src/mesa/drivers/dri/intel/intel_buffer_objects.c

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