unmap(bo);
826 brw_bo_unreference(bo);
827 }
828
829 static __DRIimage *
830 intel_create_image_with_modifiers(__DRIscreen *dri_screen,
831 int width, int height, int format,
832 const uint64_t *modifiers,
833 const unsigned count,
834 void *loaderPrivate)
835 {
836 return intel_create_image_common(dri_screen, width, height, format, 0,
837 modifiers, count, loaderPrivate);
838 }
839
840 static GLboolean
841 intel_query_image(__DRIimage *image, int attrib, int *value)
842 {
843 switch (attrib) {
844 case __DRI_IMAGE_ATTRIB_STRIDE:
845 *value = image->pitch;
846 return true;
847 case __DRI_IMAGE_ATTRIB_HANDLE:
848 *value = brw_bo_export_gem_handle(image->bo);
849 return true;
850 case __DRI_IMAGE_ATTRIB_NAME:
851 return !brw_bo_flink(image->bo, (uint32_t *) value);
852 case __DRI_IMAGE_ATTRIB_FORMAT:
853 *value = image->dri_format;
854 return true;
855 case __DRI_IMAGE_ATTRIB_WIDTH:
856 *value = image->width;
857 return true;
858 case __DRI_IMAGE_ATTRIB_HEIGHT:
859 *value = image->height;
860 return true;
861 case __DRI_IMAGE_ATTRIB_COMPONENTS:
862 if (image->planar_format == NULL)
863 return false;
864 *value = image->planar_format->components;
865 return true;
866 case __DRI_IMAGE_ATTRIB_FD:
867 return !brw_bo_gem_export_to_prime(image->bo, value);
868 case __DRI_IMAGE_ATTRIB_FOURCC:
869 return intel_lookup_fourcc(image->dri_format, value);
870 case __DRI_IMAGE_ATTRIB_NUM_PLANES:
871 if (isl_drm_modifier_has_aux(image->modifier)) {
872 assert(!image->planar_format || image->planar_format->nplanes == 1);
873 *value = 2;
874 } else if (image->planar_format) {
875 *value = image->planar_format->nplanes;
876 } else {
877 *value = 1;
878 }
879 return true;
880 case __DRI_IMAGE_ATTRIB_OFFSET:
881 *value = image->offset;
882 return true;
883 case __DRI_IMAGE_ATTRIB_MODIFIER_LOWER:
884 *value = (image->modifier & 0xffffffff);
885 return true;
886 case __DRI_IMAGE_ATTRIB_MODIFIER_UPPER:
887 *value = ((image->modifier >> 32) & 0xffffffff);
888 return true;
889
890 default:
891 return false;
892 }
893 }
894
895 static GLboolean
896 intel_query_format_modifier_attribs(__DRIscreen *dri_screen,
897 uint32_t fourcc, uint64_t modifier,
898 int attrib, uint64_t *value)
899 {
900 struct intel_screen *screen = dri_screen->driverPrivate;
901 const struct intel_image_format *f = intel_image_format_lookup(fourcc);
902
903 if (!modifier_is_supported(&screen->devinfo, f, 0, modifier))
904 return false;
905
906 switch (attrib) {
907 case __DRI_IMAGE_FORMAT_MODIFIER_ATTRIB_PLANE_COUNT:
908 *value = isl_drm_modifier_has_aux(modifier) ? 2 : f->nplanes;
909 return true;
910
911 default:
912 return false;
913 }
914 }
915
916 static __DRIimage *
917 intel_dup_image(__DRIimage *orig_image, void *loaderPrivate)
918 {
919 __DRIimage *image;
920
921 image = calloc(1, sizeof *image);
922 if (image == NULL)
923 return NULL;
924
925 brw_bo_reference(orig_image->bo);
926 image->bo = orig_image->bo;
927 image->internal_format = orig_image->internal_format;
928 image->planar_format = orig_image->planar_format;
929 image->dri_format = orig_image->dri_format;
930 image->format = orig_image->format;
931 image->modifier = orig_image->modifier;
932 image->offset = orig_image->offset;
933 image->width = orig_image->width;
934 image->height = orig_image->height;
935 image->pitch = orig_image->pitch;
936 image->tile_x = orig_image->tile_x;
937 image->tile_y = orig_image->tile_y;
938 image->has_depthstencil = orig_image->has_depthstencil;
939 image->data = loaderPrivate;
940 image->dma_buf_imported = orig_image->dma_buf_imported;
941 image->aux_offset = orig_image->aux_offset;
942 image->aux_pitch = orig_image->aux_pitch;
943
944 memcpy(image->strides, orig_image->strides, sizeof(image->strides));
945 memcpy(image->offsets, orig_image->offsets, sizeof(image->offsets));
946
947 return image;
948 }
949
950 static GLboolean
951 intel_validate_usage(__DRIimage *image, unsigned int use)
952 {
953 if (use & __DRI_IMAGE_USE_CURSOR) {
954 if (image->width != 64 || image->height != 64)
955 return GL_FALSE;
956 }
957
958 return GL_TRUE;
959 }
960
961 static __DRIimage *
962 intel_create_image_from_names(__DRIscreen *dri_screen,
963 int width, int height, int fourcc,
964 int *names, int num_names,
965 int *strides, int *offsets,
966 void *loaderPrivate)
967 {
968 const struct intel_image_format *f = NULL;
969 __DRIimage *image;
970 int i, index;
971
972 if (dri_screen == NULL || names == NULL || num_names != 1)
973 return NULL;
974
975 f = intel_image_format_lookup(fourcc);
976 if (f == NULL)
977 return NULL;
978
979 image = intel_create_image_from_name(dri_screen, width, height,
980 __DRI_IMAGE_FORMAT_NONE,
981 names[0], strides[0],
982 loaderPrivate);
983
984 if (image == NULL)
985 return NULL;
986
987 image->planar_format = f;
988 for (i = 0; i < f->nplanes; i++) {
989 index = f->planes[i].buffer_index;
990 image->offsets[index] = offsets[index];
991 image->strides[index] = strides[index];
992 }
993
994 return image;
995 }
996
997 static __DRIimage *
998 intel_create_image_from_fds_common(__DRIscreen *dri_screen,
999 int width, int height, int fourcc,
1000 uint64_t modifier, int *fds, int num_fds,
1001 int *strides, int *offsets,
1002 void *loaderPrivate)
1003 {
1004 struct intel_screen *screen = dri_screen->driverPrivate;
1005 const struct intel_image_format *f;
1006 __DRIimage *image;
1007 int i, index;
1008 bool ok;
1009
1010 if (fds == NULL || num_fds < 1)
1011 return NULL;
1012
1013 f = intel_image_format_lookup(fourcc);
1014 if (f == NULL)
1015 return NULL;
1016
1017 if (modifier != DRM_FORMAT_MOD_INVALID &&
1018 !modifier_is_supported(&screen->devinfo, f, 0, modifier))
1019 return NULL;
1020
1021 if (f->nplanes == 1)
1022 image = intel_allocate_image(screen, f->planes[0].dri_format,
1023 loaderPrivate);
1024 else
1025 image = intel_allocate_image(screen, __DRI_IMAGE_FORMAT_NONE,
1026 loaderPrivate);
1027
1028 if (image == NULL)
1029 return NULL;
1030
1031 image->width = width;
1032 image->height = height;
1033 image->pitch = strides[0];
1034
1035 image->planar_format = f;
1036
1037 if (modifier != DRM_FORMAT_MOD_INVALID) {
1038 const struct isl_drm_modifier_info *mod_info =
1039 isl_drm_modifier_get_info(modifier);
1040 uint32_t tiling = isl_tiling_to_i915_tiling(mod_info->tiling);
1041 image->bo = brw_bo_gem_create_from_prime_tiled(screen->bufmgr, fds[0],
1042 tiling, strides[0]);
1043 } else {
1044 image->bo = brw_bo_gem_create_from_prime(screen->bufmgr, fds[0]);
1045 }
1046
1047 if (image->bo == NULL) {
1048 free(image);
1049 return NULL;
1050 }
1051
1052 /* We only support all planes from the same bo.
1053 * brw_bo_gem_create_from_prime() should return the same pointer for all
1054 * fds received here */
1055 for (i = 1; i < num_fds; i++) {
1056 struct brw_bo *aux = brw_bo_gem_create_from_prime(screen->bufmgr, fds[i]);
1057 brw_bo_unreference(aux);
1058 if (aux != image->bo) {
1059 brw_bo_unreference(image->bo);
1060 free(image);
1061 return NULL;
1062 }
1063 }
1064
1065 if (modifier != DRM_FORMAT_MOD_INVALID)
1066 image->modifier = modifier;
1067 else
1068 image->modifier = tiling_to_modifier(image->bo->tiling_mode);
1069
1070 const struct isl_drm_modifier_info *mod_info =
1071 isl_drm_modifier_get_info(image->modifier);
1072
1073 int size = 0;
1074 struct isl_surf surf;
1075 for (i = 0; i < f->nplanes; i++) {
1076 index = f->planes[i].buffer_index;
1077 image->offsets[index] = offsets[index];
1078 image->strides[index] = strides[index];
1079
1080 mesa_format format = driImageFormatToGLFormat(f->planes[i].dri_format);
1081
1082 ok = isl_surf_init(&screen->isl_dev, &surf,
1083 .dim = ISL_SURF_DIM_2D,
1084 .format = brw_isl_format_for_mesa_format(format),
1085 .width = image->width >> f->planes[i].width_shift,
1086 .height = image->height >> f->planes[i].height_shift,
1087 .depth = 1,
1088 .levels = 1,
1089 .array_len = 1,
1090 .samples = 1,
1091 .row_pitch = strides[index],
1092 .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT |
1093 ISL_SURF_USAGE_TEXTURE_BIT |
1094 ISL_SURF_USAGE_STORAGE_BIT,
1095 .tiling_flags = (1 << mod_info->tiling));
1096 if (!ok) {
1097 brw_bo_unreference(image->bo);
1098 free(image);
1099 return NULL;
1100 }
1101
1102 const int end = offsets[index] + surf.size;
1103 if (size < end)
1104 size = end;
1105 }
1106
1107 if (mod_info->aux_usage == ISL_AUX_USAGE_CCS_E) {
1108 /* Even though we initialize surf in the loop above, we know that
1109 * anything with CCS_E will have exactly one plane so surf is properly
1110 * initialized when we get here.
1111 */
1112 assert(f->nplanes == 1);
1113
1114 image->aux_offset = offsets[1];
1115 image->aux_pitch = strides[1];
1116
1117 /* Scanout hardware requires that the CCS be placed after the main
1118 * surface in memory. We consider any CCS that is placed any earlier in
1119 * memory to be invalid and reject it.
1120 *
1121 * At some point in the future, this restriction may be relaxed if the
1122 * hardware becomes less strict but we may need a new modifier for that.
1123 */
1124 assert(size > 0);
1125 if (image->aux_offset < size) {
1126 brw_bo_unreference(image->bo);
1127 free(image);
1128 return NULL;
1129 }
1130
1131 struct isl_surf aux_surf;
1132 ok = isl_surf_get_ccs_surf(&screen->isl_dev, &surf, &aux_surf,
1133 image->aux_pitch);
1134 if (!ok) {
1135 brw_bo_unreference(image->bo);
1136 free(image);
1137 return NULL;
1138 }
1139
1140 const int end = image->aux_offset + aux_surf.size;
1141 if (size < end)
1142 size = end;
1143 } else {
1144 assert(mod_info->aux_usage == ISL_AUX_USAGE_NONE);
1145 }
1146
1147 /* Check that the requested image actually fits within the BO. 'size'
1148 * is already relative to the offsets, so we don't need to add that. */
1149 if (image->bo->size == 0) {
1150 image->bo->size = size;
1151 } else if (size > image->bo->size) {
1152 brw_bo_unreference(image->bo);
1153 free(image);
1154 return NULL;
1155 }
1156
1157 if (f->nplanes == 1) {
1158 image->offset = image->offsets[0];
1159 intel_image_warn_if_unaligned(image, __func__);
1160 }
1161
1162 return image;
1163 }
1164
1165 static __DRIimage *
1166 intel_create_image_from_fds(__DRIscreen *dri_screen,
1167 int width, int height, int fourcc,
1168 int *fds, int num_fds, int *strides, int *offsets,
1169 void *loaderPrivate)
1170 {
1171 return intel_create_image_from_fds_common(dri_screen, width, height, fourcc,
1172 DRM_FORMAT_MOD_INVALID,
1173 fds, num_fds, strides, offsets,
1174 loaderPrivate);
1175 }
1176
1177 static __DRIimage *
1178 intel_create_image_from_dma_bufs2(__DRIscreen *dri_screen,
1179 int width, int height,
1180 int fourcc, uint64_t modifier,
1181 int *fds, int num_fds,
1182 int *strides, int *offsets,
1183 enum __DRIYUVColorSpace yuv_color_space,
1184 enum __DRISampleRange sample_range,
1185 enum __DRIChromaSiting horizontal_siting,
1186 enum __DRIChromaSiting vertical_siting,
1187 unsigned *error,
1188 void *loaderPrivate)
1189 {
1190 __DRIimage *image;
1191 const struct intel_image_format *f = intel_image_format_lookup(fourcc);
1192
1193 if (!f) {
1194 *error = __DRI_IMAGE_ERROR_BAD_MATCH;
1195 return NULL;
1196 }
1197
1198 image = intel_create_image_from_fds_common(dri_screen, width, height,
1199 fourcc, modifier,
1200 fds, num_fds, strides, offsets,
1201 loaderPrivate);
1202
1203 /*
1204 * Invalid parameters and any inconsistencies between are assumed to be
1205 * checked by the caller. Therefore besides unsupported formats one can fail
1206 * only in allocation.
1207 */
1208 if (!image) {
1209 *error = __DRI_IMAGE_ERROR_BAD_ALLOC;
1210 return NULL;
1211 }
1212
1213 image->dma_buf_imported = true;
1214 image->yuv_color_space = yuv_color_space;
1215 image->sample_range = sample_range;
1216 image->horizontal_siting = horizontal_siting;
1217 image->vertical_siting = vertical_siting;
1218
1219 *error = __DRI_IMAGE_ERROR_SUCCESS;
1220 return image;
1221 }
1222
1223 static __DRIimage *
1224 intel_create_image_from_dma_bufs(__DRIscreen *dri_screen,
1225 int width, int height, int fourcc,
1226 int *fds, int num_fds,
1227 int *strides, int *offsets,
1228 enum __DRIYUVColorSpace yuv_color_space,
1229 enum __DRISampleRange sample_range,
1230 enum __DRIChromaSiting horizontal_siting,
1231 enum __DRIChromaSiting vertical_siting,
1232 unsigned *error,
1233 void *loaderPrivate)
1234 {
1235 return intel_create_image_from_dma_bufs2(dri_screen, width, height,
1236 fourcc, DRM_FORMAT_MOD_INVALID,
1237 fds, num_fds, strides, offsets,
1238 yuv_color_space,
1239 sample_range,
1240 horizontal_siting,
1241 vertical_siting,
1242 error,
1243 loaderPrivate);
1244 }
1245
1246 static GLboolean
1247 intel_query_dma_buf_formats(__DRIscreen *screen, int max,
1248 int *formats, int *count)
1249 {
1250 int i, j = 0;
1251
1252 if (max == 0) {
1253 *count = ARRAY_SIZE(intel_image_formats) - 1; /* not SARGB */
1254 return true;
1255 }
1256
1257 for (i = 0; i < (ARRAY_SIZE(intel_image_formats)) && j < max; i++) {
1258 if (intel_image_formats[i].fourcc == __DRI_IMAGE_FOURCC_SARGB8888)
1259 continue;
1260 formats[j++] = intel_image_formats[i].fourcc;
1261 }
1262
1263 *count = j;
1264 return true;
1265 }
1266
1267 static GLboolean
1268 intel_query_dma_buf_modifiers(__DRIscreen *_screen, int fourcc, int max,
1269 uint64_t *modifiers,
1270 unsigned int *external_only,
1271 int *count)
1272 {
1273 struct intel_screen *screen = _screen->driverPrivate;
1274 const struct intel_image_format *f;
1275 int num_mods = 0, i;
1276
1277 f = intel_image_format_lookup(fourcc);
1278 if (f == NULL)
1279 return false;
1280
1281 for (i = 0; i < ARRAY_SIZE(supported_modifiers); i++) {
1282 uint64_t modifier = supported_modifiers[i].modifier;
1283 if (!modifier_is_supported(&screen->devinfo, f, 0, modifier))
1284 continue;
1285
1286 num_mods++;
1287 if (max == 0)
1288 continue;
1289
1290 modifiers[num_mods - 1] = modifier;
1291 if (num_mods >= max)
1292 break;
1293 }
1294
1295 if (external_only != NULL) {
1296 for (i = 0; i < num_mods && i < max; i++) {
1297 if (f->components == __DRI_IMAGE_COMPONENTS_Y_U_V ||
1298 f->components == __DRI_IMAGE_COMPONENTS_Y_UV ||
1299 f->components == __DRI_IMAGE_COMPONENTS_Y_XUXV) {
1300 external_only[i] = GL_TRUE;
1301 }
1302 else {
1303 external_only[i] = GL_FALSE;
1304 }
1305 }
1306 }
1307
1308 *count = num_mods;
1309 return true;
1310 }
1311
1312 static __DRIimage *
1313 intel_from_planar(__DRIimage *parent, int plane, void *loaderPrivate)
1314 {
1315 int width, height, offset, stride, dri_format, index;
1316 const struct intel_image_format *f;
1317 __DRIimage *image;
1318
1319 if (parent == NULL) {
1320 return NULL;
1321 } else if (parent->planar_format == NULL) {
1322 const bool is_aux =
1323 isl_drm_modifier_has_aux(parent->modifier) && plane == 1;
1324 if (!is_aux)
1325 return NULL;
1326
1327 width = parent->width;
1328 height = parent->height;
1329 dri_format = parent->dri_format;
1330 offset = parent->aux_offset;
1331 stride = parent->aux_pitch;
1332 } else {
1333 /* Planar formats don't support aux buffers/images */
1334 assert(!isl_drm_modifier_has_aux(parent->modifier));
1335 f = parent->planar_format;
1336
1337 if (plane >= f->nplanes)
1338 return NULL;
1339
1340 width = parent->width >> f->planes[plane].width_shift;
1341 height = parent->height >> f->planes[plane].height_shift;
1342 dri_format = f->planes[plane].dri_format;
1343 index = f->planes[plane].buffer_index;
1344 offset = parent->offsets[index];
1345 stride = parent->strides[index];
1346
1347 if (offset + height * stride > parent->bo->size) {
1348 _mesa_warning(NULL, "intel_create_sub_image: subimage out of bounds");
1349 return NULL;
1350 }
1351 }
1352
1353 image = intel_allocate_image(parent->screen, dri_format, loaderPrivate);
1354 if (image == NULL)
1355 return NULL;
1356
1357 image->bo = parent->bo;
1358 brw_bo_reference(parent->bo);
1359 image->modifier = parent->modifier;
1360
1361 image->width = width;
1362 image->height = height;
1363 image->pitch = stride;
1364 image->offset = offset;
1365
1366 intel_image_warn_if_unaligned(image, __func__);
1367
1368 return image;
1369 }
1370
1371 static const __DRIimageExtension intelImageExtension = {
1372 .base = { __DRI_IMAGE, 16 },
1373
1374 .createImageFromName = intel_create_image_from_name,
1375 .createImageFromRenderbuffer = intel_create_image_from_renderbuffer,
1376 .destroyImage = intel_destroy_image,
1377 .createImage = intel_create_image,
1378 .queryImage = intel_query_image,
1379 .dupImage = intel_dup_image,
1380 .validateUsage = intel_validate_usage,
1381 .createImageFromNames = intel_create_image_from_names,
1382 .fromPlanar = intel_from_planar,
1383 .createImageFromTexture = intel_create_image_from_texture,
1384 .createImageFromFds = intel_create_image_from_fds,
1385 .createImageFromDmaBufs = intel_create_image_from_dma_bufs,
1386 .blitImage = NULL,
1387 .getCapabilities = NULL,
1388 .mapImage = intel_map_image,
1389 .unmapImage = intel_unmap_image,
1390 .createImageWithModifiers = intel_create_image_with_modifiers,
1391 .createImageFromDmaBufs2 = intel_create_image_from_dma_bufs2,
1392 .queryDmaBufFormats = intel_query_dma_buf_formats,
1393 .queryDmaBufModifiers = intel_query_dma_buf_modifiers,
1394 .queryDmaBufFormatModifierAttribs = intel_query_format_modifier_attribs,
1395 };
1396
1397 static uint64_t
1398 get_aperture_size(int fd)
1399 {
1400 struct drm_i915_gem_get_aperture aperture;
1401
1402 if (drmIoctl(fd, DRM_IOCTL_I915_GEM_GET_APERTURE, &aperture) != 0)
1403 return 0;
1404
1405 return aperture.aper_size;
1406 }
1407
1408 static int
1409 brw_query_renderer_integer(__DRIscreen *dri_screen,
1410 int param, unsigned int *value)
1411 {
1412 const struct intel_screen *const screen =
1413 (struct intel_screen *) dri_screen->driverPrivate;
1414
1415 switch (param) {
1416 case __DRI2_RENDERER_VENDOR_ID:
1417 value[0] = 0x8086;
1418 return 0;
1419 case __DRI2_RENDERER_DEVICE_ID:
1420 value[0] = screen->deviceID;
1421 return 0;
1422 case __DRI2_RENDERER_ACCELERATED:
1423 value[0] = 1;
1424 return 0;
1425 case __DRI2_RENDERER_VIDEO_MEMORY: {
1426 /* Once a batch uses more than 75% of the maximum mappable size, we
1427 * assume that there's some fragmentation, and we start doing extra
1428 * flushing, etc. That's the big cliff apps will care about.
1429 */
1430 const unsigned gpu_mappable_megabytes =
1431 screen->aperture_threshold / (1024 * 1024);
1432
1433 const long system_memory_pages = sysconf(_SC_PHYS_PAGES);
1434 const long system_page_size = sysconf(_SC_PAGE_SIZE);
1435
1436 if (system_memory_pages <= 0 || system_page_size <= 0)
1437 return -1;
1438
1439 const uint64_t system_memory_bytes = (uint64_t) system_memory_pages
1440 * (uint64_t) system_page_size;
1441
1442 const unsigned system_memory_megabytes =
1443 (unsigned) (system_memory_bytes / (1024 * 1024));
1444
1445 value[0] = MIN2(system_memory_megabytes, gpu_mappable_megabytes);
1446 return 0;
1447 }
1448 case __DRI2_RENDERER_UNIFIED_MEMORY_ARCHITECTURE:
1449 value[0] = 1;
1450 return 0;
1451 case __DRI2_RENDERER_HAS_TEXTURE_3D:
1452 value[0] = 1;
1453 return 0;
1454 case __DRI2_RENDERER_HAS_CONTEXT_PRIORITY:
1455 value[0] = 0;
1456 if (brw_hw_context_set_priority(screen->bufmgr,
1457 0, BRW_CONTEXT_HIGH_PRIORITY) == 0)
1458 value[0] |= __DRI2_RENDERER_HAS_CONTEXT_PRIORITY_HIGH;
1459 if (brw_hw_context_set_priority(screen->bufmgr,
1460 0, BRW_CONTEXT_LOW_PRIORITY) == 0)
1461 value[0] |= __DRI2_RENDERER_HAS_CONTEXT_PRIORITY_LOW;
1462 /* reset to default last, just in case */
1463 if (brw_hw_context_set_priority(screen->bufmgr,
1464 0, BRW_CONTEXT_MEDIUM_PRIORITY) == 0)
1465 value[0] |= __DRI2_RENDERER_HAS_CONTEXT_PRIORITY_MEDIUM;
1466 return 0;
1467 case __DRI2_RENDERER_HAS_FRAMEBUFFER_SRGB:
1468 value[0] = 1;
1469 return 0;
1470 default:
1471 return driQueryRendererIntegerCommon(dri_screen, param, value);
1472 }
1473
1474 return -1;
1475 }
1476
1477 static int
1478 brw_query_renderer_string(__DRIscreen *dri_screen,
1479 int param, const char **value)
1480 {
1481 const struct intel_screen *screen =
1482 (struct intel_screen *) dri_screen->driverPrivate;
1483
1484 switch (param) {
1485 case __DRI2_RENDERER_VENDOR_ID:
1486 value[0] = brw_vendor_string;
1487 return 0;
1488 case __DRI2_RENDERER_DEVICE_ID:
1489 value[0] = brw_get_renderer_string(screen);
1490 return 0;
1491 default:
1492 break;
1493 }
1494
1495 return -1;
1496 }
1497
1498 static void
1499 brw_set_cache_funcs(__DRIscreen *dri_screen,
1500 __DRIblobCacheSet set, __DRIblobCacheGet get)
1501 {
1502 const struct intel_screen *const screen =
1503 (struct intel_screen *) dri_screen->driverPrivate;
1504
1505 if (!screen->disk_cache)
1506 return;
1507
1508 disk_cache_set_callbacks(screen->disk_cache, set, get);
1509 }
1510
1511 static const __DRI2rendererQueryExtension intelRendererQueryExtension = {
1512 .base = { __DRI2_RENDERER_QUERY, 1 },
1513
1514 .queryInteger = brw_query_renderer_integer,
1515 .queryString = brw_query_renderer_string
1516 };
1517
1518 static const __DRIrobustnessExtension dri2Robustness = {
1519 .base = { __DRI2_ROBUSTNESS, 1 }
1520 };
1521
1522 static const __DRI2blobExtension intelBlobExtension = {
1523 .base = { __DRI2_BLOB, 1 },
1524 .set_cache_funcs = brw_set_cache_funcs
1525 };
1526
1527 static const __DRIextension *screenExtensions[] = {
1528 &intelTexBufferExtension.base,
1529 &intelFenceExtension.base,
1530 &intelFlushExtension.base,
1531 &intelImageExtension.base,
1532 &intelRendererQueryExtension.base,
1533 &dri2ConfigQueryExtension.base,
1534 &dri2NoErrorExtension.base,
1535 &intelBlobExtension.base,
1536 NULL
1537 };
1538
1539 static const __DRIextension *intelRobustScreenExtensions[] = {
1540 &intelTexBufferExtension.base,
1541 &intelFenceExtension.base,
1542 &intelFlushExtension.base,
1543 &intelImageExtension.base,
1544 &intelRendererQueryExtension.base,
1545 &dri2ConfigQueryExtension.base,
1546 &dri2Robustness.base,
1547 &dri2NoErrorExtension.base,
1548 &intelBlobExtension.base,
1549 NULL
1550 };
1551
1552 static int
1553 intel_get_param(struct intel_screen *screen, int param, int *value)
1554 {
1555 int ret = 0;
1556 struct drm_i915_getparam gp;
1557
1558 memset(&gp, 0, sizeof(gp));
1559 gp.param = param;
1560 gp.value = value;
1561
1562 if (drmIoctl(screen->driScrnPriv->fd, DRM_IOCTL_I915_GETPARAM, &gp) == -1) {
1563 ret = -errno;
1564 if (ret != -EINVAL)
1565 _mesa_warning(NULL, "drm_i915_getparam: %d", ret);
1566 }
1567
1568 return ret;
1569 }
1570
1571 static bool
1572 intel_get_boolean(struct intel_screen *screen, int param)
1573 {
1574 int value = 0;
1575 return (intel_get_param(screen, param, &value) == 0) && value;
1576 }
1577
1578 static int
1579 intel_get_integer(struct intel_screen *screen, int param)
1580 {
1581 int value = -1;
1582
1583 if (intel_get_param(screen, param, &value) == 0)
1584 return value;
1585
1586 return -1;
1587 }
1588
1589 static void
1590 intelDestroyScreen(__DRIscreen * sPriv)
1591 {
1592 struct intel_screen *screen = sPriv->driverPrivate;
1593
1594 brw_bufmgr_destroy(screen->bufmgr);
1595 driDestroyOptionInfo(&screen->optionCache);
1596
1597 disk_cache_destroy(screen->disk_cache);
1598
1599 ralloc_free(screen);
1600 sPriv->driverPrivate = NULL;
1601 }
1602
1603
1604 /**
1605 * Create a gl_framebuffer and attach it to __DRIdrawable::driverPrivate.
1606 *
1607 *_This implements driDriverAPI::createNewDrawable, which the DRI layer calls
1608 * when creating a EGLSurface, GLXDrawable, or GLXPixmap. Despite the name,
1609 * this does not allocate GPU memory.
1610 */
1611 static GLboolean
1612 intelCreateBuffer(__DRIscreen *dri_screen,
1613 __DRIdrawable * driDrawPriv,
1614 const struct gl_config * mesaVis, GLboolean isPixmap)
1615 {
1616 struct intel_renderbuffer *rb;
1617 struct intel_screen *screen = (struct intel_screen *)
1618 dri_screen->driverPrivate;
1619 mesa_format rgbFormat;
1620 unsigned num_samples =
1621 intel_quantize_num_samples(screen, mesaVis->samples);
1622
1623 if (isPixmap)
1624 return false;
1625
1626 struct gl_framebuffer *fb = CALLOC_STRUCT(gl_framebuffer);
1627 if (!fb)
1628 return false;
1629
1630 _mesa_initialize_window_framebuffer(fb, mesaVis);
1631
1632 if (screen->winsys_msaa_samples_override != -1) {
1633 num_samples = screen->winsys_msaa_samples_override;
1634 fb->Visual.samples = num_samples;
1635 }
1636
1637 if (mesaVis->redBits == 10 && mesaVis->alphaBits > 0) {
1638 rgbFormat = mesaVis->redMask == 0x3ff00000 ? MESA_FORMAT_B10G10R10A2_UNORM
1639 : MESA_FORMAT_R10G10B10A2_UNORM;
1640 } else if (mesaVis->redBits == 10) {
1641 rgbFormat = mesaVis->redMask == 0x3ff00000 ? MESA_FORMAT_B10G10R10X2_UNORM
1642 : MESA_FORMAT_R10G10B10X2_UNORM;
1643 } else if (mesaVis->redBits == 5) {
1644 rgbFormat = mesaVis->redMask == 0x1f ? MESA_FORMAT_R5G6B5_UNORM
1645 : MESA_FORMAT_B5G6R5_UNORM;
1646 } else if (mesaVis->sRGBCapable) {
1647 rgbFormat = mesaVis->redMask == 0xff ? MESA_FORMAT_R8G8B8A8_SRGB
1648 : MESA_FORMAT_B8G8R8A8_SRGB;
1649 } else if (mesaVis->alphaBits == 0) {
1650 rgbFormat = mesaVis->redMask == 0xff ? MESA_FORMAT_R8G8B8X8_UNORM
1651 : MESA_FORMAT_B8G8R8X8_UNORM;
1652 } else {
1653 rgbFormat = mesaVis->redMask == 0xff ? MESA_FORMAT_R8G8B8A8_SRGB
1654 : MESA_FORMAT_B8G8R8A8_SRGB;
1655 fb->Visual.sRGBCapable = true;
1656 }
1657
1658 /* mesaVis->sRGBCapable was set, user is asking for sRGB */
1659 bool srgb_cap_set = mesaVis->redBits >= 8 && mesaVis->sRGBCapable;
1660
1661 /* setup the hardware-based renderbuffers */
1662 rb = intel_create_winsys_renderbuffer(screen, rgbFormat, num_samples);
1663 _mesa_attach_and_own_rb(fb, BUFFER_FRONT_LEFT, &rb->Base.Base);
1664 rb->need_srgb = srgb_cap_set;
1665
1666 if (mesaVis->doubleBufferMode) {
1667 rb = intel_create_winsys_renderbuffer(screen, rgbFormat, num_samples);
1668 _mesa_attach_and_own_rb(fb, BUFFER_BACK_LEFT, &rb->Base.Base);
1669 rb->need_srgb = srgb_cap_set;
1670 }
1671
1672 /*
1673 * Assert here that the gl_config has an expected depth/stencil bit
1674 * combination: one of d24/s8, d16/s0, d0/s0. (See intelInitScreen2(),
1675 * which constructs the advertised configs.)
1676 */
1677 if (mesaVis->depthBits == 24) {
1678 assert(mesaVis->stencilBits == 8);
1679
1680 if (screen->devinfo.has_hiz_and_separate_stencil) {
1681 rb = intel_create_private_renderbuffer(screen,
1682 MESA_FORMAT_Z24_UNORM_X8_UINT,
1683 num_samples);
1684 _mesa_attach_and_own_rb(fb, BUFFER_DEPTH, &rb->Base.Base);
1685 rb = intel_create_private_renderbuffer(screen, MESA_FORMAT_S_UINT8,
1686 num_samples);
1687 _mesa_attach_and_own_rb(fb, BUFFER_STENCIL, &rb->Base.Base);
1688 } else {
1689 /*
1690 * Use combined depth/stencil. Note that the renderbuffer is
1691 * attached to two attachment points.
1692 */
1693 rb = intel_create_private_renderbuffer(screen,
1694 MESA_FORMAT_Z24_UNORM_S8_UINT,
1695 num_samples);
1696 _mesa_attach_and_own_rb(fb, BUFFER_DEPTH, &rb->Base.Base);
1697 _mesa_attach_and_reference_rb(fb, BUFFER_STENCIL, &rb->Base.Base);
1698 }
1699 }
1700 else if (mesaVis->depthBits == 16) {
1701 assert(mesaVis->stencilBits == 0);
1702 rb = intel_create_private_renderbuffer(screen, MESA_FORMAT_Z_UNORM16,
1703 num_samples);
1704 _mesa_attach_and_own_rb(fb, BUFFER_DEPTH, &rb->Base.Base);
1705 }
1706 else {
1707 assert(mesaVis->depthBits == 0);
1708 assert(mesaVis->stencilBits == 0);
1709 }
1710
1711 /* now add any/all software-based renderbuffers we may need */
1712 _swrast_add_soft_renderbuffers(fb,
1713 false, /* never sw color */
1714 false, /* never sw depth */
1715 false, /* never sw stencil */
1716 mesaVis->accumRedBits > 0,
1717 false, /* never sw alpha */
1718 false /* never sw aux */ );
1719 driDrawPriv->driverPrivate = fb;
1720
1721 return true;
1722 }
1723
1724 static void
1725 intelDestroyBuffer(__DRIdrawable * driDrawPriv)
1726 {
1727 struct gl_framebuffer *fb = driDrawPriv->driverPrivate;
1728
1729 _mesa_reference_framebuffer(&fb, NULL);
1730 }
1731
1732 static void
1733 intel_cs_timestamp_frequency(struct intel_screen *screen)
1734 {
1735 /* We shouldn't need to update gen_device_info.timestamp_frequency prior to
1736 * gen10, PCI-id is enough to figure it out.
1737 */
1738 assert(screen->devinfo.gen >= 10);
1739
1740 int ret, freq;
1741
1742 ret = intel_get_param(screen, I915_PARAM_CS_TIMESTAMP_FREQUENCY,
1743 &freq);
1744 if (ret < 0) {
1745 _mesa_warning(NULL,
1746 "Kernel 4.15 required to read the CS timestamp frequency.\n");
1747 return;
1748 }
1749
1750 screen->devinfo.timestamp_frequency = freq;
1751 }
1752
1753 static void
1754 intel_detect_sseu(struct intel_screen *screen)
1755 {
1756 assert(screen->devinfo.gen >= 8);
1757 int ret;
1758
1759 screen->subslice_total = -1;
1760 screen->eu_total = -1;
1761
1762 ret = intel_get_param(screen, I915_PARAM_SUBSLICE_TOTAL,
1763 &screen->subslice_total);
1764 if (ret < 0 && ret != -EINVAL)
1765 goto err_out;
1766
1767 ret = intel_get_param(screen,
1768 I915_PARAM_EU_TOTAL, &screen->eu_total);
1769 if (ret < 0 && ret != -EINVAL)
1770 goto err_out;
1771
1772 /* Without this information, we cannot get the right Braswell brandstrings,
1773 * and we have to use conservative numbers for GPGPU on many platforms, but
1774 * otherwise, things will just work.
1775 */
1776 if (screen->subslice_total < 1 || screen->eu_total < 1)
1777 _mesa_warning(NULL,
1778 "Kernel 4.1 required to properly query GPU properties.\n");
1779
1780 return;
1781
1782 err_out:
1783 screen->subslice_total = -1;
1784 screen->eu_total = -1;
1785 _mesa_warning(NULL, "Failed to query GPU properties (%s).\n", strerror(-ret));
1786 }
1787
1788 static bool
1789 intel_init_bufmgr(struct intel_screen *screen)
1790 {
1791 __DRIscreen *dri_screen = screen->driScrnPriv;
1792
1793 if (getenv("INTEL_NO_HW") != NULL)
1794 screen->no_hw = true;
1795
1796 screen->bufmgr = brw_bufmgr_init(&screen->devinfo, dri_screen->fd);
1797 if (screen->bufmgr == NULL) {
1798 fprintf(stderr, "[%s:%u] Error initializing buffer manager.\n",
1799 __func__, __LINE__);
1800 return false;
1801 }
1802
1803 if (!intel_get_boolean(screen, I915_PARAM_HAS_EXEC_NO_RELOC)) {
1804 fprintf(stderr, "[%s: %u] Kernel 3.9 required.\n", __func__, __LINE__);
1805 return false;
1806 }
1807
1808 return true;
1809 }
1810
1811 static bool
1812 intel_detect_swizzling(struct intel_screen *screen)
1813 {
1814 struct brw_bo *buffer;
1815 unsigned flags = 0;
1816 uint32_t aligned_pitch;
1817 uint32_t tiling = I915_TILING_X;
1818 uint32_t swizzle_mode = 0;
1819
1820 buffer = brw_bo_alloc_tiled_2d(screen->bufmgr, "swizzle test",
1821 64, 64, 4, tiling, &aligned_pitch, flags);
1822 if (buffer == NULL)
1823 return false;
1824
1825 brw_bo_get_tiling(buffer, &tiling, &swizzle_mode);
1826 brw_bo_unreference(buffer);
1827
1828 if (swizzle_mode == I915_BIT_6_SWIZZLE_NONE)
1829 return false;
1830 else
1831 return true;
1832 }
1833
1834 static int
1835 intel_detect_timestamp(struct intel_screen *screen)
1836 {
1837 uint64_t dummy = 0, last = 0;
1838 int upper, lower, loops;
1839
1840 /* On 64bit systems, some old kernels trigger a hw bug resulting in the
1841 * TIMESTAMP register being shifted and the low 32bits always zero.
1842 *
1843 * More recent kernels offer an interface to read the full 36bits
1844 * everywhere.
1845 */
1846 if (brw_reg_read(screen->bufmgr, TIMESTAMP | 1, &dummy) == 0)
1847 return 3;
1848
1849 /* Determine if we have a 32bit or 64bit kernel by inspecting the
1850 * upper 32bits for a rapidly changing timestamp.
1851 */
1852 if (brw_reg_read(screen->bufmgr, TIMESTAMP, &last))
1853 return 0;
1854
1855 upper = lower = 0;
1856 for (loops = 0; loops < 10; loops++) {
1857 /* The TIMESTAMP should change every 80ns, so several round trips
1858 * through the kernel should be enough to advance it.
1859 */
1860 if (brw_reg_read(screen->bufmgr, TIMESTAMP, &dummy))
1861 return 0;
1862
1863 upper += (dummy >> 32) != (last >> 32);
1864 if (upper > 1) /* beware 32bit counter overflow */
1865 return 2; /* upper dword holds the low 32bits of the timestamp */
1866
1867 lower += (dummy & 0xffffffff) != (last & 0xffffffff);
1868 if (lower > 1)
1869 return 1; /* timestamp is unshifted */
1870
1871 last = dummy;
1872 }
1873
1874 /* No advancement? No timestamp! */
1875 return 0;
1876 }
1877
1878 /**
1879 * Test if we can use MI_LOAD_REGISTER_MEM from an untrusted batchbuffer.
1880 *
1881 * Some combinations of hardware and kernel versions allow this feature,
1882 * while others don't. Instead of trying to enumerate every case, just
1883 * try and write a register and see if works.
1884 */
1885 static bool
1886 intel_detect_pipelined_register(struct intel_screen *screen,
1887 int reg, uint32_t expected_value, bool reset)
1888 {
1889 if (screen->no_hw)
1890 return false;
1891
1892 struct brw_bo *results, *bo;
1893 uint32_t *batch;
1894 uint32_t offset = 0;
1895 void *map;
1896 bool success = false;
1897
1898 /* Create a zero'ed temporary buffer for reading our results */
1899 results = brw_bo_alloc(screen->bufmgr, "registers", 4096, 0);
1900 if (results == NULL)
1901 goto err;
1902
1903 bo = brw_bo_alloc(screen->bufmgr, "batchbuffer", 4096, 0);
1904 if (bo == NULL)
1905 goto err_results;
1906
1907 map = brw_bo_map(NULL, bo, MAP_WRITE);
1908 if (!map)
1909 goto err_batch;
1910
1911 batch = map;
1912
1913 /* Write the register. */
1914 *batch++ = MI_LOAD_REGISTER_IMM | (3 - 2);
1915 *batch++ = reg;
1916 *batch++ = expected_value;
1917
1918 /* Save the register's value back to the buffer. */
1919 *batch++ = MI_STORE_REGISTER_MEM | (3 - 2);
1920 *batch++ = reg;
1921 struct drm_i915_gem_relocation_entry reloc = {
1922 .offset = (char *) batch - (char *) map,
1923 .delta = offset * sizeof(uint32_t),
1924 .target_handle = results->gem_handle,
1925 .read_domains = I915_GEM_DOMAIN_INSTRUCTION,
1926 .write_domain = I915_GEM_DOMAIN_INSTRUCTION,
1927 };
1928 *batch++ = reloc.presumed_offset + reloc.delta;
1929
1930 /* And afterwards clear the register */
1931 if (reset) {
1932 *batch++ = MI_LOAD_REGISTER_IMM | (3 - 2);
1933 *batch++ = reg;
1934 *batch++ = 0;
1935 }
1936
1937 *batch++ = MI_BATCH_BUFFER_END;
1938
1939 struct drm_i915_gem_exec_object2 exec_objects[2] = {
1940 {
1941 .handle = results->gem_handle,
1942 },
1943 {
1944 .handle = bo->gem_handle,
1945 .relocation_count = 1,
1946 .relocs_ptr = (uintptr_t) &reloc,
1947 }
1948 };
1949
1950 struct drm_i915_gem_execbuffer2 execbuf = {
1951 .buffers_ptr = (uintptr_t) exec_objects,
1952 .buffer_count = 2,
1953 .batch_len = ALIGN((char *) batch - (char *) map, 8),
1954 .flags = I915_EXEC_RENDER,
1955 };
1956
1957 /* Don't bother with error checking - if the execbuf fails, the
1958 * value won't be written and we'll just report that there's no access.
1959 */
1960 __DRIscreen *dri_screen = screen->driScrnPriv;
1961 drmIoctl(dri_screen->fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
1962
1963 /* Check whether the value got written. */
1964 void *results_map = brw_bo_map(NULL, results, MAP_READ);
1965 if (results_map) {
1966 success = *((uint32_t *)results_map + offset) == expected_value;
1967 brw_bo_unmap(results);
1968 }
1969
1970 err_batch:
1971 brw_bo_unreference(bo);
1972 err_results:
1973 brw_bo_unreference(results);
1974 err:
1975 return success;
1976 }
1977
1978 static bool
1979 intel_detect_pipelined_so(struct intel_screen *screen)
1980 {
1981 const struct gen_device_info *devinfo = &screen->devinfo;
1982
1983 /* Supposedly, Broadwell just works. */
1984 if (devinfo->gen >= 8)
1985 return true;
1986
1987 if (devinfo->gen <= 6)
1988 return false;
1989
1990 /* See the big explanation about command parser versions below */
1991 if (screen->cmd_parser_version >= (devinfo->is_haswell ? 7 : 2))
1992 return true;
1993
1994 /* We use SO_WRITE_OFFSET0 since you're supposed to write it (unlike the
1995 * statistics registers), and we already reset it to zero before using it.
1996 */
1997 return intel_detect_pipelined_register(screen,
1998 GEN7_SO_WRITE_OFFSET(0),
1999 0x1337d0d0,
2000 false);
2001 }
2002
2003 /**
2004 * Return array of MSAA modes supported by the hardware. The array is
2005 * zero-terminated and sorted in decreasing order.
2006 */
2007 const int*
2008 intel_supported_msaa_modes(const struct intel_screen *screen)
2009 {
2010 static const int gen9_modes[] = {16, 8, 4, 2, 0, -1};
2011 static const int gen8_modes[] = {8, 4, 2, 0, -1};
2012 static const int gen7_modes[] = {8, 4, 0, -1};
2013 static const int gen6_modes[] = {4, 0, -1};
2014 static const int gen4_modes[] = {0, -1};
2015
2016 if (screen->devinfo.gen >= 9) {
2017 return gen9_modes;
2018 } else if (screen->devinfo.gen >= 8) {
2019 return gen8_modes;
2020 } else if (screen->devinfo.gen >= 7) {
2021 return gen7_modes;
2022 } else if (screen->devinfo.gen == 6) {
2023 return gen6_modes;
2024 } else {
2025 return gen4_modes;
2026 }
2027 }
2028
2029 static unsigned
2030 intel_loader_get_cap(const __DRIscreen *dri_screen, enum dri_loader_cap cap)
2031 {
2032 if (dri_screen->dri2.loader && dri_screen->dri2.loader->base.version >= 4 &&
2033 dri_screen->dri2.loader->getCapability)
2034 return dri_screen->dri2.loader->getCapability(dri_screen->loaderPrivate, cap);
2035
2036 if (dri_screen->image.loader && dri_screen->image.loader->base.version >= 2 &&
2037 dri_screen->image.loader->getCapability)
2038 return dri_screen->image.loader->getCapability(dri_screen->loaderPrivate, cap);
2039
2040 return 0;
2041 }
2042
2043 static __DRIconfig**
2044 intel_screen_make_configs(__DRIscreen *dri_screen)
2045 {
2046 static const mesa_format formats[] = {
2047 MESA_FORMAT_B5G6R5_UNORM,
2048 MESA_FORMAT_B8G8R8A8_UNORM,
2049 MESA_FORMAT_B8G8R8X8_UNORM,
2050
2051 MESA_FORMAT_B8G8R8A8_SRGB,
2052
2053 /* For 10 bpc, 30 bit depth framebuffers. */
2054 MESA_FORMAT_B10G10R10A2_UNORM,
2055 MESA_FORMAT_B10G10R10X2_UNORM,
2056
2057 /* The 32-bit RGBA format must not precede the 32-bit BGRA format.
2058 * Likewise for RGBX and BGRX. Otherwise, the GLX client and the GLX
2059 * server may disagree on which format the GLXFBConfig represents,
2060 * resulting in swapped color channels.
2061 *
2062 * The problem, as of 2017-05-30:
2063 * When matching a GLXFBConfig to a __DRIconfig, GLX ignores the channel
2064 * order and chooses the first __DRIconfig with the expected channel
2065 * sizes. Specifically, GLX compares the GLXFBConfig's and __DRIconfig's
2066 * __DRI_ATTRIB_{CHANNEL}_SIZE but ignores __DRI_ATTRIB_{CHANNEL}_MASK.
2067 *
2068 * EGL does not suffer from this problem. It correctly compares the
2069 * channel masks when matching EGLConfig to __DRIconfig.
2070 */
2071
2072 /* Required by Android, for HAL_PIXEL_FORMAT_RGBA_8888. */
2073 MESA_FORMAT_R8G8B8A8_UNORM,
2074
2075 /* Required by Android, for HAL_PIXEL_FORMAT_RGBX_8888. */
2076 MESA_FORMAT_R8G8B8X8_UNORM,
2077 };
2078
2079 /* GLX_SWAP_COPY_OML is not supported due to page flipping. */
2080 static const GLenum back_buffer_modes[] = {
2081 __DRI_ATTRIB_SWAP_UNDEFINED, __DRI_ATTRIB_SWAP_NONE
2082 };
2083
2084 static const uint8_t singlesample_samples[1] = {0};
2085
2086 struct intel_screen *screen = dri_screen->driverPrivate;
2087 const struct gen_device_info *devinfo = &screen->devinfo;
2088 uint8_t depth_bits[4], stencil_bits[4];
2089 __DRIconfig **configs = NULL;
2090
2091 /* Expose only BGRA ordering if the loader doesn't support RGBA ordering. */
2092 unsigned num_formats;
2093 if (intel_loader_get_cap(dri_screen, DRI_LOADER_CAP_RGBA_ORDERING))
2094 num_formats = ARRAY_SIZE(formats);
2095 else
2096 num_formats = ARRAY_SIZE(formats) - 2; /* all - RGBA_ORDERING formats */
2097
2098 /* Shall we expose 10 bpc formats? */
2099 bool allow_rgb10_configs = driQueryOptionb(&screen->optionCache,
2100 "allow_rgb10_configs");
2101
2102 /* Generate singlesample configs without accumulation buffer. */
2103 for (unsigned i = 0; i < num_formats; i++) {
2104 __DRIconfig **new_configs;
2105 int num_depth_stencil_bits = 2;
2106
2107 if (!allow_rgb10_configs &&
2108 (formats[i] == MESA_FORMAT_B10G10R10A2_UNORM ||
2109 formats[i] == MESA_FORMAT_B10G10R10X2_UNORM))
2110 continue;
2111
2112 /* Starting with DRI2 protocol version 1.1 we can request a depth/stencil
2113 * buffer that has a different number of bits per pixel than the color
2114 * buffer, gen >= 6 supports this.
2115 */
2116 depth_bits[0] = 0;
2117 stencil_bits[0] = 0;
2118
2119 if (formats[i] == MESA_FORMAT_B5G6R5_UNORM) {
2120 depth_bits[1] = 16;
2121 stencil_bits[1] = 0;
2122 if (devinfo->gen >= 6) {
2123 depth_bits[2] = 24;
2124 stencil_bits[2] = 8;
2125 num_depth_stencil_bits = 3;
2126 }
2127 } else {
2128 depth_bits[1] = 24;
2129 stencil_bits[1] = 8;
2130 }
2131
2132 new_configs = driCreateConfigs(formats[i],
2133 depth_bits,
2134 stencil_bits,
2135 num_depth_stencil_bits,
2136 back_buffer_modes, 2,
2137 singlesample_samples, 1,
2138 false, false);
2139 configs = driConcatConfigs(configs, new_configs);
2140 }
2141
2142 /* Generate the minimum possible set of configs that include an
2143 * accumulation buffer.
2144 */
2145 for (unsigned i = 0; i < num_formats; i++) {
2146 __DRIconfig **new_configs;
2147
2148 if (!allow_rgb10_configs &&
2149 (formats[i] == MESA_FORMAT_B10G10R10A2_UNORM ||
2150 formats[i] == MESA_FORMAT_B10G10R10X2_UNORM))
2151 continue;
2152
2153 if (formats[i] == MESA_FORMAT_B5G6R5_UNORM) {
2154 depth_bits[0] = 16;
2155 stencil_bits[0] = 0;
2156 } else {
2157 depth_bits[0] = 24;
2158 stencil_bits[0] = 8;
2159 }
2160
2161 new_configs = driCreateConfigs(formats[i],
2162 depth_bits, stencil_bits, 1,
2163 back_buffer_modes, 1,
2164 singlesample_samples, 1,
2165 true, false);
2166 configs = driConcatConfigs(configs, new_configs);
2167 }
2168
2169 /* Generate multisample configs.
2170 *
2171 * This loop breaks early, and hence is a no-op, on gen < 6.
2172 *
2173 * Multisample configs must follow the singlesample configs in order to
2174 * work around an X server bug present in 1.12. The X server chooses to
2175 * associate the first listed RGBA888-Z24S8 config, regardless of its
2176 * sample count, with the 32-bit depth visual used for compositing.
2177 *
2178 * Only doublebuffer configs with GLX_SWAP_UNDEFINED_OML behavior are
2179 * supported. Singlebuffer configs are not supported because no one wants
2180 * them.
2181 */
2182 for (unsigned i = 0; i < num_formats; i++) {
2183 if (devinfo->gen < 6)
2184 break;
2185
2186 if (!allow_rgb10_configs &&
2187 (formats[i] == MESA_FORMAT_B10G10R10A2_UNORM ||
2188 formats[i] == MESA_FORMAT_B10G10R10X2_UNORM))
2189 continue;
2190
2191 __DRIconfig **new_configs;
2192 const int num_depth_stencil_bits = 2;
2193 int num_msaa_modes = 0;
2194 const uint8_t *multisample_samples = NULL;
2195
2196 depth_bits[0] = 0;
2197 stencil_bits[0] = 0;
2198
2199 if (formats[i] == MESA_FORMAT_B5G6R5_UNORM) {
2200 depth_bits[1] = 16;
2201 stencil_bits[1] = 0;
2202 } else {
2203 depth_bits[1] = 24;
2204 stencil_bits[1] = 8;
2205 }
2206
2207 if (devinfo->gen >= 9) {
2208 static const uint8_t multisample_samples_gen9[] = {2, 4, 8, 16};
2209 multisample_samples = multisample_samples_gen9;
2210 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen9);
2211 } else if (devinfo->gen == 8) {
2212 static const uint8_t multisample_samples_gen8[] = {2, 4, 8};
2213 multisample_samples = multisample_samples_gen8;
2214 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen8);
2215 } else if (devinfo->gen == 7) {
2216 static const uint8_t multisample_samples_gen7[] = {4, 8};
2217 multisample_samples = multisample_samples_gen7;
2218 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen7);
2219 } else if (devinfo->gen == 6) {
2220 static const uint8_t multisample_samples_gen6[] = {4};
2221 multisample_samples = multisample_samples_gen6;
2222 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen6);
2223 }
2224
2225 new_configs = driCreateConfigs(formats[i],
2226 depth_bits,
2227 stencil_bits,
2228 num_depth_stencil_bits,
2229 back_buffer_modes, 1,
2230 multisample_samples,
2231 num_msaa_modes,
2232 false, false);
2233 configs = driConcatConfigs(configs, new_configs);
2234 }
2235
2236 if (configs == NULL) {
2237 fprintf(stderr, "[%s:%u] Error creating FBConfig!\n", __func__,
2238 __LINE__);
2239 return NULL;
2240 }
2241
2242 return configs;
2243 }
2244
2245 static void
2246 set_max_gl_versions(struct intel_screen *screen)
2247 {
2248 __DRIscreen *dri_screen = screen->driScrnPriv;
2249 const bool has_astc = screen->devinfo.gen >= 9;
2250
2251 switch (screen->devinfo.gen) {
2252 case 10:
2253 case 9:
2254 case 8:
2255 dri_screen->max_gl_core_version = 45;
2256 dri_screen->max_gl_compat_version = 30;
2257 dri_screen->max_gl_es1_version = 11;
2258 dri_screen->max_gl_es2_version = has_astc ? 32 : 31;
2259 break;
2260 case 7:
2261 dri_screen->max_gl_core_version = 33;
2262 if (can_do_pipelined_register_writes(screen)) {
2263 dri_screen->max_gl_core_version = 42;
2264 if (screen->devinfo.is_haswell && can_do_compute_dispatch(screen))
2265 dri_screen->max_gl_core_version = 43;
2266 if (screen->devinfo.is_haswell && can_do_mi_math_and_lrr(screen))
2267 dri_screen->max_gl_core_version = 45;
2268 }
2269 dri_screen->max_gl_compat_version = 30;
2270 dri_screen->max_gl_es1_version = 11;
2271 dri_screen->max_gl_es2_version = screen->devinfo.is_haswell ? 31 : 30;
2272 break;
2273 case 6:
2274 dri_screen->max_gl_core_version = 33;
2275 dri_screen->max_gl_compat_version = 30;
2276 dri_screen->max_gl_es1_version = 11;
2277 dri_screen->max_gl_es2_version = 30;
2278 break;
2279 case 5:
2280 case 4:
2281 dri_screen->max_gl_core_version = 0;
2282 dri_screen->max_gl_compat_version = 21;
2283 dri_screen->max_gl_es1_version = 11;
2284 dri_screen->max_gl_es2_version = 20;
2285 break;
2286 default:
2287 unreachable("unrecognized intel_screen::gen");
2288 }
2289 }
2290
2291 /**
2292 * Return the revision (generally the revid field of the PCI header) of the
2293 * graphics device.
2294 */
2295 int
2296 intel_device_get_revision(int fd)
2297 {
2298 struct drm_i915_getparam gp;
2299 int revision;
2300 int ret;
2301
2302 memset(&gp, 0, sizeof(gp));
2303 gp.param = I915_PARAM_REVISION;
2304 gp.value = &revision;
2305
2306 ret = drmCommandWriteRead(fd, DRM_I915_GETPARAM, &gp, sizeof(gp));
2307 if (ret)
2308 revision = -1;
2309
2310 return revision;
2311 }
2312
2313 static void
2314 shader_debug_log_mesa(void *data, const char *fmt, ...)
2315 {
2316 struct brw_context *brw = (struct brw_context *)data;
2317 va_list args;
2318
2319 va_start(args, fmt);
2320 GLuint msg_id = 0;
2321 _mesa_gl_vdebug(&brw->ctx, &msg_id,
2322 MESA_DEBUG_SOURCE_SHADER_COMPILER,
2323 MESA_DEBUG_TYPE_OTHER,
2324 MESA_DEBUG_SEVERITY_NOTIFICATION, fmt, args);
2325 va_end(args);
2326 }
2327
2328 static void
2329 shader_perf_log_mesa(void *data, const char *fmt, ...)
2330 {
2331 struct brw_context *brw = (struct brw_context *)data;
2332
2333 va_list args;
2334 va_start(args, fmt);
2335
2336 if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
2337 va_list args_copy;
2338 va_copy(args_copy, args);
2339 vfprintf(stderr, fmt, args_copy);
2340 va_end(args_copy);
2341 }
2342
2343 if (brw->perf_debug) {
2344 GLuint msg_id = 0;
2345 _mesa_gl_vdebug(&brw->ctx, &msg_id,
2346 MESA_DEBUG_SOURCE_SHADER_COMPILER,
2347 MESA_DEBUG_TYPE_PERFORMANCE,
2348 MESA_DEBUG_SEVERITY_MEDIUM, fmt, args);
2349 }
2350 va_end(args);
2351 }
2352
2353 static int
2354 parse_devid_override(const char *devid_override)
2355 {
2356 static const struct {
2357 const char *name;
2358 int pci_id;
2359 } name_map[] = {
2360 { "brw", 0x2a02 },
2361 { "g4x", 0x2a42 },
2362 { "ilk", 0x0042 },
2363 { "snb", 0x0126 },
2364 { "ivb", 0x016a },
2365 { "hsw", 0x0d2e },
2366 { "byt", 0x0f33 },
2367 { "bdw", 0x162e },
2368 { "chv", 0x22B3 },
2369 { "skl", 0x1912 },
2370 { "bxt", 0x5A85 },
2371 { "kbl", 0x5912 },
2372 { "glk", 0x3185 },
2373 { "cnl", 0x5a52 },
2374 };
2375
2376 for (unsigned i = 0; i < ARRAY_SIZE(name_map); i++) {
2377 if (!strcmp(name_map[i].name, devid_override))
2378 return name_map[i].pci_id;
2379 }
2380
2381 return strtol(devid_override, NULL, 0);
2382 }
2383
2384 /**
2385 * Get the PCI ID for the device. This can be overridden by setting the
2386 * INTEL_DEVID_OVERRIDE environment variable to the desired ID.
2387 *
2388 * Returns -1 on ioctl failure.
2389 */
2390 static int
2391 get_pci_device_id(struct intel_screen *screen)
2392 {
2393 if (geteuid() == getuid()) {
2394 char *devid_override = getenv("INTEL_DEVID_OVERRIDE");
2395 if (devid_override) {
2396 screen->no_hw = true;
2397 return parse_devid_override(devid_override);
2398 }
2399 }
2400
2401 return intel_get_integer(screen, I915_PARAM_CHIPSET_ID);
2402 }
2403
2404 /**
2405 * This is the driver specific part of the createNewScreen entry point.
2406 * Called when using DRI2.
2407 *
2408 * \return the struct gl_config supported by this driver
2409 */
2410 static const
2411 __DRIconfig **intelInitScreen2(__DRIscreen *dri_screen)
2412 {
2413 struct intel_screen *screen;
2414
2415 if (dri_screen->image.loader) {
2416 } else if (dri_screen->dri2.loader->base.version <= 2 ||
2417 dri_screen->dri2.loader->getBuffersWithFormat == NULL) {
2418 fprintf(stderr,
2419 "\nERROR! DRI2 loader with getBuffersWithFormat() "
2420 "support required\n");
2421 return NULL;
2422 }
2423
2424 /* Allocate the private area */
2425 screen = rzalloc(NULL, struct intel_screen);
2426 if (!screen) {
2427 fprintf(stderr, "\nERROR! Allocating private area failed\n");
2428 return NULL;
2429 }
2430 /* parse information in __driConfigOptions */
2431 driOptionCache options;
2432 memset(&options, 0, sizeof(options));
2433
2434 driParseOptionInfo(&options, brw_config_options.xml);
2435 driParseConfigFiles(&screen->optionCache, &options, dri_screen->myNum, "i965");
2436 driDestroyOptionCache(&options);
2437
2438 screen->driScrnPriv = dri_screen;
2439 dri_screen->driverPrivate = (void *) screen;
2440
2441 screen->deviceID = get_pci_device_id(screen);
2442
2443 if (!gen_get_device_info(screen->deviceID, &screen->devinfo))
2444 return NULL;
2445
2446 if (!intel_init_bufmgr(screen))
2447 return NULL;
2448
2449 const struct gen_device_info *devinfo = &screen->devinfo;
2450
2451 brw_process_intel_debug_variable();
2452
2453 if ((INTEL_DEBUG & DEBUG_SHADER_TIME) && devinfo->gen < 7) {
2454 fprintf(stderr,
2455 "shader_time debugging requires gen7 (Ivybridge) or better.\n");
2456 INTEL_DEBUG &= ~DEBUG_SHADER_TIME;
2457 }
2458
2459 if (intel_get_integer(screen, I915_PARAM_MMAP_GTT_VERSION) >= 1) {
2460 /* Theorectically unlimited! At least for individual objects...
2461 *
2462 * Currently the entire (global) address space for all GTT maps is
2463 * limited to 64bits. That is all objects on the system that are
2464 * setup for GTT mmapping must fit within 64bits. An attempt to use
2465 * one that exceeds the limit with fail in brw_bo_map_gtt().
2466 *
2467 * Long before we hit that limit, we will be practically limited by
2468 * that any single object must fit in physical memory (RAM). The upper
2469 * limit on the CPU's address space is currently 48bits (Skylake), of
2470 * which only 39bits can be physical memory. (The GPU itself also has
2471 * a 48bit addressable virtual space.) We can fit over 32 million
2472 * objects of the current maximum allocable size before running out
2473 * of mmap space.
2474 */
2475 screen->max_gtt_map_object_size = UINT64_MAX;
2476 } else {
2477 /* Estimate the size of the mappable aperture into the GTT. There's an
2478 * ioctl to get the whole GTT size, but not one to get the mappable subset.
2479 * It turns out it's basically always 256MB, though some ancient hardware
2480 * was smaller.
2481 */
2482 uint32_t gtt_size = 256 * 1024 * 1024;
2483
2484 /* We don't want to map two objects such that a memcpy between them would
2485 * just fault one mapping in and then the other over and over forever. So
2486 * we would need to divide the GTT size by 2. Additionally, some GTT is
2487 * taken up by things like the framebuffer and the ringbuffer and such, so
2488 * be more conservative.
2489 */
2490 screen->max_gtt_map_object_size = gtt_size / 4;
2491 }
2492
2493 screen->aperture_threshold = get_aperture_size(dri_screen->fd) * 3 / 4;
2494
2495 screen->hw_has_swizzling = intel_detect_swizzling(screen);
2496 screen->hw_has_timestamp = intel_detect_timestamp(screen);
2497
2498 isl_device_init(&screen->isl_dev, &screen->devinfo,
2499 screen->hw_has_swizzling);
2500
2501 if (devinfo->gen >= 10)
2502 intel_cs_timestamp_frequency(screen);
2503
2504 /* GENs prior to 8 do not support EU/Subslice info */
2505 if (devinfo->gen >= 8) {
2506 intel_detect_sseu(screen);
2507 } else if (devinfo->gen == 7) {
2508 screen->subslice_total = 1 << (devinfo->gt - 1);
2509 }
2510
2511 /* Gen7-7.5 kernel requirements / command parser saga:
2512 *
2513 * - pre-v3.16:
2514 * Haswell and Baytrail cannot use any privileged batchbuffer features.
2515 *
2516 * Ivybridge has aliasing PPGTT on by default, which accidentally marks
2517 * all batches secure, allowing them to use any feature with no checking.
2518 * This is effectively equivalent to a command parser version of
2519 * \infinity - everything is possible.
2520 *
2521 * The command parser does not exist, and querying the version will
2522 * return -EINVAL.
2523 *
2524 * - v3.16:
2525 * The kernel enables the command parser by default, for systems with
2526 * aliasing PPGTT enabled (Ivybridge and Haswell). However, the
2527 * hardware checker is still enabled, so Haswell and Baytrail cannot
2528 * do anything.
2529 *
2530 * Ivybridge goes from "everything is possible" to "only what the
2531 * command parser allows" (if the user boots with i915.cmd_parser=0,
2532 * then everything is possible again). We can only safely use features
2533 * allowed by the supported command parser version.
2534 *
2535 * Annoyingly, I915_PARAM_CMD_PARSER_VERSION reports the static version
2536 * implemented by the kernel, even if it's turned off. So, checking
2537 * for version > 0 does not mean that you can write registers. We have
2538 * to try it and see. The version does, however, indicate the age of
2539 * the kernel.
2540 *
2541 * Instead of matching the hardware checker's behavior of converting
2542 * privileged commands to MI_NOOP, it makes execbuf2 start returning
2543 * -EINVAL, making it dangerous to try and use privileged features.
2544 *
2545 * Effective command parser versions:
2546 * - Haswell: 0 (reporting 1, writes don't work)
2547 * - Baytrail: 0 (reporting 1, writes don't work)
2548 * - Ivybridge: 1 (enabled) or infinite (disabled)
2549 *
2550 * - v3.17:
2551 * Baytrail aliasing PPGTT is enabled, making it like Ivybridge:
2552 * effectively version 1 (enabled) or infinite (disabled).
2553 *
2554 * - v3.19: f1f55cc0556031c8ee3fe99dae7251e78b9b653b
2555 * Command parser v2 supports predicate writes.
2556 *
2557 * - Haswell: 0 (reporting 1, writes don't work)
2558 * - Baytrail: 2 (enabled) or infinite (disabled)
2559 * - Ivybridge: 2 (enabled) or infinite (disabled)
2560 *
2561 * So version >= 2 is enough to know that Ivybridge and Baytrail
2562 * will work. Haswell still can't do anything.
2563 *
2564 * - v4.0: Version 3 happened. Largely not relevant.
2565 *
2566 * - v4.1: 6702cf16e0ba8b0129f5aa1b6609d4e9c70bc13b
2567 * L3 config registers are properly saved and restored as part
2568 * of the hardware context. We can approximately detect this point
2569 * in time by checking if I915_PARAM_REVISION is recognized - it
2570 * landed in a later commit, but in the same release cycle.
2571 *
2572 * - v4.2: 245054a1fe33c06ad233e0d58a27ec7b64db9284
2573 * Command parser finally gains secure batch promotion. On Haswell,
2574 * the hardware checker gets disabled, which finally allows it to do
2575 * privileged commands.
2576 *
2577 * I915_PARAM_CMD_PARSER_VERSION reports 3. Effective versions:
2578 * - Haswell: 3 (enabled) or 0 (disabled)
2579 * - Baytrail: 3 (enabled) or infinite (disabled)
2580 * - Ivybridge: 3 (enabled) or infinite (disabled)
2581 *
2582 * Unfortunately, detecting this point in time is tricky, because
2583 * no version bump happened when this important change occurred.
2584 * On Haswell, if we can write any register, then the kernel is at
2585 * least this new, and we can start trusting the version number.
2586 *
2587 * - v4.4: 2bbe6bbb0dc94fd4ce287bdac9e1bd184e23057b and
2588 * Command parser reaches version 4, allowing access to Haswell
2589 * atomic scratch and chicken3 registers. If version >= 4, we know
2590 * the kernel is new enough to support privileged features on all
2591 * hardware. However, the user might have disabled it...and the
2592 * kernel will still report version 4. So we still have to guess
2593 * and check.
2594 *
2595 * - v4.4: 7b9748cb513a6bef4af87b79f0da3ff7e8b56cd8
2596 * Command parser v5 whitelists indirect compute shader dispatch
2597 * registers, needed for OpenGL 4.3 and later.
2598 *
2599 * - v4.8:
2600 * Command parser v7 lets us use MI_MATH on Haswell.
2601 *
2602 * Additionally, the kernel begins reporting version 0 when
2603 * the command parser is disabled, allowing us to skip the
2604 * guess-and-check step on Haswell. Unfortunately, this also
2605 * means that we can no longer use it as an indicator of the
2606 * age of the kernel.
2607 */
2608 if (intel_get_param(screen, I915_PARAM_CMD_PARSER_VERSION,
2609 &screen->cmd_parser_version) < 0) {
2610 /* Command parser does not exist - getparam is unrecognized */
2611 screen->cmd_parser_version = 0;
2612 }
2613
2614 /* Kernel 4.13 retuired for exec object capture */
2615 if (intel_get_boolean(screen, I915_PARAM_HAS_EXEC_CAPTURE)) {
2616 screen->kernel_features |= KERNEL_ALLOWS_EXEC_CAPTURE;
2617 }
2618
2619 if (intel_get_boolean(screen, I915_PARAM_HAS_EXEC_BATCH_FIRST)) {
2620 screen->kernel_features |= KERNEL_ALLOWS_EXEC_BATCH_FIRST;
2621 }
2622
2623 if (!intel_detect_pipelined_so(screen)) {
2624 /* We can't do anything, so the effective version is 0. */
2625 screen->cmd_parser_version = 0;
2626 } else {
2627 screen->kernel_features |= KERNEL_ALLOWS_SOL_OFFSET_WRITES;
2628 }
2629
2630 if (devinfo->gen >= 8 || screen->cmd_parser_version >= 2)
2631 screen->kernel_features |= KERNEL_ALLOWS_PREDICATE_WRITES;
2632
2633 /* Haswell requires command parser version 4 in order to have L3
2634 * atomic scratch1 and chicken3 bits
2635 */
2636 if (devinfo->is_haswell && screen->cmd_parser_version >= 4) {
2637 screen->kernel_features |=
2638 KERNEL_ALLOWS_HSW_SCRATCH1_AND_ROW_CHICKEN3;
2639 }
2640
2641 /* Haswell requires command parser version 6 in order to write to the
2642 * MI_MATH GPR registers, and version 7 in order to use
2643 * MI_LOAD_REGISTER_REG (which all users of MI_MATH use).
2644 */
2645 if (devinfo->gen >= 8 ||
2646 (devinfo->is_haswell && screen->cmd_parser_version >= 7)) {
2647 screen->kernel_features |= KERNEL_ALLOWS_MI_MATH_AND_LRR;
2648 }
2649
2650 /* Gen7 needs at least command parser version 5 to support compute */
2651 if (devinfo->gen >= 8 || screen->cmd_parser_version >= 5)
2652 screen->kernel_features |= KERNEL_ALLOWS_COMPUTE_DISPATCH;
2653
2654 const char *force_msaa = getenv("INTEL_FORCE_MSAA");
2655 if (force_msaa) {
2656 screen->winsys_msaa_samples_override =
2657 intel_quantize_num_samples(screen, atoi(force_msaa));
2658 printf("Forcing winsys sample count to %d\n",
2659 screen->winsys_msaa_samples_override);
2660 } else {
2661 screen->winsys_msaa_samples_override = -1;
2662 }
2663
2664 set_max_gl_versions(screen);
2665
2666 /* Notification of GPU resets requires hardware contexts and a kernel new
2667 * enough to support DRM_IOCTL_I915_GET_RESET_STATS. If the ioctl is
2668 * supported, calling it with a context of 0 will either generate EPERM or
2669 * no error. If the ioctl is not supported, it always generate EINVAL.
2670 * Use this to determine whether to advertise the __DRI2_ROBUSTNESS
2671 * extension to the loader.
2672 *
2673 * Don't even try on pre-Gen6, since we don't attempt to use contexts there.
2674 */
2675 if (devinfo->gen >= 6) {
2676 struct drm_i915_reset_stats stats;
2677 memset(&stats, 0, sizeof(stats));
2678
2679 const int ret = drmIoctl(dri_screen->fd, DRM_IOCTL_I915_GET_RESET_STATS, &stats);
2680
2681 screen->has_context_reset_notification =
2682 (ret != -1 || errno != EINVAL);
2683 }
2684
2685 dri_screen->extensions = !screen->has_context_reset_notification
2686 ? screenExtensions : intelRobustScreenExtensions;
2687
2688 screen->compiler = brw_compiler_create(screen, devinfo);
2689 screen->compiler->shader_debug_log = shader_debug_log_mesa;
2690 screen->compiler->shader_perf_log = shader_perf_log_mesa;
2691 screen->compiler->constant_buffer_0_is_relative = true;
2692 screen->compiler->supports_pull_constants = true;
2693
2694 screen->has_exec_fence =
2695 intel_get_boolean(screen, I915_PARAM_HAS_EXEC_FENCE);
2696
2697 intel_screen_init_surface_formats(screen);
2698
2699 if (INTEL_DEBUG & (DEBUG_BATCH | DEBUG_SUBMIT)) {
2700 unsigned int caps = intel_get_integer(screen, I915_PARAM_HAS_SCHEDULER);
2701 if (caps) {
2702 fprintf(stderr, "Kernel scheduler detected: %08x\n", caps);
2703 if (caps & I915_SCHEDULER_CAP_PRIORITY)
2704 fprintf(stderr, " - User priority sorting enabled\n");
2705 if (caps & I915_SCHEDULER_CAP_PREEMPTION)
2706 fprintf(stderr, " - Preemption enabled\n");
2707 }
2708 }
2709
2710 brw_disk_cache_init(screen);
2711
2712 return (const __DRIconfig**) intel_screen_make_configs(dri_screen);
2713 }
2714
2715 struct intel_buffer {
2716 __DRIbuffer base;
2717 struct brw_bo *bo;
2718 };
2719
2720 static __DRIbuffer *
2721 intelAllocateBuffer(__DRIscreen *dri_screen,
2722 unsigned attachment, unsigned format,
2723 int width, int height)
2724 {
2725 struct intel_buffer *intelBuffer;
2726 struct intel_screen *screen = dri_screen->driverPrivate;
2727
2728 assert(attachment == __DRI_BUFFER_FRONT_LEFT ||
2729 attachment == __DRI_BUFFER_BACK_LEFT);
2730
2731 intelBuffer = calloc(1, sizeof *intelBuffer);
2732 if (intelBuffer == NULL)
2733 return NULL;
2734
2735 /* The front and back buffers are color buffers, which are X tiled. GEN9+
2736 * supports Y tiled and compressed buffers, but there is no way to plumb that
2737 * through to here. */
2738 uint32_t pitch;
2739 int cpp = format / 8;
2740 intelBuffer->bo = brw_bo_alloc_tiled_2d(screen->bufmgr,
2741 "intelAllocateBuffer",
2742 width,
2743 height,
2744 cpp,
2745 I915_TILING_X, &pitch,
2746 BO_ALLOC_BUSY);
2747
2748 if (intelBuffer->bo == NULL) {
2749 free(intelBuffer);
2750 return NULL;
2751 }
2752
2753 brw_bo_flink(intelBuffer->bo, &intelBuffer->base.name);
2754
2755 intelBuffer->base.attachment = attachment;
2756 intelBuffer->base.cpp = cpp;
2757 intelBuffer->base.pitch = pitch;
2758
2759 return &intelBuffer->base;
2760 }
2761
2762 static void
2763 intelReleaseBuffer(__DRIscreen *dri_screen, __DRIbuffer *buffer)
2764 {
2765 struct intel_buffer *intelBuffer = (struct intel_buffer *) buffer;
2766
2767 brw_bo_unreference(intelBuffer->bo);
2768 free(intelBuffer);
2769 }
2770
2771 static const struct __DriverAPIRec brw_driver_api = {
2772 .InitScreen = intelInitScreen2,
2773 .DestroyScreen = intelDestroyScreen,
2774 .CreateContext = brwCreateContext,
2775 .DestroyContext = intelDestroyContext,
2776 .CreateBuffer = intelCreateBuffer,
2777 .DestroyBuffer = intelDestroyBuffer,
2778 .MakeCurrent = intelMakeCurrent,
2779 .UnbindContext = intelUnbindContext,
2780 .AllocateBuffer = intelAllocateBuffer,
2781 .ReleaseBuffer = intelReleaseBuffer
2782 };
2783
2784 static const struct __DRIDriverVtableExtensionRec brw_vtable = {
2785 .base = { __DRI_DRIVER_VTABLE, 1 },
2786 .vtable = &brw_driver_api,
2787 };
2788
2789 static const __DRIextension *brw_driver_extensions[] = {
2790 &driCoreExtension.base,
2791 &driImageDriverExtension.base,
2792 &driDRI2Extension.base,
2793 &brw_vtable.base,
2794 &brw_config_options.base,
2795 NULL
2796 };
2797
2798 PUBLIC const __DRIextension **__driDriverGetExtensions_i965(void)
2799 {
2800 globalDriverAPI = &brw_driver_api;
2801
2802 return brw_driver_extensions;
2803 }